diff options
| author | trav90 <travawine@palemoon.org> | 2018-10-15 21:45:30 -0500 |
|---|---|---|
| committer | trav90 <travawine@palemoon.org> | 2018-10-15 21:45:30 -0500 |
| commit | 68569dee1416593955c1570d638b3d9250b33012 (patch) | |
| tree | d960f017cd7eba3f125b7e8a813789ee2e076310 /third_party/aom/aom_dsp/x86 | |
| parent | 07c17b6b98ed32fcecff15c083ab0fd878de3cf0 (diff) | |
| download | UXP-68569dee1416593955c1570d638b3d9250b33012.tar UXP-68569dee1416593955c1570d638b3d9250b33012.tar.gz UXP-68569dee1416593955c1570d638b3d9250b33012.tar.lz UXP-68569dee1416593955c1570d638b3d9250b33012.tar.xz UXP-68569dee1416593955c1570d638b3d9250b33012.zip | |
Import aom library
This is the reference implementation for the Alliance for Open Media's av1 video code.
The commit used was 4d668d7feb1f8abd809d1bca0418570a7f142a36.
Diffstat (limited to 'third_party/aom/aom_dsp/x86')
74 files changed, 45176 insertions, 0 deletions
diff --git a/third_party/aom/aom_dsp/x86/aom_asm_stubs.c b/third_party/aom/aom_dsp/x86/aom_asm_stubs.c new file mode 100644 index 000000000..4067b0b53 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/aom_asm_stubs.c @@ -0,0 +1,182 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include "./aom_config.h" +#include "./aom_dsp_rtcd.h" +#include "aom_dsp/x86/convolve.h" + +#if HAVE_SSE2 +filter8_1dfunction aom_filter_block1d16_v8_sse2; +filter8_1dfunction aom_filter_block1d16_h8_sse2; +filter8_1dfunction aom_filter_block1d8_v8_sse2; +filter8_1dfunction aom_filter_block1d8_h8_sse2; +filter8_1dfunction aom_filter_block1d4_v8_sse2; +filter8_1dfunction aom_filter_block1d4_h8_sse2; +filter8_1dfunction aom_filter_block1d16_v8_avg_sse2; +filter8_1dfunction aom_filter_block1d16_h8_avg_sse2; +filter8_1dfunction aom_filter_block1d8_v8_avg_sse2; +filter8_1dfunction aom_filter_block1d8_h8_avg_sse2; +filter8_1dfunction aom_filter_block1d4_v8_avg_sse2; +filter8_1dfunction aom_filter_block1d4_h8_avg_sse2; + +filter8_1dfunction aom_filter_block1d16_v2_sse2; +filter8_1dfunction aom_filter_block1d16_h2_sse2; +filter8_1dfunction aom_filter_block1d8_v2_sse2; +filter8_1dfunction aom_filter_block1d8_h2_sse2; +filter8_1dfunction aom_filter_block1d4_v2_sse2; +filter8_1dfunction aom_filter_block1d4_h2_sse2; +filter8_1dfunction aom_filter_block1d16_v2_avg_sse2; +filter8_1dfunction aom_filter_block1d16_h2_avg_sse2; +filter8_1dfunction aom_filter_block1d8_v2_avg_sse2; +filter8_1dfunction aom_filter_block1d8_h2_avg_sse2; +filter8_1dfunction aom_filter_block1d4_v2_avg_sse2; +filter8_1dfunction aom_filter_block1d4_h2_avg_sse2; + +// void aom_convolve8_horiz_sse2(const uint8_t *src, ptrdiff_t src_stride, +// uint8_t *dst, ptrdiff_t dst_stride, +// const int16_t *filter_x, int x_step_q4, +// const int16_t *filter_y, int y_step_q4, +// int w, int h); +// void aom_convolve8_vert_sse2(const uint8_t *src, ptrdiff_t src_stride, +// uint8_t *dst, ptrdiff_t dst_stride, +// const int16_t *filter_x, int x_step_q4, +// const int16_t *filter_y, int y_step_q4, +// int w, int h); +// void aom_convolve8_avg_horiz_sse2(const uint8_t *src, ptrdiff_t src_stride, +// uint8_t *dst, ptrdiff_t dst_stride, +// const int16_t *filter_x, int x_step_q4, +// const int16_t *filter_y, int y_step_q4, +// int w, int h); +// void aom_convolve8_avg_vert_sse2(const uint8_t *src, ptrdiff_t src_stride, +// uint8_t *dst, ptrdiff_t dst_stride, +// const int16_t *filter_x, int x_step_q4, +// const int16_t *filter_y, int y_step_q4, +// int w, int h); +FUN_CONV_1D(horiz, x_step_q4, filter_x, h, src, , sse2); +FUN_CONV_1D(vert, y_step_q4, filter_y, v, src - src_stride * 3, , sse2); +FUN_CONV_1D(avg_horiz, x_step_q4, filter_x, h, src, avg_, sse2); +FUN_CONV_1D(avg_vert, y_step_q4, filter_y, v, src - src_stride * 3, avg_, sse2); + +// void aom_convolve8_sse2(const uint8_t *src, ptrdiff_t src_stride, +// uint8_t *dst, ptrdiff_t dst_stride, +// const int16_t *filter_x, int x_step_q4, +// const int16_t *filter_y, int y_step_q4, +// int w, int h); +// void aom_convolve8_avg_sse2(const uint8_t *src, ptrdiff_t src_stride, +// uint8_t *dst, ptrdiff_t dst_stride, +// const int16_t *filter_x, int x_step_q4, +// const int16_t *filter_y, int y_step_q4, +// int w, int h); +FUN_CONV_2D(, sse2); +FUN_CONV_2D(avg_, sse2); + +#if CONFIG_HIGHBITDEPTH && ARCH_X86_64 +highbd_filter8_1dfunction aom_highbd_filter_block1d16_v8_sse2; +highbd_filter8_1dfunction aom_highbd_filter_block1d16_h8_sse2; +highbd_filter8_1dfunction aom_highbd_filter_block1d8_v8_sse2; +highbd_filter8_1dfunction aom_highbd_filter_block1d8_h8_sse2; +highbd_filter8_1dfunction aom_highbd_filter_block1d4_v8_sse2; +highbd_filter8_1dfunction aom_highbd_filter_block1d4_h8_sse2; +highbd_filter8_1dfunction aom_highbd_filter_block1d16_v8_avg_sse2; +highbd_filter8_1dfunction aom_highbd_filter_block1d16_h8_avg_sse2; +highbd_filter8_1dfunction aom_highbd_filter_block1d8_v8_avg_sse2; +highbd_filter8_1dfunction aom_highbd_filter_block1d8_h8_avg_sse2; +highbd_filter8_1dfunction aom_highbd_filter_block1d4_v8_avg_sse2; +highbd_filter8_1dfunction aom_highbd_filter_block1d4_h8_avg_sse2; + +highbd_filter8_1dfunction aom_highbd_filter_block1d16_v2_sse2; +highbd_filter8_1dfunction aom_highbd_filter_block1d16_h2_sse2; +highbd_filter8_1dfunction aom_highbd_filter_block1d8_v2_sse2; +highbd_filter8_1dfunction aom_highbd_filter_block1d8_h2_sse2; +highbd_filter8_1dfunction aom_highbd_filter_block1d4_v2_sse2; +highbd_filter8_1dfunction aom_highbd_filter_block1d4_h2_sse2; +highbd_filter8_1dfunction aom_highbd_filter_block1d16_v2_avg_sse2; +highbd_filter8_1dfunction aom_highbd_filter_block1d16_h2_avg_sse2; +highbd_filter8_1dfunction aom_highbd_filter_block1d8_v2_avg_sse2; +highbd_filter8_1dfunction aom_highbd_filter_block1d8_h2_avg_sse2; +highbd_filter8_1dfunction aom_highbd_filter_block1d4_v2_avg_sse2; +highbd_filter8_1dfunction aom_highbd_filter_block1d4_h2_avg_sse2; + +// void aom_highbd_convolve8_horiz_sse2(const uint8_t *src, +// ptrdiff_t src_stride, +// uint8_t *dst, +// ptrdiff_t dst_stride, +// const int16_t *filter_x, +// int x_step_q4, +// const int16_t *filter_y, +// int y_step_q4, +// int w, int h, int bd); +// void aom_highbd_convolve8_vert_sse2(const uint8_t *src, +// ptrdiff_t src_stride, +// uint8_t *dst, +// ptrdiff_t dst_stride, +// const int16_t *filter_x, +// int x_step_q4, +// const int16_t *filter_y, +// int y_step_q4, +// int w, int h, int bd); +// void aom_highbd_convolve8_avg_horiz_sse2(const uint8_t *src, +// ptrdiff_t src_stride, +// uint8_t *dst, +// ptrdiff_t dst_stride, +// const int16_t *filter_x, +// int x_step_q4, +// const int16_t *filter_y, +// int y_step_q4, +// int w, int h, int bd); +// void aom_highbd_convolve8_avg_vert_sse2(const uint8_t *src, +// ptrdiff_t src_stride, +// uint8_t *dst, +// ptrdiff_t dst_stride, +// const int16_t *filter_x, +// int x_step_q4, +// const int16_t *filter_y, +// int y_step_q4, +// int w, int h, int bd); +HIGH_FUN_CONV_1D(horiz, x_step_q4, filter_x, h, src, , sse2); +HIGH_FUN_CONV_1D(vert, y_step_q4, filter_y, v, src - src_stride * 3, , sse2); +HIGH_FUN_CONV_1D(avg_horiz, x_step_q4, filter_x, h, src, avg_, sse2); +HIGH_FUN_CONV_1D(avg_vert, y_step_q4, filter_y, v, src - src_stride * 3, avg_, + sse2); + +// void aom_highbd_convolve8_sse2(const uint8_t *src, ptrdiff_t src_stride, +// uint8_t *dst, ptrdiff_t dst_stride, +// const int16_t *filter_x, int x_step_q4, +// const int16_t *filter_y, int y_step_q4, +// int w, int h, int bd); +// void aom_highbd_convolve8_avg_sse2(const uint8_t *src, ptrdiff_t src_stride, +// uint8_t *dst, ptrdiff_t dst_stride, +// const int16_t *filter_x, int x_step_q4, +// const int16_t *filter_y, int y_step_q4, +// int w, int h, int bd); +HIGH_FUN_CONV_2D(, sse2); +HIGH_FUN_CONV_2D(avg_, sse2); + +#if CONFIG_LOOP_RESTORATION +// The SSE2 highbd convolve functions can deal with coefficients up to 32767. +// So redirect highbd_convolve8_add_src to regular highbd_convolve8. +void aom_highbd_convolve8_add_src_sse2(const uint8_t *src, ptrdiff_t src_stride, + uint8_t *dst, ptrdiff_t dst_stride, + const int16_t *filter_x, int x_step_q4, + const int16_t *filter_y, int y_step_q4, + int w, int h, int bd) { + assert(x_step_q4 == 16); + assert(y_step_q4 == 16); + ((int16_t *)filter_x)[3] += 128; + ((int16_t *)filter_y)[3] += 128; + aom_highbd_convolve8_sse2(src, src_stride, dst, dst_stride, filter_x, + x_step_q4, filter_y, y_step_q4, w, h, bd); + ((int16_t *)filter_x)[3] -= 128; + ((int16_t *)filter_y)[3] -= 128; +} +#endif // CONFIG_LOOP_RESTORATION +#endif // CONFIG_HIGHBITDEPTH && ARCH_X86_64 +#endif // HAVE_SSE2 diff --git a/third_party/aom/aom_dsp/x86/aom_convolve_copy_sse2.asm b/third_party/aom/aom_dsp/x86/aom_convolve_copy_sse2.asm new file mode 100644 index 000000000..4d3142867 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/aom_convolve_copy_sse2.asm @@ -0,0 +1,345 @@ +; +; Copyright (c) 2016, Alliance for Open Media. All rights reserved +; +; This source code is subject to the terms of the BSD 2 Clause License and +; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License +; was not distributed with this source code in the LICENSE file, you can +; obtain it at www.aomedia.org/license/software. If the Alliance for Open +; Media Patent License 1.0 was not distributed with this source code in the +; PATENTS file, you can obtain it at www.aomedia.org/license/patent. +; + +; + +%include "third_party/x86inc/x86inc.asm" + +SECTION .text + +%macro convolve_fn 1-2 +%ifidn %1, avg +%define AUX_XMM_REGS 4 +%else +%define AUX_XMM_REGS 0 +%endif +%ifidn %2, highbd +%define pavg pavgw +cglobal %2_convolve_%1, 4, 7, 4+AUX_XMM_REGS, src, src_stride, \ + dst, dst_stride, \ + fx, fxs, fy, fys, w, h, bd +%else +%define pavg pavgb +cglobal convolve_%1, 4, 7, 4+AUX_XMM_REGS, src, src_stride, \ + dst, dst_stride, \ + fx, fxs, fy, fys, w, h +%endif + mov r4d, dword wm +%ifidn %2, highbd + shl r4d, 1 + shl srcq, 1 + shl src_strideq, 1 + shl dstq, 1 + shl dst_strideq, 1 +%else + cmp r4d, 4 + je .w4 +%endif + cmp r4d, 8 + je .w8 + cmp r4d, 16 + je .w16 + cmp r4d, 32 + je .w32 + +%if CONFIG_AV1 && CONFIG_EXT_PARTITION + cmp r4d, 64 + je .w64 +%ifidn %2, highbd + cmp r4d, 128 + je .w128 + +.w256: + mov r4d, dword hm +.loop256: + movu m0, [srcq] + movu m1, [srcq+16] + movu m2, [srcq+32] + movu m3, [srcq+48] +%ifidn %1, avg + pavg m0, [dstq] + pavg m1, [dstq+16] + pavg m2, [dstq+32] + pavg m3, [dstq+48] +%endif + mova [dstq ], m0 + mova [dstq+16], m1 + mova [dstq+32], m2 + mova [dstq+48], m3 + movu m0, [srcq+64] + movu m1, [srcq+80] + movu m2, [srcq+96] + movu m3, [srcq+112] +%ifidn %1, avg + pavg m0, [dstq+64] + pavg m1, [dstq+80] + pavg m2, [dstq+96] + pavg m3, [dstq+112] +%endif + mova [dstq+64], m0 + mova [dstq+80], m1 + mova [dstq+96], m2 + mova [dstq+112], m3 + movu m0, [srcq+128] + movu m1, [srcq+128+16] + movu m2, [srcq+128+32] + movu m3, [srcq+128+48] +%ifidn %1, avg + pavg m0, [dstq+128] + pavg m1, [dstq+128+16] + pavg m2, [dstq+128+32] + pavg m3, [dstq+128+48] +%endif + mova [dstq+128 ], m0 + mova [dstq+128+16], m1 + mova [dstq+128+32], m2 + mova [dstq+128+48], m3 + movu m0, [srcq+128+64] + movu m1, [srcq+128+80] + movu m2, [srcq+128+96] + movu m3, [srcq+128+112] + add srcq, src_strideq +%ifidn %1, avg + pavg m0, [dstq+128+64] + pavg m1, [dstq+128+80] + pavg m2, [dstq+128+96] + pavg m3, [dstq+128+112] +%endif + mova [dstq+128+64], m0 + mova [dstq+128+80], m1 + mova [dstq+128+96], m2 + mova [dstq+128+112], m3 + add dstq, dst_strideq + sub r4d, 1 + jnz .loop256 + RET +%endif + +.w128: + mov r4d, dword hm +.loop128: + movu m0, [srcq] + movu m1, [srcq+16] + movu m2, [srcq+32] + movu m3, [srcq+48] +%ifidn %1, avg + pavg m0, [dstq] + pavg m1, [dstq+16] + pavg m2, [dstq+32] + pavg m3, [dstq+48] +%endif + mova [dstq ], m0 + mova [dstq+16], m1 + mova [dstq+32], m2 + mova [dstq+48], m3 + movu m0, [srcq+64] + movu m1, [srcq+80] + movu m2, [srcq+96] + movu m3, [srcq+112] + add srcq, src_strideq +%ifidn %1, avg + pavg m0, [dstq+64] + pavg m1, [dstq+80] + pavg m2, [dstq+96] + pavg m3, [dstq+112] +%endif + mova [dstq+64], m0 + mova [dstq+80], m1 + mova [dstq+96], m2 + mova [dstq+112], m3 + add dstq, dst_strideq + sub r4d, 1 + jnz .loop128 + RET + +%else ; CONFIG_AV1 && CONFIG_EXT_PARTITION + +%ifidn %2, highbd + cmp r4d, 64 + je .w64 + + mov r4d, dword hm +.loop128: + movu m0, [srcq] + movu m1, [srcq+16] + movu m2, [srcq+32] + movu m3, [srcq+48] +%ifidn %1, avg + pavg m0, [dstq] + pavg m1, [dstq+16] + pavg m2, [dstq+32] + pavg m3, [dstq+48] +%endif + mova [dstq ], m0 + mova [dstq+16], m1 + mova [dstq+32], m2 + mova [dstq+48], m3 + movu m0, [srcq+64] + movu m1, [srcq+80] + movu m2, [srcq+96] + movu m3, [srcq+112] + add srcq, src_strideq +%ifidn %1, avg + pavg m0, [dstq+64] + pavg m1, [dstq+80] + pavg m2, [dstq+96] + pavg m3, [dstq+112] +%endif + mova [dstq+64], m0 + mova [dstq+80], m1 + mova [dstq+96], m2 + mova [dstq+112], m3 + add dstq, dst_strideq + sub r4d, 1 + jnz .loop128 + RET +%endif +%endif ; CONFIG_AV1 && CONFIG_EXT_PARTITION + +.w64: + mov r4d, dword hm +.loop64: + movu m0, [srcq] + movu m1, [srcq+16] + movu m2, [srcq+32] + movu m3, [srcq+48] + add srcq, src_strideq +%ifidn %1, avg + pavg m0, [dstq] + pavg m1, [dstq+16] + pavg m2, [dstq+32] + pavg m3, [dstq+48] +%endif + mova [dstq ], m0 + mova [dstq+16], m1 + mova [dstq+32], m2 + mova [dstq+48], m3 + add dstq, dst_strideq + sub r4d, 1 + jnz .loop64 + RET + +.w32: + mov r4d, dword hm +.loop32: + movu m0, [srcq] + movu m1, [srcq+16] + movu m2, [srcq+src_strideq] + movu m3, [srcq+src_strideq+16] + lea srcq, [srcq+src_strideq*2] +%ifidn %1, avg + pavg m0, [dstq] + pavg m1, [dstq +16] + pavg m2, [dstq+dst_strideq] + pavg m3, [dstq+dst_strideq+16] +%endif + mova [dstq ], m0 + mova [dstq +16], m1 + mova [dstq+dst_strideq ], m2 + mova [dstq+dst_strideq+16], m3 + lea dstq, [dstq+dst_strideq*2] + sub r4d, 2 + jnz .loop32 + RET + +.w16: + mov r4d, dword hm + lea r5q, [src_strideq*3] + lea r6q, [dst_strideq*3] +.loop16: + movu m0, [srcq] + movu m1, [srcq+src_strideq] + movu m2, [srcq+src_strideq*2] + movu m3, [srcq+r5q] + lea srcq, [srcq+src_strideq*4] +%ifidn %1, avg + pavg m0, [dstq] + pavg m1, [dstq+dst_strideq] + pavg m2, [dstq+dst_strideq*2] + pavg m3, [dstq+r6q] +%endif + mova [dstq ], m0 + mova [dstq+dst_strideq ], m1 + mova [dstq+dst_strideq*2], m2 + mova [dstq+r6q ], m3 + lea dstq, [dstq+dst_strideq*4] + sub r4d, 4 + jnz .loop16 + RET + +.w8: + mov r4d, dword hm + lea r5q, [src_strideq*3] + lea r6q, [dst_strideq*3] +.loop8: + movh m0, [srcq] + movh m1, [srcq+src_strideq] + movh m2, [srcq+src_strideq*2] + movh m3, [srcq+r5q] + lea srcq, [srcq+src_strideq*4] +%ifidn %1, avg + movh m4, [dstq] + movh m5, [dstq+dst_strideq] + movh m6, [dstq+dst_strideq*2] + movh m7, [dstq+r6q] + pavg m0, m4 + pavg m1, m5 + pavg m2, m6 + pavg m3, m7 +%endif + movh [dstq ], m0 + movh [dstq+dst_strideq ], m1 + movh [dstq+dst_strideq*2], m2 + movh [dstq+r6q ], m3 + lea dstq, [dstq+dst_strideq*4] + sub r4d, 4 + jnz .loop8 + RET + +%ifnidn %2, highbd +.w4: + mov r4d, dword hm + lea r5q, [src_strideq*3] + lea r6q, [dst_strideq*3] +.loop4: + movd m0, [srcq] + movd m1, [srcq+src_strideq] + movd m2, [srcq+src_strideq*2] + movd m3, [srcq+r5q] + lea srcq, [srcq+src_strideq*4] +%ifidn %1, avg + movd m4, [dstq] + movd m5, [dstq+dst_strideq] + movd m6, [dstq+dst_strideq*2] + movd m7, [dstq+r6q] + pavg m0, m4 + pavg m1, m5 + pavg m2, m6 + pavg m3, m7 +%endif + movd [dstq ], m0 + movd [dstq+dst_strideq ], m1 + movd [dstq+dst_strideq*2], m2 + movd [dstq+r6q ], m3 + lea dstq, [dstq+dst_strideq*4] + sub r4d, 4 + jnz .loop4 + RET +%endif +%endmacro + +INIT_XMM sse2 +convolve_fn copy +convolve_fn avg +%if CONFIG_HIGHBITDEPTH +convolve_fn copy, highbd +convolve_fn avg, highbd +%endif diff --git a/third_party/aom/aom_dsp/x86/aom_high_subpixel_8t_sse2.asm b/third_party/aom/aom_dsp/x86/aom_high_subpixel_8t_sse2.asm new file mode 100644 index 000000000..e6d357ba3 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/aom_high_subpixel_8t_sse2.asm @@ -0,0 +1,965 @@ +; +; Copyright (c) 2016, Alliance for Open Media. All rights reserved +; +; This source code is subject to the terms of the BSD 2 Clause License and +; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License +; was not distributed with this source code in the LICENSE file, you can +; obtain it at www.aomedia.org/license/software. If the Alliance for Open +; Media Patent License 1.0 was not distributed with this source code in the +; PATENTS file, you can obtain it at www.aomedia.org/license/patent. +; + +; + + +%include "aom_ports/x86_abi_support.asm" + +;Note: tap3 and tap4 have to be applied and added after other taps to avoid +;overflow. + +%macro HIGH_GET_FILTERS_4 0 + mov rdx, arg(5) ;filter ptr + mov rcx, 0x00000040 + + movdqa xmm7, [rdx] ;load filters + pshuflw xmm0, xmm7, 0b ;k0 + pshuflw xmm1, xmm7, 01010101b ;k1 + pshuflw xmm2, xmm7, 10101010b ;k2 + pshuflw xmm3, xmm7, 11111111b ;k3 + psrldq xmm7, 8 + pshuflw xmm4, xmm7, 0b ;k4 + pshuflw xmm5, xmm7, 01010101b ;k5 + pshuflw xmm6, xmm7, 10101010b ;k6 + pshuflw xmm7, xmm7, 11111111b ;k7 + + punpcklwd xmm0, xmm6 + punpcklwd xmm2, xmm5 + punpcklwd xmm3, xmm4 + punpcklwd xmm1, xmm7 + + movdqa k0k6, xmm0 + movdqa k2k5, xmm2 + movdqa k3k4, xmm3 + movdqa k1k7, xmm1 + + movq xmm6, rcx + pshufd xmm6, xmm6, 0 + movdqa krd, xmm6 + + ;Compute max and min values of a pixel + mov rdx, 0x00010001 + movsxd rcx, DWORD PTR arg(6) ;bps + movq xmm0, rdx + movq xmm1, rcx + pshufd xmm0, xmm0, 0b + movdqa xmm2, xmm0 + psllw xmm0, xmm1 + psubw xmm0, xmm2 + pxor xmm1, xmm1 + movdqa max, xmm0 ;max value (for clamping) + movdqa min, xmm1 ;min value (for clamping) + +%endm + +%macro HIGH_APPLY_FILTER_4 1 + punpcklwd xmm0, xmm6 ;two row in one register + punpcklwd xmm1, xmm7 + punpcklwd xmm2, xmm5 + punpcklwd xmm3, xmm4 + + pmaddwd xmm0, k0k6 ;multiply the filter factors + pmaddwd xmm1, k1k7 + pmaddwd xmm2, k2k5 + pmaddwd xmm3, k3k4 + + paddd xmm0, xmm1 ;sum + paddd xmm0, xmm2 + paddd xmm0, xmm3 + + paddd xmm0, krd ;rounding + psrad xmm0, 7 ;shift + packssdw xmm0, xmm0 ;pack to word + + ;clamp the values + pminsw xmm0, max + pmaxsw xmm0, min + +%if %1 + movq xmm1, [rdi] + pavgw xmm0, xmm1 +%endif + movq [rdi], xmm0 +%endm + +%macro HIGH_GET_FILTERS 0 + mov rdx, arg(5) ;filter ptr + mov rsi, arg(0) ;src_ptr + mov rdi, arg(2) ;output_ptr + mov rcx, 0x00000040 + + movdqa xmm7, [rdx] ;load filters + pshuflw xmm0, xmm7, 0b ;k0 + pshuflw xmm1, xmm7, 01010101b ;k1 + pshuflw xmm2, xmm7, 10101010b ;k2 + pshuflw xmm3, xmm7, 11111111b ;k3 + pshufhw xmm4, xmm7, 0b ;k4 + pshufhw xmm5, xmm7, 01010101b ;k5 + pshufhw xmm6, xmm7, 10101010b ;k6 + pshufhw xmm7, xmm7, 11111111b ;k7 + punpcklqdq xmm2, xmm2 + punpcklqdq xmm3, xmm3 + punpcklwd xmm0, xmm1 + punpckhwd xmm6, xmm7 + punpckhwd xmm2, xmm5 + punpckhwd xmm3, xmm4 + + movdqa k0k1, xmm0 ;store filter factors on stack + movdqa k6k7, xmm6 + movdqa k2k5, xmm2 + movdqa k3k4, xmm3 + + movq xmm6, rcx + pshufd xmm6, xmm6, 0 + movdqa krd, xmm6 ;rounding + + ;Compute max and min values of a pixel + mov rdx, 0x00010001 + movsxd rcx, DWORD PTR arg(6) ;bps + movq xmm0, rdx + movq xmm1, rcx + pshufd xmm0, xmm0, 0b + movdqa xmm2, xmm0 + psllw xmm0, xmm1 + psubw xmm0, xmm2 + pxor xmm1, xmm1 + movdqa max, xmm0 ;max value (for clamping) + movdqa min, xmm1 ;min value (for clamping) +%endm + +%macro LOAD_VERT_8 1 + movdqu xmm0, [rsi + %1] ;0 + movdqu xmm1, [rsi + rax + %1] ;1 + movdqu xmm6, [rsi + rdx * 2 + %1] ;6 + lea rsi, [rsi + rax] + movdqu xmm7, [rsi + rdx * 2 + %1] ;7 + movdqu xmm2, [rsi + rax + %1] ;2 + movdqu xmm3, [rsi + rax * 2 + %1] ;3 + movdqu xmm4, [rsi + rdx + %1] ;4 + movdqu xmm5, [rsi + rax * 4 + %1] ;5 +%endm + +%macro HIGH_APPLY_FILTER_8 2 + movdqu temp, xmm4 + movdqa xmm4, xmm0 + punpcklwd xmm0, xmm1 + punpckhwd xmm4, xmm1 + movdqa xmm1, xmm6 + punpcklwd xmm6, xmm7 + punpckhwd xmm1, xmm7 + movdqa xmm7, xmm2 + punpcklwd xmm2, xmm5 + punpckhwd xmm7, xmm5 + + movdqu xmm5, temp + movdqu temp, xmm4 + movdqa xmm4, xmm3 + punpcklwd xmm3, xmm5 + punpckhwd xmm4, xmm5 + movdqu xmm5, temp + + pmaddwd xmm0, k0k1 + pmaddwd xmm5, k0k1 + pmaddwd xmm6, k6k7 + pmaddwd xmm1, k6k7 + pmaddwd xmm2, k2k5 + pmaddwd xmm7, k2k5 + pmaddwd xmm3, k3k4 + pmaddwd xmm4, k3k4 + + paddd xmm0, xmm6 + paddd xmm0, xmm2 + paddd xmm0, xmm3 + paddd xmm5, xmm1 + paddd xmm5, xmm7 + paddd xmm5, xmm4 + + paddd xmm0, krd ;rounding + paddd xmm5, krd + psrad xmm0, 7 ;shift + psrad xmm5, 7 + packssdw xmm0, xmm5 ;pack back to word + + ;clamp the values + pminsw xmm0, max + pmaxsw xmm0, min + +%if %1 + movdqu xmm1, [rdi + %2] + pavgw xmm0, xmm1 +%endif + movdqu [rdi + %2], xmm0 +%endm + +;void aom_filter_block1d4_v8_sse2 +;( +; unsigned char *src_ptr, +; unsigned int src_pitch, +; unsigned char *output_ptr, +; unsigned int out_pitch, +; unsigned int output_height, +; short *filter +;) +global sym(aom_highbd_filter_block1d4_v8_sse2) PRIVATE +sym(aom_highbd_filter_block1d4_v8_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 7 + SAVE_XMM 7 + push rsi + push rdi + push rbx + ; end prolog + + ALIGN_STACK 16, rax + sub rsp, 16 * 7 + %define k0k6 [rsp + 16 * 0] + %define k2k5 [rsp + 16 * 1] + %define k3k4 [rsp + 16 * 2] + %define k1k7 [rsp + 16 * 3] + %define krd [rsp + 16 * 4] + %define max [rsp + 16 * 5] + %define min [rsp + 16 * 6] + + HIGH_GET_FILTERS_4 + + mov rsi, arg(0) ;src_ptr + mov rdi, arg(2) ;output_ptr + + movsxd rax, DWORD PTR arg(1) ;pixels_per_line + movsxd rbx, DWORD PTR arg(3) ;out_pitch + lea rax, [rax + rax] ;bytes per line + lea rbx, [rbx + rbx] + lea rdx, [rax + rax * 2] + movsxd rcx, DWORD PTR arg(4) ;output_height + +.loop: + movq xmm0, [rsi] ;load src: row 0 + movq xmm1, [rsi + rax] ;1 + movq xmm6, [rsi + rdx * 2] ;6 + lea rsi, [rsi + rax] + movq xmm7, [rsi + rdx * 2] ;7 + movq xmm2, [rsi + rax] ;2 + movq xmm3, [rsi + rax * 2] ;3 + movq xmm4, [rsi + rdx] ;4 + movq xmm5, [rsi + rax * 4] ;5 + + HIGH_APPLY_FILTER_4 0 + + lea rdi, [rdi + rbx] + dec rcx + jnz .loop + + add rsp, 16 * 7 + pop rsp + pop rbx + ; begin epilog + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret + +;void aom_filter_block1d8_v8_sse2 +;( +; unsigned char *src_ptr, +; unsigned int src_pitch, +; unsigned char *output_ptr, +; unsigned int out_pitch, +; unsigned int output_height, +; short *filter +;) +global sym(aom_highbd_filter_block1d8_v8_sse2) PRIVATE +sym(aom_highbd_filter_block1d8_v8_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 7 + SAVE_XMM 7 + push rsi + push rdi + push rbx + ; end prolog + + ALIGN_STACK 16, rax + sub rsp, 16 * 8 + %define k0k1 [rsp + 16 * 0] + %define k6k7 [rsp + 16 * 1] + %define k2k5 [rsp + 16 * 2] + %define k3k4 [rsp + 16 * 3] + %define krd [rsp + 16 * 4] + %define temp [rsp + 16 * 5] + %define max [rsp + 16 * 6] + %define min [rsp + 16 * 7] + + HIGH_GET_FILTERS + + movsxd rax, DWORD PTR arg(1) ;pixels_per_line + movsxd rbx, DWORD PTR arg(3) ;out_pitch + lea rax, [rax + rax] ;bytes per line + lea rbx, [rbx + rbx] + lea rdx, [rax + rax * 2] + movsxd rcx, DWORD PTR arg(4) ;output_height + +.loop: + LOAD_VERT_8 0 + HIGH_APPLY_FILTER_8 0, 0 + + lea rdi, [rdi + rbx] + dec rcx + jnz .loop + + add rsp, 16 * 8 + pop rsp + pop rbx + ; begin epilog + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret + +;void aom_filter_block1d16_v8_sse2 +;( +; unsigned char *src_ptr, +; unsigned int src_pitch, +; unsigned char *output_ptr, +; unsigned int out_pitch, +; unsigned int output_height, +; short *filter +;) +global sym(aom_highbd_filter_block1d16_v8_sse2) PRIVATE +sym(aom_highbd_filter_block1d16_v8_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 7 + SAVE_XMM 7 + push rsi + push rdi + push rbx + ; end prolog + + ALIGN_STACK 16, rax + sub rsp, 16 * 8 + %define k0k1 [rsp + 16 * 0] + %define k6k7 [rsp + 16 * 1] + %define k2k5 [rsp + 16 * 2] + %define k3k4 [rsp + 16 * 3] + %define krd [rsp + 16 * 4] + %define temp [rsp + 16 * 5] + %define max [rsp + 16 * 6] + %define min [rsp + 16 * 7] + + HIGH_GET_FILTERS + + movsxd rax, DWORD PTR arg(1) ;pixels_per_line + movsxd rbx, DWORD PTR arg(3) ;out_pitch + lea rax, [rax + rax] ;bytes per line + lea rbx, [rbx + rbx] + lea rdx, [rax + rax * 2] + movsxd rcx, DWORD PTR arg(4) ;output_height + +.loop: + LOAD_VERT_8 0 + HIGH_APPLY_FILTER_8 0, 0 + sub rsi, rax + + LOAD_VERT_8 16 + HIGH_APPLY_FILTER_8 0, 16 + add rdi, rbx + + dec rcx + jnz .loop + + add rsp, 16 * 8 + pop rsp + pop rbx + ; begin epilog + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret + +global sym(aom_highbd_filter_block1d4_v8_avg_sse2) PRIVATE +sym(aom_highbd_filter_block1d4_v8_avg_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 7 + SAVE_XMM 7 + push rsi + push rdi + push rbx + ; end prolog + + ALIGN_STACK 16, rax + sub rsp, 16 * 7 + %define k0k6 [rsp + 16 * 0] + %define k2k5 [rsp + 16 * 1] + %define k3k4 [rsp + 16 * 2] + %define k1k7 [rsp + 16 * 3] + %define krd [rsp + 16 * 4] + %define max [rsp + 16 * 5] + %define min [rsp + 16 * 6] + + HIGH_GET_FILTERS_4 + + mov rsi, arg(0) ;src_ptr + mov rdi, arg(2) ;output_ptr + + movsxd rax, DWORD PTR arg(1) ;pixels_per_line + movsxd rbx, DWORD PTR arg(3) ;out_pitch + lea rax, [rax + rax] ;bytes per line + lea rbx, [rbx + rbx] + lea rdx, [rax + rax * 2] + movsxd rcx, DWORD PTR arg(4) ;output_height + +.loop: + movq xmm0, [rsi] ;load src: row 0 + movq xmm1, [rsi + rax] ;1 + movq xmm6, [rsi + rdx * 2] ;6 + lea rsi, [rsi + rax] + movq xmm7, [rsi + rdx * 2] ;7 + movq xmm2, [rsi + rax] ;2 + movq xmm3, [rsi + rax * 2] ;3 + movq xmm4, [rsi + rdx] ;4 + movq xmm5, [rsi + rax * 4] ;5 + + HIGH_APPLY_FILTER_4 1 + + lea rdi, [rdi + rbx] + dec rcx + jnz .loop + + add rsp, 16 * 7 + pop rsp + pop rbx + ; begin epilog + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret + +global sym(aom_highbd_filter_block1d8_v8_avg_sse2) PRIVATE +sym(aom_highbd_filter_block1d8_v8_avg_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 7 + SAVE_XMM 7 + push rsi + push rdi + push rbx + ; end prolog + + ALIGN_STACK 16, rax + sub rsp, 16 * 8 + %define k0k1 [rsp + 16 * 0] + %define k6k7 [rsp + 16 * 1] + %define k2k5 [rsp + 16 * 2] + %define k3k4 [rsp + 16 * 3] + %define krd [rsp + 16 * 4] + %define temp [rsp + 16 * 5] + %define max [rsp + 16 * 6] + %define min [rsp + 16 * 7] + + HIGH_GET_FILTERS + + movsxd rax, DWORD PTR arg(1) ;pixels_per_line + movsxd rbx, DWORD PTR arg(3) ;out_pitch + lea rax, [rax + rax] ;bytes per line + lea rbx, [rbx + rbx] + lea rdx, [rax + rax * 2] + movsxd rcx, DWORD PTR arg(4) ;output_height +.loop: + LOAD_VERT_8 0 + HIGH_APPLY_FILTER_8 1, 0 + + lea rdi, [rdi + rbx] + dec rcx + jnz .loop + + add rsp, 16 * 8 + pop rsp + pop rbx + ; begin epilog + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret + +global sym(aom_highbd_filter_block1d16_v8_avg_sse2) PRIVATE +sym(aom_highbd_filter_block1d16_v8_avg_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 7 + SAVE_XMM 7 + push rsi + push rdi + push rbx + ; end prolog + + ALIGN_STACK 16, rax + sub rsp, 16 * 8 + %define k0k1 [rsp + 16 * 0] + %define k6k7 [rsp + 16 * 1] + %define k2k5 [rsp + 16 * 2] + %define k3k4 [rsp + 16 * 3] + %define krd [rsp + 16 * 4] + %define temp [rsp + 16 * 5] + %define max [rsp + 16 * 6] + %define min [rsp + 16 * 7] + + HIGH_GET_FILTERS + + movsxd rax, DWORD PTR arg(1) ;pixels_per_line + movsxd rbx, DWORD PTR arg(3) ;out_pitch + lea rax, [rax + rax] ;bytes per line + lea rbx, [rbx + rbx] + lea rdx, [rax + rax * 2] + movsxd rcx, DWORD PTR arg(4) ;output_height +.loop: + LOAD_VERT_8 0 + HIGH_APPLY_FILTER_8 1, 0 + sub rsi, rax + + LOAD_VERT_8 16 + HIGH_APPLY_FILTER_8 1, 16 + add rdi, rbx + + dec rcx + jnz .loop + + add rsp, 16 * 8 + pop rsp + pop rbx + ; begin epilog + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret + +;void aom_filter_block1d4_h8_sse2 +;( +; unsigned char *src_ptr, +; unsigned int src_pixels_per_line, +; unsigned char *output_ptr, +; unsigned int output_pitch, +; unsigned int output_height, +; short *filter +;) +global sym(aom_highbd_filter_block1d4_h8_sse2) PRIVATE +sym(aom_highbd_filter_block1d4_h8_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 7 + SAVE_XMM 7 + push rsi + push rdi + ; end prolog + + ALIGN_STACK 16, rax + sub rsp, 16 * 7 + %define k0k6 [rsp + 16 * 0] + %define k2k5 [rsp + 16 * 1] + %define k3k4 [rsp + 16 * 2] + %define k1k7 [rsp + 16 * 3] + %define krd [rsp + 16 * 4] + %define max [rsp + 16 * 5] + %define min [rsp + 16 * 6] + + HIGH_GET_FILTERS_4 + + mov rsi, arg(0) ;src_ptr + mov rdi, arg(2) ;output_ptr + + movsxd rax, DWORD PTR arg(1) ;pixels_per_line + movsxd rdx, DWORD PTR arg(3) ;out_pitch + lea rax, [rax + rax] ;bytes per line + lea rdx, [rdx + rdx] + movsxd rcx, DWORD PTR arg(4) ;output_height + +.loop: + movdqu xmm0, [rsi - 6] ;load src + movdqu xmm4, [rsi + 2] + movdqa xmm1, xmm0 + movdqa xmm6, xmm4 + movdqa xmm7, xmm4 + movdqa xmm2, xmm0 + movdqa xmm3, xmm0 + movdqa xmm5, xmm4 + + psrldq xmm1, 2 + psrldq xmm6, 4 + psrldq xmm7, 6 + psrldq xmm2, 4 + psrldq xmm3, 6 + psrldq xmm5, 2 + + HIGH_APPLY_FILTER_4 0 + + lea rsi, [rsi + rax] + lea rdi, [rdi + rdx] + dec rcx + jnz .loop + + add rsp, 16 * 7 + pop rsp + + ; begin epilog + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret + +;void aom_filter_block1d8_h8_sse2 +;( +; unsigned char *src_ptr, +; unsigned int src_pixels_per_line, +; unsigned char *output_ptr, +; unsigned int output_pitch, +; unsigned int output_height, +; short *filter +;) +global sym(aom_highbd_filter_block1d8_h8_sse2) PRIVATE +sym(aom_highbd_filter_block1d8_h8_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 7 + SAVE_XMM 7 + push rsi + push rdi + ; end prolog + + ALIGN_STACK 16, rax + sub rsp, 16 * 8 + %define k0k1 [rsp + 16 * 0] + %define k6k7 [rsp + 16 * 1] + %define k2k5 [rsp + 16 * 2] + %define k3k4 [rsp + 16 * 3] + %define krd [rsp + 16 * 4] + %define temp [rsp + 16 * 5] + %define max [rsp + 16 * 6] + %define min [rsp + 16 * 7] + + HIGH_GET_FILTERS + + movsxd rax, DWORD PTR arg(1) ;pixels_per_line + movsxd rdx, DWORD PTR arg(3) ;out_pitch + lea rax, [rax + rax] ;bytes per line + lea rdx, [rdx + rdx] + movsxd rcx, DWORD PTR arg(4) ;output_height + +.loop: + movdqu xmm0, [rsi - 6] ;load src + movdqu xmm1, [rsi - 4] + movdqu xmm2, [rsi - 2] + movdqu xmm3, [rsi] + movdqu xmm4, [rsi + 2] + movdqu xmm5, [rsi + 4] + movdqu xmm6, [rsi + 6] + movdqu xmm7, [rsi + 8] + + HIGH_APPLY_FILTER_8 0, 0 + + lea rsi, [rsi + rax] + lea rdi, [rdi + rdx] + dec rcx + jnz .loop + + add rsp, 16 * 8 + pop rsp + + ; begin epilog + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret + +;void aom_filter_block1d16_h8_sse2 +;( +; unsigned char *src_ptr, +; unsigned int src_pixels_per_line, +; unsigned char *output_ptr, +; unsigned int output_pitch, +; unsigned int output_height, +; short *filter +;) +global sym(aom_highbd_filter_block1d16_h8_sse2) PRIVATE +sym(aom_highbd_filter_block1d16_h8_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 7 + SAVE_XMM 7 + push rsi + push rdi + ; end prolog + + ALIGN_STACK 16, rax + sub rsp, 16 * 8 + %define k0k1 [rsp + 16 * 0] + %define k6k7 [rsp + 16 * 1] + %define k2k5 [rsp + 16 * 2] + %define k3k4 [rsp + 16 * 3] + %define krd [rsp + 16 * 4] + %define temp [rsp + 16 * 5] + %define max [rsp + 16 * 6] + %define min [rsp + 16 * 7] + + HIGH_GET_FILTERS + + movsxd rax, DWORD PTR arg(1) ;pixels_per_line + movsxd rdx, DWORD PTR arg(3) ;out_pitch + lea rax, [rax + rax] ;bytes per line + lea rdx, [rdx + rdx] + movsxd rcx, DWORD PTR arg(4) ;output_height + +.loop: + movdqu xmm0, [rsi - 6] ;load src + movdqu xmm1, [rsi - 4] + movdqu xmm2, [rsi - 2] + movdqu xmm3, [rsi] + movdqu xmm4, [rsi + 2] + movdqu xmm5, [rsi + 4] + movdqu xmm6, [rsi + 6] + movdqu xmm7, [rsi + 8] + + HIGH_APPLY_FILTER_8 0, 0 + + movdqu xmm0, [rsi + 10] ;load src + movdqu xmm1, [rsi + 12] + movdqu xmm2, [rsi + 14] + movdqu xmm3, [rsi + 16] + movdqu xmm4, [rsi + 18] + movdqu xmm5, [rsi + 20] + movdqu xmm6, [rsi + 22] + movdqu xmm7, [rsi + 24] + + HIGH_APPLY_FILTER_8 0, 16 + + lea rsi, [rsi + rax] + lea rdi, [rdi + rdx] + dec rcx + jnz .loop + + add rsp, 16 * 8 + pop rsp + + ; begin epilog + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret + +global sym(aom_highbd_filter_block1d4_h8_avg_sse2) PRIVATE +sym(aom_highbd_filter_block1d4_h8_avg_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 7 + SAVE_XMM 7 + push rsi + push rdi + ; end prolog + + ALIGN_STACK 16, rax + sub rsp, 16 * 7 + %define k0k6 [rsp + 16 * 0] + %define k2k5 [rsp + 16 * 1] + %define k3k4 [rsp + 16 * 2] + %define k1k7 [rsp + 16 * 3] + %define krd [rsp + 16 * 4] + %define max [rsp + 16 * 5] + %define min [rsp + 16 * 6] + + HIGH_GET_FILTERS_4 + + mov rsi, arg(0) ;src_ptr + mov rdi, arg(2) ;output_ptr + + movsxd rax, DWORD PTR arg(1) ;pixels_per_line + movsxd rdx, DWORD PTR arg(3) ;out_pitch + lea rax, [rax + rax] ;bytes per line + lea rdx, [rdx + rdx] + movsxd rcx, DWORD PTR arg(4) ;output_height + +.loop: + movdqu xmm0, [rsi - 6] ;load src + movdqu xmm4, [rsi + 2] + movdqa xmm1, xmm0 + movdqa xmm6, xmm4 + movdqa xmm7, xmm4 + movdqa xmm2, xmm0 + movdqa xmm3, xmm0 + movdqa xmm5, xmm4 + + psrldq xmm1, 2 + psrldq xmm6, 4 + psrldq xmm7, 6 + psrldq xmm2, 4 + psrldq xmm3, 6 + psrldq xmm5, 2 + + HIGH_APPLY_FILTER_4 1 + + lea rsi, [rsi + rax] + lea rdi, [rdi + rdx] + dec rcx + jnz .loop + + add rsp, 16 * 7 + pop rsp + + ; begin epilog + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret + +global sym(aom_highbd_filter_block1d8_h8_avg_sse2) PRIVATE +sym(aom_highbd_filter_block1d8_h8_avg_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 7 + SAVE_XMM 7 + push rsi + push rdi + ; end prolog + + ALIGN_STACK 16, rax + sub rsp, 16 * 8 + %define k0k1 [rsp + 16 * 0] + %define k6k7 [rsp + 16 * 1] + %define k2k5 [rsp + 16 * 2] + %define k3k4 [rsp + 16 * 3] + %define krd [rsp + 16 * 4] + %define temp [rsp + 16 * 5] + %define max [rsp + 16 * 6] + %define min [rsp + 16 * 7] + + HIGH_GET_FILTERS + + movsxd rax, DWORD PTR arg(1) ;pixels_per_line + movsxd rdx, DWORD PTR arg(3) ;out_pitch + lea rax, [rax + rax] ;bytes per line + lea rdx, [rdx + rdx] + movsxd rcx, DWORD PTR arg(4) ;output_height + +.loop: + movdqu xmm0, [rsi - 6] ;load src + movdqu xmm1, [rsi - 4] + movdqu xmm2, [rsi - 2] + movdqu xmm3, [rsi] + movdqu xmm4, [rsi + 2] + movdqu xmm5, [rsi + 4] + movdqu xmm6, [rsi + 6] + movdqu xmm7, [rsi + 8] + + HIGH_APPLY_FILTER_8 1, 0 + + lea rsi, [rsi + rax] + lea rdi, [rdi + rdx] + dec rcx + jnz .loop + + add rsp, 16 * 8 + pop rsp + + ; begin epilog + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret + +global sym(aom_highbd_filter_block1d16_h8_avg_sse2) PRIVATE +sym(aom_highbd_filter_block1d16_h8_avg_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 7 + SAVE_XMM 7 + push rsi + push rdi + ; end prolog + + ALIGN_STACK 16, rax + sub rsp, 16 * 8 + %define k0k1 [rsp + 16 * 0] + %define k6k7 [rsp + 16 * 1] + %define k2k5 [rsp + 16 * 2] + %define k3k4 [rsp + 16 * 3] + %define krd [rsp + 16 * 4] + %define temp [rsp + 16 * 5] + %define max [rsp + 16 * 6] + %define min [rsp + 16 * 7] + + HIGH_GET_FILTERS + + movsxd rax, DWORD PTR arg(1) ;pixels_per_line + movsxd rdx, DWORD PTR arg(3) ;out_pitch + lea rax, [rax + rax] ;bytes per line + lea rdx, [rdx + rdx] + movsxd rcx, DWORD PTR arg(4) ;output_height + +.loop: + movdqu xmm0, [rsi - 6] ;load src + movdqu xmm1, [rsi - 4] + movdqu xmm2, [rsi - 2] + movdqu xmm3, [rsi] + movdqu xmm4, [rsi + 2] + movdqu xmm5, [rsi + 4] + movdqu xmm6, [rsi + 6] + movdqu xmm7, [rsi + 8] + + HIGH_APPLY_FILTER_8 1, 0 + + movdqu xmm0, [rsi + 10] ;load src + movdqu xmm1, [rsi + 12] + movdqu xmm2, [rsi + 14] + movdqu xmm3, [rsi + 16] + movdqu xmm4, [rsi + 18] + movdqu xmm5, [rsi + 20] + movdqu xmm6, [rsi + 22] + movdqu xmm7, [rsi + 24] + + HIGH_APPLY_FILTER_8 1, 16 + + lea rsi, [rsi + rax] + lea rdi, [rdi + rdx] + dec rcx + jnz .loop + + add rsp, 16 * 8 + pop rsp + + ; begin epilog + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret diff --git a/third_party/aom/aom_dsp/x86/aom_high_subpixel_bilinear_sse2.asm b/third_party/aom/aom_dsp/x86/aom_high_subpixel_bilinear_sse2.asm new file mode 100644 index 000000000..9e2ec748c --- /dev/null +++ b/third_party/aom/aom_dsp/x86/aom_high_subpixel_bilinear_sse2.asm @@ -0,0 +1,497 @@ +; +; Copyright (c) 2016, Alliance for Open Media. All rights reserved +; +; This source code is subject to the terms of the BSD 2 Clause License and +; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License +; was not distributed with this source code in the LICENSE file, you can +; obtain it at www.aomedia.org/license/software. If the Alliance for Open +; Media Patent License 1.0 was not distributed with this source code in the +; PATENTS file, you can obtain it at www.aomedia.org/license/patent. +; + +; + +%include "aom_ports/x86_abi_support.asm" + +%macro HIGH_GET_PARAM_4 0 + mov rdx, arg(5) ;filter ptr + mov rsi, arg(0) ;src_ptr + mov rdi, arg(2) ;output_ptr + mov rcx, 0x00000040 + + movdqa xmm3, [rdx] ;load filters + pshuflw xmm4, xmm3, 11111111b ;k3 + psrldq xmm3, 8 + pshuflw xmm3, xmm3, 0b ;k4 + punpcklwd xmm4, xmm3 ;k3k4 + + movq xmm3, rcx ;rounding + pshufd xmm3, xmm3, 0 + + mov rdx, 0x00010001 + movsxd rcx, DWORD PTR arg(6) ;bps + movq xmm5, rdx + movq xmm2, rcx + pshufd xmm5, xmm5, 0b + movdqa xmm1, xmm5 + psllw xmm5, xmm2 + psubw xmm5, xmm1 ;max value (for clamping) + pxor xmm2, xmm2 ;min value (for clamping) + + movsxd rax, DWORD PTR arg(1) ;pixels_per_line + movsxd rdx, DWORD PTR arg(3) ;out_pitch + movsxd rcx, DWORD PTR arg(4) ;output_height +%endm + +%macro HIGH_APPLY_FILTER_4 1 + + punpcklwd xmm0, xmm1 ;two row in one register + pmaddwd xmm0, xmm4 ;multiply the filter factors + + paddd xmm0, xmm3 ;rounding + psrad xmm0, 7 ;shift + packssdw xmm0, xmm0 ;pack to word + + ;clamp the values + pminsw xmm0, xmm5 + pmaxsw xmm0, xmm2 + +%if %1 + movq xmm1, [rdi] + pavgw xmm0, xmm1 +%endif + + movq [rdi], xmm0 + lea rsi, [rsi + 2*rax] + lea rdi, [rdi + 2*rdx] + dec rcx +%endm + +%if ARCH_X86_64 +%macro HIGH_GET_PARAM 0 + mov rdx, arg(5) ;filter ptr + mov rsi, arg(0) ;src_ptr + mov rdi, arg(2) ;output_ptr + mov rcx, 0x00000040 + + movdqa xmm6, [rdx] ;load filters + + pshuflw xmm7, xmm6, 11111111b ;k3 + pshufhw xmm6, xmm6, 0b ;k4 + psrldq xmm6, 8 + punpcklwd xmm7, xmm6 ;k3k4k3k4k3k4k3k4 + + movq xmm4, rcx ;rounding + pshufd xmm4, xmm4, 0 + + mov rdx, 0x00010001 + movsxd rcx, DWORD PTR arg(6) ;bps + movq xmm8, rdx + movq xmm5, rcx + pshufd xmm8, xmm8, 0b + movdqa xmm1, xmm8 + psllw xmm8, xmm5 + psubw xmm8, xmm1 ;max value (for clamping) + pxor xmm5, xmm5 ;min value (for clamping) + + movsxd rax, DWORD PTR arg(1) ;pixels_per_line + movsxd rdx, DWORD PTR arg(3) ;out_pitch + movsxd rcx, DWORD PTR arg(4) ;output_height +%endm + +%macro HIGH_APPLY_FILTER_8 1 + movdqa xmm6, xmm0 + punpckhwd xmm6, xmm1 + punpcklwd xmm0, xmm1 + pmaddwd xmm6, xmm7 + pmaddwd xmm0, xmm7 + + paddd xmm6, xmm4 ;rounding + paddd xmm0, xmm4 ;rounding + psrad xmm6, 7 ;shift + psrad xmm0, 7 ;shift + packssdw xmm0, xmm6 ;pack back to word + + ;clamp the values + pminsw xmm0, xmm8 + pmaxsw xmm0, xmm5 + +%if %1 + movdqu xmm1, [rdi] + pavgw xmm0, xmm1 +%endif + movdqu [rdi], xmm0 ;store the result + + lea rsi, [rsi + 2*rax] + lea rdi, [rdi + 2*rdx] + dec rcx +%endm + +%macro HIGH_APPLY_FILTER_16 1 + movdqa xmm9, xmm0 + movdqa xmm6, xmm2 + punpckhwd xmm9, xmm1 + punpckhwd xmm6, xmm3 + punpcklwd xmm0, xmm1 + punpcklwd xmm2, xmm3 + + pmaddwd xmm9, xmm7 + pmaddwd xmm6, xmm7 + pmaddwd xmm0, xmm7 + pmaddwd xmm2, xmm7 + + paddd xmm9, xmm4 ;rounding + paddd xmm6, xmm4 + paddd xmm0, xmm4 + paddd xmm2, xmm4 + + psrad xmm9, 7 ;shift + psrad xmm6, 7 + psrad xmm0, 7 + psrad xmm2, 7 + + packssdw xmm0, xmm9 ;pack back to word + packssdw xmm2, xmm6 ;pack back to word + + ;clamp the values + pminsw xmm0, xmm8 + pmaxsw xmm0, xmm5 + pminsw xmm2, xmm8 + pmaxsw xmm2, xmm5 + +%if %1 + movdqu xmm1, [rdi] + movdqu xmm3, [rdi + 16] + pavgw xmm0, xmm1 + pavgw xmm2, xmm3 +%endif + movdqu [rdi], xmm0 ;store the result + movdqu [rdi + 16], xmm2 ;store the result + + lea rsi, [rsi + 2*rax] + lea rdi, [rdi + 2*rdx] + dec rcx +%endm +%endif + +global sym(aom_highbd_filter_block1d4_v2_sse2) PRIVATE +sym(aom_highbd_filter_block1d4_v2_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 7 + push rsi + push rdi + ; end prolog + + HIGH_GET_PARAM_4 +.loop: + movq xmm0, [rsi] ;load src + movq xmm1, [rsi + 2*rax] + + HIGH_APPLY_FILTER_4 0 + jnz .loop + + ; begin epilog + pop rdi + pop rsi + UNSHADOW_ARGS + pop rbp + ret + +%if ARCH_X86_64 +global sym(aom_highbd_filter_block1d8_v2_sse2) PRIVATE +sym(aom_highbd_filter_block1d8_v2_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 7 + SAVE_XMM 8 + push rsi + push rdi + ; end prolog + + HIGH_GET_PARAM +.loop: + movdqu xmm0, [rsi] ;0 + movdqu xmm1, [rsi + 2*rax] ;1 + + HIGH_APPLY_FILTER_8 0 + jnz .loop + + ; begin epilog + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret + +global sym(aom_highbd_filter_block1d16_v2_sse2) PRIVATE +sym(aom_highbd_filter_block1d16_v2_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 7 + SAVE_XMM 9 + push rsi + push rdi + ; end prolog + + HIGH_GET_PARAM +.loop: + movdqu xmm0, [rsi] ;0 + movdqu xmm2, [rsi + 16] + movdqu xmm1, [rsi + 2*rax] ;1 + movdqu xmm3, [rsi + 2*rax + 16] + + HIGH_APPLY_FILTER_16 0 + jnz .loop + + ; begin epilog + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret +%endif + +global sym(aom_highbd_filter_block1d4_v2_avg_sse2) PRIVATE +sym(aom_highbd_filter_block1d4_v2_avg_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 7 + push rsi + push rdi + ; end prolog + + HIGH_GET_PARAM_4 +.loop: + movq xmm0, [rsi] ;load src + movq xmm1, [rsi + 2*rax] + + HIGH_APPLY_FILTER_4 1 + jnz .loop + + ; begin epilog + pop rdi + pop rsi + UNSHADOW_ARGS + pop rbp + ret + +%if ARCH_X86_64 +global sym(aom_highbd_filter_block1d8_v2_avg_sse2) PRIVATE +sym(aom_highbd_filter_block1d8_v2_avg_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 7 + SAVE_XMM 8 + push rsi + push rdi + ; end prolog + + HIGH_GET_PARAM +.loop: + movdqu xmm0, [rsi] ;0 + movdqu xmm1, [rsi + 2*rax] ;1 + + HIGH_APPLY_FILTER_8 1 + jnz .loop + + ; begin epilog + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret + +global sym(aom_highbd_filter_block1d16_v2_avg_sse2) PRIVATE +sym(aom_highbd_filter_block1d16_v2_avg_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 7 + SAVE_XMM 9 + push rsi + push rdi + ; end prolog + + HIGH_GET_PARAM +.loop: + movdqu xmm0, [rsi] ;0 + movdqu xmm1, [rsi + 2*rax] ;1 + movdqu xmm2, [rsi + 16] + movdqu xmm3, [rsi + 2*rax + 16] + + HIGH_APPLY_FILTER_16 1 + jnz .loop + + ; begin epilog + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret +%endif + +global sym(aom_highbd_filter_block1d4_h2_sse2) PRIVATE +sym(aom_highbd_filter_block1d4_h2_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 7 + push rsi + push rdi + ; end prolog + + HIGH_GET_PARAM_4 +.loop: + movdqu xmm0, [rsi] ;load src + movdqa xmm1, xmm0 + psrldq xmm1, 2 + + HIGH_APPLY_FILTER_4 0 + jnz .loop + + ; begin epilog + pop rdi + pop rsi + UNSHADOW_ARGS + pop rbp + ret + +%if ARCH_X86_64 +global sym(aom_highbd_filter_block1d8_h2_sse2) PRIVATE +sym(aom_highbd_filter_block1d8_h2_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 7 + SAVE_XMM 8 + push rsi + push rdi + ; end prolog + + HIGH_GET_PARAM +.loop: + movdqu xmm0, [rsi] ;load src + movdqu xmm1, [rsi + 2] + + HIGH_APPLY_FILTER_8 0 + jnz .loop + + ; begin epilog + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret + +global sym(aom_highbd_filter_block1d16_h2_sse2) PRIVATE +sym(aom_highbd_filter_block1d16_h2_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 7 + SAVE_XMM 9 + push rsi + push rdi + ; end prolog + + HIGH_GET_PARAM +.loop: + movdqu xmm0, [rsi] ;load src + movdqu xmm1, [rsi + 2] + movdqu xmm2, [rsi + 16] + movdqu xmm3, [rsi + 18] + + HIGH_APPLY_FILTER_16 0 + jnz .loop + + ; begin epilog + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret +%endif + +global sym(aom_highbd_filter_block1d4_h2_avg_sse2) PRIVATE +sym(aom_highbd_filter_block1d4_h2_avg_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 7 + push rsi + push rdi + ; end prolog + + HIGH_GET_PARAM_4 +.loop: + movdqu xmm0, [rsi] ;load src + movdqa xmm1, xmm0 + psrldq xmm1, 2 + + HIGH_APPLY_FILTER_4 1 + jnz .loop + + ; begin epilog + pop rdi + pop rsi + UNSHADOW_ARGS + pop rbp + ret + +%if ARCH_X86_64 +global sym(aom_highbd_filter_block1d8_h2_avg_sse2) PRIVATE +sym(aom_highbd_filter_block1d8_h2_avg_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 7 + SAVE_XMM 8 + push rsi + push rdi + ; end prolog + + HIGH_GET_PARAM +.loop: + movdqu xmm0, [rsi] ;load src + movdqu xmm1, [rsi + 2] + + HIGH_APPLY_FILTER_8 1 + jnz .loop + + ; begin epilog + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret + +global sym(aom_highbd_filter_block1d16_h2_avg_sse2) PRIVATE +sym(aom_highbd_filter_block1d16_h2_avg_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 7 + SAVE_XMM 9 + push rsi + push rdi + ; end prolog + + HIGH_GET_PARAM +.loop: + movdqu xmm0, [rsi] ;load src + movdqu xmm1, [rsi + 2] + movdqu xmm2, [rsi + 16] + movdqu xmm3, [rsi + 18] + + HIGH_APPLY_FILTER_16 1 + jnz .loop + + ; begin epilog + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret +%endif diff --git a/third_party/aom/aom_dsp/x86/aom_subpixel_8t_intrin_avx2.c b/third_party/aom/aom_dsp/x86/aom_subpixel_8t_intrin_avx2.c new file mode 100644 index 000000000..61476b8be --- /dev/null +++ b/third_party/aom/aom_dsp/x86/aom_subpixel_8t_intrin_avx2.c @@ -0,0 +1,575 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include <immintrin.h> + +#include "./aom_dsp_rtcd.h" +#include "aom_dsp/x86/convolve.h" +#include "aom_ports/mem.h" + +// filters for 16_h8 and 16_v8 +DECLARE_ALIGNED(32, static const uint8_t, filt1_global_avx2[32]) = { + 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8, + 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8 +}; + +DECLARE_ALIGNED(32, static const uint8_t, filt2_global_avx2[32]) = { + 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8, 8, 9, 9, 10, + 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8, 8, 9, 9, 10 +}; + +DECLARE_ALIGNED(32, static const uint8_t, filt3_global_avx2[32]) = { + 4, 5, 5, 6, 6, 7, 7, 8, 8, 9, 9, 10, 10, 11, 11, 12, + 4, 5, 5, 6, 6, 7, 7, 8, 8, 9, 9, 10, 10, 11, 11, 12 +}; + +DECLARE_ALIGNED(32, static const uint8_t, filt4_global_avx2[32]) = { + 6, 7, 7, 8, 8, 9, 9, 10, 10, 11, 11, 12, 12, 13, 13, 14, + 6, 7, 7, 8, 8, 9, 9, 10, 10, 11, 11, 12, 12, 13, 13, 14 +}; + +#if defined(__clang__) +#if (__clang_major__ > 0 && __clang_major__ < 3) || \ + (__clang_major__ == 3 && __clang_minor__ <= 3) || \ + (defined(__APPLE__) && defined(__apple_build_version__) && \ + ((__clang_major__ == 4 && __clang_minor__ <= 2) || \ + (__clang_major__ == 5 && __clang_minor__ == 0))) +#define MM256_BROADCASTSI128_SI256(x) \ + _mm_broadcastsi128_si256((__m128i const *)&(x)) +#else // clang > 3.3, and not 5.0 on macosx. +#define MM256_BROADCASTSI128_SI256(x) _mm256_broadcastsi128_si256(x) +#endif // clang <= 3.3 +#elif defined(__GNUC__) +#if __GNUC__ < 4 || (__GNUC__ == 4 && __GNUC_MINOR__ <= 6) +#define MM256_BROADCASTSI128_SI256(x) \ + _mm_broadcastsi128_si256((__m128i const *)&(x)) +#elif __GNUC__ == 4 && __GNUC_MINOR__ == 7 +#define MM256_BROADCASTSI128_SI256(x) _mm_broadcastsi128_si256(x) +#else // gcc > 4.7 +#define MM256_BROADCASTSI128_SI256(x) _mm256_broadcastsi128_si256(x) +#endif // gcc <= 4.6 +#else // !(gcc || clang) +#define MM256_BROADCASTSI128_SI256(x) _mm256_broadcastsi128_si256(x) +#endif // __clang__ + +static void aom_filter_block1d16_h8_avx2( + const uint8_t *src_ptr, ptrdiff_t src_pixels_per_line, uint8_t *output_ptr, + ptrdiff_t output_pitch, uint32_t output_height, const int16_t *filter) { + __m128i filtersReg; + __m256i addFilterReg64, filt1Reg, filt2Reg, filt3Reg, filt4Reg; + __m256i firstFilters, secondFilters, thirdFilters, forthFilters; + __m256i srcRegFilt32b1_1, srcRegFilt32b2_1, srcRegFilt32b2, srcRegFilt32b3; + __m256i srcReg32b1, srcReg32b2, filtersReg32; + unsigned int i; + ptrdiff_t src_stride, dst_stride; + + // create a register with 0,64,0,64,0,64,0,64,0,64,0,64,0,64,0,64 + addFilterReg64 = _mm256_set1_epi32((int)0x0400040u); + filtersReg = _mm_loadu_si128((const __m128i *)filter); + // converting the 16 bit (short) to 8 bit (byte) and have the same data + // in both lanes of 128 bit register. + filtersReg = _mm_packs_epi16(filtersReg, filtersReg); + // have the same data in both lanes of a 256 bit register + filtersReg32 = MM256_BROADCASTSI128_SI256(filtersReg); + + // duplicate only the first 16 bits (first and second byte) + // across 256 bit register + firstFilters = _mm256_shuffle_epi8(filtersReg32, _mm256_set1_epi16(0x100u)); + // duplicate only the second 16 bits (third and forth byte) + // across 256 bit register + secondFilters = _mm256_shuffle_epi8(filtersReg32, _mm256_set1_epi16(0x302u)); + // duplicate only the third 16 bits (fifth and sixth byte) + // across 256 bit register + thirdFilters = _mm256_shuffle_epi8(filtersReg32, _mm256_set1_epi16(0x504u)); + // duplicate only the forth 16 bits (seventh and eighth byte) + // across 256 bit register + forthFilters = _mm256_shuffle_epi8(filtersReg32, _mm256_set1_epi16(0x706u)); + + filt1Reg = _mm256_load_si256((__m256i const *)filt1_global_avx2); + filt2Reg = _mm256_load_si256((__m256i const *)filt2_global_avx2); + filt3Reg = _mm256_load_si256((__m256i const *)filt3_global_avx2); + filt4Reg = _mm256_load_si256((__m256i const *)filt4_global_avx2); + + // multiple the size of the source and destination stride by two + src_stride = src_pixels_per_line << 1; + dst_stride = output_pitch << 1; + for (i = output_height; i > 1; i -= 2) { + // load the 2 strides of source + srcReg32b1 = + _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(src_ptr - 3))); + srcReg32b1 = _mm256_inserti128_si256( + srcReg32b1, + _mm_loadu_si128((const __m128i *)(src_ptr + src_pixels_per_line - 3)), + 1); + + // filter the source buffer + srcRegFilt32b1_1 = _mm256_shuffle_epi8(srcReg32b1, filt1Reg); + srcRegFilt32b2 = _mm256_shuffle_epi8(srcReg32b1, filt4Reg); + + // multiply 2 adjacent elements with the filter and add the result + srcRegFilt32b1_1 = _mm256_maddubs_epi16(srcRegFilt32b1_1, firstFilters); + srcRegFilt32b2 = _mm256_maddubs_epi16(srcRegFilt32b2, forthFilters); + + // add and saturate the results together + srcRegFilt32b1_1 = _mm256_adds_epi16(srcRegFilt32b1_1, srcRegFilt32b2); + + // filter the source buffer + srcRegFilt32b3 = _mm256_shuffle_epi8(srcReg32b1, filt2Reg); + srcRegFilt32b2 = _mm256_shuffle_epi8(srcReg32b1, filt3Reg); + + // multiply 2 adjacent elements with the filter and add the result + srcRegFilt32b3 = _mm256_maddubs_epi16(srcRegFilt32b3, secondFilters); + srcRegFilt32b2 = _mm256_maddubs_epi16(srcRegFilt32b2, thirdFilters); + + // add and saturate the results together + srcRegFilt32b1_1 = _mm256_adds_epi16( + srcRegFilt32b1_1, _mm256_min_epi16(srcRegFilt32b3, srcRegFilt32b2)); + + // reading 2 strides of the next 16 bytes + // (part of it was being read by earlier read) + srcReg32b2 = + _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(src_ptr + 5))); + srcReg32b2 = _mm256_inserti128_si256( + srcReg32b2, + _mm_loadu_si128((const __m128i *)(src_ptr + src_pixels_per_line + 5)), + 1); + + // add and saturate the results together + srcRegFilt32b1_1 = _mm256_adds_epi16( + srcRegFilt32b1_1, _mm256_max_epi16(srcRegFilt32b3, srcRegFilt32b2)); + + // filter the source buffer + srcRegFilt32b2_1 = _mm256_shuffle_epi8(srcReg32b2, filt1Reg); + srcRegFilt32b2 = _mm256_shuffle_epi8(srcReg32b2, filt4Reg); + + // multiply 2 adjacent elements with the filter and add the result + srcRegFilt32b2_1 = _mm256_maddubs_epi16(srcRegFilt32b2_1, firstFilters); + srcRegFilt32b2 = _mm256_maddubs_epi16(srcRegFilt32b2, forthFilters); + + // add and saturate the results together + srcRegFilt32b2_1 = _mm256_adds_epi16(srcRegFilt32b2_1, srcRegFilt32b2); + + // filter the source buffer + srcRegFilt32b3 = _mm256_shuffle_epi8(srcReg32b2, filt2Reg); + srcRegFilt32b2 = _mm256_shuffle_epi8(srcReg32b2, filt3Reg); + + // multiply 2 adjacent elements with the filter and add the result + srcRegFilt32b3 = _mm256_maddubs_epi16(srcRegFilt32b3, secondFilters); + srcRegFilt32b2 = _mm256_maddubs_epi16(srcRegFilt32b2, thirdFilters); + + // add and saturate the results together + srcRegFilt32b2_1 = _mm256_adds_epi16( + srcRegFilt32b2_1, _mm256_min_epi16(srcRegFilt32b3, srcRegFilt32b2)); + srcRegFilt32b2_1 = _mm256_adds_epi16( + srcRegFilt32b2_1, _mm256_max_epi16(srcRegFilt32b3, srcRegFilt32b2)); + + srcRegFilt32b1_1 = _mm256_adds_epi16(srcRegFilt32b1_1, addFilterReg64); + + srcRegFilt32b2_1 = _mm256_adds_epi16(srcRegFilt32b2_1, addFilterReg64); + + // shift by 7 bit each 16 bit + srcRegFilt32b1_1 = _mm256_srai_epi16(srcRegFilt32b1_1, 7); + srcRegFilt32b2_1 = _mm256_srai_epi16(srcRegFilt32b2_1, 7); + + // shrink to 8 bit each 16 bits, the first lane contain the first + // convolve result and the second lane contain the second convolve + // result + srcRegFilt32b1_1 = _mm256_packus_epi16(srcRegFilt32b1_1, srcRegFilt32b2_1); + + src_ptr += src_stride; + + // save 16 bytes + _mm_store_si128((__m128i *)output_ptr, + _mm256_castsi256_si128(srcRegFilt32b1_1)); + + // save the next 16 bits + _mm_store_si128((__m128i *)(output_ptr + output_pitch), + _mm256_extractf128_si256(srcRegFilt32b1_1, 1)); + output_ptr += dst_stride; + } + + // if the number of strides is odd. + // process only 16 bytes + if (i > 0) { + __m128i srcReg1, srcReg2, srcRegFilt1_1, srcRegFilt2_1; + __m128i srcRegFilt2, srcRegFilt3; + + srcReg1 = _mm_loadu_si128((const __m128i *)(src_ptr - 3)); + + // filter the source buffer + srcRegFilt1_1 = _mm_shuffle_epi8(srcReg1, _mm256_castsi256_si128(filt1Reg)); + srcRegFilt2 = _mm_shuffle_epi8(srcReg1, _mm256_castsi256_si128(filt4Reg)); + + // multiply 2 adjacent elements with the filter and add the result + srcRegFilt1_1 = + _mm_maddubs_epi16(srcRegFilt1_1, _mm256_castsi256_si128(firstFilters)); + srcRegFilt2 = + _mm_maddubs_epi16(srcRegFilt2, _mm256_castsi256_si128(forthFilters)); + + // add and saturate the results together + srcRegFilt1_1 = _mm_adds_epi16(srcRegFilt1_1, srcRegFilt2); + + // filter the source buffer + srcRegFilt3 = _mm_shuffle_epi8(srcReg1, _mm256_castsi256_si128(filt2Reg)); + srcRegFilt2 = _mm_shuffle_epi8(srcReg1, _mm256_castsi256_si128(filt3Reg)); + + // multiply 2 adjacent elements with the filter and add the result + srcRegFilt3 = + _mm_maddubs_epi16(srcRegFilt3, _mm256_castsi256_si128(secondFilters)); + srcRegFilt2 = + _mm_maddubs_epi16(srcRegFilt2, _mm256_castsi256_si128(thirdFilters)); + + // add and saturate the results together + srcRegFilt1_1 = + _mm_adds_epi16(srcRegFilt1_1, _mm_min_epi16(srcRegFilt3, srcRegFilt2)); + + // reading the next 16 bytes + // (part of it was being read by earlier read) + srcReg2 = _mm_loadu_si128((const __m128i *)(src_ptr + 5)); + + // add and saturate the results together + srcRegFilt1_1 = + _mm_adds_epi16(srcRegFilt1_1, _mm_max_epi16(srcRegFilt3, srcRegFilt2)); + + // filter the source buffer + srcRegFilt2_1 = _mm_shuffle_epi8(srcReg2, _mm256_castsi256_si128(filt1Reg)); + srcRegFilt2 = _mm_shuffle_epi8(srcReg2, _mm256_castsi256_si128(filt4Reg)); + + // multiply 2 adjacent elements with the filter and add the result + srcRegFilt2_1 = + _mm_maddubs_epi16(srcRegFilt2_1, _mm256_castsi256_si128(firstFilters)); + srcRegFilt2 = + _mm_maddubs_epi16(srcRegFilt2, _mm256_castsi256_si128(forthFilters)); + + // add and saturate the results together + srcRegFilt2_1 = _mm_adds_epi16(srcRegFilt2_1, srcRegFilt2); + + // filter the source buffer + srcRegFilt3 = _mm_shuffle_epi8(srcReg2, _mm256_castsi256_si128(filt2Reg)); + srcRegFilt2 = _mm_shuffle_epi8(srcReg2, _mm256_castsi256_si128(filt3Reg)); + + // multiply 2 adjacent elements with the filter and add the result + srcRegFilt3 = + _mm_maddubs_epi16(srcRegFilt3, _mm256_castsi256_si128(secondFilters)); + srcRegFilt2 = + _mm_maddubs_epi16(srcRegFilt2, _mm256_castsi256_si128(thirdFilters)); + + // add and saturate the results together + srcRegFilt2_1 = + _mm_adds_epi16(srcRegFilt2_1, _mm_min_epi16(srcRegFilt3, srcRegFilt2)); + srcRegFilt2_1 = + _mm_adds_epi16(srcRegFilt2_1, _mm_max_epi16(srcRegFilt3, srcRegFilt2)); + + srcRegFilt1_1 = + _mm_adds_epi16(srcRegFilt1_1, _mm256_castsi256_si128(addFilterReg64)); + + srcRegFilt2_1 = + _mm_adds_epi16(srcRegFilt2_1, _mm256_castsi256_si128(addFilterReg64)); + + // shift by 7 bit each 16 bit + srcRegFilt1_1 = _mm_srai_epi16(srcRegFilt1_1, 7); + srcRegFilt2_1 = _mm_srai_epi16(srcRegFilt2_1, 7); + + // shrink to 8 bit each 16 bits, the first lane contain the first + // convolve result and the second lane contain the second convolve + // result + srcRegFilt1_1 = _mm_packus_epi16(srcRegFilt1_1, srcRegFilt2_1); + + // save 16 bytes + _mm_store_si128((__m128i *)output_ptr, srcRegFilt1_1); + } +} + +static void aom_filter_block1d16_v8_avx2( + const uint8_t *src_ptr, ptrdiff_t src_pitch, uint8_t *output_ptr, + ptrdiff_t out_pitch, uint32_t output_height, const int16_t *filter) { + __m128i filtersReg; + __m256i addFilterReg64; + __m256i srcReg32b1, srcReg32b2, srcReg32b3, srcReg32b4, srcReg32b5; + __m256i srcReg32b6, srcReg32b7, srcReg32b8, srcReg32b9, srcReg32b10; + __m256i srcReg32b11, srcReg32b12, filtersReg32; + __m256i firstFilters, secondFilters, thirdFilters, forthFilters; + unsigned int i; + ptrdiff_t src_stride, dst_stride; + + // create a register with 0,64,0,64,0,64,0,64,0,64,0,64,0,64,0,64 + addFilterReg64 = _mm256_set1_epi32((int)0x0400040u); + filtersReg = _mm_loadu_si128((const __m128i *)filter); + // converting the 16 bit (short) to 8 bit (byte) and have the + // same data in both lanes of 128 bit register. + filtersReg = _mm_packs_epi16(filtersReg, filtersReg); + // have the same data in both lanes of a 256 bit register + filtersReg32 = MM256_BROADCASTSI128_SI256(filtersReg); + + // duplicate only the first 16 bits (first and second byte) + // across 256 bit register + firstFilters = _mm256_shuffle_epi8(filtersReg32, _mm256_set1_epi16(0x100u)); + // duplicate only the second 16 bits (third and forth byte) + // across 256 bit register + secondFilters = _mm256_shuffle_epi8(filtersReg32, _mm256_set1_epi16(0x302u)); + // duplicate only the third 16 bits (fifth and sixth byte) + // across 256 bit register + thirdFilters = _mm256_shuffle_epi8(filtersReg32, _mm256_set1_epi16(0x504u)); + // duplicate only the forth 16 bits (seventh and eighth byte) + // across 256 bit register + forthFilters = _mm256_shuffle_epi8(filtersReg32, _mm256_set1_epi16(0x706u)); + + // multiple the size of the source and destination stride by two + src_stride = src_pitch << 1; + dst_stride = out_pitch << 1; + + // load 16 bytes 7 times in stride of src_pitch + srcReg32b1 = + _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(src_ptr))); + srcReg32b2 = _mm256_castsi128_si256( + _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch))); + srcReg32b3 = _mm256_castsi128_si256( + _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 2))); + srcReg32b4 = _mm256_castsi128_si256( + _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 3))); + srcReg32b5 = _mm256_castsi128_si256( + _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 4))); + srcReg32b6 = _mm256_castsi128_si256( + _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 5))); + srcReg32b7 = _mm256_castsi128_si256( + _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 6))); + + // have each consecutive loads on the same 256 register + srcReg32b1 = _mm256_inserti128_si256(srcReg32b1, + _mm256_castsi256_si128(srcReg32b2), 1); + srcReg32b2 = _mm256_inserti128_si256(srcReg32b2, + _mm256_castsi256_si128(srcReg32b3), 1); + srcReg32b3 = _mm256_inserti128_si256(srcReg32b3, + _mm256_castsi256_si128(srcReg32b4), 1); + srcReg32b4 = _mm256_inserti128_si256(srcReg32b4, + _mm256_castsi256_si128(srcReg32b5), 1); + srcReg32b5 = _mm256_inserti128_si256(srcReg32b5, + _mm256_castsi256_si128(srcReg32b6), 1); + srcReg32b6 = _mm256_inserti128_si256(srcReg32b6, + _mm256_castsi256_si128(srcReg32b7), 1); + + // merge every two consecutive registers except the last one + srcReg32b10 = _mm256_unpacklo_epi8(srcReg32b1, srcReg32b2); + srcReg32b1 = _mm256_unpackhi_epi8(srcReg32b1, srcReg32b2); + + // save + srcReg32b11 = _mm256_unpacklo_epi8(srcReg32b3, srcReg32b4); + + // save + srcReg32b3 = _mm256_unpackhi_epi8(srcReg32b3, srcReg32b4); + + // save + srcReg32b2 = _mm256_unpacklo_epi8(srcReg32b5, srcReg32b6); + + // save + srcReg32b5 = _mm256_unpackhi_epi8(srcReg32b5, srcReg32b6); + + for (i = output_height; i > 1; i -= 2) { + // load the last 2 loads of 16 bytes and have every two + // consecutive loads in the same 256 bit register + srcReg32b8 = _mm256_castsi128_si256( + _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 7))); + srcReg32b7 = _mm256_inserti128_si256(srcReg32b7, + _mm256_castsi256_si128(srcReg32b8), 1); + srcReg32b9 = _mm256_castsi128_si256( + _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 8))); + srcReg32b8 = _mm256_inserti128_si256(srcReg32b8, + _mm256_castsi256_si128(srcReg32b9), 1); + + // merge every two consecutive registers + // save + srcReg32b4 = _mm256_unpacklo_epi8(srcReg32b7, srcReg32b8); + srcReg32b7 = _mm256_unpackhi_epi8(srcReg32b7, srcReg32b8); + + // multiply 2 adjacent elements with the filter and add the result + srcReg32b10 = _mm256_maddubs_epi16(srcReg32b10, firstFilters); + srcReg32b6 = _mm256_maddubs_epi16(srcReg32b4, forthFilters); + + // add and saturate the results together + srcReg32b10 = _mm256_adds_epi16(srcReg32b10, srcReg32b6); + + // multiply 2 adjacent elements with the filter and add the result + srcReg32b8 = _mm256_maddubs_epi16(srcReg32b11, secondFilters); + srcReg32b12 = _mm256_maddubs_epi16(srcReg32b2, thirdFilters); + + // add and saturate the results together + srcReg32b10 = _mm256_adds_epi16(srcReg32b10, + _mm256_min_epi16(srcReg32b8, srcReg32b12)); + srcReg32b10 = _mm256_adds_epi16(srcReg32b10, + _mm256_max_epi16(srcReg32b8, srcReg32b12)); + + // multiply 2 adjacent elements with the filter and add the result + srcReg32b1 = _mm256_maddubs_epi16(srcReg32b1, firstFilters); + srcReg32b6 = _mm256_maddubs_epi16(srcReg32b7, forthFilters); + + srcReg32b1 = _mm256_adds_epi16(srcReg32b1, srcReg32b6); + + // multiply 2 adjacent elements with the filter and add the result + srcReg32b8 = _mm256_maddubs_epi16(srcReg32b3, secondFilters); + srcReg32b12 = _mm256_maddubs_epi16(srcReg32b5, thirdFilters); + + // add and saturate the results together + srcReg32b1 = _mm256_adds_epi16(srcReg32b1, + _mm256_min_epi16(srcReg32b8, srcReg32b12)); + srcReg32b1 = _mm256_adds_epi16(srcReg32b1, + _mm256_max_epi16(srcReg32b8, srcReg32b12)); + + srcReg32b10 = _mm256_adds_epi16(srcReg32b10, addFilterReg64); + srcReg32b1 = _mm256_adds_epi16(srcReg32b1, addFilterReg64); + + // shift by 7 bit each 16 bit + srcReg32b10 = _mm256_srai_epi16(srcReg32b10, 7); + srcReg32b1 = _mm256_srai_epi16(srcReg32b1, 7); + + // shrink to 8 bit each 16 bits, the first lane contain the first + // convolve result and the second lane contain the second convolve + // result + srcReg32b1 = _mm256_packus_epi16(srcReg32b10, srcReg32b1); + + src_ptr += src_stride; + + // save 16 bytes + _mm_store_si128((__m128i *)output_ptr, _mm256_castsi256_si128(srcReg32b1)); + + // save the next 16 bits + _mm_store_si128((__m128i *)(output_ptr + out_pitch), + _mm256_extractf128_si256(srcReg32b1, 1)); + + output_ptr += dst_stride; + + // save part of the registers for next strides + srcReg32b10 = srcReg32b11; + srcReg32b1 = srcReg32b3; + srcReg32b11 = srcReg32b2; + srcReg32b3 = srcReg32b5; + srcReg32b2 = srcReg32b4; + srcReg32b5 = srcReg32b7; + srcReg32b7 = srcReg32b9; + } + if (i > 0) { + __m128i srcRegFilt1, srcRegFilt3, srcRegFilt4, srcRegFilt5; + __m128i srcRegFilt6, srcRegFilt7, srcRegFilt8; + // load the last 16 bytes + srcRegFilt8 = _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 7)); + + // merge the last 2 results together + srcRegFilt4 = + _mm_unpacklo_epi8(_mm256_castsi256_si128(srcReg32b7), srcRegFilt8); + srcRegFilt7 = + _mm_unpackhi_epi8(_mm256_castsi256_si128(srcReg32b7), srcRegFilt8); + + // multiply 2 adjacent elements with the filter and add the result + srcRegFilt1 = _mm_maddubs_epi16(_mm256_castsi256_si128(srcReg32b10), + _mm256_castsi256_si128(firstFilters)); + srcRegFilt4 = + _mm_maddubs_epi16(srcRegFilt4, _mm256_castsi256_si128(forthFilters)); + srcRegFilt3 = _mm_maddubs_epi16(_mm256_castsi256_si128(srcReg32b1), + _mm256_castsi256_si128(firstFilters)); + srcRegFilt7 = + _mm_maddubs_epi16(srcRegFilt7, _mm256_castsi256_si128(forthFilters)); + + // add and saturate the results together + srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, srcRegFilt4); + srcRegFilt3 = _mm_adds_epi16(srcRegFilt3, srcRegFilt7); + + // multiply 2 adjacent elements with the filter and add the result + srcRegFilt4 = _mm_maddubs_epi16(_mm256_castsi256_si128(srcReg32b11), + _mm256_castsi256_si128(secondFilters)); + srcRegFilt5 = _mm_maddubs_epi16(_mm256_castsi256_si128(srcReg32b3), + _mm256_castsi256_si128(secondFilters)); + + // multiply 2 adjacent elements with the filter and add the result + srcRegFilt6 = _mm_maddubs_epi16(_mm256_castsi256_si128(srcReg32b2), + _mm256_castsi256_si128(thirdFilters)); + srcRegFilt7 = _mm_maddubs_epi16(_mm256_castsi256_si128(srcReg32b5), + _mm256_castsi256_si128(thirdFilters)); + + // add and saturate the results together + srcRegFilt1 = + _mm_adds_epi16(srcRegFilt1, _mm_min_epi16(srcRegFilt4, srcRegFilt6)); + srcRegFilt3 = + _mm_adds_epi16(srcRegFilt3, _mm_min_epi16(srcRegFilt5, srcRegFilt7)); + + // add and saturate the results together + srcRegFilt1 = + _mm_adds_epi16(srcRegFilt1, _mm_max_epi16(srcRegFilt4, srcRegFilt6)); + srcRegFilt3 = + _mm_adds_epi16(srcRegFilt3, _mm_max_epi16(srcRegFilt5, srcRegFilt7)); + + srcRegFilt1 = + _mm_adds_epi16(srcRegFilt1, _mm256_castsi256_si128(addFilterReg64)); + srcRegFilt3 = + _mm_adds_epi16(srcRegFilt3, _mm256_castsi256_si128(addFilterReg64)); + + // shift by 7 bit each 16 bit + srcRegFilt1 = _mm_srai_epi16(srcRegFilt1, 7); + srcRegFilt3 = _mm_srai_epi16(srcRegFilt3, 7); + + // shrink to 8 bit each 16 bits, the first lane contain the first + // convolve result and the second lane contain the second convolve + // result + srcRegFilt1 = _mm_packus_epi16(srcRegFilt1, srcRegFilt3); + + // save 16 bytes + _mm_store_si128((__m128i *)output_ptr, srcRegFilt1); + } +} + +#if HAVE_AVX2 && HAVE_SSSE3 +filter8_1dfunction aom_filter_block1d4_v8_ssse3; +#if ARCH_X86_64 +filter8_1dfunction aom_filter_block1d8_v8_intrin_ssse3; +filter8_1dfunction aom_filter_block1d8_h8_intrin_ssse3; +filter8_1dfunction aom_filter_block1d4_h8_intrin_ssse3; +#define aom_filter_block1d8_v8_avx2 aom_filter_block1d8_v8_intrin_ssse3 +#define aom_filter_block1d8_h8_avx2 aom_filter_block1d8_h8_intrin_ssse3 +#define aom_filter_block1d4_h8_avx2 aom_filter_block1d4_h8_intrin_ssse3 +#else // ARCH_X86 +filter8_1dfunction aom_filter_block1d8_v8_ssse3; +filter8_1dfunction aom_filter_block1d8_h8_ssse3; +filter8_1dfunction aom_filter_block1d4_h8_ssse3; +#define aom_filter_block1d8_v8_avx2 aom_filter_block1d8_v8_ssse3 +#define aom_filter_block1d8_h8_avx2 aom_filter_block1d8_h8_ssse3 +#define aom_filter_block1d4_h8_avx2 aom_filter_block1d4_h8_ssse3 +#endif // ARCH_X86_64 +filter8_1dfunction aom_filter_block1d16_v2_ssse3; +filter8_1dfunction aom_filter_block1d16_h2_ssse3; +filter8_1dfunction aom_filter_block1d8_v2_ssse3; +filter8_1dfunction aom_filter_block1d8_h2_ssse3; +filter8_1dfunction aom_filter_block1d4_v2_ssse3; +filter8_1dfunction aom_filter_block1d4_h2_ssse3; +#define aom_filter_block1d4_v8_avx2 aom_filter_block1d4_v8_ssse3 +#define aom_filter_block1d16_v2_avx2 aom_filter_block1d16_v2_ssse3 +#define aom_filter_block1d16_h2_avx2 aom_filter_block1d16_h2_ssse3 +#define aom_filter_block1d8_v2_avx2 aom_filter_block1d8_v2_ssse3 +#define aom_filter_block1d8_h2_avx2 aom_filter_block1d8_h2_ssse3 +#define aom_filter_block1d4_v2_avx2 aom_filter_block1d4_v2_ssse3 +#define aom_filter_block1d4_h2_avx2 aom_filter_block1d4_h2_ssse3 +// void aom_convolve8_horiz_avx2(const uint8_t *src, ptrdiff_t src_stride, +// uint8_t *dst, ptrdiff_t dst_stride, +// const int16_t *filter_x, int x_step_q4, +// const int16_t *filter_y, int y_step_q4, +// int w, int h); +// void aom_convolve8_vert_avx2(const uint8_t *src, ptrdiff_t src_stride, +// uint8_t *dst, ptrdiff_t dst_stride, +// const int16_t *filter_x, int x_step_q4, +// const int16_t *filter_y, int y_step_q4, +// int w, int h); +FUN_CONV_1D(horiz, x_step_q4, filter_x, h, src, , avx2); +FUN_CONV_1D(vert, y_step_q4, filter_y, v, src - src_stride * 3, , avx2); + +// void aom_convolve8_avx2(const uint8_t *src, ptrdiff_t src_stride, +// uint8_t *dst, ptrdiff_t dst_stride, +// const int16_t *filter_x, int x_step_q4, +// const int16_t *filter_y, int y_step_q4, +// int w, int h); +FUN_CONV_2D(, avx2); +#endif // HAVE_AX2 && HAVE_SSSE3 diff --git a/third_party/aom/aom_dsp/x86/aom_subpixel_8t_intrin_ssse3.c b/third_party/aom/aom_dsp/x86/aom_subpixel_8t_intrin_ssse3.c new file mode 100644 index 000000000..be37738df --- /dev/null +++ b/third_party/aom/aom_dsp/x86/aom_subpixel_8t_intrin_ssse3.c @@ -0,0 +1,920 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include <tmmintrin.h> + +#include "./aom_dsp_rtcd.h" +#include "aom_dsp/aom_filter.h" +#include "aom_dsp/x86/convolve.h" +#include "aom_mem/aom_mem.h" +#include "aom_ports/mem.h" +#include "aom_ports/emmintrin_compat.h" + +// filters only for the 4_h8 convolution +DECLARE_ALIGNED(16, static const uint8_t, filt1_4_h8[16]) = { + 0, 1, 1, 2, 2, 3, 3, 4, 2, 3, 3, 4, 4, 5, 5, 6 +}; + +DECLARE_ALIGNED(16, static const uint8_t, filt2_4_h8[16]) = { + 4, 5, 5, 6, 6, 7, 7, 8, 6, 7, 7, 8, 8, 9, 9, 10 +}; + +// filters for 8_h8 and 16_h8 +DECLARE_ALIGNED(16, static const uint8_t, filt1_global[16]) = { + 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8 +}; + +DECLARE_ALIGNED(16, static const uint8_t, filt2_global[16]) = { + 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8, 8, 9, 9, 10 +}; + +DECLARE_ALIGNED(16, static const uint8_t, filt3_global[16]) = { + 4, 5, 5, 6, 6, 7, 7, 8, 8, 9, 9, 10, 10, 11, 11, 12 +}; + +DECLARE_ALIGNED(16, static const uint8_t, filt4_global[16]) = { + 6, 7, 7, 8, 8, 9, 9, 10, 10, 11, 11, 12, 12, 13, 13, 14 +}; + +// These are reused by the avx2 intrinsics. +filter8_1dfunction aom_filter_block1d8_v8_intrin_ssse3; +filter8_1dfunction aom_filter_block1d8_h8_intrin_ssse3; +filter8_1dfunction aom_filter_block1d4_h8_intrin_ssse3; + +void aom_filter_block1d4_h8_intrin_ssse3( + const uint8_t *src_ptr, ptrdiff_t src_pixels_per_line, uint8_t *output_ptr, + ptrdiff_t output_pitch, uint32_t output_height, const int16_t *filter) { + __m128i firstFilters, secondFilters, shuffle1, shuffle2; + __m128i srcRegFilt1, srcRegFilt2, srcRegFilt3, srcRegFilt4; + __m128i addFilterReg64, filtersReg, srcReg, minReg; + unsigned int i; + + // create a register with 0,64,0,64,0,64,0,64,0,64,0,64,0,64,0,64 + addFilterReg64 = _mm_set1_epi32((int)0x0400040u); + filtersReg = _mm_loadu_si128((const __m128i *)filter); + // converting the 16 bit (short) to 8 bit (byte) and have the same data + // in both lanes of 128 bit register. + filtersReg = _mm_packs_epi16(filtersReg, filtersReg); + + // duplicate only the first 16 bits in the filter into the first lane + firstFilters = _mm_shufflelo_epi16(filtersReg, 0); + // duplicate only the third 16 bit in the filter into the first lane + secondFilters = _mm_shufflelo_epi16(filtersReg, 0xAAu); + // duplicate only the seconds 16 bits in the filter into the second lane + // firstFilters: k0 k1 k0 k1 k0 k1 k0 k1 k2 k3 k2 k3 k2 k3 k2 k3 + firstFilters = _mm_shufflehi_epi16(firstFilters, 0x55u); + // duplicate only the forth 16 bits in the filter into the second lane + // secondFilters: k4 k5 k4 k5 k4 k5 k4 k5 k6 k7 k6 k7 k6 k7 k6 k7 + secondFilters = _mm_shufflehi_epi16(secondFilters, 0xFFu); + + // loading the local filters + shuffle1 = _mm_load_si128((__m128i const *)filt1_4_h8); + shuffle2 = _mm_load_si128((__m128i const *)filt2_4_h8); + + for (i = 0; i < output_height; i++) { + srcReg = _mm_loadu_si128((const __m128i *)(src_ptr - 3)); + + // filter the source buffer + srcRegFilt1 = _mm_shuffle_epi8(srcReg, shuffle1); + srcRegFilt2 = _mm_shuffle_epi8(srcReg, shuffle2); + + // multiply 2 adjacent elements with the filter and add the result + srcRegFilt1 = _mm_maddubs_epi16(srcRegFilt1, firstFilters); + srcRegFilt2 = _mm_maddubs_epi16(srcRegFilt2, secondFilters); + + // extract the higher half of the lane + srcRegFilt3 = _mm_srli_si128(srcRegFilt1, 8); + srcRegFilt4 = _mm_srli_si128(srcRegFilt2, 8); + + minReg = _mm_min_epi16(srcRegFilt3, srcRegFilt2); + + // add and saturate all the results together + srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, srcRegFilt4); + srcRegFilt3 = _mm_max_epi16(srcRegFilt3, srcRegFilt2); + srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, minReg); + srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, srcRegFilt3); + srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, addFilterReg64); + + // shift by 7 bit each 16 bits + srcRegFilt1 = _mm_srai_epi16(srcRegFilt1, 7); + + // shrink to 8 bit each 16 bits + srcRegFilt1 = _mm_packus_epi16(srcRegFilt1, srcRegFilt1); + src_ptr += src_pixels_per_line; + + // save only 4 bytes + *((int *)&output_ptr[0]) = _mm_cvtsi128_si32(srcRegFilt1); + + output_ptr += output_pitch; + } +} + +void aom_filter_block1d8_h8_intrin_ssse3( + const uint8_t *src_ptr, ptrdiff_t src_pixels_per_line, uint8_t *output_ptr, + ptrdiff_t output_pitch, uint32_t output_height, const int16_t *filter) { + __m128i firstFilters, secondFilters, thirdFilters, forthFilters, srcReg; + __m128i filt1Reg, filt2Reg, filt3Reg, filt4Reg; + __m128i srcRegFilt1, srcRegFilt2, srcRegFilt3, srcRegFilt4; + __m128i addFilterReg64, filtersReg, minReg; + unsigned int i; + + // create a register with 0,64,0,64,0,64,0,64,0,64,0,64,0,64,0,64 + addFilterReg64 = _mm_set1_epi32((int)0x0400040u); + filtersReg = _mm_loadu_si128((const __m128i *)filter); + // converting the 16 bit (short) to 8 bit (byte) and have the same data + // in both lanes of 128 bit register. + filtersReg = _mm_packs_epi16(filtersReg, filtersReg); + + // duplicate only the first 16 bits (first and second byte) + // across 128 bit register + firstFilters = _mm_shuffle_epi8(filtersReg, _mm_set1_epi16(0x100u)); + // duplicate only the second 16 bits (third and forth byte) + // across 128 bit register + secondFilters = _mm_shuffle_epi8(filtersReg, _mm_set1_epi16(0x302u)); + // duplicate only the third 16 bits (fifth and sixth byte) + // across 128 bit register + thirdFilters = _mm_shuffle_epi8(filtersReg, _mm_set1_epi16(0x504u)); + // duplicate only the forth 16 bits (seventh and eighth byte) + // across 128 bit register + forthFilters = _mm_shuffle_epi8(filtersReg, _mm_set1_epi16(0x706u)); + + filt1Reg = _mm_load_si128((__m128i const *)filt1_global); + filt2Reg = _mm_load_si128((__m128i const *)filt2_global); + filt3Reg = _mm_load_si128((__m128i const *)filt3_global); + filt4Reg = _mm_load_si128((__m128i const *)filt4_global); + + for (i = 0; i < output_height; i++) { + srcReg = _mm_loadu_si128((const __m128i *)(src_ptr - 3)); + + // filter the source buffer + srcRegFilt1 = _mm_shuffle_epi8(srcReg, filt1Reg); + srcRegFilt2 = _mm_shuffle_epi8(srcReg, filt2Reg); + + // multiply 2 adjacent elements with the filter and add the result + srcRegFilt1 = _mm_maddubs_epi16(srcRegFilt1, firstFilters); + srcRegFilt2 = _mm_maddubs_epi16(srcRegFilt2, secondFilters); + + // filter the source buffer + srcRegFilt3 = _mm_shuffle_epi8(srcReg, filt3Reg); + srcRegFilt4 = _mm_shuffle_epi8(srcReg, filt4Reg); + + // multiply 2 adjacent elements with the filter and add the result + srcRegFilt3 = _mm_maddubs_epi16(srcRegFilt3, thirdFilters); + srcRegFilt4 = _mm_maddubs_epi16(srcRegFilt4, forthFilters); + + // add and saturate all the results together + minReg = _mm_min_epi16(srcRegFilt2, srcRegFilt3); + srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, srcRegFilt4); + + srcRegFilt2 = _mm_max_epi16(srcRegFilt2, srcRegFilt3); + srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, minReg); + srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, srcRegFilt2); + srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, addFilterReg64); + + // shift by 7 bit each 16 bits + srcRegFilt1 = _mm_srai_epi16(srcRegFilt1, 7); + + // shrink to 8 bit each 16 bits + srcRegFilt1 = _mm_packus_epi16(srcRegFilt1, srcRegFilt1); + + src_ptr += src_pixels_per_line; + + // save only 8 bytes + _mm_storel_epi64((__m128i *)&output_ptr[0], srcRegFilt1); + + output_ptr += output_pitch; + } +} + +void aom_filter_block1d8_v8_intrin_ssse3( + const uint8_t *src_ptr, ptrdiff_t src_pitch, uint8_t *output_ptr, + ptrdiff_t out_pitch, uint32_t output_height, const int16_t *filter) { + __m128i addFilterReg64, filtersReg, minReg; + __m128i firstFilters, secondFilters, thirdFilters, forthFilters; + __m128i srcRegFilt1, srcRegFilt2, srcRegFilt3, srcRegFilt5; + __m128i srcReg1, srcReg2, srcReg3, srcReg4, srcReg5, srcReg6, srcReg7; + __m128i srcReg8; + unsigned int i; + + // create a register with 0,64,0,64,0,64,0,64,0,64,0,64,0,64,0,64 + addFilterReg64 = _mm_set1_epi32((int)0x0400040u); + filtersReg = _mm_loadu_si128((const __m128i *)filter); + // converting the 16 bit (short) to 8 bit (byte) and have the same data + // in both lanes of 128 bit register. + filtersReg = _mm_packs_epi16(filtersReg, filtersReg); + + // duplicate only the first 16 bits in the filter + firstFilters = _mm_shuffle_epi8(filtersReg, _mm_set1_epi16(0x100u)); + // duplicate only the second 16 bits in the filter + secondFilters = _mm_shuffle_epi8(filtersReg, _mm_set1_epi16(0x302u)); + // duplicate only the third 16 bits in the filter + thirdFilters = _mm_shuffle_epi8(filtersReg, _mm_set1_epi16(0x504u)); + // duplicate only the forth 16 bits in the filter + forthFilters = _mm_shuffle_epi8(filtersReg, _mm_set1_epi16(0x706u)); + + // load the first 7 rows of 8 bytes + srcReg1 = _mm_loadl_epi64((const __m128i *)src_ptr); + srcReg2 = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch)); + srcReg3 = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 2)); + srcReg4 = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 3)); + srcReg5 = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 4)); + srcReg6 = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 5)); + srcReg7 = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 6)); + + for (i = 0; i < output_height; i++) { + // load the last 8 bytes + srcReg8 = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 7)); + + // merge the result together + srcRegFilt1 = _mm_unpacklo_epi8(srcReg1, srcReg2); + srcRegFilt3 = _mm_unpacklo_epi8(srcReg3, srcReg4); + + // merge the result together + srcRegFilt2 = _mm_unpacklo_epi8(srcReg5, srcReg6); + srcRegFilt5 = _mm_unpacklo_epi8(srcReg7, srcReg8); + + // multiply 2 adjacent elements with the filter and add the result + srcRegFilt1 = _mm_maddubs_epi16(srcRegFilt1, firstFilters); + srcRegFilt3 = _mm_maddubs_epi16(srcRegFilt3, secondFilters); + srcRegFilt2 = _mm_maddubs_epi16(srcRegFilt2, thirdFilters); + srcRegFilt5 = _mm_maddubs_epi16(srcRegFilt5, forthFilters); + + // add and saturate the results together + minReg = _mm_min_epi16(srcRegFilt2, srcRegFilt3); + srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, srcRegFilt5); + srcRegFilt2 = _mm_max_epi16(srcRegFilt2, srcRegFilt3); + srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, minReg); + srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, srcRegFilt2); + srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, addFilterReg64); + + // shift by 7 bit each 16 bit + srcRegFilt1 = _mm_srai_epi16(srcRegFilt1, 7); + + // shrink to 8 bit each 16 bits + srcRegFilt1 = _mm_packus_epi16(srcRegFilt1, srcRegFilt1); + + src_ptr += src_pitch; + + // shift down a row + srcReg1 = srcReg2; + srcReg2 = srcReg3; + srcReg3 = srcReg4; + srcReg4 = srcReg5; + srcReg5 = srcReg6; + srcReg6 = srcReg7; + srcReg7 = srcReg8; + + // save only 8 bytes convolve result + _mm_storel_epi64((__m128i *)&output_ptr[0], srcRegFilt1); + + output_ptr += out_pitch; + } +} + +filter8_1dfunction aom_filter_block1d16_v8_ssse3; +filter8_1dfunction aom_filter_block1d16_h8_ssse3; +filter8_1dfunction aom_filter_block1d8_v8_ssse3; +filter8_1dfunction aom_filter_block1d8_h8_ssse3; +filter8_1dfunction aom_filter_block1d4_v8_ssse3; +filter8_1dfunction aom_filter_block1d4_h8_ssse3; +filter8_1dfunction aom_filter_block1d16_v8_avg_ssse3; +filter8_1dfunction aom_filter_block1d16_h8_avg_ssse3; +filter8_1dfunction aom_filter_block1d8_v8_avg_ssse3; +filter8_1dfunction aom_filter_block1d8_h8_avg_ssse3; +filter8_1dfunction aom_filter_block1d4_v8_avg_ssse3; +filter8_1dfunction aom_filter_block1d4_h8_avg_ssse3; +#if CONFIG_LOOP_RESTORATION +filter8_1dfunction aom_filter_block1d16_v8_add_src_ssse3; +filter8_1dfunction aom_filter_block1d16_h8_add_src_ssse3; +filter8_1dfunction aom_filter_block1d8_v8_add_src_ssse3; +filter8_1dfunction aom_filter_block1d8_h8_add_src_ssse3; +filter8_1dfunction aom_filter_block1d4_v8_add_src_ssse3; +filter8_1dfunction aom_filter_block1d4_h8_add_src_ssse3; +#endif + +filter8_1dfunction aom_filter_block1d16_v2_ssse3; +filter8_1dfunction aom_filter_block1d16_h2_ssse3; +filter8_1dfunction aom_filter_block1d8_v2_ssse3; +filter8_1dfunction aom_filter_block1d8_h2_ssse3; +filter8_1dfunction aom_filter_block1d4_v2_ssse3; +filter8_1dfunction aom_filter_block1d4_h2_ssse3; +filter8_1dfunction aom_filter_block1d16_v2_avg_ssse3; +filter8_1dfunction aom_filter_block1d16_h2_avg_ssse3; +filter8_1dfunction aom_filter_block1d8_v2_avg_ssse3; +filter8_1dfunction aom_filter_block1d8_h2_avg_ssse3; +filter8_1dfunction aom_filter_block1d4_v2_avg_ssse3; +filter8_1dfunction aom_filter_block1d4_h2_avg_ssse3; + +// void aom_convolve8_horiz_ssse3(const uint8_t *src, ptrdiff_t src_stride, +// uint8_t *dst, ptrdiff_t dst_stride, +// const int16_t *filter_x, int x_step_q4, +// const int16_t *filter_y, int y_step_q4, +// int w, int h); +// void aom_convolve8_vert_ssse3(const uint8_t *src, ptrdiff_t src_stride, +// uint8_t *dst, ptrdiff_t dst_stride, +// const int16_t *filter_x, int x_step_q4, +// const int16_t *filter_y, int y_step_q4, +// int w, int h); +// void aom_convolve8_avg_horiz_ssse3(const uint8_t *src, ptrdiff_t src_stride, +// uint8_t *dst, ptrdiff_t dst_stride, +// const int16_t *filter_x, int x_step_q4, +// const int16_t *filter_y, int y_step_q4, +// int w, int h); +// void aom_convolve8_avg_vert_ssse3(const uint8_t *src, ptrdiff_t src_stride, +// uint8_t *dst, ptrdiff_t dst_stride, +// const int16_t *filter_x, int x_step_q4, +// const int16_t *filter_y, int y_step_q4, +// int w, int h); +FUN_CONV_1D(horiz, x_step_q4, filter_x, h, src, , ssse3); +FUN_CONV_1D(vert, y_step_q4, filter_y, v, src - src_stride * 3, , ssse3); +FUN_CONV_1D(avg_horiz, x_step_q4, filter_x, h, src, avg_, ssse3); +FUN_CONV_1D(avg_vert, y_step_q4, filter_y, v, src - src_stride * 3, avg_, + ssse3); + +#if CONFIG_LOOP_RESTORATION +FUN_CONV_1D_NO_BILINEAR(add_src_horiz, x_step_q4, filter_x, h, src, add_src_, + ssse3); +FUN_CONV_1D_NO_BILINEAR(add_src_vert, y_step_q4, filter_y, v, + src - src_stride * 3, add_src_, ssse3); +#endif + +#define TRANSPOSE_8X8(in0, in1, in2, in3, in4, in5, in6, in7, out0, out1, \ + out2, out3, out4, out5, out6, out7) \ + { \ + const __m128i tr0_0 = _mm_unpacklo_epi8(in0, in1); \ + const __m128i tr0_1 = _mm_unpacklo_epi8(in2, in3); \ + const __m128i tr0_2 = _mm_unpacklo_epi8(in4, in5); \ + const __m128i tr0_3 = _mm_unpacklo_epi8(in6, in7); \ + \ + const __m128i tr1_0 = _mm_unpacklo_epi16(tr0_0, tr0_1); \ + const __m128i tr1_1 = _mm_unpackhi_epi16(tr0_0, tr0_1); \ + const __m128i tr1_2 = _mm_unpacklo_epi16(tr0_2, tr0_3); \ + const __m128i tr1_3 = _mm_unpackhi_epi16(tr0_2, tr0_3); \ + \ + const __m128i tr2_0 = _mm_unpacklo_epi32(tr1_0, tr1_2); \ + const __m128i tr2_1 = _mm_unpackhi_epi32(tr1_0, tr1_2); \ + const __m128i tr2_2 = _mm_unpacklo_epi32(tr1_1, tr1_3); \ + const __m128i tr2_3 = _mm_unpackhi_epi32(tr1_1, tr1_3); \ + \ + out0 = _mm_unpacklo_epi64(tr2_0, tr2_0); \ + out1 = _mm_unpackhi_epi64(tr2_0, tr2_0); \ + out2 = _mm_unpacklo_epi64(tr2_1, tr2_1); \ + out3 = _mm_unpackhi_epi64(tr2_1, tr2_1); \ + out4 = _mm_unpacklo_epi64(tr2_2, tr2_2); \ + out5 = _mm_unpackhi_epi64(tr2_2, tr2_2); \ + out6 = _mm_unpacklo_epi64(tr2_3, tr2_3); \ + out7 = _mm_unpackhi_epi64(tr2_3, tr2_3); \ + } + +static void filter_horiz_w8_ssse3(const uint8_t *src_x, ptrdiff_t src_pitch, + uint8_t *dst, const int16_t *x_filter) { + const __m128i k_256 = _mm_set1_epi16(1 << 8); + const __m128i f_values = _mm_load_si128((const __m128i *)x_filter); + // pack and duplicate the filter values + const __m128i f1f0 = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0200u)); + const __m128i f3f2 = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0604u)); + const __m128i f5f4 = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0a08u)); + const __m128i f7f6 = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0e0cu)); + const __m128i A = _mm_loadl_epi64((const __m128i *)src_x); + const __m128i B = _mm_loadl_epi64((const __m128i *)(src_x + src_pitch)); + const __m128i C = _mm_loadl_epi64((const __m128i *)(src_x + src_pitch * 2)); + const __m128i D = _mm_loadl_epi64((const __m128i *)(src_x + src_pitch * 3)); + const __m128i E = _mm_loadl_epi64((const __m128i *)(src_x + src_pitch * 4)); + const __m128i F = _mm_loadl_epi64((const __m128i *)(src_x + src_pitch * 5)); + const __m128i G = _mm_loadl_epi64((const __m128i *)(src_x + src_pitch * 6)); + const __m128i H = _mm_loadl_epi64((const __m128i *)(src_x + src_pitch * 7)); + // 00 01 10 11 02 03 12 13 04 05 14 15 06 07 16 17 + const __m128i tr0_0 = _mm_unpacklo_epi16(A, B); + // 20 21 30 31 22 23 32 33 24 25 34 35 26 27 36 37 + const __m128i tr0_1 = _mm_unpacklo_epi16(C, D); + // 40 41 50 51 42 43 52 53 44 45 54 55 46 47 56 57 + const __m128i tr0_2 = _mm_unpacklo_epi16(E, F); + // 60 61 70 71 62 63 72 73 64 65 74 75 66 67 76 77 + const __m128i tr0_3 = _mm_unpacklo_epi16(G, H); + // 00 01 10 11 20 21 30 31 02 03 12 13 22 23 32 33 + const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1); + // 04 05 14 15 24 25 34 35 06 07 16 17 26 27 36 37 + const __m128i tr1_1 = _mm_unpackhi_epi32(tr0_0, tr0_1); + // 40 41 50 51 60 61 70 71 42 43 52 53 62 63 72 73 + const __m128i tr1_2 = _mm_unpacklo_epi32(tr0_2, tr0_3); + // 44 45 54 55 64 65 74 75 46 47 56 57 66 67 76 77 + const __m128i tr1_3 = _mm_unpackhi_epi32(tr0_2, tr0_3); + // 00 01 10 11 20 21 30 31 40 41 50 51 60 61 70 71 + const __m128i s1s0 = _mm_unpacklo_epi64(tr1_0, tr1_2); + const __m128i s3s2 = _mm_unpackhi_epi64(tr1_0, tr1_2); + const __m128i s5s4 = _mm_unpacklo_epi64(tr1_1, tr1_3); + const __m128i s7s6 = _mm_unpackhi_epi64(tr1_1, tr1_3); + // multiply 2 adjacent elements with the filter and add the result + const __m128i x0 = _mm_maddubs_epi16(s1s0, f1f0); + const __m128i x1 = _mm_maddubs_epi16(s3s2, f3f2); + const __m128i x2 = _mm_maddubs_epi16(s5s4, f5f4); + const __m128i x3 = _mm_maddubs_epi16(s7s6, f7f6); + // add and saturate the results together + const __m128i min_x2x1 = _mm_min_epi16(x2, x1); + const __m128i max_x2x1 = _mm_max_epi16(x2, x1); + __m128i temp = _mm_adds_epi16(x0, x3); + temp = _mm_adds_epi16(temp, min_x2x1); + temp = _mm_adds_epi16(temp, max_x2x1); + // round and shift by 7 bit each 16 bit + temp = _mm_mulhrs_epi16(temp, k_256); + // shrink to 8 bit each 16 bits + temp = _mm_packus_epi16(temp, temp); + // save only 8 bytes convolve result + _mm_storel_epi64((__m128i *)dst, temp); +} + +static void transpose8x8_to_dst(const uint8_t *src, ptrdiff_t src_stride, + uint8_t *dst, ptrdiff_t dst_stride) { + __m128i A, B, C, D, E, F, G, H; + + A = _mm_loadl_epi64((const __m128i *)src); + B = _mm_loadl_epi64((const __m128i *)(src + src_stride)); + C = _mm_loadl_epi64((const __m128i *)(src + src_stride * 2)); + D = _mm_loadl_epi64((const __m128i *)(src + src_stride * 3)); + E = _mm_loadl_epi64((const __m128i *)(src + src_stride * 4)); + F = _mm_loadl_epi64((const __m128i *)(src + src_stride * 5)); + G = _mm_loadl_epi64((const __m128i *)(src + src_stride * 6)); + H = _mm_loadl_epi64((const __m128i *)(src + src_stride * 7)); + + TRANSPOSE_8X8(A, B, C, D, E, F, G, H, A, B, C, D, E, F, G, H); + + _mm_storel_epi64((__m128i *)dst, A); + _mm_storel_epi64((__m128i *)(dst + dst_stride * 1), B); + _mm_storel_epi64((__m128i *)(dst + dst_stride * 2), C); + _mm_storel_epi64((__m128i *)(dst + dst_stride * 3), D); + _mm_storel_epi64((__m128i *)(dst + dst_stride * 4), E); + _mm_storel_epi64((__m128i *)(dst + dst_stride * 5), F); + _mm_storel_epi64((__m128i *)(dst + dst_stride * 6), G); + _mm_storel_epi64((__m128i *)(dst + dst_stride * 7), H); +} + +static void scaledconvolve_horiz_w8(const uint8_t *src, ptrdiff_t src_stride, + uint8_t *dst, ptrdiff_t dst_stride, + const InterpKernel *x_filters, int x0_q4, + int x_step_q4, int w, int h) { + DECLARE_ALIGNED(16, uint8_t, temp[8 * 8]); + int x, y, z; + src -= SUBPEL_TAPS / 2 - 1; + + // This function processes 8x8 areas. The intermediate height is not always + // a multiple of 8, so force it to be a multiple of 8 here. + y = h + (8 - (h & 0x7)); + + do { + int x_q4 = x0_q4; + for (x = 0; x < w; x += 8) { + // process 8 src_x steps + for (z = 0; z < 8; ++z) { + const uint8_t *const src_x = &src[x_q4 >> SUBPEL_BITS]; + const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK]; + if (x_q4 & SUBPEL_MASK) { + filter_horiz_w8_ssse3(src_x, src_stride, temp + (z * 8), x_filter); + } else { + int i; + for (i = 0; i < 8; ++i) { + temp[z * 8 + i] = src_x[i * src_stride + 3]; + } + } + x_q4 += x_step_q4; + } + + // transpose the 8x8 filters values back to dst + transpose8x8_to_dst(temp, 8, dst + x, dst_stride); + } + + src += src_stride * 8; + dst += dst_stride * 8; + } while (y -= 8); +} + +static void filter_horiz_w4_ssse3(const uint8_t *src_ptr, ptrdiff_t src_pitch, + uint8_t *dst, const int16_t *filter) { + const __m128i k_256 = _mm_set1_epi16(1 << 8); + const __m128i f_values = _mm_load_si128((const __m128i *)filter); + // pack and duplicate the filter values + const __m128i f1f0 = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0200u)); + const __m128i f3f2 = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0604u)); + const __m128i f5f4 = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0a08u)); + const __m128i f7f6 = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0e0cu)); + const __m128i A = _mm_loadl_epi64((const __m128i *)src_ptr); + const __m128i B = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch)); + const __m128i C = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 2)); + const __m128i D = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 3)); + // TRANSPOSE... + // 00 01 02 03 04 05 06 07 + // 10 11 12 13 14 15 16 17 + // 20 21 22 23 24 25 26 27 + // 30 31 32 33 34 35 36 37 + // + // TO + // + // 00 10 20 30 + // 01 11 21 31 + // 02 12 22 32 + // 03 13 23 33 + // 04 14 24 34 + // 05 15 25 35 + // 06 16 26 36 + // 07 17 27 37 + // + // 00 01 10 11 02 03 12 13 04 05 14 15 06 07 16 17 + const __m128i tr0_0 = _mm_unpacklo_epi16(A, B); + // 20 21 30 31 22 23 32 33 24 25 34 35 26 27 36 37 + const __m128i tr0_1 = _mm_unpacklo_epi16(C, D); + // 00 01 10 11 20 21 30 31 02 03 12 13 22 23 32 33 + const __m128i s1s0 = _mm_unpacklo_epi32(tr0_0, tr0_1); + // 04 05 14 15 24 25 34 35 06 07 16 17 26 27 36 37 + const __m128i s5s4 = _mm_unpackhi_epi32(tr0_0, tr0_1); + // 02 03 12 13 22 23 32 33 + const __m128i s3s2 = _mm_srli_si128(s1s0, 8); + // 06 07 16 17 26 27 36 37 + const __m128i s7s6 = _mm_srli_si128(s5s4, 8); + // multiply 2 adjacent elements with the filter and add the result + const __m128i x0 = _mm_maddubs_epi16(s1s0, f1f0); + const __m128i x1 = _mm_maddubs_epi16(s3s2, f3f2); + const __m128i x2 = _mm_maddubs_epi16(s5s4, f5f4); + const __m128i x3 = _mm_maddubs_epi16(s7s6, f7f6); + // add and saturate the results together + const __m128i min_x2x1 = _mm_min_epi16(x2, x1); + const __m128i max_x2x1 = _mm_max_epi16(x2, x1); + __m128i temp = _mm_adds_epi16(x0, x3); + temp = _mm_adds_epi16(temp, min_x2x1); + temp = _mm_adds_epi16(temp, max_x2x1); + // round and shift by 7 bit each 16 bit + temp = _mm_mulhrs_epi16(temp, k_256); + // shrink to 8 bit each 16 bits + temp = _mm_packus_epi16(temp, temp); + // save only 4 bytes + *(int *)dst = _mm_cvtsi128_si32(temp); +} + +static void transpose4x4_to_dst(const uint8_t *src, ptrdiff_t src_stride, + uint8_t *dst, ptrdiff_t dst_stride) { + __m128i A = _mm_cvtsi32_si128(*(const int *)src); + __m128i B = _mm_cvtsi32_si128(*(const int *)(src + src_stride)); + __m128i C = _mm_cvtsi32_si128(*(const int *)(src + src_stride * 2)); + __m128i D = _mm_cvtsi32_si128(*(const int *)(src + src_stride * 3)); + // 00 10 01 11 02 12 03 13 + const __m128i tr0_0 = _mm_unpacklo_epi8(A, B); + // 20 30 21 31 22 32 23 33 + const __m128i tr0_1 = _mm_unpacklo_epi8(C, D); + // 00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33 + A = _mm_unpacklo_epi16(tr0_0, tr0_1); + B = _mm_srli_si128(A, 4); + C = _mm_srli_si128(A, 8); + D = _mm_srli_si128(A, 12); + + *(int *)(dst) = _mm_cvtsi128_si32(A); + *(int *)(dst + dst_stride) = _mm_cvtsi128_si32(B); + *(int *)(dst + dst_stride * 2) = _mm_cvtsi128_si32(C); + *(int *)(dst + dst_stride * 3) = _mm_cvtsi128_si32(D); +} + +static void scaledconvolve_horiz_w4(const uint8_t *src, ptrdiff_t src_stride, + uint8_t *dst, ptrdiff_t dst_stride, + const InterpKernel *x_filters, int x0_q4, + int x_step_q4, int w, int h) { + DECLARE_ALIGNED(16, uint8_t, temp[4 * 4]); + int x, y, z; + src -= SUBPEL_TAPS / 2 - 1; + + for (y = 0; y < h; y += 4) { + int x_q4 = x0_q4; + for (x = 0; x < w; x += 4) { + // process 4 src_x steps + for (z = 0; z < 4; ++z) { + const uint8_t *const src_x = &src[x_q4 >> SUBPEL_BITS]; + const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK]; + if (x_q4 & SUBPEL_MASK) { + filter_horiz_w4_ssse3(src_x, src_stride, temp + (z * 4), x_filter); + } else { + int i; + for (i = 0; i < 4; ++i) { + temp[z * 4 + i] = src_x[i * src_stride + 3]; + } + } + x_q4 += x_step_q4; + } + + // transpose the 4x4 filters values back to dst + transpose4x4_to_dst(temp, 4, dst + x, dst_stride); + } + + src += src_stride * 4; + dst += dst_stride * 4; + } +} + +static void filter_vert_w4_ssse3(const uint8_t *src_ptr, ptrdiff_t src_pitch, + uint8_t *dst, const int16_t *filter) { + const __m128i k_256 = _mm_set1_epi16(1 << 8); + const __m128i f_values = _mm_load_si128((const __m128i *)filter); + // pack and duplicate the filter values + const __m128i f1f0 = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0200u)); + const __m128i f3f2 = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0604u)); + const __m128i f5f4 = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0a08u)); + const __m128i f7f6 = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0e0cu)); + const __m128i A = _mm_cvtsi32_si128(*(const int *)src_ptr); + const __m128i B = _mm_cvtsi32_si128(*(const int *)(src_ptr + src_pitch)); + const __m128i C = _mm_cvtsi32_si128(*(const int *)(src_ptr + src_pitch * 2)); + const __m128i D = _mm_cvtsi32_si128(*(const int *)(src_ptr + src_pitch * 3)); + const __m128i E = _mm_cvtsi32_si128(*(const int *)(src_ptr + src_pitch * 4)); + const __m128i F = _mm_cvtsi32_si128(*(const int *)(src_ptr + src_pitch * 5)); + const __m128i G = _mm_cvtsi32_si128(*(const int *)(src_ptr + src_pitch * 6)); + const __m128i H = _mm_cvtsi32_si128(*(const int *)(src_ptr + src_pitch * 7)); + const __m128i s1s0 = _mm_unpacklo_epi8(A, B); + const __m128i s3s2 = _mm_unpacklo_epi8(C, D); + const __m128i s5s4 = _mm_unpacklo_epi8(E, F); + const __m128i s7s6 = _mm_unpacklo_epi8(G, H); + // multiply 2 adjacent elements with the filter and add the result + const __m128i x0 = _mm_maddubs_epi16(s1s0, f1f0); + const __m128i x1 = _mm_maddubs_epi16(s3s2, f3f2); + const __m128i x2 = _mm_maddubs_epi16(s5s4, f5f4); + const __m128i x3 = _mm_maddubs_epi16(s7s6, f7f6); + // add and saturate the results together + const __m128i min_x2x1 = _mm_min_epi16(x2, x1); + const __m128i max_x2x1 = _mm_max_epi16(x2, x1); + __m128i temp = _mm_adds_epi16(x0, x3); + temp = _mm_adds_epi16(temp, min_x2x1); + temp = _mm_adds_epi16(temp, max_x2x1); + // round and shift by 7 bit each 16 bit + temp = _mm_mulhrs_epi16(temp, k_256); + // shrink to 8 bit each 16 bits + temp = _mm_packus_epi16(temp, temp); + // save only 4 bytes + *(int *)dst = _mm_cvtsi128_si32(temp); +} + +static void scaledconvolve_vert_w4(const uint8_t *src, ptrdiff_t src_stride, + uint8_t *dst, ptrdiff_t dst_stride, + const InterpKernel *y_filters, int y0_q4, + int y_step_q4, int w, int h) { + int y; + int y_q4 = y0_q4; + + src -= src_stride * (SUBPEL_TAPS / 2 - 1); + for (y = 0; y < h; ++y) { + const unsigned char *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride]; + const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK]; + + if (y_q4 & SUBPEL_MASK) { + filter_vert_w4_ssse3(src_y, src_stride, &dst[y * dst_stride], y_filter); + } else { + memcpy(&dst[y * dst_stride], &src_y[3 * src_stride], w); + } + + y_q4 += y_step_q4; + } +} + +static void filter_vert_w8_ssse3(const uint8_t *src_ptr, ptrdiff_t src_pitch, + uint8_t *dst, const int16_t *filter) { + const __m128i k_256 = _mm_set1_epi16(1 << 8); + const __m128i f_values = _mm_load_si128((const __m128i *)filter); + // pack and duplicate the filter values + const __m128i f1f0 = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0200u)); + const __m128i f3f2 = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0604u)); + const __m128i f5f4 = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0a08u)); + const __m128i f7f6 = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0e0cu)); + const __m128i A = _mm_loadl_epi64((const __m128i *)src_ptr); + const __m128i B = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch)); + const __m128i C = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 2)); + const __m128i D = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 3)); + const __m128i E = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 4)); + const __m128i F = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 5)); + const __m128i G = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 6)); + const __m128i H = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 7)); + const __m128i s1s0 = _mm_unpacklo_epi8(A, B); + const __m128i s3s2 = _mm_unpacklo_epi8(C, D); + const __m128i s5s4 = _mm_unpacklo_epi8(E, F); + const __m128i s7s6 = _mm_unpacklo_epi8(G, H); + // multiply 2 adjacent elements with the filter and add the result + const __m128i x0 = _mm_maddubs_epi16(s1s0, f1f0); + const __m128i x1 = _mm_maddubs_epi16(s3s2, f3f2); + const __m128i x2 = _mm_maddubs_epi16(s5s4, f5f4); + const __m128i x3 = _mm_maddubs_epi16(s7s6, f7f6); + // add and saturate the results together + const __m128i min_x2x1 = _mm_min_epi16(x2, x1); + const __m128i max_x2x1 = _mm_max_epi16(x2, x1); + __m128i temp = _mm_adds_epi16(x0, x3); + temp = _mm_adds_epi16(temp, min_x2x1); + temp = _mm_adds_epi16(temp, max_x2x1); + // round and shift by 7 bit each 16 bit + temp = _mm_mulhrs_epi16(temp, k_256); + // shrink to 8 bit each 16 bits + temp = _mm_packus_epi16(temp, temp); + // save only 8 bytes convolve result + _mm_storel_epi64((__m128i *)dst, temp); +} + +static void scaledconvolve_vert_w8(const uint8_t *src, ptrdiff_t src_stride, + uint8_t *dst, ptrdiff_t dst_stride, + const InterpKernel *y_filters, int y0_q4, + int y_step_q4, int w, int h) { + int y; + int y_q4 = y0_q4; + + src -= src_stride * (SUBPEL_TAPS / 2 - 1); + for (y = 0; y < h; ++y) { + const unsigned char *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride]; + const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK]; + if (y_q4 & SUBPEL_MASK) { + filter_vert_w8_ssse3(src_y, src_stride, &dst[y * dst_stride], y_filter); + } else { + memcpy(&dst[y * dst_stride], &src_y[3 * src_stride], w); + } + y_q4 += y_step_q4; + } +} + +static void filter_vert_w16_ssse3(const uint8_t *src_ptr, ptrdiff_t src_pitch, + uint8_t *dst, const int16_t *filter, int w) { + const __m128i k_256 = _mm_set1_epi16(1 << 8); + const __m128i f_values = _mm_load_si128((const __m128i *)filter); + // pack and duplicate the filter values + const __m128i f1f0 = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0200u)); + const __m128i f3f2 = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0604u)); + const __m128i f5f4 = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0a08u)); + const __m128i f7f6 = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0e0cu)); + int i; + + for (i = 0; i < w; i += 16) { + const __m128i A = _mm_loadu_si128((const __m128i *)src_ptr); + const __m128i B = _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch)); + const __m128i C = + _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 2)); + const __m128i D = + _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 3)); + const __m128i E = + _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 4)); + const __m128i F = + _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 5)); + const __m128i G = + _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 6)); + const __m128i H = + _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 7)); + // merge the result together + const __m128i s1s0_lo = _mm_unpacklo_epi8(A, B); + const __m128i s7s6_lo = _mm_unpacklo_epi8(G, H); + const __m128i s1s0_hi = _mm_unpackhi_epi8(A, B); + const __m128i s7s6_hi = _mm_unpackhi_epi8(G, H); + // multiply 2 adjacent elements with the filter and add the result + const __m128i x0_lo = _mm_maddubs_epi16(s1s0_lo, f1f0); + const __m128i x3_lo = _mm_maddubs_epi16(s7s6_lo, f7f6); + const __m128i x0_hi = _mm_maddubs_epi16(s1s0_hi, f1f0); + const __m128i x3_hi = _mm_maddubs_epi16(s7s6_hi, f7f6); + // add and saturate the results together + const __m128i x3x0_lo = _mm_adds_epi16(x0_lo, x3_lo); + const __m128i x3x0_hi = _mm_adds_epi16(x0_hi, x3_hi); + // merge the result together + const __m128i s3s2_lo = _mm_unpacklo_epi8(C, D); + const __m128i s3s2_hi = _mm_unpackhi_epi8(C, D); + // multiply 2 adjacent elements with the filter and add the result + const __m128i x1_lo = _mm_maddubs_epi16(s3s2_lo, f3f2); + const __m128i x1_hi = _mm_maddubs_epi16(s3s2_hi, f3f2); + // merge the result together + const __m128i s5s4_lo = _mm_unpacklo_epi8(E, F); + const __m128i s5s4_hi = _mm_unpackhi_epi8(E, F); + // multiply 2 adjacent elements with the filter and add the result + const __m128i x2_lo = _mm_maddubs_epi16(s5s4_lo, f5f4); + const __m128i x2_hi = _mm_maddubs_epi16(s5s4_hi, f5f4); + // add and saturate the results together + __m128i temp_lo = _mm_adds_epi16(x3x0_lo, _mm_min_epi16(x1_lo, x2_lo)); + __m128i temp_hi = _mm_adds_epi16(x3x0_hi, _mm_min_epi16(x1_hi, x2_hi)); + + // add and saturate the results together + temp_lo = _mm_adds_epi16(temp_lo, _mm_max_epi16(x1_lo, x2_lo)); + temp_hi = _mm_adds_epi16(temp_hi, _mm_max_epi16(x1_hi, x2_hi)); + // round and shift by 7 bit each 16 bit + temp_lo = _mm_mulhrs_epi16(temp_lo, k_256); + temp_hi = _mm_mulhrs_epi16(temp_hi, k_256); + // shrink to 8 bit each 16 bits, the first lane contain the first + // convolve result and the second lane contain the second convolve + // result + temp_hi = _mm_packus_epi16(temp_lo, temp_hi); + src_ptr += 16; + // save 16 bytes convolve result + _mm_store_si128((__m128i *)&dst[i], temp_hi); + } +} + +static void scaledconvolve_vert_w16(const uint8_t *src, ptrdiff_t src_stride, + uint8_t *dst, ptrdiff_t dst_stride, + const InterpKernel *y_filters, int y0_q4, + int y_step_q4, int w, int h) { + int y; + int y_q4 = y0_q4; + + src -= src_stride * (SUBPEL_TAPS / 2 - 1); + for (y = 0; y < h; ++y) { + const unsigned char *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride]; + const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK]; + if (y_q4 & SUBPEL_MASK) { + filter_vert_w16_ssse3(src_y, src_stride, &dst[y * dst_stride], y_filter, + w); + } else { + memcpy(&dst[y * dst_stride], &src_y[3 * src_stride], w); + } + y_q4 += y_step_q4; + } +} + +static void scaledconvolve2d(const uint8_t *src, ptrdiff_t src_stride, + uint8_t *dst, ptrdiff_t dst_stride, + const InterpKernel *const x_filters, int x0_q4, + int x_step_q4, const InterpKernel *const y_filters, + int y0_q4, int y_step_q4, int w, int h) { + // Note: Fixed size intermediate buffer, temp, places limits on parameters. + // 2d filtering proceeds in 2 steps: + // (1) Interpolate horizontally into an intermediate buffer, temp. + // (2) Interpolate temp vertically to derive the sub-pixel result. + // Deriving the maximum number of rows in the temp buffer (135): + // --Smallest scaling factor is x1/2 ==> y_step_q4 = 32 (Normative). + // --Largest block size is 64x64 pixels. + // --64 rows in the downscaled frame span a distance of (64 - 1) * 32 in the + // original frame (in 1/16th pixel units). + // --Must round-up because block may be located at sub-pixel position. + // --Require an additional SUBPEL_TAPS rows for the 8-tap filter tails. + // --((64 - 1) * 32 + 15) >> 4 + 8 = 135. + // --Require an additional 8 rows for the horiz_w8 transpose tail. + DECLARE_ALIGNED(16, uint8_t, temp[(MAX_EXT_SIZE + 8) * MAX_SB_SIZE]); + const int intermediate_height = + (((h - 1) * y_step_q4 + y0_q4) >> SUBPEL_BITS) + SUBPEL_TAPS; + + assert(w <= MAX_SB_SIZE); + assert(h <= MAX_SB_SIZE); + assert(y_step_q4 <= 32); + assert(x_step_q4 <= 32); + + if (w >= 8) { + scaledconvolve_horiz_w8(src - src_stride * (SUBPEL_TAPS / 2 - 1), + src_stride, temp, MAX_SB_SIZE, x_filters, x0_q4, + x_step_q4, w, intermediate_height); + } else { + scaledconvolve_horiz_w4(src - src_stride * (SUBPEL_TAPS / 2 - 1), + src_stride, temp, MAX_SB_SIZE, x_filters, x0_q4, + x_step_q4, w, intermediate_height); + } + + if (w >= 16) { + scaledconvolve_vert_w16(temp + MAX_SB_SIZE * (SUBPEL_TAPS / 2 - 1), + MAX_SB_SIZE, dst, dst_stride, y_filters, y0_q4, + y_step_q4, w, h); + } else if (w == 8) { + scaledconvolve_vert_w8(temp + MAX_SB_SIZE * (SUBPEL_TAPS / 2 - 1), + MAX_SB_SIZE, dst, dst_stride, y_filters, y0_q4, + y_step_q4, w, h); + } else { + scaledconvolve_vert_w4(temp + MAX_SB_SIZE * (SUBPEL_TAPS / 2 - 1), + MAX_SB_SIZE, dst, dst_stride, y_filters, y0_q4, + y_step_q4, w, h); + } +} + +static const InterpKernel *get_filter_base(const int16_t *filter) { + // NOTE: This assumes that the filter table is 256-byte aligned. + // TODO(agrange) Modify to make independent of table alignment. + return (const InterpKernel *)(((intptr_t)filter) & ~((intptr_t)0xFF)); +} + +static int get_filter_offset(const int16_t *f, const InterpKernel *base) { + return (int)((const InterpKernel *)(intptr_t)f - base); +} + +void aom_scaled_2d_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, + ptrdiff_t dst_stride, const int16_t *filter_x, + int x_step_q4, const int16_t *filter_y, int y_step_q4, + int w, int h) { + const InterpKernel *const filters_x = get_filter_base(filter_x); + const int x0_q4 = get_filter_offset(filter_x, filters_x); + + const InterpKernel *const filters_y = get_filter_base(filter_y); + const int y0_q4 = get_filter_offset(filter_y, filters_y); + + scaledconvolve2d(src, src_stride, dst, dst_stride, filters_x, x0_q4, + x_step_q4, filters_y, y0_q4, y_step_q4, w, h); +} + +// void aom_convolve8_ssse3(const uint8_t *src, ptrdiff_t src_stride, +// uint8_t *dst, ptrdiff_t dst_stride, +// const int16_t *filter_x, int x_step_q4, +// const int16_t *filter_y, int y_step_q4, +// int w, int h); +// void aom_convolve8_avg_ssse3(const uint8_t *src, ptrdiff_t src_stride, +// uint8_t *dst, ptrdiff_t dst_stride, +// const int16_t *filter_x, int x_step_q4, +// const int16_t *filter_y, int y_step_q4, +// int w, int h); +FUN_CONV_2D(, ssse3); +FUN_CONV_2D(avg_, ssse3); +#if CONFIG_LOOP_RESTORATION +FUN_CONV_2D_NO_BILINEAR(add_src_, add_src_, ssse3); +#endif diff --git a/third_party/aom/aom_dsp/x86/aom_subpixel_8t_sse2.asm b/third_party/aom/aom_dsp/x86/aom_subpixel_8t_sse2.asm new file mode 100644 index 000000000..b946010d3 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/aom_subpixel_8t_sse2.asm @@ -0,0 +1,990 @@ +; +; Copyright (c) 2016, Alliance for Open Media. All rights reserved +; +; This source code is subject to the terms of the BSD 2 Clause License and +; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License +; was not distributed with this source code in the LICENSE file, you can +; obtain it at www.aomedia.org/license/software. If the Alliance for Open +; Media Patent License 1.0 was not distributed with this source code in the +; PATENTS file, you can obtain it at www.aomedia.org/license/patent. +; + +; + + +%include "aom_ports/x86_abi_support.asm" + +;Note: tap3 and tap4 have to be applied and added after other taps to avoid +;overflow. + +%macro GET_FILTERS_4 0 + mov rdx, arg(5) ;filter ptr + mov rcx, 0x0400040 + + movdqa xmm7, [rdx] ;load filters + pshuflw xmm0, xmm7, 0b ;k0 + pshuflw xmm1, xmm7, 01010101b ;k1 + pshuflw xmm2, xmm7, 10101010b ;k2 + pshuflw xmm3, xmm7, 11111111b ;k3 + psrldq xmm7, 8 + pshuflw xmm4, xmm7, 0b ;k4 + pshuflw xmm5, xmm7, 01010101b ;k5 + pshuflw xmm6, xmm7, 10101010b ;k6 + pshuflw xmm7, xmm7, 11111111b ;k7 + + punpcklqdq xmm0, xmm1 + punpcklqdq xmm2, xmm3 + punpcklqdq xmm5, xmm4 + punpcklqdq xmm6, xmm7 + + movdqa k0k1, xmm0 + movdqa k2k3, xmm2 + movdqa k5k4, xmm5 + movdqa k6k7, xmm6 + + movq xmm6, rcx + pshufd xmm6, xmm6, 0 + movdqa krd, xmm6 + + pxor xmm7, xmm7 + movdqa zero, xmm7 +%endm + +%macro APPLY_FILTER_4 1 + punpckldq xmm0, xmm1 ;two row in one register + punpckldq xmm6, xmm7 + punpckldq xmm2, xmm3 + punpckldq xmm5, xmm4 + + punpcklbw xmm0, zero ;unpack to word + punpcklbw xmm6, zero + punpcklbw xmm2, zero + punpcklbw xmm5, zero + + pmullw xmm0, k0k1 ;multiply the filter factors + pmullw xmm6, k6k7 + pmullw xmm2, k2k3 + pmullw xmm5, k5k4 + + paddsw xmm0, xmm6 ;sum + movdqa xmm1, xmm0 + psrldq xmm1, 8 + paddsw xmm0, xmm1 + paddsw xmm0, xmm2 + psrldq xmm2, 8 + paddsw xmm0, xmm5 + psrldq xmm5, 8 + paddsw xmm0, xmm2 + paddsw xmm0, xmm5 + + paddsw xmm0, krd ;rounding + psraw xmm0, 7 ;shift + packuswb xmm0, xmm0 ;pack to byte + +%if %1 + movd xmm1, [rdi] + pavgb xmm0, xmm1 +%endif + movd [rdi], xmm0 +%endm + +%macro GET_FILTERS 0 + mov rdx, arg(5) ;filter ptr + mov rsi, arg(0) ;src_ptr + mov rdi, arg(2) ;output_ptr + mov rcx, 0x0400040 + + movdqa xmm7, [rdx] ;load filters + pshuflw xmm0, xmm7, 0b ;k0 + pshuflw xmm1, xmm7, 01010101b ;k1 + pshuflw xmm2, xmm7, 10101010b ;k2 + pshuflw xmm3, xmm7, 11111111b ;k3 + pshufhw xmm4, xmm7, 0b ;k4 + pshufhw xmm5, xmm7, 01010101b ;k5 + pshufhw xmm6, xmm7, 10101010b ;k6 + pshufhw xmm7, xmm7, 11111111b ;k7 + + punpcklwd xmm0, xmm0 + punpcklwd xmm1, xmm1 + punpcklwd xmm2, xmm2 + punpcklwd xmm3, xmm3 + punpckhwd xmm4, xmm4 + punpckhwd xmm5, xmm5 + punpckhwd xmm6, xmm6 + punpckhwd xmm7, xmm7 + + movdqa k0, xmm0 ;store filter factors on stack + movdqa k1, xmm1 + movdqa k2, xmm2 + movdqa k3, xmm3 + movdqa k4, xmm4 + movdqa k5, xmm5 + movdqa k6, xmm6 + movdqa k7, xmm7 + + movq xmm6, rcx + pshufd xmm6, xmm6, 0 + movdqa krd, xmm6 ;rounding + + pxor xmm7, xmm7 + movdqa zero, xmm7 +%endm + +%macro LOAD_VERT_8 1 + movq xmm0, [rsi + %1] ;0 + movq xmm1, [rsi + rax + %1] ;1 + movq xmm6, [rsi + rdx * 2 + %1] ;6 + lea rsi, [rsi + rax] + movq xmm7, [rsi + rdx * 2 + %1] ;7 + movq xmm2, [rsi + rax + %1] ;2 + movq xmm3, [rsi + rax * 2 + %1] ;3 + movq xmm4, [rsi + rdx + %1] ;4 + movq xmm5, [rsi + rax * 4 + %1] ;5 +%endm + +%macro APPLY_FILTER_8 2 + punpcklbw xmm0, zero + punpcklbw xmm1, zero + punpcklbw xmm6, zero + punpcklbw xmm7, zero + punpcklbw xmm2, zero + punpcklbw xmm5, zero + punpcklbw xmm3, zero + punpcklbw xmm4, zero + + pmullw xmm0, k0 + pmullw xmm1, k1 + pmullw xmm6, k6 + pmullw xmm7, k7 + pmullw xmm2, k2 + pmullw xmm5, k5 + pmullw xmm3, k3 + pmullw xmm4, k4 + + paddsw xmm0, xmm1 + paddsw xmm0, xmm6 + paddsw xmm0, xmm7 + paddsw xmm0, xmm2 + paddsw xmm0, xmm5 + paddsw xmm0, xmm3 + paddsw xmm0, xmm4 + + paddsw xmm0, krd ;rounding + psraw xmm0, 7 ;shift + packuswb xmm0, xmm0 ;pack back to byte +%if %1 + movq xmm1, [rdi + %2] + pavgb xmm0, xmm1 +%endif + movq [rdi + %2], xmm0 +%endm + +;void aom_filter_block1d4_v8_sse2 +;( +; unsigned char *src_ptr, +; unsigned int src_pitch, +; unsigned char *output_ptr, +; unsigned int out_pitch, +; unsigned int output_height, +; short *filter +;) +global sym(aom_filter_block1d4_v8_sse2) PRIVATE +sym(aom_filter_block1d4_v8_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 6 + SAVE_XMM 7 + push rsi + push rdi + push rbx + ; end prolog + + ALIGN_STACK 16, rax + sub rsp, 16 * 6 + %define k0k1 [rsp + 16 * 0] + %define k2k3 [rsp + 16 * 1] + %define k5k4 [rsp + 16 * 2] + %define k6k7 [rsp + 16 * 3] + %define krd [rsp + 16 * 4] + %define zero [rsp + 16 * 5] + + GET_FILTERS_4 + + mov rsi, arg(0) ;src_ptr + mov rdi, arg(2) ;output_ptr + + movsxd rax, DWORD PTR arg(1) ;pixels_per_line + movsxd rbx, DWORD PTR arg(3) ;out_pitch + lea rdx, [rax + rax * 2] + movsxd rcx, DWORD PTR arg(4) ;output_height + +.loop: + movd xmm0, [rsi] ;load src: row 0 + movd xmm1, [rsi + rax] ;1 + movd xmm6, [rsi + rdx * 2] ;6 + lea rsi, [rsi + rax] + movd xmm7, [rsi + rdx * 2] ;7 + movd xmm2, [rsi + rax] ;2 + movd xmm3, [rsi + rax * 2] ;3 + movd xmm4, [rsi + rdx] ;4 + movd xmm5, [rsi + rax * 4] ;5 + + APPLY_FILTER_4 0 + + lea rdi, [rdi + rbx] + dec rcx + jnz .loop + + add rsp, 16 * 6 + pop rsp + pop rbx + ; begin epilog + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret + +;void aom_filter_block1d8_v8_sse2 +;( +; unsigned char *src_ptr, +; unsigned int src_pitch, +; unsigned char *output_ptr, +; unsigned int out_pitch, +; unsigned int output_height, +; short *filter +;) +global sym(aom_filter_block1d8_v8_sse2) PRIVATE +sym(aom_filter_block1d8_v8_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 6 + SAVE_XMM 7 + push rsi + push rdi + push rbx + ; end prolog + + ALIGN_STACK 16, rax + sub rsp, 16 * 10 + %define k0 [rsp + 16 * 0] + %define k1 [rsp + 16 * 1] + %define k2 [rsp + 16 * 2] + %define k3 [rsp + 16 * 3] + %define k4 [rsp + 16 * 4] + %define k5 [rsp + 16 * 5] + %define k6 [rsp + 16 * 6] + %define k7 [rsp + 16 * 7] + %define krd [rsp + 16 * 8] + %define zero [rsp + 16 * 9] + + GET_FILTERS + + movsxd rax, DWORD PTR arg(1) ;pixels_per_line + movsxd rbx, DWORD PTR arg(3) ;out_pitch + lea rdx, [rax + rax * 2] + movsxd rcx, DWORD PTR arg(4) ;output_height + +.loop: + LOAD_VERT_8 0 + APPLY_FILTER_8 0, 0 + + lea rdi, [rdi + rbx] + dec rcx + jnz .loop + + add rsp, 16 * 10 + pop rsp + pop rbx + ; begin epilog + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret + +;void aom_filter_block1d16_v8_sse2 +;( +; unsigned char *src_ptr, +; unsigned int src_pitch, +; unsigned char *output_ptr, +; unsigned int out_pitch, +; unsigned int output_height, +; short *filter +;) +global sym(aom_filter_block1d16_v8_sse2) PRIVATE +sym(aom_filter_block1d16_v8_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 6 + SAVE_XMM 7 + push rsi + push rdi + push rbx + ; end prolog + + ALIGN_STACK 16, rax + sub rsp, 16 * 10 + %define k0 [rsp + 16 * 0] + %define k1 [rsp + 16 * 1] + %define k2 [rsp + 16 * 2] + %define k3 [rsp + 16 * 3] + %define k4 [rsp + 16 * 4] + %define k5 [rsp + 16 * 5] + %define k6 [rsp + 16 * 6] + %define k7 [rsp + 16 * 7] + %define krd [rsp + 16 * 8] + %define zero [rsp + 16 * 9] + + GET_FILTERS + + movsxd rax, DWORD PTR arg(1) ;pixels_per_line + movsxd rbx, DWORD PTR arg(3) ;out_pitch + lea rdx, [rax + rax * 2] + movsxd rcx, DWORD PTR arg(4) ;output_height + +.loop: + LOAD_VERT_8 0 + APPLY_FILTER_8 0, 0 + sub rsi, rax + + LOAD_VERT_8 8 + APPLY_FILTER_8 0, 8 + add rdi, rbx + + dec rcx + jnz .loop + + add rsp, 16 * 10 + pop rsp + pop rbx + ; begin epilog + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret + +global sym(aom_filter_block1d4_v8_avg_sse2) PRIVATE +sym(aom_filter_block1d4_v8_avg_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 6 + SAVE_XMM 7 + push rsi + push rdi + push rbx + ; end prolog + + ALIGN_STACK 16, rax + sub rsp, 16 * 6 + %define k0k1 [rsp + 16 * 0] + %define k2k3 [rsp + 16 * 1] + %define k5k4 [rsp + 16 * 2] + %define k6k7 [rsp + 16 * 3] + %define krd [rsp + 16 * 4] + %define zero [rsp + 16 * 5] + + GET_FILTERS_4 + + mov rsi, arg(0) ;src_ptr + mov rdi, arg(2) ;output_ptr + + movsxd rax, DWORD PTR arg(1) ;pixels_per_line + movsxd rbx, DWORD PTR arg(3) ;out_pitch + lea rdx, [rax + rax * 2] + movsxd rcx, DWORD PTR arg(4) ;output_height + +.loop: + movd xmm0, [rsi] ;load src: row 0 + movd xmm1, [rsi + rax] ;1 + movd xmm6, [rsi + rdx * 2] ;6 + lea rsi, [rsi + rax] + movd xmm7, [rsi + rdx * 2] ;7 + movd xmm2, [rsi + rax] ;2 + movd xmm3, [rsi + rax * 2] ;3 + movd xmm4, [rsi + rdx] ;4 + movd xmm5, [rsi + rax * 4] ;5 + + APPLY_FILTER_4 1 + + lea rdi, [rdi + rbx] + dec rcx + jnz .loop + + add rsp, 16 * 6 + pop rsp + pop rbx + ; begin epilog + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret + +global sym(aom_filter_block1d8_v8_avg_sse2) PRIVATE +sym(aom_filter_block1d8_v8_avg_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 6 + SAVE_XMM 7 + push rsi + push rdi + push rbx + ; end prolog + + ALIGN_STACK 16, rax + sub rsp, 16 * 10 + %define k0 [rsp + 16 * 0] + %define k1 [rsp + 16 * 1] + %define k2 [rsp + 16 * 2] + %define k3 [rsp + 16 * 3] + %define k4 [rsp + 16 * 4] + %define k5 [rsp + 16 * 5] + %define k6 [rsp + 16 * 6] + %define k7 [rsp + 16 * 7] + %define krd [rsp + 16 * 8] + %define zero [rsp + 16 * 9] + + GET_FILTERS + + movsxd rax, DWORD PTR arg(1) ;pixels_per_line + movsxd rbx, DWORD PTR arg(3) ;out_pitch + lea rdx, [rax + rax * 2] + movsxd rcx, DWORD PTR arg(4) ;output_height +.loop: + LOAD_VERT_8 0 + APPLY_FILTER_8 1, 0 + + lea rdi, [rdi + rbx] + dec rcx + jnz .loop + + add rsp, 16 * 10 + pop rsp + pop rbx + ; begin epilog + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret + +global sym(aom_filter_block1d16_v8_avg_sse2) PRIVATE +sym(aom_filter_block1d16_v8_avg_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 6 + SAVE_XMM 7 + push rsi + push rdi + push rbx + ; end prolog + + ALIGN_STACK 16, rax + sub rsp, 16 * 10 + %define k0 [rsp + 16 * 0] + %define k1 [rsp + 16 * 1] + %define k2 [rsp + 16 * 2] + %define k3 [rsp + 16 * 3] + %define k4 [rsp + 16 * 4] + %define k5 [rsp + 16 * 5] + %define k6 [rsp + 16 * 6] + %define k7 [rsp + 16 * 7] + %define krd [rsp + 16 * 8] + %define zero [rsp + 16 * 9] + + GET_FILTERS + + movsxd rax, DWORD PTR arg(1) ;pixels_per_line + movsxd rbx, DWORD PTR arg(3) ;out_pitch + lea rdx, [rax + rax * 2] + movsxd rcx, DWORD PTR arg(4) ;output_height +.loop: + LOAD_VERT_8 0 + APPLY_FILTER_8 1, 0 + sub rsi, rax + + LOAD_VERT_8 8 + APPLY_FILTER_8 1, 8 + add rdi, rbx + + dec rcx + jnz .loop + + add rsp, 16 * 10 + pop rsp + pop rbx + ; begin epilog + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret + +;void aom_filter_block1d4_h8_sse2 +;( +; unsigned char *src_ptr, +; unsigned int src_pixels_per_line, +; unsigned char *output_ptr, +; unsigned int output_pitch, +; unsigned int output_height, +; short *filter +;) +global sym(aom_filter_block1d4_h8_sse2) PRIVATE +sym(aom_filter_block1d4_h8_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 6 + SAVE_XMM 7 + push rsi + push rdi + ; end prolog + + ALIGN_STACK 16, rax + sub rsp, 16 * 6 + %define k0k1 [rsp + 16 * 0] + %define k2k3 [rsp + 16 * 1] + %define k5k4 [rsp + 16 * 2] + %define k6k7 [rsp + 16 * 3] + %define krd [rsp + 16 * 4] + %define zero [rsp + 16 * 5] + + GET_FILTERS_4 + + mov rsi, arg(0) ;src_ptr + mov rdi, arg(2) ;output_ptr + + movsxd rax, DWORD PTR arg(1) ;pixels_per_line + movsxd rdx, DWORD PTR arg(3) ;out_pitch + movsxd rcx, DWORD PTR arg(4) ;output_height + +.loop: + movdqu xmm0, [rsi - 3] ;load src + + movdqa xmm1, xmm0 + movdqa xmm6, xmm0 + movdqa xmm7, xmm0 + movdqa xmm2, xmm0 + movdqa xmm3, xmm0 + movdqa xmm5, xmm0 + movdqa xmm4, xmm0 + + psrldq xmm1, 1 + psrldq xmm6, 6 + psrldq xmm7, 7 + psrldq xmm2, 2 + psrldq xmm3, 3 + psrldq xmm5, 5 + psrldq xmm4, 4 + + APPLY_FILTER_4 0 + + lea rsi, [rsi + rax] + lea rdi, [rdi + rdx] + dec rcx + jnz .loop + + add rsp, 16 * 6 + pop rsp + + ; begin epilog + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret + +;void aom_filter_block1d8_h8_sse2 +;( +; unsigned char *src_ptr, +; unsigned int src_pixels_per_line, +; unsigned char *output_ptr, +; unsigned int output_pitch, +; unsigned int output_height, +; short *filter +;) +global sym(aom_filter_block1d8_h8_sse2) PRIVATE +sym(aom_filter_block1d8_h8_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 6 + SAVE_XMM 7 + push rsi + push rdi + ; end prolog + + ALIGN_STACK 16, rax + sub rsp, 16 * 10 + %define k0 [rsp + 16 * 0] + %define k1 [rsp + 16 * 1] + %define k2 [rsp + 16 * 2] + %define k3 [rsp + 16 * 3] + %define k4 [rsp + 16 * 4] + %define k5 [rsp + 16 * 5] + %define k6 [rsp + 16 * 6] + %define k7 [rsp + 16 * 7] + %define krd [rsp + 16 * 8] + %define zero [rsp + 16 * 9] + + GET_FILTERS + + movsxd rax, DWORD PTR arg(1) ;pixels_per_line + movsxd rdx, DWORD PTR arg(3) ;out_pitch + movsxd rcx, DWORD PTR arg(4) ;output_height + +.loop: + movdqu xmm0, [rsi - 3] ;load src + + movdqa xmm1, xmm0 + movdqa xmm6, xmm0 + movdqa xmm7, xmm0 + movdqa xmm2, xmm0 + movdqa xmm5, xmm0 + movdqa xmm3, xmm0 + movdqa xmm4, xmm0 + + psrldq xmm1, 1 + psrldq xmm6, 6 + psrldq xmm7, 7 + psrldq xmm2, 2 + psrldq xmm5, 5 + psrldq xmm3, 3 + psrldq xmm4, 4 + + APPLY_FILTER_8 0, 0 + + lea rsi, [rsi + rax] + lea rdi, [rdi + rdx] + dec rcx + jnz .loop + + add rsp, 16 * 10 + pop rsp + + ; begin epilog + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret + +;void aom_filter_block1d16_h8_sse2 +;( +; unsigned char *src_ptr, +; unsigned int src_pixels_per_line, +; unsigned char *output_ptr, +; unsigned int output_pitch, +; unsigned int output_height, +; short *filter +;) +global sym(aom_filter_block1d16_h8_sse2) PRIVATE +sym(aom_filter_block1d16_h8_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 6 + SAVE_XMM 7 + push rsi + push rdi + ; end prolog + + ALIGN_STACK 16, rax + sub rsp, 16 * 10 + %define k0 [rsp + 16 * 0] + %define k1 [rsp + 16 * 1] + %define k2 [rsp + 16 * 2] + %define k3 [rsp + 16 * 3] + %define k4 [rsp + 16 * 4] + %define k5 [rsp + 16 * 5] + %define k6 [rsp + 16 * 6] + %define k7 [rsp + 16 * 7] + %define krd [rsp + 16 * 8] + %define zero [rsp + 16 * 9] + + GET_FILTERS + + movsxd rax, DWORD PTR arg(1) ;pixels_per_line + movsxd rdx, DWORD PTR arg(3) ;out_pitch + movsxd rcx, DWORD PTR arg(4) ;output_height + +.loop: + movdqu xmm0, [rsi - 3] ;load src + + movdqa xmm1, xmm0 + movdqa xmm6, xmm0 + movdqa xmm7, xmm0 + movdqa xmm2, xmm0 + movdqa xmm5, xmm0 + movdqa xmm3, xmm0 + movdqa xmm4, xmm0 + + psrldq xmm1, 1 + psrldq xmm6, 6 + psrldq xmm7, 7 + psrldq xmm2, 2 + psrldq xmm5, 5 + psrldq xmm3, 3 + psrldq xmm4, 4 + + APPLY_FILTER_8 0, 0 + + movdqu xmm0, [rsi + 5] ;load src + + movdqa xmm1, xmm0 + movdqa xmm6, xmm0 + movdqa xmm7, xmm0 + movdqa xmm2, xmm0 + movdqa xmm5, xmm0 + movdqa xmm3, xmm0 + movdqa xmm4, xmm0 + + psrldq xmm1, 1 + psrldq xmm6, 6 + psrldq xmm7, 7 + psrldq xmm2, 2 + psrldq xmm5, 5 + psrldq xmm3, 3 + psrldq xmm4, 4 + + APPLY_FILTER_8 0, 8 + + lea rsi, [rsi + rax] + lea rdi, [rdi + rdx] + dec rcx + jnz .loop + + add rsp, 16 * 10 + pop rsp + + ; begin epilog + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret + +global sym(aom_filter_block1d4_h8_avg_sse2) PRIVATE +sym(aom_filter_block1d4_h8_avg_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 6 + SAVE_XMM 7 + push rsi + push rdi + ; end prolog + + ALIGN_STACK 16, rax + sub rsp, 16 * 6 + %define k0k1 [rsp + 16 * 0] + %define k2k3 [rsp + 16 * 1] + %define k5k4 [rsp + 16 * 2] + %define k6k7 [rsp + 16 * 3] + %define krd [rsp + 16 * 4] + %define zero [rsp + 16 * 5] + + GET_FILTERS_4 + + mov rsi, arg(0) ;src_ptr + mov rdi, arg(2) ;output_ptr + + movsxd rax, DWORD PTR arg(1) ;pixels_per_line + movsxd rdx, DWORD PTR arg(3) ;out_pitch + movsxd rcx, DWORD PTR arg(4) ;output_height + +.loop: + movdqu xmm0, [rsi - 3] ;load src + + movdqa xmm1, xmm0 + movdqa xmm6, xmm0 + movdqa xmm7, xmm0 + movdqa xmm2, xmm0 + movdqa xmm3, xmm0 + movdqa xmm5, xmm0 + movdqa xmm4, xmm0 + + psrldq xmm1, 1 + psrldq xmm6, 6 + psrldq xmm7, 7 + psrldq xmm2, 2 + psrldq xmm3, 3 + psrldq xmm5, 5 + psrldq xmm4, 4 + + APPLY_FILTER_4 1 + + lea rsi, [rsi + rax] + lea rdi, [rdi + rdx] + dec rcx + jnz .loop + + add rsp, 16 * 6 + pop rsp + + ; begin epilog + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret + +global sym(aom_filter_block1d8_h8_avg_sse2) PRIVATE +sym(aom_filter_block1d8_h8_avg_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 6 + SAVE_XMM 7 + push rsi + push rdi + ; end prolog + + ALIGN_STACK 16, rax + sub rsp, 16 * 10 + %define k0 [rsp + 16 * 0] + %define k1 [rsp + 16 * 1] + %define k2 [rsp + 16 * 2] + %define k3 [rsp + 16 * 3] + %define k4 [rsp + 16 * 4] + %define k5 [rsp + 16 * 5] + %define k6 [rsp + 16 * 6] + %define k7 [rsp + 16 * 7] + %define krd [rsp + 16 * 8] + %define zero [rsp + 16 * 9] + + GET_FILTERS + + movsxd rax, DWORD PTR arg(1) ;pixels_per_line + movsxd rdx, DWORD PTR arg(3) ;out_pitch + movsxd rcx, DWORD PTR arg(4) ;output_height + +.loop: + movdqu xmm0, [rsi - 3] ;load src + + movdqa xmm1, xmm0 + movdqa xmm6, xmm0 + movdqa xmm7, xmm0 + movdqa xmm2, xmm0 + movdqa xmm5, xmm0 + movdqa xmm3, xmm0 + movdqa xmm4, xmm0 + + psrldq xmm1, 1 + psrldq xmm6, 6 + psrldq xmm7, 7 + psrldq xmm2, 2 + psrldq xmm5, 5 + psrldq xmm3, 3 + psrldq xmm4, 4 + + APPLY_FILTER_8 1, 0 + + lea rsi, [rsi + rax] + lea rdi, [rdi + rdx] + dec rcx + jnz .loop + + add rsp, 16 * 10 + pop rsp + + ; begin epilog + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret + +global sym(aom_filter_block1d16_h8_avg_sse2) PRIVATE +sym(aom_filter_block1d16_h8_avg_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 6 + SAVE_XMM 7 + push rsi + push rdi + ; end prolog + + ALIGN_STACK 16, rax + sub rsp, 16 * 10 + %define k0 [rsp + 16 * 0] + %define k1 [rsp + 16 * 1] + %define k2 [rsp + 16 * 2] + %define k3 [rsp + 16 * 3] + %define k4 [rsp + 16 * 4] + %define k5 [rsp + 16 * 5] + %define k6 [rsp + 16 * 6] + %define k7 [rsp + 16 * 7] + %define krd [rsp + 16 * 8] + %define zero [rsp + 16 * 9] + + GET_FILTERS + + movsxd rax, DWORD PTR arg(1) ;pixels_per_line + movsxd rdx, DWORD PTR arg(3) ;out_pitch + movsxd rcx, DWORD PTR arg(4) ;output_height + +.loop: + movdqu xmm0, [rsi - 3] ;load src + + movdqa xmm1, xmm0 + movdqa xmm6, xmm0 + movdqa xmm7, xmm0 + movdqa xmm2, xmm0 + movdqa xmm5, xmm0 + movdqa xmm3, xmm0 + movdqa xmm4, xmm0 + + psrldq xmm1, 1 + psrldq xmm6, 6 + psrldq xmm7, 7 + psrldq xmm2, 2 + psrldq xmm5, 5 + psrldq xmm3, 3 + psrldq xmm4, 4 + + APPLY_FILTER_8 1, 0 + + movdqu xmm0, [rsi + 5] ;load src + + movdqa xmm1, xmm0 + movdqa xmm6, xmm0 + movdqa xmm7, xmm0 + movdqa xmm2, xmm0 + movdqa xmm5, xmm0 + movdqa xmm3, xmm0 + movdqa xmm4, xmm0 + + psrldq xmm1, 1 + psrldq xmm6, 6 + psrldq xmm7, 7 + psrldq xmm2, 2 + psrldq xmm5, 5 + psrldq xmm3, 3 + psrldq xmm4, 4 + + APPLY_FILTER_8 1, 8 + + lea rsi, [rsi + rax] + lea rdi, [rdi + rdx] + dec rcx + jnz .loop + + add rsp, 16 * 10 + pop rsp + + ; begin epilog + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret diff --git a/third_party/aom/aom_dsp/x86/aom_subpixel_8t_ssse3.asm b/third_party/aom/aom_dsp/x86/aom_subpixel_8t_ssse3.asm new file mode 100644 index 000000000..357f37401 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/aom_subpixel_8t_ssse3.asm @@ -0,0 +1,883 @@ +; +; Copyright (c) 2016, Alliance for Open Media. All rights reserved +; +; This source code is subject to the terms of the BSD 2 Clause License and +; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License +; was not distributed with this source code in the LICENSE file, you can +; obtain it at www.aomedia.org/license/software. If the Alliance for Open +; Media Patent License 1.0 was not distributed with this source code in the +; PATENTS file, you can obtain it at www.aomedia.org/license/patent. +; + +; + +%include "third_party/x86inc/x86inc.asm" + +SECTION_RODATA +pw_64: times 8 dw 64 +even_byte_mask: times 8 dw 0x00ff + +; %define USE_PMULHRSW +; NOTE: pmulhrsw has a latency of 5 cycles. Tests showed a performance loss +; when using this instruction. +; +; The add order below (based on ffav1) must be followed to prevent outranges. +; x = k0k1 + k4k5 +; y = k2k3 + k6k7 +; z = signed SAT(x + y) + +SECTION .text +%define LOCAL_VARS_SIZE 16*6 + +%macro SETUP_LOCAL_VARS 0 + ; TODO(slavarnway): using xmm registers for these on ARCH_X86_64 + + ; pmaddubsw has a higher latency on some platforms, this might be eased by + ; interleaving the instructions. + %define k0k1 [rsp + 16*0] + %define k2k3 [rsp + 16*1] + %define k4k5 [rsp + 16*2] + %define k6k7 [rsp + 16*3] + packsswb m4, m4 + ; TODO(slavarnway): multiple pshufb instructions had a higher latency on + ; some platforms. + pshuflw m0, m4, 0b ;k0_k1 + pshuflw m1, m4, 01010101b ;k2_k3 + pshuflw m2, m4, 10101010b ;k4_k5 + pshuflw m3, m4, 11111111b ;k6_k7 + punpcklqdq m0, m0 + punpcklqdq m1, m1 + punpcklqdq m2, m2 + punpcklqdq m3, m3 + mova k0k1, m0 + mova k2k3, m1 + mova k4k5, m2 + mova k6k7, m3 +%if ARCH_X86_64 + %define krd m12 + %define tmp0 [rsp + 16*4] + %define tmp1 [rsp + 16*5] + mova krd, [GLOBAL(pw_64)] +%else + %define krd [rsp + 16*4] +%if CONFIG_PIC=0 + mova m6, [GLOBAL(pw_64)] +%else + ; build constants without accessing global memory + pcmpeqb m6, m6 ;all ones + psrlw m6, 15 + psllw m6, 6 ;aka pw_64 +%endif + mova krd, m6 +%endif +%endm + +;------------------------------------------------------------------------------- +%if ARCH_X86_64 + %define LOCAL_VARS_SIZE_H4 0 +%else + %define LOCAL_VARS_SIZE_H4 16*4 +%endif + +%macro SUBPIX_HFILTER4 1 +cglobal filter_block1d4_%1, 6, 6, 11, LOCAL_VARS_SIZE_H4, \ + src, sstride, dst, dstride, height, filter + mova m4, [filterq] + packsswb m4, m4 +%if ARCH_X86_64 + %define k0k1k4k5 m8 + %define k2k3k6k7 m9 + %define krd m10 + mova krd, [GLOBAL(pw_64)] + pshuflw k0k1k4k5, m4, 0b ;k0_k1 + pshufhw k0k1k4k5, k0k1k4k5, 10101010b ;k0_k1_k4_k5 + pshuflw k2k3k6k7, m4, 01010101b ;k2_k3 + pshufhw k2k3k6k7, k2k3k6k7, 11111111b ;k2_k3_k6_k7 +%else + %define k0k1k4k5 [rsp + 16*0] + %define k2k3k6k7 [rsp + 16*1] + %define krd [rsp + 16*2] + pshuflw m6, m4, 0b ;k0_k1 + pshufhw m6, m6, 10101010b ;k0_k1_k4_k5 + pshuflw m7, m4, 01010101b ;k2_k3 + pshufhw m7, m7, 11111111b ;k2_k3_k6_k7 +%if CONFIG_PIC=0 + mova m1, [GLOBAL(pw_64)] +%else + ; build constants without accessing global memory + pcmpeqb m1, m1 ;all ones + psrlw m1, 15 + psllw m1, 6 ;aka pw_64 +%endif + mova k0k1k4k5, m6 + mova k2k3k6k7, m7 + mova krd, m1 +%endif + dec heightd + +.loop: + ;Do two rows at once + movu m4, [srcq - 3] + movu m5, [srcq + sstrideq - 3] + punpckhbw m1, m4, m4 + punpcklbw m4, m4 + punpckhbw m3, m5, m5 + punpcklbw m5, m5 + palignr m0, m1, m4, 1 + pmaddubsw m0, k0k1k4k5 + palignr m1, m4, 5 + pmaddubsw m1, k2k3k6k7 + palignr m2, m3, m5, 1 + pmaddubsw m2, k0k1k4k5 + palignr m3, m5, 5 + pmaddubsw m3, k2k3k6k7 + punpckhqdq m4, m0, m2 + punpcklqdq m0, m2 + punpckhqdq m5, m1, m3 + punpcklqdq m1, m3 + paddsw m0, m4 + paddsw m1, m5 +%ifidn %1, h8_avg + movd m4, [dstq] + movd m5, [dstq + dstrideq] +%endif + paddsw m0, m1 + paddsw m0, krd + psraw m0, 7 +%ifidn %1, h8_add_src + pxor m3, m3 + movu m4, [srcq] + movu m5, [srcq + sstrideq] + punpckldq m4, m5 ; Bytes 0,1,2,3 from row 0, then 0,1,2,3 from row 2 + punpcklbw m4, m3 + paddsw m0, m4 +%endif + packuswb m0, m0 + psrldq m1, m0, 4 + +%ifidn %1, h8_avg + pavgb m0, m4 + pavgb m1, m5 +%endif + movd [dstq], m0 + movd [dstq + dstrideq], m1 + + lea srcq, [srcq + sstrideq ] + prefetcht0 [srcq + 4 * sstrideq - 3] + lea srcq, [srcq + sstrideq ] + lea dstq, [dstq + 2 * dstrideq ] + prefetcht0 [srcq + 2 * sstrideq - 3] + + sub heightd, 2 + jg .loop + + ; Do last row if output_height is odd + jne .done + + movu m4, [srcq - 3] + punpckhbw m1, m4, m4 + punpcklbw m4, m4 + palignr m0, m1, m4, 1 + palignr m1, m4, 5 + pmaddubsw m0, k0k1k4k5 + pmaddubsw m1, k2k3k6k7 + psrldq m2, m0, 8 + psrldq m3, m1, 8 + paddsw m0, m2 + paddsw m1, m3 + paddsw m0, m1 + paddsw m0, krd + psraw m0, 7 +%ifidn %1, h8_add_src + pxor m3, m3 + movu m4, [srcq] + punpcklbw m4, m3 + paddsw m0, m4 +%endif + packuswb m0, m0 +%ifidn %1, h8_avg + movd m4, [dstq] + pavgb m0, m4 +%endif + movd [dstq], m0 +.done: + REP_RET +%endm + +;------------------------------------------------------------------------------- +%macro SUBPIX_HFILTER8 1 +cglobal filter_block1d8_%1, 6, 6, 14, LOCAL_VARS_SIZE, \ + src, sstride, dst, dstride, height, filter + mova m4, [filterq] + SETUP_LOCAL_VARS + dec heightd + +.loop: + ;Do two rows at once + movu m0, [srcq - 3] + movu m4, [srcq + sstrideq - 3] + punpckhbw m1, m0, m0 + punpcklbw m0, m0 + palignr m5, m1, m0, 13 + pmaddubsw m5, k6k7 + palignr m2, m1, m0, 5 + palignr m3, m1, m0, 9 + palignr m1, m0, 1 + pmaddubsw m1, k0k1 + punpckhbw m6, m4, m4 + punpcklbw m4, m4 + pmaddubsw m2, k2k3 + pmaddubsw m3, k4k5 + + palignr m7, m6, m4, 13 + palignr m0, m6, m4, 5 + pmaddubsw m7, k6k7 + paddsw m1, m3 + paddsw m2, m5 + paddsw m1, m2 +%ifidn %1, h8_avg + movh m2, [dstq] + movhps m2, [dstq + dstrideq] +%endif + palignr m5, m6, m4, 9 + palignr m6, m4, 1 + pmaddubsw m0, k2k3 + pmaddubsw m6, k0k1 + paddsw m1, krd + pmaddubsw m5, k4k5 + psraw m1, 7 + paddsw m0, m7 + paddsw m6, m5 + paddsw m6, m0 + paddsw m6, krd + psraw m6, 7 +%ifidn %1, h8_add_src + pxor m3, m3 + movu m4, [srcq] + movu m5, [srcq + sstrideq] + punpcklbw m4, m3 + punpcklbw m5, m3 + paddsw m1, m4 + paddsw m6, m5 +%endif + packuswb m1, m6 +%ifidn %1, h8_avg + pavgb m1, m2 +%endif + movh [dstq], m1 + movhps [dstq + dstrideq], m1 + + lea srcq, [srcq + sstrideq ] + prefetcht0 [srcq + 4 * sstrideq - 3] + lea srcq, [srcq + sstrideq ] + lea dstq, [dstq + 2 * dstrideq ] + prefetcht0 [srcq + 2 * sstrideq - 3] + sub heightd, 2 + jg .loop + + ; Do last row if output_height is odd + jne .done + + movu m0, [srcq - 3] + punpckhbw m3, m0, m0 + punpcklbw m0, m0 + palignr m1, m3, m0, 1 + palignr m2, m3, m0, 5 + palignr m4, m3, m0, 13 + palignr m3, m0, 9 + pmaddubsw m1, k0k1 + pmaddubsw m2, k2k3 + pmaddubsw m3, k4k5 + pmaddubsw m4, k6k7 + paddsw m1, m3 + paddsw m4, m2 + paddsw m1, m4 + paddsw m1, krd + psraw m1, 7 +%ifidn %1, h8_add_src + pxor m6, m6 + movu m5, [srcq] + punpcklbw m5, m6 + paddsw m1, m5 +%endif + packuswb m1, m1 +%ifidn %1, h8_avg + movh m0, [dstq] + pavgb m1, m0 +%endif + movh [dstq], m1 +.done: + REP_RET +%endm + +;------------------------------------------------------------------------------- +%macro SUBPIX_HFILTER16 1 +cglobal filter_block1d16_%1, 6, 6, 14, LOCAL_VARS_SIZE, \ + src, sstride, dst, dstride, height, filter + mova m4, [filterq] + SETUP_LOCAL_VARS + +.loop: + prefetcht0 [srcq + 2 * sstrideq -3] + + movu m0, [srcq - 3] + movu m4, [srcq - 2] + pmaddubsw m0, k0k1 + pmaddubsw m4, k0k1 + movu m1, [srcq - 1] + movu m5, [srcq + 0] + pmaddubsw m1, k2k3 + pmaddubsw m5, k2k3 + movu m2, [srcq + 1] + movu m6, [srcq + 2] + pmaddubsw m2, k4k5 + pmaddubsw m6, k4k5 + movu m3, [srcq + 3] + movu m7, [srcq + 4] + pmaddubsw m3, k6k7 + pmaddubsw m7, k6k7 + paddsw m0, m2 + paddsw m1, m3 + paddsw m0, m1 + paddsw m4, m6 + paddsw m5, m7 + paddsw m4, m5 + paddsw m0, krd + paddsw m4, krd + psraw m0, 7 + psraw m4, 7 +%ifidn %1, h8_add_src + movu m5, [srcq] + mova m7, m5 + pand m5, [even_byte_mask] + psrlw m7, 8 + paddsw m0, m5 + paddsw m4, m7 +%endif + packuswb m0, m0 + packuswb m4, m4 + punpcklbw m0, m4 +%ifidn %1, h8_avg + pavgb m0, [dstq] +%endif + lea srcq, [srcq + sstrideq] + mova [dstq], m0 + lea dstq, [dstq + dstrideq] + dec heightd + jnz .loop + REP_RET +%endm + +INIT_XMM ssse3 +SUBPIX_HFILTER16 h8 +SUBPIX_HFILTER16 h8_avg +SUBPIX_HFILTER8 h8 +SUBPIX_HFILTER8 h8_avg +SUBPIX_HFILTER4 h8 +SUBPIX_HFILTER4 h8_avg + +%if CONFIG_LOOP_RESTORATION +SUBPIX_HFILTER16 h8_add_src +SUBPIX_HFILTER8 h8_add_src +SUBPIX_HFILTER4 h8_add_src +%endif + +;------------------------------------------------------------------------------- + +; TODO(Linfeng): Detect cpu type and choose the code with better performance. +%define X86_SUBPIX_VFILTER_PREFER_SLOW_CELERON 1 + +%if ARCH_X86_64 && X86_SUBPIX_VFILTER_PREFER_SLOW_CELERON + %define NUM_GENERAL_REG_USED 9 +%else + %define NUM_GENERAL_REG_USED 6 +%endif + +%macro SUBPIX_VFILTER 2 +cglobal filter_block1d%2_%1, 6, NUM_GENERAL_REG_USED, 15, LOCAL_VARS_SIZE, \ + src, sstride, dst, dstride, height, filter + mova m4, [filterq] + SETUP_LOCAL_VARS + +%ifidn %2, 8 + %define movx movh +%else + %define movx movd +%endif + + dec heightd + +%if ARCH_X86 || X86_SUBPIX_VFILTER_PREFER_SLOW_CELERON + +%if ARCH_X86_64 + %define src1q r7 + %define sstride6q r8 + %define dst_stride dstrideq +%else + %define src1q filterq + %define sstride6q dstrideq + %define dst_stride dstridemp +%endif + mov src1q, srcq + add src1q, sstrideq + lea sstride6q, [sstrideq + sstrideq * 4] + add sstride6q, sstrideq ;pitch * 6 + +.loop: + ;Do two rows at once + movx m0, [srcq ] ;A + movx m1, [src1q ] ;B + punpcklbw m0, m1 ;A B + movx m2, [srcq + sstrideq * 2 ] ;C + pmaddubsw m0, k0k1 + mova m6, m2 + movx m3, [src1q + sstrideq * 2] ;D + punpcklbw m2, m3 ;C D + pmaddubsw m2, k2k3 + movx m4, [srcq + sstrideq * 4 ] ;E + mova m7, m4 + movx m5, [src1q + sstrideq * 4] ;F + punpcklbw m4, m5 ;E F + pmaddubsw m4, k4k5 + punpcklbw m1, m6 ;A B next iter + movx m6, [srcq + sstride6q ] ;G + punpcklbw m5, m6 ;E F next iter + punpcklbw m3, m7 ;C D next iter + pmaddubsw m5, k4k5 + movx m7, [src1q + sstride6q ] ;H + punpcklbw m6, m7 ;G H + pmaddubsw m6, k6k7 + pmaddubsw m3, k2k3 + pmaddubsw m1, k0k1 + paddsw m0, m4 + paddsw m2, m6 + movx m6, [srcq + sstrideq * 8 ] ;H next iter + punpcklbw m7, m6 + pmaddubsw m7, k6k7 + paddsw m0, m2 + paddsw m0, krd + psraw m0, 7 + paddsw m1, m5 +%ifidn %1, v8_add_src + pxor m6, m6 + movu m4, [srcq] + punpcklbw m4, m6 + paddsw m0, m4 +%endif + packuswb m0, m0 + + paddsw m3, m7 + paddsw m1, m3 + paddsw m1, krd + psraw m1, 7 +%ifidn %1, v8_add_src + movu m4, [src1q] + punpcklbw m4, m6 + paddsw m1, m4 +%endif + lea srcq, [srcq + sstrideq * 2 ] + lea src1q, [src1q + sstrideq * 2] + packuswb m1, m1 + +%ifidn %1, v8_avg + movx m2, [dstq] + pavgb m0, m2 +%endif + movx [dstq], m0 + add dstq, dst_stride +%ifidn %1, v8_avg + movx m3, [dstq] + pavgb m1, m3 +%endif + movx [dstq], m1 + add dstq, dst_stride + sub heightd, 2 + jg .loop + + ; Do last row if output_height is odd + jne .done + + movx m0, [srcq ] ;A + movx m1, [srcq + sstrideq ] ;B + movx m6, [srcq + sstride6q ] ;G + punpcklbw m0, m1 ;A B + movx m7, [src1q + sstride6q ] ;H + pmaddubsw m0, k0k1 + movx m2, [srcq + sstrideq * 2 ] ;C + punpcklbw m6, m7 ;G H + movx m3, [src1q + sstrideq * 2] ;D + pmaddubsw m6, k6k7 + movx m4, [srcq + sstrideq * 4 ] ;E + punpcklbw m2, m3 ;C D + movx m5, [src1q + sstrideq * 4] ;F + punpcklbw m4, m5 ;E F + pmaddubsw m2, k2k3 + pmaddubsw m4, k4k5 + paddsw m2, m6 + paddsw m0, m4 + paddsw m0, m2 + paddsw m0, krd + psraw m0, 7 +%ifidn %1, v8_add_src + pxor m6, m6 + movu m4, [srcq] + punpcklbw m4, m6 + paddsw m0, m4 +%endif + packuswb m0, m0 +%ifidn %1, v8_avg + movx m1, [dstq] + pavgb m0, m1 +%endif + movx [dstq], m0 + +%else + ; ARCH_X86_64 + + movx m0, [srcq ] ;A + movx m1, [srcq + sstrideq ] ;B + lea srcq, [srcq + sstrideq * 2 ] + movx m2, [srcq] ;C + movx m3, [srcq + sstrideq] ;D + lea srcq, [srcq + sstrideq * 2 ] + movx m4, [srcq] ;E + movx m5, [srcq + sstrideq] ;F + lea srcq, [srcq + sstrideq * 2 ] + movx m6, [srcq] ;G + punpcklbw m0, m1 ;A B + punpcklbw m1, m2 ;A B next iter + punpcklbw m2, m3 ;C D + punpcklbw m3, m4 ;C D next iter + punpcklbw m4, m5 ;E F + punpcklbw m5, m6 ;E F next iter + +.loop: + ;Do two rows at once + movx m7, [srcq + sstrideq] ;H + lea srcq, [srcq + sstrideq * 2 ] + movx m14, [srcq] ;H next iter + punpcklbw m6, m7 ;G H + punpcklbw m7, m14 ;G H next iter + pmaddubsw m8, m0, k0k1 + pmaddubsw m9, m1, k0k1 + mova m0, m2 + mova m1, m3 + pmaddubsw m10, m2, k2k3 + pmaddubsw m11, m3, k2k3 + mova m2, m4 + mova m3, m5 + pmaddubsw m4, k4k5 + pmaddubsw m5, k4k5 + paddsw m8, m4 + paddsw m9, m5 + mova m4, m6 + mova m5, m7 + pmaddubsw m6, k6k7 + pmaddubsw m7, k6k7 + paddsw m10, m6 + paddsw m11, m7 + paddsw m8, m10 + paddsw m9, m11 + mova m6, m14 + paddsw m8, krd + paddsw m9, krd + psraw m8, 7 + psraw m9, 7 +%ifidn %2, 4 + packuswb m8, m8 + packuswb m9, m9 +%else + packuswb m8, m9 +%endif + +%ifidn %1, v8_avg + movx m7, [dstq] +%ifidn %2, 4 + movx m10, [dstq + dstrideq] + pavgb m9, m10 +%else + movhpd m7, [dstq + dstrideq] +%endif + pavgb m8, m7 +%endif + movx [dstq], m8 +%ifidn %2, 4 + movx [dstq + dstrideq], m9 +%else + movhpd [dstq + dstrideq], m8 +%endif + + lea dstq, [dstq + dstrideq * 2 ] + sub heightd, 2 + jg .loop + + ; Do last row if output_height is odd + jne .done + + movx m7, [srcq + sstrideq] ;H + punpcklbw m6, m7 ;G H + pmaddubsw m0, k0k1 + pmaddubsw m2, k2k3 + pmaddubsw m4, k4k5 + pmaddubsw m6, k6k7 + paddsw m0, m4 + paddsw m2, m6 + paddsw m0, m2 + paddsw m0, krd + psraw m0, 7 + packuswb m0, m0 +%ifidn %1, v8_avg + movx m1, [dstq] + pavgb m0, m1 +%endif + movx [dstq], m0 + +%endif ; ARCH_X86_64 + +.done: + REP_RET + +%endm + +;------------------------------------------------------------------------------- +%macro SUBPIX_VFILTER16 1 +cglobal filter_block1d16_%1, 6, NUM_GENERAL_REG_USED, 16, LOCAL_VARS_SIZE, \ + src, sstride, dst, dstride, height, filter + mova m4, [filterq] + SETUP_LOCAL_VARS + +%if ARCH_X86 || X86_SUBPIX_VFILTER_PREFER_SLOW_CELERON + +%if ARCH_X86_64 + %define src1q r7 + %define sstride6q r8 + %define dst_stride dstrideq +%else + %define src1q filterq + %define sstride6q dstrideq + %define dst_stride dstridemp +%endif + lea src1q, [srcq + sstrideq] + lea sstride6q, [sstrideq + sstrideq * 4] + add sstride6q, sstrideq ;pitch * 6 + +.loop: + movh m0, [srcq ] ;A + movh m1, [src1q ] ;B + movh m2, [srcq + sstrideq * 2 ] ;C + movh m3, [src1q + sstrideq * 2] ;D + movh m4, [srcq + sstrideq * 4 ] ;E + movh m5, [src1q + sstrideq * 4] ;F + + punpcklbw m0, m1 ;A B + movh m6, [srcq + sstride6q] ;G + punpcklbw m2, m3 ;C D + movh m7, [src1q + sstride6q] ;H + punpcklbw m4, m5 ;E F + pmaddubsw m0, k0k1 + movh m3, [srcq + 8] ;A + pmaddubsw m2, k2k3 + punpcklbw m6, m7 ;G H + movh m5, [srcq + sstrideq + 8] ;B + pmaddubsw m4, k4k5 + punpcklbw m3, m5 ;A B + movh m7, [srcq + sstrideq * 2 + 8] ;C + pmaddubsw m6, k6k7 + movh m5, [src1q + sstrideq * 2 + 8] ;D + punpcklbw m7, m5 ;C D + paddsw m2, m6 + pmaddubsw m3, k0k1 + movh m1, [srcq + sstrideq * 4 + 8] ;E + paddsw m0, m4 + pmaddubsw m7, k2k3 + movh m6, [src1q + sstrideq * 4 + 8] ;F + punpcklbw m1, m6 ;E F + paddsw m0, m2 + paddsw m0, krd + movh m2, [srcq + sstride6q + 8] ;G + pmaddubsw m1, k4k5 + movh m5, [src1q + sstride6q + 8] ;H + psraw m0, 7 + punpcklbw m2, m5 ;G H + pmaddubsw m2, k6k7 + paddsw m7, m2 + paddsw m3, m1 + paddsw m3, m7 + paddsw m3, krd + psraw m3, 7 +%ifidn %1, v8_add_src + pxor m6, m6 + movu m4, [src1q + 2 * sstrideq] ; Fetch from 3 rows down + mova m5, m4 + punpcklbw m4, m6 + punpckhbw m5, m6 + paddsw m0, m4 + paddsw m3, m5 +%endif + packuswb m0, m3 + + add srcq, sstrideq + add src1q, sstrideq +%ifidn %1, v8_avg + pavgb m0, [dstq] +%endif + mova [dstq], m0 + add dstq, dst_stride + dec heightd + jnz .loop + REP_RET + +%else + ; ARCH_X86_64 + dec heightd + + movu m1, [srcq ] ;A + movu m3, [srcq + sstrideq ] ;B + lea srcq, [srcq + sstrideq * 2] + punpcklbw m0, m1, m3 ;A B + punpckhbw m1, m3 ;A B + movu m5, [srcq] ;C + punpcklbw m2, m3, m5 ;A B next iter + punpckhbw m3, m5 ;A B next iter + mova tmp0, m2 ;store to stack + mova tmp1, m3 ;store to stack + movu m7, [srcq + sstrideq] ;D + lea srcq, [srcq + sstrideq * 2] + punpcklbw m4, m5, m7 ;C D + punpckhbw m5, m7 ;C D + movu m9, [srcq] ;E + punpcklbw m6, m7, m9 ;C D next iter + punpckhbw m7, m9 ;C D next iter + movu m11, [srcq + sstrideq] ;F + lea srcq, [srcq + sstrideq * 2] + punpcklbw m8, m9, m11 ;E F + punpckhbw m9, m11 ;E F + movu m2, [srcq] ;G + punpcklbw m10, m11, m2 ;E F next iter + punpckhbw m11, m2 ;E F next iter + +.loop: + ;Do two rows at once + pmaddubsw m13, m0, k0k1 + mova m0, m4 + pmaddubsw m14, m8, k4k5 + pmaddubsw m15, m4, k2k3 + mova m4, m8 + paddsw m13, m14 + movu m3, [srcq + sstrideq] ;H + lea srcq, [srcq + sstrideq * 2] + punpcklbw m14, m2, m3 ;G H + mova m8, m14 + pmaddubsw m14, k6k7 + paddsw m15, m14 + paddsw m13, m15 + paddsw m13, krd + psraw m13, 7 + + pmaddubsw m14, m1, k0k1 + pmaddubsw m1, m9, k4k5 + pmaddubsw m15, m5, k2k3 + paddsw m14, m1 + mova m1, m5 + mova m5, m9 + punpckhbw m2, m3 ;G H + mova m9, m2 + pmaddubsw m2, k6k7 + paddsw m15, m2 + paddsw m14, m15 + paddsw m14, krd + psraw m14, 7 + packuswb m13, m14 +%ifidn %1, v8_avg + pavgb m13, [dstq] +%endif + mova [dstq], m13 + + ; next iter + pmaddubsw m15, tmp0, k0k1 + pmaddubsw m14, m10, k4k5 + pmaddubsw m13, m6, k2k3 + paddsw m15, m14 + mova tmp0, m6 + mova m6, m10 + movu m2, [srcq] ;G next iter + punpcklbw m14, m3, m2 ;G H next iter + mova m10, m14 + pmaddubsw m14, k6k7 + paddsw m13, m14 + paddsw m15, m13 + paddsw m15, krd + psraw m15, 7 + + pmaddubsw m14, tmp1, k0k1 + mova tmp1, m7 + pmaddubsw m13, m7, k2k3 + mova m7, m11 + pmaddubsw m11, k4k5 + paddsw m14, m11 + punpckhbw m3, m2 ;G H next iter + mova m11, m3 + pmaddubsw m3, k6k7 + paddsw m13, m3 + paddsw m14, m13 + paddsw m14, krd + psraw m14, 7 + packuswb m15, m14 +%ifidn %1, v8_avg + pavgb m15, [dstq + dstrideq] +%endif + mova [dstq + dstrideq], m15 + lea dstq, [dstq + dstrideq * 2] + sub heightd, 2 + jg .loop + + ; Do last row if output_height is odd + jne .done + + movu m3, [srcq + sstrideq] ;H + punpcklbw m6, m2, m3 ;G H + punpckhbw m2, m3 ;G H + pmaddubsw m0, k0k1 + pmaddubsw m1, k0k1 + pmaddubsw m4, k2k3 + pmaddubsw m5, k2k3 + pmaddubsw m8, k4k5 + pmaddubsw m9, k4k5 + pmaddubsw m6, k6k7 + pmaddubsw m2, k6k7 + paddsw m0, m8 + paddsw m1, m9 + paddsw m4, m6 + paddsw m5, m2 + paddsw m0, m4 + paddsw m1, m5 + paddsw m0, krd + paddsw m1, krd + psraw m0, 7 + psraw m1, 7 + packuswb m0, m1 +%ifidn %1, v8_avg + pavgb m0, [dstq] +%endif + mova [dstq], m0 + +.done: + REP_RET + +%endif ; ARCH_X86_64 + +%endm + +INIT_XMM ssse3 +SUBPIX_VFILTER16 v8 +SUBPIX_VFILTER16 v8_avg +SUBPIX_VFILTER v8, 8 +SUBPIX_VFILTER v8_avg, 8 +SUBPIX_VFILTER v8, 4 +SUBPIX_VFILTER v8_avg, 4 + +%if (ARCH_X86 || X86_SUBPIX_VFILTER_PREFER_SLOW_CELERON) && \ + CONFIG_LOOP_RESTORATION +SUBPIX_VFILTER16 v8_add_src +SUBPIX_VFILTER v8_add_src, 8 +SUBPIX_VFILTER v8_add_src, 4 +%endif diff --git a/third_party/aom/aom_dsp/x86/aom_subpixel_bilinear_sse2.asm b/third_party/aom/aom_dsp/x86/aom_subpixel_bilinear_sse2.asm new file mode 100644 index 000000000..8f025a8be --- /dev/null +++ b/third_party/aom/aom_dsp/x86/aom_subpixel_bilinear_sse2.asm @@ -0,0 +1,451 @@ +; +; Copyright (c) 2016, Alliance for Open Media. All rights reserved +; +; This source code is subject to the terms of the BSD 2 Clause License and +; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License +; was not distributed with this source code in the LICENSE file, you can +; obtain it at www.aomedia.org/license/software. If the Alliance for Open +; Media Patent License 1.0 was not distributed with this source code in the +; PATENTS file, you can obtain it at www.aomedia.org/license/patent. +; + +; + +%include "aom_ports/x86_abi_support.asm" + +%macro GET_PARAM_4 0 + mov rdx, arg(5) ;filter ptr + mov rsi, arg(0) ;src_ptr + mov rdi, arg(2) ;output_ptr + mov rcx, 0x0400040 + + movdqa xmm3, [rdx] ;load filters + pshuflw xmm4, xmm3, 11111111b ;k3 + psrldq xmm3, 8 + pshuflw xmm3, xmm3, 0b ;k4 + punpcklqdq xmm4, xmm3 ;k3k4 + + movq xmm3, rcx ;rounding + pshufd xmm3, xmm3, 0 + + pxor xmm2, xmm2 + + movsxd rax, DWORD PTR arg(1) ;pixels_per_line + movsxd rdx, DWORD PTR arg(3) ;out_pitch + movsxd rcx, DWORD PTR arg(4) ;output_height +%endm + +%macro APPLY_FILTER_4 1 + + punpckldq xmm0, xmm1 ;two row in one register + punpcklbw xmm0, xmm2 ;unpack to word + pmullw xmm0, xmm4 ;multiply the filter factors + + movdqa xmm1, xmm0 + psrldq xmm1, 8 + paddsw xmm0, xmm1 + + paddsw xmm0, xmm3 ;rounding + psraw xmm0, 7 ;shift + packuswb xmm0, xmm0 ;pack to byte + +%if %1 + movd xmm1, [rdi] + pavgb xmm0, xmm1 +%endif + + movd [rdi], xmm0 + lea rsi, [rsi + rax] + lea rdi, [rdi + rdx] + dec rcx +%endm + +%macro GET_PARAM 0 + mov rdx, arg(5) ;filter ptr + mov rsi, arg(0) ;src_ptr + mov rdi, arg(2) ;output_ptr + mov rcx, 0x0400040 + + movdqa xmm7, [rdx] ;load filters + + pshuflw xmm6, xmm7, 11111111b ;k3 + pshufhw xmm7, xmm7, 0b ;k4 + punpcklwd xmm6, xmm6 + punpckhwd xmm7, xmm7 + + movq xmm4, rcx ;rounding + pshufd xmm4, xmm4, 0 + + pxor xmm5, xmm5 + + movsxd rax, DWORD PTR arg(1) ;pixels_per_line + movsxd rdx, DWORD PTR arg(3) ;out_pitch + movsxd rcx, DWORD PTR arg(4) ;output_height +%endm + +%macro APPLY_FILTER_8 1 + punpcklbw xmm0, xmm5 + punpcklbw xmm1, xmm5 + + pmullw xmm0, xmm6 + pmullw xmm1, xmm7 + paddsw xmm0, xmm1 + paddsw xmm0, xmm4 ;rounding + psraw xmm0, 7 ;shift + packuswb xmm0, xmm0 ;pack back to byte +%if %1 + movq xmm1, [rdi] + pavgb xmm0, xmm1 +%endif + movq [rdi], xmm0 ;store the result + + lea rsi, [rsi + rax] + lea rdi, [rdi + rdx] + dec rcx +%endm + +%macro APPLY_FILTER_16 1 + punpcklbw xmm0, xmm5 + punpcklbw xmm1, xmm5 + punpckhbw xmm2, xmm5 + punpckhbw xmm3, xmm5 + + pmullw xmm0, xmm6 + pmullw xmm1, xmm7 + pmullw xmm2, xmm6 + pmullw xmm3, xmm7 + + paddsw xmm0, xmm1 + paddsw xmm2, xmm3 + + paddsw xmm0, xmm4 ;rounding + paddsw xmm2, xmm4 + psraw xmm0, 7 ;shift + psraw xmm2, 7 + packuswb xmm0, xmm2 ;pack back to byte +%if %1 + movdqu xmm1, [rdi] + pavgb xmm0, xmm1 +%endif + movdqu [rdi], xmm0 ;store the result + + lea rsi, [rsi + rax] + lea rdi, [rdi + rdx] + dec rcx +%endm + +global sym(aom_filter_block1d4_v2_sse2) PRIVATE +sym(aom_filter_block1d4_v2_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 6 + push rsi + push rdi + ; end prolog + + GET_PARAM_4 +.loop: + movd xmm0, [rsi] ;load src + movd xmm1, [rsi + rax] + + APPLY_FILTER_4 0 + jnz .loop + + ; begin epilog + pop rdi + pop rsi + UNSHADOW_ARGS + pop rbp + ret + +global sym(aom_filter_block1d8_v2_sse2) PRIVATE +sym(aom_filter_block1d8_v2_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 6 + SAVE_XMM 7 + push rsi + push rdi + ; end prolog + + GET_PARAM +.loop: + movq xmm0, [rsi] ;0 + movq xmm1, [rsi + rax] ;1 + + APPLY_FILTER_8 0 + jnz .loop + + ; begin epilog + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret + +global sym(aom_filter_block1d16_v2_sse2) PRIVATE +sym(aom_filter_block1d16_v2_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 6 + SAVE_XMM 7 + push rsi + push rdi + ; end prolog + + GET_PARAM +.loop: + movdqu xmm0, [rsi] ;0 + movdqu xmm1, [rsi + rax] ;1 + movdqa xmm2, xmm0 + movdqa xmm3, xmm1 + + APPLY_FILTER_16 0 + jnz .loop + + ; begin epilog + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret + +global sym(aom_filter_block1d4_v2_avg_sse2) PRIVATE +sym(aom_filter_block1d4_v2_avg_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 6 + push rsi + push rdi + ; end prolog + + GET_PARAM_4 +.loop: + movd xmm0, [rsi] ;load src + movd xmm1, [rsi + rax] + + APPLY_FILTER_4 1 + jnz .loop + + ; begin epilog + pop rdi + pop rsi + UNSHADOW_ARGS + pop rbp + ret + +global sym(aom_filter_block1d8_v2_avg_sse2) PRIVATE +sym(aom_filter_block1d8_v2_avg_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 6 + SAVE_XMM 7 + push rsi + push rdi + ; end prolog + + GET_PARAM +.loop: + movq xmm0, [rsi] ;0 + movq xmm1, [rsi + rax] ;1 + + APPLY_FILTER_8 1 + jnz .loop + + ; begin epilog + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret + +global sym(aom_filter_block1d16_v2_avg_sse2) PRIVATE +sym(aom_filter_block1d16_v2_avg_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 6 + SAVE_XMM 7 + push rsi + push rdi + ; end prolog + + GET_PARAM +.loop: + movdqu xmm0, [rsi] ;0 + movdqu xmm1, [rsi + rax] ;1 + movdqa xmm2, xmm0 + movdqa xmm3, xmm1 + + APPLY_FILTER_16 1 + jnz .loop + + ; begin epilog + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret + +global sym(aom_filter_block1d4_h2_sse2) PRIVATE +sym(aom_filter_block1d4_h2_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 6 + push rsi + push rdi + ; end prolog + + GET_PARAM_4 +.loop: + movdqu xmm0, [rsi] ;load src + movdqa xmm1, xmm0 + psrldq xmm1, 1 + + APPLY_FILTER_4 0 + jnz .loop + + ; begin epilog + pop rdi + pop rsi + UNSHADOW_ARGS + pop rbp + ret + +global sym(aom_filter_block1d8_h2_sse2) PRIVATE +sym(aom_filter_block1d8_h2_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 6 + SAVE_XMM 7 + push rsi + push rdi + ; end prolog + + GET_PARAM +.loop: + movdqu xmm0, [rsi] ;load src + movdqa xmm1, xmm0 + psrldq xmm1, 1 + + APPLY_FILTER_8 0 + jnz .loop + + ; begin epilog + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret + +global sym(aom_filter_block1d16_h2_sse2) PRIVATE +sym(aom_filter_block1d16_h2_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 6 + SAVE_XMM 7 + push rsi + push rdi + ; end prolog + + GET_PARAM +.loop: + movdqu xmm0, [rsi] ;load src + movdqu xmm1, [rsi + 1] + movdqa xmm2, xmm0 + movdqa xmm3, xmm1 + + APPLY_FILTER_16 0 + jnz .loop + + ; begin epilog + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret + +global sym(aom_filter_block1d4_h2_avg_sse2) PRIVATE +sym(aom_filter_block1d4_h2_avg_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 6 + push rsi + push rdi + ; end prolog + + GET_PARAM_4 +.loop: + movdqu xmm0, [rsi] ;load src + movdqa xmm1, xmm0 + psrldq xmm1, 1 + + APPLY_FILTER_4 1 + jnz .loop + + ; begin epilog + pop rdi + pop rsi + UNSHADOW_ARGS + pop rbp + ret + +global sym(aom_filter_block1d8_h2_avg_sse2) PRIVATE +sym(aom_filter_block1d8_h2_avg_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 6 + SAVE_XMM 7 + push rsi + push rdi + ; end prolog + + GET_PARAM +.loop: + movdqu xmm0, [rsi] ;load src + movdqa xmm1, xmm0 + psrldq xmm1, 1 + + APPLY_FILTER_8 1 + jnz .loop + + ; begin epilog + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret + +global sym(aom_filter_block1d16_h2_avg_sse2) PRIVATE +sym(aom_filter_block1d16_h2_avg_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 6 + SAVE_XMM 7 + push rsi + push rdi + ; end prolog + + GET_PARAM +.loop: + movdqu xmm0, [rsi] ;load src + movdqu xmm1, [rsi + 1] + movdqa xmm2, xmm0 + movdqa xmm3, xmm1 + + APPLY_FILTER_16 1 + jnz .loop + + ; begin epilog + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret diff --git a/third_party/aom/aom_dsp/x86/aom_subpixel_bilinear_ssse3.asm b/third_party/aom/aom_dsp/x86/aom_subpixel_bilinear_ssse3.asm new file mode 100644 index 000000000..b9b2da0be --- /dev/null +++ b/third_party/aom/aom_dsp/x86/aom_subpixel_bilinear_ssse3.asm @@ -0,0 +1,421 @@ +; +; Copyright (c) 2016, Alliance for Open Media. All rights reserved +; +; This source code is subject to the terms of the BSD 2 Clause License and +; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License +; was not distributed with this source code in the LICENSE file, you can +; obtain it at www.aomedia.org/license/software. If the Alliance for Open +; Media Patent License 1.0 was not distributed with this source code in the +; PATENTS file, you can obtain it at www.aomedia.org/license/patent. +; + +; + +%include "aom_ports/x86_abi_support.asm" + +%macro GET_PARAM_4 0 + mov rdx, arg(5) ;filter ptr + mov rsi, arg(0) ;src_ptr + mov rdi, arg(2) ;output_ptr + mov ecx, 0x01000100 + + movdqa xmm3, [rdx] ;load filters + psrldq xmm3, 6 + packsswb xmm3, xmm3 + pshuflw xmm3, xmm3, 0b ;k3_k4 + + movd xmm2, ecx ;rounding_shift + pshufd xmm2, xmm2, 0 + + movsxd rax, DWORD PTR arg(1) ;pixels_per_line + movsxd rdx, DWORD PTR arg(3) ;out_pitch + movsxd rcx, DWORD PTR arg(4) ;output_height +%endm + +%macro APPLY_FILTER_4 1 + punpcklbw xmm0, xmm1 + pmaddubsw xmm0, xmm3 + + pmulhrsw xmm0, xmm2 ;rounding(+64)+shift(>>7) + packuswb xmm0, xmm0 ;pack to byte + +%if %1 + movd xmm1, [rdi] + pavgb xmm0, xmm1 +%endif + movd [rdi], xmm0 + lea rsi, [rsi + rax] + lea rdi, [rdi + rdx] + dec rcx +%endm + +%macro GET_PARAM 0 + mov rdx, arg(5) ;filter ptr + mov rsi, arg(0) ;src_ptr + mov rdi, arg(2) ;output_ptr + mov ecx, 0x01000100 + + movdqa xmm7, [rdx] ;load filters + psrldq xmm7, 6 + packsswb xmm7, xmm7 + pshuflw xmm7, xmm7, 0b ;k3_k4 + punpcklwd xmm7, xmm7 + + movd xmm6, ecx ;rounding_shift + pshufd xmm6, xmm6, 0 + + movsxd rax, DWORD PTR arg(1) ;pixels_per_line + movsxd rdx, DWORD PTR arg(3) ;out_pitch + movsxd rcx, DWORD PTR arg(4) ;output_height +%endm + +%macro APPLY_FILTER_8 1 + punpcklbw xmm0, xmm1 + pmaddubsw xmm0, xmm7 + + pmulhrsw xmm0, xmm6 ;rounding(+64)+shift(>>7) + packuswb xmm0, xmm0 ;pack back to byte + +%if %1 + movq xmm1, [rdi] + pavgb xmm0, xmm1 +%endif + movq [rdi], xmm0 ;store the result + + lea rsi, [rsi + rax] + lea rdi, [rdi + rdx] + dec rcx +%endm + +%macro APPLY_FILTER_16 1 + punpcklbw xmm0, xmm1 + punpckhbw xmm2, xmm1 + pmaddubsw xmm0, xmm7 + pmaddubsw xmm2, xmm7 + + pmulhrsw xmm0, xmm6 ;rounding(+64)+shift(>>7) + pmulhrsw xmm2, xmm6 + packuswb xmm0, xmm2 ;pack back to byte + +%if %1 + movdqu xmm1, [rdi] + pavgb xmm0, xmm1 +%endif + movdqu [rdi], xmm0 ;store the result + + lea rsi, [rsi + rax] + lea rdi, [rdi + rdx] + dec rcx +%endm + +global sym(aom_filter_block1d4_v2_ssse3) PRIVATE +sym(aom_filter_block1d4_v2_ssse3): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 6 + push rsi + push rdi + ; end prolog + + GET_PARAM_4 +.loop: + movd xmm0, [rsi] ;load src + movd xmm1, [rsi + rax] + + APPLY_FILTER_4 0 + jnz .loop + + ; begin epilog + pop rdi + pop rsi + UNSHADOW_ARGS + pop rbp + ret + +global sym(aom_filter_block1d8_v2_ssse3) PRIVATE +sym(aom_filter_block1d8_v2_ssse3): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 6 + SAVE_XMM 7 + push rsi + push rdi + ; end prolog + + GET_PARAM +.loop: + movq xmm0, [rsi] ;0 + movq xmm1, [rsi + rax] ;1 + + APPLY_FILTER_8 0 + jnz .loop + + ; begin epilog + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret + +global sym(aom_filter_block1d16_v2_ssse3) PRIVATE +sym(aom_filter_block1d16_v2_ssse3): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 6 + SAVE_XMM 7 + push rsi + push rdi + ; end prolog + + GET_PARAM +.loop: + movdqu xmm0, [rsi] ;0 + movdqu xmm1, [rsi + rax] ;1 + movdqa xmm2, xmm0 + + APPLY_FILTER_16 0 + jnz .loop + + ; begin epilog + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret + +global sym(aom_filter_block1d4_v2_avg_ssse3) PRIVATE +sym(aom_filter_block1d4_v2_avg_ssse3): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 6 + push rsi + push rdi + ; end prolog + + GET_PARAM_4 +.loop: + movd xmm0, [rsi] ;load src + movd xmm1, [rsi + rax] + + APPLY_FILTER_4 1 + jnz .loop + + ; begin epilog + pop rdi + pop rsi + UNSHADOW_ARGS + pop rbp + ret + +global sym(aom_filter_block1d8_v2_avg_ssse3) PRIVATE +sym(aom_filter_block1d8_v2_avg_ssse3): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 6 + SAVE_XMM 7 + push rsi + push rdi + ; end prolog + + GET_PARAM +.loop: + movq xmm0, [rsi] ;0 + movq xmm1, [rsi + rax] ;1 + + APPLY_FILTER_8 1 + jnz .loop + + ; begin epilog + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret + +global sym(aom_filter_block1d16_v2_avg_ssse3) PRIVATE +sym(aom_filter_block1d16_v2_avg_ssse3): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 6 + SAVE_XMM 7 + push rsi + push rdi + ; end prolog + + GET_PARAM +.loop: + movdqu xmm0, [rsi] ;0 + movdqu xmm1, [rsi + rax] ;1 + movdqa xmm2, xmm0 + + APPLY_FILTER_16 1 + jnz .loop + + ; begin epilog + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret + +global sym(aom_filter_block1d4_h2_ssse3) PRIVATE +sym(aom_filter_block1d4_h2_ssse3): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 6 + push rsi + push rdi + ; end prolog + + GET_PARAM_4 +.loop: + movdqu xmm0, [rsi] ;load src + movdqa xmm1, xmm0 + psrldq xmm1, 1 + + APPLY_FILTER_4 0 + jnz .loop + + ; begin epilog + pop rdi + pop rsi + UNSHADOW_ARGS + pop rbp + ret + +global sym(aom_filter_block1d8_h2_ssse3) PRIVATE +sym(aom_filter_block1d8_h2_ssse3): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 6 + SAVE_XMM 7 + push rsi + push rdi + ; end prolog + + GET_PARAM +.loop: + movdqu xmm0, [rsi] ;load src + movdqa xmm1, xmm0 + psrldq xmm1, 1 + + APPLY_FILTER_8 0 + jnz .loop + + ; begin epilog + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret + +global sym(aom_filter_block1d16_h2_ssse3) PRIVATE +sym(aom_filter_block1d16_h2_ssse3): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 6 + SAVE_XMM 7 + push rsi + push rdi + ; end prolog + + GET_PARAM +.loop: + movdqu xmm0, [rsi] ;load src + movdqu xmm1, [rsi + 1] + movdqa xmm2, xmm0 + + APPLY_FILTER_16 0 + jnz .loop + + ; begin epilog + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret + +global sym(aom_filter_block1d4_h2_avg_ssse3) PRIVATE +sym(aom_filter_block1d4_h2_avg_ssse3): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 6 + push rsi + push rdi + ; end prolog + + GET_PARAM_4 +.loop: + movdqu xmm0, [rsi] ;load src + movdqa xmm1, xmm0 + psrldq xmm1, 1 + + APPLY_FILTER_4 1 + jnz .loop + + ; begin epilog + pop rdi + pop rsi + UNSHADOW_ARGS + pop rbp + ret + +global sym(aom_filter_block1d8_h2_avg_ssse3) PRIVATE +sym(aom_filter_block1d8_h2_avg_ssse3): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 6 + SAVE_XMM 7 + push rsi + push rdi + ; end prolog + + GET_PARAM +.loop: + movdqu xmm0, [rsi] ;load src + movdqa xmm1, xmm0 + psrldq xmm1, 1 + + APPLY_FILTER_8 1 + jnz .loop + + ; begin epilog + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret + +global sym(aom_filter_block1d16_h2_avg_ssse3) PRIVATE +sym(aom_filter_block1d16_h2_avg_ssse3): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 6 + SAVE_XMM 7 + push rsi + push rdi + ; end prolog + + GET_PARAM +.loop: + movdqu xmm0, [rsi] ;load src + movdqu xmm1, [rsi + 1] + movdqa xmm2, xmm0 + + APPLY_FILTER_16 1 + jnz .loop + + ; begin epilog + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret diff --git a/third_party/aom/aom_dsp/x86/avg_intrin_sse2.c b/third_party/aom/aom_dsp/x86/avg_intrin_sse2.c new file mode 100644 index 000000000..bcdc20f63 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/avg_intrin_sse2.c @@ -0,0 +1,426 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include <emmintrin.h> + +#include "aom_dsp/x86/synonyms.h" + +#include "./aom_dsp_rtcd.h" +#include "aom_ports/mem.h" + +void aom_minmax_8x8_sse2(const uint8_t *s, int p, const uint8_t *d, int dp, + int *min, int *max) { + __m128i u0, s0, d0, diff, maxabsdiff, minabsdiff, negdiff, absdiff0, absdiff; + u0 = _mm_setzero_si128(); + // Row 0 + s0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s)), u0); + d0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(d)), u0); + diff = _mm_subs_epi16(s0, d0); + negdiff = _mm_subs_epi16(u0, diff); + absdiff0 = _mm_max_epi16(diff, negdiff); + // Row 1 + s0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + p)), u0); + d0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(d + dp)), u0); + diff = _mm_subs_epi16(s0, d0); + negdiff = _mm_subs_epi16(u0, diff); + absdiff = _mm_max_epi16(diff, negdiff); + maxabsdiff = _mm_max_epi16(absdiff0, absdiff); + minabsdiff = _mm_min_epi16(absdiff0, absdiff); + // Row 2 + s0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 2 * p)), u0); + d0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(d + 2 * dp)), u0); + diff = _mm_subs_epi16(s0, d0); + negdiff = _mm_subs_epi16(u0, diff); + absdiff = _mm_max_epi16(diff, negdiff); + maxabsdiff = _mm_max_epi16(maxabsdiff, absdiff); + minabsdiff = _mm_min_epi16(minabsdiff, absdiff); + // Row 3 + s0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 3 * p)), u0); + d0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(d + 3 * dp)), u0); + diff = _mm_subs_epi16(s0, d0); + negdiff = _mm_subs_epi16(u0, diff); + absdiff = _mm_max_epi16(diff, negdiff); + maxabsdiff = _mm_max_epi16(maxabsdiff, absdiff); + minabsdiff = _mm_min_epi16(minabsdiff, absdiff); + // Row 4 + s0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 4 * p)), u0); + d0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(d + 4 * dp)), u0); + diff = _mm_subs_epi16(s0, d0); + negdiff = _mm_subs_epi16(u0, diff); + absdiff = _mm_max_epi16(diff, negdiff); + maxabsdiff = _mm_max_epi16(maxabsdiff, absdiff); + minabsdiff = _mm_min_epi16(minabsdiff, absdiff); + // Row 5 + s0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 5 * p)), u0); + d0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(d + 5 * dp)), u0); + diff = _mm_subs_epi16(s0, d0); + negdiff = _mm_subs_epi16(u0, diff); + absdiff = _mm_max_epi16(diff, negdiff); + maxabsdiff = _mm_max_epi16(maxabsdiff, absdiff); + minabsdiff = _mm_min_epi16(minabsdiff, absdiff); + // Row 6 + s0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 6 * p)), u0); + d0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(d + 6 * dp)), u0); + diff = _mm_subs_epi16(s0, d0); + negdiff = _mm_subs_epi16(u0, diff); + absdiff = _mm_max_epi16(diff, negdiff); + maxabsdiff = _mm_max_epi16(maxabsdiff, absdiff); + minabsdiff = _mm_min_epi16(minabsdiff, absdiff); + // Row 7 + s0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 7 * p)), u0); + d0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(d + 7 * dp)), u0); + diff = _mm_subs_epi16(s0, d0); + negdiff = _mm_subs_epi16(u0, diff); + absdiff = _mm_max_epi16(diff, negdiff); + maxabsdiff = _mm_max_epi16(maxabsdiff, absdiff); + minabsdiff = _mm_min_epi16(minabsdiff, absdiff); + + maxabsdiff = _mm_max_epi16(maxabsdiff, _mm_srli_si128(maxabsdiff, 8)); + maxabsdiff = _mm_max_epi16(maxabsdiff, _mm_srli_epi64(maxabsdiff, 32)); + maxabsdiff = _mm_max_epi16(maxabsdiff, _mm_srli_epi64(maxabsdiff, 16)); + *max = _mm_extract_epi16(maxabsdiff, 0); + + minabsdiff = _mm_min_epi16(minabsdiff, _mm_srli_si128(minabsdiff, 8)); + minabsdiff = _mm_min_epi16(minabsdiff, _mm_srli_epi64(minabsdiff, 32)); + minabsdiff = _mm_min_epi16(minabsdiff, _mm_srli_epi64(minabsdiff, 16)); + *min = _mm_extract_epi16(minabsdiff, 0); +} + +unsigned int aom_avg_8x8_sse2(const uint8_t *s, int p) { + __m128i s0, s1, u0; + unsigned int avg = 0; + u0 = _mm_setzero_si128(); + s0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s)), u0); + s1 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + p)), u0); + s0 = _mm_adds_epu16(s0, s1); + s1 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 2 * p)), u0); + s0 = _mm_adds_epu16(s0, s1); + s1 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 3 * p)), u0); + s0 = _mm_adds_epu16(s0, s1); + s1 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 4 * p)), u0); + s0 = _mm_adds_epu16(s0, s1); + s1 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 5 * p)), u0); + s0 = _mm_adds_epu16(s0, s1); + s1 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 6 * p)), u0); + s0 = _mm_adds_epu16(s0, s1); + s1 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 7 * p)), u0); + s0 = _mm_adds_epu16(s0, s1); + + s0 = _mm_adds_epu16(s0, _mm_srli_si128(s0, 8)); + s0 = _mm_adds_epu16(s0, _mm_srli_epi64(s0, 32)); + s0 = _mm_adds_epu16(s0, _mm_srli_epi64(s0, 16)); + avg = _mm_extract_epi16(s0, 0); + return (avg + 32) >> 6; +} + +unsigned int aom_avg_4x4_sse2(const uint8_t *s, int p) { + __m128i s0, s1, u0; + unsigned int avg = 0; + + u0 = _mm_setzero_si128(); + s0 = _mm_unpacklo_epi8(xx_loadl_32(s), u0); + s1 = _mm_unpacklo_epi8(xx_loadl_32(s + p), u0); + s0 = _mm_adds_epu16(s0, s1); + s1 = _mm_unpacklo_epi8(xx_loadl_32(s + 2 * p), u0); + s0 = _mm_adds_epu16(s0, s1); + s1 = _mm_unpacklo_epi8(xx_loadl_32(s + 3 * p), u0); + s0 = _mm_adds_epu16(s0, s1); + + s0 = _mm_adds_epu16(s0, _mm_srli_si128(s0, 4)); + s0 = _mm_adds_epu16(s0, _mm_srli_epi64(s0, 16)); + avg = _mm_extract_epi16(s0, 0); + return (avg + 8) >> 4; +} + +static void hadamard_col8_sse2(__m128i *in, int iter) { + __m128i a0 = in[0]; + __m128i a1 = in[1]; + __m128i a2 = in[2]; + __m128i a3 = in[3]; + __m128i a4 = in[4]; + __m128i a5 = in[5]; + __m128i a6 = in[6]; + __m128i a7 = in[7]; + + __m128i b0 = _mm_add_epi16(a0, a1); + __m128i b1 = _mm_sub_epi16(a0, a1); + __m128i b2 = _mm_add_epi16(a2, a3); + __m128i b3 = _mm_sub_epi16(a2, a3); + __m128i b4 = _mm_add_epi16(a4, a5); + __m128i b5 = _mm_sub_epi16(a4, a5); + __m128i b6 = _mm_add_epi16(a6, a7); + __m128i b7 = _mm_sub_epi16(a6, a7); + + a0 = _mm_add_epi16(b0, b2); + a1 = _mm_add_epi16(b1, b3); + a2 = _mm_sub_epi16(b0, b2); + a3 = _mm_sub_epi16(b1, b3); + a4 = _mm_add_epi16(b4, b6); + a5 = _mm_add_epi16(b5, b7); + a6 = _mm_sub_epi16(b4, b6); + a7 = _mm_sub_epi16(b5, b7); + + if (iter == 0) { + b0 = _mm_add_epi16(a0, a4); + b7 = _mm_add_epi16(a1, a5); + b3 = _mm_add_epi16(a2, a6); + b4 = _mm_add_epi16(a3, a7); + b2 = _mm_sub_epi16(a0, a4); + b6 = _mm_sub_epi16(a1, a5); + b1 = _mm_sub_epi16(a2, a6); + b5 = _mm_sub_epi16(a3, a7); + + a0 = _mm_unpacklo_epi16(b0, b1); + a1 = _mm_unpacklo_epi16(b2, b3); + a2 = _mm_unpackhi_epi16(b0, b1); + a3 = _mm_unpackhi_epi16(b2, b3); + a4 = _mm_unpacklo_epi16(b4, b5); + a5 = _mm_unpacklo_epi16(b6, b7); + a6 = _mm_unpackhi_epi16(b4, b5); + a7 = _mm_unpackhi_epi16(b6, b7); + + b0 = _mm_unpacklo_epi32(a0, a1); + b1 = _mm_unpacklo_epi32(a4, a5); + b2 = _mm_unpackhi_epi32(a0, a1); + b3 = _mm_unpackhi_epi32(a4, a5); + b4 = _mm_unpacklo_epi32(a2, a3); + b5 = _mm_unpacklo_epi32(a6, a7); + b6 = _mm_unpackhi_epi32(a2, a3); + b7 = _mm_unpackhi_epi32(a6, a7); + + in[0] = _mm_unpacklo_epi64(b0, b1); + in[1] = _mm_unpackhi_epi64(b0, b1); + in[2] = _mm_unpacklo_epi64(b2, b3); + in[3] = _mm_unpackhi_epi64(b2, b3); + in[4] = _mm_unpacklo_epi64(b4, b5); + in[5] = _mm_unpackhi_epi64(b4, b5); + in[6] = _mm_unpacklo_epi64(b6, b7); + in[7] = _mm_unpackhi_epi64(b6, b7); + } else { + in[0] = _mm_add_epi16(a0, a4); + in[7] = _mm_add_epi16(a1, a5); + in[3] = _mm_add_epi16(a2, a6); + in[4] = _mm_add_epi16(a3, a7); + in[2] = _mm_sub_epi16(a0, a4); + in[6] = _mm_sub_epi16(a1, a5); + in[1] = _mm_sub_epi16(a2, a6); + in[5] = _mm_sub_epi16(a3, a7); + } +} + +void aom_hadamard_8x8_sse2(int16_t const *src_diff, int src_stride, + int16_t *coeff) { + __m128i src[8]; + src[0] = _mm_load_si128((const __m128i *)src_diff); + src[1] = _mm_load_si128((const __m128i *)(src_diff += src_stride)); + src[2] = _mm_load_si128((const __m128i *)(src_diff += src_stride)); + src[3] = _mm_load_si128((const __m128i *)(src_diff += src_stride)); + src[4] = _mm_load_si128((const __m128i *)(src_diff += src_stride)); + src[5] = _mm_load_si128((const __m128i *)(src_diff += src_stride)); + src[6] = _mm_load_si128((const __m128i *)(src_diff += src_stride)); + src[7] = _mm_load_si128((const __m128i *)(src_diff += src_stride)); + + hadamard_col8_sse2(src, 0); + hadamard_col8_sse2(src, 1); + + _mm_store_si128((__m128i *)coeff, src[0]); + coeff += 8; + _mm_store_si128((__m128i *)coeff, src[1]); + coeff += 8; + _mm_store_si128((__m128i *)coeff, src[2]); + coeff += 8; + _mm_store_si128((__m128i *)coeff, src[3]); + coeff += 8; + _mm_store_si128((__m128i *)coeff, src[4]); + coeff += 8; + _mm_store_si128((__m128i *)coeff, src[5]); + coeff += 8; + _mm_store_si128((__m128i *)coeff, src[6]); + coeff += 8; + _mm_store_si128((__m128i *)coeff, src[7]); +} + +void aom_hadamard_16x16_sse2(int16_t const *src_diff, int src_stride, + int16_t *coeff) { + int idx; + for (idx = 0; idx < 4; ++idx) { + int16_t const *src_ptr = + src_diff + (idx >> 1) * 8 * src_stride + (idx & 0x01) * 8; + aom_hadamard_8x8_sse2(src_ptr, src_stride, coeff + idx * 64); + } + + for (idx = 0; idx < 64; idx += 8) { + __m128i coeff0 = _mm_load_si128((const __m128i *)coeff); + __m128i coeff1 = _mm_load_si128((const __m128i *)(coeff + 64)); + __m128i coeff2 = _mm_load_si128((const __m128i *)(coeff + 128)); + __m128i coeff3 = _mm_load_si128((const __m128i *)(coeff + 192)); + + __m128i b0 = _mm_add_epi16(coeff0, coeff1); + __m128i b1 = _mm_sub_epi16(coeff0, coeff1); + __m128i b2 = _mm_add_epi16(coeff2, coeff3); + __m128i b3 = _mm_sub_epi16(coeff2, coeff3); + + b0 = _mm_srai_epi16(b0, 1); + b1 = _mm_srai_epi16(b1, 1); + b2 = _mm_srai_epi16(b2, 1); + b3 = _mm_srai_epi16(b3, 1); + + coeff0 = _mm_add_epi16(b0, b2); + coeff1 = _mm_add_epi16(b1, b3); + _mm_store_si128((__m128i *)coeff, coeff0); + _mm_store_si128((__m128i *)(coeff + 64), coeff1); + + coeff2 = _mm_sub_epi16(b0, b2); + coeff3 = _mm_sub_epi16(b1, b3); + _mm_store_si128((__m128i *)(coeff + 128), coeff2); + _mm_store_si128((__m128i *)(coeff + 192), coeff3); + + coeff += 8; + } +} + +int aom_satd_sse2(const int16_t *coeff, int length) { + int i; + const __m128i zero = _mm_setzero_si128(); + __m128i accum = zero; + + for (i = 0; i < length; i += 8) { + const __m128i src_line = _mm_load_si128((const __m128i *)coeff); + const __m128i inv = _mm_sub_epi16(zero, src_line); + const __m128i abs = _mm_max_epi16(src_line, inv); // abs(src_line) + const __m128i abs_lo = _mm_unpacklo_epi16(abs, zero); + const __m128i abs_hi = _mm_unpackhi_epi16(abs, zero); + const __m128i sum = _mm_add_epi32(abs_lo, abs_hi); + accum = _mm_add_epi32(accum, sum); + coeff += 8; + } + + { // cascading summation of accum + __m128i hi = _mm_srli_si128(accum, 8); + accum = _mm_add_epi32(accum, hi); + hi = _mm_srli_epi64(accum, 32); + accum = _mm_add_epi32(accum, hi); + } + + return _mm_cvtsi128_si32(accum); +} + +void aom_int_pro_row_sse2(int16_t *hbuf, uint8_t const *ref, int ref_stride, + int height) { + int idx; + __m128i zero = _mm_setzero_si128(); + __m128i src_line = _mm_loadu_si128((const __m128i *)ref); + __m128i s0 = _mm_unpacklo_epi8(src_line, zero); + __m128i s1 = _mm_unpackhi_epi8(src_line, zero); + __m128i t0, t1; + int height_1 = height - 1; + ref += ref_stride; + + for (idx = 1; idx < height_1; idx += 2) { + src_line = _mm_loadu_si128((const __m128i *)ref); + t0 = _mm_unpacklo_epi8(src_line, zero); + t1 = _mm_unpackhi_epi8(src_line, zero); + s0 = _mm_adds_epu16(s0, t0); + s1 = _mm_adds_epu16(s1, t1); + ref += ref_stride; + + src_line = _mm_loadu_si128((const __m128i *)ref); + t0 = _mm_unpacklo_epi8(src_line, zero); + t1 = _mm_unpackhi_epi8(src_line, zero); + s0 = _mm_adds_epu16(s0, t0); + s1 = _mm_adds_epu16(s1, t1); + ref += ref_stride; + } + + src_line = _mm_loadu_si128((const __m128i *)ref); + t0 = _mm_unpacklo_epi8(src_line, zero); + t1 = _mm_unpackhi_epi8(src_line, zero); + s0 = _mm_adds_epu16(s0, t0); + s1 = _mm_adds_epu16(s1, t1); + + if (height == 64) { + s0 = _mm_srai_epi16(s0, 5); + s1 = _mm_srai_epi16(s1, 5); + } else if (height == 32) { + s0 = _mm_srai_epi16(s0, 4); + s1 = _mm_srai_epi16(s1, 4); + } else { + s0 = _mm_srai_epi16(s0, 3); + s1 = _mm_srai_epi16(s1, 3); + } + + _mm_storeu_si128((__m128i *)hbuf, s0); + hbuf += 8; + _mm_storeu_si128((__m128i *)hbuf, s1); +} + +int16_t aom_int_pro_col_sse2(uint8_t const *ref, int width) { + __m128i zero = _mm_setzero_si128(); + __m128i src_line = _mm_load_si128((const __m128i *)ref); + __m128i s0 = _mm_sad_epu8(src_line, zero); + __m128i s1; + int i; + + for (i = 16; i < width; i += 16) { + ref += 16; + src_line = _mm_load_si128((const __m128i *)ref); + s1 = _mm_sad_epu8(src_line, zero); + s0 = _mm_adds_epu16(s0, s1); + } + + s1 = _mm_srli_si128(s0, 8); + s0 = _mm_adds_epu16(s0, s1); + + return _mm_extract_epi16(s0, 0); +} + +int aom_vector_var_sse2(int16_t const *ref, int16_t const *src, int bwl) { + int idx; + int width = 4 << bwl; + int16_t mean; + __m128i v0 = _mm_loadu_si128((const __m128i *)ref); + __m128i v1 = _mm_load_si128((const __m128i *)src); + __m128i diff = _mm_subs_epi16(v0, v1); + __m128i sum = diff; + __m128i sse = _mm_madd_epi16(diff, diff); + + ref += 8; + src += 8; + + for (idx = 8; idx < width; idx += 8) { + v0 = _mm_loadu_si128((const __m128i *)ref); + v1 = _mm_load_si128((const __m128i *)src); + diff = _mm_subs_epi16(v0, v1); + + sum = _mm_add_epi16(sum, diff); + v0 = _mm_madd_epi16(diff, diff); + sse = _mm_add_epi32(sse, v0); + + ref += 8; + src += 8; + } + + v0 = _mm_srli_si128(sum, 8); + sum = _mm_add_epi16(sum, v0); + v0 = _mm_srli_epi64(sum, 32); + sum = _mm_add_epi16(sum, v0); + v0 = _mm_srli_epi32(sum, 16); + sum = _mm_add_epi16(sum, v0); + + v1 = _mm_srli_si128(sse, 8); + sse = _mm_add_epi32(sse, v1); + v1 = _mm_srli_epi64(sse, 32); + sse = _mm_add_epi32(sse, v1); + + mean = _mm_extract_epi16(sum, 0); + + return _mm_cvtsi128_si32(sse) - ((mean * mean) >> (bwl + 2)); +} diff --git a/third_party/aom/aom_dsp/x86/avg_ssse3_x86_64.asm b/third_party/aom/aom_dsp/x86/avg_ssse3_x86_64.asm new file mode 100644 index 000000000..b2d150296 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/avg_ssse3_x86_64.asm @@ -0,0 +1,124 @@ +; +; Copyright (c) 2016, Alliance for Open Media. All rights reserved +; +; This source code is subject to the terms of the BSD 2 Clause License and +; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License +; was not distributed with this source code in the LICENSE file, you can +; obtain it at www.aomedia.org/license/software. If the Alliance for Open +; Media Patent License 1.0 was not distributed with this source code in the +; PATENTS file, you can obtain it at www.aomedia.org/license/patent. +; + +; + +%define private_prefix aom + +%include "third_party/x86inc/x86inc.asm" + +; This file provides SSSE3 version of the hadamard transformation. Part +; of the macro definitions are originally derived from the ffmpeg project. +; The current version applies to x86 64-bit only. + +SECTION .text + +%if ARCH_X86_64 +; matrix transpose +%macro INTERLEAVE_2X 4 + punpckh%1 m%4, m%2, m%3 + punpckl%1 m%2, m%3 + SWAP %3, %4 +%endmacro + +%macro TRANSPOSE8X8 9 + INTERLEAVE_2X wd, %1, %2, %9 + INTERLEAVE_2X wd, %3, %4, %9 + INTERLEAVE_2X wd, %5, %6, %9 + INTERLEAVE_2X wd, %7, %8, %9 + + INTERLEAVE_2X dq, %1, %3, %9 + INTERLEAVE_2X dq, %2, %4, %9 + INTERLEAVE_2X dq, %5, %7, %9 + INTERLEAVE_2X dq, %6, %8, %9 + + INTERLEAVE_2X qdq, %1, %5, %9 + INTERLEAVE_2X qdq, %3, %7, %9 + INTERLEAVE_2X qdq, %2, %6, %9 + INTERLEAVE_2X qdq, %4, %8, %9 + + SWAP %2, %5 + SWAP %4, %7 +%endmacro + +%macro HMD8_1D 0 + psubw m8, m0, m1 + psubw m9, m2, m3 + paddw m0, m1 + paddw m2, m3 + SWAP 1, 8 + SWAP 3, 9 + psubw m8, m4, m5 + psubw m9, m6, m7 + paddw m4, m5 + paddw m6, m7 + SWAP 5, 8 + SWAP 7, 9 + + psubw m8, m0, m2 + psubw m9, m1, m3 + paddw m0, m2 + paddw m1, m3 + SWAP 2, 8 + SWAP 3, 9 + psubw m8, m4, m6 + psubw m9, m5, m7 + paddw m4, m6 + paddw m5, m7 + SWAP 6, 8 + SWAP 7, 9 + + psubw m8, m0, m4 + psubw m9, m1, m5 + paddw m0, m4 + paddw m1, m5 + SWAP 4, 8 + SWAP 5, 9 + psubw m8, m2, m6 + psubw m9, m3, m7 + paddw m2, m6 + paddw m3, m7 + SWAP 6, 8 + SWAP 7, 9 +%endmacro + +INIT_XMM ssse3 +cglobal hadamard_8x8, 3, 5, 10, input, stride, output + lea r3, [2 * strideq] + lea r4, [4 * strideq] + + mova m0, [inputq] + mova m1, [inputq + r3] + lea inputq, [inputq + r4] + mova m2, [inputq] + mova m3, [inputq + r3] + lea inputq, [inputq + r4] + mova m4, [inputq] + mova m5, [inputq + r3] + lea inputq, [inputq + r4] + mova m6, [inputq] + mova m7, [inputq + r3] + + HMD8_1D + TRANSPOSE8X8 0, 1, 2, 3, 4, 5, 6, 7, 9 + HMD8_1D + + mova [outputq + 0], m0 + mova [outputq + 16], m1 + mova [outputq + 32], m2 + mova [outputq + 48], m3 + mova [outputq + 64], m4 + mova [outputq + 80], m5 + mova [outputq + 96], m6 + mova [outputq + 112], m7 + + RET +%endif diff --git a/third_party/aom/aom_dsp/x86/blend_a64_hmask_sse4.c b/third_party/aom/aom_dsp/x86/blend_a64_hmask_sse4.c new file mode 100644 index 000000000..e916e4ff9 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/blend_a64_hmask_sse4.c @@ -0,0 +1,36 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include "aom/aom_integer.h" + +#include "./aom_dsp_rtcd.h" + +// To start out, just dispatch to the function using the 2D mask and +// pass mask stride as 0. This can be improved upon if necessary. + +void aom_blend_a64_hmask_sse4_1(uint8_t *dst, uint32_t dst_stride, + const uint8_t *src0, uint32_t src0_stride, + const uint8_t *src1, uint32_t src1_stride, + const uint8_t *mask, int h, int w) { + aom_blend_a64_mask_sse4_1(dst, dst_stride, src0, src0_stride, src1, + src1_stride, mask, 0, h, w, 0, 0); +} + +#if CONFIG_HIGHBITDEPTH +void aom_highbd_blend_a64_hmask_sse4_1( + uint8_t *dst_8, uint32_t dst_stride, const uint8_t *src0_8, + uint32_t src0_stride, const uint8_t *src1_8, uint32_t src1_stride, + const uint8_t *mask, int h, int w, int bd) { + aom_highbd_blend_a64_mask_sse4_1(dst_8, dst_stride, src0_8, src0_stride, + src1_8, src1_stride, mask, 0, h, w, 0, 0, + bd); +} +#endif // CONFIG_HIGHBITDEPTH diff --git a/third_party/aom/aom_dsp/x86/blend_a64_mask_sse4.c b/third_party/aom/aom_dsp/x86/blend_a64_mask_sse4.c new file mode 100644 index 000000000..68d74e517 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/blend_a64_mask_sse4.c @@ -0,0 +1,924 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include <smmintrin.h> // SSE4.1 + +#include <assert.h> + +#include "aom/aom_integer.h" +#include "aom_ports/mem.h" +#include "aom_dsp/aom_dsp_common.h" +#include "aom_dsp/blend.h" + +#include "aom_dsp/x86/synonyms.h" +#include "aom_dsp/x86/blend_sse4.h" + +#include "./aom_dsp_rtcd.h" + +////////////////////////////////////////////////////////////////////////////// +// No sub-sampling +////////////////////////////////////////////////////////////////////////////// + +static void blend_a64_mask_w4_sse4_1(uint8_t *dst, uint32_t dst_stride, + const uint8_t *src0, uint32_t src0_stride, + const uint8_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, + int h, int w) { + const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA); + + (void)w; + + do { + const __m128i v_m0_b = xx_loadl_32(mask); + const __m128i v_m0_w = _mm_cvtepu8_epi16(v_m0_b); + const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w); + + const __m128i v_res_w = blend_4(src0, src1, v_m0_w, v_m1_w); + + const __m128i v_res_b = _mm_packus_epi16(v_res_w, v_res_w); + + xx_storel_32(dst, v_res_b); + + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + mask += mask_stride; + } while (--h); +} + +static void blend_a64_mask_w8_sse4_1(uint8_t *dst, uint32_t dst_stride, + const uint8_t *src0, uint32_t src0_stride, + const uint8_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, + int h, int w) { + const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA); + + (void)w; + + do { + const __m128i v_m0_b = xx_loadl_64(mask); + const __m128i v_m0_w = _mm_cvtepu8_epi16(v_m0_b); + const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w); + + const __m128i v_res_w = blend_8(src0, src1, v_m0_w, v_m1_w); + + const __m128i v_res_b = _mm_packus_epi16(v_res_w, v_res_w); + + xx_storel_64(dst, v_res_b); + + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + mask += mask_stride; + } while (--h); +} + +static void blend_a64_mask_w16n_sse4_1( + uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, + uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int h, int w) { + const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA); + + do { + int c; + for (c = 0; c < w; c += 16) { + const __m128i v_m0l_b = xx_loadl_64(mask + c); + const __m128i v_m0h_b = xx_loadl_64(mask + c + 8); + const __m128i v_m0l_w = _mm_cvtepu8_epi16(v_m0l_b); + const __m128i v_m0h_w = _mm_cvtepu8_epi16(v_m0h_b); + const __m128i v_m1l_w = _mm_sub_epi16(v_maxval_w, v_m0l_w); + const __m128i v_m1h_w = _mm_sub_epi16(v_maxval_w, v_m0h_w); + + const __m128i v_resl_w = blend_8(src0 + c, src1 + c, v_m0l_w, v_m1l_w); + const __m128i v_resh_w = + blend_8(src0 + c + 8, src1 + c + 8, v_m0h_w, v_m1h_w); + + const __m128i v_res_b = _mm_packus_epi16(v_resl_w, v_resh_w); + + xx_storeu_128(dst + c, v_res_b); + } + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + mask += mask_stride; + } while (--h); +} + +////////////////////////////////////////////////////////////////////////////// +// Horizontal sub-sampling +////////////////////////////////////////////////////////////////////////////// + +static void blend_a64_mask_sx_w4_sse4_1( + uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, + uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int h, int w) { + const __m128i v_zmask_b = _mm_set_epi8(0, 0xff, 0, 0xff, 0, 0xff, 0, 0xff, 0, + 0xff, 0, 0xff, 0, 0xff, 0, 0xff); + const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA); + + (void)w; + + do { + const __m128i v_r_b = xx_loadl_64(mask); + const __m128i v_a_b = _mm_avg_epu8(v_r_b, _mm_srli_si128(v_r_b, 1)); + + const __m128i v_m0_w = _mm_and_si128(v_a_b, v_zmask_b); + const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w); + + const __m128i v_res_w = blend_4(src0, src1, v_m0_w, v_m1_w); + + const __m128i v_res_b = _mm_packus_epi16(v_res_w, v_res_w); + + xx_storel_32(dst, v_res_b); + + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + mask += mask_stride; + } while (--h); +} + +static void blend_a64_mask_sx_w8_sse4_1( + uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, + uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int h, int w) { + const __m128i v_zmask_b = _mm_set_epi8(0, 0xff, 0, 0xff, 0, 0xff, 0, 0xff, 0, + 0xff, 0, 0xff, 0, 0xff, 0, 0xff); + const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA); + + (void)w; + + do { + const __m128i v_r_b = xx_loadu_128(mask); + const __m128i v_a_b = _mm_avg_epu8(v_r_b, _mm_srli_si128(v_r_b, 1)); + + const __m128i v_m0_w = _mm_and_si128(v_a_b, v_zmask_b); + const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w); + + const __m128i v_res_w = blend_8(src0, src1, v_m0_w, v_m1_w); + + const __m128i v_res_b = _mm_packus_epi16(v_res_w, v_res_w); + + xx_storel_64(dst, v_res_b); + + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + mask += mask_stride; + } while (--h); +} + +static void blend_a64_mask_sx_w16n_sse4_1( + uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, + uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int h, int w) { + const __m128i v_zmask_b = _mm_set_epi8(0, 0xff, 0, 0xff, 0, 0xff, 0, 0xff, 0, + 0xff, 0, 0xff, 0, 0xff, 0, 0xff); + const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA); + + do { + int c; + for (c = 0; c < w; c += 16) { + const __m128i v_rl_b = xx_loadu_128(mask + 2 * c); + const __m128i v_rh_b = xx_loadu_128(mask + 2 * c + 16); + const __m128i v_al_b = _mm_avg_epu8(v_rl_b, _mm_srli_si128(v_rl_b, 1)); + const __m128i v_ah_b = _mm_avg_epu8(v_rh_b, _mm_srli_si128(v_rh_b, 1)); + + const __m128i v_m0l_w = _mm_and_si128(v_al_b, v_zmask_b); + const __m128i v_m0h_w = _mm_and_si128(v_ah_b, v_zmask_b); + const __m128i v_m1l_w = _mm_sub_epi16(v_maxval_w, v_m0l_w); + const __m128i v_m1h_w = _mm_sub_epi16(v_maxval_w, v_m0h_w); + + const __m128i v_resl_w = blend_8(src0 + c, src1 + c, v_m0l_w, v_m1l_w); + const __m128i v_resh_w = + blend_8(src0 + c + 8, src1 + c + 8, v_m0h_w, v_m1h_w); + + const __m128i v_res_b = _mm_packus_epi16(v_resl_w, v_resh_w); + + xx_storeu_128(dst + c, v_res_b); + } + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + mask += mask_stride; + } while (--h); +} + +////////////////////////////////////////////////////////////////////////////// +// Vertical sub-sampling +////////////////////////////////////////////////////////////////////////////// + +static void blend_a64_mask_sy_w4_sse4_1( + uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, + uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int h, int w) { + const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA); + + (void)w; + + do { + const __m128i v_ra_b = xx_loadl_32(mask); + const __m128i v_rb_b = xx_loadl_32(mask + mask_stride); + const __m128i v_a_b = _mm_avg_epu8(v_ra_b, v_rb_b); + + const __m128i v_m0_w = _mm_cvtepu8_epi16(v_a_b); + const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w); + + const __m128i v_res_w = blend_4(src0, src1, v_m0_w, v_m1_w); + + const __m128i v_res_b = _mm_packus_epi16(v_res_w, v_res_w); + + xx_storel_32(dst, v_res_b); + + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + mask += 2 * mask_stride; + } while (--h); +} + +static void blend_a64_mask_sy_w8_sse4_1( + uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, + uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int h, int w) { + const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA); + + (void)w; + + do { + const __m128i v_ra_b = xx_loadl_64(mask); + const __m128i v_rb_b = xx_loadl_64(mask + mask_stride); + const __m128i v_a_b = _mm_avg_epu8(v_ra_b, v_rb_b); + + const __m128i v_m0_w = _mm_cvtepu8_epi16(v_a_b); + const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w); + + const __m128i v_res_w = blend_8(src0, src1, v_m0_w, v_m1_w); + + const __m128i v_res_b = _mm_packus_epi16(v_res_w, v_res_w); + + xx_storel_64(dst, v_res_b); + + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + mask += 2 * mask_stride; + } while (--h); +} + +static void blend_a64_mask_sy_w16n_sse4_1( + uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, + uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int h, int w) { + const __m128i v_zero = _mm_setzero_si128(); + const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA); + + do { + int c; + for (c = 0; c < w; c += 16) { + const __m128i v_ra_b = xx_loadu_128(mask + c); + const __m128i v_rb_b = xx_loadu_128(mask + c + mask_stride); + const __m128i v_a_b = _mm_avg_epu8(v_ra_b, v_rb_b); + + const __m128i v_m0l_w = _mm_cvtepu8_epi16(v_a_b); + const __m128i v_m0h_w = _mm_unpackhi_epi8(v_a_b, v_zero); + const __m128i v_m1l_w = _mm_sub_epi16(v_maxval_w, v_m0l_w); + const __m128i v_m1h_w = _mm_sub_epi16(v_maxval_w, v_m0h_w); + + const __m128i v_resl_w = blend_8(src0 + c, src1 + c, v_m0l_w, v_m1l_w); + const __m128i v_resh_w = + blend_8(src0 + c + 8, src1 + c + 8, v_m0h_w, v_m1h_w); + + const __m128i v_res_b = _mm_packus_epi16(v_resl_w, v_resh_w); + + xx_storeu_128(dst + c, v_res_b); + } + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + mask += 2 * mask_stride; + } while (--h); +} + +////////////////////////////////////////////////////////////////////////////// +// Horizontal and Vertical sub-sampling +////////////////////////////////////////////////////////////////////////////// + +static void blend_a64_mask_sx_sy_w4_sse4_1( + uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, + uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int h, int w) { + const __m128i v_zmask_b = _mm_set_epi8(0, 0xff, 0, 0xff, 0, 0xff, 0, 0xff, 0, + 0xff, 0, 0xff, 0, 0xff, 0, 0xff); + const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA); + + (void)w; + + do { + const __m128i v_ra_b = xx_loadl_64(mask); + const __m128i v_rb_b = xx_loadl_64(mask + mask_stride); + const __m128i v_rvs_b = _mm_add_epi8(v_ra_b, v_rb_b); + const __m128i v_rvsa_w = _mm_and_si128(v_rvs_b, v_zmask_b); + const __m128i v_rvsb_w = + _mm_and_si128(_mm_srli_si128(v_rvs_b, 1), v_zmask_b); + const __m128i v_rs_w = _mm_add_epi16(v_rvsa_w, v_rvsb_w); + + const __m128i v_m0_w = xx_roundn_epu16(v_rs_w, 2); + const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w); + + const __m128i v_res_w = blend_4(src0, src1, v_m0_w, v_m1_w); + + const __m128i v_res_b = _mm_packus_epi16(v_res_w, v_res_w); + + xx_storel_32(dst, v_res_b); + + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + mask += 2 * mask_stride; + } while (--h); +} + +static void blend_a64_mask_sx_sy_w8_sse4_1( + uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, + uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int h, int w) { + const __m128i v_zmask_b = _mm_set_epi8(0, 0xff, 0, 0xff, 0, 0xff, 0, 0xff, 0, + 0xff, 0, 0xff, 0, 0xff, 0, 0xff); + const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA); + + (void)w; + + do { + const __m128i v_ra_b = xx_loadu_128(mask); + const __m128i v_rb_b = xx_loadu_128(mask + mask_stride); + const __m128i v_rvs_b = _mm_add_epi8(v_ra_b, v_rb_b); + const __m128i v_rvsa_w = _mm_and_si128(v_rvs_b, v_zmask_b); + const __m128i v_rvsb_w = + _mm_and_si128(_mm_srli_si128(v_rvs_b, 1), v_zmask_b); + const __m128i v_rs_w = _mm_add_epi16(v_rvsa_w, v_rvsb_w); + + const __m128i v_m0_w = xx_roundn_epu16(v_rs_w, 2); + const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w); + + const __m128i v_res_w = blend_8(src0, src1, v_m0_w, v_m1_w); + + const __m128i v_res_b = _mm_packus_epi16(v_res_w, v_res_w); + + xx_storel_64(dst, v_res_b); + + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + mask += 2 * mask_stride; + } while (--h); +} + +static void blend_a64_mask_sx_sy_w16n_sse4_1( + uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, + uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int h, int w) { + const __m128i v_zmask_b = _mm_set_epi8(0, 0xff, 0, 0xff, 0, 0xff, 0, 0xff, 0, + 0xff, 0, 0xff, 0, 0xff, 0, 0xff); + const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA); + + do { + int c; + for (c = 0; c < w; c += 16) { + const __m128i v_ral_b = xx_loadu_128(mask + 2 * c); + const __m128i v_rah_b = xx_loadu_128(mask + 2 * c + 16); + const __m128i v_rbl_b = xx_loadu_128(mask + mask_stride + 2 * c); + const __m128i v_rbh_b = xx_loadu_128(mask + mask_stride + 2 * c + 16); + const __m128i v_rvsl_b = _mm_add_epi8(v_ral_b, v_rbl_b); + const __m128i v_rvsh_b = _mm_add_epi8(v_rah_b, v_rbh_b); + const __m128i v_rvsal_w = _mm_and_si128(v_rvsl_b, v_zmask_b); + const __m128i v_rvsah_w = _mm_and_si128(v_rvsh_b, v_zmask_b); + const __m128i v_rvsbl_w = + _mm_and_si128(_mm_srli_si128(v_rvsl_b, 1), v_zmask_b); + const __m128i v_rvsbh_w = + _mm_and_si128(_mm_srli_si128(v_rvsh_b, 1), v_zmask_b); + const __m128i v_rsl_w = _mm_add_epi16(v_rvsal_w, v_rvsbl_w); + const __m128i v_rsh_w = _mm_add_epi16(v_rvsah_w, v_rvsbh_w); + + const __m128i v_m0l_w = xx_roundn_epu16(v_rsl_w, 2); + const __m128i v_m0h_w = xx_roundn_epu16(v_rsh_w, 2); + const __m128i v_m1l_w = _mm_sub_epi16(v_maxval_w, v_m0l_w); + const __m128i v_m1h_w = _mm_sub_epi16(v_maxval_w, v_m0h_w); + + const __m128i v_resl_w = blend_8(src0 + c, src1 + c, v_m0l_w, v_m1l_w); + const __m128i v_resh_w = + blend_8(src0 + c + 8, src1 + c + 8, v_m0h_w, v_m1h_w); + + const __m128i v_res_b = _mm_packus_epi16(v_resl_w, v_resh_w); + + xx_storeu_128(dst + c, v_res_b); + } + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + mask += 2 * mask_stride; + } while (--h); +} + +////////////////////////////////////////////////////////////////////////////// +// Dispatch +////////////////////////////////////////////////////////////////////////////// + +void aom_blend_a64_mask_sse4_1(uint8_t *dst, uint32_t dst_stride, + const uint8_t *src0, uint32_t src0_stride, + const uint8_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int h, + int w, int suby, int subx) { + typedef void (*blend_fn)( + uint8_t * dst, uint32_t dst_stride, const uint8_t *src0, + uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int h, int w); + + // Dimensions are: width_index X subx X suby + static const blend_fn blend[3][2][2] = { + { // w % 16 == 0 + { blend_a64_mask_w16n_sse4_1, blend_a64_mask_sy_w16n_sse4_1 }, + { blend_a64_mask_sx_w16n_sse4_1, blend_a64_mask_sx_sy_w16n_sse4_1 } }, + { // w == 4 + { blend_a64_mask_w4_sse4_1, blend_a64_mask_sy_w4_sse4_1 }, + { blend_a64_mask_sx_w4_sse4_1, blend_a64_mask_sx_sy_w4_sse4_1 } }, + { // w == 8 + { blend_a64_mask_w8_sse4_1, blend_a64_mask_sy_w8_sse4_1 }, + { blend_a64_mask_sx_w8_sse4_1, blend_a64_mask_sx_sy_w8_sse4_1 } } + }; + + assert(IMPLIES(src0 == dst, src0_stride == dst_stride)); + assert(IMPLIES(src1 == dst, src1_stride == dst_stride)); + + assert(h >= 1); + assert(w >= 1); + assert(IS_POWER_OF_TWO(h)); + assert(IS_POWER_OF_TWO(w)); + + if (UNLIKELY((h | w) & 3)) { // if (w <= 2 || h <= 2) + aom_blend_a64_mask_c(dst, dst_stride, src0, src0_stride, src1, src1_stride, + mask, mask_stride, h, w, suby, subx); + } else { + blend[(w >> 2) & 3][subx != 0][suby != 0](dst, dst_stride, src0, + src0_stride, src1, src1_stride, + mask, mask_stride, h, w); + } +} + +#if CONFIG_HIGHBITDEPTH +////////////////////////////////////////////////////////////////////////////// +// No sub-sampling +////////////////////////////////////////////////////////////////////////////// + +static INLINE void blend_a64_mask_bn_w4_sse4_1( + uint16_t *dst, uint32_t dst_stride, const uint16_t *src0, + uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int h, blend_unit_fn blend) { + const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA); + + do { + const __m128i v_m0_b = xx_loadl_32(mask); + const __m128i v_m0_w = _mm_cvtepu8_epi16(v_m0_b); + const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w); + + const __m128i v_res_w = blend(src0, src1, v_m0_w, v_m1_w); + + xx_storel_64(dst, v_res_w); + + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + mask += mask_stride; + } while (--h); +} + +static void blend_a64_mask_b10_w4_sse4_1( + uint16_t *dst, uint32_t dst_stride, const uint16_t *src0, + uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int h, int w) { + (void)w; + blend_a64_mask_bn_w4_sse4_1(dst, dst_stride, src0, src0_stride, src1, + src1_stride, mask, mask_stride, h, blend_4_b10); +} + +static void blend_a64_mask_b12_w4_sse4_1( + uint16_t *dst, uint32_t dst_stride, const uint16_t *src0, + uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int h, int w) { + (void)w; + blend_a64_mask_bn_w4_sse4_1(dst, dst_stride, src0, src0_stride, src1, + src1_stride, mask, mask_stride, h, blend_4_b12); +} + +static INLINE void blend_a64_mask_bn_w8n_sse4_1( + uint16_t *dst, uint32_t dst_stride, const uint16_t *src0, + uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int h, int w, + blend_unit_fn blend) { + const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA); + + do { + int c; + for (c = 0; c < w; c += 8) { + const __m128i v_m0_b = xx_loadl_64(mask + c); + const __m128i v_m0_w = _mm_cvtepu8_epi16(v_m0_b); + const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w); + + const __m128i v_res_w = blend(src0 + c, src1 + c, v_m0_w, v_m1_w); + + xx_storeu_128(dst + c, v_res_w); + } + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + mask += mask_stride; + } while (--h); +} + +static void blend_a64_mask_b10_w8n_sse4_1( + uint16_t *dst, uint32_t dst_stride, const uint16_t *src0, + uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int h, int w) { + blend_a64_mask_bn_w8n_sse4_1(dst, dst_stride, src0, src0_stride, src1, + src1_stride, mask, mask_stride, h, w, + blend_8_b10); +} + +static void blend_a64_mask_b12_w8n_sse4_1( + uint16_t *dst, uint32_t dst_stride, const uint16_t *src0, + uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int h, int w) { + blend_a64_mask_bn_w8n_sse4_1(dst, dst_stride, src0, src0_stride, src1, + src1_stride, mask, mask_stride, h, w, + blend_8_b12); +} + +////////////////////////////////////////////////////////////////////////////// +// Horizontal sub-sampling +////////////////////////////////////////////////////////////////////////////// + +static INLINE void blend_a64_mask_bn_sx_w4_sse4_1( + uint16_t *dst, uint32_t dst_stride, const uint16_t *src0, + uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int h, blend_unit_fn blend) { + const __m128i v_zmask_b = _mm_set_epi8(0, 0xff, 0, 0xff, 0, 0xff, 0, 0xff, 0, + 0xff, 0, 0xff, 0, 0xff, 0, 0xff); + const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA); + + do { + const __m128i v_r_b = xx_loadl_64(mask); + const __m128i v_a_b = _mm_avg_epu8(v_r_b, _mm_srli_si128(v_r_b, 1)); + + const __m128i v_m0_w = _mm_and_si128(v_a_b, v_zmask_b); + const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w); + + const __m128i v_res_w = blend(src0, src1, v_m0_w, v_m1_w); + + xx_storel_64(dst, v_res_w); + + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + mask += mask_stride; + } while (--h); +} + +static void blend_a64_mask_b10_sx_w4_sse4_1( + uint16_t *dst, uint32_t dst_stride, const uint16_t *src0, + uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int h, int w) { + (void)w; + blend_a64_mask_bn_sx_w4_sse4_1(dst, dst_stride, src0, src0_stride, src1, + src1_stride, mask, mask_stride, h, + blend_4_b10); +} + +static void blend_a64_mask_b12_sx_w4_sse4_1( + uint16_t *dst, uint32_t dst_stride, const uint16_t *src0, + uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int h, int w) { + (void)w; + blend_a64_mask_bn_sx_w4_sse4_1(dst, dst_stride, src0, src0_stride, src1, + src1_stride, mask, mask_stride, h, + blend_4_b12); +} + +static INLINE void blend_a64_mask_bn_sx_w8n_sse4_1( + uint16_t *dst, uint32_t dst_stride, const uint16_t *src0, + uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int h, int w, + blend_unit_fn blend) { + const __m128i v_zmask_b = _mm_set_epi8(0, 0xff, 0, 0xff, 0, 0xff, 0, 0xff, 0, + 0xff, 0, 0xff, 0, 0xff, 0, 0xff); + const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA); + + do { + int c; + for (c = 0; c < w; c += 8) { + const __m128i v_r_b = xx_loadu_128(mask + 2 * c); + const __m128i v_a_b = _mm_avg_epu8(v_r_b, _mm_srli_si128(v_r_b, 1)); + + const __m128i v_m0_w = _mm_and_si128(v_a_b, v_zmask_b); + const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w); + + const __m128i v_res_w = blend(src0 + c, src1 + c, v_m0_w, v_m1_w); + + xx_storeu_128(dst + c, v_res_w); + } + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + mask += mask_stride; + } while (--h); +} + +static void blend_a64_mask_b10_sx_w8n_sse4_1( + uint16_t *dst, uint32_t dst_stride, const uint16_t *src0, + uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int h, int w) { + blend_a64_mask_bn_sx_w8n_sse4_1(dst, dst_stride, src0, src0_stride, src1, + src1_stride, mask, mask_stride, h, w, + blend_8_b10); +} + +static void blend_a64_mask_b12_sx_w8n_sse4_1( + uint16_t *dst, uint32_t dst_stride, const uint16_t *src0, + uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int h, int w) { + blend_a64_mask_bn_sx_w8n_sse4_1(dst, dst_stride, src0, src0_stride, src1, + src1_stride, mask, mask_stride, h, w, + blend_8_b12); +} + +////////////////////////////////////////////////////////////////////////////// +// Vertical sub-sampling +////////////////////////////////////////////////////////////////////////////// + +static INLINE void blend_a64_mask_bn_sy_w4_sse4_1( + uint16_t *dst, uint32_t dst_stride, const uint16_t *src0, + uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int h, blend_unit_fn blend) { + const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA); + + do { + const __m128i v_ra_b = xx_loadl_32(mask); + const __m128i v_rb_b = xx_loadl_32(mask + mask_stride); + const __m128i v_a_b = _mm_avg_epu8(v_ra_b, v_rb_b); + + const __m128i v_m0_w = _mm_cvtepu8_epi16(v_a_b); + const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w); + + const __m128i v_res_w = blend(src0, src1, v_m0_w, v_m1_w); + + xx_storel_64(dst, v_res_w); + + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + mask += 2 * mask_stride; + } while (--h); +} + +static void blend_a64_mask_b10_sy_w4_sse4_1( + uint16_t *dst, uint32_t dst_stride, const uint16_t *src0, + uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int h, int w) { + (void)w; + blend_a64_mask_bn_sy_w4_sse4_1(dst, dst_stride, src0, src0_stride, src1, + src1_stride, mask, mask_stride, h, + blend_4_b10); +} + +static void blend_a64_mask_b12_sy_w4_sse4_1( + uint16_t *dst, uint32_t dst_stride, const uint16_t *src0, + uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int h, int w) { + (void)w; + blend_a64_mask_bn_sy_w4_sse4_1(dst, dst_stride, src0, src0_stride, src1, + src1_stride, mask, mask_stride, h, + blend_4_b12); +} + +static INLINE void blend_a64_mask_bn_sy_w8n_sse4_1( + uint16_t *dst, uint32_t dst_stride, const uint16_t *src0, + uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int h, int w, + blend_unit_fn blend) { + const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA); + + do { + int c; + for (c = 0; c < w; c += 8) { + const __m128i v_ra_b = xx_loadl_64(mask + c); + const __m128i v_rb_b = xx_loadl_64(mask + c + mask_stride); + const __m128i v_a_b = _mm_avg_epu8(v_ra_b, v_rb_b); + + const __m128i v_m0_w = _mm_cvtepu8_epi16(v_a_b); + const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w); + + const __m128i v_res_w = blend(src0 + c, src1 + c, v_m0_w, v_m1_w); + + xx_storeu_128(dst + c, v_res_w); + } + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + mask += 2 * mask_stride; + } while (--h); +} + +static void blend_a64_mask_b10_sy_w8n_sse4_1( + uint16_t *dst, uint32_t dst_stride, const uint16_t *src0, + uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int h, int w) { + blend_a64_mask_bn_sy_w8n_sse4_1(dst, dst_stride, src0, src0_stride, src1, + src1_stride, mask, mask_stride, h, w, + blend_8_b10); +} + +static void blend_a64_mask_b12_sy_w8n_sse4_1( + uint16_t *dst, uint32_t dst_stride, const uint16_t *src0, + uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int h, int w) { + blend_a64_mask_bn_sy_w8n_sse4_1(dst, dst_stride, src0, src0_stride, src1, + src1_stride, mask, mask_stride, h, w, + blend_8_b12); +} + +////////////////////////////////////////////////////////////////////////////// +// Horizontal and Vertical sub-sampling +////////////////////////////////////////////////////////////////////////////// + +static INLINE void blend_a64_mask_bn_sx_sy_w4_sse4_1( + uint16_t *dst, uint32_t dst_stride, const uint16_t *src0, + uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int h, blend_unit_fn blend) { + const __m128i v_zmask_b = _mm_set_epi8(0, 0xff, 0, 0xff, 0, 0xff, 0, 0xff, 0, + 0xff, 0, 0xff, 0, 0xff, 0, 0xff); + const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA); + + do { + const __m128i v_ra_b = xx_loadl_64(mask); + const __m128i v_rb_b = xx_loadl_64(mask + mask_stride); + const __m128i v_rvs_b = _mm_add_epi8(v_ra_b, v_rb_b); + const __m128i v_rvsa_w = _mm_and_si128(v_rvs_b, v_zmask_b); + const __m128i v_rvsb_w = + _mm_and_si128(_mm_srli_si128(v_rvs_b, 1), v_zmask_b); + const __m128i v_rs_w = _mm_add_epi16(v_rvsa_w, v_rvsb_w); + + const __m128i v_m0_w = xx_roundn_epu16(v_rs_w, 2); + const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w); + + const __m128i v_res_w = blend(src0, src1, v_m0_w, v_m1_w); + + xx_storel_64(dst, v_res_w); + + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + mask += 2 * mask_stride; + } while (--h); +} + +static void blend_a64_mask_b10_sx_sy_w4_sse4_1( + uint16_t *dst, uint32_t dst_stride, const uint16_t *src0, + uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int h, int w) { + (void)w; + blend_a64_mask_bn_sx_sy_w4_sse4_1(dst, dst_stride, src0, src0_stride, src1, + src1_stride, mask, mask_stride, h, + blend_4_b10); +} + +static void blend_a64_mask_b12_sx_sy_w4_sse4_1( + uint16_t *dst, uint32_t dst_stride, const uint16_t *src0, + uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int h, int w) { + (void)w; + blend_a64_mask_bn_sx_sy_w4_sse4_1(dst, dst_stride, src0, src0_stride, src1, + src1_stride, mask, mask_stride, h, + blend_4_b12); +} + +static INLINE void blend_a64_mask_bn_sx_sy_w8n_sse4_1( + uint16_t *dst, uint32_t dst_stride, const uint16_t *src0, + uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int h, int w, + blend_unit_fn blend) { + const __m128i v_zmask_b = _mm_set_epi8(0, 0xff, 0, 0xff, 0, 0xff, 0, 0xff, 0, + 0xff, 0, 0xff, 0, 0xff, 0, 0xff); + const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA); + + do { + int c; + for (c = 0; c < w; c += 8) { + const __m128i v_ra_b = xx_loadu_128(mask + 2 * c); + const __m128i v_rb_b = xx_loadu_128(mask + 2 * c + mask_stride); + const __m128i v_rvs_b = _mm_add_epi8(v_ra_b, v_rb_b); + const __m128i v_rvsa_w = _mm_and_si128(v_rvs_b, v_zmask_b); + const __m128i v_rvsb_w = + _mm_and_si128(_mm_srli_si128(v_rvs_b, 1), v_zmask_b); + const __m128i v_rs_w = _mm_add_epi16(v_rvsa_w, v_rvsb_w); + + const __m128i v_m0_w = xx_roundn_epu16(v_rs_w, 2); + const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w); + + const __m128i v_res_w = blend(src0 + c, src1 + c, v_m0_w, v_m1_w); + + xx_storeu_128(dst + c, v_res_w); + } + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + mask += 2 * mask_stride; + } while (--h); +} + +static void blend_a64_mask_b10_sx_sy_w8n_sse4_1( + uint16_t *dst, uint32_t dst_stride, const uint16_t *src0, + uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int h, int w) { + blend_a64_mask_bn_sx_sy_w8n_sse4_1(dst, dst_stride, src0, src0_stride, src1, + src1_stride, mask, mask_stride, h, w, + blend_8_b10); +} + +static void blend_a64_mask_b12_sx_sy_w8n_sse4_1( + uint16_t *dst, uint32_t dst_stride, const uint16_t *src0, + uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int h, int w) { + blend_a64_mask_bn_sx_sy_w8n_sse4_1(dst, dst_stride, src0, src0_stride, src1, + src1_stride, mask, mask_stride, h, w, + blend_8_b12); +} + +////////////////////////////////////////////////////////////////////////////// +// Dispatch +////////////////////////////////////////////////////////////////////////////// + +void aom_highbd_blend_a64_mask_sse4_1(uint8_t *dst_8, uint32_t dst_stride, + const uint8_t *src0_8, + uint32_t src0_stride, + const uint8_t *src1_8, + uint32_t src1_stride, const uint8_t *mask, + uint32_t mask_stride, int h, int w, + int suby, int subx, int bd) { + typedef void (*blend_fn)( + uint16_t * dst, uint32_t dst_stride, const uint16_t *src0, + uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride, + const uint8_t *mask, uint32_t mask_stride, int h, int w); + + // Dimensions are: bd_index X width_index X subx X suby + static const blend_fn blend[2][2][2][2] = { + { // bd == 8 or 10 + { // w % 8 == 0 + { blend_a64_mask_b10_w8n_sse4_1, blend_a64_mask_b10_sy_w8n_sse4_1 }, + { blend_a64_mask_b10_sx_w8n_sse4_1, + blend_a64_mask_b10_sx_sy_w8n_sse4_1 } }, + { // w == 4 + { blend_a64_mask_b10_w4_sse4_1, blend_a64_mask_b10_sy_w4_sse4_1 }, + { blend_a64_mask_b10_sx_w4_sse4_1, + blend_a64_mask_b10_sx_sy_w4_sse4_1 } } }, + { // bd == 12 + { // w % 8 == 0 + { blend_a64_mask_b12_w8n_sse4_1, blend_a64_mask_b12_sy_w8n_sse4_1 }, + { blend_a64_mask_b12_sx_w8n_sse4_1, + blend_a64_mask_b12_sx_sy_w8n_sse4_1 } }, + { // w == 4 + { blend_a64_mask_b12_w4_sse4_1, blend_a64_mask_b12_sy_w4_sse4_1 }, + { blend_a64_mask_b12_sx_w4_sse4_1, + blend_a64_mask_b12_sx_sy_w4_sse4_1 } } } + }; + + assert(IMPLIES(src0_8 == dst_8, src0_stride == dst_stride)); + assert(IMPLIES(src1_8 == dst_8, src1_stride == dst_stride)); + + assert(h >= 1); + assert(w >= 1); + assert(IS_POWER_OF_TWO(h)); + assert(IS_POWER_OF_TWO(w)); + + assert(bd == 8 || bd == 10 || bd == 12); + if (UNLIKELY((h | w) & 3)) { // if (w <= 2 || h <= 2) + aom_highbd_blend_a64_mask_c(dst_8, dst_stride, src0_8, src0_stride, src1_8, + src1_stride, mask, mask_stride, h, w, suby, + subx, bd); + } else { + uint16_t *const dst = CONVERT_TO_SHORTPTR(dst_8); + const uint16_t *const src0 = CONVERT_TO_SHORTPTR(src0_8); + const uint16_t *const src1 = CONVERT_TO_SHORTPTR(src1_8); + + blend[bd == 12][(w >> 2) & 1][subx != 0][suby != 0]( + dst, dst_stride, src0, src0_stride, src1, src1_stride, mask, + mask_stride, h, w); + } +} +#endif // CONFIG_HIGHBITDEPTH diff --git a/third_party/aom/aom_dsp/x86/blend_a64_vmask_sse4.c b/third_party/aom/aom_dsp/x86/blend_a64_vmask_sse4.c new file mode 100644 index 000000000..9dabe5b79 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/blend_a64_vmask_sse4.c @@ -0,0 +1,285 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include <smmintrin.h> // SSE4.1 + +#include <assert.h> + +#include "aom/aom_integer.h" +#include "aom_ports/mem.h" +#include "aom_dsp/aom_dsp_common.h" +#include "aom_dsp/blend.h" + +#include "aom_dsp/x86/synonyms.h" +#include "aom_dsp/x86/blend_sse4.h" + +#include "./aom_dsp_rtcd.h" + +////////////////////////////////////////////////////////////////////////////// +// Implementation - No sub-sampling +////////////////////////////////////////////////////////////////////////////// + +static void blend_a64_vmask_w4_sse4_1(uint8_t *dst, uint32_t dst_stride, + const uint8_t *src0, uint32_t src0_stride, + const uint8_t *src1, uint32_t src1_stride, + const uint8_t *mask, int h, int w) { + const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA); + + (void)w; + + do { + const __m128i v_m0_w = _mm_set1_epi16(*mask); + const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w); + + const __m128i v_res_w = blend_4(src0, src1, v_m0_w, v_m1_w); + + const __m128i v_res_b = _mm_packus_epi16(v_res_w, v_res_w); + + xx_storel_32(dst, v_res_b); + + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + mask += 1; + } while (--h); +} + +static void blend_a64_vmask_w8_sse4_1(uint8_t *dst, uint32_t dst_stride, + const uint8_t *src0, uint32_t src0_stride, + const uint8_t *src1, uint32_t src1_stride, + const uint8_t *mask, int h, int w) { + const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA); + + (void)w; + + do { + const __m128i v_m0_w = _mm_set1_epi16(*mask); + const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w); + + const __m128i v_res_w = blend_8(src0, src1, v_m0_w, v_m1_w); + + const __m128i v_res_b = _mm_packus_epi16(v_res_w, v_res_w); + + xx_storel_64(dst, v_res_b); + + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + mask += 1; + } while (--h); +} + +static void blend_a64_vmask_w16n_sse4_1(uint8_t *dst, uint32_t dst_stride, + const uint8_t *src0, + uint32_t src0_stride, + const uint8_t *src1, + uint32_t src1_stride, + const uint8_t *mask, int h, int w) { + const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA); + + do { + int c; + const __m128i v_m0_w = _mm_set1_epi16(*mask); + const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w); + for (c = 0; c < w; c += 16) { + const __m128i v_resl_w = blend_8(src0 + c, src1 + c, v_m0_w, v_m1_w); + const __m128i v_resh_w = + blend_8(src0 + c + 8, src1 + c + 8, v_m0_w, v_m1_w); + + const __m128i v_res_b = _mm_packus_epi16(v_resl_w, v_resh_w); + + xx_storeu_128(dst + c, v_res_b); + } + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + mask += 1; + } while (--h); +} + +////////////////////////////////////////////////////////////////////////////// +// Dispatch +////////////////////////////////////////////////////////////////////////////// + +void aom_blend_a64_vmask_sse4_1(uint8_t *dst, uint32_t dst_stride, + const uint8_t *src0, uint32_t src0_stride, + const uint8_t *src1, uint32_t src1_stride, + const uint8_t *mask, int h, int w) { + typedef void (*blend_fn)(uint8_t * dst, uint32_t dst_stride, + const uint8_t *src0, uint32_t src0_stride, + const uint8_t *src1, uint32_t src1_stride, + const uint8_t *mask, int h, int w); + + // Dimension: width_index + static const blend_fn blend[9] = { + blend_a64_vmask_w16n_sse4_1, // w % 16 == 0 + aom_blend_a64_vmask_c, // w == 1 + aom_blend_a64_vmask_c, // w == 2 + NULL, // INVALID + blend_a64_vmask_w4_sse4_1, // w == 4 + NULL, // INVALID + NULL, // INVALID + NULL, // INVALID + blend_a64_vmask_w8_sse4_1, // w == 8 + }; + + assert(IMPLIES(src0 == dst, src0_stride == dst_stride)); + assert(IMPLIES(src1 == dst, src1_stride == dst_stride)); + + assert(h >= 1); + assert(w >= 1); + assert(IS_POWER_OF_TWO(h)); + assert(IS_POWER_OF_TWO(w)); + + blend[w & 0xf](dst, dst_stride, src0, src0_stride, src1, src1_stride, mask, h, + w); +} + +#if CONFIG_HIGHBITDEPTH +////////////////////////////////////////////////////////////////////////////// +// Implementation - No sub-sampling +////////////////////////////////////////////////////////////////////////////// + +static INLINE void blend_a64_vmask_bn_w4_sse4_1( + uint16_t *dst, uint32_t dst_stride, const uint16_t *src0, + uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride, + const uint8_t *mask, int h, blend_unit_fn blend) { + const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA); + + do { + const __m128i v_m0_w = _mm_set1_epi16(*mask); + const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w); + + const __m128i v_res_w = blend(src0, src1, v_m0_w, v_m1_w); + + xx_storel_64(dst, v_res_w); + + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + mask += 1; + } while (--h); +} + +static void blend_a64_vmask_b10_w4_sse4_1(uint16_t *dst, uint32_t dst_stride, + const uint16_t *src0, + uint32_t src0_stride, + const uint16_t *src1, + uint32_t src1_stride, + const uint8_t *mask, int h, int w) { + (void)w; + blend_a64_vmask_bn_w4_sse4_1(dst, dst_stride, src0, src0_stride, src1, + src1_stride, mask, h, blend_4_b10); +} + +static void blend_a64_vmask_b12_w4_sse4_1(uint16_t *dst, uint32_t dst_stride, + const uint16_t *src0, + uint32_t src0_stride, + const uint16_t *src1, + uint32_t src1_stride, + const uint8_t *mask, int h, int w) { + (void)w; + blend_a64_vmask_bn_w4_sse4_1(dst, dst_stride, src0, src0_stride, src1, + src1_stride, mask, h, blend_4_b12); +} + +static INLINE void blend_a64_vmask_bn_w8n_sse4_1( + uint16_t *dst, uint32_t dst_stride, const uint16_t *src0, + uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride, + const uint8_t *mask, int h, int w, blend_unit_fn blend) { + const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA); + + do { + int c; + const __m128i v_m0_w = _mm_set1_epi16(*mask); + const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w); + for (c = 0; c < w; c += 8) { + const __m128i v_res_w = blend(src0 + c, src1 + c, v_m0_w, v_m1_w); + + xx_storeu_128(dst + c, v_res_w); + } + dst += dst_stride; + src0 += src0_stride; + src1 += src1_stride; + mask += 1; + } while (--h); +} + +static void blend_a64_vmask_b10_w8n_sse4_1(uint16_t *dst, uint32_t dst_stride, + const uint16_t *src0, + uint32_t src0_stride, + const uint16_t *src1, + uint32_t src1_stride, + const uint8_t *mask, int h, int w) { + blend_a64_vmask_bn_w8n_sse4_1(dst, dst_stride, src0, src0_stride, src1, + src1_stride, mask, h, w, blend_8_b10); +} + +static void blend_a64_vmask_b12_w8n_sse4_1(uint16_t *dst, uint32_t dst_stride, + const uint16_t *src0, + uint32_t src0_stride, + const uint16_t *src1, + uint32_t src1_stride, + const uint8_t *mask, int h, int w) { + blend_a64_vmask_bn_w8n_sse4_1(dst, dst_stride, src0, src0_stride, src1, + src1_stride, mask, h, w, blend_8_b12); +} + +////////////////////////////////////////////////////////////////////////////// +// Dispatch +////////////////////////////////////////////////////////////////////////////// + +void aom_highbd_blend_a64_vmask_sse4_1( + uint8_t *dst_8, uint32_t dst_stride, const uint8_t *src0_8, + uint32_t src0_stride, const uint8_t *src1_8, uint32_t src1_stride, + const uint8_t *mask, int h, int w, int bd) { + typedef void (*blend_fn)(uint16_t * dst, uint32_t dst_stride, + const uint16_t *src0, uint32_t src0_stride, + const uint16_t *src1, uint32_t src1_stride, + const uint8_t *mask, int h, int w); + + // Dimensions are: bd_index X width_index + static const blend_fn blend[2][2] = { + { + // bd == 8 or 10 + blend_a64_vmask_b10_w8n_sse4_1, // w % 8 == 0 + blend_a64_vmask_b10_w4_sse4_1, // w == 4 + }, + { + // bd == 12 + blend_a64_vmask_b12_w8n_sse4_1, // w % 8 == 0 + blend_a64_vmask_b12_w4_sse4_1, // w == 4 + } + }; + + assert(IMPLIES(src0_8 == dst_8, src0_stride == dst_stride)); + assert(IMPLIES(src1_8 == dst_8, src1_stride == dst_stride)); + + assert(h >= 1); + assert(w >= 1); + assert(IS_POWER_OF_TWO(h)); + assert(IS_POWER_OF_TWO(w)); + + assert(bd == 8 || bd == 10 || bd == 12); + + if (UNLIKELY((h | w) & 3)) { // if (w <= 2 || h <= 2) + aom_highbd_blend_a64_vmask_c(dst_8, dst_stride, src0_8, src0_stride, src1_8, + src1_stride, mask, h, w, bd); + } else { + uint16_t *const dst = CONVERT_TO_SHORTPTR(dst_8); + const uint16_t *const src0 = CONVERT_TO_SHORTPTR(src0_8); + const uint16_t *const src1 = CONVERT_TO_SHORTPTR(src1_8); + + blend[bd == 12][(w >> 2) & 1](dst, dst_stride, src0, src0_stride, src1, + src1_stride, mask, h, w); + } +} +#endif // CONFIG_HIGHBITDEPTH diff --git a/third_party/aom/aom_dsp/x86/blend_sse4.h b/third_party/aom/aom_dsp/x86/blend_sse4.h new file mode 100644 index 000000000..daa2b2b3a --- /dev/null +++ b/third_party/aom/aom_dsp/x86/blend_sse4.h @@ -0,0 +1,146 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#ifndef AOM_DSP_X86_BLEND_SSE4_H_ +#define AOM_DSP_X86_BLEND_SSE4_H_ + +#include "aom_dsp/blend.h" +#include "aom_dsp/x86/synonyms.h" + +////////////////////////////////////////////////////////////////////////////// +// Common kernels +////////////////////////////////////////////////////////////////////////////// + +static INLINE __m128i blend_4(const uint8_t *src0, const uint8_t *src1, + const __m128i v_m0_w, const __m128i v_m1_w) { + const __m128i v_s0_b = xx_loadl_32(src0); + const __m128i v_s1_b = xx_loadl_32(src1); + const __m128i v_s0_w = _mm_cvtepu8_epi16(v_s0_b); + const __m128i v_s1_w = _mm_cvtepu8_epi16(v_s1_b); + + const __m128i v_p0_w = _mm_mullo_epi16(v_s0_w, v_m0_w); + const __m128i v_p1_w = _mm_mullo_epi16(v_s1_w, v_m1_w); + + const __m128i v_sum_w = _mm_add_epi16(v_p0_w, v_p1_w); + + const __m128i v_res_w = xx_roundn_epu16(v_sum_w, AOM_BLEND_A64_ROUND_BITS); + + return v_res_w; +} + +static INLINE __m128i blend_8(const uint8_t *src0, const uint8_t *src1, + const __m128i v_m0_w, const __m128i v_m1_w) { + const __m128i v_s0_b = xx_loadl_64(src0); + const __m128i v_s1_b = xx_loadl_64(src1); + const __m128i v_s0_w = _mm_cvtepu8_epi16(v_s0_b); + const __m128i v_s1_w = _mm_cvtepu8_epi16(v_s1_b); + + const __m128i v_p0_w = _mm_mullo_epi16(v_s0_w, v_m0_w); + const __m128i v_p1_w = _mm_mullo_epi16(v_s1_w, v_m1_w); + + const __m128i v_sum_w = _mm_add_epi16(v_p0_w, v_p1_w); + + const __m128i v_res_w = xx_roundn_epu16(v_sum_w, AOM_BLEND_A64_ROUND_BITS); + + return v_res_w; +} + +#if CONFIG_HIGHBITDEPTH +typedef __m128i (*blend_unit_fn)(const uint16_t *src0, const uint16_t *src1, + const __m128i v_m0_w, const __m128i v_m1_w); + +static INLINE __m128i blend_4_b10(const uint16_t *src0, const uint16_t *src1, + const __m128i v_m0_w, const __m128i v_m1_w) { + const __m128i v_s0_w = xx_loadl_64(src0); + const __m128i v_s1_w = xx_loadl_64(src1); + + const __m128i v_p0_w = _mm_mullo_epi16(v_s0_w, v_m0_w); + const __m128i v_p1_w = _mm_mullo_epi16(v_s1_w, v_m1_w); + + const __m128i v_sum_w = _mm_add_epi16(v_p0_w, v_p1_w); + + const __m128i v_res_w = xx_roundn_epu16(v_sum_w, AOM_BLEND_A64_ROUND_BITS); + + return v_res_w; +} + +static INLINE __m128i blend_8_b10(const uint16_t *src0, const uint16_t *src1, + const __m128i v_m0_w, const __m128i v_m1_w) { + const __m128i v_s0_w = xx_loadu_128(src0); + const __m128i v_s1_w = xx_loadu_128(src1); + + const __m128i v_p0_w = _mm_mullo_epi16(v_s0_w, v_m0_w); + const __m128i v_p1_w = _mm_mullo_epi16(v_s1_w, v_m1_w); + + const __m128i v_sum_w = _mm_add_epi16(v_p0_w, v_p1_w); + + const __m128i v_res_w = xx_roundn_epu16(v_sum_w, AOM_BLEND_A64_ROUND_BITS); + + return v_res_w; +} + +static INLINE __m128i blend_4_b12(const uint16_t *src0, const uint16_t *src1, + const __m128i v_m0_w, const __m128i v_m1_w) { + const __m128i v_s0_w = xx_loadl_64(src0); + const __m128i v_s1_w = xx_loadl_64(src1); + + // Interleave + const __m128i v_m01_w = _mm_unpacklo_epi16(v_m0_w, v_m1_w); + const __m128i v_s01_w = _mm_unpacklo_epi16(v_s0_w, v_s1_w); + + // Multiply-Add + const __m128i v_sum_d = _mm_madd_epi16(v_s01_w, v_m01_w); + + // Scale + const __m128i v_ssum_d = + _mm_srli_epi32(v_sum_d, AOM_BLEND_A64_ROUND_BITS - 1); + + // Pack + const __m128i v_pssum_d = _mm_packs_epi32(v_ssum_d, v_ssum_d); + + // Round + const __m128i v_res_w = xx_round_epu16(v_pssum_d); + + return v_res_w; +} + +static INLINE __m128i blend_8_b12(const uint16_t *src0, const uint16_t *src1, + const __m128i v_m0_w, const __m128i v_m1_w) { + const __m128i v_s0_w = xx_loadu_128(src0); + const __m128i v_s1_w = xx_loadu_128(src1); + + // Interleave + const __m128i v_m01l_w = _mm_unpacklo_epi16(v_m0_w, v_m1_w); + const __m128i v_m01h_w = _mm_unpackhi_epi16(v_m0_w, v_m1_w); + const __m128i v_s01l_w = _mm_unpacklo_epi16(v_s0_w, v_s1_w); + const __m128i v_s01h_w = _mm_unpackhi_epi16(v_s0_w, v_s1_w); + + // Multiply-Add + const __m128i v_suml_d = _mm_madd_epi16(v_s01l_w, v_m01l_w); + const __m128i v_sumh_d = _mm_madd_epi16(v_s01h_w, v_m01h_w); + + // Scale + const __m128i v_ssuml_d = + _mm_srli_epi32(v_suml_d, AOM_BLEND_A64_ROUND_BITS - 1); + const __m128i v_ssumh_d = + _mm_srli_epi32(v_sumh_d, AOM_BLEND_A64_ROUND_BITS - 1); + + // Pack + const __m128i v_pssum_d = _mm_packs_epi32(v_ssuml_d, v_ssumh_d); + + // Round + const __m128i v_res_w = xx_round_epu16(v_pssum_d); + + return v_res_w; +} +#endif // CONFIG_HIGHBITDEPTH + +#endif // AOM_DSP_X86_BLEND_SSE4_H_ diff --git a/third_party/aom/aom_dsp/x86/convolve.h b/third_party/aom/aom_dsp/x86/convolve.h new file mode 100644 index 000000000..8641164db --- /dev/null +++ b/third_party/aom/aom_dsp/x86/convolve.h @@ -0,0 +1,288 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ +#ifndef AOM_DSP_X86_CONVOLVE_H_ +#define AOM_DSP_X86_CONVOLVE_H_ + +#include <assert.h> + +#include "./aom_config.h" +#include "aom/aom_integer.h" +#include "aom_ports/mem.h" +#include "aom_dsp/aom_convolve.h" + +typedef void filter8_1dfunction(const uint8_t *src_ptr, ptrdiff_t src_pitch, + uint8_t *output_ptr, ptrdiff_t out_pitch, + uint32_t output_height, const int16_t *filter); + +#define FUN_CONV_1D(name, step_q4, filter, dir, src_start, avg, opt) \ + void aom_convolve8_##name##_##opt( \ + const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, \ + ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, \ + const int16_t *filter_y, int y_step_q4, int w, int h) { \ + (void)filter_x; \ + (void)x_step_q4; \ + (void)filter_y; \ + (void)y_step_q4; \ + assert((-128 <= filter[3]) && (filter[3] <= 127)); \ + assert(step_q4 == 16); \ + if (filter[0] | filter[1] | filter[2]) { \ + while (w >= 16) { \ + aom_filter_block1d16_##dir##8_##avg##opt(src_start, src_stride, dst, \ + dst_stride, h, filter); \ + src += 16; \ + dst += 16; \ + w -= 16; \ + } \ + while (w >= 8) { \ + aom_filter_block1d8_##dir##8_##avg##opt(src_start, src_stride, dst, \ + dst_stride, h, filter); \ + src += 8; \ + dst += 8; \ + w -= 8; \ + } \ + while (w >= 4) { \ + aom_filter_block1d4_##dir##8_##avg##opt(src_start, src_stride, dst, \ + dst_stride, h, filter); \ + src += 4; \ + dst += 4; \ + w -= 4; \ + } \ + } else { \ + while (w >= 16) { \ + aom_filter_block1d16_##dir##2_##avg##opt(src, src_stride, dst, \ + dst_stride, h, filter); \ + src += 16; \ + dst += 16; \ + w -= 16; \ + } \ + while (w >= 8) { \ + aom_filter_block1d8_##dir##2_##avg##opt(src, src_stride, dst, \ + dst_stride, h, filter); \ + src += 8; \ + dst += 8; \ + w -= 8; \ + } \ + while (w >= 4) { \ + aom_filter_block1d4_##dir##2_##avg##opt(src, src_stride, dst, \ + dst_stride, h, filter); \ + src += 4; \ + dst += 4; \ + w -= 4; \ + } \ + } \ + if (w) { \ + aom_convolve8_##name##_c(src, src_stride, dst, dst_stride, filter_x, \ + x_step_q4, filter_y, y_step_q4, w, h); \ + } \ + } + +#define FUN_CONV_2D(avg, opt) \ + void aom_convolve8_##avg##opt( \ + const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, \ + ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, \ + const int16_t *filter_y, int y_step_q4, int w, int h) { \ + assert((-128 <= filter_x[3]) && (filter_x[3] <= 127)); \ + assert((-128 <= filter_y[3]) && (filter_y[3] <= 127)); \ + assert(w <= MAX_SB_SIZE); \ + assert(h <= MAX_SB_SIZE); \ + assert(x_step_q4 == 16); \ + assert(y_step_q4 == 16); \ + if (filter_x[0] || filter_x[1] || filter_x[2] || filter_y[0] || \ + filter_y[1] || filter_y[2]) { \ + DECLARE_ALIGNED(16, uint8_t, fdata2[MAX_SB_SIZE * (MAX_SB_SIZE + 7)]); \ + aom_convolve8_horiz_##opt(src - 3 * src_stride, src_stride, fdata2, \ + MAX_SB_SIZE, filter_x, x_step_q4, filter_y, \ + y_step_q4, w, h + 7); \ + aom_convolve8_##avg##vert_##opt(fdata2 + 3 * MAX_SB_SIZE, MAX_SB_SIZE, \ + dst, dst_stride, filter_x, x_step_q4, \ + filter_y, y_step_q4, w, h); \ + } else { \ + DECLARE_ALIGNED(16, uint8_t, fdata2[MAX_SB_SIZE * (MAX_SB_SIZE + 1)]); \ + aom_convolve8_horiz_##opt(src, src_stride, fdata2, MAX_SB_SIZE, \ + filter_x, x_step_q4, filter_y, y_step_q4, w, \ + h + 1); \ + aom_convolve8_##avg##vert_##opt(fdata2, MAX_SB_SIZE, dst, dst_stride, \ + filter_x, x_step_q4, filter_y, \ + y_step_q4, w, h); \ + } \ + } + +#if CONFIG_LOOP_RESTORATION +// convolve_add_src is only used by the Wiener filter, which will never +// end up calling the bilinear functions (it uses a symmetric filter, so +// the possible numbers of taps are 1,3,5,7) +#define FUN_CONV_1D_NO_BILINEAR(name, step_q4, filter, dir, src_start, avg, \ + opt) \ + void aom_convolve8_##name##_##opt( \ + const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, \ + ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, \ + const int16_t *filter_y, int y_step_q4, int w, int h) { \ + (void)filter_x; \ + (void)x_step_q4; \ + (void)filter_y; \ + (void)y_step_q4; \ + assert((-128 <= filter[3]) && (filter[3] <= 127)); \ + assert(step_q4 == 16); \ + while (w >= 16) { \ + aom_filter_block1d16_##dir##8_##avg##opt(src_start, src_stride, dst, \ + dst_stride, h, filter); \ + src += 16; \ + dst += 16; \ + w -= 16; \ + } \ + while (w >= 8) { \ + aom_filter_block1d8_##dir##8_##avg##opt(src_start, src_stride, dst, \ + dst_stride, h, filter); \ + src += 8; \ + dst += 8; \ + w -= 8; \ + } \ + while (w >= 4) { \ + aom_filter_block1d4_##dir##8_##avg##opt(src_start, src_stride, dst, \ + dst_stride, h, filter); \ + src += 4; \ + dst += 4; \ + w -= 4; \ + } \ + if (w) { \ + aom_convolve8_##name##_c(src, src_stride, dst, dst_stride, filter_x, \ + x_step_q4, filter_y, y_step_q4, w, h); \ + } \ + } + +#define FUN_CONV_2D_NO_BILINEAR(type, htype, opt) \ + void aom_convolve8_##type##opt( \ + const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, \ + ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, \ + const int16_t *filter_y, int y_step_q4, int w, int h) { \ + DECLARE_ALIGNED(16, uint8_t, fdata2[MAX_SB_SIZE * (MAX_SB_SIZE + 7)]); \ + assert((-128 <= filter_x[3]) && (filter_x[3] <= 127)); \ + assert((-128 <= filter_y[3]) && (filter_y[3] <= 127)); \ + assert(w <= MAX_SB_SIZE); \ + assert(h <= MAX_SB_SIZE); \ + assert(x_step_q4 == 16); \ + assert(y_step_q4 == 16); \ + aom_convolve8_##htype##horiz_##opt( \ + src - 3 * src_stride, src_stride, fdata2, MAX_SB_SIZE, filter_x, \ + x_step_q4, filter_y, y_step_q4, w, h + 7); \ + aom_convolve8_##type##vert_##opt(fdata2 + 3 * MAX_SB_SIZE, MAX_SB_SIZE, \ + dst, dst_stride, filter_x, x_step_q4, \ + filter_y, y_step_q4, w, h); \ + } +#endif + +#if CONFIG_HIGHBITDEPTH +typedef void highbd_filter8_1dfunction(const uint16_t *src_ptr, + const ptrdiff_t src_pitch, + uint16_t *output_ptr, + ptrdiff_t out_pitch, + unsigned int output_height, + const int16_t *filter, int bd); + +#define HIGH_FUN_CONV_1D(name, step_q4, filter, dir, src_start, avg, opt) \ + void aom_highbd_convolve8_##name##_##opt( \ + const uint8_t *src8, ptrdiff_t src_stride, uint8_t *dst8, \ + ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, \ + const int16_t *filter_y, int y_step_q4, int w, int h, int bd) { \ + uint16_t *src = CONVERT_TO_SHORTPTR(src8); \ + uint16_t *dst = CONVERT_TO_SHORTPTR(dst8); \ + if (step_q4 == 16 && filter[3] != 128) { \ + if (filter[0] | filter[1] | filter[2]) { \ + while (w >= 16) { \ + aom_highbd_filter_block1d16_##dir##8_##avg##opt( \ + src_start, src_stride, dst, dst_stride, h, filter, bd); \ + src += 16; \ + dst += 16; \ + w -= 16; \ + } \ + while (w >= 8) { \ + aom_highbd_filter_block1d8_##dir##8_##avg##opt( \ + src_start, src_stride, dst, dst_stride, h, filter, bd); \ + src += 8; \ + dst += 8; \ + w -= 8; \ + } \ + while (w >= 4) { \ + aom_highbd_filter_block1d4_##dir##8_##avg##opt( \ + src_start, src_stride, dst, dst_stride, h, filter, bd); \ + src += 4; \ + dst += 4; \ + w -= 4; \ + } \ + } else { \ + while (w >= 16) { \ + aom_highbd_filter_block1d16_##dir##2_##avg##opt( \ + src, src_stride, dst, dst_stride, h, filter, bd); \ + src += 16; \ + dst += 16; \ + w -= 16; \ + } \ + while (w >= 8) { \ + aom_highbd_filter_block1d8_##dir##2_##avg##opt( \ + src, src_stride, dst, dst_stride, h, filter, bd); \ + src += 8; \ + dst += 8; \ + w -= 8; \ + } \ + while (w >= 4) { \ + aom_highbd_filter_block1d4_##dir##2_##avg##opt( \ + src, src_stride, dst, dst_stride, h, filter, bd); \ + src += 4; \ + dst += 4; \ + w -= 4; \ + } \ + } \ + } \ + if (w) { \ + aom_highbd_convolve8_##name##_c( \ + CONVERT_TO_BYTEPTR(src), src_stride, CONVERT_TO_BYTEPTR(dst), \ + dst_stride, filter_x, x_step_q4, filter_y, y_step_q4, w, h, bd); \ + } \ + } + +#define HIGH_FUN_CONV_2D(avg, opt) \ + void aom_highbd_convolve8_##avg##opt( \ + const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, \ + ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, \ + const int16_t *filter_y, int y_step_q4, int w, int h, int bd) { \ + assert(w <= MAX_SB_SIZE); \ + assert(h <= MAX_SB_SIZE); \ + if (x_step_q4 == 16 && y_step_q4 == 16) { \ + if (filter_x[0] || filter_x[1] || filter_x[2] || filter_x[3] == 128 || \ + filter_y[0] || filter_y[1] || filter_y[2] || filter_y[3] == 128) { \ + DECLARE_ALIGNED(16, uint16_t, \ + fdata2[MAX_SB_SIZE * (MAX_SB_SIZE + 7)]); \ + aom_highbd_convolve8_horiz_##opt(src - 3 * src_stride, src_stride, \ + CONVERT_TO_BYTEPTR(fdata2), \ + MAX_SB_SIZE, filter_x, x_step_q4, \ + filter_y, y_step_q4, w, h + 7, bd); \ + aom_highbd_convolve8_##avg##vert_##opt( \ + CONVERT_TO_BYTEPTR(fdata2) + 3 * MAX_SB_SIZE, MAX_SB_SIZE, dst, \ + dst_stride, filter_x, x_step_q4, filter_y, y_step_q4, w, h, bd); \ + } else { \ + DECLARE_ALIGNED(16, uint16_t, \ + fdata2[MAX_SB_SIZE * (MAX_SB_SIZE + 1)]); \ + aom_highbd_convolve8_horiz_##opt( \ + src, src_stride, CONVERT_TO_BYTEPTR(fdata2), MAX_SB_SIZE, \ + filter_x, x_step_q4, filter_y, y_step_q4, w, h + 1, bd); \ + aom_highbd_convolve8_##avg##vert_##opt( \ + CONVERT_TO_BYTEPTR(fdata2), MAX_SB_SIZE, dst, dst_stride, \ + filter_x, x_step_q4, filter_y, y_step_q4, w, h, bd); \ + } \ + } else { \ + aom_highbd_convolve8_##avg##c(src, src_stride, dst, dst_stride, \ + filter_x, x_step_q4, filter_y, y_step_q4, \ + w, h, bd); \ + } \ + } +#endif // CONFIG_HIGHBITDEPTH + +#endif // AOM_DSP_X86_CONVOLVE_H_ diff --git a/third_party/aom/aom_dsp/x86/fwd_dct32_8cols_sse2.c b/third_party/aom/aom_dsp/x86/fwd_dct32_8cols_sse2.c new file mode 100644 index 000000000..b8ec08de7 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/fwd_dct32_8cols_sse2.c @@ -0,0 +1,862 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include <emmintrin.h> // SSE2 + +#include "aom_dsp/fwd_txfm.h" +#include "aom_dsp/txfm_common.h" +#include "aom_dsp/x86/txfm_common_sse2.h" + +// Apply a 32-element IDCT to 8 columns. This does not do any transposition +// of its output - the caller is expected to do that. +// The input buffers are the top and bottom halves of an 8x32 block. +void fdct32_8col(__m128i *in0, __m128i *in1) { + // Constants + // When we use them, in one case, they are all the same. In all others + // it's a pair of them that we need to repeat four times. This is done + // by constructing the 32 bit constant corresponding to that pair. + const __m128i k__cospi_p16_p16 = _mm_set1_epi16((int16_t)cospi_16_64); + const __m128i k__cospi_p16_m16 = pair_set_epi16(+cospi_16_64, -cospi_16_64); + const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64, cospi_24_64); + const __m128i k__cospi_m24_m08 = pair_set_epi16(-cospi_24_64, -cospi_8_64); + const __m128i k__cospi_p24_p08 = pair_set_epi16(+cospi_24_64, cospi_8_64); + const __m128i k__cospi_p12_p20 = pair_set_epi16(+cospi_12_64, cospi_20_64); + const __m128i k__cospi_m20_p12 = pair_set_epi16(-cospi_20_64, cospi_12_64); + const __m128i k__cospi_m04_p28 = pair_set_epi16(-cospi_4_64, cospi_28_64); + const __m128i k__cospi_p28_p04 = pair_set_epi16(+cospi_28_64, cospi_4_64); + const __m128i k__cospi_m28_m04 = pair_set_epi16(-cospi_28_64, -cospi_4_64); + const __m128i k__cospi_m12_m20 = pair_set_epi16(-cospi_12_64, -cospi_20_64); + const __m128i k__cospi_p30_p02 = pair_set_epi16(+cospi_30_64, cospi_2_64); + const __m128i k__cospi_p14_p18 = pair_set_epi16(+cospi_14_64, cospi_18_64); + const __m128i k__cospi_p22_p10 = pair_set_epi16(+cospi_22_64, cospi_10_64); + const __m128i k__cospi_p06_p26 = pair_set_epi16(+cospi_6_64, cospi_26_64); + const __m128i k__cospi_m26_p06 = pair_set_epi16(-cospi_26_64, cospi_6_64); + const __m128i k__cospi_m10_p22 = pair_set_epi16(-cospi_10_64, cospi_22_64); + const __m128i k__cospi_m18_p14 = pair_set_epi16(-cospi_18_64, cospi_14_64); + const __m128i k__cospi_m02_p30 = pair_set_epi16(-cospi_2_64, cospi_30_64); + const __m128i k__cospi_p31_p01 = pair_set_epi16(+cospi_31_64, cospi_1_64); + const __m128i k__cospi_p15_p17 = pair_set_epi16(+cospi_15_64, cospi_17_64); + const __m128i k__cospi_p23_p09 = pair_set_epi16(+cospi_23_64, cospi_9_64); + const __m128i k__cospi_p07_p25 = pair_set_epi16(+cospi_7_64, cospi_25_64); + const __m128i k__cospi_m25_p07 = pair_set_epi16(-cospi_25_64, cospi_7_64); + const __m128i k__cospi_m09_p23 = pair_set_epi16(-cospi_9_64, cospi_23_64); + const __m128i k__cospi_m17_p15 = pair_set_epi16(-cospi_17_64, cospi_15_64); + const __m128i k__cospi_m01_p31 = pair_set_epi16(-cospi_1_64, cospi_31_64); + const __m128i k__cospi_p27_p05 = pair_set_epi16(+cospi_27_64, cospi_5_64); + const __m128i k__cospi_p11_p21 = pair_set_epi16(+cospi_11_64, cospi_21_64); + const __m128i k__cospi_p19_p13 = pair_set_epi16(+cospi_19_64, cospi_13_64); + const __m128i k__cospi_p03_p29 = pair_set_epi16(+cospi_3_64, cospi_29_64); + const __m128i k__cospi_m29_p03 = pair_set_epi16(-cospi_29_64, cospi_3_64); + const __m128i k__cospi_m13_p19 = pair_set_epi16(-cospi_13_64, cospi_19_64); + const __m128i k__cospi_m21_p11 = pair_set_epi16(-cospi_21_64, cospi_11_64); + const __m128i k__cospi_m05_p27 = pair_set_epi16(-cospi_5_64, cospi_27_64); + const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING); + + __m128i step1[32]; + __m128i step2[32]; + __m128i step3[32]; + __m128i out[32]; + // Stage 1 + { + const __m128i *ina = in0; + const __m128i *inb = in1 + 15; + __m128i *step1a = &step1[0]; + __m128i *step1b = &step1[31]; + const __m128i ina0 = _mm_loadu_si128((const __m128i *)(ina)); + const __m128i ina1 = _mm_loadu_si128((const __m128i *)(ina + 1)); + const __m128i ina2 = _mm_loadu_si128((const __m128i *)(ina + 2)); + const __m128i ina3 = _mm_loadu_si128((const __m128i *)(ina + 3)); + const __m128i inb3 = _mm_loadu_si128((const __m128i *)(inb - 3)); + const __m128i inb2 = _mm_loadu_si128((const __m128i *)(inb - 2)); + const __m128i inb1 = _mm_loadu_si128((const __m128i *)(inb - 1)); + const __m128i inb0 = _mm_loadu_si128((const __m128i *)(inb)); + step1a[0] = _mm_add_epi16(ina0, inb0); + step1a[1] = _mm_add_epi16(ina1, inb1); + step1a[2] = _mm_add_epi16(ina2, inb2); + step1a[3] = _mm_add_epi16(ina3, inb3); + step1b[-3] = _mm_sub_epi16(ina3, inb3); + step1b[-2] = _mm_sub_epi16(ina2, inb2); + step1b[-1] = _mm_sub_epi16(ina1, inb1); + step1b[-0] = _mm_sub_epi16(ina0, inb0); + } + { + const __m128i *ina = in0 + 4; + const __m128i *inb = in1 + 11; + __m128i *step1a = &step1[4]; + __m128i *step1b = &step1[27]; + const __m128i ina0 = _mm_loadu_si128((const __m128i *)(ina)); + const __m128i ina1 = _mm_loadu_si128((const __m128i *)(ina + 1)); + const __m128i ina2 = _mm_loadu_si128((const __m128i *)(ina + 2)); + const __m128i ina3 = _mm_loadu_si128((const __m128i *)(ina + 3)); + const __m128i inb3 = _mm_loadu_si128((const __m128i *)(inb - 3)); + const __m128i inb2 = _mm_loadu_si128((const __m128i *)(inb - 2)); + const __m128i inb1 = _mm_loadu_si128((const __m128i *)(inb - 1)); + const __m128i inb0 = _mm_loadu_si128((const __m128i *)(inb)); + step1a[0] = _mm_add_epi16(ina0, inb0); + step1a[1] = _mm_add_epi16(ina1, inb1); + step1a[2] = _mm_add_epi16(ina2, inb2); + step1a[3] = _mm_add_epi16(ina3, inb3); + step1b[-3] = _mm_sub_epi16(ina3, inb3); + step1b[-2] = _mm_sub_epi16(ina2, inb2); + step1b[-1] = _mm_sub_epi16(ina1, inb1); + step1b[-0] = _mm_sub_epi16(ina0, inb0); + } + { + const __m128i *ina = in0 + 8; + const __m128i *inb = in1 + 7; + __m128i *step1a = &step1[8]; + __m128i *step1b = &step1[23]; + const __m128i ina0 = _mm_loadu_si128((const __m128i *)(ina)); + const __m128i ina1 = _mm_loadu_si128((const __m128i *)(ina + 1)); + const __m128i ina2 = _mm_loadu_si128((const __m128i *)(ina + 2)); + const __m128i ina3 = _mm_loadu_si128((const __m128i *)(ina + 3)); + const __m128i inb3 = _mm_loadu_si128((const __m128i *)(inb - 3)); + const __m128i inb2 = _mm_loadu_si128((const __m128i *)(inb - 2)); + const __m128i inb1 = _mm_loadu_si128((const __m128i *)(inb - 1)); + const __m128i inb0 = _mm_loadu_si128((const __m128i *)(inb)); + step1a[0] = _mm_add_epi16(ina0, inb0); + step1a[1] = _mm_add_epi16(ina1, inb1); + step1a[2] = _mm_add_epi16(ina2, inb2); + step1a[3] = _mm_add_epi16(ina3, inb3); + step1b[-3] = _mm_sub_epi16(ina3, inb3); + step1b[-2] = _mm_sub_epi16(ina2, inb2); + step1b[-1] = _mm_sub_epi16(ina1, inb1); + step1b[-0] = _mm_sub_epi16(ina0, inb0); + } + { + const __m128i *ina = in0 + 12; + const __m128i *inb = in1 + 3; + __m128i *step1a = &step1[12]; + __m128i *step1b = &step1[19]; + const __m128i ina0 = _mm_loadu_si128((const __m128i *)(ina)); + const __m128i ina1 = _mm_loadu_si128((const __m128i *)(ina + 1)); + const __m128i ina2 = _mm_loadu_si128((const __m128i *)(ina + 2)); + const __m128i ina3 = _mm_loadu_si128((const __m128i *)(ina + 3)); + const __m128i inb3 = _mm_loadu_si128((const __m128i *)(inb - 3)); + const __m128i inb2 = _mm_loadu_si128((const __m128i *)(inb - 2)); + const __m128i inb1 = _mm_loadu_si128((const __m128i *)(inb - 1)); + const __m128i inb0 = _mm_loadu_si128((const __m128i *)(inb)); + step1a[0] = _mm_add_epi16(ina0, inb0); + step1a[1] = _mm_add_epi16(ina1, inb1); + step1a[2] = _mm_add_epi16(ina2, inb2); + step1a[3] = _mm_add_epi16(ina3, inb3); + step1b[-3] = _mm_sub_epi16(ina3, inb3); + step1b[-2] = _mm_sub_epi16(ina2, inb2); + step1b[-1] = _mm_sub_epi16(ina1, inb1); + step1b[-0] = _mm_sub_epi16(ina0, inb0); + } + // Stage 2 + { + step2[0] = _mm_add_epi16(step1[0], step1[15]); + step2[1] = _mm_add_epi16(step1[1], step1[14]); + step2[2] = _mm_add_epi16(step1[2], step1[13]); + step2[3] = _mm_add_epi16(step1[3], step1[12]); + step2[4] = _mm_add_epi16(step1[4], step1[11]); + step2[5] = _mm_add_epi16(step1[5], step1[10]); + step2[6] = _mm_add_epi16(step1[6], step1[9]); + step2[7] = _mm_add_epi16(step1[7], step1[8]); + step2[8] = _mm_sub_epi16(step1[7], step1[8]); + step2[9] = _mm_sub_epi16(step1[6], step1[9]); + step2[10] = _mm_sub_epi16(step1[5], step1[10]); + step2[11] = _mm_sub_epi16(step1[4], step1[11]); + step2[12] = _mm_sub_epi16(step1[3], step1[12]); + step2[13] = _mm_sub_epi16(step1[2], step1[13]); + step2[14] = _mm_sub_epi16(step1[1], step1[14]); + step2[15] = _mm_sub_epi16(step1[0], step1[15]); + } + { + const __m128i s2_20_0 = _mm_unpacklo_epi16(step1[27], step1[20]); + const __m128i s2_20_1 = _mm_unpackhi_epi16(step1[27], step1[20]); + const __m128i s2_21_0 = _mm_unpacklo_epi16(step1[26], step1[21]); + const __m128i s2_21_1 = _mm_unpackhi_epi16(step1[26], step1[21]); + const __m128i s2_22_0 = _mm_unpacklo_epi16(step1[25], step1[22]); + const __m128i s2_22_1 = _mm_unpackhi_epi16(step1[25], step1[22]); + const __m128i s2_23_0 = _mm_unpacklo_epi16(step1[24], step1[23]); + const __m128i s2_23_1 = _mm_unpackhi_epi16(step1[24], step1[23]); + const __m128i s2_20_2 = _mm_madd_epi16(s2_20_0, k__cospi_p16_m16); + const __m128i s2_20_3 = _mm_madd_epi16(s2_20_1, k__cospi_p16_m16); + const __m128i s2_21_2 = _mm_madd_epi16(s2_21_0, k__cospi_p16_m16); + const __m128i s2_21_3 = _mm_madd_epi16(s2_21_1, k__cospi_p16_m16); + const __m128i s2_22_2 = _mm_madd_epi16(s2_22_0, k__cospi_p16_m16); + const __m128i s2_22_3 = _mm_madd_epi16(s2_22_1, k__cospi_p16_m16); + const __m128i s2_23_2 = _mm_madd_epi16(s2_23_0, k__cospi_p16_m16); + const __m128i s2_23_3 = _mm_madd_epi16(s2_23_1, k__cospi_p16_m16); + const __m128i s2_24_2 = _mm_madd_epi16(s2_23_0, k__cospi_p16_p16); + const __m128i s2_24_3 = _mm_madd_epi16(s2_23_1, k__cospi_p16_p16); + const __m128i s2_25_2 = _mm_madd_epi16(s2_22_0, k__cospi_p16_p16); + const __m128i s2_25_3 = _mm_madd_epi16(s2_22_1, k__cospi_p16_p16); + const __m128i s2_26_2 = _mm_madd_epi16(s2_21_0, k__cospi_p16_p16); + const __m128i s2_26_3 = _mm_madd_epi16(s2_21_1, k__cospi_p16_p16); + const __m128i s2_27_2 = _mm_madd_epi16(s2_20_0, k__cospi_p16_p16); + const __m128i s2_27_3 = _mm_madd_epi16(s2_20_1, k__cospi_p16_p16); + // dct_const_round_shift + const __m128i s2_20_4 = _mm_add_epi32(s2_20_2, k__DCT_CONST_ROUNDING); + const __m128i s2_20_5 = _mm_add_epi32(s2_20_3, k__DCT_CONST_ROUNDING); + const __m128i s2_21_4 = _mm_add_epi32(s2_21_2, k__DCT_CONST_ROUNDING); + const __m128i s2_21_5 = _mm_add_epi32(s2_21_3, k__DCT_CONST_ROUNDING); + const __m128i s2_22_4 = _mm_add_epi32(s2_22_2, k__DCT_CONST_ROUNDING); + const __m128i s2_22_5 = _mm_add_epi32(s2_22_3, k__DCT_CONST_ROUNDING); + const __m128i s2_23_4 = _mm_add_epi32(s2_23_2, k__DCT_CONST_ROUNDING); + const __m128i s2_23_5 = _mm_add_epi32(s2_23_3, k__DCT_CONST_ROUNDING); + const __m128i s2_24_4 = _mm_add_epi32(s2_24_2, k__DCT_CONST_ROUNDING); + const __m128i s2_24_5 = _mm_add_epi32(s2_24_3, k__DCT_CONST_ROUNDING); + const __m128i s2_25_4 = _mm_add_epi32(s2_25_2, k__DCT_CONST_ROUNDING); + const __m128i s2_25_5 = _mm_add_epi32(s2_25_3, k__DCT_CONST_ROUNDING); + const __m128i s2_26_4 = _mm_add_epi32(s2_26_2, k__DCT_CONST_ROUNDING); + const __m128i s2_26_5 = _mm_add_epi32(s2_26_3, k__DCT_CONST_ROUNDING); + const __m128i s2_27_4 = _mm_add_epi32(s2_27_2, k__DCT_CONST_ROUNDING); + const __m128i s2_27_5 = _mm_add_epi32(s2_27_3, k__DCT_CONST_ROUNDING); + const __m128i s2_20_6 = _mm_srai_epi32(s2_20_4, DCT_CONST_BITS); + const __m128i s2_20_7 = _mm_srai_epi32(s2_20_5, DCT_CONST_BITS); + const __m128i s2_21_6 = _mm_srai_epi32(s2_21_4, DCT_CONST_BITS); + const __m128i s2_21_7 = _mm_srai_epi32(s2_21_5, DCT_CONST_BITS); + const __m128i s2_22_6 = _mm_srai_epi32(s2_22_4, DCT_CONST_BITS); + const __m128i s2_22_7 = _mm_srai_epi32(s2_22_5, DCT_CONST_BITS); + const __m128i s2_23_6 = _mm_srai_epi32(s2_23_4, DCT_CONST_BITS); + const __m128i s2_23_7 = _mm_srai_epi32(s2_23_5, DCT_CONST_BITS); + const __m128i s2_24_6 = _mm_srai_epi32(s2_24_4, DCT_CONST_BITS); + const __m128i s2_24_7 = _mm_srai_epi32(s2_24_5, DCT_CONST_BITS); + const __m128i s2_25_6 = _mm_srai_epi32(s2_25_4, DCT_CONST_BITS); + const __m128i s2_25_7 = _mm_srai_epi32(s2_25_5, DCT_CONST_BITS); + const __m128i s2_26_6 = _mm_srai_epi32(s2_26_4, DCT_CONST_BITS); + const __m128i s2_26_7 = _mm_srai_epi32(s2_26_5, DCT_CONST_BITS); + const __m128i s2_27_6 = _mm_srai_epi32(s2_27_4, DCT_CONST_BITS); + const __m128i s2_27_7 = _mm_srai_epi32(s2_27_5, DCT_CONST_BITS); + // Combine + step2[20] = _mm_packs_epi32(s2_20_6, s2_20_7); + step2[21] = _mm_packs_epi32(s2_21_6, s2_21_7); + step2[22] = _mm_packs_epi32(s2_22_6, s2_22_7); + step2[23] = _mm_packs_epi32(s2_23_6, s2_23_7); + step2[24] = _mm_packs_epi32(s2_24_6, s2_24_7); + step2[25] = _mm_packs_epi32(s2_25_6, s2_25_7); + step2[26] = _mm_packs_epi32(s2_26_6, s2_26_7); + step2[27] = _mm_packs_epi32(s2_27_6, s2_27_7); + } + // Stage 3 + { + step3[0] = _mm_add_epi16(step2[(8 - 1)], step2[0]); + step3[1] = _mm_add_epi16(step2[(8 - 2)], step2[1]); + step3[2] = _mm_add_epi16(step2[(8 - 3)], step2[2]); + step3[3] = _mm_add_epi16(step2[(8 - 4)], step2[3]); + step3[4] = _mm_sub_epi16(step2[(8 - 5)], step2[4]); + step3[5] = _mm_sub_epi16(step2[(8 - 6)], step2[5]); + step3[6] = _mm_sub_epi16(step2[(8 - 7)], step2[6]); + step3[7] = _mm_sub_epi16(step2[(8 - 8)], step2[7]); + } + { + const __m128i s3_10_0 = _mm_unpacklo_epi16(step2[13], step2[10]); + const __m128i s3_10_1 = _mm_unpackhi_epi16(step2[13], step2[10]); + const __m128i s3_11_0 = _mm_unpacklo_epi16(step2[12], step2[11]); + const __m128i s3_11_1 = _mm_unpackhi_epi16(step2[12], step2[11]); + const __m128i s3_10_2 = _mm_madd_epi16(s3_10_0, k__cospi_p16_m16); + const __m128i s3_10_3 = _mm_madd_epi16(s3_10_1, k__cospi_p16_m16); + const __m128i s3_11_2 = _mm_madd_epi16(s3_11_0, k__cospi_p16_m16); + const __m128i s3_11_3 = _mm_madd_epi16(s3_11_1, k__cospi_p16_m16); + const __m128i s3_12_2 = _mm_madd_epi16(s3_11_0, k__cospi_p16_p16); + const __m128i s3_12_3 = _mm_madd_epi16(s3_11_1, k__cospi_p16_p16); + const __m128i s3_13_2 = _mm_madd_epi16(s3_10_0, k__cospi_p16_p16); + const __m128i s3_13_3 = _mm_madd_epi16(s3_10_1, k__cospi_p16_p16); + // dct_const_round_shift + const __m128i s3_10_4 = _mm_add_epi32(s3_10_2, k__DCT_CONST_ROUNDING); + const __m128i s3_10_5 = _mm_add_epi32(s3_10_3, k__DCT_CONST_ROUNDING); + const __m128i s3_11_4 = _mm_add_epi32(s3_11_2, k__DCT_CONST_ROUNDING); + const __m128i s3_11_5 = _mm_add_epi32(s3_11_3, k__DCT_CONST_ROUNDING); + const __m128i s3_12_4 = _mm_add_epi32(s3_12_2, k__DCT_CONST_ROUNDING); + const __m128i s3_12_5 = _mm_add_epi32(s3_12_3, k__DCT_CONST_ROUNDING); + const __m128i s3_13_4 = _mm_add_epi32(s3_13_2, k__DCT_CONST_ROUNDING); + const __m128i s3_13_5 = _mm_add_epi32(s3_13_3, k__DCT_CONST_ROUNDING); + const __m128i s3_10_6 = _mm_srai_epi32(s3_10_4, DCT_CONST_BITS); + const __m128i s3_10_7 = _mm_srai_epi32(s3_10_5, DCT_CONST_BITS); + const __m128i s3_11_6 = _mm_srai_epi32(s3_11_4, DCT_CONST_BITS); + const __m128i s3_11_7 = _mm_srai_epi32(s3_11_5, DCT_CONST_BITS); + const __m128i s3_12_6 = _mm_srai_epi32(s3_12_4, DCT_CONST_BITS); + const __m128i s3_12_7 = _mm_srai_epi32(s3_12_5, DCT_CONST_BITS); + const __m128i s3_13_6 = _mm_srai_epi32(s3_13_4, DCT_CONST_BITS); + const __m128i s3_13_7 = _mm_srai_epi32(s3_13_5, DCT_CONST_BITS); + // Combine + step3[10] = _mm_packs_epi32(s3_10_6, s3_10_7); + step3[11] = _mm_packs_epi32(s3_11_6, s3_11_7); + step3[12] = _mm_packs_epi32(s3_12_6, s3_12_7); + step3[13] = _mm_packs_epi32(s3_13_6, s3_13_7); + } + { + step3[16] = _mm_add_epi16(step2[23], step1[16]); + step3[17] = _mm_add_epi16(step2[22], step1[17]); + step3[18] = _mm_add_epi16(step2[21], step1[18]); + step3[19] = _mm_add_epi16(step2[20], step1[19]); + step3[20] = _mm_sub_epi16(step1[19], step2[20]); + step3[21] = _mm_sub_epi16(step1[18], step2[21]); + step3[22] = _mm_sub_epi16(step1[17], step2[22]); + step3[23] = _mm_sub_epi16(step1[16], step2[23]); + step3[24] = _mm_sub_epi16(step1[31], step2[24]); + step3[25] = _mm_sub_epi16(step1[30], step2[25]); + step3[26] = _mm_sub_epi16(step1[29], step2[26]); + step3[27] = _mm_sub_epi16(step1[28], step2[27]); + step3[28] = _mm_add_epi16(step2[27], step1[28]); + step3[29] = _mm_add_epi16(step2[26], step1[29]); + step3[30] = _mm_add_epi16(step2[25], step1[30]); + step3[31] = _mm_add_epi16(step2[24], step1[31]); + } + + // Stage 4 + { + step1[0] = _mm_add_epi16(step3[3], step3[0]); + step1[1] = _mm_add_epi16(step3[2], step3[1]); + step1[2] = _mm_sub_epi16(step3[1], step3[2]); + step1[3] = _mm_sub_epi16(step3[0], step3[3]); + step1[8] = _mm_add_epi16(step3[11], step2[8]); + step1[9] = _mm_add_epi16(step3[10], step2[9]); + step1[10] = _mm_sub_epi16(step2[9], step3[10]); + step1[11] = _mm_sub_epi16(step2[8], step3[11]); + step1[12] = _mm_sub_epi16(step2[15], step3[12]); + step1[13] = _mm_sub_epi16(step2[14], step3[13]); + step1[14] = _mm_add_epi16(step3[13], step2[14]); + step1[15] = _mm_add_epi16(step3[12], step2[15]); + } + { + const __m128i s1_05_0 = _mm_unpacklo_epi16(step3[6], step3[5]); + const __m128i s1_05_1 = _mm_unpackhi_epi16(step3[6], step3[5]); + const __m128i s1_05_2 = _mm_madd_epi16(s1_05_0, k__cospi_p16_m16); + const __m128i s1_05_3 = _mm_madd_epi16(s1_05_1, k__cospi_p16_m16); + const __m128i s1_06_2 = _mm_madd_epi16(s1_05_0, k__cospi_p16_p16); + const __m128i s1_06_3 = _mm_madd_epi16(s1_05_1, k__cospi_p16_p16); + // dct_const_round_shift + const __m128i s1_05_4 = _mm_add_epi32(s1_05_2, k__DCT_CONST_ROUNDING); + const __m128i s1_05_5 = _mm_add_epi32(s1_05_3, k__DCT_CONST_ROUNDING); + const __m128i s1_06_4 = _mm_add_epi32(s1_06_2, k__DCT_CONST_ROUNDING); + const __m128i s1_06_5 = _mm_add_epi32(s1_06_3, k__DCT_CONST_ROUNDING); + const __m128i s1_05_6 = _mm_srai_epi32(s1_05_4, DCT_CONST_BITS); + const __m128i s1_05_7 = _mm_srai_epi32(s1_05_5, DCT_CONST_BITS); + const __m128i s1_06_6 = _mm_srai_epi32(s1_06_4, DCT_CONST_BITS); + const __m128i s1_06_7 = _mm_srai_epi32(s1_06_5, DCT_CONST_BITS); + // Combine + step1[5] = _mm_packs_epi32(s1_05_6, s1_05_7); + step1[6] = _mm_packs_epi32(s1_06_6, s1_06_7); + } + { + const __m128i s1_18_0 = _mm_unpacklo_epi16(step3[18], step3[29]); + const __m128i s1_18_1 = _mm_unpackhi_epi16(step3[18], step3[29]); + const __m128i s1_19_0 = _mm_unpacklo_epi16(step3[19], step3[28]); + const __m128i s1_19_1 = _mm_unpackhi_epi16(step3[19], step3[28]); + const __m128i s1_20_0 = _mm_unpacklo_epi16(step3[20], step3[27]); + const __m128i s1_20_1 = _mm_unpackhi_epi16(step3[20], step3[27]); + const __m128i s1_21_0 = _mm_unpacklo_epi16(step3[21], step3[26]); + const __m128i s1_21_1 = _mm_unpackhi_epi16(step3[21], step3[26]); + const __m128i s1_18_2 = _mm_madd_epi16(s1_18_0, k__cospi_m08_p24); + const __m128i s1_18_3 = _mm_madd_epi16(s1_18_1, k__cospi_m08_p24); + const __m128i s1_19_2 = _mm_madd_epi16(s1_19_0, k__cospi_m08_p24); + const __m128i s1_19_3 = _mm_madd_epi16(s1_19_1, k__cospi_m08_p24); + const __m128i s1_20_2 = _mm_madd_epi16(s1_20_0, k__cospi_m24_m08); + const __m128i s1_20_3 = _mm_madd_epi16(s1_20_1, k__cospi_m24_m08); + const __m128i s1_21_2 = _mm_madd_epi16(s1_21_0, k__cospi_m24_m08); + const __m128i s1_21_3 = _mm_madd_epi16(s1_21_1, k__cospi_m24_m08); + const __m128i s1_26_2 = _mm_madd_epi16(s1_21_0, k__cospi_m08_p24); + const __m128i s1_26_3 = _mm_madd_epi16(s1_21_1, k__cospi_m08_p24); + const __m128i s1_27_2 = _mm_madd_epi16(s1_20_0, k__cospi_m08_p24); + const __m128i s1_27_3 = _mm_madd_epi16(s1_20_1, k__cospi_m08_p24); + const __m128i s1_28_2 = _mm_madd_epi16(s1_19_0, k__cospi_p24_p08); + const __m128i s1_28_3 = _mm_madd_epi16(s1_19_1, k__cospi_p24_p08); + const __m128i s1_29_2 = _mm_madd_epi16(s1_18_0, k__cospi_p24_p08); + const __m128i s1_29_3 = _mm_madd_epi16(s1_18_1, k__cospi_p24_p08); + // dct_const_round_shift + const __m128i s1_18_4 = _mm_add_epi32(s1_18_2, k__DCT_CONST_ROUNDING); + const __m128i s1_18_5 = _mm_add_epi32(s1_18_3, k__DCT_CONST_ROUNDING); + const __m128i s1_19_4 = _mm_add_epi32(s1_19_2, k__DCT_CONST_ROUNDING); + const __m128i s1_19_5 = _mm_add_epi32(s1_19_3, k__DCT_CONST_ROUNDING); + const __m128i s1_20_4 = _mm_add_epi32(s1_20_2, k__DCT_CONST_ROUNDING); + const __m128i s1_20_5 = _mm_add_epi32(s1_20_3, k__DCT_CONST_ROUNDING); + const __m128i s1_21_4 = _mm_add_epi32(s1_21_2, k__DCT_CONST_ROUNDING); + const __m128i s1_21_5 = _mm_add_epi32(s1_21_3, k__DCT_CONST_ROUNDING); + const __m128i s1_26_4 = _mm_add_epi32(s1_26_2, k__DCT_CONST_ROUNDING); + const __m128i s1_26_5 = _mm_add_epi32(s1_26_3, k__DCT_CONST_ROUNDING); + const __m128i s1_27_4 = _mm_add_epi32(s1_27_2, k__DCT_CONST_ROUNDING); + const __m128i s1_27_5 = _mm_add_epi32(s1_27_3, k__DCT_CONST_ROUNDING); + const __m128i s1_28_4 = _mm_add_epi32(s1_28_2, k__DCT_CONST_ROUNDING); + const __m128i s1_28_5 = _mm_add_epi32(s1_28_3, k__DCT_CONST_ROUNDING); + const __m128i s1_29_4 = _mm_add_epi32(s1_29_2, k__DCT_CONST_ROUNDING); + const __m128i s1_29_5 = _mm_add_epi32(s1_29_3, k__DCT_CONST_ROUNDING); + const __m128i s1_18_6 = _mm_srai_epi32(s1_18_4, DCT_CONST_BITS); + const __m128i s1_18_7 = _mm_srai_epi32(s1_18_5, DCT_CONST_BITS); + const __m128i s1_19_6 = _mm_srai_epi32(s1_19_4, DCT_CONST_BITS); + const __m128i s1_19_7 = _mm_srai_epi32(s1_19_5, DCT_CONST_BITS); + const __m128i s1_20_6 = _mm_srai_epi32(s1_20_4, DCT_CONST_BITS); + const __m128i s1_20_7 = _mm_srai_epi32(s1_20_5, DCT_CONST_BITS); + const __m128i s1_21_6 = _mm_srai_epi32(s1_21_4, DCT_CONST_BITS); + const __m128i s1_21_7 = _mm_srai_epi32(s1_21_5, DCT_CONST_BITS); + const __m128i s1_26_6 = _mm_srai_epi32(s1_26_4, DCT_CONST_BITS); + const __m128i s1_26_7 = _mm_srai_epi32(s1_26_5, DCT_CONST_BITS); + const __m128i s1_27_6 = _mm_srai_epi32(s1_27_4, DCT_CONST_BITS); + const __m128i s1_27_7 = _mm_srai_epi32(s1_27_5, DCT_CONST_BITS); + const __m128i s1_28_6 = _mm_srai_epi32(s1_28_4, DCT_CONST_BITS); + const __m128i s1_28_7 = _mm_srai_epi32(s1_28_5, DCT_CONST_BITS); + const __m128i s1_29_6 = _mm_srai_epi32(s1_29_4, DCT_CONST_BITS); + const __m128i s1_29_7 = _mm_srai_epi32(s1_29_5, DCT_CONST_BITS); + // Combine + step1[18] = _mm_packs_epi32(s1_18_6, s1_18_7); + step1[19] = _mm_packs_epi32(s1_19_6, s1_19_7); + step1[20] = _mm_packs_epi32(s1_20_6, s1_20_7); + step1[21] = _mm_packs_epi32(s1_21_6, s1_21_7); + step1[26] = _mm_packs_epi32(s1_26_6, s1_26_7); + step1[27] = _mm_packs_epi32(s1_27_6, s1_27_7); + step1[28] = _mm_packs_epi32(s1_28_6, s1_28_7); + step1[29] = _mm_packs_epi32(s1_29_6, s1_29_7); + } + // Stage 5 + { + step2[4] = _mm_add_epi16(step1[5], step3[4]); + step2[5] = _mm_sub_epi16(step3[4], step1[5]); + step2[6] = _mm_sub_epi16(step3[7], step1[6]); + step2[7] = _mm_add_epi16(step1[6], step3[7]); + } + { + const __m128i out_00_0 = _mm_unpacklo_epi16(step1[0], step1[1]); + const __m128i out_00_1 = _mm_unpackhi_epi16(step1[0], step1[1]); + const __m128i out_08_0 = _mm_unpacklo_epi16(step1[2], step1[3]); + const __m128i out_08_1 = _mm_unpackhi_epi16(step1[2], step1[3]); + const __m128i out_00_2 = _mm_madd_epi16(out_00_0, k__cospi_p16_p16); + const __m128i out_00_3 = _mm_madd_epi16(out_00_1, k__cospi_p16_p16); + const __m128i out_16_2 = _mm_madd_epi16(out_00_0, k__cospi_p16_m16); + const __m128i out_16_3 = _mm_madd_epi16(out_00_1, k__cospi_p16_m16); + const __m128i out_08_2 = _mm_madd_epi16(out_08_0, k__cospi_p24_p08); + const __m128i out_08_3 = _mm_madd_epi16(out_08_1, k__cospi_p24_p08); + const __m128i out_24_2 = _mm_madd_epi16(out_08_0, k__cospi_m08_p24); + const __m128i out_24_3 = _mm_madd_epi16(out_08_1, k__cospi_m08_p24); + // dct_const_round_shift + const __m128i out_00_4 = _mm_add_epi32(out_00_2, k__DCT_CONST_ROUNDING); + const __m128i out_00_5 = _mm_add_epi32(out_00_3, k__DCT_CONST_ROUNDING); + const __m128i out_16_4 = _mm_add_epi32(out_16_2, k__DCT_CONST_ROUNDING); + const __m128i out_16_5 = _mm_add_epi32(out_16_3, k__DCT_CONST_ROUNDING); + const __m128i out_08_4 = _mm_add_epi32(out_08_2, k__DCT_CONST_ROUNDING); + const __m128i out_08_5 = _mm_add_epi32(out_08_3, k__DCT_CONST_ROUNDING); + const __m128i out_24_4 = _mm_add_epi32(out_24_2, k__DCT_CONST_ROUNDING); + const __m128i out_24_5 = _mm_add_epi32(out_24_3, k__DCT_CONST_ROUNDING); + const __m128i out_00_6 = _mm_srai_epi32(out_00_4, DCT_CONST_BITS); + const __m128i out_00_7 = _mm_srai_epi32(out_00_5, DCT_CONST_BITS); + const __m128i out_16_6 = _mm_srai_epi32(out_16_4, DCT_CONST_BITS); + const __m128i out_16_7 = _mm_srai_epi32(out_16_5, DCT_CONST_BITS); + const __m128i out_08_6 = _mm_srai_epi32(out_08_4, DCT_CONST_BITS); + const __m128i out_08_7 = _mm_srai_epi32(out_08_5, DCT_CONST_BITS); + const __m128i out_24_6 = _mm_srai_epi32(out_24_4, DCT_CONST_BITS); + const __m128i out_24_7 = _mm_srai_epi32(out_24_5, DCT_CONST_BITS); + // Combine + out[0] = _mm_packs_epi32(out_00_6, out_00_7); + out[16] = _mm_packs_epi32(out_16_6, out_16_7); + out[8] = _mm_packs_epi32(out_08_6, out_08_7); + out[24] = _mm_packs_epi32(out_24_6, out_24_7); + } + { + const __m128i s2_09_0 = _mm_unpacklo_epi16(step1[9], step1[14]); + const __m128i s2_09_1 = _mm_unpackhi_epi16(step1[9], step1[14]); + const __m128i s2_10_0 = _mm_unpacklo_epi16(step1[10], step1[13]); + const __m128i s2_10_1 = _mm_unpackhi_epi16(step1[10], step1[13]); + const __m128i s2_09_2 = _mm_madd_epi16(s2_09_0, k__cospi_m08_p24); + const __m128i s2_09_3 = _mm_madd_epi16(s2_09_1, k__cospi_m08_p24); + const __m128i s2_10_2 = _mm_madd_epi16(s2_10_0, k__cospi_m24_m08); + const __m128i s2_10_3 = _mm_madd_epi16(s2_10_1, k__cospi_m24_m08); + const __m128i s2_13_2 = _mm_madd_epi16(s2_10_0, k__cospi_m08_p24); + const __m128i s2_13_3 = _mm_madd_epi16(s2_10_1, k__cospi_m08_p24); + const __m128i s2_14_2 = _mm_madd_epi16(s2_09_0, k__cospi_p24_p08); + const __m128i s2_14_3 = _mm_madd_epi16(s2_09_1, k__cospi_p24_p08); + // dct_const_round_shift + const __m128i s2_09_4 = _mm_add_epi32(s2_09_2, k__DCT_CONST_ROUNDING); + const __m128i s2_09_5 = _mm_add_epi32(s2_09_3, k__DCT_CONST_ROUNDING); + const __m128i s2_10_4 = _mm_add_epi32(s2_10_2, k__DCT_CONST_ROUNDING); + const __m128i s2_10_5 = _mm_add_epi32(s2_10_3, k__DCT_CONST_ROUNDING); + const __m128i s2_13_4 = _mm_add_epi32(s2_13_2, k__DCT_CONST_ROUNDING); + const __m128i s2_13_5 = _mm_add_epi32(s2_13_3, k__DCT_CONST_ROUNDING); + const __m128i s2_14_4 = _mm_add_epi32(s2_14_2, k__DCT_CONST_ROUNDING); + const __m128i s2_14_5 = _mm_add_epi32(s2_14_3, k__DCT_CONST_ROUNDING); + const __m128i s2_09_6 = _mm_srai_epi32(s2_09_4, DCT_CONST_BITS); + const __m128i s2_09_7 = _mm_srai_epi32(s2_09_5, DCT_CONST_BITS); + const __m128i s2_10_6 = _mm_srai_epi32(s2_10_4, DCT_CONST_BITS); + const __m128i s2_10_7 = _mm_srai_epi32(s2_10_5, DCT_CONST_BITS); + const __m128i s2_13_6 = _mm_srai_epi32(s2_13_4, DCT_CONST_BITS); + const __m128i s2_13_7 = _mm_srai_epi32(s2_13_5, DCT_CONST_BITS); + const __m128i s2_14_6 = _mm_srai_epi32(s2_14_4, DCT_CONST_BITS); + const __m128i s2_14_7 = _mm_srai_epi32(s2_14_5, DCT_CONST_BITS); + // Combine + step2[9] = _mm_packs_epi32(s2_09_6, s2_09_7); + step2[10] = _mm_packs_epi32(s2_10_6, s2_10_7); + step2[13] = _mm_packs_epi32(s2_13_6, s2_13_7); + step2[14] = _mm_packs_epi32(s2_14_6, s2_14_7); + } + { + step2[16] = _mm_add_epi16(step1[19], step3[16]); + step2[17] = _mm_add_epi16(step1[18], step3[17]); + step2[18] = _mm_sub_epi16(step3[17], step1[18]); + step2[19] = _mm_sub_epi16(step3[16], step1[19]); + step2[20] = _mm_sub_epi16(step3[23], step1[20]); + step2[21] = _mm_sub_epi16(step3[22], step1[21]); + step2[22] = _mm_add_epi16(step1[21], step3[22]); + step2[23] = _mm_add_epi16(step1[20], step3[23]); + step2[24] = _mm_add_epi16(step1[27], step3[24]); + step2[25] = _mm_add_epi16(step1[26], step3[25]); + step2[26] = _mm_sub_epi16(step3[25], step1[26]); + step2[27] = _mm_sub_epi16(step3[24], step1[27]); + step2[28] = _mm_sub_epi16(step3[31], step1[28]); + step2[29] = _mm_sub_epi16(step3[30], step1[29]); + step2[30] = _mm_add_epi16(step1[29], step3[30]); + step2[31] = _mm_add_epi16(step1[28], step3[31]); + } + // Stage 6 + { + const __m128i out_04_0 = _mm_unpacklo_epi16(step2[4], step2[7]); + const __m128i out_04_1 = _mm_unpackhi_epi16(step2[4], step2[7]); + const __m128i out_20_0 = _mm_unpacklo_epi16(step2[5], step2[6]); + const __m128i out_20_1 = _mm_unpackhi_epi16(step2[5], step2[6]); + const __m128i out_12_0 = _mm_unpacklo_epi16(step2[5], step2[6]); + const __m128i out_12_1 = _mm_unpackhi_epi16(step2[5], step2[6]); + const __m128i out_28_0 = _mm_unpacklo_epi16(step2[4], step2[7]); + const __m128i out_28_1 = _mm_unpackhi_epi16(step2[4], step2[7]); + const __m128i out_04_2 = _mm_madd_epi16(out_04_0, k__cospi_p28_p04); + const __m128i out_04_3 = _mm_madd_epi16(out_04_1, k__cospi_p28_p04); + const __m128i out_20_2 = _mm_madd_epi16(out_20_0, k__cospi_p12_p20); + const __m128i out_20_3 = _mm_madd_epi16(out_20_1, k__cospi_p12_p20); + const __m128i out_12_2 = _mm_madd_epi16(out_12_0, k__cospi_m20_p12); + const __m128i out_12_3 = _mm_madd_epi16(out_12_1, k__cospi_m20_p12); + const __m128i out_28_2 = _mm_madd_epi16(out_28_0, k__cospi_m04_p28); + const __m128i out_28_3 = _mm_madd_epi16(out_28_1, k__cospi_m04_p28); + // dct_const_round_shift + const __m128i out_04_4 = _mm_add_epi32(out_04_2, k__DCT_CONST_ROUNDING); + const __m128i out_04_5 = _mm_add_epi32(out_04_3, k__DCT_CONST_ROUNDING); + const __m128i out_20_4 = _mm_add_epi32(out_20_2, k__DCT_CONST_ROUNDING); + const __m128i out_20_5 = _mm_add_epi32(out_20_3, k__DCT_CONST_ROUNDING); + const __m128i out_12_4 = _mm_add_epi32(out_12_2, k__DCT_CONST_ROUNDING); + const __m128i out_12_5 = _mm_add_epi32(out_12_3, k__DCT_CONST_ROUNDING); + const __m128i out_28_4 = _mm_add_epi32(out_28_2, k__DCT_CONST_ROUNDING); + const __m128i out_28_5 = _mm_add_epi32(out_28_3, k__DCT_CONST_ROUNDING); + const __m128i out_04_6 = _mm_srai_epi32(out_04_4, DCT_CONST_BITS); + const __m128i out_04_7 = _mm_srai_epi32(out_04_5, DCT_CONST_BITS); + const __m128i out_20_6 = _mm_srai_epi32(out_20_4, DCT_CONST_BITS); + const __m128i out_20_7 = _mm_srai_epi32(out_20_5, DCT_CONST_BITS); + const __m128i out_12_6 = _mm_srai_epi32(out_12_4, DCT_CONST_BITS); + const __m128i out_12_7 = _mm_srai_epi32(out_12_5, DCT_CONST_BITS); + const __m128i out_28_6 = _mm_srai_epi32(out_28_4, DCT_CONST_BITS); + const __m128i out_28_7 = _mm_srai_epi32(out_28_5, DCT_CONST_BITS); + // Combine + out[4] = _mm_packs_epi32(out_04_6, out_04_7); + out[20] = _mm_packs_epi32(out_20_6, out_20_7); + out[12] = _mm_packs_epi32(out_12_6, out_12_7); + out[28] = _mm_packs_epi32(out_28_6, out_28_7); + } + { + step3[8] = _mm_add_epi16(step2[9], step1[8]); + step3[9] = _mm_sub_epi16(step1[8], step2[9]); + step3[10] = _mm_sub_epi16(step1[11], step2[10]); + step3[11] = _mm_add_epi16(step2[10], step1[11]); + step3[12] = _mm_add_epi16(step2[13], step1[12]); + step3[13] = _mm_sub_epi16(step1[12], step2[13]); + step3[14] = _mm_sub_epi16(step1[15], step2[14]); + step3[15] = _mm_add_epi16(step2[14], step1[15]); + } + { + const __m128i s3_17_0 = _mm_unpacklo_epi16(step2[17], step2[30]); + const __m128i s3_17_1 = _mm_unpackhi_epi16(step2[17], step2[30]); + const __m128i s3_18_0 = _mm_unpacklo_epi16(step2[18], step2[29]); + const __m128i s3_18_1 = _mm_unpackhi_epi16(step2[18], step2[29]); + const __m128i s3_21_0 = _mm_unpacklo_epi16(step2[21], step2[26]); + const __m128i s3_21_1 = _mm_unpackhi_epi16(step2[21], step2[26]); + const __m128i s3_22_0 = _mm_unpacklo_epi16(step2[22], step2[25]); + const __m128i s3_22_1 = _mm_unpackhi_epi16(step2[22], step2[25]); + const __m128i s3_17_2 = _mm_madd_epi16(s3_17_0, k__cospi_m04_p28); + const __m128i s3_17_3 = _mm_madd_epi16(s3_17_1, k__cospi_m04_p28); + const __m128i s3_18_2 = _mm_madd_epi16(s3_18_0, k__cospi_m28_m04); + const __m128i s3_18_3 = _mm_madd_epi16(s3_18_1, k__cospi_m28_m04); + const __m128i s3_21_2 = _mm_madd_epi16(s3_21_0, k__cospi_m20_p12); + const __m128i s3_21_3 = _mm_madd_epi16(s3_21_1, k__cospi_m20_p12); + const __m128i s3_22_2 = _mm_madd_epi16(s3_22_0, k__cospi_m12_m20); + const __m128i s3_22_3 = _mm_madd_epi16(s3_22_1, k__cospi_m12_m20); + const __m128i s3_25_2 = _mm_madd_epi16(s3_22_0, k__cospi_m20_p12); + const __m128i s3_25_3 = _mm_madd_epi16(s3_22_1, k__cospi_m20_p12); + const __m128i s3_26_2 = _mm_madd_epi16(s3_21_0, k__cospi_p12_p20); + const __m128i s3_26_3 = _mm_madd_epi16(s3_21_1, k__cospi_p12_p20); + const __m128i s3_29_2 = _mm_madd_epi16(s3_18_0, k__cospi_m04_p28); + const __m128i s3_29_3 = _mm_madd_epi16(s3_18_1, k__cospi_m04_p28); + const __m128i s3_30_2 = _mm_madd_epi16(s3_17_0, k__cospi_p28_p04); + const __m128i s3_30_3 = _mm_madd_epi16(s3_17_1, k__cospi_p28_p04); + // dct_const_round_shift + const __m128i s3_17_4 = _mm_add_epi32(s3_17_2, k__DCT_CONST_ROUNDING); + const __m128i s3_17_5 = _mm_add_epi32(s3_17_3, k__DCT_CONST_ROUNDING); + const __m128i s3_18_4 = _mm_add_epi32(s3_18_2, k__DCT_CONST_ROUNDING); + const __m128i s3_18_5 = _mm_add_epi32(s3_18_3, k__DCT_CONST_ROUNDING); + const __m128i s3_21_4 = _mm_add_epi32(s3_21_2, k__DCT_CONST_ROUNDING); + const __m128i s3_21_5 = _mm_add_epi32(s3_21_3, k__DCT_CONST_ROUNDING); + const __m128i s3_22_4 = _mm_add_epi32(s3_22_2, k__DCT_CONST_ROUNDING); + const __m128i s3_22_5 = _mm_add_epi32(s3_22_3, k__DCT_CONST_ROUNDING); + const __m128i s3_17_6 = _mm_srai_epi32(s3_17_4, DCT_CONST_BITS); + const __m128i s3_17_7 = _mm_srai_epi32(s3_17_5, DCT_CONST_BITS); + const __m128i s3_18_6 = _mm_srai_epi32(s3_18_4, DCT_CONST_BITS); + const __m128i s3_18_7 = _mm_srai_epi32(s3_18_5, DCT_CONST_BITS); + const __m128i s3_21_6 = _mm_srai_epi32(s3_21_4, DCT_CONST_BITS); + const __m128i s3_21_7 = _mm_srai_epi32(s3_21_5, DCT_CONST_BITS); + const __m128i s3_22_6 = _mm_srai_epi32(s3_22_4, DCT_CONST_BITS); + const __m128i s3_22_7 = _mm_srai_epi32(s3_22_5, DCT_CONST_BITS); + const __m128i s3_25_4 = _mm_add_epi32(s3_25_2, k__DCT_CONST_ROUNDING); + const __m128i s3_25_5 = _mm_add_epi32(s3_25_3, k__DCT_CONST_ROUNDING); + const __m128i s3_26_4 = _mm_add_epi32(s3_26_2, k__DCT_CONST_ROUNDING); + const __m128i s3_26_5 = _mm_add_epi32(s3_26_3, k__DCT_CONST_ROUNDING); + const __m128i s3_29_4 = _mm_add_epi32(s3_29_2, k__DCT_CONST_ROUNDING); + const __m128i s3_29_5 = _mm_add_epi32(s3_29_3, k__DCT_CONST_ROUNDING); + const __m128i s3_30_4 = _mm_add_epi32(s3_30_2, k__DCT_CONST_ROUNDING); + const __m128i s3_30_5 = _mm_add_epi32(s3_30_3, k__DCT_CONST_ROUNDING); + const __m128i s3_25_6 = _mm_srai_epi32(s3_25_4, DCT_CONST_BITS); + const __m128i s3_25_7 = _mm_srai_epi32(s3_25_5, DCT_CONST_BITS); + const __m128i s3_26_6 = _mm_srai_epi32(s3_26_4, DCT_CONST_BITS); + const __m128i s3_26_7 = _mm_srai_epi32(s3_26_5, DCT_CONST_BITS); + const __m128i s3_29_6 = _mm_srai_epi32(s3_29_4, DCT_CONST_BITS); + const __m128i s3_29_7 = _mm_srai_epi32(s3_29_5, DCT_CONST_BITS); + const __m128i s3_30_6 = _mm_srai_epi32(s3_30_4, DCT_CONST_BITS); + const __m128i s3_30_7 = _mm_srai_epi32(s3_30_5, DCT_CONST_BITS); + // Combine + step3[17] = _mm_packs_epi32(s3_17_6, s3_17_7); + step3[18] = _mm_packs_epi32(s3_18_6, s3_18_7); + step3[21] = _mm_packs_epi32(s3_21_6, s3_21_7); + step3[22] = _mm_packs_epi32(s3_22_6, s3_22_7); + // Combine + step3[25] = _mm_packs_epi32(s3_25_6, s3_25_7); + step3[26] = _mm_packs_epi32(s3_26_6, s3_26_7); + step3[29] = _mm_packs_epi32(s3_29_6, s3_29_7); + step3[30] = _mm_packs_epi32(s3_30_6, s3_30_7); + } + // Stage 7 + { + const __m128i out_02_0 = _mm_unpacklo_epi16(step3[8], step3[15]); + const __m128i out_02_1 = _mm_unpackhi_epi16(step3[8], step3[15]); + const __m128i out_18_0 = _mm_unpacklo_epi16(step3[9], step3[14]); + const __m128i out_18_1 = _mm_unpackhi_epi16(step3[9], step3[14]); + const __m128i out_10_0 = _mm_unpacklo_epi16(step3[10], step3[13]); + const __m128i out_10_1 = _mm_unpackhi_epi16(step3[10], step3[13]); + const __m128i out_26_0 = _mm_unpacklo_epi16(step3[11], step3[12]); + const __m128i out_26_1 = _mm_unpackhi_epi16(step3[11], step3[12]); + const __m128i out_02_2 = _mm_madd_epi16(out_02_0, k__cospi_p30_p02); + const __m128i out_02_3 = _mm_madd_epi16(out_02_1, k__cospi_p30_p02); + const __m128i out_18_2 = _mm_madd_epi16(out_18_0, k__cospi_p14_p18); + const __m128i out_18_3 = _mm_madd_epi16(out_18_1, k__cospi_p14_p18); + const __m128i out_10_2 = _mm_madd_epi16(out_10_0, k__cospi_p22_p10); + const __m128i out_10_3 = _mm_madd_epi16(out_10_1, k__cospi_p22_p10); + const __m128i out_26_2 = _mm_madd_epi16(out_26_0, k__cospi_p06_p26); + const __m128i out_26_3 = _mm_madd_epi16(out_26_1, k__cospi_p06_p26); + const __m128i out_06_2 = _mm_madd_epi16(out_26_0, k__cospi_m26_p06); + const __m128i out_06_3 = _mm_madd_epi16(out_26_1, k__cospi_m26_p06); + const __m128i out_22_2 = _mm_madd_epi16(out_10_0, k__cospi_m10_p22); + const __m128i out_22_3 = _mm_madd_epi16(out_10_1, k__cospi_m10_p22); + const __m128i out_14_2 = _mm_madd_epi16(out_18_0, k__cospi_m18_p14); + const __m128i out_14_3 = _mm_madd_epi16(out_18_1, k__cospi_m18_p14); + const __m128i out_30_2 = _mm_madd_epi16(out_02_0, k__cospi_m02_p30); + const __m128i out_30_3 = _mm_madd_epi16(out_02_1, k__cospi_m02_p30); + // dct_const_round_shift + const __m128i out_02_4 = _mm_add_epi32(out_02_2, k__DCT_CONST_ROUNDING); + const __m128i out_02_5 = _mm_add_epi32(out_02_3, k__DCT_CONST_ROUNDING); + const __m128i out_18_4 = _mm_add_epi32(out_18_2, k__DCT_CONST_ROUNDING); + const __m128i out_18_5 = _mm_add_epi32(out_18_3, k__DCT_CONST_ROUNDING); + const __m128i out_10_4 = _mm_add_epi32(out_10_2, k__DCT_CONST_ROUNDING); + const __m128i out_10_5 = _mm_add_epi32(out_10_3, k__DCT_CONST_ROUNDING); + const __m128i out_26_4 = _mm_add_epi32(out_26_2, k__DCT_CONST_ROUNDING); + const __m128i out_26_5 = _mm_add_epi32(out_26_3, k__DCT_CONST_ROUNDING); + const __m128i out_06_4 = _mm_add_epi32(out_06_2, k__DCT_CONST_ROUNDING); + const __m128i out_06_5 = _mm_add_epi32(out_06_3, k__DCT_CONST_ROUNDING); + const __m128i out_22_4 = _mm_add_epi32(out_22_2, k__DCT_CONST_ROUNDING); + const __m128i out_22_5 = _mm_add_epi32(out_22_3, k__DCT_CONST_ROUNDING); + const __m128i out_14_4 = _mm_add_epi32(out_14_2, k__DCT_CONST_ROUNDING); + const __m128i out_14_5 = _mm_add_epi32(out_14_3, k__DCT_CONST_ROUNDING); + const __m128i out_30_4 = _mm_add_epi32(out_30_2, k__DCT_CONST_ROUNDING); + const __m128i out_30_5 = _mm_add_epi32(out_30_3, k__DCT_CONST_ROUNDING); + const __m128i out_02_6 = _mm_srai_epi32(out_02_4, DCT_CONST_BITS); + const __m128i out_02_7 = _mm_srai_epi32(out_02_5, DCT_CONST_BITS); + const __m128i out_18_6 = _mm_srai_epi32(out_18_4, DCT_CONST_BITS); + const __m128i out_18_7 = _mm_srai_epi32(out_18_5, DCT_CONST_BITS); + const __m128i out_10_6 = _mm_srai_epi32(out_10_4, DCT_CONST_BITS); + const __m128i out_10_7 = _mm_srai_epi32(out_10_5, DCT_CONST_BITS); + const __m128i out_26_6 = _mm_srai_epi32(out_26_4, DCT_CONST_BITS); + const __m128i out_26_7 = _mm_srai_epi32(out_26_5, DCT_CONST_BITS); + const __m128i out_06_6 = _mm_srai_epi32(out_06_4, DCT_CONST_BITS); + const __m128i out_06_7 = _mm_srai_epi32(out_06_5, DCT_CONST_BITS); + const __m128i out_22_6 = _mm_srai_epi32(out_22_4, DCT_CONST_BITS); + const __m128i out_22_7 = _mm_srai_epi32(out_22_5, DCT_CONST_BITS); + const __m128i out_14_6 = _mm_srai_epi32(out_14_4, DCT_CONST_BITS); + const __m128i out_14_7 = _mm_srai_epi32(out_14_5, DCT_CONST_BITS); + const __m128i out_30_6 = _mm_srai_epi32(out_30_4, DCT_CONST_BITS); + const __m128i out_30_7 = _mm_srai_epi32(out_30_5, DCT_CONST_BITS); + // Combine + out[2] = _mm_packs_epi32(out_02_6, out_02_7); + out[18] = _mm_packs_epi32(out_18_6, out_18_7); + out[10] = _mm_packs_epi32(out_10_6, out_10_7); + out[26] = _mm_packs_epi32(out_26_6, out_26_7); + out[6] = _mm_packs_epi32(out_06_6, out_06_7); + out[22] = _mm_packs_epi32(out_22_6, out_22_7); + out[14] = _mm_packs_epi32(out_14_6, out_14_7); + out[30] = _mm_packs_epi32(out_30_6, out_30_7); + } + { + step1[16] = _mm_add_epi16(step3[17], step2[16]); + step1[17] = _mm_sub_epi16(step2[16], step3[17]); + step1[18] = _mm_sub_epi16(step2[19], step3[18]); + step1[19] = _mm_add_epi16(step3[18], step2[19]); + step1[20] = _mm_add_epi16(step3[21], step2[20]); + step1[21] = _mm_sub_epi16(step2[20], step3[21]); + step1[22] = _mm_sub_epi16(step2[23], step3[22]); + step1[23] = _mm_add_epi16(step3[22], step2[23]); + step1[24] = _mm_add_epi16(step3[25], step2[24]); + step1[25] = _mm_sub_epi16(step2[24], step3[25]); + step1[26] = _mm_sub_epi16(step2[27], step3[26]); + step1[27] = _mm_add_epi16(step3[26], step2[27]); + step1[28] = _mm_add_epi16(step3[29], step2[28]); + step1[29] = _mm_sub_epi16(step2[28], step3[29]); + step1[30] = _mm_sub_epi16(step2[31], step3[30]); + step1[31] = _mm_add_epi16(step3[30], step2[31]); + } + // Final stage --- outputs indices are bit-reversed. + { + const __m128i out_01_0 = _mm_unpacklo_epi16(step1[16], step1[31]); + const __m128i out_01_1 = _mm_unpackhi_epi16(step1[16], step1[31]); + const __m128i out_17_0 = _mm_unpacklo_epi16(step1[17], step1[30]); + const __m128i out_17_1 = _mm_unpackhi_epi16(step1[17], step1[30]); + const __m128i out_09_0 = _mm_unpacklo_epi16(step1[18], step1[29]); + const __m128i out_09_1 = _mm_unpackhi_epi16(step1[18], step1[29]); + const __m128i out_25_0 = _mm_unpacklo_epi16(step1[19], step1[28]); + const __m128i out_25_1 = _mm_unpackhi_epi16(step1[19], step1[28]); + const __m128i out_01_2 = _mm_madd_epi16(out_01_0, k__cospi_p31_p01); + const __m128i out_01_3 = _mm_madd_epi16(out_01_1, k__cospi_p31_p01); + const __m128i out_17_2 = _mm_madd_epi16(out_17_0, k__cospi_p15_p17); + const __m128i out_17_3 = _mm_madd_epi16(out_17_1, k__cospi_p15_p17); + const __m128i out_09_2 = _mm_madd_epi16(out_09_0, k__cospi_p23_p09); + const __m128i out_09_3 = _mm_madd_epi16(out_09_1, k__cospi_p23_p09); + const __m128i out_25_2 = _mm_madd_epi16(out_25_0, k__cospi_p07_p25); + const __m128i out_25_3 = _mm_madd_epi16(out_25_1, k__cospi_p07_p25); + const __m128i out_07_2 = _mm_madd_epi16(out_25_0, k__cospi_m25_p07); + const __m128i out_07_3 = _mm_madd_epi16(out_25_1, k__cospi_m25_p07); + const __m128i out_23_2 = _mm_madd_epi16(out_09_0, k__cospi_m09_p23); + const __m128i out_23_3 = _mm_madd_epi16(out_09_1, k__cospi_m09_p23); + const __m128i out_15_2 = _mm_madd_epi16(out_17_0, k__cospi_m17_p15); + const __m128i out_15_3 = _mm_madd_epi16(out_17_1, k__cospi_m17_p15); + const __m128i out_31_2 = _mm_madd_epi16(out_01_0, k__cospi_m01_p31); + const __m128i out_31_3 = _mm_madd_epi16(out_01_1, k__cospi_m01_p31); + // dct_const_round_shift + const __m128i out_01_4 = _mm_add_epi32(out_01_2, k__DCT_CONST_ROUNDING); + const __m128i out_01_5 = _mm_add_epi32(out_01_3, k__DCT_CONST_ROUNDING); + const __m128i out_17_4 = _mm_add_epi32(out_17_2, k__DCT_CONST_ROUNDING); + const __m128i out_17_5 = _mm_add_epi32(out_17_3, k__DCT_CONST_ROUNDING); + const __m128i out_09_4 = _mm_add_epi32(out_09_2, k__DCT_CONST_ROUNDING); + const __m128i out_09_5 = _mm_add_epi32(out_09_3, k__DCT_CONST_ROUNDING); + const __m128i out_25_4 = _mm_add_epi32(out_25_2, k__DCT_CONST_ROUNDING); + const __m128i out_25_5 = _mm_add_epi32(out_25_3, k__DCT_CONST_ROUNDING); + const __m128i out_07_4 = _mm_add_epi32(out_07_2, k__DCT_CONST_ROUNDING); + const __m128i out_07_5 = _mm_add_epi32(out_07_3, k__DCT_CONST_ROUNDING); + const __m128i out_23_4 = _mm_add_epi32(out_23_2, k__DCT_CONST_ROUNDING); + const __m128i out_23_5 = _mm_add_epi32(out_23_3, k__DCT_CONST_ROUNDING); + const __m128i out_15_4 = _mm_add_epi32(out_15_2, k__DCT_CONST_ROUNDING); + const __m128i out_15_5 = _mm_add_epi32(out_15_3, k__DCT_CONST_ROUNDING); + const __m128i out_31_4 = _mm_add_epi32(out_31_2, k__DCT_CONST_ROUNDING); + const __m128i out_31_5 = _mm_add_epi32(out_31_3, k__DCT_CONST_ROUNDING); + const __m128i out_01_6 = _mm_srai_epi32(out_01_4, DCT_CONST_BITS); + const __m128i out_01_7 = _mm_srai_epi32(out_01_5, DCT_CONST_BITS); + const __m128i out_17_6 = _mm_srai_epi32(out_17_4, DCT_CONST_BITS); + const __m128i out_17_7 = _mm_srai_epi32(out_17_5, DCT_CONST_BITS); + const __m128i out_09_6 = _mm_srai_epi32(out_09_4, DCT_CONST_BITS); + const __m128i out_09_7 = _mm_srai_epi32(out_09_5, DCT_CONST_BITS); + const __m128i out_25_6 = _mm_srai_epi32(out_25_4, DCT_CONST_BITS); + const __m128i out_25_7 = _mm_srai_epi32(out_25_5, DCT_CONST_BITS); + const __m128i out_07_6 = _mm_srai_epi32(out_07_4, DCT_CONST_BITS); + const __m128i out_07_7 = _mm_srai_epi32(out_07_5, DCT_CONST_BITS); + const __m128i out_23_6 = _mm_srai_epi32(out_23_4, DCT_CONST_BITS); + const __m128i out_23_7 = _mm_srai_epi32(out_23_5, DCT_CONST_BITS); + const __m128i out_15_6 = _mm_srai_epi32(out_15_4, DCT_CONST_BITS); + const __m128i out_15_7 = _mm_srai_epi32(out_15_5, DCT_CONST_BITS); + const __m128i out_31_6 = _mm_srai_epi32(out_31_4, DCT_CONST_BITS); + const __m128i out_31_7 = _mm_srai_epi32(out_31_5, DCT_CONST_BITS); + // Combine + out[1] = _mm_packs_epi32(out_01_6, out_01_7); + out[17] = _mm_packs_epi32(out_17_6, out_17_7); + out[9] = _mm_packs_epi32(out_09_6, out_09_7); + out[25] = _mm_packs_epi32(out_25_6, out_25_7); + out[7] = _mm_packs_epi32(out_07_6, out_07_7); + out[23] = _mm_packs_epi32(out_23_6, out_23_7); + out[15] = _mm_packs_epi32(out_15_6, out_15_7); + out[31] = _mm_packs_epi32(out_31_6, out_31_7); + } + { + const __m128i out_05_0 = _mm_unpacklo_epi16(step1[20], step1[27]); + const __m128i out_05_1 = _mm_unpackhi_epi16(step1[20], step1[27]); + const __m128i out_21_0 = _mm_unpacklo_epi16(step1[21], step1[26]); + const __m128i out_21_1 = _mm_unpackhi_epi16(step1[21], step1[26]); + const __m128i out_13_0 = _mm_unpacklo_epi16(step1[22], step1[25]); + const __m128i out_13_1 = _mm_unpackhi_epi16(step1[22], step1[25]); + const __m128i out_29_0 = _mm_unpacklo_epi16(step1[23], step1[24]); + const __m128i out_29_1 = _mm_unpackhi_epi16(step1[23], step1[24]); + const __m128i out_05_2 = _mm_madd_epi16(out_05_0, k__cospi_p27_p05); + const __m128i out_05_3 = _mm_madd_epi16(out_05_1, k__cospi_p27_p05); + const __m128i out_21_2 = _mm_madd_epi16(out_21_0, k__cospi_p11_p21); + const __m128i out_21_3 = _mm_madd_epi16(out_21_1, k__cospi_p11_p21); + const __m128i out_13_2 = _mm_madd_epi16(out_13_0, k__cospi_p19_p13); + const __m128i out_13_3 = _mm_madd_epi16(out_13_1, k__cospi_p19_p13); + const __m128i out_29_2 = _mm_madd_epi16(out_29_0, k__cospi_p03_p29); + const __m128i out_29_3 = _mm_madd_epi16(out_29_1, k__cospi_p03_p29); + const __m128i out_03_2 = _mm_madd_epi16(out_29_0, k__cospi_m29_p03); + const __m128i out_03_3 = _mm_madd_epi16(out_29_1, k__cospi_m29_p03); + const __m128i out_19_2 = _mm_madd_epi16(out_13_0, k__cospi_m13_p19); + const __m128i out_19_3 = _mm_madd_epi16(out_13_1, k__cospi_m13_p19); + const __m128i out_11_2 = _mm_madd_epi16(out_21_0, k__cospi_m21_p11); + const __m128i out_11_3 = _mm_madd_epi16(out_21_1, k__cospi_m21_p11); + const __m128i out_27_2 = _mm_madd_epi16(out_05_0, k__cospi_m05_p27); + const __m128i out_27_3 = _mm_madd_epi16(out_05_1, k__cospi_m05_p27); + // dct_const_round_shift + const __m128i out_05_4 = _mm_add_epi32(out_05_2, k__DCT_CONST_ROUNDING); + const __m128i out_05_5 = _mm_add_epi32(out_05_3, k__DCT_CONST_ROUNDING); + const __m128i out_21_4 = _mm_add_epi32(out_21_2, k__DCT_CONST_ROUNDING); + const __m128i out_21_5 = _mm_add_epi32(out_21_3, k__DCT_CONST_ROUNDING); + const __m128i out_13_4 = _mm_add_epi32(out_13_2, k__DCT_CONST_ROUNDING); + const __m128i out_13_5 = _mm_add_epi32(out_13_3, k__DCT_CONST_ROUNDING); + const __m128i out_29_4 = _mm_add_epi32(out_29_2, k__DCT_CONST_ROUNDING); + const __m128i out_29_5 = _mm_add_epi32(out_29_3, k__DCT_CONST_ROUNDING); + const __m128i out_03_4 = _mm_add_epi32(out_03_2, k__DCT_CONST_ROUNDING); + const __m128i out_03_5 = _mm_add_epi32(out_03_3, k__DCT_CONST_ROUNDING); + const __m128i out_19_4 = _mm_add_epi32(out_19_2, k__DCT_CONST_ROUNDING); + const __m128i out_19_5 = _mm_add_epi32(out_19_3, k__DCT_CONST_ROUNDING); + const __m128i out_11_4 = _mm_add_epi32(out_11_2, k__DCT_CONST_ROUNDING); + const __m128i out_11_5 = _mm_add_epi32(out_11_3, k__DCT_CONST_ROUNDING); + const __m128i out_27_4 = _mm_add_epi32(out_27_2, k__DCT_CONST_ROUNDING); + const __m128i out_27_5 = _mm_add_epi32(out_27_3, k__DCT_CONST_ROUNDING); + const __m128i out_05_6 = _mm_srai_epi32(out_05_4, DCT_CONST_BITS); + const __m128i out_05_7 = _mm_srai_epi32(out_05_5, DCT_CONST_BITS); + const __m128i out_21_6 = _mm_srai_epi32(out_21_4, DCT_CONST_BITS); + const __m128i out_21_7 = _mm_srai_epi32(out_21_5, DCT_CONST_BITS); + const __m128i out_13_6 = _mm_srai_epi32(out_13_4, DCT_CONST_BITS); + const __m128i out_13_7 = _mm_srai_epi32(out_13_5, DCT_CONST_BITS); + const __m128i out_29_6 = _mm_srai_epi32(out_29_4, DCT_CONST_BITS); + const __m128i out_29_7 = _mm_srai_epi32(out_29_5, DCT_CONST_BITS); + const __m128i out_03_6 = _mm_srai_epi32(out_03_4, DCT_CONST_BITS); + const __m128i out_03_7 = _mm_srai_epi32(out_03_5, DCT_CONST_BITS); + const __m128i out_19_6 = _mm_srai_epi32(out_19_4, DCT_CONST_BITS); + const __m128i out_19_7 = _mm_srai_epi32(out_19_5, DCT_CONST_BITS); + const __m128i out_11_6 = _mm_srai_epi32(out_11_4, DCT_CONST_BITS); + const __m128i out_11_7 = _mm_srai_epi32(out_11_5, DCT_CONST_BITS); + const __m128i out_27_6 = _mm_srai_epi32(out_27_4, DCT_CONST_BITS); + const __m128i out_27_7 = _mm_srai_epi32(out_27_5, DCT_CONST_BITS); + // Combine + out[5] = _mm_packs_epi32(out_05_6, out_05_7); + out[21] = _mm_packs_epi32(out_21_6, out_21_7); + out[13] = _mm_packs_epi32(out_13_6, out_13_7); + out[29] = _mm_packs_epi32(out_29_6, out_29_7); + out[3] = _mm_packs_epi32(out_03_6, out_03_7); + out[19] = _mm_packs_epi32(out_19_6, out_19_7); + out[11] = _mm_packs_epi32(out_11_6, out_11_7); + out[27] = _mm_packs_epi32(out_27_6, out_27_7); + } + + // Output results + { + int j; + for (j = 0; j < 16; ++j) { + _mm_storeu_si128((__m128i *)(in0 + j), out[j]); + _mm_storeu_si128((__m128i *)(in1 + j), out[j + 16]); + } + } +} // NOLINT diff --git a/third_party/aom/aom_dsp/x86/fwd_dct32x32_impl_avx2.h b/third_party/aom/aom_dsp/x86/fwd_dct32x32_impl_avx2.h new file mode 100644 index 000000000..216739581 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/fwd_dct32x32_impl_avx2.h @@ -0,0 +1,3022 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include <immintrin.h> // AVX2 + +#include "aom_dsp/txfm_common.h" +#include "aom_dsp/x86/txfm_common_intrin.h" +#include "aom_dsp/x86/txfm_common_avx2.h" + +#if FDCT32x32_HIGH_PRECISION +static INLINE __m256i k_madd_epi32_avx2(__m256i a, __m256i b) { + __m256i buf0, buf1; + buf0 = _mm256_mul_epu32(a, b); + a = _mm256_srli_epi64(a, 32); + b = _mm256_srli_epi64(b, 32); + buf1 = _mm256_mul_epu32(a, b); + return _mm256_add_epi64(buf0, buf1); +} + +static INLINE __m256i k_packs_epi64_avx2(__m256i a, __m256i b) { + __m256i buf0 = _mm256_shuffle_epi32(a, _MM_SHUFFLE(0, 0, 2, 0)); + __m256i buf1 = _mm256_shuffle_epi32(b, _MM_SHUFFLE(0, 0, 2, 0)); + return _mm256_unpacklo_epi64(buf0, buf1); +} +#endif + +#ifndef STORE_COEFF_FUNC +#define STORE_COEFF_FUNC +static void store_coeff(const __m256i *coeff, tran_low_t *curr, + tran_low_t *next) { + __m128i u = _mm256_castsi256_si128(*coeff); + storeu_output(&u, curr); + u = _mm256_extractf128_si256(*coeff, 1); + storeu_output(&u, next); +} +#endif + +void FDCT32x32_2D_AVX2(const int16_t *input, tran_low_t *output_org, + int stride) { + // Calculate pre-multiplied strides + const int str1 = stride; + const int str2 = 2 * stride; + const int str3 = 2 * stride + str1; + // We need an intermediate buffer between passes. + DECLARE_ALIGNED(32, int16_t, intermediate[32 * 32]); + // Constants + // When we use them, in one case, they are all the same. In all others + // it's a pair of them that we need to repeat four times. This is done + // by constructing the 32 bit constant corresponding to that pair. + const __m256i k__cospi_p16_p16 = _mm256_set1_epi16((int16_t)cospi_16_64); + const __m256i k__cospi_p16_m16 = + pair256_set_epi16(+cospi_16_64, -cospi_16_64); + const __m256i k__cospi_m08_p24 = pair256_set_epi16(-cospi_8_64, cospi_24_64); + const __m256i k__cospi_m24_m08 = pair256_set_epi16(-cospi_24_64, -cospi_8_64); + const __m256i k__cospi_p24_p08 = pair256_set_epi16(+cospi_24_64, cospi_8_64); + const __m256i k__cospi_p12_p20 = pair256_set_epi16(+cospi_12_64, cospi_20_64); + const __m256i k__cospi_m20_p12 = pair256_set_epi16(-cospi_20_64, cospi_12_64); + const __m256i k__cospi_m04_p28 = pair256_set_epi16(-cospi_4_64, cospi_28_64); + const __m256i k__cospi_p28_p04 = pair256_set_epi16(+cospi_28_64, cospi_4_64); + const __m256i k__cospi_m28_m04 = pair256_set_epi16(-cospi_28_64, -cospi_4_64); + const __m256i k__cospi_m12_m20 = + pair256_set_epi16(-cospi_12_64, -cospi_20_64); + const __m256i k__cospi_p30_p02 = pair256_set_epi16(+cospi_30_64, cospi_2_64); + const __m256i k__cospi_p14_p18 = pair256_set_epi16(+cospi_14_64, cospi_18_64); + const __m256i k__cospi_p22_p10 = pair256_set_epi16(+cospi_22_64, cospi_10_64); + const __m256i k__cospi_p06_p26 = pair256_set_epi16(+cospi_6_64, cospi_26_64); + const __m256i k__cospi_m26_p06 = pair256_set_epi16(-cospi_26_64, cospi_6_64); + const __m256i k__cospi_m10_p22 = pair256_set_epi16(-cospi_10_64, cospi_22_64); + const __m256i k__cospi_m18_p14 = pair256_set_epi16(-cospi_18_64, cospi_14_64); + const __m256i k__cospi_m02_p30 = pair256_set_epi16(-cospi_2_64, cospi_30_64); + const __m256i k__cospi_p31_p01 = pair256_set_epi16(+cospi_31_64, cospi_1_64); + const __m256i k__cospi_p15_p17 = pair256_set_epi16(+cospi_15_64, cospi_17_64); + const __m256i k__cospi_p23_p09 = pair256_set_epi16(+cospi_23_64, cospi_9_64); + const __m256i k__cospi_p07_p25 = pair256_set_epi16(+cospi_7_64, cospi_25_64); + const __m256i k__cospi_m25_p07 = pair256_set_epi16(-cospi_25_64, cospi_7_64); + const __m256i k__cospi_m09_p23 = pair256_set_epi16(-cospi_9_64, cospi_23_64); + const __m256i k__cospi_m17_p15 = pair256_set_epi16(-cospi_17_64, cospi_15_64); + const __m256i k__cospi_m01_p31 = pair256_set_epi16(-cospi_1_64, cospi_31_64); + const __m256i k__cospi_p27_p05 = pair256_set_epi16(+cospi_27_64, cospi_5_64); + const __m256i k__cospi_p11_p21 = pair256_set_epi16(+cospi_11_64, cospi_21_64); + const __m256i k__cospi_p19_p13 = pair256_set_epi16(+cospi_19_64, cospi_13_64); + const __m256i k__cospi_p03_p29 = pair256_set_epi16(+cospi_3_64, cospi_29_64); + const __m256i k__cospi_m29_p03 = pair256_set_epi16(-cospi_29_64, cospi_3_64); + const __m256i k__cospi_m13_p19 = pair256_set_epi16(-cospi_13_64, cospi_19_64); + const __m256i k__cospi_m21_p11 = pair256_set_epi16(-cospi_21_64, cospi_11_64); + const __m256i k__cospi_m05_p27 = pair256_set_epi16(-cospi_5_64, cospi_27_64); + const __m256i k__DCT_CONST_ROUNDING = _mm256_set1_epi32(DCT_CONST_ROUNDING); + const __m256i kZero = _mm256_set1_epi16(0); + const __m256i kOne = _mm256_set1_epi16(1); + // Do the two transform/transpose passes + int pass; + for (pass = 0; pass < 2; ++pass) { + // We process sixteen columns (transposed rows in second pass) at a time. + int column_start; + for (column_start = 0; column_start < 32; column_start += 16) { + __m256i step1[32]; + __m256i step2[32]; + __m256i step3[32]; + __m256i out[32]; + // Stage 1 + // Note: even though all the loads below are aligned, using the aligned + // intrinsic make the code slightly slower. + if (0 == pass) { + const int16_t *in = &input[column_start]; + // step1[i] = (in[ 0 * stride] + in[(32 - 1) * stride]) << 2; + // Note: the next four blocks could be in a loop. That would help the + // instruction cache but is actually slower. + { + const int16_t *ina = in + 0 * str1; + const int16_t *inb = in + 31 * str1; + __m256i *step1a = &step1[0]; + __m256i *step1b = &step1[31]; + const __m256i ina0 = _mm256_loadu_si256((const __m256i *)(ina)); + const __m256i ina1 = + _mm256_loadu_si256((const __m256i *)(ina + str1)); + const __m256i ina2 = + _mm256_loadu_si256((const __m256i *)(ina + str2)); + const __m256i ina3 = + _mm256_loadu_si256((const __m256i *)(ina + str3)); + const __m256i inb3 = + _mm256_loadu_si256((const __m256i *)(inb - str3)); + const __m256i inb2 = + _mm256_loadu_si256((const __m256i *)(inb - str2)); + const __m256i inb1 = + _mm256_loadu_si256((const __m256i *)(inb - str1)); + const __m256i inb0 = _mm256_loadu_si256((const __m256i *)(inb)); + step1a[0] = _mm256_add_epi16(ina0, inb0); + step1a[1] = _mm256_add_epi16(ina1, inb1); + step1a[2] = _mm256_add_epi16(ina2, inb2); + step1a[3] = _mm256_add_epi16(ina3, inb3); + step1b[-3] = _mm256_sub_epi16(ina3, inb3); + step1b[-2] = _mm256_sub_epi16(ina2, inb2); + step1b[-1] = _mm256_sub_epi16(ina1, inb1); + step1b[-0] = _mm256_sub_epi16(ina0, inb0); + step1a[0] = _mm256_slli_epi16(step1a[0], 2); + step1a[1] = _mm256_slli_epi16(step1a[1], 2); + step1a[2] = _mm256_slli_epi16(step1a[2], 2); + step1a[3] = _mm256_slli_epi16(step1a[3], 2); + step1b[-3] = _mm256_slli_epi16(step1b[-3], 2); + step1b[-2] = _mm256_slli_epi16(step1b[-2], 2); + step1b[-1] = _mm256_slli_epi16(step1b[-1], 2); + step1b[-0] = _mm256_slli_epi16(step1b[-0], 2); + } + { + const int16_t *ina = in + 4 * str1; + const int16_t *inb = in + 27 * str1; + __m256i *step1a = &step1[4]; + __m256i *step1b = &step1[27]; + const __m256i ina0 = _mm256_loadu_si256((const __m256i *)(ina)); + const __m256i ina1 = + _mm256_loadu_si256((const __m256i *)(ina + str1)); + const __m256i ina2 = + _mm256_loadu_si256((const __m256i *)(ina + str2)); + const __m256i ina3 = + _mm256_loadu_si256((const __m256i *)(ina + str3)); + const __m256i inb3 = + _mm256_loadu_si256((const __m256i *)(inb - str3)); + const __m256i inb2 = + _mm256_loadu_si256((const __m256i *)(inb - str2)); + const __m256i inb1 = + _mm256_loadu_si256((const __m256i *)(inb - str1)); + const __m256i inb0 = _mm256_loadu_si256((const __m256i *)(inb)); + step1a[0] = _mm256_add_epi16(ina0, inb0); + step1a[1] = _mm256_add_epi16(ina1, inb1); + step1a[2] = _mm256_add_epi16(ina2, inb2); + step1a[3] = _mm256_add_epi16(ina3, inb3); + step1b[-3] = _mm256_sub_epi16(ina3, inb3); + step1b[-2] = _mm256_sub_epi16(ina2, inb2); + step1b[-1] = _mm256_sub_epi16(ina1, inb1); + step1b[-0] = _mm256_sub_epi16(ina0, inb0); + step1a[0] = _mm256_slli_epi16(step1a[0], 2); + step1a[1] = _mm256_slli_epi16(step1a[1], 2); + step1a[2] = _mm256_slli_epi16(step1a[2], 2); + step1a[3] = _mm256_slli_epi16(step1a[3], 2); + step1b[-3] = _mm256_slli_epi16(step1b[-3], 2); + step1b[-2] = _mm256_slli_epi16(step1b[-2], 2); + step1b[-1] = _mm256_slli_epi16(step1b[-1], 2); + step1b[-0] = _mm256_slli_epi16(step1b[-0], 2); + } + { + const int16_t *ina = in + 8 * str1; + const int16_t *inb = in + 23 * str1; + __m256i *step1a = &step1[8]; + __m256i *step1b = &step1[23]; + const __m256i ina0 = _mm256_loadu_si256((const __m256i *)(ina)); + const __m256i ina1 = + _mm256_loadu_si256((const __m256i *)(ina + str1)); + const __m256i ina2 = + _mm256_loadu_si256((const __m256i *)(ina + str2)); + const __m256i ina3 = + _mm256_loadu_si256((const __m256i *)(ina + str3)); + const __m256i inb3 = + _mm256_loadu_si256((const __m256i *)(inb - str3)); + const __m256i inb2 = + _mm256_loadu_si256((const __m256i *)(inb - str2)); + const __m256i inb1 = + _mm256_loadu_si256((const __m256i *)(inb - str1)); + const __m256i inb0 = _mm256_loadu_si256((const __m256i *)(inb)); + step1a[0] = _mm256_add_epi16(ina0, inb0); + step1a[1] = _mm256_add_epi16(ina1, inb1); + step1a[2] = _mm256_add_epi16(ina2, inb2); + step1a[3] = _mm256_add_epi16(ina3, inb3); + step1b[-3] = _mm256_sub_epi16(ina3, inb3); + step1b[-2] = _mm256_sub_epi16(ina2, inb2); + step1b[-1] = _mm256_sub_epi16(ina1, inb1); + step1b[-0] = _mm256_sub_epi16(ina0, inb0); + step1a[0] = _mm256_slli_epi16(step1a[0], 2); + step1a[1] = _mm256_slli_epi16(step1a[1], 2); + step1a[2] = _mm256_slli_epi16(step1a[2], 2); + step1a[3] = _mm256_slli_epi16(step1a[3], 2); + step1b[-3] = _mm256_slli_epi16(step1b[-3], 2); + step1b[-2] = _mm256_slli_epi16(step1b[-2], 2); + step1b[-1] = _mm256_slli_epi16(step1b[-1], 2); + step1b[-0] = _mm256_slli_epi16(step1b[-0], 2); + } + { + const int16_t *ina = in + 12 * str1; + const int16_t *inb = in + 19 * str1; + __m256i *step1a = &step1[12]; + __m256i *step1b = &step1[19]; + const __m256i ina0 = _mm256_loadu_si256((const __m256i *)(ina)); + const __m256i ina1 = + _mm256_loadu_si256((const __m256i *)(ina + str1)); + const __m256i ina2 = + _mm256_loadu_si256((const __m256i *)(ina + str2)); + const __m256i ina3 = + _mm256_loadu_si256((const __m256i *)(ina + str3)); + const __m256i inb3 = + _mm256_loadu_si256((const __m256i *)(inb - str3)); + const __m256i inb2 = + _mm256_loadu_si256((const __m256i *)(inb - str2)); + const __m256i inb1 = + _mm256_loadu_si256((const __m256i *)(inb - str1)); + const __m256i inb0 = _mm256_loadu_si256((const __m256i *)(inb)); + step1a[0] = _mm256_add_epi16(ina0, inb0); + step1a[1] = _mm256_add_epi16(ina1, inb1); + step1a[2] = _mm256_add_epi16(ina2, inb2); + step1a[3] = _mm256_add_epi16(ina3, inb3); + step1b[-3] = _mm256_sub_epi16(ina3, inb3); + step1b[-2] = _mm256_sub_epi16(ina2, inb2); + step1b[-1] = _mm256_sub_epi16(ina1, inb1); + step1b[-0] = _mm256_sub_epi16(ina0, inb0); + step1a[0] = _mm256_slli_epi16(step1a[0], 2); + step1a[1] = _mm256_slli_epi16(step1a[1], 2); + step1a[2] = _mm256_slli_epi16(step1a[2], 2); + step1a[3] = _mm256_slli_epi16(step1a[3], 2); + step1b[-3] = _mm256_slli_epi16(step1b[-3], 2); + step1b[-2] = _mm256_slli_epi16(step1b[-2], 2); + step1b[-1] = _mm256_slli_epi16(step1b[-1], 2); + step1b[-0] = _mm256_slli_epi16(step1b[-0], 2); + } + } else { + int16_t *in = &intermediate[column_start]; + // step1[i] = in[ 0 * 32] + in[(32 - 1) * 32]; + // Note: using the same approach as above to have common offset is + // counter-productive as all offsets can be calculated at compile + // time. + // Note: the next four blocks could be in a loop. That would help the + // instruction cache but is actually slower. + { + __m256i in00 = _mm256_loadu_si256((const __m256i *)(in + 0 * 32)); + __m256i in01 = _mm256_loadu_si256((const __m256i *)(in + 1 * 32)); + __m256i in02 = _mm256_loadu_si256((const __m256i *)(in + 2 * 32)); + __m256i in03 = _mm256_loadu_si256((const __m256i *)(in + 3 * 32)); + __m256i in28 = _mm256_loadu_si256((const __m256i *)(in + 28 * 32)); + __m256i in29 = _mm256_loadu_si256((const __m256i *)(in + 29 * 32)); + __m256i in30 = _mm256_loadu_si256((const __m256i *)(in + 30 * 32)); + __m256i in31 = _mm256_loadu_si256((const __m256i *)(in + 31 * 32)); + step1[0] = _mm256_add_epi16(in00, in31); + step1[1] = _mm256_add_epi16(in01, in30); + step1[2] = _mm256_add_epi16(in02, in29); + step1[3] = _mm256_add_epi16(in03, in28); + step1[28] = _mm256_sub_epi16(in03, in28); + step1[29] = _mm256_sub_epi16(in02, in29); + step1[30] = _mm256_sub_epi16(in01, in30); + step1[31] = _mm256_sub_epi16(in00, in31); + } + { + __m256i in04 = _mm256_loadu_si256((const __m256i *)(in + 4 * 32)); + __m256i in05 = _mm256_loadu_si256((const __m256i *)(in + 5 * 32)); + __m256i in06 = _mm256_loadu_si256((const __m256i *)(in + 6 * 32)); + __m256i in07 = _mm256_loadu_si256((const __m256i *)(in + 7 * 32)); + __m256i in24 = _mm256_loadu_si256((const __m256i *)(in + 24 * 32)); + __m256i in25 = _mm256_loadu_si256((const __m256i *)(in + 25 * 32)); + __m256i in26 = _mm256_loadu_si256((const __m256i *)(in + 26 * 32)); + __m256i in27 = _mm256_loadu_si256((const __m256i *)(in + 27 * 32)); + step1[4] = _mm256_add_epi16(in04, in27); + step1[5] = _mm256_add_epi16(in05, in26); + step1[6] = _mm256_add_epi16(in06, in25); + step1[7] = _mm256_add_epi16(in07, in24); + step1[24] = _mm256_sub_epi16(in07, in24); + step1[25] = _mm256_sub_epi16(in06, in25); + step1[26] = _mm256_sub_epi16(in05, in26); + step1[27] = _mm256_sub_epi16(in04, in27); + } + { + __m256i in08 = _mm256_loadu_si256((const __m256i *)(in + 8 * 32)); + __m256i in09 = _mm256_loadu_si256((const __m256i *)(in + 9 * 32)); + __m256i in10 = _mm256_loadu_si256((const __m256i *)(in + 10 * 32)); + __m256i in11 = _mm256_loadu_si256((const __m256i *)(in + 11 * 32)); + __m256i in20 = _mm256_loadu_si256((const __m256i *)(in + 20 * 32)); + __m256i in21 = _mm256_loadu_si256((const __m256i *)(in + 21 * 32)); + __m256i in22 = _mm256_loadu_si256((const __m256i *)(in + 22 * 32)); + __m256i in23 = _mm256_loadu_si256((const __m256i *)(in + 23 * 32)); + step1[8] = _mm256_add_epi16(in08, in23); + step1[9] = _mm256_add_epi16(in09, in22); + step1[10] = _mm256_add_epi16(in10, in21); + step1[11] = _mm256_add_epi16(in11, in20); + step1[20] = _mm256_sub_epi16(in11, in20); + step1[21] = _mm256_sub_epi16(in10, in21); + step1[22] = _mm256_sub_epi16(in09, in22); + step1[23] = _mm256_sub_epi16(in08, in23); + } + { + __m256i in12 = _mm256_loadu_si256((const __m256i *)(in + 12 * 32)); + __m256i in13 = _mm256_loadu_si256((const __m256i *)(in + 13 * 32)); + __m256i in14 = _mm256_loadu_si256((const __m256i *)(in + 14 * 32)); + __m256i in15 = _mm256_loadu_si256((const __m256i *)(in + 15 * 32)); + __m256i in16 = _mm256_loadu_si256((const __m256i *)(in + 16 * 32)); + __m256i in17 = _mm256_loadu_si256((const __m256i *)(in + 17 * 32)); + __m256i in18 = _mm256_loadu_si256((const __m256i *)(in + 18 * 32)); + __m256i in19 = _mm256_loadu_si256((const __m256i *)(in + 19 * 32)); + step1[12] = _mm256_add_epi16(in12, in19); + step1[13] = _mm256_add_epi16(in13, in18); + step1[14] = _mm256_add_epi16(in14, in17); + step1[15] = _mm256_add_epi16(in15, in16); + step1[16] = _mm256_sub_epi16(in15, in16); + step1[17] = _mm256_sub_epi16(in14, in17); + step1[18] = _mm256_sub_epi16(in13, in18); + step1[19] = _mm256_sub_epi16(in12, in19); + } + } + // Stage 2 + { + step2[0] = _mm256_add_epi16(step1[0], step1[15]); + step2[1] = _mm256_add_epi16(step1[1], step1[14]); + step2[2] = _mm256_add_epi16(step1[2], step1[13]); + step2[3] = _mm256_add_epi16(step1[3], step1[12]); + step2[4] = _mm256_add_epi16(step1[4], step1[11]); + step2[5] = _mm256_add_epi16(step1[5], step1[10]); + step2[6] = _mm256_add_epi16(step1[6], step1[9]); + step2[7] = _mm256_add_epi16(step1[7], step1[8]); + step2[8] = _mm256_sub_epi16(step1[7], step1[8]); + step2[9] = _mm256_sub_epi16(step1[6], step1[9]); + step2[10] = _mm256_sub_epi16(step1[5], step1[10]); + step2[11] = _mm256_sub_epi16(step1[4], step1[11]); + step2[12] = _mm256_sub_epi16(step1[3], step1[12]); + step2[13] = _mm256_sub_epi16(step1[2], step1[13]); + step2[14] = _mm256_sub_epi16(step1[1], step1[14]); + step2[15] = _mm256_sub_epi16(step1[0], step1[15]); + } + { + const __m256i s2_20_0 = _mm256_unpacklo_epi16(step1[27], step1[20]); + const __m256i s2_20_1 = _mm256_unpackhi_epi16(step1[27], step1[20]); + const __m256i s2_21_0 = _mm256_unpacklo_epi16(step1[26], step1[21]); + const __m256i s2_21_1 = _mm256_unpackhi_epi16(step1[26], step1[21]); + const __m256i s2_22_0 = _mm256_unpacklo_epi16(step1[25], step1[22]); + const __m256i s2_22_1 = _mm256_unpackhi_epi16(step1[25], step1[22]); + const __m256i s2_23_0 = _mm256_unpacklo_epi16(step1[24], step1[23]); + const __m256i s2_23_1 = _mm256_unpackhi_epi16(step1[24], step1[23]); + const __m256i s2_20_2 = _mm256_madd_epi16(s2_20_0, k__cospi_p16_m16); + const __m256i s2_20_3 = _mm256_madd_epi16(s2_20_1, k__cospi_p16_m16); + const __m256i s2_21_2 = _mm256_madd_epi16(s2_21_0, k__cospi_p16_m16); + const __m256i s2_21_3 = _mm256_madd_epi16(s2_21_1, k__cospi_p16_m16); + const __m256i s2_22_2 = _mm256_madd_epi16(s2_22_0, k__cospi_p16_m16); + const __m256i s2_22_3 = _mm256_madd_epi16(s2_22_1, k__cospi_p16_m16); + const __m256i s2_23_2 = _mm256_madd_epi16(s2_23_0, k__cospi_p16_m16); + const __m256i s2_23_3 = _mm256_madd_epi16(s2_23_1, k__cospi_p16_m16); + const __m256i s2_24_2 = _mm256_madd_epi16(s2_23_0, k__cospi_p16_p16); + const __m256i s2_24_3 = _mm256_madd_epi16(s2_23_1, k__cospi_p16_p16); + const __m256i s2_25_2 = _mm256_madd_epi16(s2_22_0, k__cospi_p16_p16); + const __m256i s2_25_3 = _mm256_madd_epi16(s2_22_1, k__cospi_p16_p16); + const __m256i s2_26_2 = _mm256_madd_epi16(s2_21_0, k__cospi_p16_p16); + const __m256i s2_26_3 = _mm256_madd_epi16(s2_21_1, k__cospi_p16_p16); + const __m256i s2_27_2 = _mm256_madd_epi16(s2_20_0, k__cospi_p16_p16); + const __m256i s2_27_3 = _mm256_madd_epi16(s2_20_1, k__cospi_p16_p16); + // dct_const_round_shift + const __m256i s2_20_4 = + _mm256_add_epi32(s2_20_2, k__DCT_CONST_ROUNDING); + const __m256i s2_20_5 = + _mm256_add_epi32(s2_20_3, k__DCT_CONST_ROUNDING); + const __m256i s2_21_4 = + _mm256_add_epi32(s2_21_2, k__DCT_CONST_ROUNDING); + const __m256i s2_21_5 = + _mm256_add_epi32(s2_21_3, k__DCT_CONST_ROUNDING); + const __m256i s2_22_4 = + _mm256_add_epi32(s2_22_2, k__DCT_CONST_ROUNDING); + const __m256i s2_22_5 = + _mm256_add_epi32(s2_22_3, k__DCT_CONST_ROUNDING); + const __m256i s2_23_4 = + _mm256_add_epi32(s2_23_2, k__DCT_CONST_ROUNDING); + const __m256i s2_23_5 = + _mm256_add_epi32(s2_23_3, k__DCT_CONST_ROUNDING); + const __m256i s2_24_4 = + _mm256_add_epi32(s2_24_2, k__DCT_CONST_ROUNDING); + const __m256i s2_24_5 = + _mm256_add_epi32(s2_24_3, k__DCT_CONST_ROUNDING); + const __m256i s2_25_4 = + _mm256_add_epi32(s2_25_2, k__DCT_CONST_ROUNDING); + const __m256i s2_25_5 = + _mm256_add_epi32(s2_25_3, k__DCT_CONST_ROUNDING); + const __m256i s2_26_4 = + _mm256_add_epi32(s2_26_2, k__DCT_CONST_ROUNDING); + const __m256i s2_26_5 = + _mm256_add_epi32(s2_26_3, k__DCT_CONST_ROUNDING); + const __m256i s2_27_4 = + _mm256_add_epi32(s2_27_2, k__DCT_CONST_ROUNDING); + const __m256i s2_27_5 = + _mm256_add_epi32(s2_27_3, k__DCT_CONST_ROUNDING); + const __m256i s2_20_6 = _mm256_srai_epi32(s2_20_4, DCT_CONST_BITS); + const __m256i s2_20_7 = _mm256_srai_epi32(s2_20_5, DCT_CONST_BITS); + const __m256i s2_21_6 = _mm256_srai_epi32(s2_21_4, DCT_CONST_BITS); + const __m256i s2_21_7 = _mm256_srai_epi32(s2_21_5, DCT_CONST_BITS); + const __m256i s2_22_6 = _mm256_srai_epi32(s2_22_4, DCT_CONST_BITS); + const __m256i s2_22_7 = _mm256_srai_epi32(s2_22_5, DCT_CONST_BITS); + const __m256i s2_23_6 = _mm256_srai_epi32(s2_23_4, DCT_CONST_BITS); + const __m256i s2_23_7 = _mm256_srai_epi32(s2_23_5, DCT_CONST_BITS); + const __m256i s2_24_6 = _mm256_srai_epi32(s2_24_4, DCT_CONST_BITS); + const __m256i s2_24_7 = _mm256_srai_epi32(s2_24_5, DCT_CONST_BITS); + const __m256i s2_25_6 = _mm256_srai_epi32(s2_25_4, DCT_CONST_BITS); + const __m256i s2_25_7 = _mm256_srai_epi32(s2_25_5, DCT_CONST_BITS); + const __m256i s2_26_6 = _mm256_srai_epi32(s2_26_4, DCT_CONST_BITS); + const __m256i s2_26_7 = _mm256_srai_epi32(s2_26_5, DCT_CONST_BITS); + const __m256i s2_27_6 = _mm256_srai_epi32(s2_27_4, DCT_CONST_BITS); + const __m256i s2_27_7 = _mm256_srai_epi32(s2_27_5, DCT_CONST_BITS); + // Combine + step2[20] = _mm256_packs_epi32(s2_20_6, s2_20_7); + step2[21] = _mm256_packs_epi32(s2_21_6, s2_21_7); + step2[22] = _mm256_packs_epi32(s2_22_6, s2_22_7); + step2[23] = _mm256_packs_epi32(s2_23_6, s2_23_7); + step2[24] = _mm256_packs_epi32(s2_24_6, s2_24_7); + step2[25] = _mm256_packs_epi32(s2_25_6, s2_25_7); + step2[26] = _mm256_packs_epi32(s2_26_6, s2_26_7); + step2[27] = _mm256_packs_epi32(s2_27_6, s2_27_7); + } + +#if !FDCT32x32_HIGH_PRECISION + // dump the magnitude by half, hence the intermediate values are within + // the range of 16 bits. + if (1 == pass) { + __m256i s3_00_0 = _mm256_cmpgt_epi16(kZero, step2[0]); + __m256i s3_01_0 = _mm256_cmpgt_epi16(kZero, step2[1]); + __m256i s3_02_0 = _mm256_cmpgt_epi16(kZero, step2[2]); + __m256i s3_03_0 = _mm256_cmpgt_epi16(kZero, step2[3]); + __m256i s3_04_0 = _mm256_cmpgt_epi16(kZero, step2[4]); + __m256i s3_05_0 = _mm256_cmpgt_epi16(kZero, step2[5]); + __m256i s3_06_0 = _mm256_cmpgt_epi16(kZero, step2[6]); + __m256i s3_07_0 = _mm256_cmpgt_epi16(kZero, step2[7]); + __m256i s2_08_0 = _mm256_cmpgt_epi16(kZero, step2[8]); + __m256i s2_09_0 = _mm256_cmpgt_epi16(kZero, step2[9]); + __m256i s3_10_0 = _mm256_cmpgt_epi16(kZero, step2[10]); + __m256i s3_11_0 = _mm256_cmpgt_epi16(kZero, step2[11]); + __m256i s3_12_0 = _mm256_cmpgt_epi16(kZero, step2[12]); + __m256i s3_13_0 = _mm256_cmpgt_epi16(kZero, step2[13]); + __m256i s2_14_0 = _mm256_cmpgt_epi16(kZero, step2[14]); + __m256i s2_15_0 = _mm256_cmpgt_epi16(kZero, step2[15]); + __m256i s3_16_0 = _mm256_cmpgt_epi16(kZero, step1[16]); + __m256i s3_17_0 = _mm256_cmpgt_epi16(kZero, step1[17]); + __m256i s3_18_0 = _mm256_cmpgt_epi16(kZero, step1[18]); + __m256i s3_19_0 = _mm256_cmpgt_epi16(kZero, step1[19]); + __m256i s3_20_0 = _mm256_cmpgt_epi16(kZero, step2[20]); + __m256i s3_21_0 = _mm256_cmpgt_epi16(kZero, step2[21]); + __m256i s3_22_0 = _mm256_cmpgt_epi16(kZero, step2[22]); + __m256i s3_23_0 = _mm256_cmpgt_epi16(kZero, step2[23]); + __m256i s3_24_0 = _mm256_cmpgt_epi16(kZero, step2[24]); + __m256i s3_25_0 = _mm256_cmpgt_epi16(kZero, step2[25]); + __m256i s3_26_0 = _mm256_cmpgt_epi16(kZero, step2[26]); + __m256i s3_27_0 = _mm256_cmpgt_epi16(kZero, step2[27]); + __m256i s3_28_0 = _mm256_cmpgt_epi16(kZero, step1[28]); + __m256i s3_29_0 = _mm256_cmpgt_epi16(kZero, step1[29]); + __m256i s3_30_0 = _mm256_cmpgt_epi16(kZero, step1[30]); + __m256i s3_31_0 = _mm256_cmpgt_epi16(kZero, step1[31]); + + step2[0] = _mm256_sub_epi16(step2[0], s3_00_0); + step2[1] = _mm256_sub_epi16(step2[1], s3_01_0); + step2[2] = _mm256_sub_epi16(step2[2], s3_02_0); + step2[3] = _mm256_sub_epi16(step2[3], s3_03_0); + step2[4] = _mm256_sub_epi16(step2[4], s3_04_0); + step2[5] = _mm256_sub_epi16(step2[5], s3_05_0); + step2[6] = _mm256_sub_epi16(step2[6], s3_06_0); + step2[7] = _mm256_sub_epi16(step2[7], s3_07_0); + step2[8] = _mm256_sub_epi16(step2[8], s2_08_0); + step2[9] = _mm256_sub_epi16(step2[9], s2_09_0); + step2[10] = _mm256_sub_epi16(step2[10], s3_10_0); + step2[11] = _mm256_sub_epi16(step2[11], s3_11_0); + step2[12] = _mm256_sub_epi16(step2[12], s3_12_0); + step2[13] = _mm256_sub_epi16(step2[13], s3_13_0); + step2[14] = _mm256_sub_epi16(step2[14], s2_14_0); + step2[15] = _mm256_sub_epi16(step2[15], s2_15_0); + step1[16] = _mm256_sub_epi16(step1[16], s3_16_0); + step1[17] = _mm256_sub_epi16(step1[17], s3_17_0); + step1[18] = _mm256_sub_epi16(step1[18], s3_18_0); + step1[19] = _mm256_sub_epi16(step1[19], s3_19_0); + step2[20] = _mm256_sub_epi16(step2[20], s3_20_0); + step2[21] = _mm256_sub_epi16(step2[21], s3_21_0); + step2[22] = _mm256_sub_epi16(step2[22], s3_22_0); + step2[23] = _mm256_sub_epi16(step2[23], s3_23_0); + step2[24] = _mm256_sub_epi16(step2[24], s3_24_0); + step2[25] = _mm256_sub_epi16(step2[25], s3_25_0); + step2[26] = _mm256_sub_epi16(step2[26], s3_26_0); + step2[27] = _mm256_sub_epi16(step2[27], s3_27_0); + step1[28] = _mm256_sub_epi16(step1[28], s3_28_0); + step1[29] = _mm256_sub_epi16(step1[29], s3_29_0); + step1[30] = _mm256_sub_epi16(step1[30], s3_30_0); + step1[31] = _mm256_sub_epi16(step1[31], s3_31_0); + + step2[0] = _mm256_add_epi16(step2[0], kOne); + step2[1] = _mm256_add_epi16(step2[1], kOne); + step2[2] = _mm256_add_epi16(step2[2], kOne); + step2[3] = _mm256_add_epi16(step2[3], kOne); + step2[4] = _mm256_add_epi16(step2[4], kOne); + step2[5] = _mm256_add_epi16(step2[5], kOne); + step2[6] = _mm256_add_epi16(step2[6], kOne); + step2[7] = _mm256_add_epi16(step2[7], kOne); + step2[8] = _mm256_add_epi16(step2[8], kOne); + step2[9] = _mm256_add_epi16(step2[9], kOne); + step2[10] = _mm256_add_epi16(step2[10], kOne); + step2[11] = _mm256_add_epi16(step2[11], kOne); + step2[12] = _mm256_add_epi16(step2[12], kOne); + step2[13] = _mm256_add_epi16(step2[13], kOne); + step2[14] = _mm256_add_epi16(step2[14], kOne); + step2[15] = _mm256_add_epi16(step2[15], kOne); + step1[16] = _mm256_add_epi16(step1[16], kOne); + step1[17] = _mm256_add_epi16(step1[17], kOne); + step1[18] = _mm256_add_epi16(step1[18], kOne); + step1[19] = _mm256_add_epi16(step1[19], kOne); + step2[20] = _mm256_add_epi16(step2[20], kOne); + step2[21] = _mm256_add_epi16(step2[21], kOne); + step2[22] = _mm256_add_epi16(step2[22], kOne); + step2[23] = _mm256_add_epi16(step2[23], kOne); + step2[24] = _mm256_add_epi16(step2[24], kOne); + step2[25] = _mm256_add_epi16(step2[25], kOne); + step2[26] = _mm256_add_epi16(step2[26], kOne); + step2[27] = _mm256_add_epi16(step2[27], kOne); + step1[28] = _mm256_add_epi16(step1[28], kOne); + step1[29] = _mm256_add_epi16(step1[29], kOne); + step1[30] = _mm256_add_epi16(step1[30], kOne); + step1[31] = _mm256_add_epi16(step1[31], kOne); + + step2[0] = _mm256_srai_epi16(step2[0], 2); + step2[1] = _mm256_srai_epi16(step2[1], 2); + step2[2] = _mm256_srai_epi16(step2[2], 2); + step2[3] = _mm256_srai_epi16(step2[3], 2); + step2[4] = _mm256_srai_epi16(step2[4], 2); + step2[5] = _mm256_srai_epi16(step2[5], 2); + step2[6] = _mm256_srai_epi16(step2[6], 2); + step2[7] = _mm256_srai_epi16(step2[7], 2); + step2[8] = _mm256_srai_epi16(step2[8], 2); + step2[9] = _mm256_srai_epi16(step2[9], 2); + step2[10] = _mm256_srai_epi16(step2[10], 2); + step2[11] = _mm256_srai_epi16(step2[11], 2); + step2[12] = _mm256_srai_epi16(step2[12], 2); + step2[13] = _mm256_srai_epi16(step2[13], 2); + step2[14] = _mm256_srai_epi16(step2[14], 2); + step2[15] = _mm256_srai_epi16(step2[15], 2); + step1[16] = _mm256_srai_epi16(step1[16], 2); + step1[17] = _mm256_srai_epi16(step1[17], 2); + step1[18] = _mm256_srai_epi16(step1[18], 2); + step1[19] = _mm256_srai_epi16(step1[19], 2); + step2[20] = _mm256_srai_epi16(step2[20], 2); + step2[21] = _mm256_srai_epi16(step2[21], 2); + step2[22] = _mm256_srai_epi16(step2[22], 2); + step2[23] = _mm256_srai_epi16(step2[23], 2); + step2[24] = _mm256_srai_epi16(step2[24], 2); + step2[25] = _mm256_srai_epi16(step2[25], 2); + step2[26] = _mm256_srai_epi16(step2[26], 2); + step2[27] = _mm256_srai_epi16(step2[27], 2); + step1[28] = _mm256_srai_epi16(step1[28], 2); + step1[29] = _mm256_srai_epi16(step1[29], 2); + step1[30] = _mm256_srai_epi16(step1[30], 2); + step1[31] = _mm256_srai_epi16(step1[31], 2); + } +#endif + +#if FDCT32x32_HIGH_PRECISION + if (pass == 0) { +#endif + // Stage 3 + { + step3[0] = _mm256_add_epi16(step2[(8 - 1)], step2[0]); + step3[1] = _mm256_add_epi16(step2[(8 - 2)], step2[1]); + step3[2] = _mm256_add_epi16(step2[(8 - 3)], step2[2]); + step3[3] = _mm256_add_epi16(step2[(8 - 4)], step2[3]); + step3[4] = _mm256_sub_epi16(step2[(8 - 5)], step2[4]); + step3[5] = _mm256_sub_epi16(step2[(8 - 6)], step2[5]); + step3[6] = _mm256_sub_epi16(step2[(8 - 7)], step2[6]); + step3[7] = _mm256_sub_epi16(step2[(8 - 8)], step2[7]); + } + { + const __m256i s3_10_0 = _mm256_unpacklo_epi16(step2[13], step2[10]); + const __m256i s3_10_1 = _mm256_unpackhi_epi16(step2[13], step2[10]); + const __m256i s3_11_0 = _mm256_unpacklo_epi16(step2[12], step2[11]); + const __m256i s3_11_1 = _mm256_unpackhi_epi16(step2[12], step2[11]); + const __m256i s3_10_2 = _mm256_madd_epi16(s3_10_0, k__cospi_p16_m16); + const __m256i s3_10_3 = _mm256_madd_epi16(s3_10_1, k__cospi_p16_m16); + const __m256i s3_11_2 = _mm256_madd_epi16(s3_11_0, k__cospi_p16_m16); + const __m256i s3_11_3 = _mm256_madd_epi16(s3_11_1, k__cospi_p16_m16); + const __m256i s3_12_2 = _mm256_madd_epi16(s3_11_0, k__cospi_p16_p16); + const __m256i s3_12_3 = _mm256_madd_epi16(s3_11_1, k__cospi_p16_p16); + const __m256i s3_13_2 = _mm256_madd_epi16(s3_10_0, k__cospi_p16_p16); + const __m256i s3_13_3 = _mm256_madd_epi16(s3_10_1, k__cospi_p16_p16); + // dct_const_round_shift + const __m256i s3_10_4 = + _mm256_add_epi32(s3_10_2, k__DCT_CONST_ROUNDING); + const __m256i s3_10_5 = + _mm256_add_epi32(s3_10_3, k__DCT_CONST_ROUNDING); + const __m256i s3_11_4 = + _mm256_add_epi32(s3_11_2, k__DCT_CONST_ROUNDING); + const __m256i s3_11_5 = + _mm256_add_epi32(s3_11_3, k__DCT_CONST_ROUNDING); + const __m256i s3_12_4 = + _mm256_add_epi32(s3_12_2, k__DCT_CONST_ROUNDING); + const __m256i s3_12_5 = + _mm256_add_epi32(s3_12_3, k__DCT_CONST_ROUNDING); + const __m256i s3_13_4 = + _mm256_add_epi32(s3_13_2, k__DCT_CONST_ROUNDING); + const __m256i s3_13_5 = + _mm256_add_epi32(s3_13_3, k__DCT_CONST_ROUNDING); + const __m256i s3_10_6 = _mm256_srai_epi32(s3_10_4, DCT_CONST_BITS); + const __m256i s3_10_7 = _mm256_srai_epi32(s3_10_5, DCT_CONST_BITS); + const __m256i s3_11_6 = _mm256_srai_epi32(s3_11_4, DCT_CONST_BITS); + const __m256i s3_11_7 = _mm256_srai_epi32(s3_11_5, DCT_CONST_BITS); + const __m256i s3_12_6 = _mm256_srai_epi32(s3_12_4, DCT_CONST_BITS); + const __m256i s3_12_7 = _mm256_srai_epi32(s3_12_5, DCT_CONST_BITS); + const __m256i s3_13_6 = _mm256_srai_epi32(s3_13_4, DCT_CONST_BITS); + const __m256i s3_13_7 = _mm256_srai_epi32(s3_13_5, DCT_CONST_BITS); + // Combine + step3[10] = _mm256_packs_epi32(s3_10_6, s3_10_7); + step3[11] = _mm256_packs_epi32(s3_11_6, s3_11_7); + step3[12] = _mm256_packs_epi32(s3_12_6, s3_12_7); + step3[13] = _mm256_packs_epi32(s3_13_6, s3_13_7); + } + { + step3[16] = _mm256_add_epi16(step2[23], step1[16]); + step3[17] = _mm256_add_epi16(step2[22], step1[17]); + step3[18] = _mm256_add_epi16(step2[21], step1[18]); + step3[19] = _mm256_add_epi16(step2[20], step1[19]); + step3[20] = _mm256_sub_epi16(step1[19], step2[20]); + step3[21] = _mm256_sub_epi16(step1[18], step2[21]); + step3[22] = _mm256_sub_epi16(step1[17], step2[22]); + step3[23] = _mm256_sub_epi16(step1[16], step2[23]); + step3[24] = _mm256_sub_epi16(step1[31], step2[24]); + step3[25] = _mm256_sub_epi16(step1[30], step2[25]); + step3[26] = _mm256_sub_epi16(step1[29], step2[26]); + step3[27] = _mm256_sub_epi16(step1[28], step2[27]); + step3[28] = _mm256_add_epi16(step2[27], step1[28]); + step3[29] = _mm256_add_epi16(step2[26], step1[29]); + step3[30] = _mm256_add_epi16(step2[25], step1[30]); + step3[31] = _mm256_add_epi16(step2[24], step1[31]); + } + + // Stage 4 + { + step1[0] = _mm256_add_epi16(step3[3], step3[0]); + step1[1] = _mm256_add_epi16(step3[2], step3[1]); + step1[2] = _mm256_sub_epi16(step3[1], step3[2]); + step1[3] = _mm256_sub_epi16(step3[0], step3[3]); + step1[8] = _mm256_add_epi16(step3[11], step2[8]); + step1[9] = _mm256_add_epi16(step3[10], step2[9]); + step1[10] = _mm256_sub_epi16(step2[9], step3[10]); + step1[11] = _mm256_sub_epi16(step2[8], step3[11]); + step1[12] = _mm256_sub_epi16(step2[15], step3[12]); + step1[13] = _mm256_sub_epi16(step2[14], step3[13]); + step1[14] = _mm256_add_epi16(step3[13], step2[14]); + step1[15] = _mm256_add_epi16(step3[12], step2[15]); + } + { + const __m256i s1_05_0 = _mm256_unpacklo_epi16(step3[6], step3[5]); + const __m256i s1_05_1 = _mm256_unpackhi_epi16(step3[6], step3[5]); + const __m256i s1_05_2 = _mm256_madd_epi16(s1_05_0, k__cospi_p16_m16); + const __m256i s1_05_3 = _mm256_madd_epi16(s1_05_1, k__cospi_p16_m16); + const __m256i s1_06_2 = _mm256_madd_epi16(s1_05_0, k__cospi_p16_p16); + const __m256i s1_06_3 = _mm256_madd_epi16(s1_05_1, k__cospi_p16_p16); + // dct_const_round_shift + const __m256i s1_05_4 = + _mm256_add_epi32(s1_05_2, k__DCT_CONST_ROUNDING); + const __m256i s1_05_5 = + _mm256_add_epi32(s1_05_3, k__DCT_CONST_ROUNDING); + const __m256i s1_06_4 = + _mm256_add_epi32(s1_06_2, k__DCT_CONST_ROUNDING); + const __m256i s1_06_5 = + _mm256_add_epi32(s1_06_3, k__DCT_CONST_ROUNDING); + const __m256i s1_05_6 = _mm256_srai_epi32(s1_05_4, DCT_CONST_BITS); + const __m256i s1_05_7 = _mm256_srai_epi32(s1_05_5, DCT_CONST_BITS); + const __m256i s1_06_6 = _mm256_srai_epi32(s1_06_4, DCT_CONST_BITS); + const __m256i s1_06_7 = _mm256_srai_epi32(s1_06_5, DCT_CONST_BITS); + // Combine + step1[5] = _mm256_packs_epi32(s1_05_6, s1_05_7); + step1[6] = _mm256_packs_epi32(s1_06_6, s1_06_7); + } + { + const __m256i s1_18_0 = _mm256_unpacklo_epi16(step3[18], step3[29]); + const __m256i s1_18_1 = _mm256_unpackhi_epi16(step3[18], step3[29]); + const __m256i s1_19_0 = _mm256_unpacklo_epi16(step3[19], step3[28]); + const __m256i s1_19_1 = _mm256_unpackhi_epi16(step3[19], step3[28]); + const __m256i s1_20_0 = _mm256_unpacklo_epi16(step3[20], step3[27]); + const __m256i s1_20_1 = _mm256_unpackhi_epi16(step3[20], step3[27]); + const __m256i s1_21_0 = _mm256_unpacklo_epi16(step3[21], step3[26]); + const __m256i s1_21_1 = _mm256_unpackhi_epi16(step3[21], step3[26]); + const __m256i s1_18_2 = _mm256_madd_epi16(s1_18_0, k__cospi_m08_p24); + const __m256i s1_18_3 = _mm256_madd_epi16(s1_18_1, k__cospi_m08_p24); + const __m256i s1_19_2 = _mm256_madd_epi16(s1_19_0, k__cospi_m08_p24); + const __m256i s1_19_3 = _mm256_madd_epi16(s1_19_1, k__cospi_m08_p24); + const __m256i s1_20_2 = _mm256_madd_epi16(s1_20_0, k__cospi_m24_m08); + const __m256i s1_20_3 = _mm256_madd_epi16(s1_20_1, k__cospi_m24_m08); + const __m256i s1_21_2 = _mm256_madd_epi16(s1_21_0, k__cospi_m24_m08); + const __m256i s1_21_3 = _mm256_madd_epi16(s1_21_1, k__cospi_m24_m08); + const __m256i s1_26_2 = _mm256_madd_epi16(s1_21_0, k__cospi_m08_p24); + const __m256i s1_26_3 = _mm256_madd_epi16(s1_21_1, k__cospi_m08_p24); + const __m256i s1_27_2 = _mm256_madd_epi16(s1_20_0, k__cospi_m08_p24); + const __m256i s1_27_3 = _mm256_madd_epi16(s1_20_1, k__cospi_m08_p24); + const __m256i s1_28_2 = _mm256_madd_epi16(s1_19_0, k__cospi_p24_p08); + const __m256i s1_28_3 = _mm256_madd_epi16(s1_19_1, k__cospi_p24_p08); + const __m256i s1_29_2 = _mm256_madd_epi16(s1_18_0, k__cospi_p24_p08); + const __m256i s1_29_3 = _mm256_madd_epi16(s1_18_1, k__cospi_p24_p08); + // dct_const_round_shift + const __m256i s1_18_4 = + _mm256_add_epi32(s1_18_2, k__DCT_CONST_ROUNDING); + const __m256i s1_18_5 = + _mm256_add_epi32(s1_18_3, k__DCT_CONST_ROUNDING); + const __m256i s1_19_4 = + _mm256_add_epi32(s1_19_2, k__DCT_CONST_ROUNDING); + const __m256i s1_19_5 = + _mm256_add_epi32(s1_19_3, k__DCT_CONST_ROUNDING); + const __m256i s1_20_4 = + _mm256_add_epi32(s1_20_2, k__DCT_CONST_ROUNDING); + const __m256i s1_20_5 = + _mm256_add_epi32(s1_20_3, k__DCT_CONST_ROUNDING); + const __m256i s1_21_4 = + _mm256_add_epi32(s1_21_2, k__DCT_CONST_ROUNDING); + const __m256i s1_21_5 = + _mm256_add_epi32(s1_21_3, k__DCT_CONST_ROUNDING); + const __m256i s1_26_4 = + _mm256_add_epi32(s1_26_2, k__DCT_CONST_ROUNDING); + const __m256i s1_26_5 = + _mm256_add_epi32(s1_26_3, k__DCT_CONST_ROUNDING); + const __m256i s1_27_4 = + _mm256_add_epi32(s1_27_2, k__DCT_CONST_ROUNDING); + const __m256i s1_27_5 = + _mm256_add_epi32(s1_27_3, k__DCT_CONST_ROUNDING); + const __m256i s1_28_4 = + _mm256_add_epi32(s1_28_2, k__DCT_CONST_ROUNDING); + const __m256i s1_28_5 = + _mm256_add_epi32(s1_28_3, k__DCT_CONST_ROUNDING); + const __m256i s1_29_4 = + _mm256_add_epi32(s1_29_2, k__DCT_CONST_ROUNDING); + const __m256i s1_29_5 = + _mm256_add_epi32(s1_29_3, k__DCT_CONST_ROUNDING); + const __m256i s1_18_6 = _mm256_srai_epi32(s1_18_4, DCT_CONST_BITS); + const __m256i s1_18_7 = _mm256_srai_epi32(s1_18_5, DCT_CONST_BITS); + const __m256i s1_19_6 = _mm256_srai_epi32(s1_19_4, DCT_CONST_BITS); + const __m256i s1_19_7 = _mm256_srai_epi32(s1_19_5, DCT_CONST_BITS); + const __m256i s1_20_6 = _mm256_srai_epi32(s1_20_4, DCT_CONST_BITS); + const __m256i s1_20_7 = _mm256_srai_epi32(s1_20_5, DCT_CONST_BITS); + const __m256i s1_21_6 = _mm256_srai_epi32(s1_21_4, DCT_CONST_BITS); + const __m256i s1_21_7 = _mm256_srai_epi32(s1_21_5, DCT_CONST_BITS); + const __m256i s1_26_6 = _mm256_srai_epi32(s1_26_4, DCT_CONST_BITS); + const __m256i s1_26_7 = _mm256_srai_epi32(s1_26_5, DCT_CONST_BITS); + const __m256i s1_27_6 = _mm256_srai_epi32(s1_27_4, DCT_CONST_BITS); + const __m256i s1_27_7 = _mm256_srai_epi32(s1_27_5, DCT_CONST_BITS); + const __m256i s1_28_6 = _mm256_srai_epi32(s1_28_4, DCT_CONST_BITS); + const __m256i s1_28_7 = _mm256_srai_epi32(s1_28_5, DCT_CONST_BITS); + const __m256i s1_29_6 = _mm256_srai_epi32(s1_29_4, DCT_CONST_BITS); + const __m256i s1_29_7 = _mm256_srai_epi32(s1_29_5, DCT_CONST_BITS); + // Combine + step1[18] = _mm256_packs_epi32(s1_18_6, s1_18_7); + step1[19] = _mm256_packs_epi32(s1_19_6, s1_19_7); + step1[20] = _mm256_packs_epi32(s1_20_6, s1_20_7); + step1[21] = _mm256_packs_epi32(s1_21_6, s1_21_7); + step1[26] = _mm256_packs_epi32(s1_26_6, s1_26_7); + step1[27] = _mm256_packs_epi32(s1_27_6, s1_27_7); + step1[28] = _mm256_packs_epi32(s1_28_6, s1_28_7); + step1[29] = _mm256_packs_epi32(s1_29_6, s1_29_7); + } + // Stage 5 + { + step2[4] = _mm256_add_epi16(step1[5], step3[4]); + step2[5] = _mm256_sub_epi16(step3[4], step1[5]); + step2[6] = _mm256_sub_epi16(step3[7], step1[6]); + step2[7] = _mm256_add_epi16(step1[6], step3[7]); + } + { + const __m256i out_00_0 = _mm256_unpacklo_epi16(step1[0], step1[1]); + const __m256i out_00_1 = _mm256_unpackhi_epi16(step1[0], step1[1]); + const __m256i out_08_0 = _mm256_unpacklo_epi16(step1[2], step1[3]); + const __m256i out_08_1 = _mm256_unpackhi_epi16(step1[2], step1[3]); + const __m256i out_00_2 = + _mm256_madd_epi16(out_00_0, k__cospi_p16_p16); + const __m256i out_00_3 = + _mm256_madd_epi16(out_00_1, k__cospi_p16_p16); + const __m256i out_16_2 = + _mm256_madd_epi16(out_00_0, k__cospi_p16_m16); + const __m256i out_16_3 = + _mm256_madd_epi16(out_00_1, k__cospi_p16_m16); + const __m256i out_08_2 = + _mm256_madd_epi16(out_08_0, k__cospi_p24_p08); + const __m256i out_08_3 = + _mm256_madd_epi16(out_08_1, k__cospi_p24_p08); + const __m256i out_24_2 = + _mm256_madd_epi16(out_08_0, k__cospi_m08_p24); + const __m256i out_24_3 = + _mm256_madd_epi16(out_08_1, k__cospi_m08_p24); + // dct_const_round_shift + const __m256i out_00_4 = + _mm256_add_epi32(out_00_2, k__DCT_CONST_ROUNDING); + const __m256i out_00_5 = + _mm256_add_epi32(out_00_3, k__DCT_CONST_ROUNDING); + const __m256i out_16_4 = + _mm256_add_epi32(out_16_2, k__DCT_CONST_ROUNDING); + const __m256i out_16_5 = + _mm256_add_epi32(out_16_3, k__DCT_CONST_ROUNDING); + const __m256i out_08_4 = + _mm256_add_epi32(out_08_2, k__DCT_CONST_ROUNDING); + const __m256i out_08_5 = + _mm256_add_epi32(out_08_3, k__DCT_CONST_ROUNDING); + const __m256i out_24_4 = + _mm256_add_epi32(out_24_2, k__DCT_CONST_ROUNDING); + const __m256i out_24_5 = + _mm256_add_epi32(out_24_3, k__DCT_CONST_ROUNDING); + const __m256i out_00_6 = _mm256_srai_epi32(out_00_4, DCT_CONST_BITS); + const __m256i out_00_7 = _mm256_srai_epi32(out_00_5, DCT_CONST_BITS); + const __m256i out_16_6 = _mm256_srai_epi32(out_16_4, DCT_CONST_BITS); + const __m256i out_16_7 = _mm256_srai_epi32(out_16_5, DCT_CONST_BITS); + const __m256i out_08_6 = _mm256_srai_epi32(out_08_4, DCT_CONST_BITS); + const __m256i out_08_7 = _mm256_srai_epi32(out_08_5, DCT_CONST_BITS); + const __m256i out_24_6 = _mm256_srai_epi32(out_24_4, DCT_CONST_BITS); + const __m256i out_24_7 = _mm256_srai_epi32(out_24_5, DCT_CONST_BITS); + // Combine + out[0] = _mm256_packs_epi32(out_00_6, out_00_7); + out[16] = _mm256_packs_epi32(out_16_6, out_16_7); + out[8] = _mm256_packs_epi32(out_08_6, out_08_7); + out[24] = _mm256_packs_epi32(out_24_6, out_24_7); + } + { + const __m256i s2_09_0 = _mm256_unpacklo_epi16(step1[9], step1[14]); + const __m256i s2_09_1 = _mm256_unpackhi_epi16(step1[9], step1[14]); + const __m256i s2_10_0 = _mm256_unpacklo_epi16(step1[10], step1[13]); + const __m256i s2_10_1 = _mm256_unpackhi_epi16(step1[10], step1[13]); + const __m256i s2_09_2 = _mm256_madd_epi16(s2_09_0, k__cospi_m08_p24); + const __m256i s2_09_3 = _mm256_madd_epi16(s2_09_1, k__cospi_m08_p24); + const __m256i s2_10_2 = _mm256_madd_epi16(s2_10_0, k__cospi_m24_m08); + const __m256i s2_10_3 = _mm256_madd_epi16(s2_10_1, k__cospi_m24_m08); + const __m256i s2_13_2 = _mm256_madd_epi16(s2_10_0, k__cospi_m08_p24); + const __m256i s2_13_3 = _mm256_madd_epi16(s2_10_1, k__cospi_m08_p24); + const __m256i s2_14_2 = _mm256_madd_epi16(s2_09_0, k__cospi_p24_p08); + const __m256i s2_14_3 = _mm256_madd_epi16(s2_09_1, k__cospi_p24_p08); + // dct_const_round_shift + const __m256i s2_09_4 = + _mm256_add_epi32(s2_09_2, k__DCT_CONST_ROUNDING); + const __m256i s2_09_5 = + _mm256_add_epi32(s2_09_3, k__DCT_CONST_ROUNDING); + const __m256i s2_10_4 = + _mm256_add_epi32(s2_10_2, k__DCT_CONST_ROUNDING); + const __m256i s2_10_5 = + _mm256_add_epi32(s2_10_3, k__DCT_CONST_ROUNDING); + const __m256i s2_13_4 = + _mm256_add_epi32(s2_13_2, k__DCT_CONST_ROUNDING); + const __m256i s2_13_5 = + _mm256_add_epi32(s2_13_3, k__DCT_CONST_ROUNDING); + const __m256i s2_14_4 = + _mm256_add_epi32(s2_14_2, k__DCT_CONST_ROUNDING); + const __m256i s2_14_5 = + _mm256_add_epi32(s2_14_3, k__DCT_CONST_ROUNDING); + const __m256i s2_09_6 = _mm256_srai_epi32(s2_09_4, DCT_CONST_BITS); + const __m256i s2_09_7 = _mm256_srai_epi32(s2_09_5, DCT_CONST_BITS); + const __m256i s2_10_6 = _mm256_srai_epi32(s2_10_4, DCT_CONST_BITS); + const __m256i s2_10_7 = _mm256_srai_epi32(s2_10_5, DCT_CONST_BITS); + const __m256i s2_13_6 = _mm256_srai_epi32(s2_13_4, DCT_CONST_BITS); + const __m256i s2_13_7 = _mm256_srai_epi32(s2_13_5, DCT_CONST_BITS); + const __m256i s2_14_6 = _mm256_srai_epi32(s2_14_4, DCT_CONST_BITS); + const __m256i s2_14_7 = _mm256_srai_epi32(s2_14_5, DCT_CONST_BITS); + // Combine + step2[9] = _mm256_packs_epi32(s2_09_6, s2_09_7); + step2[10] = _mm256_packs_epi32(s2_10_6, s2_10_7); + step2[13] = _mm256_packs_epi32(s2_13_6, s2_13_7); + step2[14] = _mm256_packs_epi32(s2_14_6, s2_14_7); + } + { + step2[16] = _mm256_add_epi16(step1[19], step3[16]); + step2[17] = _mm256_add_epi16(step1[18], step3[17]); + step2[18] = _mm256_sub_epi16(step3[17], step1[18]); + step2[19] = _mm256_sub_epi16(step3[16], step1[19]); + step2[20] = _mm256_sub_epi16(step3[23], step1[20]); + step2[21] = _mm256_sub_epi16(step3[22], step1[21]); + step2[22] = _mm256_add_epi16(step1[21], step3[22]); + step2[23] = _mm256_add_epi16(step1[20], step3[23]); + step2[24] = _mm256_add_epi16(step1[27], step3[24]); + step2[25] = _mm256_add_epi16(step1[26], step3[25]); + step2[26] = _mm256_sub_epi16(step3[25], step1[26]); + step2[27] = _mm256_sub_epi16(step3[24], step1[27]); + step2[28] = _mm256_sub_epi16(step3[31], step1[28]); + step2[29] = _mm256_sub_epi16(step3[30], step1[29]); + step2[30] = _mm256_add_epi16(step1[29], step3[30]); + step2[31] = _mm256_add_epi16(step1[28], step3[31]); + } + // Stage 6 + { + const __m256i out_04_0 = _mm256_unpacklo_epi16(step2[4], step2[7]); + const __m256i out_04_1 = _mm256_unpackhi_epi16(step2[4], step2[7]); + const __m256i out_20_0 = _mm256_unpacklo_epi16(step2[5], step2[6]); + const __m256i out_20_1 = _mm256_unpackhi_epi16(step2[5], step2[6]); + const __m256i out_12_0 = _mm256_unpacklo_epi16(step2[5], step2[6]); + const __m256i out_12_1 = _mm256_unpackhi_epi16(step2[5], step2[6]); + const __m256i out_28_0 = _mm256_unpacklo_epi16(step2[4], step2[7]); + const __m256i out_28_1 = _mm256_unpackhi_epi16(step2[4], step2[7]); + const __m256i out_04_2 = + _mm256_madd_epi16(out_04_0, k__cospi_p28_p04); + const __m256i out_04_3 = + _mm256_madd_epi16(out_04_1, k__cospi_p28_p04); + const __m256i out_20_2 = + _mm256_madd_epi16(out_20_0, k__cospi_p12_p20); + const __m256i out_20_3 = + _mm256_madd_epi16(out_20_1, k__cospi_p12_p20); + const __m256i out_12_2 = + _mm256_madd_epi16(out_12_0, k__cospi_m20_p12); + const __m256i out_12_3 = + _mm256_madd_epi16(out_12_1, k__cospi_m20_p12); + const __m256i out_28_2 = + _mm256_madd_epi16(out_28_0, k__cospi_m04_p28); + const __m256i out_28_3 = + _mm256_madd_epi16(out_28_1, k__cospi_m04_p28); + // dct_const_round_shift + const __m256i out_04_4 = + _mm256_add_epi32(out_04_2, k__DCT_CONST_ROUNDING); + const __m256i out_04_5 = + _mm256_add_epi32(out_04_3, k__DCT_CONST_ROUNDING); + const __m256i out_20_4 = + _mm256_add_epi32(out_20_2, k__DCT_CONST_ROUNDING); + const __m256i out_20_5 = + _mm256_add_epi32(out_20_3, k__DCT_CONST_ROUNDING); + const __m256i out_12_4 = + _mm256_add_epi32(out_12_2, k__DCT_CONST_ROUNDING); + const __m256i out_12_5 = + _mm256_add_epi32(out_12_3, k__DCT_CONST_ROUNDING); + const __m256i out_28_4 = + _mm256_add_epi32(out_28_2, k__DCT_CONST_ROUNDING); + const __m256i out_28_5 = + _mm256_add_epi32(out_28_3, k__DCT_CONST_ROUNDING); + const __m256i out_04_6 = _mm256_srai_epi32(out_04_4, DCT_CONST_BITS); + const __m256i out_04_7 = _mm256_srai_epi32(out_04_5, DCT_CONST_BITS); + const __m256i out_20_6 = _mm256_srai_epi32(out_20_4, DCT_CONST_BITS); + const __m256i out_20_7 = _mm256_srai_epi32(out_20_5, DCT_CONST_BITS); + const __m256i out_12_6 = _mm256_srai_epi32(out_12_4, DCT_CONST_BITS); + const __m256i out_12_7 = _mm256_srai_epi32(out_12_5, DCT_CONST_BITS); + const __m256i out_28_6 = _mm256_srai_epi32(out_28_4, DCT_CONST_BITS); + const __m256i out_28_7 = _mm256_srai_epi32(out_28_5, DCT_CONST_BITS); + // Combine + out[4] = _mm256_packs_epi32(out_04_6, out_04_7); + out[20] = _mm256_packs_epi32(out_20_6, out_20_7); + out[12] = _mm256_packs_epi32(out_12_6, out_12_7); + out[28] = _mm256_packs_epi32(out_28_6, out_28_7); + } + { + step3[8] = _mm256_add_epi16(step2[9], step1[8]); + step3[9] = _mm256_sub_epi16(step1[8], step2[9]); + step3[10] = _mm256_sub_epi16(step1[11], step2[10]); + step3[11] = _mm256_add_epi16(step2[10], step1[11]); + step3[12] = _mm256_add_epi16(step2[13], step1[12]); + step3[13] = _mm256_sub_epi16(step1[12], step2[13]); + step3[14] = _mm256_sub_epi16(step1[15], step2[14]); + step3[15] = _mm256_add_epi16(step2[14], step1[15]); + } + { + const __m256i s3_17_0 = _mm256_unpacklo_epi16(step2[17], step2[30]); + const __m256i s3_17_1 = _mm256_unpackhi_epi16(step2[17], step2[30]); + const __m256i s3_18_0 = _mm256_unpacklo_epi16(step2[18], step2[29]); + const __m256i s3_18_1 = _mm256_unpackhi_epi16(step2[18], step2[29]); + const __m256i s3_21_0 = _mm256_unpacklo_epi16(step2[21], step2[26]); + const __m256i s3_21_1 = _mm256_unpackhi_epi16(step2[21], step2[26]); + const __m256i s3_22_0 = _mm256_unpacklo_epi16(step2[22], step2[25]); + const __m256i s3_22_1 = _mm256_unpackhi_epi16(step2[22], step2[25]); + const __m256i s3_17_2 = _mm256_madd_epi16(s3_17_0, k__cospi_m04_p28); + const __m256i s3_17_3 = _mm256_madd_epi16(s3_17_1, k__cospi_m04_p28); + const __m256i s3_18_2 = _mm256_madd_epi16(s3_18_0, k__cospi_m28_m04); + const __m256i s3_18_3 = _mm256_madd_epi16(s3_18_1, k__cospi_m28_m04); + const __m256i s3_21_2 = _mm256_madd_epi16(s3_21_0, k__cospi_m20_p12); + const __m256i s3_21_3 = _mm256_madd_epi16(s3_21_1, k__cospi_m20_p12); + const __m256i s3_22_2 = _mm256_madd_epi16(s3_22_0, k__cospi_m12_m20); + const __m256i s3_22_3 = _mm256_madd_epi16(s3_22_1, k__cospi_m12_m20); + const __m256i s3_25_2 = _mm256_madd_epi16(s3_22_0, k__cospi_m20_p12); + const __m256i s3_25_3 = _mm256_madd_epi16(s3_22_1, k__cospi_m20_p12); + const __m256i s3_26_2 = _mm256_madd_epi16(s3_21_0, k__cospi_p12_p20); + const __m256i s3_26_3 = _mm256_madd_epi16(s3_21_1, k__cospi_p12_p20); + const __m256i s3_29_2 = _mm256_madd_epi16(s3_18_0, k__cospi_m04_p28); + const __m256i s3_29_3 = _mm256_madd_epi16(s3_18_1, k__cospi_m04_p28); + const __m256i s3_30_2 = _mm256_madd_epi16(s3_17_0, k__cospi_p28_p04); + const __m256i s3_30_3 = _mm256_madd_epi16(s3_17_1, k__cospi_p28_p04); + // dct_const_round_shift + const __m256i s3_17_4 = + _mm256_add_epi32(s3_17_2, k__DCT_CONST_ROUNDING); + const __m256i s3_17_5 = + _mm256_add_epi32(s3_17_3, k__DCT_CONST_ROUNDING); + const __m256i s3_18_4 = + _mm256_add_epi32(s3_18_2, k__DCT_CONST_ROUNDING); + const __m256i s3_18_5 = + _mm256_add_epi32(s3_18_3, k__DCT_CONST_ROUNDING); + const __m256i s3_21_4 = + _mm256_add_epi32(s3_21_2, k__DCT_CONST_ROUNDING); + const __m256i s3_21_5 = + _mm256_add_epi32(s3_21_3, k__DCT_CONST_ROUNDING); + const __m256i s3_22_4 = + _mm256_add_epi32(s3_22_2, k__DCT_CONST_ROUNDING); + const __m256i s3_22_5 = + _mm256_add_epi32(s3_22_3, k__DCT_CONST_ROUNDING); + const __m256i s3_17_6 = _mm256_srai_epi32(s3_17_4, DCT_CONST_BITS); + const __m256i s3_17_7 = _mm256_srai_epi32(s3_17_5, DCT_CONST_BITS); + const __m256i s3_18_6 = _mm256_srai_epi32(s3_18_4, DCT_CONST_BITS); + const __m256i s3_18_7 = _mm256_srai_epi32(s3_18_5, DCT_CONST_BITS); + const __m256i s3_21_6 = _mm256_srai_epi32(s3_21_4, DCT_CONST_BITS); + const __m256i s3_21_7 = _mm256_srai_epi32(s3_21_5, DCT_CONST_BITS); + const __m256i s3_22_6 = _mm256_srai_epi32(s3_22_4, DCT_CONST_BITS); + const __m256i s3_22_7 = _mm256_srai_epi32(s3_22_5, DCT_CONST_BITS); + const __m256i s3_25_4 = + _mm256_add_epi32(s3_25_2, k__DCT_CONST_ROUNDING); + const __m256i s3_25_5 = + _mm256_add_epi32(s3_25_3, k__DCT_CONST_ROUNDING); + const __m256i s3_26_4 = + _mm256_add_epi32(s3_26_2, k__DCT_CONST_ROUNDING); + const __m256i s3_26_5 = + _mm256_add_epi32(s3_26_3, k__DCT_CONST_ROUNDING); + const __m256i s3_29_4 = + _mm256_add_epi32(s3_29_2, k__DCT_CONST_ROUNDING); + const __m256i s3_29_5 = + _mm256_add_epi32(s3_29_3, k__DCT_CONST_ROUNDING); + const __m256i s3_30_4 = + _mm256_add_epi32(s3_30_2, k__DCT_CONST_ROUNDING); + const __m256i s3_30_5 = + _mm256_add_epi32(s3_30_3, k__DCT_CONST_ROUNDING); + const __m256i s3_25_6 = _mm256_srai_epi32(s3_25_4, DCT_CONST_BITS); + const __m256i s3_25_7 = _mm256_srai_epi32(s3_25_5, DCT_CONST_BITS); + const __m256i s3_26_6 = _mm256_srai_epi32(s3_26_4, DCT_CONST_BITS); + const __m256i s3_26_7 = _mm256_srai_epi32(s3_26_5, DCT_CONST_BITS); + const __m256i s3_29_6 = _mm256_srai_epi32(s3_29_4, DCT_CONST_BITS); + const __m256i s3_29_7 = _mm256_srai_epi32(s3_29_5, DCT_CONST_BITS); + const __m256i s3_30_6 = _mm256_srai_epi32(s3_30_4, DCT_CONST_BITS); + const __m256i s3_30_7 = _mm256_srai_epi32(s3_30_5, DCT_CONST_BITS); + // Combine + step3[17] = _mm256_packs_epi32(s3_17_6, s3_17_7); + step3[18] = _mm256_packs_epi32(s3_18_6, s3_18_7); + step3[21] = _mm256_packs_epi32(s3_21_6, s3_21_7); + step3[22] = _mm256_packs_epi32(s3_22_6, s3_22_7); + // Combine + step3[25] = _mm256_packs_epi32(s3_25_6, s3_25_7); + step3[26] = _mm256_packs_epi32(s3_26_6, s3_26_7); + step3[29] = _mm256_packs_epi32(s3_29_6, s3_29_7); + step3[30] = _mm256_packs_epi32(s3_30_6, s3_30_7); + } + // Stage 7 + { + const __m256i out_02_0 = _mm256_unpacklo_epi16(step3[8], step3[15]); + const __m256i out_02_1 = _mm256_unpackhi_epi16(step3[8], step3[15]); + const __m256i out_18_0 = _mm256_unpacklo_epi16(step3[9], step3[14]); + const __m256i out_18_1 = _mm256_unpackhi_epi16(step3[9], step3[14]); + const __m256i out_10_0 = _mm256_unpacklo_epi16(step3[10], step3[13]); + const __m256i out_10_1 = _mm256_unpackhi_epi16(step3[10], step3[13]); + const __m256i out_26_0 = _mm256_unpacklo_epi16(step3[11], step3[12]); + const __m256i out_26_1 = _mm256_unpackhi_epi16(step3[11], step3[12]); + const __m256i out_02_2 = + _mm256_madd_epi16(out_02_0, k__cospi_p30_p02); + const __m256i out_02_3 = + _mm256_madd_epi16(out_02_1, k__cospi_p30_p02); + const __m256i out_18_2 = + _mm256_madd_epi16(out_18_0, k__cospi_p14_p18); + const __m256i out_18_3 = + _mm256_madd_epi16(out_18_1, k__cospi_p14_p18); + const __m256i out_10_2 = + _mm256_madd_epi16(out_10_0, k__cospi_p22_p10); + const __m256i out_10_3 = + _mm256_madd_epi16(out_10_1, k__cospi_p22_p10); + const __m256i out_26_2 = + _mm256_madd_epi16(out_26_0, k__cospi_p06_p26); + const __m256i out_26_3 = + _mm256_madd_epi16(out_26_1, k__cospi_p06_p26); + const __m256i out_06_2 = + _mm256_madd_epi16(out_26_0, k__cospi_m26_p06); + const __m256i out_06_3 = + _mm256_madd_epi16(out_26_1, k__cospi_m26_p06); + const __m256i out_22_2 = + _mm256_madd_epi16(out_10_0, k__cospi_m10_p22); + const __m256i out_22_3 = + _mm256_madd_epi16(out_10_1, k__cospi_m10_p22); + const __m256i out_14_2 = + _mm256_madd_epi16(out_18_0, k__cospi_m18_p14); + const __m256i out_14_3 = + _mm256_madd_epi16(out_18_1, k__cospi_m18_p14); + const __m256i out_30_2 = + _mm256_madd_epi16(out_02_0, k__cospi_m02_p30); + const __m256i out_30_3 = + _mm256_madd_epi16(out_02_1, k__cospi_m02_p30); + // dct_const_round_shift + const __m256i out_02_4 = + _mm256_add_epi32(out_02_2, k__DCT_CONST_ROUNDING); + const __m256i out_02_5 = + _mm256_add_epi32(out_02_3, k__DCT_CONST_ROUNDING); + const __m256i out_18_4 = + _mm256_add_epi32(out_18_2, k__DCT_CONST_ROUNDING); + const __m256i out_18_5 = + _mm256_add_epi32(out_18_3, k__DCT_CONST_ROUNDING); + const __m256i out_10_4 = + _mm256_add_epi32(out_10_2, k__DCT_CONST_ROUNDING); + const __m256i out_10_5 = + _mm256_add_epi32(out_10_3, k__DCT_CONST_ROUNDING); + const __m256i out_26_4 = + _mm256_add_epi32(out_26_2, k__DCT_CONST_ROUNDING); + const __m256i out_26_5 = + _mm256_add_epi32(out_26_3, k__DCT_CONST_ROUNDING); + const __m256i out_06_4 = + _mm256_add_epi32(out_06_2, k__DCT_CONST_ROUNDING); + const __m256i out_06_5 = + _mm256_add_epi32(out_06_3, k__DCT_CONST_ROUNDING); + const __m256i out_22_4 = + _mm256_add_epi32(out_22_2, k__DCT_CONST_ROUNDING); + const __m256i out_22_5 = + _mm256_add_epi32(out_22_3, k__DCT_CONST_ROUNDING); + const __m256i out_14_4 = + _mm256_add_epi32(out_14_2, k__DCT_CONST_ROUNDING); + const __m256i out_14_5 = + _mm256_add_epi32(out_14_3, k__DCT_CONST_ROUNDING); + const __m256i out_30_4 = + _mm256_add_epi32(out_30_2, k__DCT_CONST_ROUNDING); + const __m256i out_30_5 = + _mm256_add_epi32(out_30_3, k__DCT_CONST_ROUNDING); + const __m256i out_02_6 = _mm256_srai_epi32(out_02_4, DCT_CONST_BITS); + const __m256i out_02_7 = _mm256_srai_epi32(out_02_5, DCT_CONST_BITS); + const __m256i out_18_6 = _mm256_srai_epi32(out_18_4, DCT_CONST_BITS); + const __m256i out_18_7 = _mm256_srai_epi32(out_18_5, DCT_CONST_BITS); + const __m256i out_10_6 = _mm256_srai_epi32(out_10_4, DCT_CONST_BITS); + const __m256i out_10_7 = _mm256_srai_epi32(out_10_5, DCT_CONST_BITS); + const __m256i out_26_6 = _mm256_srai_epi32(out_26_4, DCT_CONST_BITS); + const __m256i out_26_7 = _mm256_srai_epi32(out_26_5, DCT_CONST_BITS); + const __m256i out_06_6 = _mm256_srai_epi32(out_06_4, DCT_CONST_BITS); + const __m256i out_06_7 = _mm256_srai_epi32(out_06_5, DCT_CONST_BITS); + const __m256i out_22_6 = _mm256_srai_epi32(out_22_4, DCT_CONST_BITS); + const __m256i out_22_7 = _mm256_srai_epi32(out_22_5, DCT_CONST_BITS); + const __m256i out_14_6 = _mm256_srai_epi32(out_14_4, DCT_CONST_BITS); + const __m256i out_14_7 = _mm256_srai_epi32(out_14_5, DCT_CONST_BITS); + const __m256i out_30_6 = _mm256_srai_epi32(out_30_4, DCT_CONST_BITS); + const __m256i out_30_7 = _mm256_srai_epi32(out_30_5, DCT_CONST_BITS); + // Combine + out[2] = _mm256_packs_epi32(out_02_6, out_02_7); + out[18] = _mm256_packs_epi32(out_18_6, out_18_7); + out[10] = _mm256_packs_epi32(out_10_6, out_10_7); + out[26] = _mm256_packs_epi32(out_26_6, out_26_7); + out[6] = _mm256_packs_epi32(out_06_6, out_06_7); + out[22] = _mm256_packs_epi32(out_22_6, out_22_7); + out[14] = _mm256_packs_epi32(out_14_6, out_14_7); + out[30] = _mm256_packs_epi32(out_30_6, out_30_7); + } + { + step1[16] = _mm256_add_epi16(step3[17], step2[16]); + step1[17] = _mm256_sub_epi16(step2[16], step3[17]); + step1[18] = _mm256_sub_epi16(step2[19], step3[18]); + step1[19] = _mm256_add_epi16(step3[18], step2[19]); + step1[20] = _mm256_add_epi16(step3[21], step2[20]); + step1[21] = _mm256_sub_epi16(step2[20], step3[21]); + step1[22] = _mm256_sub_epi16(step2[23], step3[22]); + step1[23] = _mm256_add_epi16(step3[22], step2[23]); + step1[24] = _mm256_add_epi16(step3[25], step2[24]); + step1[25] = _mm256_sub_epi16(step2[24], step3[25]); + step1[26] = _mm256_sub_epi16(step2[27], step3[26]); + step1[27] = _mm256_add_epi16(step3[26], step2[27]); + step1[28] = _mm256_add_epi16(step3[29], step2[28]); + step1[29] = _mm256_sub_epi16(step2[28], step3[29]); + step1[30] = _mm256_sub_epi16(step2[31], step3[30]); + step1[31] = _mm256_add_epi16(step3[30], step2[31]); + } + // Final stage --- outputs indices are bit-reversed. + { + const __m256i out_01_0 = _mm256_unpacklo_epi16(step1[16], step1[31]); + const __m256i out_01_1 = _mm256_unpackhi_epi16(step1[16], step1[31]); + const __m256i out_17_0 = _mm256_unpacklo_epi16(step1[17], step1[30]); + const __m256i out_17_1 = _mm256_unpackhi_epi16(step1[17], step1[30]); + const __m256i out_09_0 = _mm256_unpacklo_epi16(step1[18], step1[29]); + const __m256i out_09_1 = _mm256_unpackhi_epi16(step1[18], step1[29]); + const __m256i out_25_0 = _mm256_unpacklo_epi16(step1[19], step1[28]); + const __m256i out_25_1 = _mm256_unpackhi_epi16(step1[19], step1[28]); + const __m256i out_01_2 = + _mm256_madd_epi16(out_01_0, k__cospi_p31_p01); + const __m256i out_01_3 = + _mm256_madd_epi16(out_01_1, k__cospi_p31_p01); + const __m256i out_17_2 = + _mm256_madd_epi16(out_17_0, k__cospi_p15_p17); + const __m256i out_17_3 = + _mm256_madd_epi16(out_17_1, k__cospi_p15_p17); + const __m256i out_09_2 = + _mm256_madd_epi16(out_09_0, k__cospi_p23_p09); + const __m256i out_09_3 = + _mm256_madd_epi16(out_09_1, k__cospi_p23_p09); + const __m256i out_25_2 = + _mm256_madd_epi16(out_25_0, k__cospi_p07_p25); + const __m256i out_25_3 = + _mm256_madd_epi16(out_25_1, k__cospi_p07_p25); + const __m256i out_07_2 = + _mm256_madd_epi16(out_25_0, k__cospi_m25_p07); + const __m256i out_07_3 = + _mm256_madd_epi16(out_25_1, k__cospi_m25_p07); + const __m256i out_23_2 = + _mm256_madd_epi16(out_09_0, k__cospi_m09_p23); + const __m256i out_23_3 = + _mm256_madd_epi16(out_09_1, k__cospi_m09_p23); + const __m256i out_15_2 = + _mm256_madd_epi16(out_17_0, k__cospi_m17_p15); + const __m256i out_15_3 = + _mm256_madd_epi16(out_17_1, k__cospi_m17_p15); + const __m256i out_31_2 = + _mm256_madd_epi16(out_01_0, k__cospi_m01_p31); + const __m256i out_31_3 = + _mm256_madd_epi16(out_01_1, k__cospi_m01_p31); + // dct_const_round_shift + const __m256i out_01_4 = + _mm256_add_epi32(out_01_2, k__DCT_CONST_ROUNDING); + const __m256i out_01_5 = + _mm256_add_epi32(out_01_3, k__DCT_CONST_ROUNDING); + const __m256i out_17_4 = + _mm256_add_epi32(out_17_2, k__DCT_CONST_ROUNDING); + const __m256i out_17_5 = + _mm256_add_epi32(out_17_3, k__DCT_CONST_ROUNDING); + const __m256i out_09_4 = + _mm256_add_epi32(out_09_2, k__DCT_CONST_ROUNDING); + const __m256i out_09_5 = + _mm256_add_epi32(out_09_3, k__DCT_CONST_ROUNDING); + const __m256i out_25_4 = + _mm256_add_epi32(out_25_2, k__DCT_CONST_ROUNDING); + const __m256i out_25_5 = + _mm256_add_epi32(out_25_3, k__DCT_CONST_ROUNDING); + const __m256i out_07_4 = + _mm256_add_epi32(out_07_2, k__DCT_CONST_ROUNDING); + const __m256i out_07_5 = + _mm256_add_epi32(out_07_3, k__DCT_CONST_ROUNDING); + const __m256i out_23_4 = + _mm256_add_epi32(out_23_2, k__DCT_CONST_ROUNDING); + const __m256i out_23_5 = + _mm256_add_epi32(out_23_3, k__DCT_CONST_ROUNDING); + const __m256i out_15_4 = + _mm256_add_epi32(out_15_2, k__DCT_CONST_ROUNDING); + const __m256i out_15_5 = + _mm256_add_epi32(out_15_3, k__DCT_CONST_ROUNDING); + const __m256i out_31_4 = + _mm256_add_epi32(out_31_2, k__DCT_CONST_ROUNDING); + const __m256i out_31_5 = + _mm256_add_epi32(out_31_3, k__DCT_CONST_ROUNDING); + const __m256i out_01_6 = _mm256_srai_epi32(out_01_4, DCT_CONST_BITS); + const __m256i out_01_7 = _mm256_srai_epi32(out_01_5, DCT_CONST_BITS); + const __m256i out_17_6 = _mm256_srai_epi32(out_17_4, DCT_CONST_BITS); + const __m256i out_17_7 = _mm256_srai_epi32(out_17_5, DCT_CONST_BITS); + const __m256i out_09_6 = _mm256_srai_epi32(out_09_4, DCT_CONST_BITS); + const __m256i out_09_7 = _mm256_srai_epi32(out_09_5, DCT_CONST_BITS); + const __m256i out_25_6 = _mm256_srai_epi32(out_25_4, DCT_CONST_BITS); + const __m256i out_25_7 = _mm256_srai_epi32(out_25_5, DCT_CONST_BITS); + const __m256i out_07_6 = _mm256_srai_epi32(out_07_4, DCT_CONST_BITS); + const __m256i out_07_7 = _mm256_srai_epi32(out_07_5, DCT_CONST_BITS); + const __m256i out_23_6 = _mm256_srai_epi32(out_23_4, DCT_CONST_BITS); + const __m256i out_23_7 = _mm256_srai_epi32(out_23_5, DCT_CONST_BITS); + const __m256i out_15_6 = _mm256_srai_epi32(out_15_4, DCT_CONST_BITS); + const __m256i out_15_7 = _mm256_srai_epi32(out_15_5, DCT_CONST_BITS); + const __m256i out_31_6 = _mm256_srai_epi32(out_31_4, DCT_CONST_BITS); + const __m256i out_31_7 = _mm256_srai_epi32(out_31_5, DCT_CONST_BITS); + // Combine + out[1] = _mm256_packs_epi32(out_01_6, out_01_7); + out[17] = _mm256_packs_epi32(out_17_6, out_17_7); + out[9] = _mm256_packs_epi32(out_09_6, out_09_7); + out[25] = _mm256_packs_epi32(out_25_6, out_25_7); + out[7] = _mm256_packs_epi32(out_07_6, out_07_7); + out[23] = _mm256_packs_epi32(out_23_6, out_23_7); + out[15] = _mm256_packs_epi32(out_15_6, out_15_7); + out[31] = _mm256_packs_epi32(out_31_6, out_31_7); + } + { + const __m256i out_05_0 = _mm256_unpacklo_epi16(step1[20], step1[27]); + const __m256i out_05_1 = _mm256_unpackhi_epi16(step1[20], step1[27]); + const __m256i out_21_0 = _mm256_unpacklo_epi16(step1[21], step1[26]); + const __m256i out_21_1 = _mm256_unpackhi_epi16(step1[21], step1[26]); + const __m256i out_13_0 = _mm256_unpacklo_epi16(step1[22], step1[25]); + const __m256i out_13_1 = _mm256_unpackhi_epi16(step1[22], step1[25]); + const __m256i out_29_0 = _mm256_unpacklo_epi16(step1[23], step1[24]); + const __m256i out_29_1 = _mm256_unpackhi_epi16(step1[23], step1[24]); + const __m256i out_05_2 = + _mm256_madd_epi16(out_05_0, k__cospi_p27_p05); + const __m256i out_05_3 = + _mm256_madd_epi16(out_05_1, k__cospi_p27_p05); + const __m256i out_21_2 = + _mm256_madd_epi16(out_21_0, k__cospi_p11_p21); + const __m256i out_21_3 = + _mm256_madd_epi16(out_21_1, k__cospi_p11_p21); + const __m256i out_13_2 = + _mm256_madd_epi16(out_13_0, k__cospi_p19_p13); + const __m256i out_13_3 = + _mm256_madd_epi16(out_13_1, k__cospi_p19_p13); + const __m256i out_29_2 = + _mm256_madd_epi16(out_29_0, k__cospi_p03_p29); + const __m256i out_29_3 = + _mm256_madd_epi16(out_29_1, k__cospi_p03_p29); + const __m256i out_03_2 = + _mm256_madd_epi16(out_29_0, k__cospi_m29_p03); + const __m256i out_03_3 = + _mm256_madd_epi16(out_29_1, k__cospi_m29_p03); + const __m256i out_19_2 = + _mm256_madd_epi16(out_13_0, k__cospi_m13_p19); + const __m256i out_19_3 = + _mm256_madd_epi16(out_13_1, k__cospi_m13_p19); + const __m256i out_11_2 = + _mm256_madd_epi16(out_21_0, k__cospi_m21_p11); + const __m256i out_11_3 = + _mm256_madd_epi16(out_21_1, k__cospi_m21_p11); + const __m256i out_27_2 = + _mm256_madd_epi16(out_05_0, k__cospi_m05_p27); + const __m256i out_27_3 = + _mm256_madd_epi16(out_05_1, k__cospi_m05_p27); + // dct_const_round_shift + const __m256i out_05_4 = + _mm256_add_epi32(out_05_2, k__DCT_CONST_ROUNDING); + const __m256i out_05_5 = + _mm256_add_epi32(out_05_3, k__DCT_CONST_ROUNDING); + const __m256i out_21_4 = + _mm256_add_epi32(out_21_2, k__DCT_CONST_ROUNDING); + const __m256i out_21_5 = + _mm256_add_epi32(out_21_3, k__DCT_CONST_ROUNDING); + const __m256i out_13_4 = + _mm256_add_epi32(out_13_2, k__DCT_CONST_ROUNDING); + const __m256i out_13_5 = + _mm256_add_epi32(out_13_3, k__DCT_CONST_ROUNDING); + const __m256i out_29_4 = + _mm256_add_epi32(out_29_2, k__DCT_CONST_ROUNDING); + const __m256i out_29_5 = + _mm256_add_epi32(out_29_3, k__DCT_CONST_ROUNDING); + const __m256i out_03_4 = + _mm256_add_epi32(out_03_2, k__DCT_CONST_ROUNDING); + const __m256i out_03_5 = + _mm256_add_epi32(out_03_3, k__DCT_CONST_ROUNDING); + const __m256i out_19_4 = + _mm256_add_epi32(out_19_2, k__DCT_CONST_ROUNDING); + const __m256i out_19_5 = + _mm256_add_epi32(out_19_3, k__DCT_CONST_ROUNDING); + const __m256i out_11_4 = + _mm256_add_epi32(out_11_2, k__DCT_CONST_ROUNDING); + const __m256i out_11_5 = + _mm256_add_epi32(out_11_3, k__DCT_CONST_ROUNDING); + const __m256i out_27_4 = + _mm256_add_epi32(out_27_2, k__DCT_CONST_ROUNDING); + const __m256i out_27_5 = + _mm256_add_epi32(out_27_3, k__DCT_CONST_ROUNDING); + const __m256i out_05_6 = _mm256_srai_epi32(out_05_4, DCT_CONST_BITS); + const __m256i out_05_7 = _mm256_srai_epi32(out_05_5, DCT_CONST_BITS); + const __m256i out_21_6 = _mm256_srai_epi32(out_21_4, DCT_CONST_BITS); + const __m256i out_21_7 = _mm256_srai_epi32(out_21_5, DCT_CONST_BITS); + const __m256i out_13_6 = _mm256_srai_epi32(out_13_4, DCT_CONST_BITS); + const __m256i out_13_7 = _mm256_srai_epi32(out_13_5, DCT_CONST_BITS); + const __m256i out_29_6 = _mm256_srai_epi32(out_29_4, DCT_CONST_BITS); + const __m256i out_29_7 = _mm256_srai_epi32(out_29_5, DCT_CONST_BITS); + const __m256i out_03_6 = _mm256_srai_epi32(out_03_4, DCT_CONST_BITS); + const __m256i out_03_7 = _mm256_srai_epi32(out_03_5, DCT_CONST_BITS); + const __m256i out_19_6 = _mm256_srai_epi32(out_19_4, DCT_CONST_BITS); + const __m256i out_19_7 = _mm256_srai_epi32(out_19_5, DCT_CONST_BITS); + const __m256i out_11_6 = _mm256_srai_epi32(out_11_4, DCT_CONST_BITS); + const __m256i out_11_7 = _mm256_srai_epi32(out_11_5, DCT_CONST_BITS); + const __m256i out_27_6 = _mm256_srai_epi32(out_27_4, DCT_CONST_BITS); + const __m256i out_27_7 = _mm256_srai_epi32(out_27_5, DCT_CONST_BITS); + // Combine + out[5] = _mm256_packs_epi32(out_05_6, out_05_7); + out[21] = _mm256_packs_epi32(out_21_6, out_21_7); + out[13] = _mm256_packs_epi32(out_13_6, out_13_7); + out[29] = _mm256_packs_epi32(out_29_6, out_29_7); + out[3] = _mm256_packs_epi32(out_03_6, out_03_7); + out[19] = _mm256_packs_epi32(out_19_6, out_19_7); + out[11] = _mm256_packs_epi32(out_11_6, out_11_7); + out[27] = _mm256_packs_epi32(out_27_6, out_27_7); + } +#if FDCT32x32_HIGH_PRECISION + } else { + __m256i lstep1[64], lstep2[64], lstep3[64]; + __m256i u[32], v[32], sign[16]; + const __m256i K32One = _mm256_set_epi32(1, 1, 1, 1, 1, 1, 1, 1); + // start using 32-bit operations + // stage 3 + { + // expanding to 32-bit length priori to addition operations + lstep2[0] = _mm256_unpacklo_epi16(step2[0], kZero); + lstep2[1] = _mm256_unpackhi_epi16(step2[0], kZero); + lstep2[2] = _mm256_unpacklo_epi16(step2[1], kZero); + lstep2[3] = _mm256_unpackhi_epi16(step2[1], kZero); + lstep2[4] = _mm256_unpacklo_epi16(step2[2], kZero); + lstep2[5] = _mm256_unpackhi_epi16(step2[2], kZero); + lstep2[6] = _mm256_unpacklo_epi16(step2[3], kZero); + lstep2[7] = _mm256_unpackhi_epi16(step2[3], kZero); + lstep2[8] = _mm256_unpacklo_epi16(step2[4], kZero); + lstep2[9] = _mm256_unpackhi_epi16(step2[4], kZero); + lstep2[10] = _mm256_unpacklo_epi16(step2[5], kZero); + lstep2[11] = _mm256_unpackhi_epi16(step2[5], kZero); + lstep2[12] = _mm256_unpacklo_epi16(step2[6], kZero); + lstep2[13] = _mm256_unpackhi_epi16(step2[6], kZero); + lstep2[14] = _mm256_unpacklo_epi16(step2[7], kZero); + lstep2[15] = _mm256_unpackhi_epi16(step2[7], kZero); + lstep2[0] = _mm256_madd_epi16(lstep2[0], kOne); + lstep2[1] = _mm256_madd_epi16(lstep2[1], kOne); + lstep2[2] = _mm256_madd_epi16(lstep2[2], kOne); + lstep2[3] = _mm256_madd_epi16(lstep2[3], kOne); + lstep2[4] = _mm256_madd_epi16(lstep2[4], kOne); + lstep2[5] = _mm256_madd_epi16(lstep2[5], kOne); + lstep2[6] = _mm256_madd_epi16(lstep2[6], kOne); + lstep2[7] = _mm256_madd_epi16(lstep2[7], kOne); + lstep2[8] = _mm256_madd_epi16(lstep2[8], kOne); + lstep2[9] = _mm256_madd_epi16(lstep2[9], kOne); + lstep2[10] = _mm256_madd_epi16(lstep2[10], kOne); + lstep2[11] = _mm256_madd_epi16(lstep2[11], kOne); + lstep2[12] = _mm256_madd_epi16(lstep2[12], kOne); + lstep2[13] = _mm256_madd_epi16(lstep2[13], kOne); + lstep2[14] = _mm256_madd_epi16(lstep2[14], kOne); + lstep2[15] = _mm256_madd_epi16(lstep2[15], kOne); + + lstep3[0] = _mm256_add_epi32(lstep2[14], lstep2[0]); + lstep3[1] = _mm256_add_epi32(lstep2[15], lstep2[1]); + lstep3[2] = _mm256_add_epi32(lstep2[12], lstep2[2]); + lstep3[3] = _mm256_add_epi32(lstep2[13], lstep2[3]); + lstep3[4] = _mm256_add_epi32(lstep2[10], lstep2[4]); + lstep3[5] = _mm256_add_epi32(lstep2[11], lstep2[5]); + lstep3[6] = _mm256_add_epi32(lstep2[8], lstep2[6]); + lstep3[7] = _mm256_add_epi32(lstep2[9], lstep2[7]); + lstep3[8] = _mm256_sub_epi32(lstep2[6], lstep2[8]); + lstep3[9] = _mm256_sub_epi32(lstep2[7], lstep2[9]); + lstep3[10] = _mm256_sub_epi32(lstep2[4], lstep2[10]); + lstep3[11] = _mm256_sub_epi32(lstep2[5], lstep2[11]); + lstep3[12] = _mm256_sub_epi32(lstep2[2], lstep2[12]); + lstep3[13] = _mm256_sub_epi32(lstep2[3], lstep2[13]); + lstep3[14] = _mm256_sub_epi32(lstep2[0], lstep2[14]); + lstep3[15] = _mm256_sub_epi32(lstep2[1], lstep2[15]); + } + { + const __m256i s3_10_0 = _mm256_unpacklo_epi16(step2[13], step2[10]); + const __m256i s3_10_1 = _mm256_unpackhi_epi16(step2[13], step2[10]); + const __m256i s3_11_0 = _mm256_unpacklo_epi16(step2[12], step2[11]); + const __m256i s3_11_1 = _mm256_unpackhi_epi16(step2[12], step2[11]); + const __m256i s3_10_2 = _mm256_madd_epi16(s3_10_0, k__cospi_p16_m16); + const __m256i s3_10_3 = _mm256_madd_epi16(s3_10_1, k__cospi_p16_m16); + const __m256i s3_11_2 = _mm256_madd_epi16(s3_11_0, k__cospi_p16_m16); + const __m256i s3_11_3 = _mm256_madd_epi16(s3_11_1, k__cospi_p16_m16); + const __m256i s3_12_2 = _mm256_madd_epi16(s3_11_0, k__cospi_p16_p16); + const __m256i s3_12_3 = _mm256_madd_epi16(s3_11_1, k__cospi_p16_p16); + const __m256i s3_13_2 = _mm256_madd_epi16(s3_10_0, k__cospi_p16_p16); + const __m256i s3_13_3 = _mm256_madd_epi16(s3_10_1, k__cospi_p16_p16); + // dct_const_round_shift + const __m256i s3_10_4 = + _mm256_add_epi32(s3_10_2, k__DCT_CONST_ROUNDING); + const __m256i s3_10_5 = + _mm256_add_epi32(s3_10_3, k__DCT_CONST_ROUNDING); + const __m256i s3_11_4 = + _mm256_add_epi32(s3_11_2, k__DCT_CONST_ROUNDING); + const __m256i s3_11_5 = + _mm256_add_epi32(s3_11_3, k__DCT_CONST_ROUNDING); + const __m256i s3_12_4 = + _mm256_add_epi32(s3_12_2, k__DCT_CONST_ROUNDING); + const __m256i s3_12_5 = + _mm256_add_epi32(s3_12_3, k__DCT_CONST_ROUNDING); + const __m256i s3_13_4 = + _mm256_add_epi32(s3_13_2, k__DCT_CONST_ROUNDING); + const __m256i s3_13_5 = + _mm256_add_epi32(s3_13_3, k__DCT_CONST_ROUNDING); + lstep3[20] = _mm256_srai_epi32(s3_10_4, DCT_CONST_BITS); + lstep3[21] = _mm256_srai_epi32(s3_10_5, DCT_CONST_BITS); + lstep3[22] = _mm256_srai_epi32(s3_11_4, DCT_CONST_BITS); + lstep3[23] = _mm256_srai_epi32(s3_11_5, DCT_CONST_BITS); + lstep3[24] = _mm256_srai_epi32(s3_12_4, DCT_CONST_BITS); + lstep3[25] = _mm256_srai_epi32(s3_12_5, DCT_CONST_BITS); + lstep3[26] = _mm256_srai_epi32(s3_13_4, DCT_CONST_BITS); + lstep3[27] = _mm256_srai_epi32(s3_13_5, DCT_CONST_BITS); + } + { + lstep2[40] = _mm256_unpacklo_epi16(step2[20], kZero); + lstep2[41] = _mm256_unpackhi_epi16(step2[20], kZero); + lstep2[42] = _mm256_unpacklo_epi16(step2[21], kZero); + lstep2[43] = _mm256_unpackhi_epi16(step2[21], kZero); + lstep2[44] = _mm256_unpacklo_epi16(step2[22], kZero); + lstep2[45] = _mm256_unpackhi_epi16(step2[22], kZero); + lstep2[46] = _mm256_unpacklo_epi16(step2[23], kZero); + lstep2[47] = _mm256_unpackhi_epi16(step2[23], kZero); + lstep2[48] = _mm256_unpacklo_epi16(step2[24], kZero); + lstep2[49] = _mm256_unpackhi_epi16(step2[24], kZero); + lstep2[50] = _mm256_unpacklo_epi16(step2[25], kZero); + lstep2[51] = _mm256_unpackhi_epi16(step2[25], kZero); + lstep2[52] = _mm256_unpacklo_epi16(step2[26], kZero); + lstep2[53] = _mm256_unpackhi_epi16(step2[26], kZero); + lstep2[54] = _mm256_unpacklo_epi16(step2[27], kZero); + lstep2[55] = _mm256_unpackhi_epi16(step2[27], kZero); + lstep2[40] = _mm256_madd_epi16(lstep2[40], kOne); + lstep2[41] = _mm256_madd_epi16(lstep2[41], kOne); + lstep2[42] = _mm256_madd_epi16(lstep2[42], kOne); + lstep2[43] = _mm256_madd_epi16(lstep2[43], kOne); + lstep2[44] = _mm256_madd_epi16(lstep2[44], kOne); + lstep2[45] = _mm256_madd_epi16(lstep2[45], kOne); + lstep2[46] = _mm256_madd_epi16(lstep2[46], kOne); + lstep2[47] = _mm256_madd_epi16(lstep2[47], kOne); + lstep2[48] = _mm256_madd_epi16(lstep2[48], kOne); + lstep2[49] = _mm256_madd_epi16(lstep2[49], kOne); + lstep2[50] = _mm256_madd_epi16(lstep2[50], kOne); + lstep2[51] = _mm256_madd_epi16(lstep2[51], kOne); + lstep2[52] = _mm256_madd_epi16(lstep2[52], kOne); + lstep2[53] = _mm256_madd_epi16(lstep2[53], kOne); + lstep2[54] = _mm256_madd_epi16(lstep2[54], kOne); + lstep2[55] = _mm256_madd_epi16(lstep2[55], kOne); + + lstep1[32] = _mm256_unpacklo_epi16(step1[16], kZero); + lstep1[33] = _mm256_unpackhi_epi16(step1[16], kZero); + lstep1[34] = _mm256_unpacklo_epi16(step1[17], kZero); + lstep1[35] = _mm256_unpackhi_epi16(step1[17], kZero); + lstep1[36] = _mm256_unpacklo_epi16(step1[18], kZero); + lstep1[37] = _mm256_unpackhi_epi16(step1[18], kZero); + lstep1[38] = _mm256_unpacklo_epi16(step1[19], kZero); + lstep1[39] = _mm256_unpackhi_epi16(step1[19], kZero); + lstep1[56] = _mm256_unpacklo_epi16(step1[28], kZero); + lstep1[57] = _mm256_unpackhi_epi16(step1[28], kZero); + lstep1[58] = _mm256_unpacklo_epi16(step1[29], kZero); + lstep1[59] = _mm256_unpackhi_epi16(step1[29], kZero); + lstep1[60] = _mm256_unpacklo_epi16(step1[30], kZero); + lstep1[61] = _mm256_unpackhi_epi16(step1[30], kZero); + lstep1[62] = _mm256_unpacklo_epi16(step1[31], kZero); + lstep1[63] = _mm256_unpackhi_epi16(step1[31], kZero); + lstep1[32] = _mm256_madd_epi16(lstep1[32], kOne); + lstep1[33] = _mm256_madd_epi16(lstep1[33], kOne); + lstep1[34] = _mm256_madd_epi16(lstep1[34], kOne); + lstep1[35] = _mm256_madd_epi16(lstep1[35], kOne); + lstep1[36] = _mm256_madd_epi16(lstep1[36], kOne); + lstep1[37] = _mm256_madd_epi16(lstep1[37], kOne); + lstep1[38] = _mm256_madd_epi16(lstep1[38], kOne); + lstep1[39] = _mm256_madd_epi16(lstep1[39], kOne); + lstep1[56] = _mm256_madd_epi16(lstep1[56], kOne); + lstep1[57] = _mm256_madd_epi16(lstep1[57], kOne); + lstep1[58] = _mm256_madd_epi16(lstep1[58], kOne); + lstep1[59] = _mm256_madd_epi16(lstep1[59], kOne); + lstep1[60] = _mm256_madd_epi16(lstep1[60], kOne); + lstep1[61] = _mm256_madd_epi16(lstep1[61], kOne); + lstep1[62] = _mm256_madd_epi16(lstep1[62], kOne); + lstep1[63] = _mm256_madd_epi16(lstep1[63], kOne); + + lstep3[32] = _mm256_add_epi32(lstep2[46], lstep1[32]); + lstep3[33] = _mm256_add_epi32(lstep2[47], lstep1[33]); + + lstep3[34] = _mm256_add_epi32(lstep2[44], lstep1[34]); + lstep3[35] = _mm256_add_epi32(lstep2[45], lstep1[35]); + lstep3[36] = _mm256_add_epi32(lstep2[42], lstep1[36]); + lstep3[37] = _mm256_add_epi32(lstep2[43], lstep1[37]); + lstep3[38] = _mm256_add_epi32(lstep2[40], lstep1[38]); + lstep3[39] = _mm256_add_epi32(lstep2[41], lstep1[39]); + lstep3[40] = _mm256_sub_epi32(lstep1[38], lstep2[40]); + lstep3[41] = _mm256_sub_epi32(lstep1[39], lstep2[41]); + lstep3[42] = _mm256_sub_epi32(lstep1[36], lstep2[42]); + lstep3[43] = _mm256_sub_epi32(lstep1[37], lstep2[43]); + lstep3[44] = _mm256_sub_epi32(lstep1[34], lstep2[44]); + lstep3[45] = _mm256_sub_epi32(lstep1[35], lstep2[45]); + lstep3[46] = _mm256_sub_epi32(lstep1[32], lstep2[46]); + lstep3[47] = _mm256_sub_epi32(lstep1[33], lstep2[47]); + lstep3[48] = _mm256_sub_epi32(lstep1[62], lstep2[48]); + lstep3[49] = _mm256_sub_epi32(lstep1[63], lstep2[49]); + lstep3[50] = _mm256_sub_epi32(lstep1[60], lstep2[50]); + lstep3[51] = _mm256_sub_epi32(lstep1[61], lstep2[51]); + lstep3[52] = _mm256_sub_epi32(lstep1[58], lstep2[52]); + lstep3[53] = _mm256_sub_epi32(lstep1[59], lstep2[53]); + lstep3[54] = _mm256_sub_epi32(lstep1[56], lstep2[54]); + lstep3[55] = _mm256_sub_epi32(lstep1[57], lstep2[55]); + lstep3[56] = _mm256_add_epi32(lstep2[54], lstep1[56]); + lstep3[57] = _mm256_add_epi32(lstep2[55], lstep1[57]); + lstep3[58] = _mm256_add_epi32(lstep2[52], lstep1[58]); + lstep3[59] = _mm256_add_epi32(lstep2[53], lstep1[59]); + lstep3[60] = _mm256_add_epi32(lstep2[50], lstep1[60]); + lstep3[61] = _mm256_add_epi32(lstep2[51], lstep1[61]); + lstep3[62] = _mm256_add_epi32(lstep2[48], lstep1[62]); + lstep3[63] = _mm256_add_epi32(lstep2[49], lstep1[63]); + } + + // stage 4 + { + // expanding to 32-bit length priori to addition operations + lstep2[16] = _mm256_unpacklo_epi16(step2[8], kZero); + lstep2[17] = _mm256_unpackhi_epi16(step2[8], kZero); + lstep2[18] = _mm256_unpacklo_epi16(step2[9], kZero); + lstep2[19] = _mm256_unpackhi_epi16(step2[9], kZero); + lstep2[28] = _mm256_unpacklo_epi16(step2[14], kZero); + lstep2[29] = _mm256_unpackhi_epi16(step2[14], kZero); + lstep2[30] = _mm256_unpacklo_epi16(step2[15], kZero); + lstep2[31] = _mm256_unpackhi_epi16(step2[15], kZero); + lstep2[16] = _mm256_madd_epi16(lstep2[16], kOne); + lstep2[17] = _mm256_madd_epi16(lstep2[17], kOne); + lstep2[18] = _mm256_madd_epi16(lstep2[18], kOne); + lstep2[19] = _mm256_madd_epi16(lstep2[19], kOne); + lstep2[28] = _mm256_madd_epi16(lstep2[28], kOne); + lstep2[29] = _mm256_madd_epi16(lstep2[29], kOne); + lstep2[30] = _mm256_madd_epi16(lstep2[30], kOne); + lstep2[31] = _mm256_madd_epi16(lstep2[31], kOne); + + lstep1[0] = _mm256_add_epi32(lstep3[6], lstep3[0]); + lstep1[1] = _mm256_add_epi32(lstep3[7], lstep3[1]); + lstep1[2] = _mm256_add_epi32(lstep3[4], lstep3[2]); + lstep1[3] = _mm256_add_epi32(lstep3[5], lstep3[3]); + lstep1[4] = _mm256_sub_epi32(lstep3[2], lstep3[4]); + lstep1[5] = _mm256_sub_epi32(lstep3[3], lstep3[5]); + lstep1[6] = _mm256_sub_epi32(lstep3[0], lstep3[6]); + lstep1[7] = _mm256_sub_epi32(lstep3[1], lstep3[7]); + lstep1[16] = _mm256_add_epi32(lstep3[22], lstep2[16]); + lstep1[17] = _mm256_add_epi32(lstep3[23], lstep2[17]); + lstep1[18] = _mm256_add_epi32(lstep3[20], lstep2[18]); + lstep1[19] = _mm256_add_epi32(lstep3[21], lstep2[19]); + lstep1[20] = _mm256_sub_epi32(lstep2[18], lstep3[20]); + lstep1[21] = _mm256_sub_epi32(lstep2[19], lstep3[21]); + lstep1[22] = _mm256_sub_epi32(lstep2[16], lstep3[22]); + lstep1[23] = _mm256_sub_epi32(lstep2[17], lstep3[23]); + lstep1[24] = _mm256_sub_epi32(lstep2[30], lstep3[24]); + lstep1[25] = _mm256_sub_epi32(lstep2[31], lstep3[25]); + lstep1[26] = _mm256_sub_epi32(lstep2[28], lstep3[26]); + lstep1[27] = _mm256_sub_epi32(lstep2[29], lstep3[27]); + lstep1[28] = _mm256_add_epi32(lstep3[26], lstep2[28]); + lstep1[29] = _mm256_add_epi32(lstep3[27], lstep2[29]); + lstep1[30] = _mm256_add_epi32(lstep3[24], lstep2[30]); + lstep1[31] = _mm256_add_epi32(lstep3[25], lstep2[31]); + } + { + // to be continued... + // + const __m256i k32_p16_p16 = + pair256_set_epi32(cospi_16_64, cospi_16_64); + const __m256i k32_p16_m16 = + pair256_set_epi32(cospi_16_64, -cospi_16_64); + + u[0] = _mm256_unpacklo_epi32(lstep3[12], lstep3[10]); + u[1] = _mm256_unpackhi_epi32(lstep3[12], lstep3[10]); + u[2] = _mm256_unpacklo_epi32(lstep3[13], lstep3[11]); + u[3] = _mm256_unpackhi_epi32(lstep3[13], lstep3[11]); + + // TODO(jingning): manually inline k_madd_epi32_avx2_ to further hide + // instruction latency. + v[0] = k_madd_epi32_avx2(u[0], k32_p16_m16); + v[1] = k_madd_epi32_avx2(u[1], k32_p16_m16); + v[2] = k_madd_epi32_avx2(u[2], k32_p16_m16); + v[3] = k_madd_epi32_avx2(u[3], k32_p16_m16); + v[4] = k_madd_epi32_avx2(u[0], k32_p16_p16); + v[5] = k_madd_epi32_avx2(u[1], k32_p16_p16); + v[6] = k_madd_epi32_avx2(u[2], k32_p16_p16); + v[7] = k_madd_epi32_avx2(u[3], k32_p16_p16); + + u[0] = k_packs_epi64_avx2(v[0], v[1]); + u[1] = k_packs_epi64_avx2(v[2], v[3]); + u[2] = k_packs_epi64_avx2(v[4], v[5]); + u[3] = k_packs_epi64_avx2(v[6], v[7]); + + v[0] = _mm256_add_epi32(u[0], k__DCT_CONST_ROUNDING); + v[1] = _mm256_add_epi32(u[1], k__DCT_CONST_ROUNDING); + v[2] = _mm256_add_epi32(u[2], k__DCT_CONST_ROUNDING); + v[3] = _mm256_add_epi32(u[3], k__DCT_CONST_ROUNDING); + + lstep1[10] = _mm256_srai_epi32(v[0], DCT_CONST_BITS); + lstep1[11] = _mm256_srai_epi32(v[1], DCT_CONST_BITS); + lstep1[12] = _mm256_srai_epi32(v[2], DCT_CONST_BITS); + lstep1[13] = _mm256_srai_epi32(v[3], DCT_CONST_BITS); + } + { + const __m256i k32_m08_p24 = + pair256_set_epi32(-cospi_8_64, cospi_24_64); + const __m256i k32_m24_m08 = + pair256_set_epi32(-cospi_24_64, -cospi_8_64); + const __m256i k32_p24_p08 = + pair256_set_epi32(cospi_24_64, cospi_8_64); + + u[0] = _mm256_unpacklo_epi32(lstep3[36], lstep3[58]); + u[1] = _mm256_unpackhi_epi32(lstep3[36], lstep3[58]); + u[2] = _mm256_unpacklo_epi32(lstep3[37], lstep3[59]); + u[3] = _mm256_unpackhi_epi32(lstep3[37], lstep3[59]); + u[4] = _mm256_unpacklo_epi32(lstep3[38], lstep3[56]); + u[5] = _mm256_unpackhi_epi32(lstep3[38], lstep3[56]); + u[6] = _mm256_unpacklo_epi32(lstep3[39], lstep3[57]); + u[7] = _mm256_unpackhi_epi32(lstep3[39], lstep3[57]); + u[8] = _mm256_unpacklo_epi32(lstep3[40], lstep3[54]); + u[9] = _mm256_unpackhi_epi32(lstep3[40], lstep3[54]); + u[10] = _mm256_unpacklo_epi32(lstep3[41], lstep3[55]); + u[11] = _mm256_unpackhi_epi32(lstep3[41], lstep3[55]); + u[12] = _mm256_unpacklo_epi32(lstep3[42], lstep3[52]); + u[13] = _mm256_unpackhi_epi32(lstep3[42], lstep3[52]); + u[14] = _mm256_unpacklo_epi32(lstep3[43], lstep3[53]); + u[15] = _mm256_unpackhi_epi32(lstep3[43], lstep3[53]); + + v[0] = k_madd_epi32_avx2(u[0], k32_m08_p24); + v[1] = k_madd_epi32_avx2(u[1], k32_m08_p24); + v[2] = k_madd_epi32_avx2(u[2], k32_m08_p24); + v[3] = k_madd_epi32_avx2(u[3], k32_m08_p24); + v[4] = k_madd_epi32_avx2(u[4], k32_m08_p24); + v[5] = k_madd_epi32_avx2(u[5], k32_m08_p24); + v[6] = k_madd_epi32_avx2(u[6], k32_m08_p24); + v[7] = k_madd_epi32_avx2(u[7], k32_m08_p24); + v[8] = k_madd_epi32_avx2(u[8], k32_m24_m08); + v[9] = k_madd_epi32_avx2(u[9], k32_m24_m08); + v[10] = k_madd_epi32_avx2(u[10], k32_m24_m08); + v[11] = k_madd_epi32_avx2(u[11], k32_m24_m08); + v[12] = k_madd_epi32_avx2(u[12], k32_m24_m08); + v[13] = k_madd_epi32_avx2(u[13], k32_m24_m08); + v[14] = k_madd_epi32_avx2(u[14], k32_m24_m08); + v[15] = k_madd_epi32_avx2(u[15], k32_m24_m08); + v[16] = k_madd_epi32_avx2(u[12], k32_m08_p24); + v[17] = k_madd_epi32_avx2(u[13], k32_m08_p24); + v[18] = k_madd_epi32_avx2(u[14], k32_m08_p24); + v[19] = k_madd_epi32_avx2(u[15], k32_m08_p24); + v[20] = k_madd_epi32_avx2(u[8], k32_m08_p24); + v[21] = k_madd_epi32_avx2(u[9], k32_m08_p24); + v[22] = k_madd_epi32_avx2(u[10], k32_m08_p24); + v[23] = k_madd_epi32_avx2(u[11], k32_m08_p24); + v[24] = k_madd_epi32_avx2(u[4], k32_p24_p08); + v[25] = k_madd_epi32_avx2(u[5], k32_p24_p08); + v[26] = k_madd_epi32_avx2(u[6], k32_p24_p08); + v[27] = k_madd_epi32_avx2(u[7], k32_p24_p08); + v[28] = k_madd_epi32_avx2(u[0], k32_p24_p08); + v[29] = k_madd_epi32_avx2(u[1], k32_p24_p08); + v[30] = k_madd_epi32_avx2(u[2], k32_p24_p08); + v[31] = k_madd_epi32_avx2(u[3], k32_p24_p08); + + u[0] = k_packs_epi64_avx2(v[0], v[1]); + u[1] = k_packs_epi64_avx2(v[2], v[3]); + u[2] = k_packs_epi64_avx2(v[4], v[5]); + u[3] = k_packs_epi64_avx2(v[6], v[7]); + u[4] = k_packs_epi64_avx2(v[8], v[9]); + u[5] = k_packs_epi64_avx2(v[10], v[11]); + u[6] = k_packs_epi64_avx2(v[12], v[13]); + u[7] = k_packs_epi64_avx2(v[14], v[15]); + u[8] = k_packs_epi64_avx2(v[16], v[17]); + u[9] = k_packs_epi64_avx2(v[18], v[19]); + u[10] = k_packs_epi64_avx2(v[20], v[21]); + u[11] = k_packs_epi64_avx2(v[22], v[23]); + u[12] = k_packs_epi64_avx2(v[24], v[25]); + u[13] = k_packs_epi64_avx2(v[26], v[27]); + u[14] = k_packs_epi64_avx2(v[28], v[29]); + u[15] = k_packs_epi64_avx2(v[30], v[31]); + + v[0] = _mm256_add_epi32(u[0], k__DCT_CONST_ROUNDING); + v[1] = _mm256_add_epi32(u[1], k__DCT_CONST_ROUNDING); + v[2] = _mm256_add_epi32(u[2], k__DCT_CONST_ROUNDING); + v[3] = _mm256_add_epi32(u[3], k__DCT_CONST_ROUNDING); + v[4] = _mm256_add_epi32(u[4], k__DCT_CONST_ROUNDING); + v[5] = _mm256_add_epi32(u[5], k__DCT_CONST_ROUNDING); + v[6] = _mm256_add_epi32(u[6], k__DCT_CONST_ROUNDING); + v[7] = _mm256_add_epi32(u[7], k__DCT_CONST_ROUNDING); + v[8] = _mm256_add_epi32(u[8], k__DCT_CONST_ROUNDING); + v[9] = _mm256_add_epi32(u[9], k__DCT_CONST_ROUNDING); + v[10] = _mm256_add_epi32(u[10], k__DCT_CONST_ROUNDING); + v[11] = _mm256_add_epi32(u[11], k__DCT_CONST_ROUNDING); + v[12] = _mm256_add_epi32(u[12], k__DCT_CONST_ROUNDING); + v[13] = _mm256_add_epi32(u[13], k__DCT_CONST_ROUNDING); + v[14] = _mm256_add_epi32(u[14], k__DCT_CONST_ROUNDING); + v[15] = _mm256_add_epi32(u[15], k__DCT_CONST_ROUNDING); + + lstep1[36] = _mm256_srai_epi32(v[0], DCT_CONST_BITS); + lstep1[37] = _mm256_srai_epi32(v[1], DCT_CONST_BITS); + lstep1[38] = _mm256_srai_epi32(v[2], DCT_CONST_BITS); + lstep1[39] = _mm256_srai_epi32(v[3], DCT_CONST_BITS); + lstep1[40] = _mm256_srai_epi32(v[4], DCT_CONST_BITS); + lstep1[41] = _mm256_srai_epi32(v[5], DCT_CONST_BITS); + lstep1[42] = _mm256_srai_epi32(v[6], DCT_CONST_BITS); + lstep1[43] = _mm256_srai_epi32(v[7], DCT_CONST_BITS); + lstep1[52] = _mm256_srai_epi32(v[8], DCT_CONST_BITS); + lstep1[53] = _mm256_srai_epi32(v[9], DCT_CONST_BITS); + lstep1[54] = _mm256_srai_epi32(v[10], DCT_CONST_BITS); + lstep1[55] = _mm256_srai_epi32(v[11], DCT_CONST_BITS); + lstep1[56] = _mm256_srai_epi32(v[12], DCT_CONST_BITS); + lstep1[57] = _mm256_srai_epi32(v[13], DCT_CONST_BITS); + lstep1[58] = _mm256_srai_epi32(v[14], DCT_CONST_BITS); + lstep1[59] = _mm256_srai_epi32(v[15], DCT_CONST_BITS); + } + // stage 5 + { + lstep2[8] = _mm256_add_epi32(lstep1[10], lstep3[8]); + lstep2[9] = _mm256_add_epi32(lstep1[11], lstep3[9]); + lstep2[10] = _mm256_sub_epi32(lstep3[8], lstep1[10]); + lstep2[11] = _mm256_sub_epi32(lstep3[9], lstep1[11]); + lstep2[12] = _mm256_sub_epi32(lstep3[14], lstep1[12]); + lstep2[13] = _mm256_sub_epi32(lstep3[15], lstep1[13]); + lstep2[14] = _mm256_add_epi32(lstep1[12], lstep3[14]); + lstep2[15] = _mm256_add_epi32(lstep1[13], lstep3[15]); + } + { + const __m256i k32_p16_p16 = + pair256_set_epi32(cospi_16_64, cospi_16_64); + const __m256i k32_p16_m16 = + pair256_set_epi32(cospi_16_64, -cospi_16_64); + const __m256i k32_p24_p08 = + pair256_set_epi32(cospi_24_64, cospi_8_64); + const __m256i k32_m08_p24 = + pair256_set_epi32(-cospi_8_64, cospi_24_64); + + u[0] = _mm256_unpacklo_epi32(lstep1[0], lstep1[2]); + u[1] = _mm256_unpackhi_epi32(lstep1[0], lstep1[2]); + u[2] = _mm256_unpacklo_epi32(lstep1[1], lstep1[3]); + u[3] = _mm256_unpackhi_epi32(lstep1[1], lstep1[3]); + u[4] = _mm256_unpacklo_epi32(lstep1[4], lstep1[6]); + u[5] = _mm256_unpackhi_epi32(lstep1[4], lstep1[6]); + u[6] = _mm256_unpacklo_epi32(lstep1[5], lstep1[7]); + u[7] = _mm256_unpackhi_epi32(lstep1[5], lstep1[7]); + + // TODO(jingning): manually inline k_madd_epi32_avx2_ to further hide + // instruction latency. + v[0] = k_madd_epi32_avx2(u[0], k32_p16_p16); + v[1] = k_madd_epi32_avx2(u[1], k32_p16_p16); + v[2] = k_madd_epi32_avx2(u[2], k32_p16_p16); + v[3] = k_madd_epi32_avx2(u[3], k32_p16_p16); + v[4] = k_madd_epi32_avx2(u[0], k32_p16_m16); + v[5] = k_madd_epi32_avx2(u[1], k32_p16_m16); + v[6] = k_madd_epi32_avx2(u[2], k32_p16_m16); + v[7] = k_madd_epi32_avx2(u[3], k32_p16_m16); + v[8] = k_madd_epi32_avx2(u[4], k32_p24_p08); + v[9] = k_madd_epi32_avx2(u[5], k32_p24_p08); + v[10] = k_madd_epi32_avx2(u[6], k32_p24_p08); + v[11] = k_madd_epi32_avx2(u[7], k32_p24_p08); + v[12] = k_madd_epi32_avx2(u[4], k32_m08_p24); + v[13] = k_madd_epi32_avx2(u[5], k32_m08_p24); + v[14] = k_madd_epi32_avx2(u[6], k32_m08_p24); + v[15] = k_madd_epi32_avx2(u[7], k32_m08_p24); + + u[0] = k_packs_epi64_avx2(v[0], v[1]); + u[1] = k_packs_epi64_avx2(v[2], v[3]); + u[2] = k_packs_epi64_avx2(v[4], v[5]); + u[3] = k_packs_epi64_avx2(v[6], v[7]); + u[4] = k_packs_epi64_avx2(v[8], v[9]); + u[5] = k_packs_epi64_avx2(v[10], v[11]); + u[6] = k_packs_epi64_avx2(v[12], v[13]); + u[7] = k_packs_epi64_avx2(v[14], v[15]); + + v[0] = _mm256_add_epi32(u[0], k__DCT_CONST_ROUNDING); + v[1] = _mm256_add_epi32(u[1], k__DCT_CONST_ROUNDING); + v[2] = _mm256_add_epi32(u[2], k__DCT_CONST_ROUNDING); + v[3] = _mm256_add_epi32(u[3], k__DCT_CONST_ROUNDING); + v[4] = _mm256_add_epi32(u[4], k__DCT_CONST_ROUNDING); + v[5] = _mm256_add_epi32(u[5], k__DCT_CONST_ROUNDING); + v[6] = _mm256_add_epi32(u[6], k__DCT_CONST_ROUNDING); + v[7] = _mm256_add_epi32(u[7], k__DCT_CONST_ROUNDING); + + u[0] = _mm256_srai_epi32(v[0], DCT_CONST_BITS); + u[1] = _mm256_srai_epi32(v[1], DCT_CONST_BITS); + u[2] = _mm256_srai_epi32(v[2], DCT_CONST_BITS); + u[3] = _mm256_srai_epi32(v[3], DCT_CONST_BITS); + u[4] = _mm256_srai_epi32(v[4], DCT_CONST_BITS); + u[5] = _mm256_srai_epi32(v[5], DCT_CONST_BITS); + u[6] = _mm256_srai_epi32(v[6], DCT_CONST_BITS); + u[7] = _mm256_srai_epi32(v[7], DCT_CONST_BITS); + + sign[0] = _mm256_cmpgt_epi32(kZero, u[0]); + sign[1] = _mm256_cmpgt_epi32(kZero, u[1]); + sign[2] = _mm256_cmpgt_epi32(kZero, u[2]); + sign[3] = _mm256_cmpgt_epi32(kZero, u[3]); + sign[4] = _mm256_cmpgt_epi32(kZero, u[4]); + sign[5] = _mm256_cmpgt_epi32(kZero, u[5]); + sign[6] = _mm256_cmpgt_epi32(kZero, u[6]); + sign[7] = _mm256_cmpgt_epi32(kZero, u[7]); + + u[0] = _mm256_sub_epi32(u[0], sign[0]); + u[1] = _mm256_sub_epi32(u[1], sign[1]); + u[2] = _mm256_sub_epi32(u[2], sign[2]); + u[3] = _mm256_sub_epi32(u[3], sign[3]); + u[4] = _mm256_sub_epi32(u[4], sign[4]); + u[5] = _mm256_sub_epi32(u[5], sign[5]); + u[6] = _mm256_sub_epi32(u[6], sign[6]); + u[7] = _mm256_sub_epi32(u[7], sign[7]); + + u[0] = _mm256_add_epi32(u[0], K32One); + u[1] = _mm256_add_epi32(u[1], K32One); + u[2] = _mm256_add_epi32(u[2], K32One); + u[3] = _mm256_add_epi32(u[3], K32One); + u[4] = _mm256_add_epi32(u[4], K32One); + u[5] = _mm256_add_epi32(u[5], K32One); + u[6] = _mm256_add_epi32(u[6], K32One); + u[7] = _mm256_add_epi32(u[7], K32One); + + u[0] = _mm256_srai_epi32(u[0], 2); + u[1] = _mm256_srai_epi32(u[1], 2); + u[2] = _mm256_srai_epi32(u[2], 2); + u[3] = _mm256_srai_epi32(u[3], 2); + u[4] = _mm256_srai_epi32(u[4], 2); + u[5] = _mm256_srai_epi32(u[5], 2); + u[6] = _mm256_srai_epi32(u[6], 2); + u[7] = _mm256_srai_epi32(u[7], 2); + + // Combine + out[0] = _mm256_packs_epi32(u[0], u[1]); + out[16] = _mm256_packs_epi32(u[2], u[3]); + out[8] = _mm256_packs_epi32(u[4], u[5]); + out[24] = _mm256_packs_epi32(u[6], u[7]); + } + { + const __m256i k32_m08_p24 = + pair256_set_epi32(-cospi_8_64, cospi_24_64); + const __m256i k32_m24_m08 = + pair256_set_epi32(-cospi_24_64, -cospi_8_64); + const __m256i k32_p24_p08 = + pair256_set_epi32(cospi_24_64, cospi_8_64); + + u[0] = _mm256_unpacklo_epi32(lstep1[18], lstep1[28]); + u[1] = _mm256_unpackhi_epi32(lstep1[18], lstep1[28]); + u[2] = _mm256_unpacklo_epi32(lstep1[19], lstep1[29]); + u[3] = _mm256_unpackhi_epi32(lstep1[19], lstep1[29]); + u[4] = _mm256_unpacklo_epi32(lstep1[20], lstep1[26]); + u[5] = _mm256_unpackhi_epi32(lstep1[20], lstep1[26]); + u[6] = _mm256_unpacklo_epi32(lstep1[21], lstep1[27]); + u[7] = _mm256_unpackhi_epi32(lstep1[21], lstep1[27]); + + v[0] = k_madd_epi32_avx2(u[0], k32_m08_p24); + v[1] = k_madd_epi32_avx2(u[1], k32_m08_p24); + v[2] = k_madd_epi32_avx2(u[2], k32_m08_p24); + v[3] = k_madd_epi32_avx2(u[3], k32_m08_p24); + v[4] = k_madd_epi32_avx2(u[4], k32_m24_m08); + v[5] = k_madd_epi32_avx2(u[5], k32_m24_m08); + v[6] = k_madd_epi32_avx2(u[6], k32_m24_m08); + v[7] = k_madd_epi32_avx2(u[7], k32_m24_m08); + v[8] = k_madd_epi32_avx2(u[4], k32_m08_p24); + v[9] = k_madd_epi32_avx2(u[5], k32_m08_p24); + v[10] = k_madd_epi32_avx2(u[6], k32_m08_p24); + v[11] = k_madd_epi32_avx2(u[7], k32_m08_p24); + v[12] = k_madd_epi32_avx2(u[0], k32_p24_p08); + v[13] = k_madd_epi32_avx2(u[1], k32_p24_p08); + v[14] = k_madd_epi32_avx2(u[2], k32_p24_p08); + v[15] = k_madd_epi32_avx2(u[3], k32_p24_p08); + + u[0] = k_packs_epi64_avx2(v[0], v[1]); + u[1] = k_packs_epi64_avx2(v[2], v[3]); + u[2] = k_packs_epi64_avx2(v[4], v[5]); + u[3] = k_packs_epi64_avx2(v[6], v[7]); + u[4] = k_packs_epi64_avx2(v[8], v[9]); + u[5] = k_packs_epi64_avx2(v[10], v[11]); + u[6] = k_packs_epi64_avx2(v[12], v[13]); + u[7] = k_packs_epi64_avx2(v[14], v[15]); + + u[0] = _mm256_add_epi32(u[0], k__DCT_CONST_ROUNDING); + u[1] = _mm256_add_epi32(u[1], k__DCT_CONST_ROUNDING); + u[2] = _mm256_add_epi32(u[2], k__DCT_CONST_ROUNDING); + u[3] = _mm256_add_epi32(u[3], k__DCT_CONST_ROUNDING); + u[4] = _mm256_add_epi32(u[4], k__DCT_CONST_ROUNDING); + u[5] = _mm256_add_epi32(u[5], k__DCT_CONST_ROUNDING); + u[6] = _mm256_add_epi32(u[6], k__DCT_CONST_ROUNDING); + u[7] = _mm256_add_epi32(u[7], k__DCT_CONST_ROUNDING); + + lstep2[18] = _mm256_srai_epi32(u[0], DCT_CONST_BITS); + lstep2[19] = _mm256_srai_epi32(u[1], DCT_CONST_BITS); + lstep2[20] = _mm256_srai_epi32(u[2], DCT_CONST_BITS); + lstep2[21] = _mm256_srai_epi32(u[3], DCT_CONST_BITS); + lstep2[26] = _mm256_srai_epi32(u[4], DCT_CONST_BITS); + lstep2[27] = _mm256_srai_epi32(u[5], DCT_CONST_BITS); + lstep2[28] = _mm256_srai_epi32(u[6], DCT_CONST_BITS); + lstep2[29] = _mm256_srai_epi32(u[7], DCT_CONST_BITS); + } + { + lstep2[32] = _mm256_add_epi32(lstep1[38], lstep3[32]); + lstep2[33] = _mm256_add_epi32(lstep1[39], lstep3[33]); + lstep2[34] = _mm256_add_epi32(lstep1[36], lstep3[34]); + lstep2[35] = _mm256_add_epi32(lstep1[37], lstep3[35]); + lstep2[36] = _mm256_sub_epi32(lstep3[34], lstep1[36]); + lstep2[37] = _mm256_sub_epi32(lstep3[35], lstep1[37]); + lstep2[38] = _mm256_sub_epi32(lstep3[32], lstep1[38]); + lstep2[39] = _mm256_sub_epi32(lstep3[33], lstep1[39]); + lstep2[40] = _mm256_sub_epi32(lstep3[46], lstep1[40]); + lstep2[41] = _mm256_sub_epi32(lstep3[47], lstep1[41]); + lstep2[42] = _mm256_sub_epi32(lstep3[44], lstep1[42]); + lstep2[43] = _mm256_sub_epi32(lstep3[45], lstep1[43]); + lstep2[44] = _mm256_add_epi32(lstep1[42], lstep3[44]); + lstep2[45] = _mm256_add_epi32(lstep1[43], lstep3[45]); + lstep2[46] = _mm256_add_epi32(lstep1[40], lstep3[46]); + lstep2[47] = _mm256_add_epi32(lstep1[41], lstep3[47]); + lstep2[48] = _mm256_add_epi32(lstep1[54], lstep3[48]); + lstep2[49] = _mm256_add_epi32(lstep1[55], lstep3[49]); + lstep2[50] = _mm256_add_epi32(lstep1[52], lstep3[50]); + lstep2[51] = _mm256_add_epi32(lstep1[53], lstep3[51]); + lstep2[52] = _mm256_sub_epi32(lstep3[50], lstep1[52]); + lstep2[53] = _mm256_sub_epi32(lstep3[51], lstep1[53]); + lstep2[54] = _mm256_sub_epi32(lstep3[48], lstep1[54]); + lstep2[55] = _mm256_sub_epi32(lstep3[49], lstep1[55]); + lstep2[56] = _mm256_sub_epi32(lstep3[62], lstep1[56]); + lstep2[57] = _mm256_sub_epi32(lstep3[63], lstep1[57]); + lstep2[58] = _mm256_sub_epi32(lstep3[60], lstep1[58]); + lstep2[59] = _mm256_sub_epi32(lstep3[61], lstep1[59]); + lstep2[60] = _mm256_add_epi32(lstep1[58], lstep3[60]); + lstep2[61] = _mm256_add_epi32(lstep1[59], lstep3[61]); + lstep2[62] = _mm256_add_epi32(lstep1[56], lstep3[62]); + lstep2[63] = _mm256_add_epi32(lstep1[57], lstep3[63]); + } + // stage 6 + { + const __m256i k32_p28_p04 = + pair256_set_epi32(cospi_28_64, cospi_4_64); + const __m256i k32_p12_p20 = + pair256_set_epi32(cospi_12_64, cospi_20_64); + const __m256i k32_m20_p12 = + pair256_set_epi32(-cospi_20_64, cospi_12_64); + const __m256i k32_m04_p28 = + pair256_set_epi32(-cospi_4_64, cospi_28_64); + + u[0] = _mm256_unpacklo_epi32(lstep2[8], lstep2[14]); + u[1] = _mm256_unpackhi_epi32(lstep2[8], lstep2[14]); + u[2] = _mm256_unpacklo_epi32(lstep2[9], lstep2[15]); + u[3] = _mm256_unpackhi_epi32(lstep2[9], lstep2[15]); + u[4] = _mm256_unpacklo_epi32(lstep2[10], lstep2[12]); + u[5] = _mm256_unpackhi_epi32(lstep2[10], lstep2[12]); + u[6] = _mm256_unpacklo_epi32(lstep2[11], lstep2[13]); + u[7] = _mm256_unpackhi_epi32(lstep2[11], lstep2[13]); + u[8] = _mm256_unpacklo_epi32(lstep2[10], lstep2[12]); + u[9] = _mm256_unpackhi_epi32(lstep2[10], lstep2[12]); + u[10] = _mm256_unpacklo_epi32(lstep2[11], lstep2[13]); + u[11] = _mm256_unpackhi_epi32(lstep2[11], lstep2[13]); + u[12] = _mm256_unpacklo_epi32(lstep2[8], lstep2[14]); + u[13] = _mm256_unpackhi_epi32(lstep2[8], lstep2[14]); + u[14] = _mm256_unpacklo_epi32(lstep2[9], lstep2[15]); + u[15] = _mm256_unpackhi_epi32(lstep2[9], lstep2[15]); + + v[0] = k_madd_epi32_avx2(u[0], k32_p28_p04); + v[1] = k_madd_epi32_avx2(u[1], k32_p28_p04); + v[2] = k_madd_epi32_avx2(u[2], k32_p28_p04); + v[3] = k_madd_epi32_avx2(u[3], k32_p28_p04); + v[4] = k_madd_epi32_avx2(u[4], k32_p12_p20); + v[5] = k_madd_epi32_avx2(u[5], k32_p12_p20); + v[6] = k_madd_epi32_avx2(u[6], k32_p12_p20); + v[7] = k_madd_epi32_avx2(u[7], k32_p12_p20); + v[8] = k_madd_epi32_avx2(u[8], k32_m20_p12); + v[9] = k_madd_epi32_avx2(u[9], k32_m20_p12); + v[10] = k_madd_epi32_avx2(u[10], k32_m20_p12); + v[11] = k_madd_epi32_avx2(u[11], k32_m20_p12); + v[12] = k_madd_epi32_avx2(u[12], k32_m04_p28); + v[13] = k_madd_epi32_avx2(u[13], k32_m04_p28); + v[14] = k_madd_epi32_avx2(u[14], k32_m04_p28); + v[15] = k_madd_epi32_avx2(u[15], k32_m04_p28); + + u[0] = k_packs_epi64_avx2(v[0], v[1]); + u[1] = k_packs_epi64_avx2(v[2], v[3]); + u[2] = k_packs_epi64_avx2(v[4], v[5]); + u[3] = k_packs_epi64_avx2(v[6], v[7]); + u[4] = k_packs_epi64_avx2(v[8], v[9]); + u[5] = k_packs_epi64_avx2(v[10], v[11]); + u[6] = k_packs_epi64_avx2(v[12], v[13]); + u[7] = k_packs_epi64_avx2(v[14], v[15]); + + v[0] = _mm256_add_epi32(u[0], k__DCT_CONST_ROUNDING); + v[1] = _mm256_add_epi32(u[1], k__DCT_CONST_ROUNDING); + v[2] = _mm256_add_epi32(u[2], k__DCT_CONST_ROUNDING); + v[3] = _mm256_add_epi32(u[3], k__DCT_CONST_ROUNDING); + v[4] = _mm256_add_epi32(u[4], k__DCT_CONST_ROUNDING); + v[5] = _mm256_add_epi32(u[5], k__DCT_CONST_ROUNDING); + v[6] = _mm256_add_epi32(u[6], k__DCT_CONST_ROUNDING); + v[7] = _mm256_add_epi32(u[7], k__DCT_CONST_ROUNDING); + + u[0] = _mm256_srai_epi32(v[0], DCT_CONST_BITS); + u[1] = _mm256_srai_epi32(v[1], DCT_CONST_BITS); + u[2] = _mm256_srai_epi32(v[2], DCT_CONST_BITS); + u[3] = _mm256_srai_epi32(v[3], DCT_CONST_BITS); + u[4] = _mm256_srai_epi32(v[4], DCT_CONST_BITS); + u[5] = _mm256_srai_epi32(v[5], DCT_CONST_BITS); + u[6] = _mm256_srai_epi32(v[6], DCT_CONST_BITS); + u[7] = _mm256_srai_epi32(v[7], DCT_CONST_BITS); + + sign[0] = _mm256_cmpgt_epi32(kZero, u[0]); + sign[1] = _mm256_cmpgt_epi32(kZero, u[1]); + sign[2] = _mm256_cmpgt_epi32(kZero, u[2]); + sign[3] = _mm256_cmpgt_epi32(kZero, u[3]); + sign[4] = _mm256_cmpgt_epi32(kZero, u[4]); + sign[5] = _mm256_cmpgt_epi32(kZero, u[5]); + sign[6] = _mm256_cmpgt_epi32(kZero, u[6]); + sign[7] = _mm256_cmpgt_epi32(kZero, u[7]); + + u[0] = _mm256_sub_epi32(u[0], sign[0]); + u[1] = _mm256_sub_epi32(u[1], sign[1]); + u[2] = _mm256_sub_epi32(u[2], sign[2]); + u[3] = _mm256_sub_epi32(u[3], sign[3]); + u[4] = _mm256_sub_epi32(u[4], sign[4]); + u[5] = _mm256_sub_epi32(u[5], sign[5]); + u[6] = _mm256_sub_epi32(u[6], sign[6]); + u[7] = _mm256_sub_epi32(u[7], sign[7]); + + u[0] = _mm256_add_epi32(u[0], K32One); + u[1] = _mm256_add_epi32(u[1], K32One); + u[2] = _mm256_add_epi32(u[2], K32One); + u[3] = _mm256_add_epi32(u[3], K32One); + u[4] = _mm256_add_epi32(u[4], K32One); + u[5] = _mm256_add_epi32(u[5], K32One); + u[6] = _mm256_add_epi32(u[6], K32One); + u[7] = _mm256_add_epi32(u[7], K32One); + + u[0] = _mm256_srai_epi32(u[0], 2); + u[1] = _mm256_srai_epi32(u[1], 2); + u[2] = _mm256_srai_epi32(u[2], 2); + u[3] = _mm256_srai_epi32(u[3], 2); + u[4] = _mm256_srai_epi32(u[4], 2); + u[5] = _mm256_srai_epi32(u[5], 2); + u[6] = _mm256_srai_epi32(u[6], 2); + u[7] = _mm256_srai_epi32(u[7], 2); + + out[4] = _mm256_packs_epi32(u[0], u[1]); + out[20] = _mm256_packs_epi32(u[2], u[3]); + out[12] = _mm256_packs_epi32(u[4], u[5]); + out[28] = _mm256_packs_epi32(u[6], u[7]); + } + { + lstep3[16] = _mm256_add_epi32(lstep2[18], lstep1[16]); + lstep3[17] = _mm256_add_epi32(lstep2[19], lstep1[17]); + lstep3[18] = _mm256_sub_epi32(lstep1[16], lstep2[18]); + lstep3[19] = _mm256_sub_epi32(lstep1[17], lstep2[19]); + lstep3[20] = _mm256_sub_epi32(lstep1[22], lstep2[20]); + lstep3[21] = _mm256_sub_epi32(lstep1[23], lstep2[21]); + lstep3[22] = _mm256_add_epi32(lstep2[20], lstep1[22]); + lstep3[23] = _mm256_add_epi32(lstep2[21], lstep1[23]); + lstep3[24] = _mm256_add_epi32(lstep2[26], lstep1[24]); + lstep3[25] = _mm256_add_epi32(lstep2[27], lstep1[25]); + lstep3[26] = _mm256_sub_epi32(lstep1[24], lstep2[26]); + lstep3[27] = _mm256_sub_epi32(lstep1[25], lstep2[27]); + lstep3[28] = _mm256_sub_epi32(lstep1[30], lstep2[28]); + lstep3[29] = _mm256_sub_epi32(lstep1[31], lstep2[29]); + lstep3[30] = _mm256_add_epi32(lstep2[28], lstep1[30]); + lstep3[31] = _mm256_add_epi32(lstep2[29], lstep1[31]); + } + { + const __m256i k32_m04_p28 = + pair256_set_epi32(-cospi_4_64, cospi_28_64); + const __m256i k32_m28_m04 = + pair256_set_epi32(-cospi_28_64, -cospi_4_64); + const __m256i k32_m20_p12 = + pair256_set_epi32(-cospi_20_64, cospi_12_64); + const __m256i k32_m12_m20 = + pair256_set_epi32(-cospi_12_64, -cospi_20_64); + const __m256i k32_p12_p20 = + pair256_set_epi32(cospi_12_64, cospi_20_64); + const __m256i k32_p28_p04 = + pair256_set_epi32(cospi_28_64, cospi_4_64); + + u[0] = _mm256_unpacklo_epi32(lstep2[34], lstep2[60]); + u[1] = _mm256_unpackhi_epi32(lstep2[34], lstep2[60]); + u[2] = _mm256_unpacklo_epi32(lstep2[35], lstep2[61]); + u[3] = _mm256_unpackhi_epi32(lstep2[35], lstep2[61]); + u[4] = _mm256_unpacklo_epi32(lstep2[36], lstep2[58]); + u[5] = _mm256_unpackhi_epi32(lstep2[36], lstep2[58]); + u[6] = _mm256_unpacklo_epi32(lstep2[37], lstep2[59]); + u[7] = _mm256_unpackhi_epi32(lstep2[37], lstep2[59]); + u[8] = _mm256_unpacklo_epi32(lstep2[42], lstep2[52]); + u[9] = _mm256_unpackhi_epi32(lstep2[42], lstep2[52]); + u[10] = _mm256_unpacklo_epi32(lstep2[43], lstep2[53]); + u[11] = _mm256_unpackhi_epi32(lstep2[43], lstep2[53]); + u[12] = _mm256_unpacklo_epi32(lstep2[44], lstep2[50]); + u[13] = _mm256_unpackhi_epi32(lstep2[44], lstep2[50]); + u[14] = _mm256_unpacklo_epi32(lstep2[45], lstep2[51]); + u[15] = _mm256_unpackhi_epi32(lstep2[45], lstep2[51]); + + v[0] = k_madd_epi32_avx2(u[0], k32_m04_p28); + v[1] = k_madd_epi32_avx2(u[1], k32_m04_p28); + v[2] = k_madd_epi32_avx2(u[2], k32_m04_p28); + v[3] = k_madd_epi32_avx2(u[3], k32_m04_p28); + v[4] = k_madd_epi32_avx2(u[4], k32_m28_m04); + v[5] = k_madd_epi32_avx2(u[5], k32_m28_m04); + v[6] = k_madd_epi32_avx2(u[6], k32_m28_m04); + v[7] = k_madd_epi32_avx2(u[7], k32_m28_m04); + v[8] = k_madd_epi32_avx2(u[8], k32_m20_p12); + v[9] = k_madd_epi32_avx2(u[9], k32_m20_p12); + v[10] = k_madd_epi32_avx2(u[10], k32_m20_p12); + v[11] = k_madd_epi32_avx2(u[11], k32_m20_p12); + v[12] = k_madd_epi32_avx2(u[12], k32_m12_m20); + v[13] = k_madd_epi32_avx2(u[13], k32_m12_m20); + v[14] = k_madd_epi32_avx2(u[14], k32_m12_m20); + v[15] = k_madd_epi32_avx2(u[15], k32_m12_m20); + v[16] = k_madd_epi32_avx2(u[12], k32_m20_p12); + v[17] = k_madd_epi32_avx2(u[13], k32_m20_p12); + v[18] = k_madd_epi32_avx2(u[14], k32_m20_p12); + v[19] = k_madd_epi32_avx2(u[15], k32_m20_p12); + v[20] = k_madd_epi32_avx2(u[8], k32_p12_p20); + v[21] = k_madd_epi32_avx2(u[9], k32_p12_p20); + v[22] = k_madd_epi32_avx2(u[10], k32_p12_p20); + v[23] = k_madd_epi32_avx2(u[11], k32_p12_p20); + v[24] = k_madd_epi32_avx2(u[4], k32_m04_p28); + v[25] = k_madd_epi32_avx2(u[5], k32_m04_p28); + v[26] = k_madd_epi32_avx2(u[6], k32_m04_p28); + v[27] = k_madd_epi32_avx2(u[7], k32_m04_p28); + v[28] = k_madd_epi32_avx2(u[0], k32_p28_p04); + v[29] = k_madd_epi32_avx2(u[1], k32_p28_p04); + v[30] = k_madd_epi32_avx2(u[2], k32_p28_p04); + v[31] = k_madd_epi32_avx2(u[3], k32_p28_p04); + + u[0] = k_packs_epi64_avx2(v[0], v[1]); + u[1] = k_packs_epi64_avx2(v[2], v[3]); + u[2] = k_packs_epi64_avx2(v[4], v[5]); + u[3] = k_packs_epi64_avx2(v[6], v[7]); + u[4] = k_packs_epi64_avx2(v[8], v[9]); + u[5] = k_packs_epi64_avx2(v[10], v[11]); + u[6] = k_packs_epi64_avx2(v[12], v[13]); + u[7] = k_packs_epi64_avx2(v[14], v[15]); + u[8] = k_packs_epi64_avx2(v[16], v[17]); + u[9] = k_packs_epi64_avx2(v[18], v[19]); + u[10] = k_packs_epi64_avx2(v[20], v[21]); + u[11] = k_packs_epi64_avx2(v[22], v[23]); + u[12] = k_packs_epi64_avx2(v[24], v[25]); + u[13] = k_packs_epi64_avx2(v[26], v[27]); + u[14] = k_packs_epi64_avx2(v[28], v[29]); + u[15] = k_packs_epi64_avx2(v[30], v[31]); + + v[0] = _mm256_add_epi32(u[0], k__DCT_CONST_ROUNDING); + v[1] = _mm256_add_epi32(u[1], k__DCT_CONST_ROUNDING); + v[2] = _mm256_add_epi32(u[2], k__DCT_CONST_ROUNDING); + v[3] = _mm256_add_epi32(u[3], k__DCT_CONST_ROUNDING); + v[4] = _mm256_add_epi32(u[4], k__DCT_CONST_ROUNDING); + v[5] = _mm256_add_epi32(u[5], k__DCT_CONST_ROUNDING); + v[6] = _mm256_add_epi32(u[6], k__DCT_CONST_ROUNDING); + v[7] = _mm256_add_epi32(u[7], k__DCT_CONST_ROUNDING); + v[8] = _mm256_add_epi32(u[8], k__DCT_CONST_ROUNDING); + v[9] = _mm256_add_epi32(u[9], k__DCT_CONST_ROUNDING); + v[10] = _mm256_add_epi32(u[10], k__DCT_CONST_ROUNDING); + v[11] = _mm256_add_epi32(u[11], k__DCT_CONST_ROUNDING); + v[12] = _mm256_add_epi32(u[12], k__DCT_CONST_ROUNDING); + v[13] = _mm256_add_epi32(u[13], k__DCT_CONST_ROUNDING); + v[14] = _mm256_add_epi32(u[14], k__DCT_CONST_ROUNDING); + v[15] = _mm256_add_epi32(u[15], k__DCT_CONST_ROUNDING); + + lstep3[34] = _mm256_srai_epi32(v[0], DCT_CONST_BITS); + lstep3[35] = _mm256_srai_epi32(v[1], DCT_CONST_BITS); + lstep3[36] = _mm256_srai_epi32(v[2], DCT_CONST_BITS); + lstep3[37] = _mm256_srai_epi32(v[3], DCT_CONST_BITS); + lstep3[42] = _mm256_srai_epi32(v[4], DCT_CONST_BITS); + lstep3[43] = _mm256_srai_epi32(v[5], DCT_CONST_BITS); + lstep3[44] = _mm256_srai_epi32(v[6], DCT_CONST_BITS); + lstep3[45] = _mm256_srai_epi32(v[7], DCT_CONST_BITS); + lstep3[50] = _mm256_srai_epi32(v[8], DCT_CONST_BITS); + lstep3[51] = _mm256_srai_epi32(v[9], DCT_CONST_BITS); + lstep3[52] = _mm256_srai_epi32(v[10], DCT_CONST_BITS); + lstep3[53] = _mm256_srai_epi32(v[11], DCT_CONST_BITS); + lstep3[58] = _mm256_srai_epi32(v[12], DCT_CONST_BITS); + lstep3[59] = _mm256_srai_epi32(v[13], DCT_CONST_BITS); + lstep3[60] = _mm256_srai_epi32(v[14], DCT_CONST_BITS); + lstep3[61] = _mm256_srai_epi32(v[15], DCT_CONST_BITS); + } + // stage 7 + { + const __m256i k32_p30_p02 = + pair256_set_epi32(cospi_30_64, cospi_2_64); + const __m256i k32_p14_p18 = + pair256_set_epi32(cospi_14_64, cospi_18_64); + const __m256i k32_p22_p10 = + pair256_set_epi32(cospi_22_64, cospi_10_64); + const __m256i k32_p06_p26 = + pair256_set_epi32(cospi_6_64, cospi_26_64); + const __m256i k32_m26_p06 = + pair256_set_epi32(-cospi_26_64, cospi_6_64); + const __m256i k32_m10_p22 = + pair256_set_epi32(-cospi_10_64, cospi_22_64); + const __m256i k32_m18_p14 = + pair256_set_epi32(-cospi_18_64, cospi_14_64); + const __m256i k32_m02_p30 = + pair256_set_epi32(-cospi_2_64, cospi_30_64); + + u[0] = _mm256_unpacklo_epi32(lstep3[16], lstep3[30]); + u[1] = _mm256_unpackhi_epi32(lstep3[16], lstep3[30]); + u[2] = _mm256_unpacklo_epi32(lstep3[17], lstep3[31]); + u[3] = _mm256_unpackhi_epi32(lstep3[17], lstep3[31]); + u[4] = _mm256_unpacklo_epi32(lstep3[18], lstep3[28]); + u[5] = _mm256_unpackhi_epi32(lstep3[18], lstep3[28]); + u[6] = _mm256_unpacklo_epi32(lstep3[19], lstep3[29]); + u[7] = _mm256_unpackhi_epi32(lstep3[19], lstep3[29]); + u[8] = _mm256_unpacklo_epi32(lstep3[20], lstep3[26]); + u[9] = _mm256_unpackhi_epi32(lstep3[20], lstep3[26]); + u[10] = _mm256_unpacklo_epi32(lstep3[21], lstep3[27]); + u[11] = _mm256_unpackhi_epi32(lstep3[21], lstep3[27]); + u[12] = _mm256_unpacklo_epi32(lstep3[22], lstep3[24]); + u[13] = _mm256_unpackhi_epi32(lstep3[22], lstep3[24]); + u[14] = _mm256_unpacklo_epi32(lstep3[23], lstep3[25]); + u[15] = _mm256_unpackhi_epi32(lstep3[23], lstep3[25]); + + v[0] = k_madd_epi32_avx2(u[0], k32_p30_p02); + v[1] = k_madd_epi32_avx2(u[1], k32_p30_p02); + v[2] = k_madd_epi32_avx2(u[2], k32_p30_p02); + v[3] = k_madd_epi32_avx2(u[3], k32_p30_p02); + v[4] = k_madd_epi32_avx2(u[4], k32_p14_p18); + v[5] = k_madd_epi32_avx2(u[5], k32_p14_p18); + v[6] = k_madd_epi32_avx2(u[6], k32_p14_p18); + v[7] = k_madd_epi32_avx2(u[7], k32_p14_p18); + v[8] = k_madd_epi32_avx2(u[8], k32_p22_p10); + v[9] = k_madd_epi32_avx2(u[9], k32_p22_p10); + v[10] = k_madd_epi32_avx2(u[10], k32_p22_p10); + v[11] = k_madd_epi32_avx2(u[11], k32_p22_p10); + v[12] = k_madd_epi32_avx2(u[12], k32_p06_p26); + v[13] = k_madd_epi32_avx2(u[13], k32_p06_p26); + v[14] = k_madd_epi32_avx2(u[14], k32_p06_p26); + v[15] = k_madd_epi32_avx2(u[15], k32_p06_p26); + v[16] = k_madd_epi32_avx2(u[12], k32_m26_p06); + v[17] = k_madd_epi32_avx2(u[13], k32_m26_p06); + v[18] = k_madd_epi32_avx2(u[14], k32_m26_p06); + v[19] = k_madd_epi32_avx2(u[15], k32_m26_p06); + v[20] = k_madd_epi32_avx2(u[8], k32_m10_p22); + v[21] = k_madd_epi32_avx2(u[9], k32_m10_p22); + v[22] = k_madd_epi32_avx2(u[10], k32_m10_p22); + v[23] = k_madd_epi32_avx2(u[11], k32_m10_p22); + v[24] = k_madd_epi32_avx2(u[4], k32_m18_p14); + v[25] = k_madd_epi32_avx2(u[5], k32_m18_p14); + v[26] = k_madd_epi32_avx2(u[6], k32_m18_p14); + v[27] = k_madd_epi32_avx2(u[7], k32_m18_p14); + v[28] = k_madd_epi32_avx2(u[0], k32_m02_p30); + v[29] = k_madd_epi32_avx2(u[1], k32_m02_p30); + v[30] = k_madd_epi32_avx2(u[2], k32_m02_p30); + v[31] = k_madd_epi32_avx2(u[3], k32_m02_p30); + + u[0] = k_packs_epi64_avx2(v[0], v[1]); + u[1] = k_packs_epi64_avx2(v[2], v[3]); + u[2] = k_packs_epi64_avx2(v[4], v[5]); + u[3] = k_packs_epi64_avx2(v[6], v[7]); + u[4] = k_packs_epi64_avx2(v[8], v[9]); + u[5] = k_packs_epi64_avx2(v[10], v[11]); + u[6] = k_packs_epi64_avx2(v[12], v[13]); + u[7] = k_packs_epi64_avx2(v[14], v[15]); + u[8] = k_packs_epi64_avx2(v[16], v[17]); + u[9] = k_packs_epi64_avx2(v[18], v[19]); + u[10] = k_packs_epi64_avx2(v[20], v[21]); + u[11] = k_packs_epi64_avx2(v[22], v[23]); + u[12] = k_packs_epi64_avx2(v[24], v[25]); + u[13] = k_packs_epi64_avx2(v[26], v[27]); + u[14] = k_packs_epi64_avx2(v[28], v[29]); + u[15] = k_packs_epi64_avx2(v[30], v[31]); + + v[0] = _mm256_add_epi32(u[0], k__DCT_CONST_ROUNDING); + v[1] = _mm256_add_epi32(u[1], k__DCT_CONST_ROUNDING); + v[2] = _mm256_add_epi32(u[2], k__DCT_CONST_ROUNDING); + v[3] = _mm256_add_epi32(u[3], k__DCT_CONST_ROUNDING); + v[4] = _mm256_add_epi32(u[4], k__DCT_CONST_ROUNDING); + v[5] = _mm256_add_epi32(u[5], k__DCT_CONST_ROUNDING); + v[6] = _mm256_add_epi32(u[6], k__DCT_CONST_ROUNDING); + v[7] = _mm256_add_epi32(u[7], k__DCT_CONST_ROUNDING); + v[8] = _mm256_add_epi32(u[8], k__DCT_CONST_ROUNDING); + v[9] = _mm256_add_epi32(u[9], k__DCT_CONST_ROUNDING); + v[10] = _mm256_add_epi32(u[10], k__DCT_CONST_ROUNDING); + v[11] = _mm256_add_epi32(u[11], k__DCT_CONST_ROUNDING); + v[12] = _mm256_add_epi32(u[12], k__DCT_CONST_ROUNDING); + v[13] = _mm256_add_epi32(u[13], k__DCT_CONST_ROUNDING); + v[14] = _mm256_add_epi32(u[14], k__DCT_CONST_ROUNDING); + v[15] = _mm256_add_epi32(u[15], k__DCT_CONST_ROUNDING); + + u[0] = _mm256_srai_epi32(v[0], DCT_CONST_BITS); + u[1] = _mm256_srai_epi32(v[1], DCT_CONST_BITS); + u[2] = _mm256_srai_epi32(v[2], DCT_CONST_BITS); + u[3] = _mm256_srai_epi32(v[3], DCT_CONST_BITS); + u[4] = _mm256_srai_epi32(v[4], DCT_CONST_BITS); + u[5] = _mm256_srai_epi32(v[5], DCT_CONST_BITS); + u[6] = _mm256_srai_epi32(v[6], DCT_CONST_BITS); + u[7] = _mm256_srai_epi32(v[7], DCT_CONST_BITS); + u[8] = _mm256_srai_epi32(v[8], DCT_CONST_BITS); + u[9] = _mm256_srai_epi32(v[9], DCT_CONST_BITS); + u[10] = _mm256_srai_epi32(v[10], DCT_CONST_BITS); + u[11] = _mm256_srai_epi32(v[11], DCT_CONST_BITS); + u[12] = _mm256_srai_epi32(v[12], DCT_CONST_BITS); + u[13] = _mm256_srai_epi32(v[13], DCT_CONST_BITS); + u[14] = _mm256_srai_epi32(v[14], DCT_CONST_BITS); + u[15] = _mm256_srai_epi32(v[15], DCT_CONST_BITS); + + v[0] = _mm256_cmpgt_epi32(kZero, u[0]); + v[1] = _mm256_cmpgt_epi32(kZero, u[1]); + v[2] = _mm256_cmpgt_epi32(kZero, u[2]); + v[3] = _mm256_cmpgt_epi32(kZero, u[3]); + v[4] = _mm256_cmpgt_epi32(kZero, u[4]); + v[5] = _mm256_cmpgt_epi32(kZero, u[5]); + v[6] = _mm256_cmpgt_epi32(kZero, u[6]); + v[7] = _mm256_cmpgt_epi32(kZero, u[7]); + v[8] = _mm256_cmpgt_epi32(kZero, u[8]); + v[9] = _mm256_cmpgt_epi32(kZero, u[9]); + v[10] = _mm256_cmpgt_epi32(kZero, u[10]); + v[11] = _mm256_cmpgt_epi32(kZero, u[11]); + v[12] = _mm256_cmpgt_epi32(kZero, u[12]); + v[13] = _mm256_cmpgt_epi32(kZero, u[13]); + v[14] = _mm256_cmpgt_epi32(kZero, u[14]); + v[15] = _mm256_cmpgt_epi32(kZero, u[15]); + + u[0] = _mm256_sub_epi32(u[0], v[0]); + u[1] = _mm256_sub_epi32(u[1], v[1]); + u[2] = _mm256_sub_epi32(u[2], v[2]); + u[3] = _mm256_sub_epi32(u[3], v[3]); + u[4] = _mm256_sub_epi32(u[4], v[4]); + u[5] = _mm256_sub_epi32(u[5], v[5]); + u[6] = _mm256_sub_epi32(u[6], v[6]); + u[7] = _mm256_sub_epi32(u[7], v[7]); + u[8] = _mm256_sub_epi32(u[8], v[8]); + u[9] = _mm256_sub_epi32(u[9], v[9]); + u[10] = _mm256_sub_epi32(u[10], v[10]); + u[11] = _mm256_sub_epi32(u[11], v[11]); + u[12] = _mm256_sub_epi32(u[12], v[12]); + u[13] = _mm256_sub_epi32(u[13], v[13]); + u[14] = _mm256_sub_epi32(u[14], v[14]); + u[15] = _mm256_sub_epi32(u[15], v[15]); + + v[0] = _mm256_add_epi32(u[0], K32One); + v[1] = _mm256_add_epi32(u[1], K32One); + v[2] = _mm256_add_epi32(u[2], K32One); + v[3] = _mm256_add_epi32(u[3], K32One); + v[4] = _mm256_add_epi32(u[4], K32One); + v[5] = _mm256_add_epi32(u[5], K32One); + v[6] = _mm256_add_epi32(u[6], K32One); + v[7] = _mm256_add_epi32(u[7], K32One); + v[8] = _mm256_add_epi32(u[8], K32One); + v[9] = _mm256_add_epi32(u[9], K32One); + v[10] = _mm256_add_epi32(u[10], K32One); + v[11] = _mm256_add_epi32(u[11], K32One); + v[12] = _mm256_add_epi32(u[12], K32One); + v[13] = _mm256_add_epi32(u[13], K32One); + v[14] = _mm256_add_epi32(u[14], K32One); + v[15] = _mm256_add_epi32(u[15], K32One); + + u[0] = _mm256_srai_epi32(v[0], 2); + u[1] = _mm256_srai_epi32(v[1], 2); + u[2] = _mm256_srai_epi32(v[2], 2); + u[3] = _mm256_srai_epi32(v[3], 2); + u[4] = _mm256_srai_epi32(v[4], 2); + u[5] = _mm256_srai_epi32(v[5], 2); + u[6] = _mm256_srai_epi32(v[6], 2); + u[7] = _mm256_srai_epi32(v[7], 2); + u[8] = _mm256_srai_epi32(v[8], 2); + u[9] = _mm256_srai_epi32(v[9], 2); + u[10] = _mm256_srai_epi32(v[10], 2); + u[11] = _mm256_srai_epi32(v[11], 2); + u[12] = _mm256_srai_epi32(v[12], 2); + u[13] = _mm256_srai_epi32(v[13], 2); + u[14] = _mm256_srai_epi32(v[14], 2); + u[15] = _mm256_srai_epi32(v[15], 2); + + out[2] = _mm256_packs_epi32(u[0], u[1]); + out[18] = _mm256_packs_epi32(u[2], u[3]); + out[10] = _mm256_packs_epi32(u[4], u[5]); + out[26] = _mm256_packs_epi32(u[6], u[7]); + out[6] = _mm256_packs_epi32(u[8], u[9]); + out[22] = _mm256_packs_epi32(u[10], u[11]); + out[14] = _mm256_packs_epi32(u[12], u[13]); + out[30] = _mm256_packs_epi32(u[14], u[15]); + } + { + lstep1[32] = _mm256_add_epi32(lstep3[34], lstep2[32]); + lstep1[33] = _mm256_add_epi32(lstep3[35], lstep2[33]); + lstep1[34] = _mm256_sub_epi32(lstep2[32], lstep3[34]); + lstep1[35] = _mm256_sub_epi32(lstep2[33], lstep3[35]); + lstep1[36] = _mm256_sub_epi32(lstep2[38], lstep3[36]); + lstep1[37] = _mm256_sub_epi32(lstep2[39], lstep3[37]); + lstep1[38] = _mm256_add_epi32(lstep3[36], lstep2[38]); + lstep1[39] = _mm256_add_epi32(lstep3[37], lstep2[39]); + lstep1[40] = _mm256_add_epi32(lstep3[42], lstep2[40]); + lstep1[41] = _mm256_add_epi32(lstep3[43], lstep2[41]); + lstep1[42] = _mm256_sub_epi32(lstep2[40], lstep3[42]); + lstep1[43] = _mm256_sub_epi32(lstep2[41], lstep3[43]); + lstep1[44] = _mm256_sub_epi32(lstep2[46], lstep3[44]); + lstep1[45] = _mm256_sub_epi32(lstep2[47], lstep3[45]); + lstep1[46] = _mm256_add_epi32(lstep3[44], lstep2[46]); + lstep1[47] = _mm256_add_epi32(lstep3[45], lstep2[47]); + lstep1[48] = _mm256_add_epi32(lstep3[50], lstep2[48]); + lstep1[49] = _mm256_add_epi32(lstep3[51], lstep2[49]); + lstep1[50] = _mm256_sub_epi32(lstep2[48], lstep3[50]); + lstep1[51] = _mm256_sub_epi32(lstep2[49], lstep3[51]); + lstep1[52] = _mm256_sub_epi32(lstep2[54], lstep3[52]); + lstep1[53] = _mm256_sub_epi32(lstep2[55], lstep3[53]); + lstep1[54] = _mm256_add_epi32(lstep3[52], lstep2[54]); + lstep1[55] = _mm256_add_epi32(lstep3[53], lstep2[55]); + lstep1[56] = _mm256_add_epi32(lstep3[58], lstep2[56]); + lstep1[57] = _mm256_add_epi32(lstep3[59], lstep2[57]); + lstep1[58] = _mm256_sub_epi32(lstep2[56], lstep3[58]); + lstep1[59] = _mm256_sub_epi32(lstep2[57], lstep3[59]); + lstep1[60] = _mm256_sub_epi32(lstep2[62], lstep3[60]); + lstep1[61] = _mm256_sub_epi32(lstep2[63], lstep3[61]); + lstep1[62] = _mm256_add_epi32(lstep3[60], lstep2[62]); + lstep1[63] = _mm256_add_epi32(lstep3[61], lstep2[63]); + } + // stage 8 + { + const __m256i k32_p31_p01 = + pair256_set_epi32(cospi_31_64, cospi_1_64); + const __m256i k32_p15_p17 = + pair256_set_epi32(cospi_15_64, cospi_17_64); + const __m256i k32_p23_p09 = + pair256_set_epi32(cospi_23_64, cospi_9_64); + const __m256i k32_p07_p25 = + pair256_set_epi32(cospi_7_64, cospi_25_64); + const __m256i k32_m25_p07 = + pair256_set_epi32(-cospi_25_64, cospi_7_64); + const __m256i k32_m09_p23 = + pair256_set_epi32(-cospi_9_64, cospi_23_64); + const __m256i k32_m17_p15 = + pair256_set_epi32(-cospi_17_64, cospi_15_64); + const __m256i k32_m01_p31 = + pair256_set_epi32(-cospi_1_64, cospi_31_64); + + u[0] = _mm256_unpacklo_epi32(lstep1[32], lstep1[62]); + u[1] = _mm256_unpackhi_epi32(lstep1[32], lstep1[62]); + u[2] = _mm256_unpacklo_epi32(lstep1[33], lstep1[63]); + u[3] = _mm256_unpackhi_epi32(lstep1[33], lstep1[63]); + u[4] = _mm256_unpacklo_epi32(lstep1[34], lstep1[60]); + u[5] = _mm256_unpackhi_epi32(lstep1[34], lstep1[60]); + u[6] = _mm256_unpacklo_epi32(lstep1[35], lstep1[61]); + u[7] = _mm256_unpackhi_epi32(lstep1[35], lstep1[61]); + u[8] = _mm256_unpacklo_epi32(lstep1[36], lstep1[58]); + u[9] = _mm256_unpackhi_epi32(lstep1[36], lstep1[58]); + u[10] = _mm256_unpacklo_epi32(lstep1[37], lstep1[59]); + u[11] = _mm256_unpackhi_epi32(lstep1[37], lstep1[59]); + u[12] = _mm256_unpacklo_epi32(lstep1[38], lstep1[56]); + u[13] = _mm256_unpackhi_epi32(lstep1[38], lstep1[56]); + u[14] = _mm256_unpacklo_epi32(lstep1[39], lstep1[57]); + u[15] = _mm256_unpackhi_epi32(lstep1[39], lstep1[57]); + + v[0] = k_madd_epi32_avx2(u[0], k32_p31_p01); + v[1] = k_madd_epi32_avx2(u[1], k32_p31_p01); + v[2] = k_madd_epi32_avx2(u[2], k32_p31_p01); + v[3] = k_madd_epi32_avx2(u[3], k32_p31_p01); + v[4] = k_madd_epi32_avx2(u[4], k32_p15_p17); + v[5] = k_madd_epi32_avx2(u[5], k32_p15_p17); + v[6] = k_madd_epi32_avx2(u[6], k32_p15_p17); + v[7] = k_madd_epi32_avx2(u[7], k32_p15_p17); + v[8] = k_madd_epi32_avx2(u[8], k32_p23_p09); + v[9] = k_madd_epi32_avx2(u[9], k32_p23_p09); + v[10] = k_madd_epi32_avx2(u[10], k32_p23_p09); + v[11] = k_madd_epi32_avx2(u[11], k32_p23_p09); + v[12] = k_madd_epi32_avx2(u[12], k32_p07_p25); + v[13] = k_madd_epi32_avx2(u[13], k32_p07_p25); + v[14] = k_madd_epi32_avx2(u[14], k32_p07_p25); + v[15] = k_madd_epi32_avx2(u[15], k32_p07_p25); + v[16] = k_madd_epi32_avx2(u[12], k32_m25_p07); + v[17] = k_madd_epi32_avx2(u[13], k32_m25_p07); + v[18] = k_madd_epi32_avx2(u[14], k32_m25_p07); + v[19] = k_madd_epi32_avx2(u[15], k32_m25_p07); + v[20] = k_madd_epi32_avx2(u[8], k32_m09_p23); + v[21] = k_madd_epi32_avx2(u[9], k32_m09_p23); + v[22] = k_madd_epi32_avx2(u[10], k32_m09_p23); + v[23] = k_madd_epi32_avx2(u[11], k32_m09_p23); + v[24] = k_madd_epi32_avx2(u[4], k32_m17_p15); + v[25] = k_madd_epi32_avx2(u[5], k32_m17_p15); + v[26] = k_madd_epi32_avx2(u[6], k32_m17_p15); + v[27] = k_madd_epi32_avx2(u[7], k32_m17_p15); + v[28] = k_madd_epi32_avx2(u[0], k32_m01_p31); + v[29] = k_madd_epi32_avx2(u[1], k32_m01_p31); + v[30] = k_madd_epi32_avx2(u[2], k32_m01_p31); + v[31] = k_madd_epi32_avx2(u[3], k32_m01_p31); + + u[0] = k_packs_epi64_avx2(v[0], v[1]); + u[1] = k_packs_epi64_avx2(v[2], v[3]); + u[2] = k_packs_epi64_avx2(v[4], v[5]); + u[3] = k_packs_epi64_avx2(v[6], v[7]); + u[4] = k_packs_epi64_avx2(v[8], v[9]); + u[5] = k_packs_epi64_avx2(v[10], v[11]); + u[6] = k_packs_epi64_avx2(v[12], v[13]); + u[7] = k_packs_epi64_avx2(v[14], v[15]); + u[8] = k_packs_epi64_avx2(v[16], v[17]); + u[9] = k_packs_epi64_avx2(v[18], v[19]); + u[10] = k_packs_epi64_avx2(v[20], v[21]); + u[11] = k_packs_epi64_avx2(v[22], v[23]); + u[12] = k_packs_epi64_avx2(v[24], v[25]); + u[13] = k_packs_epi64_avx2(v[26], v[27]); + u[14] = k_packs_epi64_avx2(v[28], v[29]); + u[15] = k_packs_epi64_avx2(v[30], v[31]); + + v[0] = _mm256_add_epi32(u[0], k__DCT_CONST_ROUNDING); + v[1] = _mm256_add_epi32(u[1], k__DCT_CONST_ROUNDING); + v[2] = _mm256_add_epi32(u[2], k__DCT_CONST_ROUNDING); + v[3] = _mm256_add_epi32(u[3], k__DCT_CONST_ROUNDING); + v[4] = _mm256_add_epi32(u[4], k__DCT_CONST_ROUNDING); + v[5] = _mm256_add_epi32(u[5], k__DCT_CONST_ROUNDING); + v[6] = _mm256_add_epi32(u[6], k__DCT_CONST_ROUNDING); + v[7] = _mm256_add_epi32(u[7], k__DCT_CONST_ROUNDING); + v[8] = _mm256_add_epi32(u[8], k__DCT_CONST_ROUNDING); + v[9] = _mm256_add_epi32(u[9], k__DCT_CONST_ROUNDING); + v[10] = _mm256_add_epi32(u[10], k__DCT_CONST_ROUNDING); + v[11] = _mm256_add_epi32(u[11], k__DCT_CONST_ROUNDING); + v[12] = _mm256_add_epi32(u[12], k__DCT_CONST_ROUNDING); + v[13] = _mm256_add_epi32(u[13], k__DCT_CONST_ROUNDING); + v[14] = _mm256_add_epi32(u[14], k__DCT_CONST_ROUNDING); + v[15] = _mm256_add_epi32(u[15], k__DCT_CONST_ROUNDING); + + u[0] = _mm256_srai_epi32(v[0], DCT_CONST_BITS); + u[1] = _mm256_srai_epi32(v[1], DCT_CONST_BITS); + u[2] = _mm256_srai_epi32(v[2], DCT_CONST_BITS); + u[3] = _mm256_srai_epi32(v[3], DCT_CONST_BITS); + u[4] = _mm256_srai_epi32(v[4], DCT_CONST_BITS); + u[5] = _mm256_srai_epi32(v[5], DCT_CONST_BITS); + u[6] = _mm256_srai_epi32(v[6], DCT_CONST_BITS); + u[7] = _mm256_srai_epi32(v[7], DCT_CONST_BITS); + u[8] = _mm256_srai_epi32(v[8], DCT_CONST_BITS); + u[9] = _mm256_srai_epi32(v[9], DCT_CONST_BITS); + u[10] = _mm256_srai_epi32(v[10], DCT_CONST_BITS); + u[11] = _mm256_srai_epi32(v[11], DCT_CONST_BITS); + u[12] = _mm256_srai_epi32(v[12], DCT_CONST_BITS); + u[13] = _mm256_srai_epi32(v[13], DCT_CONST_BITS); + u[14] = _mm256_srai_epi32(v[14], DCT_CONST_BITS); + u[15] = _mm256_srai_epi32(v[15], DCT_CONST_BITS); + + v[0] = _mm256_cmpgt_epi32(kZero, u[0]); + v[1] = _mm256_cmpgt_epi32(kZero, u[1]); + v[2] = _mm256_cmpgt_epi32(kZero, u[2]); + v[3] = _mm256_cmpgt_epi32(kZero, u[3]); + v[4] = _mm256_cmpgt_epi32(kZero, u[4]); + v[5] = _mm256_cmpgt_epi32(kZero, u[5]); + v[6] = _mm256_cmpgt_epi32(kZero, u[6]); + v[7] = _mm256_cmpgt_epi32(kZero, u[7]); + v[8] = _mm256_cmpgt_epi32(kZero, u[8]); + v[9] = _mm256_cmpgt_epi32(kZero, u[9]); + v[10] = _mm256_cmpgt_epi32(kZero, u[10]); + v[11] = _mm256_cmpgt_epi32(kZero, u[11]); + v[12] = _mm256_cmpgt_epi32(kZero, u[12]); + v[13] = _mm256_cmpgt_epi32(kZero, u[13]); + v[14] = _mm256_cmpgt_epi32(kZero, u[14]); + v[15] = _mm256_cmpgt_epi32(kZero, u[15]); + + u[0] = _mm256_sub_epi32(u[0], v[0]); + u[1] = _mm256_sub_epi32(u[1], v[1]); + u[2] = _mm256_sub_epi32(u[2], v[2]); + u[3] = _mm256_sub_epi32(u[3], v[3]); + u[4] = _mm256_sub_epi32(u[4], v[4]); + u[5] = _mm256_sub_epi32(u[5], v[5]); + u[6] = _mm256_sub_epi32(u[6], v[6]); + u[7] = _mm256_sub_epi32(u[7], v[7]); + u[8] = _mm256_sub_epi32(u[8], v[8]); + u[9] = _mm256_sub_epi32(u[9], v[9]); + u[10] = _mm256_sub_epi32(u[10], v[10]); + u[11] = _mm256_sub_epi32(u[11], v[11]); + u[12] = _mm256_sub_epi32(u[12], v[12]); + u[13] = _mm256_sub_epi32(u[13], v[13]); + u[14] = _mm256_sub_epi32(u[14], v[14]); + u[15] = _mm256_sub_epi32(u[15], v[15]); + + v[0] = _mm256_add_epi32(u[0], K32One); + v[1] = _mm256_add_epi32(u[1], K32One); + v[2] = _mm256_add_epi32(u[2], K32One); + v[3] = _mm256_add_epi32(u[3], K32One); + v[4] = _mm256_add_epi32(u[4], K32One); + v[5] = _mm256_add_epi32(u[5], K32One); + v[6] = _mm256_add_epi32(u[6], K32One); + v[7] = _mm256_add_epi32(u[7], K32One); + v[8] = _mm256_add_epi32(u[8], K32One); + v[9] = _mm256_add_epi32(u[9], K32One); + v[10] = _mm256_add_epi32(u[10], K32One); + v[11] = _mm256_add_epi32(u[11], K32One); + v[12] = _mm256_add_epi32(u[12], K32One); + v[13] = _mm256_add_epi32(u[13], K32One); + v[14] = _mm256_add_epi32(u[14], K32One); + v[15] = _mm256_add_epi32(u[15], K32One); + + u[0] = _mm256_srai_epi32(v[0], 2); + u[1] = _mm256_srai_epi32(v[1], 2); + u[2] = _mm256_srai_epi32(v[2], 2); + u[3] = _mm256_srai_epi32(v[3], 2); + u[4] = _mm256_srai_epi32(v[4], 2); + u[5] = _mm256_srai_epi32(v[5], 2); + u[6] = _mm256_srai_epi32(v[6], 2); + u[7] = _mm256_srai_epi32(v[7], 2); + u[8] = _mm256_srai_epi32(v[8], 2); + u[9] = _mm256_srai_epi32(v[9], 2); + u[10] = _mm256_srai_epi32(v[10], 2); + u[11] = _mm256_srai_epi32(v[11], 2); + u[12] = _mm256_srai_epi32(v[12], 2); + u[13] = _mm256_srai_epi32(v[13], 2); + u[14] = _mm256_srai_epi32(v[14], 2); + u[15] = _mm256_srai_epi32(v[15], 2); + + out[1] = _mm256_packs_epi32(u[0], u[1]); + out[17] = _mm256_packs_epi32(u[2], u[3]); + out[9] = _mm256_packs_epi32(u[4], u[5]); + out[25] = _mm256_packs_epi32(u[6], u[7]); + out[7] = _mm256_packs_epi32(u[8], u[9]); + out[23] = _mm256_packs_epi32(u[10], u[11]); + out[15] = _mm256_packs_epi32(u[12], u[13]); + out[31] = _mm256_packs_epi32(u[14], u[15]); + } + { + const __m256i k32_p27_p05 = + pair256_set_epi32(cospi_27_64, cospi_5_64); + const __m256i k32_p11_p21 = + pair256_set_epi32(cospi_11_64, cospi_21_64); + const __m256i k32_p19_p13 = + pair256_set_epi32(cospi_19_64, cospi_13_64); + const __m256i k32_p03_p29 = + pair256_set_epi32(cospi_3_64, cospi_29_64); + const __m256i k32_m29_p03 = + pair256_set_epi32(-cospi_29_64, cospi_3_64); + const __m256i k32_m13_p19 = + pair256_set_epi32(-cospi_13_64, cospi_19_64); + const __m256i k32_m21_p11 = + pair256_set_epi32(-cospi_21_64, cospi_11_64); + const __m256i k32_m05_p27 = + pair256_set_epi32(-cospi_5_64, cospi_27_64); + + u[0] = _mm256_unpacklo_epi32(lstep1[40], lstep1[54]); + u[1] = _mm256_unpackhi_epi32(lstep1[40], lstep1[54]); + u[2] = _mm256_unpacklo_epi32(lstep1[41], lstep1[55]); + u[3] = _mm256_unpackhi_epi32(lstep1[41], lstep1[55]); + u[4] = _mm256_unpacklo_epi32(lstep1[42], lstep1[52]); + u[5] = _mm256_unpackhi_epi32(lstep1[42], lstep1[52]); + u[6] = _mm256_unpacklo_epi32(lstep1[43], lstep1[53]); + u[7] = _mm256_unpackhi_epi32(lstep1[43], lstep1[53]); + u[8] = _mm256_unpacklo_epi32(lstep1[44], lstep1[50]); + u[9] = _mm256_unpackhi_epi32(lstep1[44], lstep1[50]); + u[10] = _mm256_unpacklo_epi32(lstep1[45], lstep1[51]); + u[11] = _mm256_unpackhi_epi32(lstep1[45], lstep1[51]); + u[12] = _mm256_unpacklo_epi32(lstep1[46], lstep1[48]); + u[13] = _mm256_unpackhi_epi32(lstep1[46], lstep1[48]); + u[14] = _mm256_unpacklo_epi32(lstep1[47], lstep1[49]); + u[15] = _mm256_unpackhi_epi32(lstep1[47], lstep1[49]); + + v[0] = k_madd_epi32_avx2(u[0], k32_p27_p05); + v[1] = k_madd_epi32_avx2(u[1], k32_p27_p05); + v[2] = k_madd_epi32_avx2(u[2], k32_p27_p05); + v[3] = k_madd_epi32_avx2(u[3], k32_p27_p05); + v[4] = k_madd_epi32_avx2(u[4], k32_p11_p21); + v[5] = k_madd_epi32_avx2(u[5], k32_p11_p21); + v[6] = k_madd_epi32_avx2(u[6], k32_p11_p21); + v[7] = k_madd_epi32_avx2(u[7], k32_p11_p21); + v[8] = k_madd_epi32_avx2(u[8], k32_p19_p13); + v[9] = k_madd_epi32_avx2(u[9], k32_p19_p13); + v[10] = k_madd_epi32_avx2(u[10], k32_p19_p13); + v[11] = k_madd_epi32_avx2(u[11], k32_p19_p13); + v[12] = k_madd_epi32_avx2(u[12], k32_p03_p29); + v[13] = k_madd_epi32_avx2(u[13], k32_p03_p29); + v[14] = k_madd_epi32_avx2(u[14], k32_p03_p29); + v[15] = k_madd_epi32_avx2(u[15], k32_p03_p29); + v[16] = k_madd_epi32_avx2(u[12], k32_m29_p03); + v[17] = k_madd_epi32_avx2(u[13], k32_m29_p03); + v[18] = k_madd_epi32_avx2(u[14], k32_m29_p03); + v[19] = k_madd_epi32_avx2(u[15], k32_m29_p03); + v[20] = k_madd_epi32_avx2(u[8], k32_m13_p19); + v[21] = k_madd_epi32_avx2(u[9], k32_m13_p19); + v[22] = k_madd_epi32_avx2(u[10], k32_m13_p19); + v[23] = k_madd_epi32_avx2(u[11], k32_m13_p19); + v[24] = k_madd_epi32_avx2(u[4], k32_m21_p11); + v[25] = k_madd_epi32_avx2(u[5], k32_m21_p11); + v[26] = k_madd_epi32_avx2(u[6], k32_m21_p11); + v[27] = k_madd_epi32_avx2(u[7], k32_m21_p11); + v[28] = k_madd_epi32_avx2(u[0], k32_m05_p27); + v[29] = k_madd_epi32_avx2(u[1], k32_m05_p27); + v[30] = k_madd_epi32_avx2(u[2], k32_m05_p27); + v[31] = k_madd_epi32_avx2(u[3], k32_m05_p27); + + u[0] = k_packs_epi64_avx2(v[0], v[1]); + u[1] = k_packs_epi64_avx2(v[2], v[3]); + u[2] = k_packs_epi64_avx2(v[4], v[5]); + u[3] = k_packs_epi64_avx2(v[6], v[7]); + u[4] = k_packs_epi64_avx2(v[8], v[9]); + u[5] = k_packs_epi64_avx2(v[10], v[11]); + u[6] = k_packs_epi64_avx2(v[12], v[13]); + u[7] = k_packs_epi64_avx2(v[14], v[15]); + u[8] = k_packs_epi64_avx2(v[16], v[17]); + u[9] = k_packs_epi64_avx2(v[18], v[19]); + u[10] = k_packs_epi64_avx2(v[20], v[21]); + u[11] = k_packs_epi64_avx2(v[22], v[23]); + u[12] = k_packs_epi64_avx2(v[24], v[25]); + u[13] = k_packs_epi64_avx2(v[26], v[27]); + u[14] = k_packs_epi64_avx2(v[28], v[29]); + u[15] = k_packs_epi64_avx2(v[30], v[31]); + + v[0] = _mm256_add_epi32(u[0], k__DCT_CONST_ROUNDING); + v[1] = _mm256_add_epi32(u[1], k__DCT_CONST_ROUNDING); + v[2] = _mm256_add_epi32(u[2], k__DCT_CONST_ROUNDING); + v[3] = _mm256_add_epi32(u[3], k__DCT_CONST_ROUNDING); + v[4] = _mm256_add_epi32(u[4], k__DCT_CONST_ROUNDING); + v[5] = _mm256_add_epi32(u[5], k__DCT_CONST_ROUNDING); + v[6] = _mm256_add_epi32(u[6], k__DCT_CONST_ROUNDING); + v[7] = _mm256_add_epi32(u[7], k__DCT_CONST_ROUNDING); + v[8] = _mm256_add_epi32(u[8], k__DCT_CONST_ROUNDING); + v[9] = _mm256_add_epi32(u[9], k__DCT_CONST_ROUNDING); + v[10] = _mm256_add_epi32(u[10], k__DCT_CONST_ROUNDING); + v[11] = _mm256_add_epi32(u[11], k__DCT_CONST_ROUNDING); + v[12] = _mm256_add_epi32(u[12], k__DCT_CONST_ROUNDING); + v[13] = _mm256_add_epi32(u[13], k__DCT_CONST_ROUNDING); + v[14] = _mm256_add_epi32(u[14], k__DCT_CONST_ROUNDING); + v[15] = _mm256_add_epi32(u[15], k__DCT_CONST_ROUNDING); + + u[0] = _mm256_srai_epi32(v[0], DCT_CONST_BITS); + u[1] = _mm256_srai_epi32(v[1], DCT_CONST_BITS); + u[2] = _mm256_srai_epi32(v[2], DCT_CONST_BITS); + u[3] = _mm256_srai_epi32(v[3], DCT_CONST_BITS); + u[4] = _mm256_srai_epi32(v[4], DCT_CONST_BITS); + u[5] = _mm256_srai_epi32(v[5], DCT_CONST_BITS); + u[6] = _mm256_srai_epi32(v[6], DCT_CONST_BITS); + u[7] = _mm256_srai_epi32(v[7], DCT_CONST_BITS); + u[8] = _mm256_srai_epi32(v[8], DCT_CONST_BITS); + u[9] = _mm256_srai_epi32(v[9], DCT_CONST_BITS); + u[10] = _mm256_srai_epi32(v[10], DCT_CONST_BITS); + u[11] = _mm256_srai_epi32(v[11], DCT_CONST_BITS); + u[12] = _mm256_srai_epi32(v[12], DCT_CONST_BITS); + u[13] = _mm256_srai_epi32(v[13], DCT_CONST_BITS); + u[14] = _mm256_srai_epi32(v[14], DCT_CONST_BITS); + u[15] = _mm256_srai_epi32(v[15], DCT_CONST_BITS); + + v[0] = _mm256_cmpgt_epi32(kZero, u[0]); + v[1] = _mm256_cmpgt_epi32(kZero, u[1]); + v[2] = _mm256_cmpgt_epi32(kZero, u[2]); + v[3] = _mm256_cmpgt_epi32(kZero, u[3]); + v[4] = _mm256_cmpgt_epi32(kZero, u[4]); + v[5] = _mm256_cmpgt_epi32(kZero, u[5]); + v[6] = _mm256_cmpgt_epi32(kZero, u[6]); + v[7] = _mm256_cmpgt_epi32(kZero, u[7]); + v[8] = _mm256_cmpgt_epi32(kZero, u[8]); + v[9] = _mm256_cmpgt_epi32(kZero, u[9]); + v[10] = _mm256_cmpgt_epi32(kZero, u[10]); + v[11] = _mm256_cmpgt_epi32(kZero, u[11]); + v[12] = _mm256_cmpgt_epi32(kZero, u[12]); + v[13] = _mm256_cmpgt_epi32(kZero, u[13]); + v[14] = _mm256_cmpgt_epi32(kZero, u[14]); + v[15] = _mm256_cmpgt_epi32(kZero, u[15]); + + u[0] = _mm256_sub_epi32(u[0], v[0]); + u[1] = _mm256_sub_epi32(u[1], v[1]); + u[2] = _mm256_sub_epi32(u[2], v[2]); + u[3] = _mm256_sub_epi32(u[3], v[3]); + u[4] = _mm256_sub_epi32(u[4], v[4]); + u[5] = _mm256_sub_epi32(u[5], v[5]); + u[6] = _mm256_sub_epi32(u[6], v[6]); + u[7] = _mm256_sub_epi32(u[7], v[7]); + u[8] = _mm256_sub_epi32(u[8], v[8]); + u[9] = _mm256_sub_epi32(u[9], v[9]); + u[10] = _mm256_sub_epi32(u[10], v[10]); + u[11] = _mm256_sub_epi32(u[11], v[11]); + u[12] = _mm256_sub_epi32(u[12], v[12]); + u[13] = _mm256_sub_epi32(u[13], v[13]); + u[14] = _mm256_sub_epi32(u[14], v[14]); + u[15] = _mm256_sub_epi32(u[15], v[15]); + + v[0] = _mm256_add_epi32(u[0], K32One); + v[1] = _mm256_add_epi32(u[1], K32One); + v[2] = _mm256_add_epi32(u[2], K32One); + v[3] = _mm256_add_epi32(u[3], K32One); + v[4] = _mm256_add_epi32(u[4], K32One); + v[5] = _mm256_add_epi32(u[5], K32One); + v[6] = _mm256_add_epi32(u[6], K32One); + v[7] = _mm256_add_epi32(u[7], K32One); + v[8] = _mm256_add_epi32(u[8], K32One); + v[9] = _mm256_add_epi32(u[9], K32One); + v[10] = _mm256_add_epi32(u[10], K32One); + v[11] = _mm256_add_epi32(u[11], K32One); + v[12] = _mm256_add_epi32(u[12], K32One); + v[13] = _mm256_add_epi32(u[13], K32One); + v[14] = _mm256_add_epi32(u[14], K32One); + v[15] = _mm256_add_epi32(u[15], K32One); + + u[0] = _mm256_srai_epi32(v[0], 2); + u[1] = _mm256_srai_epi32(v[1], 2); + u[2] = _mm256_srai_epi32(v[2], 2); + u[3] = _mm256_srai_epi32(v[3], 2); + u[4] = _mm256_srai_epi32(v[4], 2); + u[5] = _mm256_srai_epi32(v[5], 2); + u[6] = _mm256_srai_epi32(v[6], 2); + u[7] = _mm256_srai_epi32(v[7], 2); + u[8] = _mm256_srai_epi32(v[8], 2); + u[9] = _mm256_srai_epi32(v[9], 2); + u[10] = _mm256_srai_epi32(v[10], 2); + u[11] = _mm256_srai_epi32(v[11], 2); + u[12] = _mm256_srai_epi32(v[12], 2); + u[13] = _mm256_srai_epi32(v[13], 2); + u[14] = _mm256_srai_epi32(v[14], 2); + u[15] = _mm256_srai_epi32(v[15], 2); + + out[5] = _mm256_packs_epi32(u[0], u[1]); + out[21] = _mm256_packs_epi32(u[2], u[3]); + out[13] = _mm256_packs_epi32(u[4], u[5]); + out[29] = _mm256_packs_epi32(u[6], u[7]); + out[3] = _mm256_packs_epi32(u[8], u[9]); + out[19] = _mm256_packs_epi32(u[10], u[11]); + out[11] = _mm256_packs_epi32(u[12], u[13]); + out[27] = _mm256_packs_epi32(u[14], u[15]); + } + } +#endif + // Transpose the results, do it as four 8x8 transposes. + { + int transpose_block; + int16_t *output_currStep, *output_nextStep; + tran_low_t *curr_out, *next_out; + // Pass 0 + output_currStep = &intermediate[column_start * 32]; + output_nextStep = &intermediate[(column_start + 8) * 32]; + // Pass 1 + curr_out = &output_org[column_start * 32]; + next_out = &output_org[(column_start + 8) * 32]; + + for (transpose_block = 0; transpose_block < 4; ++transpose_block) { + __m256i *this_out = &out[8 * transpose_block]; + // 00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 + // 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 + // 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 + // 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 + // 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 + // 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 + // 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 + // 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 + const __m256i tr0_0 = _mm256_unpacklo_epi16(this_out[0], this_out[1]); + const __m256i tr0_1 = _mm256_unpacklo_epi16(this_out[2], this_out[3]); + const __m256i tr0_2 = _mm256_unpackhi_epi16(this_out[0], this_out[1]); + const __m256i tr0_3 = _mm256_unpackhi_epi16(this_out[2], this_out[3]); + const __m256i tr0_4 = _mm256_unpacklo_epi16(this_out[4], this_out[5]); + const __m256i tr0_5 = _mm256_unpacklo_epi16(this_out[6], this_out[7]); + const __m256i tr0_6 = _mm256_unpackhi_epi16(this_out[4], this_out[5]); + const __m256i tr0_7 = _mm256_unpackhi_epi16(this_out[6], this_out[7]); + // 00 20 01 21 02 22 03 23 08 28 09 29 10 30 11 31 + // 40 60 41 61 42 62 43 63 48 68 49 69 50 70 51 71 + // 04 24 05 25 06 26 07 27 12 32 13 33 14 34 15 35 + // 44 64 45 65 46 66 47 67 52 72 53 73 54 74 55 75 + // 80 100 81 101 82 102 83 103 88 108 89 109 90 110 91 101 + // 120 140 121 141 122 142 123 143 128 148 129 149 130 150 131 151 + // 84 104 85 105 86 106 87 107 92 112 93 113 94 114 95 115 + // 124 144 125 145 126 146 127 147 132 152 133 153 134 154 135 155 + + const __m256i tr1_0 = _mm256_unpacklo_epi32(tr0_0, tr0_1); + const __m256i tr1_1 = _mm256_unpacklo_epi32(tr0_2, tr0_3); + const __m256i tr1_2 = _mm256_unpackhi_epi32(tr0_0, tr0_1); + const __m256i tr1_3 = _mm256_unpackhi_epi32(tr0_2, tr0_3); + const __m256i tr1_4 = _mm256_unpacklo_epi32(tr0_4, tr0_5); + const __m256i tr1_5 = _mm256_unpacklo_epi32(tr0_6, tr0_7); + const __m256i tr1_6 = _mm256_unpackhi_epi32(tr0_4, tr0_5); + const __m256i tr1_7 = _mm256_unpackhi_epi32(tr0_6, tr0_7); + // 00 20 40 60 01 21 41 61 08 28 48 68 09 29 49 69 + // 04 24 44 64 05 25 45 65 12 32 52 72 13 33 53 73 + // 02 22 42 62 03 23 43 63 10 30 50 70 11 31 51 71 + // 06 26 46 66 07 27 47 67 14 34 54 74 15 35 55 75 + // 80 100 120 140 81 101 121 141 88 108 128 148 89 109 129 149 + // 84 104 124 144 85 105 125 145 92 112 132 152 93 113 133 153 + // 82 102 122 142 83 103 123 143 90 110 130 150 91 101 131 151 + // 86 106 126 146 87 107 127 147 94 114 134 154 95 115 135 155 + __m256i tr2_0 = _mm256_unpacklo_epi64(tr1_0, tr1_4); + __m256i tr2_1 = _mm256_unpackhi_epi64(tr1_0, tr1_4); + __m256i tr2_2 = _mm256_unpacklo_epi64(tr1_2, tr1_6); + __m256i tr2_3 = _mm256_unpackhi_epi64(tr1_2, tr1_6); + __m256i tr2_4 = _mm256_unpacklo_epi64(tr1_1, tr1_5); + __m256i tr2_5 = _mm256_unpackhi_epi64(tr1_1, tr1_5); + __m256i tr2_6 = _mm256_unpacklo_epi64(tr1_3, tr1_7); + __m256i tr2_7 = _mm256_unpackhi_epi64(tr1_3, tr1_7); + // 00 20 40 60 80 100 120 140 08 28 48 68 88 108 128 148 + // 01 21 41 61 81 101 121 141 09 29 49 69 89 109 129 149 + // 02 22 42 62 82 102 122 142 10 30 50 70 90 110 130 150 + // 03 23 43 63 83 103 123 143 11 31 51 71 91 101 131 151 + // 04 24 44 64 84 104 124 144 12 32 52 72 92 112 132 152 + // 05 25 45 65 85 105 125 145 13 33 53 73 93 113 133 153 + // 06 26 46 66 86 106 126 146 14 34 54 74 94 114 134 154 + // 07 27 47 67 87 107 127 147 15 35 55 75 95 115 135 155 + if (0 == pass) { + // output[j] = (output[j] + 1 + (output[j] > 0)) >> 2; + // TODO(cd): see quality impact of only doing + // output[j] = (output[j] + 1) >> 2; + // which would remove the code between here ... + __m256i tr2_0_0 = _mm256_cmpgt_epi16(tr2_0, kZero); + __m256i tr2_1_0 = _mm256_cmpgt_epi16(tr2_1, kZero); + __m256i tr2_2_0 = _mm256_cmpgt_epi16(tr2_2, kZero); + __m256i tr2_3_0 = _mm256_cmpgt_epi16(tr2_3, kZero); + __m256i tr2_4_0 = _mm256_cmpgt_epi16(tr2_4, kZero); + __m256i tr2_5_0 = _mm256_cmpgt_epi16(tr2_5, kZero); + __m256i tr2_6_0 = _mm256_cmpgt_epi16(tr2_6, kZero); + __m256i tr2_7_0 = _mm256_cmpgt_epi16(tr2_7, kZero); + tr2_0 = _mm256_sub_epi16(tr2_0, tr2_0_0); + tr2_1 = _mm256_sub_epi16(tr2_1, tr2_1_0); + tr2_2 = _mm256_sub_epi16(tr2_2, tr2_2_0); + tr2_3 = _mm256_sub_epi16(tr2_3, tr2_3_0); + tr2_4 = _mm256_sub_epi16(tr2_4, tr2_4_0); + tr2_5 = _mm256_sub_epi16(tr2_5, tr2_5_0); + tr2_6 = _mm256_sub_epi16(tr2_6, tr2_6_0); + tr2_7 = _mm256_sub_epi16(tr2_7, tr2_7_0); + // ... and here. + // PS: also change code in av1/encoder/av1_dct.c + tr2_0 = _mm256_add_epi16(tr2_0, kOne); + tr2_1 = _mm256_add_epi16(tr2_1, kOne); + tr2_2 = _mm256_add_epi16(tr2_2, kOne); + tr2_3 = _mm256_add_epi16(tr2_3, kOne); + tr2_4 = _mm256_add_epi16(tr2_4, kOne); + tr2_5 = _mm256_add_epi16(tr2_5, kOne); + tr2_6 = _mm256_add_epi16(tr2_6, kOne); + tr2_7 = _mm256_add_epi16(tr2_7, kOne); + tr2_0 = _mm256_srai_epi16(tr2_0, 2); + tr2_1 = _mm256_srai_epi16(tr2_1, 2); + tr2_2 = _mm256_srai_epi16(tr2_2, 2); + tr2_3 = _mm256_srai_epi16(tr2_3, 2); + tr2_4 = _mm256_srai_epi16(tr2_4, 2); + tr2_5 = _mm256_srai_epi16(tr2_5, 2); + tr2_6 = _mm256_srai_epi16(tr2_6, 2); + tr2_7 = _mm256_srai_epi16(tr2_7, 2); + } + if (0 == pass) { + // Note: even though all these stores are aligned, using the aligned + // intrinsic make the code slightly slower. + _mm_storeu_si128((__m128i *)(output_currStep + 0 * 32), + _mm256_castsi256_si128(tr2_0)); + _mm_storeu_si128((__m128i *)(output_currStep + 1 * 32), + _mm256_castsi256_si128(tr2_1)); + _mm_storeu_si128((__m128i *)(output_currStep + 2 * 32), + _mm256_castsi256_si128(tr2_2)); + _mm_storeu_si128((__m128i *)(output_currStep + 3 * 32), + _mm256_castsi256_si128(tr2_3)); + _mm_storeu_si128((__m128i *)(output_currStep + 4 * 32), + _mm256_castsi256_si128(tr2_4)); + _mm_storeu_si128((__m128i *)(output_currStep + 5 * 32), + _mm256_castsi256_si128(tr2_5)); + _mm_storeu_si128((__m128i *)(output_currStep + 6 * 32), + _mm256_castsi256_si128(tr2_6)); + _mm_storeu_si128((__m128i *)(output_currStep + 7 * 32), + _mm256_castsi256_si128(tr2_7)); + + _mm_storeu_si128((__m128i *)(output_nextStep + 0 * 32), + _mm256_extractf128_si256(tr2_0, 1)); + _mm_storeu_si128((__m128i *)(output_nextStep + 1 * 32), + _mm256_extractf128_si256(tr2_1, 1)); + _mm_storeu_si128((__m128i *)(output_nextStep + 2 * 32), + _mm256_extractf128_si256(tr2_2, 1)); + _mm_storeu_si128((__m128i *)(output_nextStep + 3 * 32), + _mm256_extractf128_si256(tr2_3, 1)); + _mm_storeu_si128((__m128i *)(output_nextStep + 4 * 32), + _mm256_extractf128_si256(tr2_4, 1)); + _mm_storeu_si128((__m128i *)(output_nextStep + 5 * 32), + _mm256_extractf128_si256(tr2_5, 1)); + _mm_storeu_si128((__m128i *)(output_nextStep + 6 * 32), + _mm256_extractf128_si256(tr2_6, 1)); + _mm_storeu_si128((__m128i *)(output_nextStep + 7 * 32), + _mm256_extractf128_si256(tr2_7, 1)); + // Process next 8x8 + output_currStep += 8; + output_nextStep += 8; + } + if (1 == pass) { + store_coeff(&tr2_0, curr_out + 0 * 32, next_out + 0 * 32); + store_coeff(&tr2_1, curr_out + 1 * 32, next_out + 1 * 32); + store_coeff(&tr2_2, curr_out + 2 * 32, next_out + 2 * 32); + store_coeff(&tr2_3, curr_out + 3 * 32, next_out + 3 * 32); + store_coeff(&tr2_4, curr_out + 4 * 32, next_out + 4 * 32); + store_coeff(&tr2_5, curr_out + 5 * 32, next_out + 5 * 32); + store_coeff(&tr2_6, curr_out + 6 * 32, next_out + 6 * 32); + store_coeff(&tr2_7, curr_out + 7 * 32, next_out + 7 * 32); + curr_out += 8; + next_out += 8; + } + } + } + } + } + _mm256_zeroupper(); +} // NOLINT diff --git a/third_party/aom/aom_dsp/x86/fwd_dct32x32_impl_sse2.h b/third_party/aom/aom_dsp/x86/fwd_dct32x32_impl_sse2.h new file mode 100644 index 000000000..69dd6af11 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/fwd_dct32x32_impl_sse2.h @@ -0,0 +1,3201 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include <emmintrin.h> // SSE2 + +#include "aom_dsp/fwd_txfm.h" +#include "aom_dsp/txfm_common.h" +#include "aom_dsp/x86/txfm_common_sse2.h" + +// TODO(jingning) The high bit-depth version needs re-work for performance. +// The current SSE2 implementation also causes cross reference to the static +// functions in the C implementation file. +#if DCT_HIGH_BIT_DEPTH +#define ADD_EPI16 _mm_adds_epi16 +#define SUB_EPI16 _mm_subs_epi16 +#if FDCT32x32_HIGH_PRECISION +void aom_fdct32x32_rows_c(const int16_t *intermediate, tran_low_t *out) { + int i, j; + for (i = 0; i < 32; ++i) { + tran_high_t temp_in[32], temp_out[32]; + for (j = 0; j < 32; ++j) temp_in[j] = intermediate[j * 32 + i]; + aom_fdct32(temp_in, temp_out, 0); + for (j = 0; j < 32; ++j) + out[j + i * 32] = + (tran_low_t)((temp_out[j] + 1 + (temp_out[j] < 0)) >> 2); + } +} +#define HIGH_FDCT32x32_2D_C aom_highbd_fdct32x32_c +#define HIGH_FDCT32x32_2D_ROWS_C aom_fdct32x32_rows_c +#else +void aom_fdct32x32_rd_rows_c(const int16_t *intermediate, tran_low_t *out) { + int i, j; + for (i = 0; i < 32; ++i) { + tran_high_t temp_in[32], temp_out[32]; + for (j = 0; j < 32; ++j) temp_in[j] = intermediate[j * 32 + i]; + aom_fdct32(temp_in, temp_out, 1); + for (j = 0; j < 32; ++j) out[j + i * 32] = (tran_low_t)temp_out[j]; + } +} +#define HIGH_FDCT32x32_2D_C aom_highbd_fdct32x32_rd_c +#define HIGH_FDCT32x32_2D_ROWS_C aom_fdct32x32_rd_rows_c +#endif // FDCT32x32_HIGH_PRECISION +#else +#define ADD_EPI16 _mm_add_epi16 +#define SUB_EPI16 _mm_sub_epi16 +#endif // DCT_HIGH_BIT_DEPTH + +void FDCT32x32_2D(const int16_t *input, tran_low_t *output_org, int stride) { + // Calculate pre-multiplied strides + const int str1 = stride; + const int str2 = 2 * stride; + const int str3 = 2 * stride + str1; + // We need an intermediate buffer between passes. + DECLARE_ALIGNED(16, int16_t, intermediate[32 * 32]); + // Constants + // When we use them, in one case, they are all the same. In all others + // it's a pair of them that we need to repeat four times. This is done + // by constructing the 32 bit constant corresponding to that pair. + const __m128i k__cospi_p16_p16 = _mm_set1_epi16((int16_t)cospi_16_64); + const __m128i k__cospi_p16_m16 = pair_set_epi16(+cospi_16_64, -cospi_16_64); + const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64, cospi_24_64); + const __m128i k__cospi_m24_m08 = pair_set_epi16(-cospi_24_64, -cospi_8_64); + const __m128i k__cospi_p24_p08 = pair_set_epi16(+cospi_24_64, cospi_8_64); + const __m128i k__cospi_p12_p20 = pair_set_epi16(+cospi_12_64, cospi_20_64); + const __m128i k__cospi_m20_p12 = pair_set_epi16(-cospi_20_64, cospi_12_64); + const __m128i k__cospi_m04_p28 = pair_set_epi16(-cospi_4_64, cospi_28_64); + const __m128i k__cospi_p28_p04 = pair_set_epi16(+cospi_28_64, cospi_4_64); + const __m128i k__cospi_m28_m04 = pair_set_epi16(-cospi_28_64, -cospi_4_64); + const __m128i k__cospi_m12_m20 = pair_set_epi16(-cospi_12_64, -cospi_20_64); + const __m128i k__cospi_p30_p02 = pair_set_epi16(+cospi_30_64, cospi_2_64); + const __m128i k__cospi_p14_p18 = pair_set_epi16(+cospi_14_64, cospi_18_64); + const __m128i k__cospi_p22_p10 = pair_set_epi16(+cospi_22_64, cospi_10_64); + const __m128i k__cospi_p06_p26 = pair_set_epi16(+cospi_6_64, cospi_26_64); + const __m128i k__cospi_m26_p06 = pair_set_epi16(-cospi_26_64, cospi_6_64); + const __m128i k__cospi_m10_p22 = pair_set_epi16(-cospi_10_64, cospi_22_64); + const __m128i k__cospi_m18_p14 = pair_set_epi16(-cospi_18_64, cospi_14_64); + const __m128i k__cospi_m02_p30 = pair_set_epi16(-cospi_2_64, cospi_30_64); + const __m128i k__cospi_p31_p01 = pair_set_epi16(+cospi_31_64, cospi_1_64); + const __m128i k__cospi_p15_p17 = pair_set_epi16(+cospi_15_64, cospi_17_64); + const __m128i k__cospi_p23_p09 = pair_set_epi16(+cospi_23_64, cospi_9_64); + const __m128i k__cospi_p07_p25 = pair_set_epi16(+cospi_7_64, cospi_25_64); + const __m128i k__cospi_m25_p07 = pair_set_epi16(-cospi_25_64, cospi_7_64); + const __m128i k__cospi_m09_p23 = pair_set_epi16(-cospi_9_64, cospi_23_64); + const __m128i k__cospi_m17_p15 = pair_set_epi16(-cospi_17_64, cospi_15_64); + const __m128i k__cospi_m01_p31 = pair_set_epi16(-cospi_1_64, cospi_31_64); + const __m128i k__cospi_p27_p05 = pair_set_epi16(+cospi_27_64, cospi_5_64); + const __m128i k__cospi_p11_p21 = pair_set_epi16(+cospi_11_64, cospi_21_64); + const __m128i k__cospi_p19_p13 = pair_set_epi16(+cospi_19_64, cospi_13_64); + const __m128i k__cospi_p03_p29 = pair_set_epi16(+cospi_3_64, cospi_29_64); + const __m128i k__cospi_m29_p03 = pair_set_epi16(-cospi_29_64, cospi_3_64); + const __m128i k__cospi_m13_p19 = pair_set_epi16(-cospi_13_64, cospi_19_64); + const __m128i k__cospi_m21_p11 = pair_set_epi16(-cospi_21_64, cospi_11_64); + const __m128i k__cospi_m05_p27 = pair_set_epi16(-cospi_5_64, cospi_27_64); + const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING); + const __m128i kZero = _mm_set1_epi16(0); + const __m128i kOne = _mm_set1_epi16(1); + // Do the two transform/transpose passes + int pass; +#if DCT_HIGH_BIT_DEPTH + int overflow; +#endif + for (pass = 0; pass < 2; ++pass) { + // We process eight columns (transposed rows in second pass) at a time. + int column_start; + for (column_start = 0; column_start < 32; column_start += 8) { + __m128i step1[32]; + __m128i step2[32]; + __m128i step3[32]; + __m128i out[32]; + // Stage 1 + // Note: even though all the loads below are aligned, using the aligned + // intrinsic make the code slightly slower. + if (0 == pass) { + const int16_t *in = &input[column_start]; + // step1[i] = (in[ 0 * stride] + in[(32 - 1) * stride]) << 2; + // Note: the next four blocks could be in a loop. That would help the + // instruction cache but is actually slower. + { + const int16_t *ina = in + 0 * str1; + const int16_t *inb = in + 31 * str1; + __m128i *step1a = &step1[0]; + __m128i *step1b = &step1[31]; + const __m128i ina0 = _mm_loadu_si128((const __m128i *)(ina)); + const __m128i ina1 = _mm_loadu_si128((const __m128i *)(ina + str1)); + const __m128i ina2 = _mm_loadu_si128((const __m128i *)(ina + str2)); + const __m128i ina3 = _mm_loadu_si128((const __m128i *)(ina + str3)); + const __m128i inb3 = _mm_loadu_si128((const __m128i *)(inb - str3)); + const __m128i inb2 = _mm_loadu_si128((const __m128i *)(inb - str2)); + const __m128i inb1 = _mm_loadu_si128((const __m128i *)(inb - str1)); + const __m128i inb0 = _mm_loadu_si128((const __m128i *)(inb)); + step1a[0] = _mm_add_epi16(ina0, inb0); + step1a[1] = _mm_add_epi16(ina1, inb1); + step1a[2] = _mm_add_epi16(ina2, inb2); + step1a[3] = _mm_add_epi16(ina3, inb3); + step1b[-3] = _mm_sub_epi16(ina3, inb3); + step1b[-2] = _mm_sub_epi16(ina2, inb2); + step1b[-1] = _mm_sub_epi16(ina1, inb1); + step1b[-0] = _mm_sub_epi16(ina0, inb0); + step1a[0] = _mm_slli_epi16(step1a[0], 2); + step1a[1] = _mm_slli_epi16(step1a[1], 2); + step1a[2] = _mm_slli_epi16(step1a[2], 2); + step1a[3] = _mm_slli_epi16(step1a[3], 2); + step1b[-3] = _mm_slli_epi16(step1b[-3], 2); + step1b[-2] = _mm_slli_epi16(step1b[-2], 2); + step1b[-1] = _mm_slli_epi16(step1b[-1], 2); + step1b[-0] = _mm_slli_epi16(step1b[-0], 2); + } + { + const int16_t *ina = in + 4 * str1; + const int16_t *inb = in + 27 * str1; + __m128i *step1a = &step1[4]; + __m128i *step1b = &step1[27]; + const __m128i ina0 = _mm_loadu_si128((const __m128i *)(ina)); + const __m128i ina1 = _mm_loadu_si128((const __m128i *)(ina + str1)); + const __m128i ina2 = _mm_loadu_si128((const __m128i *)(ina + str2)); + const __m128i ina3 = _mm_loadu_si128((const __m128i *)(ina + str3)); + const __m128i inb3 = _mm_loadu_si128((const __m128i *)(inb - str3)); + const __m128i inb2 = _mm_loadu_si128((const __m128i *)(inb - str2)); + const __m128i inb1 = _mm_loadu_si128((const __m128i *)(inb - str1)); + const __m128i inb0 = _mm_loadu_si128((const __m128i *)(inb)); + step1a[0] = _mm_add_epi16(ina0, inb0); + step1a[1] = _mm_add_epi16(ina1, inb1); + step1a[2] = _mm_add_epi16(ina2, inb2); + step1a[3] = _mm_add_epi16(ina3, inb3); + step1b[-3] = _mm_sub_epi16(ina3, inb3); + step1b[-2] = _mm_sub_epi16(ina2, inb2); + step1b[-1] = _mm_sub_epi16(ina1, inb1); + step1b[-0] = _mm_sub_epi16(ina0, inb0); + step1a[0] = _mm_slli_epi16(step1a[0], 2); + step1a[1] = _mm_slli_epi16(step1a[1], 2); + step1a[2] = _mm_slli_epi16(step1a[2], 2); + step1a[3] = _mm_slli_epi16(step1a[3], 2); + step1b[-3] = _mm_slli_epi16(step1b[-3], 2); + step1b[-2] = _mm_slli_epi16(step1b[-2], 2); + step1b[-1] = _mm_slli_epi16(step1b[-1], 2); + step1b[-0] = _mm_slli_epi16(step1b[-0], 2); + } + { + const int16_t *ina = in + 8 * str1; + const int16_t *inb = in + 23 * str1; + __m128i *step1a = &step1[8]; + __m128i *step1b = &step1[23]; + const __m128i ina0 = _mm_loadu_si128((const __m128i *)(ina)); + const __m128i ina1 = _mm_loadu_si128((const __m128i *)(ina + str1)); + const __m128i ina2 = _mm_loadu_si128((const __m128i *)(ina + str2)); + const __m128i ina3 = _mm_loadu_si128((const __m128i *)(ina + str3)); + const __m128i inb3 = _mm_loadu_si128((const __m128i *)(inb - str3)); + const __m128i inb2 = _mm_loadu_si128((const __m128i *)(inb - str2)); + const __m128i inb1 = _mm_loadu_si128((const __m128i *)(inb - str1)); + const __m128i inb0 = _mm_loadu_si128((const __m128i *)(inb)); + step1a[0] = _mm_add_epi16(ina0, inb0); + step1a[1] = _mm_add_epi16(ina1, inb1); + step1a[2] = _mm_add_epi16(ina2, inb2); + step1a[3] = _mm_add_epi16(ina3, inb3); + step1b[-3] = _mm_sub_epi16(ina3, inb3); + step1b[-2] = _mm_sub_epi16(ina2, inb2); + step1b[-1] = _mm_sub_epi16(ina1, inb1); + step1b[-0] = _mm_sub_epi16(ina0, inb0); + step1a[0] = _mm_slli_epi16(step1a[0], 2); + step1a[1] = _mm_slli_epi16(step1a[1], 2); + step1a[2] = _mm_slli_epi16(step1a[2], 2); + step1a[3] = _mm_slli_epi16(step1a[3], 2); + step1b[-3] = _mm_slli_epi16(step1b[-3], 2); + step1b[-2] = _mm_slli_epi16(step1b[-2], 2); + step1b[-1] = _mm_slli_epi16(step1b[-1], 2); + step1b[-0] = _mm_slli_epi16(step1b[-0], 2); + } + { + const int16_t *ina = in + 12 * str1; + const int16_t *inb = in + 19 * str1; + __m128i *step1a = &step1[12]; + __m128i *step1b = &step1[19]; + const __m128i ina0 = _mm_loadu_si128((const __m128i *)(ina)); + const __m128i ina1 = _mm_loadu_si128((const __m128i *)(ina + str1)); + const __m128i ina2 = _mm_loadu_si128((const __m128i *)(ina + str2)); + const __m128i ina3 = _mm_loadu_si128((const __m128i *)(ina + str3)); + const __m128i inb3 = _mm_loadu_si128((const __m128i *)(inb - str3)); + const __m128i inb2 = _mm_loadu_si128((const __m128i *)(inb - str2)); + const __m128i inb1 = _mm_loadu_si128((const __m128i *)(inb - str1)); + const __m128i inb0 = _mm_loadu_si128((const __m128i *)(inb)); + step1a[0] = _mm_add_epi16(ina0, inb0); + step1a[1] = _mm_add_epi16(ina1, inb1); + step1a[2] = _mm_add_epi16(ina2, inb2); + step1a[3] = _mm_add_epi16(ina3, inb3); + step1b[-3] = _mm_sub_epi16(ina3, inb3); + step1b[-2] = _mm_sub_epi16(ina2, inb2); + step1b[-1] = _mm_sub_epi16(ina1, inb1); + step1b[-0] = _mm_sub_epi16(ina0, inb0); + step1a[0] = _mm_slli_epi16(step1a[0], 2); + step1a[1] = _mm_slli_epi16(step1a[1], 2); + step1a[2] = _mm_slli_epi16(step1a[2], 2); + step1a[3] = _mm_slli_epi16(step1a[3], 2); + step1b[-3] = _mm_slli_epi16(step1b[-3], 2); + step1b[-2] = _mm_slli_epi16(step1b[-2], 2); + step1b[-1] = _mm_slli_epi16(step1b[-1], 2); + step1b[-0] = _mm_slli_epi16(step1b[-0], 2); + } + } else { + int16_t *in = &intermediate[column_start]; + // step1[i] = in[ 0 * 32] + in[(32 - 1) * 32]; + // Note: using the same approach as above to have common offset is + // counter-productive as all offsets can be calculated at compile + // time. + // Note: the next four blocks could be in a loop. That would help the + // instruction cache but is actually slower. + { + __m128i in00 = _mm_loadu_si128((const __m128i *)(in + 0 * 32)); + __m128i in01 = _mm_loadu_si128((const __m128i *)(in + 1 * 32)); + __m128i in02 = _mm_loadu_si128((const __m128i *)(in + 2 * 32)); + __m128i in03 = _mm_loadu_si128((const __m128i *)(in + 3 * 32)); + __m128i in28 = _mm_loadu_si128((const __m128i *)(in + 28 * 32)); + __m128i in29 = _mm_loadu_si128((const __m128i *)(in + 29 * 32)); + __m128i in30 = _mm_loadu_si128((const __m128i *)(in + 30 * 32)); + __m128i in31 = _mm_loadu_si128((const __m128i *)(in + 31 * 32)); + step1[0] = ADD_EPI16(in00, in31); + step1[1] = ADD_EPI16(in01, in30); + step1[2] = ADD_EPI16(in02, in29); + step1[3] = ADD_EPI16(in03, in28); + step1[28] = SUB_EPI16(in03, in28); + step1[29] = SUB_EPI16(in02, in29); + step1[30] = SUB_EPI16(in01, in30); + step1[31] = SUB_EPI16(in00, in31); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x8(&step1[0], &step1[1], &step1[2], + &step1[3], &step1[28], &step1[29], + &step1[30], &step1[31]); + if (overflow) { + HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + } + { + __m128i in04 = _mm_loadu_si128((const __m128i *)(in + 4 * 32)); + __m128i in05 = _mm_loadu_si128((const __m128i *)(in + 5 * 32)); + __m128i in06 = _mm_loadu_si128((const __m128i *)(in + 6 * 32)); + __m128i in07 = _mm_loadu_si128((const __m128i *)(in + 7 * 32)); + __m128i in24 = _mm_loadu_si128((const __m128i *)(in + 24 * 32)); + __m128i in25 = _mm_loadu_si128((const __m128i *)(in + 25 * 32)); + __m128i in26 = _mm_loadu_si128((const __m128i *)(in + 26 * 32)); + __m128i in27 = _mm_loadu_si128((const __m128i *)(in + 27 * 32)); + step1[4] = ADD_EPI16(in04, in27); + step1[5] = ADD_EPI16(in05, in26); + step1[6] = ADD_EPI16(in06, in25); + step1[7] = ADD_EPI16(in07, in24); + step1[24] = SUB_EPI16(in07, in24); + step1[25] = SUB_EPI16(in06, in25); + step1[26] = SUB_EPI16(in05, in26); + step1[27] = SUB_EPI16(in04, in27); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x8(&step1[4], &step1[5], &step1[6], + &step1[7], &step1[24], &step1[25], + &step1[26], &step1[27]); + if (overflow) { + HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + } + { + __m128i in08 = _mm_loadu_si128((const __m128i *)(in + 8 * 32)); + __m128i in09 = _mm_loadu_si128((const __m128i *)(in + 9 * 32)); + __m128i in10 = _mm_loadu_si128((const __m128i *)(in + 10 * 32)); + __m128i in11 = _mm_loadu_si128((const __m128i *)(in + 11 * 32)); + __m128i in20 = _mm_loadu_si128((const __m128i *)(in + 20 * 32)); + __m128i in21 = _mm_loadu_si128((const __m128i *)(in + 21 * 32)); + __m128i in22 = _mm_loadu_si128((const __m128i *)(in + 22 * 32)); + __m128i in23 = _mm_loadu_si128((const __m128i *)(in + 23 * 32)); + step1[8] = ADD_EPI16(in08, in23); + step1[9] = ADD_EPI16(in09, in22); + step1[10] = ADD_EPI16(in10, in21); + step1[11] = ADD_EPI16(in11, in20); + step1[20] = SUB_EPI16(in11, in20); + step1[21] = SUB_EPI16(in10, in21); + step1[22] = SUB_EPI16(in09, in22); + step1[23] = SUB_EPI16(in08, in23); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x8(&step1[8], &step1[9], &step1[10], + &step1[11], &step1[20], &step1[21], + &step1[22], &step1[23]); + if (overflow) { + HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + } + { + __m128i in12 = _mm_loadu_si128((const __m128i *)(in + 12 * 32)); + __m128i in13 = _mm_loadu_si128((const __m128i *)(in + 13 * 32)); + __m128i in14 = _mm_loadu_si128((const __m128i *)(in + 14 * 32)); + __m128i in15 = _mm_loadu_si128((const __m128i *)(in + 15 * 32)); + __m128i in16 = _mm_loadu_si128((const __m128i *)(in + 16 * 32)); + __m128i in17 = _mm_loadu_si128((const __m128i *)(in + 17 * 32)); + __m128i in18 = _mm_loadu_si128((const __m128i *)(in + 18 * 32)); + __m128i in19 = _mm_loadu_si128((const __m128i *)(in + 19 * 32)); + step1[12] = ADD_EPI16(in12, in19); + step1[13] = ADD_EPI16(in13, in18); + step1[14] = ADD_EPI16(in14, in17); + step1[15] = ADD_EPI16(in15, in16); + step1[16] = SUB_EPI16(in15, in16); + step1[17] = SUB_EPI16(in14, in17); + step1[18] = SUB_EPI16(in13, in18); + step1[19] = SUB_EPI16(in12, in19); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x8(&step1[12], &step1[13], &step1[14], + &step1[15], &step1[16], &step1[17], + &step1[18], &step1[19]); + if (overflow) { + HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + } + } + // Stage 2 + { + step2[0] = ADD_EPI16(step1[0], step1[15]); + step2[1] = ADD_EPI16(step1[1], step1[14]); + step2[2] = ADD_EPI16(step1[2], step1[13]); + step2[3] = ADD_EPI16(step1[3], step1[12]); + step2[4] = ADD_EPI16(step1[4], step1[11]); + step2[5] = ADD_EPI16(step1[5], step1[10]); + step2[6] = ADD_EPI16(step1[6], step1[9]); + step2[7] = ADD_EPI16(step1[7], step1[8]); + step2[8] = SUB_EPI16(step1[7], step1[8]); + step2[9] = SUB_EPI16(step1[6], step1[9]); + step2[10] = SUB_EPI16(step1[5], step1[10]); + step2[11] = SUB_EPI16(step1[4], step1[11]); + step2[12] = SUB_EPI16(step1[3], step1[12]); + step2[13] = SUB_EPI16(step1[2], step1[13]); + step2[14] = SUB_EPI16(step1[1], step1[14]); + step2[15] = SUB_EPI16(step1[0], step1[15]); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x16( + &step2[0], &step2[1], &step2[2], &step2[3], &step2[4], &step2[5], + &step2[6], &step2[7], &step2[8], &step2[9], &step2[10], &step2[11], + &step2[12], &step2[13], &step2[14], &step2[15]); + if (overflow) { + if (pass == 0) + HIGH_FDCT32x32_2D_C(input, output_org, stride); + else + HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + } + { + const __m128i s2_20_0 = _mm_unpacklo_epi16(step1[27], step1[20]); + const __m128i s2_20_1 = _mm_unpackhi_epi16(step1[27], step1[20]); + const __m128i s2_21_0 = _mm_unpacklo_epi16(step1[26], step1[21]); + const __m128i s2_21_1 = _mm_unpackhi_epi16(step1[26], step1[21]); + const __m128i s2_22_0 = _mm_unpacklo_epi16(step1[25], step1[22]); + const __m128i s2_22_1 = _mm_unpackhi_epi16(step1[25], step1[22]); + const __m128i s2_23_0 = _mm_unpacklo_epi16(step1[24], step1[23]); + const __m128i s2_23_1 = _mm_unpackhi_epi16(step1[24], step1[23]); + const __m128i s2_20_2 = _mm_madd_epi16(s2_20_0, k__cospi_p16_m16); + const __m128i s2_20_3 = _mm_madd_epi16(s2_20_1, k__cospi_p16_m16); + const __m128i s2_21_2 = _mm_madd_epi16(s2_21_0, k__cospi_p16_m16); + const __m128i s2_21_3 = _mm_madd_epi16(s2_21_1, k__cospi_p16_m16); + const __m128i s2_22_2 = _mm_madd_epi16(s2_22_0, k__cospi_p16_m16); + const __m128i s2_22_3 = _mm_madd_epi16(s2_22_1, k__cospi_p16_m16); + const __m128i s2_23_2 = _mm_madd_epi16(s2_23_0, k__cospi_p16_m16); + const __m128i s2_23_3 = _mm_madd_epi16(s2_23_1, k__cospi_p16_m16); + const __m128i s2_24_2 = _mm_madd_epi16(s2_23_0, k__cospi_p16_p16); + const __m128i s2_24_3 = _mm_madd_epi16(s2_23_1, k__cospi_p16_p16); + const __m128i s2_25_2 = _mm_madd_epi16(s2_22_0, k__cospi_p16_p16); + const __m128i s2_25_3 = _mm_madd_epi16(s2_22_1, k__cospi_p16_p16); + const __m128i s2_26_2 = _mm_madd_epi16(s2_21_0, k__cospi_p16_p16); + const __m128i s2_26_3 = _mm_madd_epi16(s2_21_1, k__cospi_p16_p16); + const __m128i s2_27_2 = _mm_madd_epi16(s2_20_0, k__cospi_p16_p16); + const __m128i s2_27_3 = _mm_madd_epi16(s2_20_1, k__cospi_p16_p16); + // dct_const_round_shift + const __m128i s2_20_4 = _mm_add_epi32(s2_20_2, k__DCT_CONST_ROUNDING); + const __m128i s2_20_5 = _mm_add_epi32(s2_20_3, k__DCT_CONST_ROUNDING); + const __m128i s2_21_4 = _mm_add_epi32(s2_21_2, k__DCT_CONST_ROUNDING); + const __m128i s2_21_5 = _mm_add_epi32(s2_21_3, k__DCT_CONST_ROUNDING); + const __m128i s2_22_4 = _mm_add_epi32(s2_22_2, k__DCT_CONST_ROUNDING); + const __m128i s2_22_5 = _mm_add_epi32(s2_22_3, k__DCT_CONST_ROUNDING); + const __m128i s2_23_4 = _mm_add_epi32(s2_23_2, k__DCT_CONST_ROUNDING); + const __m128i s2_23_5 = _mm_add_epi32(s2_23_3, k__DCT_CONST_ROUNDING); + const __m128i s2_24_4 = _mm_add_epi32(s2_24_2, k__DCT_CONST_ROUNDING); + const __m128i s2_24_5 = _mm_add_epi32(s2_24_3, k__DCT_CONST_ROUNDING); + const __m128i s2_25_4 = _mm_add_epi32(s2_25_2, k__DCT_CONST_ROUNDING); + const __m128i s2_25_5 = _mm_add_epi32(s2_25_3, k__DCT_CONST_ROUNDING); + const __m128i s2_26_4 = _mm_add_epi32(s2_26_2, k__DCT_CONST_ROUNDING); + const __m128i s2_26_5 = _mm_add_epi32(s2_26_3, k__DCT_CONST_ROUNDING); + const __m128i s2_27_4 = _mm_add_epi32(s2_27_2, k__DCT_CONST_ROUNDING); + const __m128i s2_27_5 = _mm_add_epi32(s2_27_3, k__DCT_CONST_ROUNDING); + const __m128i s2_20_6 = _mm_srai_epi32(s2_20_4, DCT_CONST_BITS); + const __m128i s2_20_7 = _mm_srai_epi32(s2_20_5, DCT_CONST_BITS); + const __m128i s2_21_6 = _mm_srai_epi32(s2_21_4, DCT_CONST_BITS); + const __m128i s2_21_7 = _mm_srai_epi32(s2_21_5, DCT_CONST_BITS); + const __m128i s2_22_6 = _mm_srai_epi32(s2_22_4, DCT_CONST_BITS); + const __m128i s2_22_7 = _mm_srai_epi32(s2_22_5, DCT_CONST_BITS); + const __m128i s2_23_6 = _mm_srai_epi32(s2_23_4, DCT_CONST_BITS); + const __m128i s2_23_7 = _mm_srai_epi32(s2_23_5, DCT_CONST_BITS); + const __m128i s2_24_6 = _mm_srai_epi32(s2_24_4, DCT_CONST_BITS); + const __m128i s2_24_7 = _mm_srai_epi32(s2_24_5, DCT_CONST_BITS); + const __m128i s2_25_6 = _mm_srai_epi32(s2_25_4, DCT_CONST_BITS); + const __m128i s2_25_7 = _mm_srai_epi32(s2_25_5, DCT_CONST_BITS); + const __m128i s2_26_6 = _mm_srai_epi32(s2_26_4, DCT_CONST_BITS); + const __m128i s2_26_7 = _mm_srai_epi32(s2_26_5, DCT_CONST_BITS); + const __m128i s2_27_6 = _mm_srai_epi32(s2_27_4, DCT_CONST_BITS); + const __m128i s2_27_7 = _mm_srai_epi32(s2_27_5, DCT_CONST_BITS); + // Combine + step2[20] = _mm_packs_epi32(s2_20_6, s2_20_7); + step2[21] = _mm_packs_epi32(s2_21_6, s2_21_7); + step2[22] = _mm_packs_epi32(s2_22_6, s2_22_7); + step2[23] = _mm_packs_epi32(s2_23_6, s2_23_7); + step2[24] = _mm_packs_epi32(s2_24_6, s2_24_7); + step2[25] = _mm_packs_epi32(s2_25_6, s2_25_7); + step2[26] = _mm_packs_epi32(s2_26_6, s2_26_7); + step2[27] = _mm_packs_epi32(s2_27_6, s2_27_7); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x8(&step2[20], &step2[21], &step2[22], + &step2[23], &step2[24], &step2[25], + &step2[26], &step2[27]); + if (overflow) { + if (pass == 0) + HIGH_FDCT32x32_2D_C(input, output_org, stride); + else + HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + } + +#if !FDCT32x32_HIGH_PRECISION + // dump the magnitude by half, hence the intermediate values are within + // the range of 16 bits. + if (1 == pass) { + __m128i s3_00_0 = _mm_cmplt_epi16(step2[0], kZero); + __m128i s3_01_0 = _mm_cmplt_epi16(step2[1], kZero); + __m128i s3_02_0 = _mm_cmplt_epi16(step2[2], kZero); + __m128i s3_03_0 = _mm_cmplt_epi16(step2[3], kZero); + __m128i s3_04_0 = _mm_cmplt_epi16(step2[4], kZero); + __m128i s3_05_0 = _mm_cmplt_epi16(step2[5], kZero); + __m128i s3_06_0 = _mm_cmplt_epi16(step2[6], kZero); + __m128i s3_07_0 = _mm_cmplt_epi16(step2[7], kZero); + __m128i s2_08_0 = _mm_cmplt_epi16(step2[8], kZero); + __m128i s2_09_0 = _mm_cmplt_epi16(step2[9], kZero); + __m128i s3_10_0 = _mm_cmplt_epi16(step2[10], kZero); + __m128i s3_11_0 = _mm_cmplt_epi16(step2[11], kZero); + __m128i s3_12_0 = _mm_cmplt_epi16(step2[12], kZero); + __m128i s3_13_0 = _mm_cmplt_epi16(step2[13], kZero); + __m128i s2_14_0 = _mm_cmplt_epi16(step2[14], kZero); + __m128i s2_15_0 = _mm_cmplt_epi16(step2[15], kZero); + __m128i s3_16_0 = _mm_cmplt_epi16(step1[16], kZero); + __m128i s3_17_0 = _mm_cmplt_epi16(step1[17], kZero); + __m128i s3_18_0 = _mm_cmplt_epi16(step1[18], kZero); + __m128i s3_19_0 = _mm_cmplt_epi16(step1[19], kZero); + __m128i s3_20_0 = _mm_cmplt_epi16(step2[20], kZero); + __m128i s3_21_0 = _mm_cmplt_epi16(step2[21], kZero); + __m128i s3_22_0 = _mm_cmplt_epi16(step2[22], kZero); + __m128i s3_23_0 = _mm_cmplt_epi16(step2[23], kZero); + __m128i s3_24_0 = _mm_cmplt_epi16(step2[24], kZero); + __m128i s3_25_0 = _mm_cmplt_epi16(step2[25], kZero); + __m128i s3_26_0 = _mm_cmplt_epi16(step2[26], kZero); + __m128i s3_27_0 = _mm_cmplt_epi16(step2[27], kZero); + __m128i s3_28_0 = _mm_cmplt_epi16(step1[28], kZero); + __m128i s3_29_0 = _mm_cmplt_epi16(step1[29], kZero); + __m128i s3_30_0 = _mm_cmplt_epi16(step1[30], kZero); + __m128i s3_31_0 = _mm_cmplt_epi16(step1[31], kZero); + + step2[0] = SUB_EPI16(step2[0], s3_00_0); + step2[1] = SUB_EPI16(step2[1], s3_01_0); + step2[2] = SUB_EPI16(step2[2], s3_02_0); + step2[3] = SUB_EPI16(step2[3], s3_03_0); + step2[4] = SUB_EPI16(step2[4], s3_04_0); + step2[5] = SUB_EPI16(step2[5], s3_05_0); + step2[6] = SUB_EPI16(step2[6], s3_06_0); + step2[7] = SUB_EPI16(step2[7], s3_07_0); + step2[8] = SUB_EPI16(step2[8], s2_08_0); + step2[9] = SUB_EPI16(step2[9], s2_09_0); + step2[10] = SUB_EPI16(step2[10], s3_10_0); + step2[11] = SUB_EPI16(step2[11], s3_11_0); + step2[12] = SUB_EPI16(step2[12], s3_12_0); + step2[13] = SUB_EPI16(step2[13], s3_13_0); + step2[14] = SUB_EPI16(step2[14], s2_14_0); + step2[15] = SUB_EPI16(step2[15], s2_15_0); + step1[16] = SUB_EPI16(step1[16], s3_16_0); + step1[17] = SUB_EPI16(step1[17], s3_17_0); + step1[18] = SUB_EPI16(step1[18], s3_18_0); + step1[19] = SUB_EPI16(step1[19], s3_19_0); + step2[20] = SUB_EPI16(step2[20], s3_20_0); + step2[21] = SUB_EPI16(step2[21], s3_21_0); + step2[22] = SUB_EPI16(step2[22], s3_22_0); + step2[23] = SUB_EPI16(step2[23], s3_23_0); + step2[24] = SUB_EPI16(step2[24], s3_24_0); + step2[25] = SUB_EPI16(step2[25], s3_25_0); + step2[26] = SUB_EPI16(step2[26], s3_26_0); + step2[27] = SUB_EPI16(step2[27], s3_27_0); + step1[28] = SUB_EPI16(step1[28], s3_28_0); + step1[29] = SUB_EPI16(step1[29], s3_29_0); + step1[30] = SUB_EPI16(step1[30], s3_30_0); + step1[31] = SUB_EPI16(step1[31], s3_31_0); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x32( + &step2[0], &step2[1], &step2[2], &step2[3], &step2[4], &step2[5], + &step2[6], &step2[7], &step2[8], &step2[9], &step2[10], &step2[11], + &step2[12], &step2[13], &step2[14], &step2[15], &step1[16], + &step1[17], &step1[18], &step1[19], &step2[20], &step2[21], + &step2[22], &step2[23], &step2[24], &step2[25], &step2[26], + &step2[27], &step1[28], &step1[29], &step1[30], &step1[31]); + if (overflow) { + HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + step2[0] = _mm_add_epi16(step2[0], kOne); + step2[1] = _mm_add_epi16(step2[1], kOne); + step2[2] = _mm_add_epi16(step2[2], kOne); + step2[3] = _mm_add_epi16(step2[3], kOne); + step2[4] = _mm_add_epi16(step2[4], kOne); + step2[5] = _mm_add_epi16(step2[5], kOne); + step2[6] = _mm_add_epi16(step2[6], kOne); + step2[7] = _mm_add_epi16(step2[7], kOne); + step2[8] = _mm_add_epi16(step2[8], kOne); + step2[9] = _mm_add_epi16(step2[9], kOne); + step2[10] = _mm_add_epi16(step2[10], kOne); + step2[11] = _mm_add_epi16(step2[11], kOne); + step2[12] = _mm_add_epi16(step2[12], kOne); + step2[13] = _mm_add_epi16(step2[13], kOne); + step2[14] = _mm_add_epi16(step2[14], kOne); + step2[15] = _mm_add_epi16(step2[15], kOne); + step1[16] = _mm_add_epi16(step1[16], kOne); + step1[17] = _mm_add_epi16(step1[17], kOne); + step1[18] = _mm_add_epi16(step1[18], kOne); + step1[19] = _mm_add_epi16(step1[19], kOne); + step2[20] = _mm_add_epi16(step2[20], kOne); + step2[21] = _mm_add_epi16(step2[21], kOne); + step2[22] = _mm_add_epi16(step2[22], kOne); + step2[23] = _mm_add_epi16(step2[23], kOne); + step2[24] = _mm_add_epi16(step2[24], kOne); + step2[25] = _mm_add_epi16(step2[25], kOne); + step2[26] = _mm_add_epi16(step2[26], kOne); + step2[27] = _mm_add_epi16(step2[27], kOne); + step1[28] = _mm_add_epi16(step1[28], kOne); + step1[29] = _mm_add_epi16(step1[29], kOne); + step1[30] = _mm_add_epi16(step1[30], kOne); + step1[31] = _mm_add_epi16(step1[31], kOne); + + step2[0] = _mm_srai_epi16(step2[0], 2); + step2[1] = _mm_srai_epi16(step2[1], 2); + step2[2] = _mm_srai_epi16(step2[2], 2); + step2[3] = _mm_srai_epi16(step2[3], 2); + step2[4] = _mm_srai_epi16(step2[4], 2); + step2[5] = _mm_srai_epi16(step2[5], 2); + step2[6] = _mm_srai_epi16(step2[6], 2); + step2[7] = _mm_srai_epi16(step2[7], 2); + step2[8] = _mm_srai_epi16(step2[8], 2); + step2[9] = _mm_srai_epi16(step2[9], 2); + step2[10] = _mm_srai_epi16(step2[10], 2); + step2[11] = _mm_srai_epi16(step2[11], 2); + step2[12] = _mm_srai_epi16(step2[12], 2); + step2[13] = _mm_srai_epi16(step2[13], 2); + step2[14] = _mm_srai_epi16(step2[14], 2); + step2[15] = _mm_srai_epi16(step2[15], 2); + step1[16] = _mm_srai_epi16(step1[16], 2); + step1[17] = _mm_srai_epi16(step1[17], 2); + step1[18] = _mm_srai_epi16(step1[18], 2); + step1[19] = _mm_srai_epi16(step1[19], 2); + step2[20] = _mm_srai_epi16(step2[20], 2); + step2[21] = _mm_srai_epi16(step2[21], 2); + step2[22] = _mm_srai_epi16(step2[22], 2); + step2[23] = _mm_srai_epi16(step2[23], 2); + step2[24] = _mm_srai_epi16(step2[24], 2); + step2[25] = _mm_srai_epi16(step2[25], 2); + step2[26] = _mm_srai_epi16(step2[26], 2); + step2[27] = _mm_srai_epi16(step2[27], 2); + step1[28] = _mm_srai_epi16(step1[28], 2); + step1[29] = _mm_srai_epi16(step1[29], 2); + step1[30] = _mm_srai_epi16(step1[30], 2); + step1[31] = _mm_srai_epi16(step1[31], 2); + } +#endif // !FDCT32x32_HIGH_PRECISION + +#if FDCT32x32_HIGH_PRECISION + if (pass == 0) { +#endif + // Stage 3 + { + step3[0] = ADD_EPI16(step2[(8 - 1)], step2[0]); + step3[1] = ADD_EPI16(step2[(8 - 2)], step2[1]); + step3[2] = ADD_EPI16(step2[(8 - 3)], step2[2]); + step3[3] = ADD_EPI16(step2[(8 - 4)], step2[3]); + step3[4] = SUB_EPI16(step2[(8 - 5)], step2[4]); + step3[5] = SUB_EPI16(step2[(8 - 6)], step2[5]); + step3[6] = SUB_EPI16(step2[(8 - 7)], step2[6]); + step3[7] = SUB_EPI16(step2[(8 - 8)], step2[7]); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x8(&step3[0], &step3[1], &step3[2], + &step3[3], &step3[4], &step3[5], + &step3[6], &step3[7]); + if (overflow) { + if (pass == 0) + HIGH_FDCT32x32_2D_C(input, output_org, stride); + else + HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + } + { + const __m128i s3_10_0 = _mm_unpacklo_epi16(step2[13], step2[10]); + const __m128i s3_10_1 = _mm_unpackhi_epi16(step2[13], step2[10]); + const __m128i s3_11_0 = _mm_unpacklo_epi16(step2[12], step2[11]); + const __m128i s3_11_1 = _mm_unpackhi_epi16(step2[12], step2[11]); + const __m128i s3_10_2 = _mm_madd_epi16(s3_10_0, k__cospi_p16_m16); + const __m128i s3_10_3 = _mm_madd_epi16(s3_10_1, k__cospi_p16_m16); + const __m128i s3_11_2 = _mm_madd_epi16(s3_11_0, k__cospi_p16_m16); + const __m128i s3_11_3 = _mm_madd_epi16(s3_11_1, k__cospi_p16_m16); + const __m128i s3_12_2 = _mm_madd_epi16(s3_11_0, k__cospi_p16_p16); + const __m128i s3_12_3 = _mm_madd_epi16(s3_11_1, k__cospi_p16_p16); + const __m128i s3_13_2 = _mm_madd_epi16(s3_10_0, k__cospi_p16_p16); + const __m128i s3_13_3 = _mm_madd_epi16(s3_10_1, k__cospi_p16_p16); + // dct_const_round_shift + const __m128i s3_10_4 = _mm_add_epi32(s3_10_2, k__DCT_CONST_ROUNDING); + const __m128i s3_10_5 = _mm_add_epi32(s3_10_3, k__DCT_CONST_ROUNDING); + const __m128i s3_11_4 = _mm_add_epi32(s3_11_2, k__DCT_CONST_ROUNDING); + const __m128i s3_11_5 = _mm_add_epi32(s3_11_3, k__DCT_CONST_ROUNDING); + const __m128i s3_12_4 = _mm_add_epi32(s3_12_2, k__DCT_CONST_ROUNDING); + const __m128i s3_12_5 = _mm_add_epi32(s3_12_3, k__DCT_CONST_ROUNDING); + const __m128i s3_13_4 = _mm_add_epi32(s3_13_2, k__DCT_CONST_ROUNDING); + const __m128i s3_13_5 = _mm_add_epi32(s3_13_3, k__DCT_CONST_ROUNDING); + const __m128i s3_10_6 = _mm_srai_epi32(s3_10_4, DCT_CONST_BITS); + const __m128i s3_10_7 = _mm_srai_epi32(s3_10_5, DCT_CONST_BITS); + const __m128i s3_11_6 = _mm_srai_epi32(s3_11_4, DCT_CONST_BITS); + const __m128i s3_11_7 = _mm_srai_epi32(s3_11_5, DCT_CONST_BITS); + const __m128i s3_12_6 = _mm_srai_epi32(s3_12_4, DCT_CONST_BITS); + const __m128i s3_12_7 = _mm_srai_epi32(s3_12_5, DCT_CONST_BITS); + const __m128i s3_13_6 = _mm_srai_epi32(s3_13_4, DCT_CONST_BITS); + const __m128i s3_13_7 = _mm_srai_epi32(s3_13_5, DCT_CONST_BITS); + // Combine + step3[10] = _mm_packs_epi32(s3_10_6, s3_10_7); + step3[11] = _mm_packs_epi32(s3_11_6, s3_11_7); + step3[12] = _mm_packs_epi32(s3_12_6, s3_12_7); + step3[13] = _mm_packs_epi32(s3_13_6, s3_13_7); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x4(&step3[10], &step3[11], &step3[12], + &step3[13]); + if (overflow) { + if (pass == 0) + HIGH_FDCT32x32_2D_C(input, output_org, stride); + else + HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + } + { + step3[16] = ADD_EPI16(step2[23], step1[16]); + step3[17] = ADD_EPI16(step2[22], step1[17]); + step3[18] = ADD_EPI16(step2[21], step1[18]); + step3[19] = ADD_EPI16(step2[20], step1[19]); + step3[20] = SUB_EPI16(step1[19], step2[20]); + step3[21] = SUB_EPI16(step1[18], step2[21]); + step3[22] = SUB_EPI16(step1[17], step2[22]); + step3[23] = SUB_EPI16(step1[16], step2[23]); + step3[24] = SUB_EPI16(step1[31], step2[24]); + step3[25] = SUB_EPI16(step1[30], step2[25]); + step3[26] = SUB_EPI16(step1[29], step2[26]); + step3[27] = SUB_EPI16(step1[28], step2[27]); + step3[28] = ADD_EPI16(step2[27], step1[28]); + step3[29] = ADD_EPI16(step2[26], step1[29]); + step3[30] = ADD_EPI16(step2[25], step1[30]); + step3[31] = ADD_EPI16(step2[24], step1[31]); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x16( + &step3[16], &step3[17], &step3[18], &step3[19], &step3[20], + &step3[21], &step3[22], &step3[23], &step3[24], &step3[25], + &step3[26], &step3[27], &step3[28], &step3[29], &step3[30], + &step3[31]); + if (overflow) { + if (pass == 0) + HIGH_FDCT32x32_2D_C(input, output_org, stride); + else + HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + } + + // Stage 4 + { + step1[0] = ADD_EPI16(step3[3], step3[0]); + step1[1] = ADD_EPI16(step3[2], step3[1]); + step1[2] = SUB_EPI16(step3[1], step3[2]); + step1[3] = SUB_EPI16(step3[0], step3[3]); + step1[8] = ADD_EPI16(step3[11], step2[8]); + step1[9] = ADD_EPI16(step3[10], step2[9]); + step1[10] = SUB_EPI16(step2[9], step3[10]); + step1[11] = SUB_EPI16(step2[8], step3[11]); + step1[12] = SUB_EPI16(step2[15], step3[12]); + step1[13] = SUB_EPI16(step2[14], step3[13]); + step1[14] = ADD_EPI16(step3[13], step2[14]); + step1[15] = ADD_EPI16(step3[12], step2[15]); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x16( + &step1[0], &step1[1], &step1[2], &step1[3], &step1[4], &step1[5], + &step1[6], &step1[7], &step1[8], &step1[9], &step1[10], + &step1[11], &step1[12], &step1[13], &step1[14], &step1[15]); + if (overflow) { + if (pass == 0) + HIGH_FDCT32x32_2D_C(input, output_org, stride); + else + HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + } + { + const __m128i s1_05_0 = _mm_unpacklo_epi16(step3[6], step3[5]); + const __m128i s1_05_1 = _mm_unpackhi_epi16(step3[6], step3[5]); + const __m128i s1_05_2 = _mm_madd_epi16(s1_05_0, k__cospi_p16_m16); + const __m128i s1_05_3 = _mm_madd_epi16(s1_05_1, k__cospi_p16_m16); + const __m128i s1_06_2 = _mm_madd_epi16(s1_05_0, k__cospi_p16_p16); + const __m128i s1_06_3 = _mm_madd_epi16(s1_05_1, k__cospi_p16_p16); + // dct_const_round_shift + const __m128i s1_05_4 = _mm_add_epi32(s1_05_2, k__DCT_CONST_ROUNDING); + const __m128i s1_05_5 = _mm_add_epi32(s1_05_3, k__DCT_CONST_ROUNDING); + const __m128i s1_06_4 = _mm_add_epi32(s1_06_2, k__DCT_CONST_ROUNDING); + const __m128i s1_06_5 = _mm_add_epi32(s1_06_3, k__DCT_CONST_ROUNDING); + const __m128i s1_05_6 = _mm_srai_epi32(s1_05_4, DCT_CONST_BITS); + const __m128i s1_05_7 = _mm_srai_epi32(s1_05_5, DCT_CONST_BITS); + const __m128i s1_06_6 = _mm_srai_epi32(s1_06_4, DCT_CONST_BITS); + const __m128i s1_06_7 = _mm_srai_epi32(s1_06_5, DCT_CONST_BITS); + // Combine + step1[5] = _mm_packs_epi32(s1_05_6, s1_05_7); + step1[6] = _mm_packs_epi32(s1_06_6, s1_06_7); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x2(&step1[5], &step1[6]); + if (overflow) { + if (pass == 0) + HIGH_FDCT32x32_2D_C(input, output_org, stride); + else + HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + } + { + const __m128i s1_18_0 = _mm_unpacklo_epi16(step3[18], step3[29]); + const __m128i s1_18_1 = _mm_unpackhi_epi16(step3[18], step3[29]); + const __m128i s1_19_0 = _mm_unpacklo_epi16(step3[19], step3[28]); + const __m128i s1_19_1 = _mm_unpackhi_epi16(step3[19], step3[28]); + const __m128i s1_20_0 = _mm_unpacklo_epi16(step3[20], step3[27]); + const __m128i s1_20_1 = _mm_unpackhi_epi16(step3[20], step3[27]); + const __m128i s1_21_0 = _mm_unpacklo_epi16(step3[21], step3[26]); + const __m128i s1_21_1 = _mm_unpackhi_epi16(step3[21], step3[26]); + const __m128i s1_18_2 = _mm_madd_epi16(s1_18_0, k__cospi_m08_p24); + const __m128i s1_18_3 = _mm_madd_epi16(s1_18_1, k__cospi_m08_p24); + const __m128i s1_19_2 = _mm_madd_epi16(s1_19_0, k__cospi_m08_p24); + const __m128i s1_19_3 = _mm_madd_epi16(s1_19_1, k__cospi_m08_p24); + const __m128i s1_20_2 = _mm_madd_epi16(s1_20_0, k__cospi_m24_m08); + const __m128i s1_20_3 = _mm_madd_epi16(s1_20_1, k__cospi_m24_m08); + const __m128i s1_21_2 = _mm_madd_epi16(s1_21_0, k__cospi_m24_m08); + const __m128i s1_21_3 = _mm_madd_epi16(s1_21_1, k__cospi_m24_m08); + const __m128i s1_26_2 = _mm_madd_epi16(s1_21_0, k__cospi_m08_p24); + const __m128i s1_26_3 = _mm_madd_epi16(s1_21_1, k__cospi_m08_p24); + const __m128i s1_27_2 = _mm_madd_epi16(s1_20_0, k__cospi_m08_p24); + const __m128i s1_27_3 = _mm_madd_epi16(s1_20_1, k__cospi_m08_p24); + const __m128i s1_28_2 = _mm_madd_epi16(s1_19_0, k__cospi_p24_p08); + const __m128i s1_28_3 = _mm_madd_epi16(s1_19_1, k__cospi_p24_p08); + const __m128i s1_29_2 = _mm_madd_epi16(s1_18_0, k__cospi_p24_p08); + const __m128i s1_29_3 = _mm_madd_epi16(s1_18_1, k__cospi_p24_p08); + // dct_const_round_shift + const __m128i s1_18_4 = _mm_add_epi32(s1_18_2, k__DCT_CONST_ROUNDING); + const __m128i s1_18_5 = _mm_add_epi32(s1_18_3, k__DCT_CONST_ROUNDING); + const __m128i s1_19_4 = _mm_add_epi32(s1_19_2, k__DCT_CONST_ROUNDING); + const __m128i s1_19_5 = _mm_add_epi32(s1_19_3, k__DCT_CONST_ROUNDING); + const __m128i s1_20_4 = _mm_add_epi32(s1_20_2, k__DCT_CONST_ROUNDING); + const __m128i s1_20_5 = _mm_add_epi32(s1_20_3, k__DCT_CONST_ROUNDING); + const __m128i s1_21_4 = _mm_add_epi32(s1_21_2, k__DCT_CONST_ROUNDING); + const __m128i s1_21_5 = _mm_add_epi32(s1_21_3, k__DCT_CONST_ROUNDING); + const __m128i s1_26_4 = _mm_add_epi32(s1_26_2, k__DCT_CONST_ROUNDING); + const __m128i s1_26_5 = _mm_add_epi32(s1_26_3, k__DCT_CONST_ROUNDING); + const __m128i s1_27_4 = _mm_add_epi32(s1_27_2, k__DCT_CONST_ROUNDING); + const __m128i s1_27_5 = _mm_add_epi32(s1_27_3, k__DCT_CONST_ROUNDING); + const __m128i s1_28_4 = _mm_add_epi32(s1_28_2, k__DCT_CONST_ROUNDING); + const __m128i s1_28_5 = _mm_add_epi32(s1_28_3, k__DCT_CONST_ROUNDING); + const __m128i s1_29_4 = _mm_add_epi32(s1_29_2, k__DCT_CONST_ROUNDING); + const __m128i s1_29_5 = _mm_add_epi32(s1_29_3, k__DCT_CONST_ROUNDING); + const __m128i s1_18_6 = _mm_srai_epi32(s1_18_4, DCT_CONST_BITS); + const __m128i s1_18_7 = _mm_srai_epi32(s1_18_5, DCT_CONST_BITS); + const __m128i s1_19_6 = _mm_srai_epi32(s1_19_4, DCT_CONST_BITS); + const __m128i s1_19_7 = _mm_srai_epi32(s1_19_5, DCT_CONST_BITS); + const __m128i s1_20_6 = _mm_srai_epi32(s1_20_4, DCT_CONST_BITS); + const __m128i s1_20_7 = _mm_srai_epi32(s1_20_5, DCT_CONST_BITS); + const __m128i s1_21_6 = _mm_srai_epi32(s1_21_4, DCT_CONST_BITS); + const __m128i s1_21_7 = _mm_srai_epi32(s1_21_5, DCT_CONST_BITS); + const __m128i s1_26_6 = _mm_srai_epi32(s1_26_4, DCT_CONST_BITS); + const __m128i s1_26_7 = _mm_srai_epi32(s1_26_5, DCT_CONST_BITS); + const __m128i s1_27_6 = _mm_srai_epi32(s1_27_4, DCT_CONST_BITS); + const __m128i s1_27_7 = _mm_srai_epi32(s1_27_5, DCT_CONST_BITS); + const __m128i s1_28_6 = _mm_srai_epi32(s1_28_4, DCT_CONST_BITS); + const __m128i s1_28_7 = _mm_srai_epi32(s1_28_5, DCT_CONST_BITS); + const __m128i s1_29_6 = _mm_srai_epi32(s1_29_4, DCT_CONST_BITS); + const __m128i s1_29_7 = _mm_srai_epi32(s1_29_5, DCT_CONST_BITS); + // Combine + step1[18] = _mm_packs_epi32(s1_18_6, s1_18_7); + step1[19] = _mm_packs_epi32(s1_19_6, s1_19_7); + step1[20] = _mm_packs_epi32(s1_20_6, s1_20_7); + step1[21] = _mm_packs_epi32(s1_21_6, s1_21_7); + step1[26] = _mm_packs_epi32(s1_26_6, s1_26_7); + step1[27] = _mm_packs_epi32(s1_27_6, s1_27_7); + step1[28] = _mm_packs_epi32(s1_28_6, s1_28_7); + step1[29] = _mm_packs_epi32(s1_29_6, s1_29_7); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x8(&step1[18], &step1[19], &step1[20], + &step1[21], &step1[26], &step1[27], + &step1[28], &step1[29]); + if (overflow) { + if (pass == 0) + HIGH_FDCT32x32_2D_C(input, output_org, stride); + else + HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + } + // Stage 5 + { + step2[4] = ADD_EPI16(step1[5], step3[4]); + step2[5] = SUB_EPI16(step3[4], step1[5]); + step2[6] = SUB_EPI16(step3[7], step1[6]); + step2[7] = ADD_EPI16(step1[6], step3[7]); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x4(&step2[4], &step2[5], &step2[6], + &step2[7]); + if (overflow) { + if (pass == 0) + HIGH_FDCT32x32_2D_C(input, output_org, stride); + else + HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + } + { + const __m128i out_00_0 = _mm_unpacklo_epi16(step1[0], step1[1]); + const __m128i out_00_1 = _mm_unpackhi_epi16(step1[0], step1[1]); + const __m128i out_08_0 = _mm_unpacklo_epi16(step1[2], step1[3]); + const __m128i out_08_1 = _mm_unpackhi_epi16(step1[2], step1[3]); + const __m128i out_00_2 = _mm_madd_epi16(out_00_0, k__cospi_p16_p16); + const __m128i out_00_3 = _mm_madd_epi16(out_00_1, k__cospi_p16_p16); + const __m128i out_16_2 = _mm_madd_epi16(out_00_0, k__cospi_p16_m16); + const __m128i out_16_3 = _mm_madd_epi16(out_00_1, k__cospi_p16_m16); + const __m128i out_08_2 = _mm_madd_epi16(out_08_0, k__cospi_p24_p08); + const __m128i out_08_3 = _mm_madd_epi16(out_08_1, k__cospi_p24_p08); + const __m128i out_24_2 = _mm_madd_epi16(out_08_0, k__cospi_m08_p24); + const __m128i out_24_3 = _mm_madd_epi16(out_08_1, k__cospi_m08_p24); + // dct_const_round_shift + const __m128i out_00_4 = + _mm_add_epi32(out_00_2, k__DCT_CONST_ROUNDING); + const __m128i out_00_5 = + _mm_add_epi32(out_00_3, k__DCT_CONST_ROUNDING); + const __m128i out_16_4 = + _mm_add_epi32(out_16_2, k__DCT_CONST_ROUNDING); + const __m128i out_16_5 = + _mm_add_epi32(out_16_3, k__DCT_CONST_ROUNDING); + const __m128i out_08_4 = + _mm_add_epi32(out_08_2, k__DCT_CONST_ROUNDING); + const __m128i out_08_5 = + _mm_add_epi32(out_08_3, k__DCT_CONST_ROUNDING); + const __m128i out_24_4 = + _mm_add_epi32(out_24_2, k__DCT_CONST_ROUNDING); + const __m128i out_24_5 = + _mm_add_epi32(out_24_3, k__DCT_CONST_ROUNDING); + const __m128i out_00_6 = _mm_srai_epi32(out_00_4, DCT_CONST_BITS); + const __m128i out_00_7 = _mm_srai_epi32(out_00_5, DCT_CONST_BITS); + const __m128i out_16_6 = _mm_srai_epi32(out_16_4, DCT_CONST_BITS); + const __m128i out_16_7 = _mm_srai_epi32(out_16_5, DCT_CONST_BITS); + const __m128i out_08_6 = _mm_srai_epi32(out_08_4, DCT_CONST_BITS); + const __m128i out_08_7 = _mm_srai_epi32(out_08_5, DCT_CONST_BITS); + const __m128i out_24_6 = _mm_srai_epi32(out_24_4, DCT_CONST_BITS); + const __m128i out_24_7 = _mm_srai_epi32(out_24_5, DCT_CONST_BITS); + // Combine + out[0] = _mm_packs_epi32(out_00_6, out_00_7); + out[16] = _mm_packs_epi32(out_16_6, out_16_7); + out[8] = _mm_packs_epi32(out_08_6, out_08_7); + out[24] = _mm_packs_epi32(out_24_6, out_24_7); +#if DCT_HIGH_BIT_DEPTH + overflow = + check_epi16_overflow_x4(&out[0], &out[16], &out[8], &out[24]); + if (overflow) { + if (pass == 0) + HIGH_FDCT32x32_2D_C(input, output_org, stride); + else + HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + } + { + const __m128i s2_09_0 = _mm_unpacklo_epi16(step1[9], step1[14]); + const __m128i s2_09_1 = _mm_unpackhi_epi16(step1[9], step1[14]); + const __m128i s2_10_0 = _mm_unpacklo_epi16(step1[10], step1[13]); + const __m128i s2_10_1 = _mm_unpackhi_epi16(step1[10], step1[13]); + const __m128i s2_09_2 = _mm_madd_epi16(s2_09_0, k__cospi_m08_p24); + const __m128i s2_09_3 = _mm_madd_epi16(s2_09_1, k__cospi_m08_p24); + const __m128i s2_10_2 = _mm_madd_epi16(s2_10_0, k__cospi_m24_m08); + const __m128i s2_10_3 = _mm_madd_epi16(s2_10_1, k__cospi_m24_m08); + const __m128i s2_13_2 = _mm_madd_epi16(s2_10_0, k__cospi_m08_p24); + const __m128i s2_13_3 = _mm_madd_epi16(s2_10_1, k__cospi_m08_p24); + const __m128i s2_14_2 = _mm_madd_epi16(s2_09_0, k__cospi_p24_p08); + const __m128i s2_14_3 = _mm_madd_epi16(s2_09_1, k__cospi_p24_p08); + // dct_const_round_shift + const __m128i s2_09_4 = _mm_add_epi32(s2_09_2, k__DCT_CONST_ROUNDING); + const __m128i s2_09_5 = _mm_add_epi32(s2_09_3, k__DCT_CONST_ROUNDING); + const __m128i s2_10_4 = _mm_add_epi32(s2_10_2, k__DCT_CONST_ROUNDING); + const __m128i s2_10_5 = _mm_add_epi32(s2_10_3, k__DCT_CONST_ROUNDING); + const __m128i s2_13_4 = _mm_add_epi32(s2_13_2, k__DCT_CONST_ROUNDING); + const __m128i s2_13_5 = _mm_add_epi32(s2_13_3, k__DCT_CONST_ROUNDING); + const __m128i s2_14_4 = _mm_add_epi32(s2_14_2, k__DCT_CONST_ROUNDING); + const __m128i s2_14_5 = _mm_add_epi32(s2_14_3, k__DCT_CONST_ROUNDING); + const __m128i s2_09_6 = _mm_srai_epi32(s2_09_4, DCT_CONST_BITS); + const __m128i s2_09_7 = _mm_srai_epi32(s2_09_5, DCT_CONST_BITS); + const __m128i s2_10_6 = _mm_srai_epi32(s2_10_4, DCT_CONST_BITS); + const __m128i s2_10_7 = _mm_srai_epi32(s2_10_5, DCT_CONST_BITS); + const __m128i s2_13_6 = _mm_srai_epi32(s2_13_4, DCT_CONST_BITS); + const __m128i s2_13_7 = _mm_srai_epi32(s2_13_5, DCT_CONST_BITS); + const __m128i s2_14_6 = _mm_srai_epi32(s2_14_4, DCT_CONST_BITS); + const __m128i s2_14_7 = _mm_srai_epi32(s2_14_5, DCT_CONST_BITS); + // Combine + step2[9] = _mm_packs_epi32(s2_09_6, s2_09_7); + step2[10] = _mm_packs_epi32(s2_10_6, s2_10_7); + step2[13] = _mm_packs_epi32(s2_13_6, s2_13_7); + step2[14] = _mm_packs_epi32(s2_14_6, s2_14_7); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x4(&step2[9], &step2[10], &step2[13], + &step2[14]); + if (overflow) { + if (pass == 0) + HIGH_FDCT32x32_2D_C(input, output_org, stride); + else + HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + } + { + step2[16] = ADD_EPI16(step1[19], step3[16]); + step2[17] = ADD_EPI16(step1[18], step3[17]); + step2[18] = SUB_EPI16(step3[17], step1[18]); + step2[19] = SUB_EPI16(step3[16], step1[19]); + step2[20] = SUB_EPI16(step3[23], step1[20]); + step2[21] = SUB_EPI16(step3[22], step1[21]); + step2[22] = ADD_EPI16(step1[21], step3[22]); + step2[23] = ADD_EPI16(step1[20], step3[23]); + step2[24] = ADD_EPI16(step1[27], step3[24]); + step2[25] = ADD_EPI16(step1[26], step3[25]); + step2[26] = SUB_EPI16(step3[25], step1[26]); + step2[27] = SUB_EPI16(step3[24], step1[27]); + step2[28] = SUB_EPI16(step3[31], step1[28]); + step2[29] = SUB_EPI16(step3[30], step1[29]); + step2[30] = ADD_EPI16(step1[29], step3[30]); + step2[31] = ADD_EPI16(step1[28], step3[31]); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x16( + &step2[16], &step2[17], &step2[18], &step2[19], &step2[20], + &step2[21], &step2[22], &step2[23], &step2[24], &step2[25], + &step2[26], &step2[27], &step2[28], &step2[29], &step2[30], + &step2[31]); + if (overflow) { + if (pass == 0) + HIGH_FDCT32x32_2D_C(input, output_org, stride); + else + HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + } + // Stage 6 + { + const __m128i out_04_0 = _mm_unpacklo_epi16(step2[4], step2[7]); + const __m128i out_04_1 = _mm_unpackhi_epi16(step2[4], step2[7]); + const __m128i out_20_0 = _mm_unpacklo_epi16(step2[5], step2[6]); + const __m128i out_20_1 = _mm_unpackhi_epi16(step2[5], step2[6]); + const __m128i out_12_0 = _mm_unpacklo_epi16(step2[5], step2[6]); + const __m128i out_12_1 = _mm_unpackhi_epi16(step2[5], step2[6]); + const __m128i out_28_0 = _mm_unpacklo_epi16(step2[4], step2[7]); + const __m128i out_28_1 = _mm_unpackhi_epi16(step2[4], step2[7]); + const __m128i out_04_2 = _mm_madd_epi16(out_04_0, k__cospi_p28_p04); + const __m128i out_04_3 = _mm_madd_epi16(out_04_1, k__cospi_p28_p04); + const __m128i out_20_2 = _mm_madd_epi16(out_20_0, k__cospi_p12_p20); + const __m128i out_20_3 = _mm_madd_epi16(out_20_1, k__cospi_p12_p20); + const __m128i out_12_2 = _mm_madd_epi16(out_12_0, k__cospi_m20_p12); + const __m128i out_12_3 = _mm_madd_epi16(out_12_1, k__cospi_m20_p12); + const __m128i out_28_2 = _mm_madd_epi16(out_28_0, k__cospi_m04_p28); + const __m128i out_28_3 = _mm_madd_epi16(out_28_1, k__cospi_m04_p28); + // dct_const_round_shift + const __m128i out_04_4 = + _mm_add_epi32(out_04_2, k__DCT_CONST_ROUNDING); + const __m128i out_04_5 = + _mm_add_epi32(out_04_3, k__DCT_CONST_ROUNDING); + const __m128i out_20_4 = + _mm_add_epi32(out_20_2, k__DCT_CONST_ROUNDING); + const __m128i out_20_5 = + _mm_add_epi32(out_20_3, k__DCT_CONST_ROUNDING); + const __m128i out_12_4 = + _mm_add_epi32(out_12_2, k__DCT_CONST_ROUNDING); + const __m128i out_12_5 = + _mm_add_epi32(out_12_3, k__DCT_CONST_ROUNDING); + const __m128i out_28_4 = + _mm_add_epi32(out_28_2, k__DCT_CONST_ROUNDING); + const __m128i out_28_5 = + _mm_add_epi32(out_28_3, k__DCT_CONST_ROUNDING); + const __m128i out_04_6 = _mm_srai_epi32(out_04_4, DCT_CONST_BITS); + const __m128i out_04_7 = _mm_srai_epi32(out_04_5, DCT_CONST_BITS); + const __m128i out_20_6 = _mm_srai_epi32(out_20_4, DCT_CONST_BITS); + const __m128i out_20_7 = _mm_srai_epi32(out_20_5, DCT_CONST_BITS); + const __m128i out_12_6 = _mm_srai_epi32(out_12_4, DCT_CONST_BITS); + const __m128i out_12_7 = _mm_srai_epi32(out_12_5, DCT_CONST_BITS); + const __m128i out_28_6 = _mm_srai_epi32(out_28_4, DCT_CONST_BITS); + const __m128i out_28_7 = _mm_srai_epi32(out_28_5, DCT_CONST_BITS); + // Combine + out[4] = _mm_packs_epi32(out_04_6, out_04_7); + out[20] = _mm_packs_epi32(out_20_6, out_20_7); + out[12] = _mm_packs_epi32(out_12_6, out_12_7); + out[28] = _mm_packs_epi32(out_28_6, out_28_7); +#if DCT_HIGH_BIT_DEPTH + overflow = + check_epi16_overflow_x4(&out[4], &out[20], &out[12], &out[28]); + if (overflow) { + if (pass == 0) + HIGH_FDCT32x32_2D_C(input, output_org, stride); + else + HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + } + { + step3[8] = ADD_EPI16(step2[9], step1[8]); + step3[9] = SUB_EPI16(step1[8], step2[9]); + step3[10] = SUB_EPI16(step1[11], step2[10]); + step3[11] = ADD_EPI16(step2[10], step1[11]); + step3[12] = ADD_EPI16(step2[13], step1[12]); + step3[13] = SUB_EPI16(step1[12], step2[13]); + step3[14] = SUB_EPI16(step1[15], step2[14]); + step3[15] = ADD_EPI16(step2[14], step1[15]); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x8(&step3[8], &step3[9], &step3[10], + &step3[11], &step3[12], &step3[13], + &step3[14], &step3[15]); + if (overflow) { + if (pass == 0) + HIGH_FDCT32x32_2D_C(input, output_org, stride); + else + HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + } + { + const __m128i s3_17_0 = _mm_unpacklo_epi16(step2[17], step2[30]); + const __m128i s3_17_1 = _mm_unpackhi_epi16(step2[17], step2[30]); + const __m128i s3_18_0 = _mm_unpacklo_epi16(step2[18], step2[29]); + const __m128i s3_18_1 = _mm_unpackhi_epi16(step2[18], step2[29]); + const __m128i s3_21_0 = _mm_unpacklo_epi16(step2[21], step2[26]); + const __m128i s3_21_1 = _mm_unpackhi_epi16(step2[21], step2[26]); + const __m128i s3_22_0 = _mm_unpacklo_epi16(step2[22], step2[25]); + const __m128i s3_22_1 = _mm_unpackhi_epi16(step2[22], step2[25]); + const __m128i s3_17_2 = _mm_madd_epi16(s3_17_0, k__cospi_m04_p28); + const __m128i s3_17_3 = _mm_madd_epi16(s3_17_1, k__cospi_m04_p28); + const __m128i s3_18_2 = _mm_madd_epi16(s3_18_0, k__cospi_m28_m04); + const __m128i s3_18_3 = _mm_madd_epi16(s3_18_1, k__cospi_m28_m04); + const __m128i s3_21_2 = _mm_madd_epi16(s3_21_0, k__cospi_m20_p12); + const __m128i s3_21_3 = _mm_madd_epi16(s3_21_1, k__cospi_m20_p12); + const __m128i s3_22_2 = _mm_madd_epi16(s3_22_0, k__cospi_m12_m20); + const __m128i s3_22_3 = _mm_madd_epi16(s3_22_1, k__cospi_m12_m20); + const __m128i s3_25_2 = _mm_madd_epi16(s3_22_0, k__cospi_m20_p12); + const __m128i s3_25_3 = _mm_madd_epi16(s3_22_1, k__cospi_m20_p12); + const __m128i s3_26_2 = _mm_madd_epi16(s3_21_0, k__cospi_p12_p20); + const __m128i s3_26_3 = _mm_madd_epi16(s3_21_1, k__cospi_p12_p20); + const __m128i s3_29_2 = _mm_madd_epi16(s3_18_0, k__cospi_m04_p28); + const __m128i s3_29_3 = _mm_madd_epi16(s3_18_1, k__cospi_m04_p28); + const __m128i s3_30_2 = _mm_madd_epi16(s3_17_0, k__cospi_p28_p04); + const __m128i s3_30_3 = _mm_madd_epi16(s3_17_1, k__cospi_p28_p04); + // dct_const_round_shift + const __m128i s3_17_4 = _mm_add_epi32(s3_17_2, k__DCT_CONST_ROUNDING); + const __m128i s3_17_5 = _mm_add_epi32(s3_17_3, k__DCT_CONST_ROUNDING); + const __m128i s3_18_4 = _mm_add_epi32(s3_18_2, k__DCT_CONST_ROUNDING); + const __m128i s3_18_5 = _mm_add_epi32(s3_18_3, k__DCT_CONST_ROUNDING); + const __m128i s3_21_4 = _mm_add_epi32(s3_21_2, k__DCT_CONST_ROUNDING); + const __m128i s3_21_5 = _mm_add_epi32(s3_21_3, k__DCT_CONST_ROUNDING); + const __m128i s3_22_4 = _mm_add_epi32(s3_22_2, k__DCT_CONST_ROUNDING); + const __m128i s3_22_5 = _mm_add_epi32(s3_22_3, k__DCT_CONST_ROUNDING); + const __m128i s3_17_6 = _mm_srai_epi32(s3_17_4, DCT_CONST_BITS); + const __m128i s3_17_7 = _mm_srai_epi32(s3_17_5, DCT_CONST_BITS); + const __m128i s3_18_6 = _mm_srai_epi32(s3_18_4, DCT_CONST_BITS); + const __m128i s3_18_7 = _mm_srai_epi32(s3_18_5, DCT_CONST_BITS); + const __m128i s3_21_6 = _mm_srai_epi32(s3_21_4, DCT_CONST_BITS); + const __m128i s3_21_7 = _mm_srai_epi32(s3_21_5, DCT_CONST_BITS); + const __m128i s3_22_6 = _mm_srai_epi32(s3_22_4, DCT_CONST_BITS); + const __m128i s3_22_7 = _mm_srai_epi32(s3_22_5, DCT_CONST_BITS); + const __m128i s3_25_4 = _mm_add_epi32(s3_25_2, k__DCT_CONST_ROUNDING); + const __m128i s3_25_5 = _mm_add_epi32(s3_25_3, k__DCT_CONST_ROUNDING); + const __m128i s3_26_4 = _mm_add_epi32(s3_26_2, k__DCT_CONST_ROUNDING); + const __m128i s3_26_5 = _mm_add_epi32(s3_26_3, k__DCT_CONST_ROUNDING); + const __m128i s3_29_4 = _mm_add_epi32(s3_29_2, k__DCT_CONST_ROUNDING); + const __m128i s3_29_5 = _mm_add_epi32(s3_29_3, k__DCT_CONST_ROUNDING); + const __m128i s3_30_4 = _mm_add_epi32(s3_30_2, k__DCT_CONST_ROUNDING); + const __m128i s3_30_5 = _mm_add_epi32(s3_30_3, k__DCT_CONST_ROUNDING); + const __m128i s3_25_6 = _mm_srai_epi32(s3_25_4, DCT_CONST_BITS); + const __m128i s3_25_7 = _mm_srai_epi32(s3_25_5, DCT_CONST_BITS); + const __m128i s3_26_6 = _mm_srai_epi32(s3_26_4, DCT_CONST_BITS); + const __m128i s3_26_7 = _mm_srai_epi32(s3_26_5, DCT_CONST_BITS); + const __m128i s3_29_6 = _mm_srai_epi32(s3_29_4, DCT_CONST_BITS); + const __m128i s3_29_7 = _mm_srai_epi32(s3_29_5, DCT_CONST_BITS); + const __m128i s3_30_6 = _mm_srai_epi32(s3_30_4, DCT_CONST_BITS); + const __m128i s3_30_7 = _mm_srai_epi32(s3_30_5, DCT_CONST_BITS); + // Combine + step3[17] = _mm_packs_epi32(s3_17_6, s3_17_7); + step3[18] = _mm_packs_epi32(s3_18_6, s3_18_7); + step3[21] = _mm_packs_epi32(s3_21_6, s3_21_7); + step3[22] = _mm_packs_epi32(s3_22_6, s3_22_7); + // Combine + step3[25] = _mm_packs_epi32(s3_25_6, s3_25_7); + step3[26] = _mm_packs_epi32(s3_26_6, s3_26_7); + step3[29] = _mm_packs_epi32(s3_29_6, s3_29_7); + step3[30] = _mm_packs_epi32(s3_30_6, s3_30_7); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x8(&step3[17], &step3[18], &step3[21], + &step3[22], &step3[25], &step3[26], + &step3[29], &step3[30]); + if (overflow) { + if (pass == 0) + HIGH_FDCT32x32_2D_C(input, output_org, stride); + else + HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + } + // Stage 7 + { + const __m128i out_02_0 = _mm_unpacklo_epi16(step3[8], step3[15]); + const __m128i out_02_1 = _mm_unpackhi_epi16(step3[8], step3[15]); + const __m128i out_18_0 = _mm_unpacklo_epi16(step3[9], step3[14]); + const __m128i out_18_1 = _mm_unpackhi_epi16(step3[9], step3[14]); + const __m128i out_10_0 = _mm_unpacklo_epi16(step3[10], step3[13]); + const __m128i out_10_1 = _mm_unpackhi_epi16(step3[10], step3[13]); + const __m128i out_26_0 = _mm_unpacklo_epi16(step3[11], step3[12]); + const __m128i out_26_1 = _mm_unpackhi_epi16(step3[11], step3[12]); + const __m128i out_02_2 = _mm_madd_epi16(out_02_0, k__cospi_p30_p02); + const __m128i out_02_3 = _mm_madd_epi16(out_02_1, k__cospi_p30_p02); + const __m128i out_18_2 = _mm_madd_epi16(out_18_0, k__cospi_p14_p18); + const __m128i out_18_3 = _mm_madd_epi16(out_18_1, k__cospi_p14_p18); + const __m128i out_10_2 = _mm_madd_epi16(out_10_0, k__cospi_p22_p10); + const __m128i out_10_3 = _mm_madd_epi16(out_10_1, k__cospi_p22_p10); + const __m128i out_26_2 = _mm_madd_epi16(out_26_0, k__cospi_p06_p26); + const __m128i out_26_3 = _mm_madd_epi16(out_26_1, k__cospi_p06_p26); + const __m128i out_06_2 = _mm_madd_epi16(out_26_0, k__cospi_m26_p06); + const __m128i out_06_3 = _mm_madd_epi16(out_26_1, k__cospi_m26_p06); + const __m128i out_22_2 = _mm_madd_epi16(out_10_0, k__cospi_m10_p22); + const __m128i out_22_3 = _mm_madd_epi16(out_10_1, k__cospi_m10_p22); + const __m128i out_14_2 = _mm_madd_epi16(out_18_0, k__cospi_m18_p14); + const __m128i out_14_3 = _mm_madd_epi16(out_18_1, k__cospi_m18_p14); + const __m128i out_30_2 = _mm_madd_epi16(out_02_0, k__cospi_m02_p30); + const __m128i out_30_3 = _mm_madd_epi16(out_02_1, k__cospi_m02_p30); + // dct_const_round_shift + const __m128i out_02_4 = + _mm_add_epi32(out_02_2, k__DCT_CONST_ROUNDING); + const __m128i out_02_5 = + _mm_add_epi32(out_02_3, k__DCT_CONST_ROUNDING); + const __m128i out_18_4 = + _mm_add_epi32(out_18_2, k__DCT_CONST_ROUNDING); + const __m128i out_18_5 = + _mm_add_epi32(out_18_3, k__DCT_CONST_ROUNDING); + const __m128i out_10_4 = + _mm_add_epi32(out_10_2, k__DCT_CONST_ROUNDING); + const __m128i out_10_5 = + _mm_add_epi32(out_10_3, k__DCT_CONST_ROUNDING); + const __m128i out_26_4 = + _mm_add_epi32(out_26_2, k__DCT_CONST_ROUNDING); + const __m128i out_26_5 = + _mm_add_epi32(out_26_3, k__DCT_CONST_ROUNDING); + const __m128i out_06_4 = + _mm_add_epi32(out_06_2, k__DCT_CONST_ROUNDING); + const __m128i out_06_5 = + _mm_add_epi32(out_06_3, k__DCT_CONST_ROUNDING); + const __m128i out_22_4 = + _mm_add_epi32(out_22_2, k__DCT_CONST_ROUNDING); + const __m128i out_22_5 = + _mm_add_epi32(out_22_3, k__DCT_CONST_ROUNDING); + const __m128i out_14_4 = + _mm_add_epi32(out_14_2, k__DCT_CONST_ROUNDING); + const __m128i out_14_5 = + _mm_add_epi32(out_14_3, k__DCT_CONST_ROUNDING); + const __m128i out_30_4 = + _mm_add_epi32(out_30_2, k__DCT_CONST_ROUNDING); + const __m128i out_30_5 = + _mm_add_epi32(out_30_3, k__DCT_CONST_ROUNDING); + const __m128i out_02_6 = _mm_srai_epi32(out_02_4, DCT_CONST_BITS); + const __m128i out_02_7 = _mm_srai_epi32(out_02_5, DCT_CONST_BITS); + const __m128i out_18_6 = _mm_srai_epi32(out_18_4, DCT_CONST_BITS); + const __m128i out_18_7 = _mm_srai_epi32(out_18_5, DCT_CONST_BITS); + const __m128i out_10_6 = _mm_srai_epi32(out_10_4, DCT_CONST_BITS); + const __m128i out_10_7 = _mm_srai_epi32(out_10_5, DCT_CONST_BITS); + const __m128i out_26_6 = _mm_srai_epi32(out_26_4, DCT_CONST_BITS); + const __m128i out_26_7 = _mm_srai_epi32(out_26_5, DCT_CONST_BITS); + const __m128i out_06_6 = _mm_srai_epi32(out_06_4, DCT_CONST_BITS); + const __m128i out_06_7 = _mm_srai_epi32(out_06_5, DCT_CONST_BITS); + const __m128i out_22_6 = _mm_srai_epi32(out_22_4, DCT_CONST_BITS); + const __m128i out_22_7 = _mm_srai_epi32(out_22_5, DCT_CONST_BITS); + const __m128i out_14_6 = _mm_srai_epi32(out_14_4, DCT_CONST_BITS); + const __m128i out_14_7 = _mm_srai_epi32(out_14_5, DCT_CONST_BITS); + const __m128i out_30_6 = _mm_srai_epi32(out_30_4, DCT_CONST_BITS); + const __m128i out_30_7 = _mm_srai_epi32(out_30_5, DCT_CONST_BITS); + // Combine + out[2] = _mm_packs_epi32(out_02_6, out_02_7); + out[18] = _mm_packs_epi32(out_18_6, out_18_7); + out[10] = _mm_packs_epi32(out_10_6, out_10_7); + out[26] = _mm_packs_epi32(out_26_6, out_26_7); + out[6] = _mm_packs_epi32(out_06_6, out_06_7); + out[22] = _mm_packs_epi32(out_22_6, out_22_7); + out[14] = _mm_packs_epi32(out_14_6, out_14_7); + out[30] = _mm_packs_epi32(out_30_6, out_30_7); +#if DCT_HIGH_BIT_DEPTH + overflow = + check_epi16_overflow_x8(&out[2], &out[18], &out[10], &out[26], + &out[6], &out[22], &out[14], &out[30]); + if (overflow) { + if (pass == 0) + HIGH_FDCT32x32_2D_C(input, output_org, stride); + else + HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + } + { + step1[16] = ADD_EPI16(step3[17], step2[16]); + step1[17] = SUB_EPI16(step2[16], step3[17]); + step1[18] = SUB_EPI16(step2[19], step3[18]); + step1[19] = ADD_EPI16(step3[18], step2[19]); + step1[20] = ADD_EPI16(step3[21], step2[20]); + step1[21] = SUB_EPI16(step2[20], step3[21]); + step1[22] = SUB_EPI16(step2[23], step3[22]); + step1[23] = ADD_EPI16(step3[22], step2[23]); + step1[24] = ADD_EPI16(step3[25], step2[24]); + step1[25] = SUB_EPI16(step2[24], step3[25]); + step1[26] = SUB_EPI16(step2[27], step3[26]); + step1[27] = ADD_EPI16(step3[26], step2[27]); + step1[28] = ADD_EPI16(step3[29], step2[28]); + step1[29] = SUB_EPI16(step2[28], step3[29]); + step1[30] = SUB_EPI16(step2[31], step3[30]); + step1[31] = ADD_EPI16(step3[30], step2[31]); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x16( + &step1[16], &step1[17], &step1[18], &step1[19], &step1[20], + &step1[21], &step1[22], &step1[23], &step1[24], &step1[25], + &step1[26], &step1[27], &step1[28], &step1[29], &step1[30], + &step1[31]); + if (overflow) { + if (pass == 0) + HIGH_FDCT32x32_2D_C(input, output_org, stride); + else + HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + } + // Final stage --- outputs indices are bit-reversed. + { + const __m128i out_01_0 = _mm_unpacklo_epi16(step1[16], step1[31]); + const __m128i out_01_1 = _mm_unpackhi_epi16(step1[16], step1[31]); + const __m128i out_17_0 = _mm_unpacklo_epi16(step1[17], step1[30]); + const __m128i out_17_1 = _mm_unpackhi_epi16(step1[17], step1[30]); + const __m128i out_09_0 = _mm_unpacklo_epi16(step1[18], step1[29]); + const __m128i out_09_1 = _mm_unpackhi_epi16(step1[18], step1[29]); + const __m128i out_25_0 = _mm_unpacklo_epi16(step1[19], step1[28]); + const __m128i out_25_1 = _mm_unpackhi_epi16(step1[19], step1[28]); + const __m128i out_01_2 = _mm_madd_epi16(out_01_0, k__cospi_p31_p01); + const __m128i out_01_3 = _mm_madd_epi16(out_01_1, k__cospi_p31_p01); + const __m128i out_17_2 = _mm_madd_epi16(out_17_0, k__cospi_p15_p17); + const __m128i out_17_3 = _mm_madd_epi16(out_17_1, k__cospi_p15_p17); + const __m128i out_09_2 = _mm_madd_epi16(out_09_0, k__cospi_p23_p09); + const __m128i out_09_3 = _mm_madd_epi16(out_09_1, k__cospi_p23_p09); + const __m128i out_25_2 = _mm_madd_epi16(out_25_0, k__cospi_p07_p25); + const __m128i out_25_3 = _mm_madd_epi16(out_25_1, k__cospi_p07_p25); + const __m128i out_07_2 = _mm_madd_epi16(out_25_0, k__cospi_m25_p07); + const __m128i out_07_3 = _mm_madd_epi16(out_25_1, k__cospi_m25_p07); + const __m128i out_23_2 = _mm_madd_epi16(out_09_0, k__cospi_m09_p23); + const __m128i out_23_3 = _mm_madd_epi16(out_09_1, k__cospi_m09_p23); + const __m128i out_15_2 = _mm_madd_epi16(out_17_0, k__cospi_m17_p15); + const __m128i out_15_3 = _mm_madd_epi16(out_17_1, k__cospi_m17_p15); + const __m128i out_31_2 = _mm_madd_epi16(out_01_0, k__cospi_m01_p31); + const __m128i out_31_3 = _mm_madd_epi16(out_01_1, k__cospi_m01_p31); + // dct_const_round_shift + const __m128i out_01_4 = + _mm_add_epi32(out_01_2, k__DCT_CONST_ROUNDING); + const __m128i out_01_5 = + _mm_add_epi32(out_01_3, k__DCT_CONST_ROUNDING); + const __m128i out_17_4 = + _mm_add_epi32(out_17_2, k__DCT_CONST_ROUNDING); + const __m128i out_17_5 = + _mm_add_epi32(out_17_3, k__DCT_CONST_ROUNDING); + const __m128i out_09_4 = + _mm_add_epi32(out_09_2, k__DCT_CONST_ROUNDING); + const __m128i out_09_5 = + _mm_add_epi32(out_09_3, k__DCT_CONST_ROUNDING); + const __m128i out_25_4 = + _mm_add_epi32(out_25_2, k__DCT_CONST_ROUNDING); + const __m128i out_25_5 = + _mm_add_epi32(out_25_3, k__DCT_CONST_ROUNDING); + const __m128i out_07_4 = + _mm_add_epi32(out_07_2, k__DCT_CONST_ROUNDING); + const __m128i out_07_5 = + _mm_add_epi32(out_07_3, k__DCT_CONST_ROUNDING); + const __m128i out_23_4 = + _mm_add_epi32(out_23_2, k__DCT_CONST_ROUNDING); + const __m128i out_23_5 = + _mm_add_epi32(out_23_3, k__DCT_CONST_ROUNDING); + const __m128i out_15_4 = + _mm_add_epi32(out_15_2, k__DCT_CONST_ROUNDING); + const __m128i out_15_5 = + _mm_add_epi32(out_15_3, k__DCT_CONST_ROUNDING); + const __m128i out_31_4 = + _mm_add_epi32(out_31_2, k__DCT_CONST_ROUNDING); + const __m128i out_31_5 = + _mm_add_epi32(out_31_3, k__DCT_CONST_ROUNDING); + const __m128i out_01_6 = _mm_srai_epi32(out_01_4, DCT_CONST_BITS); + const __m128i out_01_7 = _mm_srai_epi32(out_01_5, DCT_CONST_BITS); + const __m128i out_17_6 = _mm_srai_epi32(out_17_4, DCT_CONST_BITS); + const __m128i out_17_7 = _mm_srai_epi32(out_17_5, DCT_CONST_BITS); + const __m128i out_09_6 = _mm_srai_epi32(out_09_4, DCT_CONST_BITS); + const __m128i out_09_7 = _mm_srai_epi32(out_09_5, DCT_CONST_BITS); + const __m128i out_25_6 = _mm_srai_epi32(out_25_4, DCT_CONST_BITS); + const __m128i out_25_7 = _mm_srai_epi32(out_25_5, DCT_CONST_BITS); + const __m128i out_07_6 = _mm_srai_epi32(out_07_4, DCT_CONST_BITS); + const __m128i out_07_7 = _mm_srai_epi32(out_07_5, DCT_CONST_BITS); + const __m128i out_23_6 = _mm_srai_epi32(out_23_4, DCT_CONST_BITS); + const __m128i out_23_7 = _mm_srai_epi32(out_23_5, DCT_CONST_BITS); + const __m128i out_15_6 = _mm_srai_epi32(out_15_4, DCT_CONST_BITS); + const __m128i out_15_7 = _mm_srai_epi32(out_15_5, DCT_CONST_BITS); + const __m128i out_31_6 = _mm_srai_epi32(out_31_4, DCT_CONST_BITS); + const __m128i out_31_7 = _mm_srai_epi32(out_31_5, DCT_CONST_BITS); + // Combine + out[1] = _mm_packs_epi32(out_01_6, out_01_7); + out[17] = _mm_packs_epi32(out_17_6, out_17_7); + out[9] = _mm_packs_epi32(out_09_6, out_09_7); + out[25] = _mm_packs_epi32(out_25_6, out_25_7); + out[7] = _mm_packs_epi32(out_07_6, out_07_7); + out[23] = _mm_packs_epi32(out_23_6, out_23_7); + out[15] = _mm_packs_epi32(out_15_6, out_15_7); + out[31] = _mm_packs_epi32(out_31_6, out_31_7); +#if DCT_HIGH_BIT_DEPTH + overflow = + check_epi16_overflow_x8(&out[1], &out[17], &out[9], &out[25], + &out[7], &out[23], &out[15], &out[31]); + if (overflow) { + if (pass == 0) + HIGH_FDCT32x32_2D_C(input, output_org, stride); + else + HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + } + { + const __m128i out_05_0 = _mm_unpacklo_epi16(step1[20], step1[27]); + const __m128i out_05_1 = _mm_unpackhi_epi16(step1[20], step1[27]); + const __m128i out_21_0 = _mm_unpacklo_epi16(step1[21], step1[26]); + const __m128i out_21_1 = _mm_unpackhi_epi16(step1[21], step1[26]); + const __m128i out_13_0 = _mm_unpacklo_epi16(step1[22], step1[25]); + const __m128i out_13_1 = _mm_unpackhi_epi16(step1[22], step1[25]); + const __m128i out_29_0 = _mm_unpacklo_epi16(step1[23], step1[24]); + const __m128i out_29_1 = _mm_unpackhi_epi16(step1[23], step1[24]); + const __m128i out_05_2 = _mm_madd_epi16(out_05_0, k__cospi_p27_p05); + const __m128i out_05_3 = _mm_madd_epi16(out_05_1, k__cospi_p27_p05); + const __m128i out_21_2 = _mm_madd_epi16(out_21_0, k__cospi_p11_p21); + const __m128i out_21_3 = _mm_madd_epi16(out_21_1, k__cospi_p11_p21); + const __m128i out_13_2 = _mm_madd_epi16(out_13_0, k__cospi_p19_p13); + const __m128i out_13_3 = _mm_madd_epi16(out_13_1, k__cospi_p19_p13); + const __m128i out_29_2 = _mm_madd_epi16(out_29_0, k__cospi_p03_p29); + const __m128i out_29_3 = _mm_madd_epi16(out_29_1, k__cospi_p03_p29); + const __m128i out_03_2 = _mm_madd_epi16(out_29_0, k__cospi_m29_p03); + const __m128i out_03_3 = _mm_madd_epi16(out_29_1, k__cospi_m29_p03); + const __m128i out_19_2 = _mm_madd_epi16(out_13_0, k__cospi_m13_p19); + const __m128i out_19_3 = _mm_madd_epi16(out_13_1, k__cospi_m13_p19); + const __m128i out_11_2 = _mm_madd_epi16(out_21_0, k__cospi_m21_p11); + const __m128i out_11_3 = _mm_madd_epi16(out_21_1, k__cospi_m21_p11); + const __m128i out_27_2 = _mm_madd_epi16(out_05_0, k__cospi_m05_p27); + const __m128i out_27_3 = _mm_madd_epi16(out_05_1, k__cospi_m05_p27); + // dct_const_round_shift + const __m128i out_05_4 = + _mm_add_epi32(out_05_2, k__DCT_CONST_ROUNDING); + const __m128i out_05_5 = + _mm_add_epi32(out_05_3, k__DCT_CONST_ROUNDING); + const __m128i out_21_4 = + _mm_add_epi32(out_21_2, k__DCT_CONST_ROUNDING); + const __m128i out_21_5 = + _mm_add_epi32(out_21_3, k__DCT_CONST_ROUNDING); + const __m128i out_13_4 = + _mm_add_epi32(out_13_2, k__DCT_CONST_ROUNDING); + const __m128i out_13_5 = + _mm_add_epi32(out_13_3, k__DCT_CONST_ROUNDING); + const __m128i out_29_4 = + _mm_add_epi32(out_29_2, k__DCT_CONST_ROUNDING); + const __m128i out_29_5 = + _mm_add_epi32(out_29_3, k__DCT_CONST_ROUNDING); + const __m128i out_03_4 = + _mm_add_epi32(out_03_2, k__DCT_CONST_ROUNDING); + const __m128i out_03_5 = + _mm_add_epi32(out_03_3, k__DCT_CONST_ROUNDING); + const __m128i out_19_4 = + _mm_add_epi32(out_19_2, k__DCT_CONST_ROUNDING); + const __m128i out_19_5 = + _mm_add_epi32(out_19_3, k__DCT_CONST_ROUNDING); + const __m128i out_11_4 = + _mm_add_epi32(out_11_2, k__DCT_CONST_ROUNDING); + const __m128i out_11_5 = + _mm_add_epi32(out_11_3, k__DCT_CONST_ROUNDING); + const __m128i out_27_4 = + _mm_add_epi32(out_27_2, k__DCT_CONST_ROUNDING); + const __m128i out_27_5 = + _mm_add_epi32(out_27_3, k__DCT_CONST_ROUNDING); + const __m128i out_05_6 = _mm_srai_epi32(out_05_4, DCT_CONST_BITS); + const __m128i out_05_7 = _mm_srai_epi32(out_05_5, DCT_CONST_BITS); + const __m128i out_21_6 = _mm_srai_epi32(out_21_4, DCT_CONST_BITS); + const __m128i out_21_7 = _mm_srai_epi32(out_21_5, DCT_CONST_BITS); + const __m128i out_13_6 = _mm_srai_epi32(out_13_4, DCT_CONST_BITS); + const __m128i out_13_7 = _mm_srai_epi32(out_13_5, DCT_CONST_BITS); + const __m128i out_29_6 = _mm_srai_epi32(out_29_4, DCT_CONST_BITS); + const __m128i out_29_7 = _mm_srai_epi32(out_29_5, DCT_CONST_BITS); + const __m128i out_03_6 = _mm_srai_epi32(out_03_4, DCT_CONST_BITS); + const __m128i out_03_7 = _mm_srai_epi32(out_03_5, DCT_CONST_BITS); + const __m128i out_19_6 = _mm_srai_epi32(out_19_4, DCT_CONST_BITS); + const __m128i out_19_7 = _mm_srai_epi32(out_19_5, DCT_CONST_BITS); + const __m128i out_11_6 = _mm_srai_epi32(out_11_4, DCT_CONST_BITS); + const __m128i out_11_7 = _mm_srai_epi32(out_11_5, DCT_CONST_BITS); + const __m128i out_27_6 = _mm_srai_epi32(out_27_4, DCT_CONST_BITS); + const __m128i out_27_7 = _mm_srai_epi32(out_27_5, DCT_CONST_BITS); + // Combine + out[5] = _mm_packs_epi32(out_05_6, out_05_7); + out[21] = _mm_packs_epi32(out_21_6, out_21_7); + out[13] = _mm_packs_epi32(out_13_6, out_13_7); + out[29] = _mm_packs_epi32(out_29_6, out_29_7); + out[3] = _mm_packs_epi32(out_03_6, out_03_7); + out[19] = _mm_packs_epi32(out_19_6, out_19_7); + out[11] = _mm_packs_epi32(out_11_6, out_11_7); + out[27] = _mm_packs_epi32(out_27_6, out_27_7); +#if DCT_HIGH_BIT_DEPTH + overflow = + check_epi16_overflow_x8(&out[5], &out[21], &out[13], &out[29], + &out[3], &out[19], &out[11], &out[27]); + if (overflow) { + if (pass == 0) + HIGH_FDCT32x32_2D_C(input, output_org, stride); + else + HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + } +#if FDCT32x32_HIGH_PRECISION + } else { + __m128i lstep1[64], lstep2[64], lstep3[64]; + __m128i u[32], v[32], sign[16]; + const __m128i K32One = _mm_set_epi32(1, 1, 1, 1); + // start using 32-bit operations + // stage 3 + { + // expanding to 32-bit length priori to addition operations + lstep2[0] = _mm_unpacklo_epi16(step2[0], kZero); + lstep2[1] = _mm_unpackhi_epi16(step2[0], kZero); + lstep2[2] = _mm_unpacklo_epi16(step2[1], kZero); + lstep2[3] = _mm_unpackhi_epi16(step2[1], kZero); + lstep2[4] = _mm_unpacklo_epi16(step2[2], kZero); + lstep2[5] = _mm_unpackhi_epi16(step2[2], kZero); + lstep2[6] = _mm_unpacklo_epi16(step2[3], kZero); + lstep2[7] = _mm_unpackhi_epi16(step2[3], kZero); + lstep2[8] = _mm_unpacklo_epi16(step2[4], kZero); + lstep2[9] = _mm_unpackhi_epi16(step2[4], kZero); + lstep2[10] = _mm_unpacklo_epi16(step2[5], kZero); + lstep2[11] = _mm_unpackhi_epi16(step2[5], kZero); + lstep2[12] = _mm_unpacklo_epi16(step2[6], kZero); + lstep2[13] = _mm_unpackhi_epi16(step2[6], kZero); + lstep2[14] = _mm_unpacklo_epi16(step2[7], kZero); + lstep2[15] = _mm_unpackhi_epi16(step2[7], kZero); + lstep2[0] = _mm_madd_epi16(lstep2[0], kOne); + lstep2[1] = _mm_madd_epi16(lstep2[1], kOne); + lstep2[2] = _mm_madd_epi16(lstep2[2], kOne); + lstep2[3] = _mm_madd_epi16(lstep2[3], kOne); + lstep2[4] = _mm_madd_epi16(lstep2[4], kOne); + lstep2[5] = _mm_madd_epi16(lstep2[5], kOne); + lstep2[6] = _mm_madd_epi16(lstep2[6], kOne); + lstep2[7] = _mm_madd_epi16(lstep2[7], kOne); + lstep2[8] = _mm_madd_epi16(lstep2[8], kOne); + lstep2[9] = _mm_madd_epi16(lstep2[9], kOne); + lstep2[10] = _mm_madd_epi16(lstep2[10], kOne); + lstep2[11] = _mm_madd_epi16(lstep2[11], kOne); + lstep2[12] = _mm_madd_epi16(lstep2[12], kOne); + lstep2[13] = _mm_madd_epi16(lstep2[13], kOne); + lstep2[14] = _mm_madd_epi16(lstep2[14], kOne); + lstep2[15] = _mm_madd_epi16(lstep2[15], kOne); + + lstep3[0] = _mm_add_epi32(lstep2[14], lstep2[0]); + lstep3[1] = _mm_add_epi32(lstep2[15], lstep2[1]); + lstep3[2] = _mm_add_epi32(lstep2[12], lstep2[2]); + lstep3[3] = _mm_add_epi32(lstep2[13], lstep2[3]); + lstep3[4] = _mm_add_epi32(lstep2[10], lstep2[4]); + lstep3[5] = _mm_add_epi32(lstep2[11], lstep2[5]); + lstep3[6] = _mm_add_epi32(lstep2[8], lstep2[6]); + lstep3[7] = _mm_add_epi32(lstep2[9], lstep2[7]); + lstep3[8] = _mm_sub_epi32(lstep2[6], lstep2[8]); + lstep3[9] = _mm_sub_epi32(lstep2[7], lstep2[9]); + lstep3[10] = _mm_sub_epi32(lstep2[4], lstep2[10]); + lstep3[11] = _mm_sub_epi32(lstep2[5], lstep2[11]); + lstep3[12] = _mm_sub_epi32(lstep2[2], lstep2[12]); + lstep3[13] = _mm_sub_epi32(lstep2[3], lstep2[13]); + lstep3[14] = _mm_sub_epi32(lstep2[0], lstep2[14]); + lstep3[15] = _mm_sub_epi32(lstep2[1], lstep2[15]); + } + { + const __m128i s3_10_0 = _mm_unpacklo_epi16(step2[13], step2[10]); + const __m128i s3_10_1 = _mm_unpackhi_epi16(step2[13], step2[10]); + const __m128i s3_11_0 = _mm_unpacklo_epi16(step2[12], step2[11]); + const __m128i s3_11_1 = _mm_unpackhi_epi16(step2[12], step2[11]); + const __m128i s3_10_2 = _mm_madd_epi16(s3_10_0, k__cospi_p16_m16); + const __m128i s3_10_3 = _mm_madd_epi16(s3_10_1, k__cospi_p16_m16); + const __m128i s3_11_2 = _mm_madd_epi16(s3_11_0, k__cospi_p16_m16); + const __m128i s3_11_3 = _mm_madd_epi16(s3_11_1, k__cospi_p16_m16); + const __m128i s3_12_2 = _mm_madd_epi16(s3_11_0, k__cospi_p16_p16); + const __m128i s3_12_3 = _mm_madd_epi16(s3_11_1, k__cospi_p16_p16); + const __m128i s3_13_2 = _mm_madd_epi16(s3_10_0, k__cospi_p16_p16); + const __m128i s3_13_3 = _mm_madd_epi16(s3_10_1, k__cospi_p16_p16); + // dct_const_round_shift + const __m128i s3_10_4 = _mm_add_epi32(s3_10_2, k__DCT_CONST_ROUNDING); + const __m128i s3_10_5 = _mm_add_epi32(s3_10_3, k__DCT_CONST_ROUNDING); + const __m128i s3_11_4 = _mm_add_epi32(s3_11_2, k__DCT_CONST_ROUNDING); + const __m128i s3_11_5 = _mm_add_epi32(s3_11_3, k__DCT_CONST_ROUNDING); + const __m128i s3_12_4 = _mm_add_epi32(s3_12_2, k__DCT_CONST_ROUNDING); + const __m128i s3_12_5 = _mm_add_epi32(s3_12_3, k__DCT_CONST_ROUNDING); + const __m128i s3_13_4 = _mm_add_epi32(s3_13_2, k__DCT_CONST_ROUNDING); + const __m128i s3_13_5 = _mm_add_epi32(s3_13_3, k__DCT_CONST_ROUNDING); + lstep3[20] = _mm_srai_epi32(s3_10_4, DCT_CONST_BITS); + lstep3[21] = _mm_srai_epi32(s3_10_5, DCT_CONST_BITS); + lstep3[22] = _mm_srai_epi32(s3_11_4, DCT_CONST_BITS); + lstep3[23] = _mm_srai_epi32(s3_11_5, DCT_CONST_BITS); + lstep3[24] = _mm_srai_epi32(s3_12_4, DCT_CONST_BITS); + lstep3[25] = _mm_srai_epi32(s3_12_5, DCT_CONST_BITS); + lstep3[26] = _mm_srai_epi32(s3_13_4, DCT_CONST_BITS); + lstep3[27] = _mm_srai_epi32(s3_13_5, DCT_CONST_BITS); + } + { + lstep2[40] = _mm_unpacklo_epi16(step2[20], kZero); + lstep2[41] = _mm_unpackhi_epi16(step2[20], kZero); + lstep2[42] = _mm_unpacklo_epi16(step2[21], kZero); + lstep2[43] = _mm_unpackhi_epi16(step2[21], kZero); + lstep2[44] = _mm_unpacklo_epi16(step2[22], kZero); + lstep2[45] = _mm_unpackhi_epi16(step2[22], kZero); + lstep2[46] = _mm_unpacklo_epi16(step2[23], kZero); + lstep2[47] = _mm_unpackhi_epi16(step2[23], kZero); + lstep2[48] = _mm_unpacklo_epi16(step2[24], kZero); + lstep2[49] = _mm_unpackhi_epi16(step2[24], kZero); + lstep2[50] = _mm_unpacklo_epi16(step2[25], kZero); + lstep2[51] = _mm_unpackhi_epi16(step2[25], kZero); + lstep2[52] = _mm_unpacklo_epi16(step2[26], kZero); + lstep2[53] = _mm_unpackhi_epi16(step2[26], kZero); + lstep2[54] = _mm_unpacklo_epi16(step2[27], kZero); + lstep2[55] = _mm_unpackhi_epi16(step2[27], kZero); + lstep2[40] = _mm_madd_epi16(lstep2[40], kOne); + lstep2[41] = _mm_madd_epi16(lstep2[41], kOne); + lstep2[42] = _mm_madd_epi16(lstep2[42], kOne); + lstep2[43] = _mm_madd_epi16(lstep2[43], kOne); + lstep2[44] = _mm_madd_epi16(lstep2[44], kOne); + lstep2[45] = _mm_madd_epi16(lstep2[45], kOne); + lstep2[46] = _mm_madd_epi16(lstep2[46], kOne); + lstep2[47] = _mm_madd_epi16(lstep2[47], kOne); + lstep2[48] = _mm_madd_epi16(lstep2[48], kOne); + lstep2[49] = _mm_madd_epi16(lstep2[49], kOne); + lstep2[50] = _mm_madd_epi16(lstep2[50], kOne); + lstep2[51] = _mm_madd_epi16(lstep2[51], kOne); + lstep2[52] = _mm_madd_epi16(lstep2[52], kOne); + lstep2[53] = _mm_madd_epi16(lstep2[53], kOne); + lstep2[54] = _mm_madd_epi16(lstep2[54], kOne); + lstep2[55] = _mm_madd_epi16(lstep2[55], kOne); + + lstep1[32] = _mm_unpacklo_epi16(step1[16], kZero); + lstep1[33] = _mm_unpackhi_epi16(step1[16], kZero); + lstep1[34] = _mm_unpacklo_epi16(step1[17], kZero); + lstep1[35] = _mm_unpackhi_epi16(step1[17], kZero); + lstep1[36] = _mm_unpacklo_epi16(step1[18], kZero); + lstep1[37] = _mm_unpackhi_epi16(step1[18], kZero); + lstep1[38] = _mm_unpacklo_epi16(step1[19], kZero); + lstep1[39] = _mm_unpackhi_epi16(step1[19], kZero); + lstep1[56] = _mm_unpacklo_epi16(step1[28], kZero); + lstep1[57] = _mm_unpackhi_epi16(step1[28], kZero); + lstep1[58] = _mm_unpacklo_epi16(step1[29], kZero); + lstep1[59] = _mm_unpackhi_epi16(step1[29], kZero); + lstep1[60] = _mm_unpacklo_epi16(step1[30], kZero); + lstep1[61] = _mm_unpackhi_epi16(step1[30], kZero); + lstep1[62] = _mm_unpacklo_epi16(step1[31], kZero); + lstep1[63] = _mm_unpackhi_epi16(step1[31], kZero); + lstep1[32] = _mm_madd_epi16(lstep1[32], kOne); + lstep1[33] = _mm_madd_epi16(lstep1[33], kOne); + lstep1[34] = _mm_madd_epi16(lstep1[34], kOne); + lstep1[35] = _mm_madd_epi16(lstep1[35], kOne); + lstep1[36] = _mm_madd_epi16(lstep1[36], kOne); + lstep1[37] = _mm_madd_epi16(lstep1[37], kOne); + lstep1[38] = _mm_madd_epi16(lstep1[38], kOne); + lstep1[39] = _mm_madd_epi16(lstep1[39], kOne); + lstep1[56] = _mm_madd_epi16(lstep1[56], kOne); + lstep1[57] = _mm_madd_epi16(lstep1[57], kOne); + lstep1[58] = _mm_madd_epi16(lstep1[58], kOne); + lstep1[59] = _mm_madd_epi16(lstep1[59], kOne); + lstep1[60] = _mm_madd_epi16(lstep1[60], kOne); + lstep1[61] = _mm_madd_epi16(lstep1[61], kOne); + lstep1[62] = _mm_madd_epi16(lstep1[62], kOne); + lstep1[63] = _mm_madd_epi16(lstep1[63], kOne); + + lstep3[32] = _mm_add_epi32(lstep2[46], lstep1[32]); + lstep3[33] = _mm_add_epi32(lstep2[47], lstep1[33]); + + lstep3[34] = _mm_add_epi32(lstep2[44], lstep1[34]); + lstep3[35] = _mm_add_epi32(lstep2[45], lstep1[35]); + lstep3[36] = _mm_add_epi32(lstep2[42], lstep1[36]); + lstep3[37] = _mm_add_epi32(lstep2[43], lstep1[37]); + lstep3[38] = _mm_add_epi32(lstep2[40], lstep1[38]); + lstep3[39] = _mm_add_epi32(lstep2[41], lstep1[39]); + lstep3[40] = _mm_sub_epi32(lstep1[38], lstep2[40]); + lstep3[41] = _mm_sub_epi32(lstep1[39], lstep2[41]); + lstep3[42] = _mm_sub_epi32(lstep1[36], lstep2[42]); + lstep3[43] = _mm_sub_epi32(lstep1[37], lstep2[43]); + lstep3[44] = _mm_sub_epi32(lstep1[34], lstep2[44]); + lstep3[45] = _mm_sub_epi32(lstep1[35], lstep2[45]); + lstep3[46] = _mm_sub_epi32(lstep1[32], lstep2[46]); + lstep3[47] = _mm_sub_epi32(lstep1[33], lstep2[47]); + lstep3[48] = _mm_sub_epi32(lstep1[62], lstep2[48]); + lstep3[49] = _mm_sub_epi32(lstep1[63], lstep2[49]); + lstep3[50] = _mm_sub_epi32(lstep1[60], lstep2[50]); + lstep3[51] = _mm_sub_epi32(lstep1[61], lstep2[51]); + lstep3[52] = _mm_sub_epi32(lstep1[58], lstep2[52]); + lstep3[53] = _mm_sub_epi32(lstep1[59], lstep2[53]); + lstep3[54] = _mm_sub_epi32(lstep1[56], lstep2[54]); + lstep3[55] = _mm_sub_epi32(lstep1[57], lstep2[55]); + lstep3[56] = _mm_add_epi32(lstep2[54], lstep1[56]); + lstep3[57] = _mm_add_epi32(lstep2[55], lstep1[57]); + lstep3[58] = _mm_add_epi32(lstep2[52], lstep1[58]); + lstep3[59] = _mm_add_epi32(lstep2[53], lstep1[59]); + lstep3[60] = _mm_add_epi32(lstep2[50], lstep1[60]); + lstep3[61] = _mm_add_epi32(lstep2[51], lstep1[61]); + lstep3[62] = _mm_add_epi32(lstep2[48], lstep1[62]); + lstep3[63] = _mm_add_epi32(lstep2[49], lstep1[63]); + } + + // stage 4 + { + // expanding to 32-bit length priori to addition operations + lstep2[16] = _mm_unpacklo_epi16(step2[8], kZero); + lstep2[17] = _mm_unpackhi_epi16(step2[8], kZero); + lstep2[18] = _mm_unpacklo_epi16(step2[9], kZero); + lstep2[19] = _mm_unpackhi_epi16(step2[9], kZero); + lstep2[28] = _mm_unpacklo_epi16(step2[14], kZero); + lstep2[29] = _mm_unpackhi_epi16(step2[14], kZero); + lstep2[30] = _mm_unpacklo_epi16(step2[15], kZero); + lstep2[31] = _mm_unpackhi_epi16(step2[15], kZero); + lstep2[16] = _mm_madd_epi16(lstep2[16], kOne); + lstep2[17] = _mm_madd_epi16(lstep2[17], kOne); + lstep2[18] = _mm_madd_epi16(lstep2[18], kOne); + lstep2[19] = _mm_madd_epi16(lstep2[19], kOne); + lstep2[28] = _mm_madd_epi16(lstep2[28], kOne); + lstep2[29] = _mm_madd_epi16(lstep2[29], kOne); + lstep2[30] = _mm_madd_epi16(lstep2[30], kOne); + lstep2[31] = _mm_madd_epi16(lstep2[31], kOne); + + lstep1[0] = _mm_add_epi32(lstep3[6], lstep3[0]); + lstep1[1] = _mm_add_epi32(lstep3[7], lstep3[1]); + lstep1[2] = _mm_add_epi32(lstep3[4], lstep3[2]); + lstep1[3] = _mm_add_epi32(lstep3[5], lstep3[3]); + lstep1[4] = _mm_sub_epi32(lstep3[2], lstep3[4]); + lstep1[5] = _mm_sub_epi32(lstep3[3], lstep3[5]); + lstep1[6] = _mm_sub_epi32(lstep3[0], lstep3[6]); + lstep1[7] = _mm_sub_epi32(lstep3[1], lstep3[7]); + lstep1[16] = _mm_add_epi32(lstep3[22], lstep2[16]); + lstep1[17] = _mm_add_epi32(lstep3[23], lstep2[17]); + lstep1[18] = _mm_add_epi32(lstep3[20], lstep2[18]); + lstep1[19] = _mm_add_epi32(lstep3[21], lstep2[19]); + lstep1[20] = _mm_sub_epi32(lstep2[18], lstep3[20]); + lstep1[21] = _mm_sub_epi32(lstep2[19], lstep3[21]); + lstep1[22] = _mm_sub_epi32(lstep2[16], lstep3[22]); + lstep1[23] = _mm_sub_epi32(lstep2[17], lstep3[23]); + lstep1[24] = _mm_sub_epi32(lstep2[30], lstep3[24]); + lstep1[25] = _mm_sub_epi32(lstep2[31], lstep3[25]); + lstep1[26] = _mm_sub_epi32(lstep2[28], lstep3[26]); + lstep1[27] = _mm_sub_epi32(lstep2[29], lstep3[27]); + lstep1[28] = _mm_add_epi32(lstep3[26], lstep2[28]); + lstep1[29] = _mm_add_epi32(lstep3[27], lstep2[29]); + lstep1[30] = _mm_add_epi32(lstep3[24], lstep2[30]); + lstep1[31] = _mm_add_epi32(lstep3[25], lstep2[31]); + } + { + // to be continued... + // + const __m128i k32_p16_p16 = pair_set_epi32(cospi_16_64, cospi_16_64); + const __m128i k32_p16_m16 = pair_set_epi32(cospi_16_64, -cospi_16_64); + + u[0] = _mm_unpacklo_epi32(lstep3[12], lstep3[10]); + u[1] = _mm_unpackhi_epi32(lstep3[12], lstep3[10]); + u[2] = _mm_unpacklo_epi32(lstep3[13], lstep3[11]); + u[3] = _mm_unpackhi_epi32(lstep3[13], lstep3[11]); + + // TODO(jingning): manually inline k_madd_epi32_ to further hide + // instruction latency. + v[0] = k_madd_epi32(u[0], k32_p16_m16); + v[1] = k_madd_epi32(u[1], k32_p16_m16); + v[2] = k_madd_epi32(u[2], k32_p16_m16); + v[3] = k_madd_epi32(u[3], k32_p16_m16); + v[4] = k_madd_epi32(u[0], k32_p16_p16); + v[5] = k_madd_epi32(u[1], k32_p16_p16); + v[6] = k_madd_epi32(u[2], k32_p16_p16); + v[7] = k_madd_epi32(u[3], k32_p16_p16); +#if DCT_HIGH_BIT_DEPTH + overflow = k_check_epi32_overflow_8(&v[0], &v[1], &v[2], &v[3], &v[4], + &v[5], &v[6], &v[7], &kZero); + if (overflow) { + HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + u[0] = k_packs_epi64(v[0], v[1]); + u[1] = k_packs_epi64(v[2], v[3]); + u[2] = k_packs_epi64(v[4], v[5]); + u[3] = k_packs_epi64(v[6], v[7]); + + v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING); + v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING); + v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING); + v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING); + + lstep1[10] = _mm_srai_epi32(v[0], DCT_CONST_BITS); + lstep1[11] = _mm_srai_epi32(v[1], DCT_CONST_BITS); + lstep1[12] = _mm_srai_epi32(v[2], DCT_CONST_BITS); + lstep1[13] = _mm_srai_epi32(v[3], DCT_CONST_BITS); + } + { + const __m128i k32_m08_p24 = pair_set_epi32(-cospi_8_64, cospi_24_64); + const __m128i k32_m24_m08 = pair_set_epi32(-cospi_24_64, -cospi_8_64); + const __m128i k32_p24_p08 = pair_set_epi32(cospi_24_64, cospi_8_64); + + u[0] = _mm_unpacklo_epi32(lstep3[36], lstep3[58]); + u[1] = _mm_unpackhi_epi32(lstep3[36], lstep3[58]); + u[2] = _mm_unpacklo_epi32(lstep3[37], lstep3[59]); + u[3] = _mm_unpackhi_epi32(lstep3[37], lstep3[59]); + u[4] = _mm_unpacklo_epi32(lstep3[38], lstep3[56]); + u[5] = _mm_unpackhi_epi32(lstep3[38], lstep3[56]); + u[6] = _mm_unpacklo_epi32(lstep3[39], lstep3[57]); + u[7] = _mm_unpackhi_epi32(lstep3[39], lstep3[57]); + u[8] = _mm_unpacklo_epi32(lstep3[40], lstep3[54]); + u[9] = _mm_unpackhi_epi32(lstep3[40], lstep3[54]); + u[10] = _mm_unpacklo_epi32(lstep3[41], lstep3[55]); + u[11] = _mm_unpackhi_epi32(lstep3[41], lstep3[55]); + u[12] = _mm_unpacklo_epi32(lstep3[42], lstep3[52]); + u[13] = _mm_unpackhi_epi32(lstep3[42], lstep3[52]); + u[14] = _mm_unpacklo_epi32(lstep3[43], lstep3[53]); + u[15] = _mm_unpackhi_epi32(lstep3[43], lstep3[53]); + + v[0] = k_madd_epi32(u[0], k32_m08_p24); + v[1] = k_madd_epi32(u[1], k32_m08_p24); + v[2] = k_madd_epi32(u[2], k32_m08_p24); + v[3] = k_madd_epi32(u[3], k32_m08_p24); + v[4] = k_madd_epi32(u[4], k32_m08_p24); + v[5] = k_madd_epi32(u[5], k32_m08_p24); + v[6] = k_madd_epi32(u[6], k32_m08_p24); + v[7] = k_madd_epi32(u[7], k32_m08_p24); + v[8] = k_madd_epi32(u[8], k32_m24_m08); + v[9] = k_madd_epi32(u[9], k32_m24_m08); + v[10] = k_madd_epi32(u[10], k32_m24_m08); + v[11] = k_madd_epi32(u[11], k32_m24_m08); + v[12] = k_madd_epi32(u[12], k32_m24_m08); + v[13] = k_madd_epi32(u[13], k32_m24_m08); + v[14] = k_madd_epi32(u[14], k32_m24_m08); + v[15] = k_madd_epi32(u[15], k32_m24_m08); + v[16] = k_madd_epi32(u[12], k32_m08_p24); + v[17] = k_madd_epi32(u[13], k32_m08_p24); + v[18] = k_madd_epi32(u[14], k32_m08_p24); + v[19] = k_madd_epi32(u[15], k32_m08_p24); + v[20] = k_madd_epi32(u[8], k32_m08_p24); + v[21] = k_madd_epi32(u[9], k32_m08_p24); + v[22] = k_madd_epi32(u[10], k32_m08_p24); + v[23] = k_madd_epi32(u[11], k32_m08_p24); + v[24] = k_madd_epi32(u[4], k32_p24_p08); + v[25] = k_madd_epi32(u[5], k32_p24_p08); + v[26] = k_madd_epi32(u[6], k32_p24_p08); + v[27] = k_madd_epi32(u[7], k32_p24_p08); + v[28] = k_madd_epi32(u[0], k32_p24_p08); + v[29] = k_madd_epi32(u[1], k32_p24_p08); + v[30] = k_madd_epi32(u[2], k32_p24_p08); + v[31] = k_madd_epi32(u[3], k32_p24_p08); + +#if DCT_HIGH_BIT_DEPTH + overflow = k_check_epi32_overflow_32( + &v[0], &v[1], &v[2], &v[3], &v[4], &v[5], &v[6], &v[7], &v[8], + &v[9], &v[10], &v[11], &v[12], &v[13], &v[14], &v[15], &v[16], + &v[17], &v[18], &v[19], &v[20], &v[21], &v[22], &v[23], &v[24], + &v[25], &v[26], &v[27], &v[28], &v[29], &v[30], &v[31], &kZero); + if (overflow) { + HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + u[0] = k_packs_epi64(v[0], v[1]); + u[1] = k_packs_epi64(v[2], v[3]); + u[2] = k_packs_epi64(v[4], v[5]); + u[3] = k_packs_epi64(v[6], v[7]); + u[4] = k_packs_epi64(v[8], v[9]); + u[5] = k_packs_epi64(v[10], v[11]); + u[6] = k_packs_epi64(v[12], v[13]); + u[7] = k_packs_epi64(v[14], v[15]); + u[8] = k_packs_epi64(v[16], v[17]); + u[9] = k_packs_epi64(v[18], v[19]); + u[10] = k_packs_epi64(v[20], v[21]); + u[11] = k_packs_epi64(v[22], v[23]); + u[12] = k_packs_epi64(v[24], v[25]); + u[13] = k_packs_epi64(v[26], v[27]); + u[14] = k_packs_epi64(v[28], v[29]); + u[15] = k_packs_epi64(v[30], v[31]); + + v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING); + v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING); + v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING); + v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING); + v[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING); + v[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING); + v[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING); + v[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING); + v[8] = _mm_add_epi32(u[8], k__DCT_CONST_ROUNDING); + v[9] = _mm_add_epi32(u[9], k__DCT_CONST_ROUNDING); + v[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING); + v[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING); + v[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING); + v[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING); + v[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING); + v[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING); + + lstep1[36] = _mm_srai_epi32(v[0], DCT_CONST_BITS); + lstep1[37] = _mm_srai_epi32(v[1], DCT_CONST_BITS); + lstep1[38] = _mm_srai_epi32(v[2], DCT_CONST_BITS); + lstep1[39] = _mm_srai_epi32(v[3], DCT_CONST_BITS); + lstep1[40] = _mm_srai_epi32(v[4], DCT_CONST_BITS); + lstep1[41] = _mm_srai_epi32(v[5], DCT_CONST_BITS); + lstep1[42] = _mm_srai_epi32(v[6], DCT_CONST_BITS); + lstep1[43] = _mm_srai_epi32(v[7], DCT_CONST_BITS); + lstep1[52] = _mm_srai_epi32(v[8], DCT_CONST_BITS); + lstep1[53] = _mm_srai_epi32(v[9], DCT_CONST_BITS); + lstep1[54] = _mm_srai_epi32(v[10], DCT_CONST_BITS); + lstep1[55] = _mm_srai_epi32(v[11], DCT_CONST_BITS); + lstep1[56] = _mm_srai_epi32(v[12], DCT_CONST_BITS); + lstep1[57] = _mm_srai_epi32(v[13], DCT_CONST_BITS); + lstep1[58] = _mm_srai_epi32(v[14], DCT_CONST_BITS); + lstep1[59] = _mm_srai_epi32(v[15], DCT_CONST_BITS); + } + // stage 5 + { + lstep2[8] = _mm_add_epi32(lstep1[10], lstep3[8]); + lstep2[9] = _mm_add_epi32(lstep1[11], lstep3[9]); + lstep2[10] = _mm_sub_epi32(lstep3[8], lstep1[10]); + lstep2[11] = _mm_sub_epi32(lstep3[9], lstep1[11]); + lstep2[12] = _mm_sub_epi32(lstep3[14], lstep1[12]); + lstep2[13] = _mm_sub_epi32(lstep3[15], lstep1[13]); + lstep2[14] = _mm_add_epi32(lstep1[12], lstep3[14]); + lstep2[15] = _mm_add_epi32(lstep1[13], lstep3[15]); + } + { + const __m128i k32_p16_p16 = pair_set_epi32(cospi_16_64, cospi_16_64); + const __m128i k32_p16_m16 = pair_set_epi32(cospi_16_64, -cospi_16_64); + const __m128i k32_p24_p08 = pair_set_epi32(cospi_24_64, cospi_8_64); + const __m128i k32_m08_p24 = pair_set_epi32(-cospi_8_64, cospi_24_64); + + u[0] = _mm_unpacklo_epi32(lstep1[0], lstep1[2]); + u[1] = _mm_unpackhi_epi32(lstep1[0], lstep1[2]); + u[2] = _mm_unpacklo_epi32(lstep1[1], lstep1[3]); + u[3] = _mm_unpackhi_epi32(lstep1[1], lstep1[3]); + u[4] = _mm_unpacklo_epi32(lstep1[4], lstep1[6]); + u[5] = _mm_unpackhi_epi32(lstep1[4], lstep1[6]); + u[6] = _mm_unpacklo_epi32(lstep1[5], lstep1[7]); + u[7] = _mm_unpackhi_epi32(lstep1[5], lstep1[7]); + + // TODO(jingning): manually inline k_madd_epi32_ to further hide + // instruction latency. + v[0] = k_madd_epi32(u[0], k32_p16_p16); + v[1] = k_madd_epi32(u[1], k32_p16_p16); + v[2] = k_madd_epi32(u[2], k32_p16_p16); + v[3] = k_madd_epi32(u[3], k32_p16_p16); + v[4] = k_madd_epi32(u[0], k32_p16_m16); + v[5] = k_madd_epi32(u[1], k32_p16_m16); + v[6] = k_madd_epi32(u[2], k32_p16_m16); + v[7] = k_madd_epi32(u[3], k32_p16_m16); + v[8] = k_madd_epi32(u[4], k32_p24_p08); + v[9] = k_madd_epi32(u[5], k32_p24_p08); + v[10] = k_madd_epi32(u[6], k32_p24_p08); + v[11] = k_madd_epi32(u[7], k32_p24_p08); + v[12] = k_madd_epi32(u[4], k32_m08_p24); + v[13] = k_madd_epi32(u[5], k32_m08_p24); + v[14] = k_madd_epi32(u[6], k32_m08_p24); + v[15] = k_madd_epi32(u[7], k32_m08_p24); + +#if DCT_HIGH_BIT_DEPTH + overflow = k_check_epi32_overflow_16( + &v[0], &v[1], &v[2], &v[3], &v[4], &v[5], &v[6], &v[7], &v[8], + &v[9], &v[10], &v[11], &v[12], &v[13], &v[14], &v[15], &kZero); + if (overflow) { + HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + u[0] = k_packs_epi64(v[0], v[1]); + u[1] = k_packs_epi64(v[2], v[3]); + u[2] = k_packs_epi64(v[4], v[5]); + u[3] = k_packs_epi64(v[6], v[7]); + u[4] = k_packs_epi64(v[8], v[9]); + u[5] = k_packs_epi64(v[10], v[11]); + u[6] = k_packs_epi64(v[12], v[13]); + u[7] = k_packs_epi64(v[14], v[15]); + + v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING); + v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING); + v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING); + v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING); + v[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING); + v[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING); + v[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING); + v[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING); + + u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS); + u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS); + u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS); + u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS); + u[4] = _mm_srai_epi32(v[4], DCT_CONST_BITS); + u[5] = _mm_srai_epi32(v[5], DCT_CONST_BITS); + u[6] = _mm_srai_epi32(v[6], DCT_CONST_BITS); + u[7] = _mm_srai_epi32(v[7], DCT_CONST_BITS); + + sign[0] = _mm_cmplt_epi32(u[0], kZero); + sign[1] = _mm_cmplt_epi32(u[1], kZero); + sign[2] = _mm_cmplt_epi32(u[2], kZero); + sign[3] = _mm_cmplt_epi32(u[3], kZero); + sign[4] = _mm_cmplt_epi32(u[4], kZero); + sign[5] = _mm_cmplt_epi32(u[5], kZero); + sign[6] = _mm_cmplt_epi32(u[6], kZero); + sign[7] = _mm_cmplt_epi32(u[7], kZero); + + u[0] = _mm_sub_epi32(u[0], sign[0]); + u[1] = _mm_sub_epi32(u[1], sign[1]); + u[2] = _mm_sub_epi32(u[2], sign[2]); + u[3] = _mm_sub_epi32(u[3], sign[3]); + u[4] = _mm_sub_epi32(u[4], sign[4]); + u[5] = _mm_sub_epi32(u[5], sign[5]); + u[6] = _mm_sub_epi32(u[6], sign[6]); + u[7] = _mm_sub_epi32(u[7], sign[7]); + + u[0] = _mm_add_epi32(u[0], K32One); + u[1] = _mm_add_epi32(u[1], K32One); + u[2] = _mm_add_epi32(u[2], K32One); + u[3] = _mm_add_epi32(u[3], K32One); + u[4] = _mm_add_epi32(u[4], K32One); + u[5] = _mm_add_epi32(u[5], K32One); + u[6] = _mm_add_epi32(u[6], K32One); + u[7] = _mm_add_epi32(u[7], K32One); + + u[0] = _mm_srai_epi32(u[0], 2); + u[1] = _mm_srai_epi32(u[1], 2); + u[2] = _mm_srai_epi32(u[2], 2); + u[3] = _mm_srai_epi32(u[3], 2); + u[4] = _mm_srai_epi32(u[4], 2); + u[5] = _mm_srai_epi32(u[5], 2); + u[6] = _mm_srai_epi32(u[6], 2); + u[7] = _mm_srai_epi32(u[7], 2); + + // Combine + out[0] = _mm_packs_epi32(u[0], u[1]); + out[16] = _mm_packs_epi32(u[2], u[3]); + out[8] = _mm_packs_epi32(u[4], u[5]); + out[24] = _mm_packs_epi32(u[6], u[7]); +#if DCT_HIGH_BIT_DEPTH + overflow = + check_epi16_overflow_x4(&out[0], &out[16], &out[8], &out[24]); + if (overflow) { + HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + } + { + const __m128i k32_m08_p24 = pair_set_epi32(-cospi_8_64, cospi_24_64); + const __m128i k32_m24_m08 = pair_set_epi32(-cospi_24_64, -cospi_8_64); + const __m128i k32_p24_p08 = pair_set_epi32(cospi_24_64, cospi_8_64); + + u[0] = _mm_unpacklo_epi32(lstep1[18], lstep1[28]); + u[1] = _mm_unpackhi_epi32(lstep1[18], lstep1[28]); + u[2] = _mm_unpacklo_epi32(lstep1[19], lstep1[29]); + u[3] = _mm_unpackhi_epi32(lstep1[19], lstep1[29]); + u[4] = _mm_unpacklo_epi32(lstep1[20], lstep1[26]); + u[5] = _mm_unpackhi_epi32(lstep1[20], lstep1[26]); + u[6] = _mm_unpacklo_epi32(lstep1[21], lstep1[27]); + u[7] = _mm_unpackhi_epi32(lstep1[21], lstep1[27]); + + v[0] = k_madd_epi32(u[0], k32_m08_p24); + v[1] = k_madd_epi32(u[1], k32_m08_p24); + v[2] = k_madd_epi32(u[2], k32_m08_p24); + v[3] = k_madd_epi32(u[3], k32_m08_p24); + v[4] = k_madd_epi32(u[4], k32_m24_m08); + v[5] = k_madd_epi32(u[5], k32_m24_m08); + v[6] = k_madd_epi32(u[6], k32_m24_m08); + v[7] = k_madd_epi32(u[7], k32_m24_m08); + v[8] = k_madd_epi32(u[4], k32_m08_p24); + v[9] = k_madd_epi32(u[5], k32_m08_p24); + v[10] = k_madd_epi32(u[6], k32_m08_p24); + v[11] = k_madd_epi32(u[7], k32_m08_p24); + v[12] = k_madd_epi32(u[0], k32_p24_p08); + v[13] = k_madd_epi32(u[1], k32_p24_p08); + v[14] = k_madd_epi32(u[2], k32_p24_p08); + v[15] = k_madd_epi32(u[3], k32_p24_p08); + +#if DCT_HIGH_BIT_DEPTH + overflow = k_check_epi32_overflow_16( + &v[0], &v[1], &v[2], &v[3], &v[4], &v[5], &v[6], &v[7], &v[8], + &v[9], &v[10], &v[11], &v[12], &v[13], &v[14], &v[15], &kZero); + if (overflow) { + HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + u[0] = k_packs_epi64(v[0], v[1]); + u[1] = k_packs_epi64(v[2], v[3]); + u[2] = k_packs_epi64(v[4], v[5]); + u[3] = k_packs_epi64(v[6], v[7]); + u[4] = k_packs_epi64(v[8], v[9]); + u[5] = k_packs_epi64(v[10], v[11]); + u[6] = k_packs_epi64(v[12], v[13]); + u[7] = k_packs_epi64(v[14], v[15]); + + u[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING); + u[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING); + u[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING); + u[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING); + u[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING); + u[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING); + u[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING); + u[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING); + + lstep2[18] = _mm_srai_epi32(u[0], DCT_CONST_BITS); + lstep2[19] = _mm_srai_epi32(u[1], DCT_CONST_BITS); + lstep2[20] = _mm_srai_epi32(u[2], DCT_CONST_BITS); + lstep2[21] = _mm_srai_epi32(u[3], DCT_CONST_BITS); + lstep2[26] = _mm_srai_epi32(u[4], DCT_CONST_BITS); + lstep2[27] = _mm_srai_epi32(u[5], DCT_CONST_BITS); + lstep2[28] = _mm_srai_epi32(u[6], DCT_CONST_BITS); + lstep2[29] = _mm_srai_epi32(u[7], DCT_CONST_BITS); + } + { + lstep2[32] = _mm_add_epi32(lstep1[38], lstep3[32]); + lstep2[33] = _mm_add_epi32(lstep1[39], lstep3[33]); + lstep2[34] = _mm_add_epi32(lstep1[36], lstep3[34]); + lstep2[35] = _mm_add_epi32(lstep1[37], lstep3[35]); + lstep2[36] = _mm_sub_epi32(lstep3[34], lstep1[36]); + lstep2[37] = _mm_sub_epi32(lstep3[35], lstep1[37]); + lstep2[38] = _mm_sub_epi32(lstep3[32], lstep1[38]); + lstep2[39] = _mm_sub_epi32(lstep3[33], lstep1[39]); + lstep2[40] = _mm_sub_epi32(lstep3[46], lstep1[40]); + lstep2[41] = _mm_sub_epi32(lstep3[47], lstep1[41]); + lstep2[42] = _mm_sub_epi32(lstep3[44], lstep1[42]); + lstep2[43] = _mm_sub_epi32(lstep3[45], lstep1[43]); + lstep2[44] = _mm_add_epi32(lstep1[42], lstep3[44]); + lstep2[45] = _mm_add_epi32(lstep1[43], lstep3[45]); + lstep2[46] = _mm_add_epi32(lstep1[40], lstep3[46]); + lstep2[47] = _mm_add_epi32(lstep1[41], lstep3[47]); + lstep2[48] = _mm_add_epi32(lstep1[54], lstep3[48]); + lstep2[49] = _mm_add_epi32(lstep1[55], lstep3[49]); + lstep2[50] = _mm_add_epi32(lstep1[52], lstep3[50]); + lstep2[51] = _mm_add_epi32(lstep1[53], lstep3[51]); + lstep2[52] = _mm_sub_epi32(lstep3[50], lstep1[52]); + lstep2[53] = _mm_sub_epi32(lstep3[51], lstep1[53]); + lstep2[54] = _mm_sub_epi32(lstep3[48], lstep1[54]); + lstep2[55] = _mm_sub_epi32(lstep3[49], lstep1[55]); + lstep2[56] = _mm_sub_epi32(lstep3[62], lstep1[56]); + lstep2[57] = _mm_sub_epi32(lstep3[63], lstep1[57]); + lstep2[58] = _mm_sub_epi32(lstep3[60], lstep1[58]); + lstep2[59] = _mm_sub_epi32(lstep3[61], lstep1[59]); + lstep2[60] = _mm_add_epi32(lstep1[58], lstep3[60]); + lstep2[61] = _mm_add_epi32(lstep1[59], lstep3[61]); + lstep2[62] = _mm_add_epi32(lstep1[56], lstep3[62]); + lstep2[63] = _mm_add_epi32(lstep1[57], lstep3[63]); + } + // stage 6 + { + const __m128i k32_p28_p04 = pair_set_epi32(cospi_28_64, cospi_4_64); + const __m128i k32_p12_p20 = pair_set_epi32(cospi_12_64, cospi_20_64); + const __m128i k32_m20_p12 = pair_set_epi32(-cospi_20_64, cospi_12_64); + const __m128i k32_m04_p28 = pair_set_epi32(-cospi_4_64, cospi_28_64); + + u[0] = _mm_unpacklo_epi32(lstep2[8], lstep2[14]); + u[1] = _mm_unpackhi_epi32(lstep2[8], lstep2[14]); + u[2] = _mm_unpacklo_epi32(lstep2[9], lstep2[15]); + u[3] = _mm_unpackhi_epi32(lstep2[9], lstep2[15]); + u[4] = _mm_unpacklo_epi32(lstep2[10], lstep2[12]); + u[5] = _mm_unpackhi_epi32(lstep2[10], lstep2[12]); + u[6] = _mm_unpacklo_epi32(lstep2[11], lstep2[13]); + u[7] = _mm_unpackhi_epi32(lstep2[11], lstep2[13]); + u[8] = _mm_unpacklo_epi32(lstep2[10], lstep2[12]); + u[9] = _mm_unpackhi_epi32(lstep2[10], lstep2[12]); + u[10] = _mm_unpacklo_epi32(lstep2[11], lstep2[13]); + u[11] = _mm_unpackhi_epi32(lstep2[11], lstep2[13]); + u[12] = _mm_unpacklo_epi32(lstep2[8], lstep2[14]); + u[13] = _mm_unpackhi_epi32(lstep2[8], lstep2[14]); + u[14] = _mm_unpacklo_epi32(lstep2[9], lstep2[15]); + u[15] = _mm_unpackhi_epi32(lstep2[9], lstep2[15]); + + v[0] = k_madd_epi32(u[0], k32_p28_p04); + v[1] = k_madd_epi32(u[1], k32_p28_p04); + v[2] = k_madd_epi32(u[2], k32_p28_p04); + v[3] = k_madd_epi32(u[3], k32_p28_p04); + v[4] = k_madd_epi32(u[4], k32_p12_p20); + v[5] = k_madd_epi32(u[5], k32_p12_p20); + v[6] = k_madd_epi32(u[6], k32_p12_p20); + v[7] = k_madd_epi32(u[7], k32_p12_p20); + v[8] = k_madd_epi32(u[8], k32_m20_p12); + v[9] = k_madd_epi32(u[9], k32_m20_p12); + v[10] = k_madd_epi32(u[10], k32_m20_p12); + v[11] = k_madd_epi32(u[11], k32_m20_p12); + v[12] = k_madd_epi32(u[12], k32_m04_p28); + v[13] = k_madd_epi32(u[13], k32_m04_p28); + v[14] = k_madd_epi32(u[14], k32_m04_p28); + v[15] = k_madd_epi32(u[15], k32_m04_p28); + +#if DCT_HIGH_BIT_DEPTH + overflow = k_check_epi32_overflow_16( + &v[0], &v[1], &v[2], &v[3], &v[4], &v[5], &v[6], &v[7], &v[8], + &v[9], &v[10], &v[11], &v[12], &v[13], &v[14], &v[15], &kZero); + if (overflow) { + HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + u[0] = k_packs_epi64(v[0], v[1]); + u[1] = k_packs_epi64(v[2], v[3]); + u[2] = k_packs_epi64(v[4], v[5]); + u[3] = k_packs_epi64(v[6], v[7]); + u[4] = k_packs_epi64(v[8], v[9]); + u[5] = k_packs_epi64(v[10], v[11]); + u[6] = k_packs_epi64(v[12], v[13]); + u[7] = k_packs_epi64(v[14], v[15]); + + v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING); + v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING); + v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING); + v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING); + v[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING); + v[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING); + v[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING); + v[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING); + + u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS); + u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS); + u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS); + u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS); + u[4] = _mm_srai_epi32(v[4], DCT_CONST_BITS); + u[5] = _mm_srai_epi32(v[5], DCT_CONST_BITS); + u[6] = _mm_srai_epi32(v[6], DCT_CONST_BITS); + u[7] = _mm_srai_epi32(v[7], DCT_CONST_BITS); + + sign[0] = _mm_cmplt_epi32(u[0], kZero); + sign[1] = _mm_cmplt_epi32(u[1], kZero); + sign[2] = _mm_cmplt_epi32(u[2], kZero); + sign[3] = _mm_cmplt_epi32(u[3], kZero); + sign[4] = _mm_cmplt_epi32(u[4], kZero); + sign[5] = _mm_cmplt_epi32(u[5], kZero); + sign[6] = _mm_cmplt_epi32(u[6], kZero); + sign[7] = _mm_cmplt_epi32(u[7], kZero); + + u[0] = _mm_sub_epi32(u[0], sign[0]); + u[1] = _mm_sub_epi32(u[1], sign[1]); + u[2] = _mm_sub_epi32(u[2], sign[2]); + u[3] = _mm_sub_epi32(u[3], sign[3]); + u[4] = _mm_sub_epi32(u[4], sign[4]); + u[5] = _mm_sub_epi32(u[5], sign[5]); + u[6] = _mm_sub_epi32(u[6], sign[6]); + u[7] = _mm_sub_epi32(u[7], sign[7]); + + u[0] = _mm_add_epi32(u[0], K32One); + u[1] = _mm_add_epi32(u[1], K32One); + u[2] = _mm_add_epi32(u[2], K32One); + u[3] = _mm_add_epi32(u[3], K32One); + u[4] = _mm_add_epi32(u[4], K32One); + u[5] = _mm_add_epi32(u[5], K32One); + u[6] = _mm_add_epi32(u[6], K32One); + u[7] = _mm_add_epi32(u[7], K32One); + + u[0] = _mm_srai_epi32(u[0], 2); + u[1] = _mm_srai_epi32(u[1], 2); + u[2] = _mm_srai_epi32(u[2], 2); + u[3] = _mm_srai_epi32(u[3], 2); + u[4] = _mm_srai_epi32(u[4], 2); + u[5] = _mm_srai_epi32(u[5], 2); + u[6] = _mm_srai_epi32(u[6], 2); + u[7] = _mm_srai_epi32(u[7], 2); + + out[4] = _mm_packs_epi32(u[0], u[1]); + out[20] = _mm_packs_epi32(u[2], u[3]); + out[12] = _mm_packs_epi32(u[4], u[5]); + out[28] = _mm_packs_epi32(u[6], u[7]); +#if DCT_HIGH_BIT_DEPTH + overflow = + check_epi16_overflow_x4(&out[4], &out[20], &out[12], &out[28]); + if (overflow) { + HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + } + { + lstep3[16] = _mm_add_epi32(lstep2[18], lstep1[16]); + lstep3[17] = _mm_add_epi32(lstep2[19], lstep1[17]); + lstep3[18] = _mm_sub_epi32(lstep1[16], lstep2[18]); + lstep3[19] = _mm_sub_epi32(lstep1[17], lstep2[19]); + lstep3[20] = _mm_sub_epi32(lstep1[22], lstep2[20]); + lstep3[21] = _mm_sub_epi32(lstep1[23], lstep2[21]); + lstep3[22] = _mm_add_epi32(lstep2[20], lstep1[22]); + lstep3[23] = _mm_add_epi32(lstep2[21], lstep1[23]); + lstep3[24] = _mm_add_epi32(lstep2[26], lstep1[24]); + lstep3[25] = _mm_add_epi32(lstep2[27], lstep1[25]); + lstep3[26] = _mm_sub_epi32(lstep1[24], lstep2[26]); + lstep3[27] = _mm_sub_epi32(lstep1[25], lstep2[27]); + lstep3[28] = _mm_sub_epi32(lstep1[30], lstep2[28]); + lstep3[29] = _mm_sub_epi32(lstep1[31], lstep2[29]); + lstep3[30] = _mm_add_epi32(lstep2[28], lstep1[30]); + lstep3[31] = _mm_add_epi32(lstep2[29], lstep1[31]); + } + { + const __m128i k32_m04_p28 = pair_set_epi32(-cospi_4_64, cospi_28_64); + const __m128i k32_m28_m04 = pair_set_epi32(-cospi_28_64, -cospi_4_64); + const __m128i k32_m20_p12 = pair_set_epi32(-cospi_20_64, cospi_12_64); + const __m128i k32_m12_m20 = + pair_set_epi32(-cospi_12_64, -cospi_20_64); + const __m128i k32_p12_p20 = pair_set_epi32(cospi_12_64, cospi_20_64); + const __m128i k32_p28_p04 = pair_set_epi32(cospi_28_64, cospi_4_64); + + u[0] = _mm_unpacklo_epi32(lstep2[34], lstep2[60]); + u[1] = _mm_unpackhi_epi32(lstep2[34], lstep2[60]); + u[2] = _mm_unpacklo_epi32(lstep2[35], lstep2[61]); + u[3] = _mm_unpackhi_epi32(lstep2[35], lstep2[61]); + u[4] = _mm_unpacklo_epi32(lstep2[36], lstep2[58]); + u[5] = _mm_unpackhi_epi32(lstep2[36], lstep2[58]); + u[6] = _mm_unpacklo_epi32(lstep2[37], lstep2[59]); + u[7] = _mm_unpackhi_epi32(lstep2[37], lstep2[59]); + u[8] = _mm_unpacklo_epi32(lstep2[42], lstep2[52]); + u[9] = _mm_unpackhi_epi32(lstep2[42], lstep2[52]); + u[10] = _mm_unpacklo_epi32(lstep2[43], lstep2[53]); + u[11] = _mm_unpackhi_epi32(lstep2[43], lstep2[53]); + u[12] = _mm_unpacklo_epi32(lstep2[44], lstep2[50]); + u[13] = _mm_unpackhi_epi32(lstep2[44], lstep2[50]); + u[14] = _mm_unpacklo_epi32(lstep2[45], lstep2[51]); + u[15] = _mm_unpackhi_epi32(lstep2[45], lstep2[51]); + + v[0] = k_madd_epi32(u[0], k32_m04_p28); + v[1] = k_madd_epi32(u[1], k32_m04_p28); + v[2] = k_madd_epi32(u[2], k32_m04_p28); + v[3] = k_madd_epi32(u[3], k32_m04_p28); + v[4] = k_madd_epi32(u[4], k32_m28_m04); + v[5] = k_madd_epi32(u[5], k32_m28_m04); + v[6] = k_madd_epi32(u[6], k32_m28_m04); + v[7] = k_madd_epi32(u[7], k32_m28_m04); + v[8] = k_madd_epi32(u[8], k32_m20_p12); + v[9] = k_madd_epi32(u[9], k32_m20_p12); + v[10] = k_madd_epi32(u[10], k32_m20_p12); + v[11] = k_madd_epi32(u[11], k32_m20_p12); + v[12] = k_madd_epi32(u[12], k32_m12_m20); + v[13] = k_madd_epi32(u[13], k32_m12_m20); + v[14] = k_madd_epi32(u[14], k32_m12_m20); + v[15] = k_madd_epi32(u[15], k32_m12_m20); + v[16] = k_madd_epi32(u[12], k32_m20_p12); + v[17] = k_madd_epi32(u[13], k32_m20_p12); + v[18] = k_madd_epi32(u[14], k32_m20_p12); + v[19] = k_madd_epi32(u[15], k32_m20_p12); + v[20] = k_madd_epi32(u[8], k32_p12_p20); + v[21] = k_madd_epi32(u[9], k32_p12_p20); + v[22] = k_madd_epi32(u[10], k32_p12_p20); + v[23] = k_madd_epi32(u[11], k32_p12_p20); + v[24] = k_madd_epi32(u[4], k32_m04_p28); + v[25] = k_madd_epi32(u[5], k32_m04_p28); + v[26] = k_madd_epi32(u[6], k32_m04_p28); + v[27] = k_madd_epi32(u[7], k32_m04_p28); + v[28] = k_madd_epi32(u[0], k32_p28_p04); + v[29] = k_madd_epi32(u[1], k32_p28_p04); + v[30] = k_madd_epi32(u[2], k32_p28_p04); + v[31] = k_madd_epi32(u[3], k32_p28_p04); + +#if DCT_HIGH_BIT_DEPTH + overflow = k_check_epi32_overflow_32( + &v[0], &v[1], &v[2], &v[3], &v[4], &v[5], &v[6], &v[7], &v[8], + &v[9], &v[10], &v[11], &v[12], &v[13], &v[14], &v[15], &v[16], + &v[17], &v[18], &v[19], &v[20], &v[21], &v[22], &v[23], &v[24], + &v[25], &v[26], &v[27], &v[28], &v[29], &v[30], &v[31], &kZero); + if (overflow) { + HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + u[0] = k_packs_epi64(v[0], v[1]); + u[1] = k_packs_epi64(v[2], v[3]); + u[2] = k_packs_epi64(v[4], v[5]); + u[3] = k_packs_epi64(v[6], v[7]); + u[4] = k_packs_epi64(v[8], v[9]); + u[5] = k_packs_epi64(v[10], v[11]); + u[6] = k_packs_epi64(v[12], v[13]); + u[7] = k_packs_epi64(v[14], v[15]); + u[8] = k_packs_epi64(v[16], v[17]); + u[9] = k_packs_epi64(v[18], v[19]); + u[10] = k_packs_epi64(v[20], v[21]); + u[11] = k_packs_epi64(v[22], v[23]); + u[12] = k_packs_epi64(v[24], v[25]); + u[13] = k_packs_epi64(v[26], v[27]); + u[14] = k_packs_epi64(v[28], v[29]); + u[15] = k_packs_epi64(v[30], v[31]); + + v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING); + v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING); + v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING); + v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING); + v[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING); + v[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING); + v[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING); + v[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING); + v[8] = _mm_add_epi32(u[8], k__DCT_CONST_ROUNDING); + v[9] = _mm_add_epi32(u[9], k__DCT_CONST_ROUNDING); + v[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING); + v[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING); + v[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING); + v[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING); + v[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING); + v[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING); + + lstep3[34] = _mm_srai_epi32(v[0], DCT_CONST_BITS); + lstep3[35] = _mm_srai_epi32(v[1], DCT_CONST_BITS); + lstep3[36] = _mm_srai_epi32(v[2], DCT_CONST_BITS); + lstep3[37] = _mm_srai_epi32(v[3], DCT_CONST_BITS); + lstep3[42] = _mm_srai_epi32(v[4], DCT_CONST_BITS); + lstep3[43] = _mm_srai_epi32(v[5], DCT_CONST_BITS); + lstep3[44] = _mm_srai_epi32(v[6], DCT_CONST_BITS); + lstep3[45] = _mm_srai_epi32(v[7], DCT_CONST_BITS); + lstep3[50] = _mm_srai_epi32(v[8], DCT_CONST_BITS); + lstep3[51] = _mm_srai_epi32(v[9], DCT_CONST_BITS); + lstep3[52] = _mm_srai_epi32(v[10], DCT_CONST_BITS); + lstep3[53] = _mm_srai_epi32(v[11], DCT_CONST_BITS); + lstep3[58] = _mm_srai_epi32(v[12], DCT_CONST_BITS); + lstep3[59] = _mm_srai_epi32(v[13], DCT_CONST_BITS); + lstep3[60] = _mm_srai_epi32(v[14], DCT_CONST_BITS); + lstep3[61] = _mm_srai_epi32(v[15], DCT_CONST_BITS); + } + // stage 7 + { + const __m128i k32_p30_p02 = pair_set_epi32(cospi_30_64, cospi_2_64); + const __m128i k32_p14_p18 = pair_set_epi32(cospi_14_64, cospi_18_64); + const __m128i k32_p22_p10 = pair_set_epi32(cospi_22_64, cospi_10_64); + const __m128i k32_p06_p26 = pair_set_epi32(cospi_6_64, cospi_26_64); + const __m128i k32_m26_p06 = pair_set_epi32(-cospi_26_64, cospi_6_64); + const __m128i k32_m10_p22 = pair_set_epi32(-cospi_10_64, cospi_22_64); + const __m128i k32_m18_p14 = pair_set_epi32(-cospi_18_64, cospi_14_64); + const __m128i k32_m02_p30 = pair_set_epi32(-cospi_2_64, cospi_30_64); + + u[0] = _mm_unpacklo_epi32(lstep3[16], lstep3[30]); + u[1] = _mm_unpackhi_epi32(lstep3[16], lstep3[30]); + u[2] = _mm_unpacklo_epi32(lstep3[17], lstep3[31]); + u[3] = _mm_unpackhi_epi32(lstep3[17], lstep3[31]); + u[4] = _mm_unpacklo_epi32(lstep3[18], lstep3[28]); + u[5] = _mm_unpackhi_epi32(lstep3[18], lstep3[28]); + u[6] = _mm_unpacklo_epi32(lstep3[19], lstep3[29]); + u[7] = _mm_unpackhi_epi32(lstep3[19], lstep3[29]); + u[8] = _mm_unpacklo_epi32(lstep3[20], lstep3[26]); + u[9] = _mm_unpackhi_epi32(lstep3[20], lstep3[26]); + u[10] = _mm_unpacklo_epi32(lstep3[21], lstep3[27]); + u[11] = _mm_unpackhi_epi32(lstep3[21], lstep3[27]); + u[12] = _mm_unpacklo_epi32(lstep3[22], lstep3[24]); + u[13] = _mm_unpackhi_epi32(lstep3[22], lstep3[24]); + u[14] = _mm_unpacklo_epi32(lstep3[23], lstep3[25]); + u[15] = _mm_unpackhi_epi32(lstep3[23], lstep3[25]); + + v[0] = k_madd_epi32(u[0], k32_p30_p02); + v[1] = k_madd_epi32(u[1], k32_p30_p02); + v[2] = k_madd_epi32(u[2], k32_p30_p02); + v[3] = k_madd_epi32(u[3], k32_p30_p02); + v[4] = k_madd_epi32(u[4], k32_p14_p18); + v[5] = k_madd_epi32(u[5], k32_p14_p18); + v[6] = k_madd_epi32(u[6], k32_p14_p18); + v[7] = k_madd_epi32(u[7], k32_p14_p18); + v[8] = k_madd_epi32(u[8], k32_p22_p10); + v[9] = k_madd_epi32(u[9], k32_p22_p10); + v[10] = k_madd_epi32(u[10], k32_p22_p10); + v[11] = k_madd_epi32(u[11], k32_p22_p10); + v[12] = k_madd_epi32(u[12], k32_p06_p26); + v[13] = k_madd_epi32(u[13], k32_p06_p26); + v[14] = k_madd_epi32(u[14], k32_p06_p26); + v[15] = k_madd_epi32(u[15], k32_p06_p26); + v[16] = k_madd_epi32(u[12], k32_m26_p06); + v[17] = k_madd_epi32(u[13], k32_m26_p06); + v[18] = k_madd_epi32(u[14], k32_m26_p06); + v[19] = k_madd_epi32(u[15], k32_m26_p06); + v[20] = k_madd_epi32(u[8], k32_m10_p22); + v[21] = k_madd_epi32(u[9], k32_m10_p22); + v[22] = k_madd_epi32(u[10], k32_m10_p22); + v[23] = k_madd_epi32(u[11], k32_m10_p22); + v[24] = k_madd_epi32(u[4], k32_m18_p14); + v[25] = k_madd_epi32(u[5], k32_m18_p14); + v[26] = k_madd_epi32(u[6], k32_m18_p14); + v[27] = k_madd_epi32(u[7], k32_m18_p14); + v[28] = k_madd_epi32(u[0], k32_m02_p30); + v[29] = k_madd_epi32(u[1], k32_m02_p30); + v[30] = k_madd_epi32(u[2], k32_m02_p30); + v[31] = k_madd_epi32(u[3], k32_m02_p30); + +#if DCT_HIGH_BIT_DEPTH + overflow = k_check_epi32_overflow_32( + &v[0], &v[1], &v[2], &v[3], &v[4], &v[5], &v[6], &v[7], &v[8], + &v[9], &v[10], &v[11], &v[12], &v[13], &v[14], &v[15], &v[16], + &v[17], &v[18], &v[19], &v[20], &v[21], &v[22], &v[23], &v[24], + &v[25], &v[26], &v[27], &v[28], &v[29], &v[30], &v[31], &kZero); + if (overflow) { + HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + u[0] = k_packs_epi64(v[0], v[1]); + u[1] = k_packs_epi64(v[2], v[3]); + u[2] = k_packs_epi64(v[4], v[5]); + u[3] = k_packs_epi64(v[6], v[7]); + u[4] = k_packs_epi64(v[8], v[9]); + u[5] = k_packs_epi64(v[10], v[11]); + u[6] = k_packs_epi64(v[12], v[13]); + u[7] = k_packs_epi64(v[14], v[15]); + u[8] = k_packs_epi64(v[16], v[17]); + u[9] = k_packs_epi64(v[18], v[19]); + u[10] = k_packs_epi64(v[20], v[21]); + u[11] = k_packs_epi64(v[22], v[23]); + u[12] = k_packs_epi64(v[24], v[25]); + u[13] = k_packs_epi64(v[26], v[27]); + u[14] = k_packs_epi64(v[28], v[29]); + u[15] = k_packs_epi64(v[30], v[31]); + + v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING); + v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING); + v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING); + v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING); + v[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING); + v[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING); + v[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING); + v[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING); + v[8] = _mm_add_epi32(u[8], k__DCT_CONST_ROUNDING); + v[9] = _mm_add_epi32(u[9], k__DCT_CONST_ROUNDING); + v[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING); + v[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING); + v[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING); + v[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING); + v[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING); + v[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING); + + u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS); + u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS); + u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS); + u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS); + u[4] = _mm_srai_epi32(v[4], DCT_CONST_BITS); + u[5] = _mm_srai_epi32(v[5], DCT_CONST_BITS); + u[6] = _mm_srai_epi32(v[6], DCT_CONST_BITS); + u[7] = _mm_srai_epi32(v[7], DCT_CONST_BITS); + u[8] = _mm_srai_epi32(v[8], DCT_CONST_BITS); + u[9] = _mm_srai_epi32(v[9], DCT_CONST_BITS); + u[10] = _mm_srai_epi32(v[10], DCT_CONST_BITS); + u[11] = _mm_srai_epi32(v[11], DCT_CONST_BITS); + u[12] = _mm_srai_epi32(v[12], DCT_CONST_BITS); + u[13] = _mm_srai_epi32(v[13], DCT_CONST_BITS); + u[14] = _mm_srai_epi32(v[14], DCT_CONST_BITS); + u[15] = _mm_srai_epi32(v[15], DCT_CONST_BITS); + + v[0] = _mm_cmplt_epi32(u[0], kZero); + v[1] = _mm_cmplt_epi32(u[1], kZero); + v[2] = _mm_cmplt_epi32(u[2], kZero); + v[3] = _mm_cmplt_epi32(u[3], kZero); + v[4] = _mm_cmplt_epi32(u[4], kZero); + v[5] = _mm_cmplt_epi32(u[5], kZero); + v[6] = _mm_cmplt_epi32(u[6], kZero); + v[7] = _mm_cmplt_epi32(u[7], kZero); + v[8] = _mm_cmplt_epi32(u[8], kZero); + v[9] = _mm_cmplt_epi32(u[9], kZero); + v[10] = _mm_cmplt_epi32(u[10], kZero); + v[11] = _mm_cmplt_epi32(u[11], kZero); + v[12] = _mm_cmplt_epi32(u[12], kZero); + v[13] = _mm_cmplt_epi32(u[13], kZero); + v[14] = _mm_cmplt_epi32(u[14], kZero); + v[15] = _mm_cmplt_epi32(u[15], kZero); + + u[0] = _mm_sub_epi32(u[0], v[0]); + u[1] = _mm_sub_epi32(u[1], v[1]); + u[2] = _mm_sub_epi32(u[2], v[2]); + u[3] = _mm_sub_epi32(u[3], v[3]); + u[4] = _mm_sub_epi32(u[4], v[4]); + u[5] = _mm_sub_epi32(u[5], v[5]); + u[6] = _mm_sub_epi32(u[6], v[6]); + u[7] = _mm_sub_epi32(u[7], v[7]); + u[8] = _mm_sub_epi32(u[8], v[8]); + u[9] = _mm_sub_epi32(u[9], v[9]); + u[10] = _mm_sub_epi32(u[10], v[10]); + u[11] = _mm_sub_epi32(u[11], v[11]); + u[12] = _mm_sub_epi32(u[12], v[12]); + u[13] = _mm_sub_epi32(u[13], v[13]); + u[14] = _mm_sub_epi32(u[14], v[14]); + u[15] = _mm_sub_epi32(u[15], v[15]); + + v[0] = _mm_add_epi32(u[0], K32One); + v[1] = _mm_add_epi32(u[1], K32One); + v[2] = _mm_add_epi32(u[2], K32One); + v[3] = _mm_add_epi32(u[3], K32One); + v[4] = _mm_add_epi32(u[4], K32One); + v[5] = _mm_add_epi32(u[5], K32One); + v[6] = _mm_add_epi32(u[6], K32One); + v[7] = _mm_add_epi32(u[7], K32One); + v[8] = _mm_add_epi32(u[8], K32One); + v[9] = _mm_add_epi32(u[9], K32One); + v[10] = _mm_add_epi32(u[10], K32One); + v[11] = _mm_add_epi32(u[11], K32One); + v[12] = _mm_add_epi32(u[12], K32One); + v[13] = _mm_add_epi32(u[13], K32One); + v[14] = _mm_add_epi32(u[14], K32One); + v[15] = _mm_add_epi32(u[15], K32One); + + u[0] = _mm_srai_epi32(v[0], 2); + u[1] = _mm_srai_epi32(v[1], 2); + u[2] = _mm_srai_epi32(v[2], 2); + u[3] = _mm_srai_epi32(v[3], 2); + u[4] = _mm_srai_epi32(v[4], 2); + u[5] = _mm_srai_epi32(v[5], 2); + u[6] = _mm_srai_epi32(v[6], 2); + u[7] = _mm_srai_epi32(v[7], 2); + u[8] = _mm_srai_epi32(v[8], 2); + u[9] = _mm_srai_epi32(v[9], 2); + u[10] = _mm_srai_epi32(v[10], 2); + u[11] = _mm_srai_epi32(v[11], 2); + u[12] = _mm_srai_epi32(v[12], 2); + u[13] = _mm_srai_epi32(v[13], 2); + u[14] = _mm_srai_epi32(v[14], 2); + u[15] = _mm_srai_epi32(v[15], 2); + + out[2] = _mm_packs_epi32(u[0], u[1]); + out[18] = _mm_packs_epi32(u[2], u[3]); + out[10] = _mm_packs_epi32(u[4], u[5]); + out[26] = _mm_packs_epi32(u[6], u[7]); + out[6] = _mm_packs_epi32(u[8], u[9]); + out[22] = _mm_packs_epi32(u[10], u[11]); + out[14] = _mm_packs_epi32(u[12], u[13]); + out[30] = _mm_packs_epi32(u[14], u[15]); +#if DCT_HIGH_BIT_DEPTH + overflow = + check_epi16_overflow_x8(&out[2], &out[18], &out[10], &out[26], + &out[6], &out[22], &out[14], &out[30]); + if (overflow) { + HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + } + { + lstep1[32] = _mm_add_epi32(lstep3[34], lstep2[32]); + lstep1[33] = _mm_add_epi32(lstep3[35], lstep2[33]); + lstep1[34] = _mm_sub_epi32(lstep2[32], lstep3[34]); + lstep1[35] = _mm_sub_epi32(lstep2[33], lstep3[35]); + lstep1[36] = _mm_sub_epi32(lstep2[38], lstep3[36]); + lstep1[37] = _mm_sub_epi32(lstep2[39], lstep3[37]); + lstep1[38] = _mm_add_epi32(lstep3[36], lstep2[38]); + lstep1[39] = _mm_add_epi32(lstep3[37], lstep2[39]); + lstep1[40] = _mm_add_epi32(lstep3[42], lstep2[40]); + lstep1[41] = _mm_add_epi32(lstep3[43], lstep2[41]); + lstep1[42] = _mm_sub_epi32(lstep2[40], lstep3[42]); + lstep1[43] = _mm_sub_epi32(lstep2[41], lstep3[43]); + lstep1[44] = _mm_sub_epi32(lstep2[46], lstep3[44]); + lstep1[45] = _mm_sub_epi32(lstep2[47], lstep3[45]); + lstep1[46] = _mm_add_epi32(lstep3[44], lstep2[46]); + lstep1[47] = _mm_add_epi32(lstep3[45], lstep2[47]); + lstep1[48] = _mm_add_epi32(lstep3[50], lstep2[48]); + lstep1[49] = _mm_add_epi32(lstep3[51], lstep2[49]); + lstep1[50] = _mm_sub_epi32(lstep2[48], lstep3[50]); + lstep1[51] = _mm_sub_epi32(lstep2[49], lstep3[51]); + lstep1[52] = _mm_sub_epi32(lstep2[54], lstep3[52]); + lstep1[53] = _mm_sub_epi32(lstep2[55], lstep3[53]); + lstep1[54] = _mm_add_epi32(lstep3[52], lstep2[54]); + lstep1[55] = _mm_add_epi32(lstep3[53], lstep2[55]); + lstep1[56] = _mm_add_epi32(lstep3[58], lstep2[56]); + lstep1[57] = _mm_add_epi32(lstep3[59], lstep2[57]); + lstep1[58] = _mm_sub_epi32(lstep2[56], lstep3[58]); + lstep1[59] = _mm_sub_epi32(lstep2[57], lstep3[59]); + lstep1[60] = _mm_sub_epi32(lstep2[62], lstep3[60]); + lstep1[61] = _mm_sub_epi32(lstep2[63], lstep3[61]); + lstep1[62] = _mm_add_epi32(lstep3[60], lstep2[62]); + lstep1[63] = _mm_add_epi32(lstep3[61], lstep2[63]); + } + // stage 8 + { + const __m128i k32_p31_p01 = pair_set_epi32(cospi_31_64, cospi_1_64); + const __m128i k32_p15_p17 = pair_set_epi32(cospi_15_64, cospi_17_64); + const __m128i k32_p23_p09 = pair_set_epi32(cospi_23_64, cospi_9_64); + const __m128i k32_p07_p25 = pair_set_epi32(cospi_7_64, cospi_25_64); + const __m128i k32_m25_p07 = pair_set_epi32(-cospi_25_64, cospi_7_64); + const __m128i k32_m09_p23 = pair_set_epi32(-cospi_9_64, cospi_23_64); + const __m128i k32_m17_p15 = pair_set_epi32(-cospi_17_64, cospi_15_64); + const __m128i k32_m01_p31 = pair_set_epi32(-cospi_1_64, cospi_31_64); + + u[0] = _mm_unpacklo_epi32(lstep1[32], lstep1[62]); + u[1] = _mm_unpackhi_epi32(lstep1[32], lstep1[62]); + u[2] = _mm_unpacklo_epi32(lstep1[33], lstep1[63]); + u[3] = _mm_unpackhi_epi32(lstep1[33], lstep1[63]); + u[4] = _mm_unpacklo_epi32(lstep1[34], lstep1[60]); + u[5] = _mm_unpackhi_epi32(lstep1[34], lstep1[60]); + u[6] = _mm_unpacklo_epi32(lstep1[35], lstep1[61]); + u[7] = _mm_unpackhi_epi32(lstep1[35], lstep1[61]); + u[8] = _mm_unpacklo_epi32(lstep1[36], lstep1[58]); + u[9] = _mm_unpackhi_epi32(lstep1[36], lstep1[58]); + u[10] = _mm_unpacklo_epi32(lstep1[37], lstep1[59]); + u[11] = _mm_unpackhi_epi32(lstep1[37], lstep1[59]); + u[12] = _mm_unpacklo_epi32(lstep1[38], lstep1[56]); + u[13] = _mm_unpackhi_epi32(lstep1[38], lstep1[56]); + u[14] = _mm_unpacklo_epi32(lstep1[39], lstep1[57]); + u[15] = _mm_unpackhi_epi32(lstep1[39], lstep1[57]); + + v[0] = k_madd_epi32(u[0], k32_p31_p01); + v[1] = k_madd_epi32(u[1], k32_p31_p01); + v[2] = k_madd_epi32(u[2], k32_p31_p01); + v[3] = k_madd_epi32(u[3], k32_p31_p01); + v[4] = k_madd_epi32(u[4], k32_p15_p17); + v[5] = k_madd_epi32(u[5], k32_p15_p17); + v[6] = k_madd_epi32(u[6], k32_p15_p17); + v[7] = k_madd_epi32(u[7], k32_p15_p17); + v[8] = k_madd_epi32(u[8], k32_p23_p09); + v[9] = k_madd_epi32(u[9], k32_p23_p09); + v[10] = k_madd_epi32(u[10], k32_p23_p09); + v[11] = k_madd_epi32(u[11], k32_p23_p09); + v[12] = k_madd_epi32(u[12], k32_p07_p25); + v[13] = k_madd_epi32(u[13], k32_p07_p25); + v[14] = k_madd_epi32(u[14], k32_p07_p25); + v[15] = k_madd_epi32(u[15], k32_p07_p25); + v[16] = k_madd_epi32(u[12], k32_m25_p07); + v[17] = k_madd_epi32(u[13], k32_m25_p07); + v[18] = k_madd_epi32(u[14], k32_m25_p07); + v[19] = k_madd_epi32(u[15], k32_m25_p07); + v[20] = k_madd_epi32(u[8], k32_m09_p23); + v[21] = k_madd_epi32(u[9], k32_m09_p23); + v[22] = k_madd_epi32(u[10], k32_m09_p23); + v[23] = k_madd_epi32(u[11], k32_m09_p23); + v[24] = k_madd_epi32(u[4], k32_m17_p15); + v[25] = k_madd_epi32(u[5], k32_m17_p15); + v[26] = k_madd_epi32(u[6], k32_m17_p15); + v[27] = k_madd_epi32(u[7], k32_m17_p15); + v[28] = k_madd_epi32(u[0], k32_m01_p31); + v[29] = k_madd_epi32(u[1], k32_m01_p31); + v[30] = k_madd_epi32(u[2], k32_m01_p31); + v[31] = k_madd_epi32(u[3], k32_m01_p31); + +#if DCT_HIGH_BIT_DEPTH + overflow = k_check_epi32_overflow_32( + &v[0], &v[1], &v[2], &v[3], &v[4], &v[5], &v[6], &v[7], &v[8], + &v[9], &v[10], &v[11], &v[12], &v[13], &v[14], &v[15], &v[16], + &v[17], &v[18], &v[19], &v[20], &v[21], &v[22], &v[23], &v[24], + &v[25], &v[26], &v[27], &v[28], &v[29], &v[30], &v[31], &kZero); + if (overflow) { + HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + u[0] = k_packs_epi64(v[0], v[1]); + u[1] = k_packs_epi64(v[2], v[3]); + u[2] = k_packs_epi64(v[4], v[5]); + u[3] = k_packs_epi64(v[6], v[7]); + u[4] = k_packs_epi64(v[8], v[9]); + u[5] = k_packs_epi64(v[10], v[11]); + u[6] = k_packs_epi64(v[12], v[13]); + u[7] = k_packs_epi64(v[14], v[15]); + u[8] = k_packs_epi64(v[16], v[17]); + u[9] = k_packs_epi64(v[18], v[19]); + u[10] = k_packs_epi64(v[20], v[21]); + u[11] = k_packs_epi64(v[22], v[23]); + u[12] = k_packs_epi64(v[24], v[25]); + u[13] = k_packs_epi64(v[26], v[27]); + u[14] = k_packs_epi64(v[28], v[29]); + u[15] = k_packs_epi64(v[30], v[31]); + + v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING); + v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING); + v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING); + v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING); + v[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING); + v[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING); + v[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING); + v[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING); + v[8] = _mm_add_epi32(u[8], k__DCT_CONST_ROUNDING); + v[9] = _mm_add_epi32(u[9], k__DCT_CONST_ROUNDING); + v[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING); + v[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING); + v[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING); + v[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING); + v[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING); + v[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING); + + u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS); + u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS); + u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS); + u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS); + u[4] = _mm_srai_epi32(v[4], DCT_CONST_BITS); + u[5] = _mm_srai_epi32(v[5], DCT_CONST_BITS); + u[6] = _mm_srai_epi32(v[6], DCT_CONST_BITS); + u[7] = _mm_srai_epi32(v[7], DCT_CONST_BITS); + u[8] = _mm_srai_epi32(v[8], DCT_CONST_BITS); + u[9] = _mm_srai_epi32(v[9], DCT_CONST_BITS); + u[10] = _mm_srai_epi32(v[10], DCT_CONST_BITS); + u[11] = _mm_srai_epi32(v[11], DCT_CONST_BITS); + u[12] = _mm_srai_epi32(v[12], DCT_CONST_BITS); + u[13] = _mm_srai_epi32(v[13], DCT_CONST_BITS); + u[14] = _mm_srai_epi32(v[14], DCT_CONST_BITS); + u[15] = _mm_srai_epi32(v[15], DCT_CONST_BITS); + + v[0] = _mm_cmplt_epi32(u[0], kZero); + v[1] = _mm_cmplt_epi32(u[1], kZero); + v[2] = _mm_cmplt_epi32(u[2], kZero); + v[3] = _mm_cmplt_epi32(u[3], kZero); + v[4] = _mm_cmplt_epi32(u[4], kZero); + v[5] = _mm_cmplt_epi32(u[5], kZero); + v[6] = _mm_cmplt_epi32(u[6], kZero); + v[7] = _mm_cmplt_epi32(u[7], kZero); + v[8] = _mm_cmplt_epi32(u[8], kZero); + v[9] = _mm_cmplt_epi32(u[9], kZero); + v[10] = _mm_cmplt_epi32(u[10], kZero); + v[11] = _mm_cmplt_epi32(u[11], kZero); + v[12] = _mm_cmplt_epi32(u[12], kZero); + v[13] = _mm_cmplt_epi32(u[13], kZero); + v[14] = _mm_cmplt_epi32(u[14], kZero); + v[15] = _mm_cmplt_epi32(u[15], kZero); + + u[0] = _mm_sub_epi32(u[0], v[0]); + u[1] = _mm_sub_epi32(u[1], v[1]); + u[2] = _mm_sub_epi32(u[2], v[2]); + u[3] = _mm_sub_epi32(u[3], v[3]); + u[4] = _mm_sub_epi32(u[4], v[4]); + u[5] = _mm_sub_epi32(u[5], v[5]); + u[6] = _mm_sub_epi32(u[6], v[6]); + u[7] = _mm_sub_epi32(u[7], v[7]); + u[8] = _mm_sub_epi32(u[8], v[8]); + u[9] = _mm_sub_epi32(u[9], v[9]); + u[10] = _mm_sub_epi32(u[10], v[10]); + u[11] = _mm_sub_epi32(u[11], v[11]); + u[12] = _mm_sub_epi32(u[12], v[12]); + u[13] = _mm_sub_epi32(u[13], v[13]); + u[14] = _mm_sub_epi32(u[14], v[14]); + u[15] = _mm_sub_epi32(u[15], v[15]); + + v[0] = _mm_add_epi32(u[0], K32One); + v[1] = _mm_add_epi32(u[1], K32One); + v[2] = _mm_add_epi32(u[2], K32One); + v[3] = _mm_add_epi32(u[3], K32One); + v[4] = _mm_add_epi32(u[4], K32One); + v[5] = _mm_add_epi32(u[5], K32One); + v[6] = _mm_add_epi32(u[6], K32One); + v[7] = _mm_add_epi32(u[7], K32One); + v[8] = _mm_add_epi32(u[8], K32One); + v[9] = _mm_add_epi32(u[9], K32One); + v[10] = _mm_add_epi32(u[10], K32One); + v[11] = _mm_add_epi32(u[11], K32One); + v[12] = _mm_add_epi32(u[12], K32One); + v[13] = _mm_add_epi32(u[13], K32One); + v[14] = _mm_add_epi32(u[14], K32One); + v[15] = _mm_add_epi32(u[15], K32One); + + u[0] = _mm_srai_epi32(v[0], 2); + u[1] = _mm_srai_epi32(v[1], 2); + u[2] = _mm_srai_epi32(v[2], 2); + u[3] = _mm_srai_epi32(v[3], 2); + u[4] = _mm_srai_epi32(v[4], 2); + u[5] = _mm_srai_epi32(v[5], 2); + u[6] = _mm_srai_epi32(v[6], 2); + u[7] = _mm_srai_epi32(v[7], 2); + u[8] = _mm_srai_epi32(v[8], 2); + u[9] = _mm_srai_epi32(v[9], 2); + u[10] = _mm_srai_epi32(v[10], 2); + u[11] = _mm_srai_epi32(v[11], 2); + u[12] = _mm_srai_epi32(v[12], 2); + u[13] = _mm_srai_epi32(v[13], 2); + u[14] = _mm_srai_epi32(v[14], 2); + u[15] = _mm_srai_epi32(v[15], 2); + + out[1] = _mm_packs_epi32(u[0], u[1]); + out[17] = _mm_packs_epi32(u[2], u[3]); + out[9] = _mm_packs_epi32(u[4], u[5]); + out[25] = _mm_packs_epi32(u[6], u[7]); + out[7] = _mm_packs_epi32(u[8], u[9]); + out[23] = _mm_packs_epi32(u[10], u[11]); + out[15] = _mm_packs_epi32(u[12], u[13]); + out[31] = _mm_packs_epi32(u[14], u[15]); +#if DCT_HIGH_BIT_DEPTH + overflow = + check_epi16_overflow_x8(&out[1], &out[17], &out[9], &out[25], + &out[7], &out[23], &out[15], &out[31]); + if (overflow) { + HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + } + { + const __m128i k32_p27_p05 = pair_set_epi32(cospi_27_64, cospi_5_64); + const __m128i k32_p11_p21 = pair_set_epi32(cospi_11_64, cospi_21_64); + const __m128i k32_p19_p13 = pair_set_epi32(cospi_19_64, cospi_13_64); + const __m128i k32_p03_p29 = pair_set_epi32(cospi_3_64, cospi_29_64); + const __m128i k32_m29_p03 = pair_set_epi32(-cospi_29_64, cospi_3_64); + const __m128i k32_m13_p19 = pair_set_epi32(-cospi_13_64, cospi_19_64); + const __m128i k32_m21_p11 = pair_set_epi32(-cospi_21_64, cospi_11_64); + const __m128i k32_m05_p27 = pair_set_epi32(-cospi_5_64, cospi_27_64); + + u[0] = _mm_unpacklo_epi32(lstep1[40], lstep1[54]); + u[1] = _mm_unpackhi_epi32(lstep1[40], lstep1[54]); + u[2] = _mm_unpacklo_epi32(lstep1[41], lstep1[55]); + u[3] = _mm_unpackhi_epi32(lstep1[41], lstep1[55]); + u[4] = _mm_unpacklo_epi32(lstep1[42], lstep1[52]); + u[5] = _mm_unpackhi_epi32(lstep1[42], lstep1[52]); + u[6] = _mm_unpacklo_epi32(lstep1[43], lstep1[53]); + u[7] = _mm_unpackhi_epi32(lstep1[43], lstep1[53]); + u[8] = _mm_unpacklo_epi32(lstep1[44], lstep1[50]); + u[9] = _mm_unpackhi_epi32(lstep1[44], lstep1[50]); + u[10] = _mm_unpacklo_epi32(lstep1[45], lstep1[51]); + u[11] = _mm_unpackhi_epi32(lstep1[45], lstep1[51]); + u[12] = _mm_unpacklo_epi32(lstep1[46], lstep1[48]); + u[13] = _mm_unpackhi_epi32(lstep1[46], lstep1[48]); + u[14] = _mm_unpacklo_epi32(lstep1[47], lstep1[49]); + u[15] = _mm_unpackhi_epi32(lstep1[47], lstep1[49]); + + v[0] = k_madd_epi32(u[0], k32_p27_p05); + v[1] = k_madd_epi32(u[1], k32_p27_p05); + v[2] = k_madd_epi32(u[2], k32_p27_p05); + v[3] = k_madd_epi32(u[3], k32_p27_p05); + v[4] = k_madd_epi32(u[4], k32_p11_p21); + v[5] = k_madd_epi32(u[5], k32_p11_p21); + v[6] = k_madd_epi32(u[6], k32_p11_p21); + v[7] = k_madd_epi32(u[7], k32_p11_p21); + v[8] = k_madd_epi32(u[8], k32_p19_p13); + v[9] = k_madd_epi32(u[9], k32_p19_p13); + v[10] = k_madd_epi32(u[10], k32_p19_p13); + v[11] = k_madd_epi32(u[11], k32_p19_p13); + v[12] = k_madd_epi32(u[12], k32_p03_p29); + v[13] = k_madd_epi32(u[13], k32_p03_p29); + v[14] = k_madd_epi32(u[14], k32_p03_p29); + v[15] = k_madd_epi32(u[15], k32_p03_p29); + v[16] = k_madd_epi32(u[12], k32_m29_p03); + v[17] = k_madd_epi32(u[13], k32_m29_p03); + v[18] = k_madd_epi32(u[14], k32_m29_p03); + v[19] = k_madd_epi32(u[15], k32_m29_p03); + v[20] = k_madd_epi32(u[8], k32_m13_p19); + v[21] = k_madd_epi32(u[9], k32_m13_p19); + v[22] = k_madd_epi32(u[10], k32_m13_p19); + v[23] = k_madd_epi32(u[11], k32_m13_p19); + v[24] = k_madd_epi32(u[4], k32_m21_p11); + v[25] = k_madd_epi32(u[5], k32_m21_p11); + v[26] = k_madd_epi32(u[6], k32_m21_p11); + v[27] = k_madd_epi32(u[7], k32_m21_p11); + v[28] = k_madd_epi32(u[0], k32_m05_p27); + v[29] = k_madd_epi32(u[1], k32_m05_p27); + v[30] = k_madd_epi32(u[2], k32_m05_p27); + v[31] = k_madd_epi32(u[3], k32_m05_p27); + +#if DCT_HIGH_BIT_DEPTH + overflow = k_check_epi32_overflow_32( + &v[0], &v[1], &v[2], &v[3], &v[4], &v[5], &v[6], &v[7], &v[8], + &v[9], &v[10], &v[11], &v[12], &v[13], &v[14], &v[15], &v[16], + &v[17], &v[18], &v[19], &v[20], &v[21], &v[22], &v[23], &v[24], + &v[25], &v[26], &v[27], &v[28], &v[29], &v[30], &v[31], &kZero); + if (overflow) { + HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + u[0] = k_packs_epi64(v[0], v[1]); + u[1] = k_packs_epi64(v[2], v[3]); + u[2] = k_packs_epi64(v[4], v[5]); + u[3] = k_packs_epi64(v[6], v[7]); + u[4] = k_packs_epi64(v[8], v[9]); + u[5] = k_packs_epi64(v[10], v[11]); + u[6] = k_packs_epi64(v[12], v[13]); + u[7] = k_packs_epi64(v[14], v[15]); + u[8] = k_packs_epi64(v[16], v[17]); + u[9] = k_packs_epi64(v[18], v[19]); + u[10] = k_packs_epi64(v[20], v[21]); + u[11] = k_packs_epi64(v[22], v[23]); + u[12] = k_packs_epi64(v[24], v[25]); + u[13] = k_packs_epi64(v[26], v[27]); + u[14] = k_packs_epi64(v[28], v[29]); + u[15] = k_packs_epi64(v[30], v[31]); + + v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING); + v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING); + v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING); + v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING); + v[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING); + v[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING); + v[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING); + v[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING); + v[8] = _mm_add_epi32(u[8], k__DCT_CONST_ROUNDING); + v[9] = _mm_add_epi32(u[9], k__DCT_CONST_ROUNDING); + v[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING); + v[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING); + v[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING); + v[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING); + v[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING); + v[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING); + + u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS); + u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS); + u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS); + u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS); + u[4] = _mm_srai_epi32(v[4], DCT_CONST_BITS); + u[5] = _mm_srai_epi32(v[5], DCT_CONST_BITS); + u[6] = _mm_srai_epi32(v[6], DCT_CONST_BITS); + u[7] = _mm_srai_epi32(v[7], DCT_CONST_BITS); + u[8] = _mm_srai_epi32(v[8], DCT_CONST_BITS); + u[9] = _mm_srai_epi32(v[9], DCT_CONST_BITS); + u[10] = _mm_srai_epi32(v[10], DCT_CONST_BITS); + u[11] = _mm_srai_epi32(v[11], DCT_CONST_BITS); + u[12] = _mm_srai_epi32(v[12], DCT_CONST_BITS); + u[13] = _mm_srai_epi32(v[13], DCT_CONST_BITS); + u[14] = _mm_srai_epi32(v[14], DCT_CONST_BITS); + u[15] = _mm_srai_epi32(v[15], DCT_CONST_BITS); + + v[0] = _mm_cmplt_epi32(u[0], kZero); + v[1] = _mm_cmplt_epi32(u[1], kZero); + v[2] = _mm_cmplt_epi32(u[2], kZero); + v[3] = _mm_cmplt_epi32(u[3], kZero); + v[4] = _mm_cmplt_epi32(u[4], kZero); + v[5] = _mm_cmplt_epi32(u[5], kZero); + v[6] = _mm_cmplt_epi32(u[6], kZero); + v[7] = _mm_cmplt_epi32(u[7], kZero); + v[8] = _mm_cmplt_epi32(u[8], kZero); + v[9] = _mm_cmplt_epi32(u[9], kZero); + v[10] = _mm_cmplt_epi32(u[10], kZero); + v[11] = _mm_cmplt_epi32(u[11], kZero); + v[12] = _mm_cmplt_epi32(u[12], kZero); + v[13] = _mm_cmplt_epi32(u[13], kZero); + v[14] = _mm_cmplt_epi32(u[14], kZero); + v[15] = _mm_cmplt_epi32(u[15], kZero); + + u[0] = _mm_sub_epi32(u[0], v[0]); + u[1] = _mm_sub_epi32(u[1], v[1]); + u[2] = _mm_sub_epi32(u[2], v[2]); + u[3] = _mm_sub_epi32(u[3], v[3]); + u[4] = _mm_sub_epi32(u[4], v[4]); + u[5] = _mm_sub_epi32(u[5], v[5]); + u[6] = _mm_sub_epi32(u[6], v[6]); + u[7] = _mm_sub_epi32(u[7], v[7]); + u[8] = _mm_sub_epi32(u[8], v[8]); + u[9] = _mm_sub_epi32(u[9], v[9]); + u[10] = _mm_sub_epi32(u[10], v[10]); + u[11] = _mm_sub_epi32(u[11], v[11]); + u[12] = _mm_sub_epi32(u[12], v[12]); + u[13] = _mm_sub_epi32(u[13], v[13]); + u[14] = _mm_sub_epi32(u[14], v[14]); + u[15] = _mm_sub_epi32(u[15], v[15]); + + v[0] = _mm_add_epi32(u[0], K32One); + v[1] = _mm_add_epi32(u[1], K32One); + v[2] = _mm_add_epi32(u[2], K32One); + v[3] = _mm_add_epi32(u[3], K32One); + v[4] = _mm_add_epi32(u[4], K32One); + v[5] = _mm_add_epi32(u[5], K32One); + v[6] = _mm_add_epi32(u[6], K32One); + v[7] = _mm_add_epi32(u[7], K32One); + v[8] = _mm_add_epi32(u[8], K32One); + v[9] = _mm_add_epi32(u[9], K32One); + v[10] = _mm_add_epi32(u[10], K32One); + v[11] = _mm_add_epi32(u[11], K32One); + v[12] = _mm_add_epi32(u[12], K32One); + v[13] = _mm_add_epi32(u[13], K32One); + v[14] = _mm_add_epi32(u[14], K32One); + v[15] = _mm_add_epi32(u[15], K32One); + + u[0] = _mm_srai_epi32(v[0], 2); + u[1] = _mm_srai_epi32(v[1], 2); + u[2] = _mm_srai_epi32(v[2], 2); + u[3] = _mm_srai_epi32(v[3], 2); + u[4] = _mm_srai_epi32(v[4], 2); + u[5] = _mm_srai_epi32(v[5], 2); + u[6] = _mm_srai_epi32(v[6], 2); + u[7] = _mm_srai_epi32(v[7], 2); + u[8] = _mm_srai_epi32(v[8], 2); + u[9] = _mm_srai_epi32(v[9], 2); + u[10] = _mm_srai_epi32(v[10], 2); + u[11] = _mm_srai_epi32(v[11], 2); + u[12] = _mm_srai_epi32(v[12], 2); + u[13] = _mm_srai_epi32(v[13], 2); + u[14] = _mm_srai_epi32(v[14], 2); + u[15] = _mm_srai_epi32(v[15], 2); + + out[5] = _mm_packs_epi32(u[0], u[1]); + out[21] = _mm_packs_epi32(u[2], u[3]); + out[13] = _mm_packs_epi32(u[4], u[5]); + out[29] = _mm_packs_epi32(u[6], u[7]); + out[3] = _mm_packs_epi32(u[8], u[9]); + out[19] = _mm_packs_epi32(u[10], u[11]); + out[11] = _mm_packs_epi32(u[12], u[13]); + out[27] = _mm_packs_epi32(u[14], u[15]); +#if DCT_HIGH_BIT_DEPTH + overflow = + check_epi16_overflow_x8(&out[5], &out[21], &out[13], &out[29], + &out[3], &out[19], &out[11], &out[27]); + if (overflow) { + HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + } + } +#endif // FDCT32x32_HIGH_PRECISION + // Transpose the results, do it as four 8x8 transposes. + { + int transpose_block; + int16_t *output0 = &intermediate[column_start * 32]; + tran_low_t *output1 = &output_org[column_start * 32]; + for (transpose_block = 0; transpose_block < 4; ++transpose_block) { + __m128i *this_out = &out[8 * transpose_block]; + // 00 01 02 03 04 05 06 07 + // 10 11 12 13 14 15 16 17 + // 20 21 22 23 24 25 26 27 + // 30 31 32 33 34 35 36 37 + // 40 41 42 43 44 45 46 47 + // 50 51 52 53 54 55 56 57 + // 60 61 62 63 64 65 66 67 + // 70 71 72 73 74 75 76 77 + const __m128i tr0_0 = _mm_unpacklo_epi16(this_out[0], this_out[1]); + const __m128i tr0_1 = _mm_unpacklo_epi16(this_out[2], this_out[3]); + const __m128i tr0_2 = _mm_unpackhi_epi16(this_out[0], this_out[1]); + const __m128i tr0_3 = _mm_unpackhi_epi16(this_out[2], this_out[3]); + const __m128i tr0_4 = _mm_unpacklo_epi16(this_out[4], this_out[5]); + const __m128i tr0_5 = _mm_unpacklo_epi16(this_out[6], this_out[7]); + const __m128i tr0_6 = _mm_unpackhi_epi16(this_out[4], this_out[5]); + const __m128i tr0_7 = _mm_unpackhi_epi16(this_out[6], this_out[7]); + // 00 10 01 11 02 12 03 13 + // 20 30 21 31 22 32 23 33 + // 04 14 05 15 06 16 07 17 + // 24 34 25 35 26 36 27 37 + // 40 50 41 51 42 52 43 53 + // 60 70 61 71 62 72 63 73 + // 54 54 55 55 56 56 57 57 + // 64 74 65 75 66 76 67 77 + const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1); + const __m128i tr1_1 = _mm_unpacklo_epi32(tr0_2, tr0_3); + const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1); + const __m128i tr1_3 = _mm_unpackhi_epi32(tr0_2, tr0_3); + const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_4, tr0_5); + const __m128i tr1_5 = _mm_unpacklo_epi32(tr0_6, tr0_7); + const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_4, tr0_5); + const __m128i tr1_7 = _mm_unpackhi_epi32(tr0_6, tr0_7); + // 00 10 20 30 01 11 21 31 + // 40 50 60 70 41 51 61 71 + // 02 12 22 32 03 13 23 33 + // 42 52 62 72 43 53 63 73 + // 04 14 24 34 05 15 21 36 + // 44 54 64 74 45 55 61 76 + // 06 16 26 36 07 17 27 37 + // 46 56 66 76 47 57 67 77 + __m128i tr2_0 = _mm_unpacklo_epi64(tr1_0, tr1_4); + __m128i tr2_1 = _mm_unpackhi_epi64(tr1_0, tr1_4); + __m128i tr2_2 = _mm_unpacklo_epi64(tr1_2, tr1_6); + __m128i tr2_3 = _mm_unpackhi_epi64(tr1_2, tr1_6); + __m128i tr2_4 = _mm_unpacklo_epi64(tr1_1, tr1_5); + __m128i tr2_5 = _mm_unpackhi_epi64(tr1_1, tr1_5); + __m128i tr2_6 = _mm_unpacklo_epi64(tr1_3, tr1_7); + __m128i tr2_7 = _mm_unpackhi_epi64(tr1_3, tr1_7); + // 00 10 20 30 40 50 60 70 + // 01 11 21 31 41 51 61 71 + // 02 12 22 32 42 52 62 72 + // 03 13 23 33 43 53 63 73 + // 04 14 24 34 44 54 64 74 + // 05 15 25 35 45 55 65 75 + // 06 16 26 36 46 56 66 76 + // 07 17 27 37 47 57 67 77 + if (0 == pass) { + // output[j] = (output[j] + 1 + (output[j] > 0)) >> 2; + // TODO(cd): see quality impact of only doing + // output[j] = (output[j] + 1) >> 2; + // which would remove the code between here ... + __m128i tr2_0_0 = _mm_cmpgt_epi16(tr2_0, kZero); + __m128i tr2_1_0 = _mm_cmpgt_epi16(tr2_1, kZero); + __m128i tr2_2_0 = _mm_cmpgt_epi16(tr2_2, kZero); + __m128i tr2_3_0 = _mm_cmpgt_epi16(tr2_3, kZero); + __m128i tr2_4_0 = _mm_cmpgt_epi16(tr2_4, kZero); + __m128i tr2_5_0 = _mm_cmpgt_epi16(tr2_5, kZero); + __m128i tr2_6_0 = _mm_cmpgt_epi16(tr2_6, kZero); + __m128i tr2_7_0 = _mm_cmpgt_epi16(tr2_7, kZero); + tr2_0 = _mm_sub_epi16(tr2_0, tr2_0_0); + tr2_1 = _mm_sub_epi16(tr2_1, tr2_1_0); + tr2_2 = _mm_sub_epi16(tr2_2, tr2_2_0); + tr2_3 = _mm_sub_epi16(tr2_3, tr2_3_0); + tr2_4 = _mm_sub_epi16(tr2_4, tr2_4_0); + tr2_5 = _mm_sub_epi16(tr2_5, tr2_5_0); + tr2_6 = _mm_sub_epi16(tr2_6, tr2_6_0); + tr2_7 = _mm_sub_epi16(tr2_7, tr2_7_0); + // ... and here. + // PS: also change code in av1/encoder/av1_dct.c + tr2_0 = _mm_add_epi16(tr2_0, kOne); + tr2_1 = _mm_add_epi16(tr2_1, kOne); + tr2_2 = _mm_add_epi16(tr2_2, kOne); + tr2_3 = _mm_add_epi16(tr2_3, kOne); + tr2_4 = _mm_add_epi16(tr2_4, kOne); + tr2_5 = _mm_add_epi16(tr2_5, kOne); + tr2_6 = _mm_add_epi16(tr2_6, kOne); + tr2_7 = _mm_add_epi16(tr2_7, kOne); + tr2_0 = _mm_srai_epi16(tr2_0, 2); + tr2_1 = _mm_srai_epi16(tr2_1, 2); + tr2_2 = _mm_srai_epi16(tr2_2, 2); + tr2_3 = _mm_srai_epi16(tr2_3, 2); + tr2_4 = _mm_srai_epi16(tr2_4, 2); + tr2_5 = _mm_srai_epi16(tr2_5, 2); + tr2_6 = _mm_srai_epi16(tr2_6, 2); + tr2_7 = _mm_srai_epi16(tr2_7, 2); + } + // Note: even though all these stores are aligned, using the aligned + // intrinsic make the code slightly slower. + if (pass == 0) { + _mm_storeu_si128((__m128i *)(output0 + 0 * 32), tr2_0); + _mm_storeu_si128((__m128i *)(output0 + 1 * 32), tr2_1); + _mm_storeu_si128((__m128i *)(output0 + 2 * 32), tr2_2); + _mm_storeu_si128((__m128i *)(output0 + 3 * 32), tr2_3); + _mm_storeu_si128((__m128i *)(output0 + 4 * 32), tr2_4); + _mm_storeu_si128((__m128i *)(output0 + 5 * 32), tr2_5); + _mm_storeu_si128((__m128i *)(output0 + 6 * 32), tr2_6); + _mm_storeu_si128((__m128i *)(output0 + 7 * 32), tr2_7); + // Process next 8x8 + output0 += 8; + } else { + storeu_output(&tr2_0, (output1 + 0 * 32)); + storeu_output(&tr2_1, (output1 + 1 * 32)); + storeu_output(&tr2_2, (output1 + 2 * 32)); + storeu_output(&tr2_3, (output1 + 3 * 32)); + storeu_output(&tr2_4, (output1 + 4 * 32)); + storeu_output(&tr2_5, (output1 + 5 * 32)); + storeu_output(&tr2_6, (output1 + 6 * 32)); + storeu_output(&tr2_7, (output1 + 7 * 32)); + // Process next 8x8 + output1 += 8; + } + } + } + } + } +} // NOLINT + +#undef ADD_EPI16 +#undef SUB_EPI16 +#undef HIGH_FDCT32x32_2D_C +#undef HIGH_FDCT32x32_2D_ROWS_C diff --git a/third_party/aom/aom_dsp/x86/fwd_txfm_avx2.c b/third_party/aom/aom_dsp/x86/fwd_txfm_avx2.c new file mode 100644 index 000000000..670f864d0 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/fwd_txfm_avx2.c @@ -0,0 +1,24 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include "./aom_config.h" + +#define FDCT32x32_2D_AVX2 aom_fdct32x32_rd_avx2 +#define FDCT32x32_HIGH_PRECISION 0 +#include "aom_dsp/x86/fwd_dct32x32_impl_avx2.h" +#undef FDCT32x32_2D_AVX2 +#undef FDCT32x32_HIGH_PRECISION + +#define FDCT32x32_2D_AVX2 aom_fdct32x32_avx2 +#define FDCT32x32_HIGH_PRECISION 1 +#include "aom_dsp/x86/fwd_dct32x32_impl_avx2.h" // NOLINT +#undef FDCT32x32_2D_AVX2 +#undef FDCT32x32_HIGH_PRECISION diff --git a/third_party/aom/aom_dsp/x86/fwd_txfm_avx2.h b/third_party/aom/aom_dsp/x86/fwd_txfm_avx2.h new file mode 100644 index 000000000..d3aceae00 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/fwd_txfm_avx2.h @@ -0,0 +1,35 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#ifndef AOM_DSP_X86_FWD_TXFM_AVX2_H +#define AOM_DSP_X86_FWD_TXFM_AVX2_H + +#include "./aom_config.h" + +static INLINE void storeu_output_avx2(const __m256i *coeff, tran_low_t *out) { +#if CONFIG_HIGHBITDEPTH + const __m256i zero = _mm256_setzero_si256(); + const __m256i sign = _mm256_cmpgt_epi16(zero, *coeff); + + __m256i x0 = _mm256_unpacklo_epi16(*coeff, sign); + __m256i x1 = _mm256_unpackhi_epi16(*coeff, sign); + + __m256i y0 = _mm256_permute2x128_si256(x0, x1, 0x20); + __m256i y1 = _mm256_permute2x128_si256(x0, x1, 0x31); + + _mm256_storeu_si256((__m256i *)out, y0); + _mm256_storeu_si256((__m256i *)(out + 8), y1); +#else + _mm256_storeu_si256((__m256i *)out, *coeff); +#endif +} + +#endif // AOM_DSP_X86_FWD_TXFM_AVX2_H diff --git a/third_party/aom/aom_dsp/x86/fwd_txfm_impl_sse2.h b/third_party/aom/aom_dsp/x86/fwd_txfm_impl_sse2.h new file mode 100644 index 000000000..7bb1db70a --- /dev/null +++ b/third_party/aom/aom_dsp/x86/fwd_txfm_impl_sse2.h @@ -0,0 +1,1014 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include <emmintrin.h> // SSE2 + +#include "./aom_dsp_rtcd.h" +#include "aom_dsp/txfm_common.h" +#include "aom_dsp/x86/fwd_txfm_sse2.h" +#include "aom_dsp/x86/txfm_common_sse2.h" +#include "aom_ports/mem.h" + +// TODO(jingning) The high bit-depth functions need rework for performance. +// After we properly fix the high bit-depth function implementations, this +// file's dependency should be substantially simplified. +#if DCT_HIGH_BIT_DEPTH +#define ADD_EPI16 _mm_adds_epi16 +#define SUB_EPI16 _mm_subs_epi16 + +#else +#define ADD_EPI16 _mm_add_epi16 +#define SUB_EPI16 _mm_sub_epi16 +#endif + +void FDCT4x4_2D(const int16_t *input, tran_low_t *output, int stride) { + // This 2D transform implements 4 vertical 1D transforms followed + // by 4 horizontal 1D transforms. The multiplies and adds are as given + // by Chen, Smith and Fralick ('77). The commands for moving the data + // around have been minimized by hand. + // For the purposes of the comments, the 16 inputs are referred to at i0 + // through iF (in raster order), intermediate variables are a0, b0, c0 + // through f, and correspond to the in-place computations mapped to input + // locations. The outputs, o0 through oF are labeled according to the + // output locations. + + // Constants + // These are the coefficients used for the multiplies. + // In the comments, pN means cos(N pi /64) and mN is -cos(N pi /64), + // where cospi_N_64 = cos(N pi /64) + const __m128i k__cospi_A = + octa_set_epi16(cospi_16_64, cospi_16_64, cospi_16_64, cospi_16_64, + cospi_16_64, -cospi_16_64, cospi_16_64, -cospi_16_64); + const __m128i k__cospi_B = + octa_set_epi16(cospi_16_64, -cospi_16_64, cospi_16_64, -cospi_16_64, + cospi_16_64, cospi_16_64, cospi_16_64, cospi_16_64); + const __m128i k__cospi_C = + octa_set_epi16(cospi_8_64, cospi_24_64, cospi_8_64, cospi_24_64, + cospi_24_64, -cospi_8_64, cospi_24_64, -cospi_8_64); + const __m128i k__cospi_D = + octa_set_epi16(cospi_24_64, -cospi_8_64, cospi_24_64, -cospi_8_64, + cospi_8_64, cospi_24_64, cospi_8_64, cospi_24_64); + const __m128i k__cospi_E = + octa_set_epi16(cospi_16_64, cospi_16_64, cospi_16_64, cospi_16_64, + cospi_16_64, cospi_16_64, cospi_16_64, cospi_16_64); + const __m128i k__cospi_F = + octa_set_epi16(cospi_16_64, -cospi_16_64, cospi_16_64, -cospi_16_64, + cospi_16_64, -cospi_16_64, cospi_16_64, -cospi_16_64); + const __m128i k__cospi_G = + octa_set_epi16(cospi_8_64, cospi_24_64, cospi_8_64, cospi_24_64, + -cospi_8_64, -cospi_24_64, -cospi_8_64, -cospi_24_64); + const __m128i k__cospi_H = + octa_set_epi16(cospi_24_64, -cospi_8_64, cospi_24_64, -cospi_8_64, + -cospi_24_64, cospi_8_64, -cospi_24_64, cospi_8_64); + + const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING); + // This second rounding constant saves doing some extra adds at the end + const __m128i k__DCT_CONST_ROUNDING2 = + _mm_set1_epi32(DCT_CONST_ROUNDING + (DCT_CONST_ROUNDING << 1)); + const int DCT_CONST_BITS2 = DCT_CONST_BITS + 2; + const __m128i k__nonzero_bias_a = _mm_setr_epi16(0, 1, 1, 1, 1, 1, 1, 1); + const __m128i k__nonzero_bias_b = _mm_setr_epi16(1, 0, 0, 0, 0, 0, 0, 0); + __m128i in0, in1; +#if DCT_HIGH_BIT_DEPTH + __m128i cmp0, cmp1; + int test, overflow; +#endif + + // Load inputs. + in0 = _mm_loadl_epi64((const __m128i *)(input + 0 * stride)); + in1 = _mm_loadl_epi64((const __m128i *)(input + 1 * stride)); + // in0 = [i0 i1 i2 i3 iC iD iE iF] + // in1 = [i4 i5 i6 i7 i8 i9 iA iB] + in1 = _mm_unpacklo_epi64( + in1, _mm_loadl_epi64((const __m128i *)(input + 2 * stride))); + in0 = _mm_unpacklo_epi64( + in0, _mm_loadl_epi64((const __m128i *)(input + 3 * stride))); +#if DCT_HIGH_BIT_DEPTH + // Check inputs small enough to use optimised code + cmp0 = _mm_xor_si128(_mm_cmpgt_epi16(in0, _mm_set1_epi16(0x3ff)), + _mm_cmplt_epi16(in0, _mm_set1_epi16(0xfc00))); + cmp1 = _mm_xor_si128(_mm_cmpgt_epi16(in1, _mm_set1_epi16(0x3ff)), + _mm_cmplt_epi16(in1, _mm_set1_epi16(0xfc00))); + test = _mm_movemask_epi8(_mm_or_si128(cmp0, cmp1)); + if (test) { + aom_highbd_fdct4x4_c(input, output, stride); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + + // multiply by 16 to give some extra precision + in0 = _mm_slli_epi16(in0, 4); + in1 = _mm_slli_epi16(in1, 4); + // if (i == 0 && input[0]) input[0] += 1; + // add 1 to the upper left pixel if it is non-zero, which helps reduce + // the round-trip error + { + // The mask will only contain whether the first value is zero, all + // other comparison will fail as something shifted by 4 (above << 4) + // can never be equal to one. To increment in the non-zero case, we + // add the mask and one for the first element: + // - if zero, mask = -1, v = v - 1 + 1 = v + // - if non-zero, mask = 0, v = v + 0 + 1 = v + 1 + __m128i mask = _mm_cmpeq_epi16(in0, k__nonzero_bias_a); + in0 = _mm_add_epi16(in0, mask); + in0 = _mm_add_epi16(in0, k__nonzero_bias_b); + } + // There are 4 total stages, alternating between an add/subtract stage + // followed by an multiply-and-add stage. + { + // Stage 1: Add/subtract + + // in0 = [i0 i1 i2 i3 iC iD iE iF] + // in1 = [i4 i5 i6 i7 i8 i9 iA iB] + const __m128i r0 = _mm_unpacklo_epi16(in0, in1); + const __m128i r1 = _mm_unpackhi_epi16(in0, in1); + // r0 = [i0 i4 i1 i5 i2 i6 i3 i7] + // r1 = [iC i8 iD i9 iE iA iF iB] + const __m128i r2 = _mm_shuffle_epi32(r0, 0xB4); + const __m128i r3 = _mm_shuffle_epi32(r1, 0xB4); + // r2 = [i0 i4 i1 i5 i3 i7 i2 i6] + // r3 = [iC i8 iD i9 iF iB iE iA] + + const __m128i t0 = _mm_add_epi16(r2, r3); + const __m128i t1 = _mm_sub_epi16(r2, r3); + // t0 = [a0 a4 a1 a5 a3 a7 a2 a6] + // t1 = [aC a8 aD a9 aF aB aE aA] + + // Stage 2: multiply by constants (which gets us into 32 bits). + // The constants needed here are: + // k__cospi_A = [p16 p16 p16 p16 p16 m16 p16 m16] + // k__cospi_B = [p16 m16 p16 m16 p16 p16 p16 p16] + // k__cospi_C = [p08 p24 p08 p24 p24 m08 p24 m08] + // k__cospi_D = [p24 m08 p24 m08 p08 p24 p08 p24] + const __m128i u0 = _mm_madd_epi16(t0, k__cospi_A); + const __m128i u2 = _mm_madd_epi16(t0, k__cospi_B); + const __m128i u1 = _mm_madd_epi16(t1, k__cospi_C); + const __m128i u3 = _mm_madd_epi16(t1, k__cospi_D); + // Then add and right-shift to get back to 16-bit range + const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING); + const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING); + const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING); + const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING); + const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS); + const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS); + const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS); + const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS); + // w0 = [b0 b1 b7 b6] + // w1 = [b8 b9 bF bE] + // w2 = [b4 b5 b3 b2] + // w3 = [bC bD bB bA] + const __m128i x0 = _mm_packs_epi32(w0, w1); + const __m128i x1 = _mm_packs_epi32(w2, w3); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x2(&x0, &x1); + if (overflow) { + aom_highbd_fdct4x4_c(input, output, stride); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + // x0 = [b0 b1 b7 b6 b8 b9 bF bE] + // x1 = [b4 b5 b3 b2 bC bD bB bA] + in0 = _mm_shuffle_epi32(x0, 0xD8); + in1 = _mm_shuffle_epi32(x1, 0x8D); + // in0 = [b0 b1 b8 b9 b7 b6 bF bE] + // in1 = [b3 b2 bB bA b4 b5 bC bD] + } + { + // vertical DCTs finished. Now we do the horizontal DCTs. + // Stage 3: Add/subtract + + // t0 = [c0 c1 c8 c9 c4 c5 cC cD] + // t1 = [c3 c2 cB cA -c7 -c6 -cF -cE] + const __m128i t0 = ADD_EPI16(in0, in1); + const __m128i t1 = SUB_EPI16(in0, in1); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x2(&t0, &t1); + if (overflow) { + aom_highbd_fdct4x4_c(input, output, stride); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + + // Stage 4: multiply by constants (which gets us into 32 bits). + { + // The constants needed here are: + // k__cospi_E = [p16 p16 p16 p16 p16 p16 p16 p16] + // k__cospi_F = [p16 m16 p16 m16 p16 m16 p16 m16] + // k__cospi_G = [p08 p24 p08 p24 m08 m24 m08 m24] + // k__cospi_H = [p24 m08 p24 m08 m24 p08 m24 p08] + const __m128i u0 = _mm_madd_epi16(t0, k__cospi_E); + const __m128i u1 = _mm_madd_epi16(t0, k__cospi_F); + const __m128i u2 = _mm_madd_epi16(t1, k__cospi_G); + const __m128i u3 = _mm_madd_epi16(t1, k__cospi_H); + // Then add and right-shift to get back to 16-bit range + // but this combines the final right-shift as well to save operations + // This unusual rounding operations is to maintain bit-accurate + // compatibility with the c version of this function which has two + // rounding steps in a row. + const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING2); + const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING2); + const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING2); + const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING2); + const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS2); + const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS2); + const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS2); + const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS2); + // w0 = [o0 o4 o8 oC] + // w1 = [o2 o6 oA oE] + // w2 = [o1 o5 o9 oD] + // w3 = [o3 o7 oB oF] + // remember the o's are numbered according to the correct output location + const __m128i x0 = _mm_packs_epi32(w0, w1); + const __m128i x1 = _mm_packs_epi32(w2, w3); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x2(&x0, &x1); + if (overflow) { + aom_highbd_fdct4x4_c(input, output, stride); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + { + // x0 = [o0 o4 o8 oC o2 o6 oA oE] + // x1 = [o1 o5 o9 oD o3 o7 oB oF] + const __m128i y0 = _mm_unpacklo_epi16(x0, x1); + const __m128i y1 = _mm_unpackhi_epi16(x0, x1); + // y0 = [o0 o1 o4 o5 o8 o9 oC oD] + // y1 = [o2 o3 o6 o7 oA oB oE oF] + in0 = _mm_unpacklo_epi32(y0, y1); + // in0 = [o0 o1 o2 o3 o4 o5 o6 o7] + in1 = _mm_unpackhi_epi32(y0, y1); + // in1 = [o8 o9 oA oB oC oD oE oF] + } + } + } + // Post-condition (v + 1) >> 2 is now incorporated into previous + // add and right-shift commands. Only 2 store instructions needed + // because we are using the fact that 1/3 are stored just after 0/2. + storeu_output(&in0, output + 0 * 4); + storeu_output(&in1, output + 2 * 4); +} + +void FDCT8x8_2D(const int16_t *input, tran_low_t *output, int stride) { + int pass; + // Constants + // When we use them, in one case, they are all the same. In all others + // it's a pair of them that we need to repeat four times. This is done + // by constructing the 32 bit constant corresponding to that pair. + const __m128i k__cospi_p16_p16 = _mm_set1_epi16((int16_t)cospi_16_64); + const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64); + const __m128i k__cospi_p24_p08 = pair_set_epi16(cospi_24_64, cospi_8_64); + const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64, cospi_24_64); + const __m128i k__cospi_p28_p04 = pair_set_epi16(cospi_28_64, cospi_4_64); + const __m128i k__cospi_m04_p28 = pair_set_epi16(-cospi_4_64, cospi_28_64); + const __m128i k__cospi_p12_p20 = pair_set_epi16(cospi_12_64, cospi_20_64); + const __m128i k__cospi_m20_p12 = pair_set_epi16(-cospi_20_64, cospi_12_64); + const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING); +#if DCT_HIGH_BIT_DEPTH + int overflow; +#endif + // Load input + __m128i in0 = _mm_load_si128((const __m128i *)(input + 0 * stride)); + __m128i in1 = _mm_load_si128((const __m128i *)(input + 1 * stride)); + __m128i in2 = _mm_load_si128((const __m128i *)(input + 2 * stride)); + __m128i in3 = _mm_load_si128((const __m128i *)(input + 3 * stride)); + __m128i in4 = _mm_load_si128((const __m128i *)(input + 4 * stride)); + __m128i in5 = _mm_load_si128((const __m128i *)(input + 5 * stride)); + __m128i in6 = _mm_load_si128((const __m128i *)(input + 6 * stride)); + __m128i in7 = _mm_load_si128((const __m128i *)(input + 7 * stride)); + // Pre-condition input (shift by two) + in0 = _mm_slli_epi16(in0, 2); + in1 = _mm_slli_epi16(in1, 2); + in2 = _mm_slli_epi16(in2, 2); + in3 = _mm_slli_epi16(in3, 2); + in4 = _mm_slli_epi16(in4, 2); + in5 = _mm_slli_epi16(in5, 2); + in6 = _mm_slli_epi16(in6, 2); + in7 = _mm_slli_epi16(in7, 2); + + // We do two passes, first the columns, then the rows. The results of the + // first pass are transposed so that the same column code can be reused. The + // results of the second pass are also transposed so that the rows (processed + // as columns) are put back in row positions. + for (pass = 0; pass < 2; pass++) { + // To store results of each pass before the transpose. + __m128i res0, res1, res2, res3, res4, res5, res6, res7; + // Add/subtract + const __m128i q0 = ADD_EPI16(in0, in7); + const __m128i q1 = ADD_EPI16(in1, in6); + const __m128i q2 = ADD_EPI16(in2, in5); + const __m128i q3 = ADD_EPI16(in3, in4); + const __m128i q4 = SUB_EPI16(in3, in4); + const __m128i q5 = SUB_EPI16(in2, in5); + const __m128i q6 = SUB_EPI16(in1, in6); + const __m128i q7 = SUB_EPI16(in0, in7); +#if DCT_HIGH_BIT_DEPTH + if (pass == 1) { + overflow = + check_epi16_overflow_x8(&q0, &q1, &q2, &q3, &q4, &q5, &q6, &q7); + if (overflow) { + aom_highbd_fdct8x8_c(input, output, stride); + return; + } + } +#endif // DCT_HIGH_BIT_DEPTH + // Work on first four results + { + // Add/subtract + const __m128i r0 = ADD_EPI16(q0, q3); + const __m128i r1 = ADD_EPI16(q1, q2); + const __m128i r2 = SUB_EPI16(q1, q2); + const __m128i r3 = SUB_EPI16(q0, q3); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x4(&r0, &r1, &r2, &r3); + if (overflow) { + aom_highbd_fdct8x8_c(input, output, stride); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + // Interleave to do the multiply by constants which gets us into 32bits + { + const __m128i t0 = _mm_unpacklo_epi16(r0, r1); + const __m128i t1 = _mm_unpackhi_epi16(r0, r1); + const __m128i t2 = _mm_unpacklo_epi16(r2, r3); + const __m128i t3 = _mm_unpackhi_epi16(r2, r3); + const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p16_p16); + const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p16_p16); + const __m128i u2 = _mm_madd_epi16(t0, k__cospi_p16_m16); + const __m128i u3 = _mm_madd_epi16(t1, k__cospi_p16_m16); + const __m128i u4 = _mm_madd_epi16(t2, k__cospi_p24_p08); + const __m128i u5 = _mm_madd_epi16(t3, k__cospi_p24_p08); + const __m128i u6 = _mm_madd_epi16(t2, k__cospi_m08_p24); + const __m128i u7 = _mm_madd_epi16(t3, k__cospi_m08_p24); + // dct_const_round_shift + const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING); + const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING); + const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING); + const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING); + const __m128i v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING); + const __m128i v5 = _mm_add_epi32(u5, k__DCT_CONST_ROUNDING); + const __m128i v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING); + const __m128i v7 = _mm_add_epi32(u7, k__DCT_CONST_ROUNDING); + const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS); + const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS); + const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS); + const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS); + const __m128i w4 = _mm_srai_epi32(v4, DCT_CONST_BITS); + const __m128i w5 = _mm_srai_epi32(v5, DCT_CONST_BITS); + const __m128i w6 = _mm_srai_epi32(v6, DCT_CONST_BITS); + const __m128i w7 = _mm_srai_epi32(v7, DCT_CONST_BITS); + // Combine + res0 = _mm_packs_epi32(w0, w1); + res4 = _mm_packs_epi32(w2, w3); + res2 = _mm_packs_epi32(w4, w5); + res6 = _mm_packs_epi32(w6, w7); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x4(&res0, &res4, &res2, &res6); + if (overflow) { + aom_highbd_fdct8x8_c(input, output, stride); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + } + } + // Work on next four results + { + // Interleave to do the multiply by constants which gets us into 32bits + const __m128i d0 = _mm_unpacklo_epi16(q6, q5); + const __m128i d1 = _mm_unpackhi_epi16(q6, q5); + const __m128i e0 = _mm_madd_epi16(d0, k__cospi_p16_m16); + const __m128i e1 = _mm_madd_epi16(d1, k__cospi_p16_m16); + const __m128i e2 = _mm_madd_epi16(d0, k__cospi_p16_p16); + const __m128i e3 = _mm_madd_epi16(d1, k__cospi_p16_p16); + // dct_const_round_shift + const __m128i f0 = _mm_add_epi32(e0, k__DCT_CONST_ROUNDING); + const __m128i f1 = _mm_add_epi32(e1, k__DCT_CONST_ROUNDING); + const __m128i f2 = _mm_add_epi32(e2, k__DCT_CONST_ROUNDING); + const __m128i f3 = _mm_add_epi32(e3, k__DCT_CONST_ROUNDING); + const __m128i s0 = _mm_srai_epi32(f0, DCT_CONST_BITS); + const __m128i s1 = _mm_srai_epi32(f1, DCT_CONST_BITS); + const __m128i s2 = _mm_srai_epi32(f2, DCT_CONST_BITS); + const __m128i s3 = _mm_srai_epi32(f3, DCT_CONST_BITS); + // Combine + const __m128i r0 = _mm_packs_epi32(s0, s1); + const __m128i r1 = _mm_packs_epi32(s2, s3); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x2(&r0, &r1); + if (overflow) { + aom_highbd_fdct8x8_c(input, output, stride); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + { + // Add/subtract + const __m128i x0 = ADD_EPI16(q4, r0); + const __m128i x1 = SUB_EPI16(q4, r0); + const __m128i x2 = SUB_EPI16(q7, r1); + const __m128i x3 = ADD_EPI16(q7, r1); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x4(&x0, &x1, &x2, &x3); + if (overflow) { + aom_highbd_fdct8x8_c(input, output, stride); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + // Interleave to do the multiply by constants which gets us into 32bits + { + const __m128i t0 = _mm_unpacklo_epi16(x0, x3); + const __m128i t1 = _mm_unpackhi_epi16(x0, x3); + const __m128i t2 = _mm_unpacklo_epi16(x1, x2); + const __m128i t3 = _mm_unpackhi_epi16(x1, x2); + const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p28_p04); + const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p28_p04); + const __m128i u2 = _mm_madd_epi16(t0, k__cospi_m04_p28); + const __m128i u3 = _mm_madd_epi16(t1, k__cospi_m04_p28); + const __m128i u4 = _mm_madd_epi16(t2, k__cospi_p12_p20); + const __m128i u5 = _mm_madd_epi16(t3, k__cospi_p12_p20); + const __m128i u6 = _mm_madd_epi16(t2, k__cospi_m20_p12); + const __m128i u7 = _mm_madd_epi16(t3, k__cospi_m20_p12); + // dct_const_round_shift + const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING); + const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING); + const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING); + const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING); + const __m128i v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING); + const __m128i v5 = _mm_add_epi32(u5, k__DCT_CONST_ROUNDING); + const __m128i v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING); + const __m128i v7 = _mm_add_epi32(u7, k__DCT_CONST_ROUNDING); + const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS); + const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS); + const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS); + const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS); + const __m128i w4 = _mm_srai_epi32(v4, DCT_CONST_BITS); + const __m128i w5 = _mm_srai_epi32(v5, DCT_CONST_BITS); + const __m128i w6 = _mm_srai_epi32(v6, DCT_CONST_BITS); + const __m128i w7 = _mm_srai_epi32(v7, DCT_CONST_BITS); + // Combine + res1 = _mm_packs_epi32(w0, w1); + res7 = _mm_packs_epi32(w2, w3); + res5 = _mm_packs_epi32(w4, w5); + res3 = _mm_packs_epi32(w6, w7); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x4(&res1, &res7, &res5, &res3); + if (overflow) { + aom_highbd_fdct8x8_c(input, output, stride); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + } + } + } + // Transpose the 8x8. + { + // 00 01 02 03 04 05 06 07 + // 10 11 12 13 14 15 16 17 + // 20 21 22 23 24 25 26 27 + // 30 31 32 33 34 35 36 37 + // 40 41 42 43 44 45 46 47 + // 50 51 52 53 54 55 56 57 + // 60 61 62 63 64 65 66 67 + // 70 71 72 73 74 75 76 77 + const __m128i tr0_0 = _mm_unpacklo_epi16(res0, res1); + const __m128i tr0_1 = _mm_unpacklo_epi16(res2, res3); + const __m128i tr0_2 = _mm_unpackhi_epi16(res0, res1); + const __m128i tr0_3 = _mm_unpackhi_epi16(res2, res3); + const __m128i tr0_4 = _mm_unpacklo_epi16(res4, res5); + const __m128i tr0_5 = _mm_unpacklo_epi16(res6, res7); + const __m128i tr0_6 = _mm_unpackhi_epi16(res4, res5); + const __m128i tr0_7 = _mm_unpackhi_epi16(res6, res7); + // 00 10 01 11 02 12 03 13 + // 20 30 21 31 22 32 23 33 + // 04 14 05 15 06 16 07 17 + // 24 34 25 35 26 36 27 37 + // 40 50 41 51 42 52 43 53 + // 60 70 61 71 62 72 63 73 + // 54 54 55 55 56 56 57 57 + // 64 74 65 75 66 76 67 77 + const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1); + const __m128i tr1_1 = _mm_unpacklo_epi32(tr0_2, tr0_3); + const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1); + const __m128i tr1_3 = _mm_unpackhi_epi32(tr0_2, tr0_3); + const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_4, tr0_5); + const __m128i tr1_5 = _mm_unpacklo_epi32(tr0_6, tr0_7); + const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_4, tr0_5); + const __m128i tr1_7 = _mm_unpackhi_epi32(tr0_6, tr0_7); + // 00 10 20 30 01 11 21 31 + // 40 50 60 70 41 51 61 71 + // 02 12 22 32 03 13 23 33 + // 42 52 62 72 43 53 63 73 + // 04 14 24 34 05 15 21 36 + // 44 54 64 74 45 55 61 76 + // 06 16 26 36 07 17 27 37 + // 46 56 66 76 47 57 67 77 + in0 = _mm_unpacklo_epi64(tr1_0, tr1_4); + in1 = _mm_unpackhi_epi64(tr1_0, tr1_4); + in2 = _mm_unpacklo_epi64(tr1_2, tr1_6); + in3 = _mm_unpackhi_epi64(tr1_2, tr1_6); + in4 = _mm_unpacklo_epi64(tr1_1, tr1_5); + in5 = _mm_unpackhi_epi64(tr1_1, tr1_5); + in6 = _mm_unpacklo_epi64(tr1_3, tr1_7); + in7 = _mm_unpackhi_epi64(tr1_3, tr1_7); + // 00 10 20 30 40 50 60 70 + // 01 11 21 31 41 51 61 71 + // 02 12 22 32 42 52 62 72 + // 03 13 23 33 43 53 63 73 + // 04 14 24 34 44 54 64 74 + // 05 15 25 35 45 55 65 75 + // 06 16 26 36 46 56 66 76 + // 07 17 27 37 47 57 67 77 + } + } + // Post-condition output and store it + { + // Post-condition (division by two) + // division of two 16 bits signed numbers using shifts + // n / 2 = (n - (n >> 15)) >> 1 + const __m128i sign_in0 = _mm_srai_epi16(in0, 15); + const __m128i sign_in1 = _mm_srai_epi16(in1, 15); + const __m128i sign_in2 = _mm_srai_epi16(in2, 15); + const __m128i sign_in3 = _mm_srai_epi16(in3, 15); + const __m128i sign_in4 = _mm_srai_epi16(in4, 15); + const __m128i sign_in5 = _mm_srai_epi16(in5, 15); + const __m128i sign_in6 = _mm_srai_epi16(in6, 15); + const __m128i sign_in7 = _mm_srai_epi16(in7, 15); + in0 = _mm_sub_epi16(in0, sign_in0); + in1 = _mm_sub_epi16(in1, sign_in1); + in2 = _mm_sub_epi16(in2, sign_in2); + in3 = _mm_sub_epi16(in3, sign_in3); + in4 = _mm_sub_epi16(in4, sign_in4); + in5 = _mm_sub_epi16(in5, sign_in5); + in6 = _mm_sub_epi16(in6, sign_in6); + in7 = _mm_sub_epi16(in7, sign_in7); + in0 = _mm_srai_epi16(in0, 1); + in1 = _mm_srai_epi16(in1, 1); + in2 = _mm_srai_epi16(in2, 1); + in3 = _mm_srai_epi16(in3, 1); + in4 = _mm_srai_epi16(in4, 1); + in5 = _mm_srai_epi16(in5, 1); + in6 = _mm_srai_epi16(in6, 1); + in7 = _mm_srai_epi16(in7, 1); + // store results + store_output(&in0, (output + 0 * 8)); + store_output(&in1, (output + 1 * 8)); + store_output(&in2, (output + 2 * 8)); + store_output(&in3, (output + 3 * 8)); + store_output(&in4, (output + 4 * 8)); + store_output(&in5, (output + 5 * 8)); + store_output(&in6, (output + 6 * 8)); + store_output(&in7, (output + 7 * 8)); + } +} + +void FDCT16x16_2D(const int16_t *input, tran_low_t *output, int stride) { + // The 2D transform is done with two passes which are actually pretty + // similar. In the first one, we transform the columns and transpose + // the results. In the second one, we transform the rows. To achieve that, + // as the first pass results are transposed, we transpose the columns (that + // is the transposed rows) and transpose the results (so that it goes back + // in normal/row positions). + int pass; + // We need an intermediate buffer between passes. + DECLARE_ALIGNED(16, int16_t, intermediate[256]); + const int16_t *in = input; + int16_t *out0 = intermediate; + tran_low_t *out1 = output; + // Constants + // When we use them, in one case, they are all the same. In all others + // it's a pair of them that we need to repeat four times. This is done + // by constructing the 32 bit constant corresponding to that pair. + const __m128i k__cospi_p16_p16 = _mm_set1_epi16((int16_t)cospi_16_64); + const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64); + const __m128i k__cospi_p24_p08 = pair_set_epi16(cospi_24_64, cospi_8_64); + const __m128i k__cospi_p08_m24 = pair_set_epi16(cospi_8_64, -cospi_24_64); + const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64, cospi_24_64); + const __m128i k__cospi_p28_p04 = pair_set_epi16(cospi_28_64, cospi_4_64); + const __m128i k__cospi_m04_p28 = pair_set_epi16(-cospi_4_64, cospi_28_64); + const __m128i k__cospi_p12_p20 = pair_set_epi16(cospi_12_64, cospi_20_64); + const __m128i k__cospi_m20_p12 = pair_set_epi16(-cospi_20_64, cospi_12_64); + const __m128i k__cospi_p30_p02 = pair_set_epi16(cospi_30_64, cospi_2_64); + const __m128i k__cospi_p14_p18 = pair_set_epi16(cospi_14_64, cospi_18_64); + const __m128i k__cospi_m02_p30 = pair_set_epi16(-cospi_2_64, cospi_30_64); + const __m128i k__cospi_m18_p14 = pair_set_epi16(-cospi_18_64, cospi_14_64); + const __m128i k__cospi_p22_p10 = pair_set_epi16(cospi_22_64, cospi_10_64); + const __m128i k__cospi_p06_p26 = pair_set_epi16(cospi_6_64, cospi_26_64); + const __m128i k__cospi_m10_p22 = pair_set_epi16(-cospi_10_64, cospi_22_64); + const __m128i k__cospi_m26_p06 = pair_set_epi16(-cospi_26_64, cospi_6_64); + const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING); + const __m128i kOne = _mm_set1_epi16(1); + // Do the two transform/transpose passes + for (pass = 0; pass < 2; ++pass) { + // We process eight columns (transposed rows in second pass) at a time. + int column_start; +#if DCT_HIGH_BIT_DEPTH + int overflow; +#endif + for (column_start = 0; column_start < 16; column_start += 8) { + __m128i in00, in01, in02, in03, in04, in05, in06, in07; + __m128i in08, in09, in10, in11, in12, in13, in14, in15; + __m128i input0, input1, input2, input3, input4, input5, input6, input7; + __m128i step1_0, step1_1, step1_2, step1_3; + __m128i step1_4, step1_5, step1_6, step1_7; + __m128i step2_1, step2_2, step2_3, step2_4, step2_5, step2_6; + __m128i step3_0, step3_1, step3_2, step3_3; + __m128i step3_4, step3_5, step3_6, step3_7; + __m128i res00, res01, res02, res03, res04, res05, res06, res07; + __m128i res08, res09, res10, res11, res12, res13, res14, res15; + // Load and pre-condition input. + if (0 == pass) { + in00 = _mm_load_si128((const __m128i *)(in + 0 * stride)); + in01 = _mm_load_si128((const __m128i *)(in + 1 * stride)); + in02 = _mm_load_si128((const __m128i *)(in + 2 * stride)); + in03 = _mm_load_si128((const __m128i *)(in + 3 * stride)); + in04 = _mm_load_si128((const __m128i *)(in + 4 * stride)); + in05 = _mm_load_si128((const __m128i *)(in + 5 * stride)); + in06 = _mm_load_si128((const __m128i *)(in + 6 * stride)); + in07 = _mm_load_si128((const __m128i *)(in + 7 * stride)); + in08 = _mm_load_si128((const __m128i *)(in + 8 * stride)); + in09 = _mm_load_si128((const __m128i *)(in + 9 * stride)); + in10 = _mm_load_si128((const __m128i *)(in + 10 * stride)); + in11 = _mm_load_si128((const __m128i *)(in + 11 * stride)); + in12 = _mm_load_si128((const __m128i *)(in + 12 * stride)); + in13 = _mm_load_si128((const __m128i *)(in + 13 * stride)); + in14 = _mm_load_si128((const __m128i *)(in + 14 * stride)); + in15 = _mm_load_si128((const __m128i *)(in + 15 * stride)); + // x = x << 2 + in00 = _mm_slli_epi16(in00, 2); + in01 = _mm_slli_epi16(in01, 2); + in02 = _mm_slli_epi16(in02, 2); + in03 = _mm_slli_epi16(in03, 2); + in04 = _mm_slli_epi16(in04, 2); + in05 = _mm_slli_epi16(in05, 2); + in06 = _mm_slli_epi16(in06, 2); + in07 = _mm_slli_epi16(in07, 2); + in08 = _mm_slli_epi16(in08, 2); + in09 = _mm_slli_epi16(in09, 2); + in10 = _mm_slli_epi16(in10, 2); + in11 = _mm_slli_epi16(in11, 2); + in12 = _mm_slli_epi16(in12, 2); + in13 = _mm_slli_epi16(in13, 2); + in14 = _mm_slli_epi16(in14, 2); + in15 = _mm_slli_epi16(in15, 2); + } else { + in00 = _mm_load_si128((const __m128i *)(in + 0 * 16)); + in01 = _mm_load_si128((const __m128i *)(in + 1 * 16)); + in02 = _mm_load_si128((const __m128i *)(in + 2 * 16)); + in03 = _mm_load_si128((const __m128i *)(in + 3 * 16)); + in04 = _mm_load_si128((const __m128i *)(in + 4 * 16)); + in05 = _mm_load_si128((const __m128i *)(in + 5 * 16)); + in06 = _mm_load_si128((const __m128i *)(in + 6 * 16)); + in07 = _mm_load_si128((const __m128i *)(in + 7 * 16)); + in08 = _mm_load_si128((const __m128i *)(in + 8 * 16)); + in09 = _mm_load_si128((const __m128i *)(in + 9 * 16)); + in10 = _mm_load_si128((const __m128i *)(in + 10 * 16)); + in11 = _mm_load_si128((const __m128i *)(in + 11 * 16)); + in12 = _mm_load_si128((const __m128i *)(in + 12 * 16)); + in13 = _mm_load_si128((const __m128i *)(in + 13 * 16)); + in14 = _mm_load_si128((const __m128i *)(in + 14 * 16)); + in15 = _mm_load_si128((const __m128i *)(in + 15 * 16)); + // x = (x + 1) >> 2 + in00 = _mm_add_epi16(in00, kOne); + in01 = _mm_add_epi16(in01, kOne); + in02 = _mm_add_epi16(in02, kOne); + in03 = _mm_add_epi16(in03, kOne); + in04 = _mm_add_epi16(in04, kOne); + in05 = _mm_add_epi16(in05, kOne); + in06 = _mm_add_epi16(in06, kOne); + in07 = _mm_add_epi16(in07, kOne); + in08 = _mm_add_epi16(in08, kOne); + in09 = _mm_add_epi16(in09, kOne); + in10 = _mm_add_epi16(in10, kOne); + in11 = _mm_add_epi16(in11, kOne); + in12 = _mm_add_epi16(in12, kOne); + in13 = _mm_add_epi16(in13, kOne); + in14 = _mm_add_epi16(in14, kOne); + in15 = _mm_add_epi16(in15, kOne); + in00 = _mm_srai_epi16(in00, 2); + in01 = _mm_srai_epi16(in01, 2); + in02 = _mm_srai_epi16(in02, 2); + in03 = _mm_srai_epi16(in03, 2); + in04 = _mm_srai_epi16(in04, 2); + in05 = _mm_srai_epi16(in05, 2); + in06 = _mm_srai_epi16(in06, 2); + in07 = _mm_srai_epi16(in07, 2); + in08 = _mm_srai_epi16(in08, 2); + in09 = _mm_srai_epi16(in09, 2); + in10 = _mm_srai_epi16(in10, 2); + in11 = _mm_srai_epi16(in11, 2); + in12 = _mm_srai_epi16(in12, 2); + in13 = _mm_srai_epi16(in13, 2); + in14 = _mm_srai_epi16(in14, 2); + in15 = _mm_srai_epi16(in15, 2); + } + in += 8; + // Calculate input for the first 8 results. + { + input0 = ADD_EPI16(in00, in15); + input1 = ADD_EPI16(in01, in14); + input2 = ADD_EPI16(in02, in13); + input3 = ADD_EPI16(in03, in12); + input4 = ADD_EPI16(in04, in11); + input5 = ADD_EPI16(in05, in10); + input6 = ADD_EPI16(in06, in09); + input7 = ADD_EPI16(in07, in08); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x8(&input0, &input1, &input2, &input3, + &input4, &input5, &input6, &input7); + if (overflow) { + aom_highbd_fdct16x16_c(input, output, stride); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + } + // Calculate input for the next 8 results. + { + step1_0 = SUB_EPI16(in07, in08); + step1_1 = SUB_EPI16(in06, in09); + step1_2 = SUB_EPI16(in05, in10); + step1_3 = SUB_EPI16(in04, in11); + step1_4 = SUB_EPI16(in03, in12); + step1_5 = SUB_EPI16(in02, in13); + step1_6 = SUB_EPI16(in01, in14); + step1_7 = SUB_EPI16(in00, in15); +#if DCT_HIGH_BIT_DEPTH + overflow = + check_epi16_overflow_x8(&step1_0, &step1_1, &step1_2, &step1_3, + &step1_4, &step1_5, &step1_6, &step1_7); + if (overflow) { + aom_highbd_fdct16x16_c(input, output, stride); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + } + // Work on the first eight values; fdct8(input, even_results); + { + // Add/subtract + const __m128i q0 = ADD_EPI16(input0, input7); + const __m128i q1 = ADD_EPI16(input1, input6); + const __m128i q2 = ADD_EPI16(input2, input5); + const __m128i q3 = ADD_EPI16(input3, input4); + const __m128i q4 = SUB_EPI16(input3, input4); + const __m128i q5 = SUB_EPI16(input2, input5); + const __m128i q6 = SUB_EPI16(input1, input6); + const __m128i q7 = SUB_EPI16(input0, input7); +#if DCT_HIGH_BIT_DEPTH + overflow = + check_epi16_overflow_x8(&q0, &q1, &q2, &q3, &q4, &q5, &q6, &q7); + if (overflow) { + aom_highbd_fdct16x16_c(input, output, stride); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + // Work on first four results + { + // Add/subtract + const __m128i r0 = ADD_EPI16(q0, q3); + const __m128i r1 = ADD_EPI16(q1, q2); + const __m128i r2 = SUB_EPI16(q1, q2); + const __m128i r3 = SUB_EPI16(q0, q3); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x4(&r0, &r1, &r2, &r3); + if (overflow) { + aom_highbd_fdct16x16_c(input, output, stride); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + // Interleave to do the multiply by constants which gets us + // into 32 bits. + { + const __m128i t0 = _mm_unpacklo_epi16(r0, r1); + const __m128i t1 = _mm_unpackhi_epi16(r0, r1); + const __m128i t2 = _mm_unpacklo_epi16(r2, r3); + const __m128i t3 = _mm_unpackhi_epi16(r2, r3); + res00 = mult_round_shift(&t0, &t1, &k__cospi_p16_p16, + &k__DCT_CONST_ROUNDING, DCT_CONST_BITS); + res08 = mult_round_shift(&t0, &t1, &k__cospi_p16_m16, + &k__DCT_CONST_ROUNDING, DCT_CONST_BITS); + res04 = mult_round_shift(&t2, &t3, &k__cospi_p24_p08, + &k__DCT_CONST_ROUNDING, DCT_CONST_BITS); + res12 = mult_round_shift(&t2, &t3, &k__cospi_m08_p24, + &k__DCT_CONST_ROUNDING, DCT_CONST_BITS); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x4(&res00, &res08, &res04, &res12); + if (overflow) { + aom_highbd_fdct16x16_c(input, output, stride); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + } + } + // Work on next four results + { + // Interleave to do the multiply by constants which gets us + // into 32 bits. + const __m128i d0 = _mm_unpacklo_epi16(q6, q5); + const __m128i d1 = _mm_unpackhi_epi16(q6, q5); + const __m128i r0 = + mult_round_shift(&d0, &d1, &k__cospi_p16_m16, + &k__DCT_CONST_ROUNDING, DCT_CONST_BITS); + const __m128i r1 = + mult_round_shift(&d0, &d1, &k__cospi_p16_p16, + &k__DCT_CONST_ROUNDING, DCT_CONST_BITS); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x2(&r0, &r1); + if (overflow) { + aom_highbd_fdct16x16_c(input, output, stride); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + { + // Add/subtract + const __m128i x0 = ADD_EPI16(q4, r0); + const __m128i x1 = SUB_EPI16(q4, r0); + const __m128i x2 = SUB_EPI16(q7, r1); + const __m128i x3 = ADD_EPI16(q7, r1); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x4(&x0, &x1, &x2, &x3); + if (overflow) { + aom_highbd_fdct16x16_c(input, output, stride); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + // Interleave to do the multiply by constants which gets us + // into 32 bits. + { + const __m128i t0 = _mm_unpacklo_epi16(x0, x3); + const __m128i t1 = _mm_unpackhi_epi16(x0, x3); + const __m128i t2 = _mm_unpacklo_epi16(x1, x2); + const __m128i t3 = _mm_unpackhi_epi16(x1, x2); + res02 = mult_round_shift(&t0, &t1, &k__cospi_p28_p04, + &k__DCT_CONST_ROUNDING, DCT_CONST_BITS); + res14 = mult_round_shift(&t0, &t1, &k__cospi_m04_p28, + &k__DCT_CONST_ROUNDING, DCT_CONST_BITS); + res10 = mult_round_shift(&t2, &t3, &k__cospi_p12_p20, + &k__DCT_CONST_ROUNDING, DCT_CONST_BITS); + res06 = mult_round_shift(&t2, &t3, &k__cospi_m20_p12, + &k__DCT_CONST_ROUNDING, DCT_CONST_BITS); +#if DCT_HIGH_BIT_DEPTH + overflow = + check_epi16_overflow_x4(&res02, &res14, &res10, &res06); + if (overflow) { + aom_highbd_fdct16x16_c(input, output, stride); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + } + } + } + } + // Work on the next eight values; step1 -> odd_results + { + // step 2 + { + const __m128i t0 = _mm_unpacklo_epi16(step1_5, step1_2); + const __m128i t1 = _mm_unpackhi_epi16(step1_5, step1_2); + const __m128i t2 = _mm_unpacklo_epi16(step1_4, step1_3); + const __m128i t3 = _mm_unpackhi_epi16(step1_4, step1_3); + step2_2 = mult_round_shift(&t0, &t1, &k__cospi_p16_m16, + &k__DCT_CONST_ROUNDING, DCT_CONST_BITS); + step2_3 = mult_round_shift(&t2, &t3, &k__cospi_p16_m16, + &k__DCT_CONST_ROUNDING, DCT_CONST_BITS); + step2_5 = mult_round_shift(&t0, &t1, &k__cospi_p16_p16, + &k__DCT_CONST_ROUNDING, DCT_CONST_BITS); + step2_4 = mult_round_shift(&t2, &t3, &k__cospi_p16_p16, + &k__DCT_CONST_ROUNDING, DCT_CONST_BITS); +#if DCT_HIGH_BIT_DEPTH + overflow = + check_epi16_overflow_x4(&step2_2, &step2_3, &step2_5, &step2_4); + if (overflow) { + aom_highbd_fdct16x16_c(input, output, stride); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + } + // step 3 + { + step3_0 = ADD_EPI16(step1_0, step2_3); + step3_1 = ADD_EPI16(step1_1, step2_2); + step3_2 = SUB_EPI16(step1_1, step2_2); + step3_3 = SUB_EPI16(step1_0, step2_3); + step3_4 = SUB_EPI16(step1_7, step2_4); + step3_5 = SUB_EPI16(step1_6, step2_5); + step3_6 = ADD_EPI16(step1_6, step2_5); + step3_7 = ADD_EPI16(step1_7, step2_4); +#if DCT_HIGH_BIT_DEPTH + overflow = + check_epi16_overflow_x8(&step3_0, &step3_1, &step3_2, &step3_3, + &step3_4, &step3_5, &step3_6, &step3_7); + if (overflow) { + aom_highbd_fdct16x16_c(input, output, stride); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + } + // step 4 + { + const __m128i t0 = _mm_unpacklo_epi16(step3_1, step3_6); + const __m128i t1 = _mm_unpackhi_epi16(step3_1, step3_6); + const __m128i t2 = _mm_unpacklo_epi16(step3_2, step3_5); + const __m128i t3 = _mm_unpackhi_epi16(step3_2, step3_5); + step2_1 = mult_round_shift(&t0, &t1, &k__cospi_m08_p24, + &k__DCT_CONST_ROUNDING, DCT_CONST_BITS); + step2_2 = mult_round_shift(&t2, &t3, &k__cospi_p24_p08, + &k__DCT_CONST_ROUNDING, DCT_CONST_BITS); + step2_6 = mult_round_shift(&t0, &t1, &k__cospi_p24_p08, + &k__DCT_CONST_ROUNDING, DCT_CONST_BITS); + step2_5 = mult_round_shift(&t2, &t3, &k__cospi_p08_m24, + &k__DCT_CONST_ROUNDING, DCT_CONST_BITS); +#if DCT_HIGH_BIT_DEPTH + overflow = + check_epi16_overflow_x4(&step2_1, &step2_2, &step2_6, &step2_5); + if (overflow) { + aom_highbd_fdct16x16_c(input, output, stride); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + } + // step 5 + { + step1_0 = ADD_EPI16(step3_0, step2_1); + step1_1 = SUB_EPI16(step3_0, step2_1); + step1_2 = ADD_EPI16(step3_3, step2_2); + step1_3 = SUB_EPI16(step3_3, step2_2); + step1_4 = SUB_EPI16(step3_4, step2_5); + step1_5 = ADD_EPI16(step3_4, step2_5); + step1_6 = SUB_EPI16(step3_7, step2_6); + step1_7 = ADD_EPI16(step3_7, step2_6); +#if DCT_HIGH_BIT_DEPTH + overflow = + check_epi16_overflow_x8(&step1_0, &step1_1, &step1_2, &step1_3, + &step1_4, &step1_5, &step1_6, &step1_7); + if (overflow) { + aom_highbd_fdct16x16_c(input, output, stride); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + } + // step 6 + { + const __m128i t0 = _mm_unpacklo_epi16(step1_0, step1_7); + const __m128i t1 = _mm_unpackhi_epi16(step1_0, step1_7); + const __m128i t2 = _mm_unpacklo_epi16(step1_1, step1_6); + const __m128i t3 = _mm_unpackhi_epi16(step1_1, step1_6); + res01 = mult_round_shift(&t0, &t1, &k__cospi_p30_p02, + &k__DCT_CONST_ROUNDING, DCT_CONST_BITS); + res09 = mult_round_shift(&t2, &t3, &k__cospi_p14_p18, + &k__DCT_CONST_ROUNDING, DCT_CONST_BITS); + res15 = mult_round_shift(&t0, &t1, &k__cospi_m02_p30, + &k__DCT_CONST_ROUNDING, DCT_CONST_BITS); + res07 = mult_round_shift(&t2, &t3, &k__cospi_m18_p14, + &k__DCT_CONST_ROUNDING, DCT_CONST_BITS); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x4(&res01, &res09, &res15, &res07); + if (overflow) { + aom_highbd_fdct16x16_c(input, output, stride); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + } + { + const __m128i t0 = _mm_unpacklo_epi16(step1_2, step1_5); + const __m128i t1 = _mm_unpackhi_epi16(step1_2, step1_5); + const __m128i t2 = _mm_unpacklo_epi16(step1_3, step1_4); + const __m128i t3 = _mm_unpackhi_epi16(step1_3, step1_4); + res05 = mult_round_shift(&t0, &t1, &k__cospi_p22_p10, + &k__DCT_CONST_ROUNDING, DCT_CONST_BITS); + res13 = mult_round_shift(&t2, &t3, &k__cospi_p06_p26, + &k__DCT_CONST_ROUNDING, DCT_CONST_BITS); + res11 = mult_round_shift(&t0, &t1, &k__cospi_m10_p22, + &k__DCT_CONST_ROUNDING, DCT_CONST_BITS); + res03 = mult_round_shift(&t2, &t3, &k__cospi_m26_p06, + &k__DCT_CONST_ROUNDING, DCT_CONST_BITS); +#if DCT_HIGH_BIT_DEPTH + overflow = check_epi16_overflow_x4(&res05, &res13, &res11, &res03); + if (overflow) { + aom_highbd_fdct16x16_c(input, output, stride); + return; + } +#endif // DCT_HIGH_BIT_DEPTH + } + } + // Transpose the results, do it as two 8x8 transposes. + transpose_and_output8x8(&res00, &res01, &res02, &res03, &res04, &res05, + &res06, &res07, pass, out0, out1); + transpose_and_output8x8(&res08, &res09, &res10, &res11, &res12, &res13, + &res14, &res15, pass, out0 + 8, out1 + 8); + if (pass == 0) { + out0 += 8 * 16; + } else { + out1 += 8 * 16; + } + } + // Setup in/out for next pass. + in = intermediate; + } +} + +#undef ADD_EPI16 +#undef SUB_EPI16 diff --git a/third_party/aom/aom_dsp/x86/fwd_txfm_sse2.c b/third_party/aom/aom_dsp/x86/fwd_txfm_sse2.c new file mode 100644 index 000000000..a337e618d --- /dev/null +++ b/third_party/aom/aom_dsp/x86/fwd_txfm_sse2.c @@ -0,0 +1,273 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include <emmintrin.h> // SSE2 + +#include "./aom_config.h" +#include "./aom_dsp_rtcd.h" +#include "aom_dsp/aom_dsp_common.h" +#include "aom_dsp/x86/fwd_txfm_sse2.h" + +void aom_fdct4x4_1_sse2(const int16_t *input, tran_low_t *output, int stride) { + __m128i in0, in1; + __m128i tmp; + const __m128i zero = _mm_setzero_si128(); + in0 = _mm_loadl_epi64((const __m128i *)(input + 0 * stride)); + in1 = _mm_loadl_epi64((const __m128i *)(input + 1 * stride)); + in1 = _mm_unpacklo_epi64( + in1, _mm_loadl_epi64((const __m128i *)(input + 2 * stride))); + in0 = _mm_unpacklo_epi64( + in0, _mm_loadl_epi64((const __m128i *)(input + 3 * stride))); + + tmp = _mm_add_epi16(in0, in1); + in0 = _mm_unpacklo_epi16(zero, tmp); + in1 = _mm_unpackhi_epi16(zero, tmp); + in0 = _mm_srai_epi32(in0, 16); + in1 = _mm_srai_epi32(in1, 16); + + tmp = _mm_add_epi32(in0, in1); + in0 = _mm_unpacklo_epi32(tmp, zero); + in1 = _mm_unpackhi_epi32(tmp, zero); + + tmp = _mm_add_epi32(in0, in1); + in0 = _mm_srli_si128(tmp, 8); + + in1 = _mm_add_epi32(tmp, in0); + in0 = _mm_slli_epi32(in1, 1); + output[0] = (tran_low_t)_mm_cvtsi128_si32(in0); +} + +void aom_fdct8x8_1_sse2(const int16_t *input, tran_low_t *output, int stride) { + __m128i in0 = _mm_load_si128((const __m128i *)(input + 0 * stride)); + __m128i in1 = _mm_load_si128((const __m128i *)(input + 1 * stride)); + __m128i in2 = _mm_load_si128((const __m128i *)(input + 2 * stride)); + __m128i in3 = _mm_load_si128((const __m128i *)(input + 3 * stride)); + __m128i u0, u1, sum; + + u0 = _mm_add_epi16(in0, in1); + u1 = _mm_add_epi16(in2, in3); + + in0 = _mm_load_si128((const __m128i *)(input + 4 * stride)); + in1 = _mm_load_si128((const __m128i *)(input + 5 * stride)); + in2 = _mm_load_si128((const __m128i *)(input + 6 * stride)); + in3 = _mm_load_si128((const __m128i *)(input + 7 * stride)); + + sum = _mm_add_epi16(u0, u1); + + in0 = _mm_add_epi16(in0, in1); + in2 = _mm_add_epi16(in2, in3); + sum = _mm_add_epi16(sum, in0); + + u0 = _mm_setzero_si128(); + sum = _mm_add_epi16(sum, in2); + + in0 = _mm_unpacklo_epi16(u0, sum); + in1 = _mm_unpackhi_epi16(u0, sum); + in0 = _mm_srai_epi32(in0, 16); + in1 = _mm_srai_epi32(in1, 16); + + sum = _mm_add_epi32(in0, in1); + in0 = _mm_unpacklo_epi32(sum, u0); + in1 = _mm_unpackhi_epi32(sum, u0); + + sum = _mm_add_epi32(in0, in1); + in0 = _mm_srli_si128(sum, 8); + + in1 = _mm_add_epi32(sum, in0); + output[0] = (tran_low_t)_mm_cvtsi128_si32(in1); +} + +void aom_fdct16x16_1_sse2(const int16_t *input, tran_low_t *output, + int stride) { + __m128i in0, in1, in2, in3; + __m128i u0, u1; + __m128i sum = _mm_setzero_si128(); + int i; + + for (i = 0; i < 2; ++i) { + in0 = _mm_load_si128((const __m128i *)(input + 0 * stride + 0)); + in1 = _mm_load_si128((const __m128i *)(input + 0 * stride + 8)); + in2 = _mm_load_si128((const __m128i *)(input + 1 * stride + 0)); + in3 = _mm_load_si128((const __m128i *)(input + 1 * stride + 8)); + + u0 = _mm_add_epi16(in0, in1); + u1 = _mm_add_epi16(in2, in3); + sum = _mm_add_epi16(sum, u0); + + in0 = _mm_load_si128((const __m128i *)(input + 2 * stride + 0)); + in1 = _mm_load_si128((const __m128i *)(input + 2 * stride + 8)); + in2 = _mm_load_si128((const __m128i *)(input + 3 * stride + 0)); + in3 = _mm_load_si128((const __m128i *)(input + 3 * stride + 8)); + + sum = _mm_add_epi16(sum, u1); + u0 = _mm_add_epi16(in0, in1); + u1 = _mm_add_epi16(in2, in3); + sum = _mm_add_epi16(sum, u0); + + in0 = _mm_load_si128((const __m128i *)(input + 4 * stride + 0)); + in1 = _mm_load_si128((const __m128i *)(input + 4 * stride + 8)); + in2 = _mm_load_si128((const __m128i *)(input + 5 * stride + 0)); + in3 = _mm_load_si128((const __m128i *)(input + 5 * stride + 8)); + + sum = _mm_add_epi16(sum, u1); + u0 = _mm_add_epi16(in0, in1); + u1 = _mm_add_epi16(in2, in3); + sum = _mm_add_epi16(sum, u0); + + in0 = _mm_load_si128((const __m128i *)(input + 6 * stride + 0)); + in1 = _mm_load_si128((const __m128i *)(input + 6 * stride + 8)); + in2 = _mm_load_si128((const __m128i *)(input + 7 * stride + 0)); + in3 = _mm_load_si128((const __m128i *)(input + 7 * stride + 8)); + + sum = _mm_add_epi16(sum, u1); + u0 = _mm_add_epi16(in0, in1); + u1 = _mm_add_epi16(in2, in3); + sum = _mm_add_epi16(sum, u0); + + sum = _mm_add_epi16(sum, u1); + input += 8 * stride; + } + + u0 = _mm_setzero_si128(); + in0 = _mm_unpacklo_epi16(u0, sum); + in1 = _mm_unpackhi_epi16(u0, sum); + in0 = _mm_srai_epi32(in0, 16); + in1 = _mm_srai_epi32(in1, 16); + + sum = _mm_add_epi32(in0, in1); + in0 = _mm_unpacklo_epi32(sum, u0); + in1 = _mm_unpackhi_epi32(sum, u0); + + sum = _mm_add_epi32(in0, in1); + in0 = _mm_srli_si128(sum, 8); + + in1 = _mm_add_epi32(sum, in0); + in1 = _mm_srai_epi32(in1, 1); + output[0] = (tran_low_t)_mm_cvtsi128_si32(in1); +} + +void aom_fdct32x32_1_sse2(const int16_t *input, tran_low_t *output, + int stride) { + __m128i in0, in1, in2, in3; + __m128i u0, u1; + __m128i sum = _mm_setzero_si128(); + int i; + + for (i = 0; i < 8; ++i) { + in0 = _mm_load_si128((const __m128i *)(input + 0)); + in1 = _mm_load_si128((const __m128i *)(input + 8)); + in2 = _mm_load_si128((const __m128i *)(input + 16)); + in3 = _mm_load_si128((const __m128i *)(input + 24)); + + input += stride; + u0 = _mm_add_epi16(in0, in1); + u1 = _mm_add_epi16(in2, in3); + sum = _mm_add_epi16(sum, u0); + + in0 = _mm_load_si128((const __m128i *)(input + 0)); + in1 = _mm_load_si128((const __m128i *)(input + 8)); + in2 = _mm_load_si128((const __m128i *)(input + 16)); + in3 = _mm_load_si128((const __m128i *)(input + 24)); + + input += stride; + sum = _mm_add_epi16(sum, u1); + u0 = _mm_add_epi16(in0, in1); + u1 = _mm_add_epi16(in2, in3); + sum = _mm_add_epi16(sum, u0); + + in0 = _mm_load_si128((const __m128i *)(input + 0)); + in1 = _mm_load_si128((const __m128i *)(input + 8)); + in2 = _mm_load_si128((const __m128i *)(input + 16)); + in3 = _mm_load_si128((const __m128i *)(input + 24)); + + input += stride; + sum = _mm_add_epi16(sum, u1); + u0 = _mm_add_epi16(in0, in1); + u1 = _mm_add_epi16(in2, in3); + sum = _mm_add_epi16(sum, u0); + + in0 = _mm_load_si128((const __m128i *)(input + 0)); + in1 = _mm_load_si128((const __m128i *)(input + 8)); + in2 = _mm_load_si128((const __m128i *)(input + 16)); + in3 = _mm_load_si128((const __m128i *)(input + 24)); + + input += stride; + sum = _mm_add_epi16(sum, u1); + u0 = _mm_add_epi16(in0, in1); + u1 = _mm_add_epi16(in2, in3); + sum = _mm_add_epi16(sum, u0); + + sum = _mm_add_epi16(sum, u1); + } + + u0 = _mm_setzero_si128(); + in0 = _mm_unpacklo_epi16(u0, sum); + in1 = _mm_unpackhi_epi16(u0, sum); + in0 = _mm_srai_epi32(in0, 16); + in1 = _mm_srai_epi32(in1, 16); + + sum = _mm_add_epi32(in0, in1); + in0 = _mm_unpacklo_epi32(sum, u0); + in1 = _mm_unpackhi_epi32(sum, u0); + + sum = _mm_add_epi32(in0, in1); + in0 = _mm_srli_si128(sum, 8); + + in1 = _mm_add_epi32(sum, in0); + in1 = _mm_srai_epi32(in1, 3); + output[0] = (tran_low_t)_mm_cvtsi128_si32(in1); +} + +#define DCT_HIGH_BIT_DEPTH 0 +#define FDCT4x4_2D aom_fdct4x4_sse2 +#define FDCT8x8_2D aom_fdct8x8_sse2 +#define FDCT16x16_2D aom_fdct16x16_sse2 +#include "aom_dsp/x86/fwd_txfm_impl_sse2.h" +#undef FDCT4x4_2D +#undef FDCT8x8_2D +#undef FDCT16x16_2D + +#define FDCT32x32_2D aom_fdct32x32_rd_sse2 +#define FDCT32x32_HIGH_PRECISION 0 +#include "aom_dsp/x86/fwd_dct32x32_impl_sse2.h" +#undef FDCT32x32_2D +#undef FDCT32x32_HIGH_PRECISION + +#define FDCT32x32_2D aom_fdct32x32_sse2 +#define FDCT32x32_HIGH_PRECISION 1 +#include "aom_dsp/x86/fwd_dct32x32_impl_sse2.h" // NOLINT +#undef FDCT32x32_2D +#undef FDCT32x32_HIGH_PRECISION +#undef DCT_HIGH_BIT_DEPTH + +#if CONFIG_HIGHBITDEPTH +#define DCT_HIGH_BIT_DEPTH 1 +#define FDCT4x4_2D aom_highbd_fdct4x4_sse2 +#define FDCT8x8_2D aom_highbd_fdct8x8_sse2 +#define FDCT16x16_2D aom_highbd_fdct16x16_sse2 +#include "aom_dsp/x86/fwd_txfm_impl_sse2.h" // NOLINT +#undef FDCT4x4_2D +#undef FDCT8x8_2D +#undef FDCT16x16_2D + +#define FDCT32x32_2D aom_highbd_fdct32x32_rd_sse2 +#define FDCT32x32_HIGH_PRECISION 0 +#include "aom_dsp/x86/fwd_dct32x32_impl_sse2.h" // NOLINT +#undef FDCT32x32_2D +#undef FDCT32x32_HIGH_PRECISION + +#define FDCT32x32_2D aom_highbd_fdct32x32_sse2 +#define FDCT32x32_HIGH_PRECISION 1 +#include "aom_dsp/x86/fwd_dct32x32_impl_sse2.h" // NOLINT +#undef FDCT32x32_2D +#undef FDCT32x32_HIGH_PRECISION +#undef DCT_HIGH_BIT_DEPTH +#endif // CONFIG_HIGHBITDEPTH diff --git a/third_party/aom/aom_dsp/x86/fwd_txfm_sse2.h b/third_party/aom/aom_dsp/x86/fwd_txfm_sse2.h new file mode 100644 index 000000000..26b2db2e0 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/fwd_txfm_sse2.h @@ -0,0 +1,362 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#ifndef AOM_DSP_X86_FWD_TXFM_SSE2_H_ +#define AOM_DSP_X86_FWD_TXFM_SSE2_H_ + +#include "aom_dsp/x86/txfm_common_intrin.h" + +#ifdef __cplusplus +extern "C" { +#endif + +#define pair_set_epi32(a, b) \ + _mm_set_epi32((int)(b), (int)(a), (int)(b), (int)(a)) + +static INLINE __m128i k_madd_epi32(__m128i a, __m128i b) { + __m128i buf0, buf1; + buf0 = _mm_mul_epu32(a, b); + a = _mm_srli_epi64(a, 32); + b = _mm_srli_epi64(b, 32); + buf1 = _mm_mul_epu32(a, b); + return _mm_add_epi64(buf0, buf1); +} + +static INLINE __m128i k_packs_epi64(__m128i a, __m128i b) { + __m128i buf0 = _mm_shuffle_epi32(a, _MM_SHUFFLE(0, 0, 2, 0)); + __m128i buf1 = _mm_shuffle_epi32(b, _MM_SHUFFLE(0, 0, 2, 0)); + return _mm_unpacklo_epi64(buf0, buf1); +} + +static INLINE int check_epi16_overflow_x2(const __m128i *preg0, + const __m128i *preg1) { + const __m128i max_overflow = _mm_set1_epi16(0x7fff); + const __m128i min_overflow = _mm_set1_epi16(0x8000); + __m128i cmp0 = _mm_or_si128(_mm_cmpeq_epi16(*preg0, max_overflow), + _mm_cmpeq_epi16(*preg0, min_overflow)); + __m128i cmp1 = _mm_or_si128(_mm_cmpeq_epi16(*preg1, max_overflow), + _mm_cmpeq_epi16(*preg1, min_overflow)); + cmp0 = _mm_or_si128(cmp0, cmp1); + return _mm_movemask_epi8(cmp0); +} + +static INLINE int check_epi16_overflow_x4(const __m128i *preg0, + const __m128i *preg1, + const __m128i *preg2, + const __m128i *preg3) { + const __m128i max_overflow = _mm_set1_epi16(0x7fff); + const __m128i min_overflow = _mm_set1_epi16(0x8000); + __m128i cmp0 = _mm_or_si128(_mm_cmpeq_epi16(*preg0, max_overflow), + _mm_cmpeq_epi16(*preg0, min_overflow)); + __m128i cmp1 = _mm_or_si128(_mm_cmpeq_epi16(*preg1, max_overflow), + _mm_cmpeq_epi16(*preg1, min_overflow)); + __m128i cmp2 = _mm_or_si128(_mm_cmpeq_epi16(*preg2, max_overflow), + _mm_cmpeq_epi16(*preg2, min_overflow)); + __m128i cmp3 = _mm_or_si128(_mm_cmpeq_epi16(*preg3, max_overflow), + _mm_cmpeq_epi16(*preg3, min_overflow)); + cmp0 = _mm_or_si128(_mm_or_si128(cmp0, cmp1), _mm_or_si128(cmp2, cmp3)); + return _mm_movemask_epi8(cmp0); +} + +static INLINE int check_epi16_overflow_x8( + const __m128i *preg0, const __m128i *preg1, const __m128i *preg2, + const __m128i *preg3, const __m128i *preg4, const __m128i *preg5, + const __m128i *preg6, const __m128i *preg7) { + int res0, res1; + res0 = check_epi16_overflow_x4(preg0, preg1, preg2, preg3); + res1 = check_epi16_overflow_x4(preg4, preg5, preg6, preg7); + return res0 + res1; +} + +static INLINE int check_epi16_overflow_x12( + const __m128i *preg0, const __m128i *preg1, const __m128i *preg2, + const __m128i *preg3, const __m128i *preg4, const __m128i *preg5, + const __m128i *preg6, const __m128i *preg7, const __m128i *preg8, + const __m128i *preg9, const __m128i *preg10, const __m128i *preg11) { + int res0, res1; + res0 = check_epi16_overflow_x4(preg0, preg1, preg2, preg3); + res1 = check_epi16_overflow_x4(preg4, preg5, preg6, preg7); + if (!res0) res0 = check_epi16_overflow_x4(preg8, preg9, preg10, preg11); + return res0 + res1; +} + +static INLINE int check_epi16_overflow_x16( + const __m128i *preg0, const __m128i *preg1, const __m128i *preg2, + const __m128i *preg3, const __m128i *preg4, const __m128i *preg5, + const __m128i *preg6, const __m128i *preg7, const __m128i *preg8, + const __m128i *preg9, const __m128i *preg10, const __m128i *preg11, + const __m128i *preg12, const __m128i *preg13, const __m128i *preg14, + const __m128i *preg15) { + int res0, res1; + res0 = check_epi16_overflow_x4(preg0, preg1, preg2, preg3); + res1 = check_epi16_overflow_x4(preg4, preg5, preg6, preg7); + if (!res0) { + res0 = check_epi16_overflow_x4(preg8, preg9, preg10, preg11); + if (!res1) res1 = check_epi16_overflow_x4(preg12, preg13, preg14, preg15); + } + return res0 + res1; +} + +static INLINE int check_epi16_overflow_x32( + const __m128i *preg0, const __m128i *preg1, const __m128i *preg2, + const __m128i *preg3, const __m128i *preg4, const __m128i *preg5, + const __m128i *preg6, const __m128i *preg7, const __m128i *preg8, + const __m128i *preg9, const __m128i *preg10, const __m128i *preg11, + const __m128i *preg12, const __m128i *preg13, const __m128i *preg14, + const __m128i *preg15, const __m128i *preg16, const __m128i *preg17, + const __m128i *preg18, const __m128i *preg19, const __m128i *preg20, + const __m128i *preg21, const __m128i *preg22, const __m128i *preg23, + const __m128i *preg24, const __m128i *preg25, const __m128i *preg26, + const __m128i *preg27, const __m128i *preg28, const __m128i *preg29, + const __m128i *preg30, const __m128i *preg31) { + int res0, res1; + res0 = check_epi16_overflow_x4(preg0, preg1, preg2, preg3); + res1 = check_epi16_overflow_x4(preg4, preg5, preg6, preg7); + if (!res0) { + res0 = check_epi16_overflow_x4(preg8, preg9, preg10, preg11); + if (!res1) { + res1 = check_epi16_overflow_x4(preg12, preg13, preg14, preg15); + if (!res0) { + res0 = check_epi16_overflow_x4(preg16, preg17, preg18, preg19); + if (!res1) { + res1 = check_epi16_overflow_x4(preg20, preg21, preg22, preg23); + if (!res0) { + res0 = check_epi16_overflow_x4(preg24, preg25, preg26, preg27); + if (!res1) + res1 = check_epi16_overflow_x4(preg28, preg29, preg30, preg31); + } + } + } + } + } + return res0 + res1; +} + +static INLINE int k_check_epi32_overflow_4(const __m128i *preg0, + const __m128i *preg1, + const __m128i *preg2, + const __m128i *preg3, + const __m128i *zero) { + __m128i minus_one = _mm_set1_epi32(-1); + // Check for overflows + __m128i reg0_shifted = _mm_slli_epi64(*preg0, 1); + __m128i reg1_shifted = _mm_slli_epi64(*preg1, 1); + __m128i reg2_shifted = _mm_slli_epi64(*preg2, 1); + __m128i reg3_shifted = _mm_slli_epi64(*preg3, 1); + __m128i reg0_top_dwords = + _mm_shuffle_epi32(reg0_shifted, _MM_SHUFFLE(0, 0, 3, 1)); + __m128i reg1_top_dwords = + _mm_shuffle_epi32(reg1_shifted, _MM_SHUFFLE(0, 0, 3, 1)); + __m128i reg2_top_dwords = + _mm_shuffle_epi32(reg2_shifted, _MM_SHUFFLE(0, 0, 3, 1)); + __m128i reg3_top_dwords = + _mm_shuffle_epi32(reg3_shifted, _MM_SHUFFLE(0, 0, 3, 1)); + __m128i top_dwords_01 = _mm_unpacklo_epi64(reg0_top_dwords, reg1_top_dwords); + __m128i top_dwords_23 = _mm_unpacklo_epi64(reg2_top_dwords, reg3_top_dwords); + __m128i valid_positve_01 = _mm_cmpeq_epi32(top_dwords_01, *zero); + __m128i valid_positve_23 = _mm_cmpeq_epi32(top_dwords_23, *zero); + __m128i valid_negative_01 = _mm_cmpeq_epi32(top_dwords_01, minus_one); + __m128i valid_negative_23 = _mm_cmpeq_epi32(top_dwords_23, minus_one); + int overflow_01 = + _mm_movemask_epi8(_mm_cmpeq_epi32(valid_positve_01, valid_negative_01)); + int overflow_23 = + _mm_movemask_epi8(_mm_cmpeq_epi32(valid_positve_23, valid_negative_23)); + return (overflow_01 + overflow_23); +} + +static INLINE int k_check_epi32_overflow_8( + const __m128i *preg0, const __m128i *preg1, const __m128i *preg2, + const __m128i *preg3, const __m128i *preg4, const __m128i *preg5, + const __m128i *preg6, const __m128i *preg7, const __m128i *zero) { + int overflow = k_check_epi32_overflow_4(preg0, preg1, preg2, preg3, zero); + if (!overflow) { + overflow = k_check_epi32_overflow_4(preg4, preg5, preg6, preg7, zero); + } + return overflow; +} + +static INLINE int k_check_epi32_overflow_16( + const __m128i *preg0, const __m128i *preg1, const __m128i *preg2, + const __m128i *preg3, const __m128i *preg4, const __m128i *preg5, + const __m128i *preg6, const __m128i *preg7, const __m128i *preg8, + const __m128i *preg9, const __m128i *preg10, const __m128i *preg11, + const __m128i *preg12, const __m128i *preg13, const __m128i *preg14, + const __m128i *preg15, const __m128i *zero) { + int overflow = k_check_epi32_overflow_4(preg0, preg1, preg2, preg3, zero); + if (!overflow) { + overflow = k_check_epi32_overflow_4(preg4, preg5, preg6, preg7, zero); + if (!overflow) { + overflow = k_check_epi32_overflow_4(preg8, preg9, preg10, preg11, zero); + if (!overflow) { + overflow = + k_check_epi32_overflow_4(preg12, preg13, preg14, preg15, zero); + } + } + } + return overflow; +} + +static INLINE int k_check_epi32_overflow_32( + const __m128i *preg0, const __m128i *preg1, const __m128i *preg2, + const __m128i *preg3, const __m128i *preg4, const __m128i *preg5, + const __m128i *preg6, const __m128i *preg7, const __m128i *preg8, + const __m128i *preg9, const __m128i *preg10, const __m128i *preg11, + const __m128i *preg12, const __m128i *preg13, const __m128i *preg14, + const __m128i *preg15, const __m128i *preg16, const __m128i *preg17, + const __m128i *preg18, const __m128i *preg19, const __m128i *preg20, + const __m128i *preg21, const __m128i *preg22, const __m128i *preg23, + const __m128i *preg24, const __m128i *preg25, const __m128i *preg26, + const __m128i *preg27, const __m128i *preg28, const __m128i *preg29, + const __m128i *preg30, const __m128i *preg31, const __m128i *zero) { + int overflow = k_check_epi32_overflow_4(preg0, preg1, preg2, preg3, zero); + if (!overflow) { + overflow = k_check_epi32_overflow_4(preg4, preg5, preg6, preg7, zero); + if (!overflow) { + overflow = k_check_epi32_overflow_4(preg8, preg9, preg10, preg11, zero); + if (!overflow) { + overflow = + k_check_epi32_overflow_4(preg12, preg13, preg14, preg15, zero); + if (!overflow) { + overflow = + k_check_epi32_overflow_4(preg16, preg17, preg18, preg19, zero); + if (!overflow) { + overflow = + k_check_epi32_overflow_4(preg20, preg21, preg22, preg23, zero); + if (!overflow) { + overflow = k_check_epi32_overflow_4(preg24, preg25, preg26, + preg27, zero); + if (!overflow) { + overflow = k_check_epi32_overflow_4(preg28, preg29, preg30, + preg31, zero); + } + } + } + } + } + } + } + return overflow; +} + +static INLINE void store_output(const __m128i *poutput, tran_low_t *dst_ptr) { +#if CONFIG_HIGHBITDEPTH + const __m128i zero = _mm_setzero_si128(); + const __m128i sign_bits = _mm_cmplt_epi16(*poutput, zero); + __m128i out0 = _mm_unpacklo_epi16(*poutput, sign_bits); + __m128i out1 = _mm_unpackhi_epi16(*poutput, sign_bits); + _mm_store_si128((__m128i *)(dst_ptr), out0); + _mm_store_si128((__m128i *)(dst_ptr + 4), out1); +#else + _mm_store_si128((__m128i *)(dst_ptr), *poutput); +#endif // CONFIG_HIGHBITDEPTH +} + +static INLINE __m128i mult_round_shift(const __m128i *pin0, const __m128i *pin1, + const __m128i *pmultiplier, + const __m128i *prounding, int shift) { + const __m128i u0 = _mm_madd_epi16(*pin0, *pmultiplier); + const __m128i u1 = _mm_madd_epi16(*pin1, *pmultiplier); + const __m128i v0 = _mm_add_epi32(u0, *prounding); + const __m128i v1 = _mm_add_epi32(u1, *prounding); + const __m128i w0 = _mm_srai_epi32(v0, shift); + const __m128i w1 = _mm_srai_epi32(v1, shift); + return _mm_packs_epi32(w0, w1); +} + +static INLINE void transpose_and_output8x8( + const __m128i *pin00, const __m128i *pin01, const __m128i *pin02, + const __m128i *pin03, const __m128i *pin04, const __m128i *pin05, + const __m128i *pin06, const __m128i *pin07, int pass, int16_t *out0_ptr, + tran_low_t *out1_ptr) { + // 00 01 02 03 04 05 06 07 + // 10 11 12 13 14 15 16 17 + // 20 21 22 23 24 25 26 27 + // 30 31 32 33 34 35 36 37 + // 40 41 42 43 44 45 46 47 + // 50 51 52 53 54 55 56 57 + // 60 61 62 63 64 65 66 67 + // 70 71 72 73 74 75 76 77 + const __m128i tr0_0 = _mm_unpacklo_epi16(*pin00, *pin01); + const __m128i tr0_1 = _mm_unpacklo_epi16(*pin02, *pin03); + const __m128i tr0_2 = _mm_unpackhi_epi16(*pin00, *pin01); + const __m128i tr0_3 = _mm_unpackhi_epi16(*pin02, *pin03); + const __m128i tr0_4 = _mm_unpacklo_epi16(*pin04, *pin05); + const __m128i tr0_5 = _mm_unpacklo_epi16(*pin06, *pin07); + const __m128i tr0_6 = _mm_unpackhi_epi16(*pin04, *pin05); + const __m128i tr0_7 = _mm_unpackhi_epi16(*pin06, *pin07); + // 00 10 01 11 02 12 03 13 + // 20 30 21 31 22 32 23 33 + // 04 14 05 15 06 16 07 17 + // 24 34 25 35 26 36 27 37 + // 40 50 41 51 42 52 43 53 + // 60 70 61 71 62 72 63 73 + // 54 54 55 55 56 56 57 57 + // 64 74 65 75 66 76 67 77 + const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1); + const __m128i tr1_1 = _mm_unpacklo_epi32(tr0_2, tr0_3); + const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1); + const __m128i tr1_3 = _mm_unpackhi_epi32(tr0_2, tr0_3); + const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_4, tr0_5); + const __m128i tr1_5 = _mm_unpacklo_epi32(tr0_6, tr0_7); + const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_4, tr0_5); + const __m128i tr1_7 = _mm_unpackhi_epi32(tr0_6, tr0_7); + // 00 10 20 30 01 11 21 31 + // 40 50 60 70 41 51 61 71 + // 02 12 22 32 03 13 23 33 + // 42 52 62 72 43 53 63 73 + // 04 14 24 34 05 15 21 36 + // 44 54 64 74 45 55 61 76 + // 06 16 26 36 07 17 27 37 + // 46 56 66 76 47 57 67 77 + const __m128i tr2_0 = _mm_unpacklo_epi64(tr1_0, tr1_4); + const __m128i tr2_1 = _mm_unpackhi_epi64(tr1_0, tr1_4); + const __m128i tr2_2 = _mm_unpacklo_epi64(tr1_2, tr1_6); + const __m128i tr2_3 = _mm_unpackhi_epi64(tr1_2, tr1_6); + const __m128i tr2_4 = _mm_unpacklo_epi64(tr1_1, tr1_5); + const __m128i tr2_5 = _mm_unpackhi_epi64(tr1_1, tr1_5); + const __m128i tr2_6 = _mm_unpacklo_epi64(tr1_3, tr1_7); + const __m128i tr2_7 = _mm_unpackhi_epi64(tr1_3, tr1_7); + // 00 10 20 30 40 50 60 70 + // 01 11 21 31 41 51 61 71 + // 02 12 22 32 42 52 62 72 + // 03 13 23 33 43 53 63 73 + // 04 14 24 34 44 54 64 74 + // 05 15 25 35 45 55 65 75 + // 06 16 26 36 46 56 66 76 + // 07 17 27 37 47 57 67 77 + if (pass == 0) { + _mm_storeu_si128((__m128i *)(out0_ptr + 0 * 16), tr2_0); + _mm_storeu_si128((__m128i *)(out0_ptr + 1 * 16), tr2_1); + _mm_storeu_si128((__m128i *)(out0_ptr + 2 * 16), tr2_2); + _mm_storeu_si128((__m128i *)(out0_ptr + 3 * 16), tr2_3); + _mm_storeu_si128((__m128i *)(out0_ptr + 4 * 16), tr2_4); + _mm_storeu_si128((__m128i *)(out0_ptr + 5 * 16), tr2_5); + _mm_storeu_si128((__m128i *)(out0_ptr + 6 * 16), tr2_6); + _mm_storeu_si128((__m128i *)(out0_ptr + 7 * 16), tr2_7); + } else { + storeu_output(&tr2_0, (out1_ptr + 0 * 16)); + storeu_output(&tr2_1, (out1_ptr + 1 * 16)); + storeu_output(&tr2_2, (out1_ptr + 2 * 16)); + storeu_output(&tr2_3, (out1_ptr + 3 * 16)); + storeu_output(&tr2_4, (out1_ptr + 4 * 16)); + storeu_output(&tr2_5, (out1_ptr + 5 * 16)); + storeu_output(&tr2_6, (out1_ptr + 6 * 16)); + storeu_output(&tr2_7, (out1_ptr + 7 * 16)); + } +} + +void fdct32_8col(__m128i *in0, __m128i *in1); + +#ifdef __cplusplus +} // extern "C" +#endif + +#endif // AOM_DSP_X86_FWD_TXFM_SSE2_H_ diff --git a/third_party/aom/aom_dsp/x86/fwd_txfm_ssse3_x86_64.asm b/third_party/aom/aom_dsp/x86/fwd_txfm_ssse3_x86_64.asm new file mode 100644 index 000000000..8fa1c04d0 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/fwd_txfm_ssse3_x86_64.asm @@ -0,0 +1,204 @@ +; +; Copyright (c) 2016, Alliance for Open Media. All rights reserved +; +; This source code is subject to the terms of the BSD 2 Clause License and +; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License +; was not distributed with this source code in the LICENSE file, you can +; obtain it at www.aomedia.org/license/software. If the Alliance for Open +; Media Patent License 1.0 was not distributed with this source code in the +; PATENTS file, you can obtain it at www.aomedia.org/license/patent. +; + +; + +%include "third_party/x86inc/x86inc.asm" + +; This file provides SSSE3 version of the forward transformation. Part +; of the macro definitions are originally derived from the ffmpeg project. +; The current version applies to x86 64-bit only. + +SECTION_RODATA + +pw_11585x2: times 8 dw 23170 +pd_8192: times 4 dd 8192 + +%macro TRANSFORM_COEFFS 2 +pw_%1_%2: dw %1, %2, %1, %2, %1, %2, %1, %2 +pw_%2_m%1: dw %2, -%1, %2, -%1, %2, -%1, %2, -%1 +%endmacro + +TRANSFORM_COEFFS 11585, 11585 +TRANSFORM_COEFFS 15137, 6270 +TRANSFORM_COEFFS 16069, 3196 +TRANSFORM_COEFFS 9102, 13623 + +SECTION .text + +%if ARCH_X86_64 +%macro SUM_SUB 3 + psubw m%3, m%1, m%2 + paddw m%1, m%2 + SWAP %2, %3 +%endmacro + +; butterfly operation +%macro MUL_ADD_2X 6 ; dst1, dst2, src, round, coefs1, coefs2 + pmaddwd m%1, m%3, %5 + pmaddwd m%2, m%3, %6 + paddd m%1, %4 + paddd m%2, %4 + psrad m%1, 14 + psrad m%2, 14 +%endmacro + +%macro BUTTERFLY_4X 7 ; dst1, dst2, coef1, coef2, round, tmp1, tmp2 + punpckhwd m%6, m%2, m%1 + MUL_ADD_2X %7, %6, %6, %5, [pw_%4_%3], [pw_%3_m%4] + punpcklwd m%2, m%1 + MUL_ADD_2X %1, %2, %2, %5, [pw_%4_%3], [pw_%3_m%4] + packssdw m%1, m%7 + packssdw m%2, m%6 +%endmacro + +; matrix transpose +%macro INTERLEAVE_2X 4 + punpckh%1 m%4, m%2, m%3 + punpckl%1 m%2, m%3 + SWAP %3, %4 +%endmacro + +%macro TRANSPOSE8X8 9 + INTERLEAVE_2X wd, %1, %2, %9 + INTERLEAVE_2X wd, %3, %4, %9 + INTERLEAVE_2X wd, %5, %6, %9 + INTERLEAVE_2X wd, %7, %8, %9 + + INTERLEAVE_2X dq, %1, %3, %9 + INTERLEAVE_2X dq, %2, %4, %9 + INTERLEAVE_2X dq, %5, %7, %9 + INTERLEAVE_2X dq, %6, %8, %9 + + INTERLEAVE_2X qdq, %1, %5, %9 + INTERLEAVE_2X qdq, %3, %7, %9 + INTERLEAVE_2X qdq, %2, %6, %9 + INTERLEAVE_2X qdq, %4, %8, %9 + + SWAP %2, %5 + SWAP %4, %7 +%endmacro + +; 1D forward 8x8 DCT transform +%macro FDCT8_1D 1 + SUM_SUB 0, 7, 9 + SUM_SUB 1, 6, 9 + SUM_SUB 2, 5, 9 + SUM_SUB 3, 4, 9 + + SUM_SUB 0, 3, 9 + SUM_SUB 1, 2, 9 + SUM_SUB 6, 5, 9 +%if %1 == 0 + SUM_SUB 0, 1, 9 +%endif + + BUTTERFLY_4X 2, 3, 6270, 15137, m8, 9, 10 + + pmulhrsw m6, m12 + pmulhrsw m5, m12 +%if %1 == 0 + pmulhrsw m0, m12 + pmulhrsw m1, m12 +%else + BUTTERFLY_4X 1, 0, 11585, 11585, m8, 9, 10 + SWAP 0, 1 +%endif + + SUM_SUB 4, 5, 9 + SUM_SUB 7, 6, 9 + BUTTERFLY_4X 4, 7, 3196, 16069, m8, 9, 10 + BUTTERFLY_4X 5, 6, 13623, 9102, m8, 9, 10 + SWAP 1, 4 + SWAP 3, 6 +%endmacro + +%macro DIVIDE_ROUND_2X 4 ; dst1, dst2, tmp1, tmp2 + psraw m%3, m%1, 15 + psraw m%4, m%2, 15 + psubw m%1, m%3 + psubw m%2, m%4 + psraw m%1, 1 + psraw m%2, 1 +%endmacro + +%macro STORE_OUTPUT 2 ; index, result +%if CONFIG_HIGHBITDEPTH + ; const __m128i sign_bits = _mm_cmplt_epi16(*poutput, zero); + ; __m128i out0 = _mm_unpacklo_epi16(*poutput, sign_bits); + ; __m128i out1 = _mm_unpackhi_epi16(*poutput, sign_bits); + ; _mm_store_si128((__m128i *)(dst_ptr), out0); + ; _mm_store_si128((__m128i *)(dst_ptr + 4), out1); + pxor m11, m11 + pcmpgtw m11, m%2 + movdqa m12, m%2 + punpcklwd m%2, m11 + punpckhwd m12, m11 + mova [outputq + 4*%1 + 0], m%2 + mova [outputq + 4*%1 + 16], m12 +%else + mova [outputq + 2*%1], m%2 +%endif +%endmacro + +INIT_XMM ssse3 +cglobal fdct8x8, 3, 5, 13, input, output, stride + + mova m8, [pd_8192] + mova m12, [pw_11585x2] + + lea r3, [2 * strideq] + lea r4, [4 * strideq] + mova m0, [inputq] + mova m1, [inputq + r3] + lea inputq, [inputq + r4] + mova m2, [inputq] + mova m3, [inputq + r3] + lea inputq, [inputq + r4] + mova m4, [inputq] + mova m5, [inputq + r3] + lea inputq, [inputq + r4] + mova m6, [inputq] + mova m7, [inputq + r3] + + ; left shift by 2 to increase forward transformation precision + psllw m0, 2 + psllw m1, 2 + psllw m2, 2 + psllw m3, 2 + psllw m4, 2 + psllw m5, 2 + psllw m6, 2 + psllw m7, 2 + + ; column transform + FDCT8_1D 0 + TRANSPOSE8X8 0, 1, 2, 3, 4, 5, 6, 7, 9 + + FDCT8_1D 1 + TRANSPOSE8X8 0, 1, 2, 3, 4, 5, 6, 7, 9 + + DIVIDE_ROUND_2X 0, 1, 9, 10 + DIVIDE_ROUND_2X 2, 3, 9, 10 + DIVIDE_ROUND_2X 4, 5, 9, 10 + DIVIDE_ROUND_2X 6, 7, 9, 10 + + STORE_OUTPUT 0, 0 + STORE_OUTPUT 8, 1 + STORE_OUTPUT 16, 2 + STORE_OUTPUT 24, 3 + STORE_OUTPUT 32, 4 + STORE_OUTPUT 40, 5 + STORE_OUTPUT 48, 6 + STORE_OUTPUT 56, 7 + + RET +%endif diff --git a/third_party/aom/aom_dsp/x86/halfpix_variance_impl_sse2.asm b/third_party/aom/aom_dsp/x86/halfpix_variance_impl_sse2.asm new file mode 100644 index 000000000..60446b086 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/halfpix_variance_impl_sse2.asm @@ -0,0 +1,349 @@ +; +; Copyright (c) 2016, Alliance for Open Media. All rights reserved +; +; This source code is subject to the terms of the BSD 2 Clause License and +; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License +; was not distributed with this source code in the LICENSE file, you can +; obtain it at www.aomedia.org/license/software. If the Alliance for Open +; Media Patent License 1.0 was not distributed with this source code in the +; PATENTS file, you can obtain it at www.aomedia.org/license/patent. +; + +; + +%include "aom_ports/x86_abi_support.asm" + +;void aom_half_horiz_vert_variance16x_h_sse2(unsigned char *ref, +; int ref_stride, +; unsigned char *src, +; int src_stride, +; unsigned int height, +; int *sum, +; unsigned int *sumsquared) +global sym(aom_half_horiz_vert_variance16x_h_sse2) PRIVATE +sym(aom_half_horiz_vert_variance16x_h_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 7 + SAVE_XMM 7 + GET_GOT rbx + push rsi + push rdi + ; end prolog + + pxor xmm6, xmm6 ; error accumulator + pxor xmm7, xmm7 ; sse eaccumulator + mov rsi, arg(0) ;ref + + mov rdi, arg(2) ;src + movsxd rcx, dword ptr arg(4) ;height + movsxd rax, dword ptr arg(1) ;ref_stride + movsxd rdx, dword ptr arg(3) ;src_stride + + pxor xmm0, xmm0 ; + + movdqu xmm5, XMMWORD PTR [rsi] + movdqu xmm3, XMMWORD PTR [rsi+1] + pavgb xmm5, xmm3 ; xmm5 = avg(xmm1,xmm3) horizontal line 1 + + lea rsi, [rsi + rax] + +aom_half_horiz_vert_variance16x_h_1: + movdqu xmm1, XMMWORD PTR [rsi] ; + movdqu xmm2, XMMWORD PTR [rsi+1] ; + pavgb xmm1, xmm2 ; xmm1 = avg(xmm1,xmm3) horizontal line i+1 + + pavgb xmm5, xmm1 ; xmm = vertical average of the above + + movdqa xmm4, xmm5 + punpcklbw xmm5, xmm0 ; xmm5 = words of above + punpckhbw xmm4, xmm0 + + movq xmm3, QWORD PTR [rdi] ; xmm3 = d0,d1,d2..d7 + punpcklbw xmm3, xmm0 ; xmm3 = words of above + psubw xmm5, xmm3 ; xmm5 -= xmm3 + + movq xmm3, QWORD PTR [rdi+8] + punpcklbw xmm3, xmm0 + psubw xmm4, xmm3 + + paddw xmm6, xmm5 ; xmm6 += accumulated column differences + paddw xmm6, xmm4 + pmaddwd xmm5, xmm5 ; xmm5 *= xmm5 + pmaddwd xmm4, xmm4 + paddd xmm7, xmm5 ; xmm7 += accumulated square column differences + paddd xmm7, xmm4 + + movdqa xmm5, xmm1 ; save xmm1 for use on the next row + + lea rsi, [rsi + rax] + lea rdi, [rdi + rdx] + + sub rcx, 1 ; + jnz aom_half_horiz_vert_variance16x_h_1 ; + + pxor xmm1, xmm1 + pxor xmm5, xmm5 + + punpcklwd xmm0, xmm6 + punpckhwd xmm1, xmm6 + psrad xmm0, 16 + psrad xmm1, 16 + paddd xmm0, xmm1 + movdqa xmm1, xmm0 + + movdqa xmm6, xmm7 + punpckldq xmm6, xmm5 + punpckhdq xmm7, xmm5 + paddd xmm6, xmm7 + + punpckldq xmm0, xmm5 + punpckhdq xmm1, xmm5 + paddd xmm0, xmm1 + + movdqa xmm7, xmm6 + movdqa xmm1, xmm0 + + psrldq xmm7, 8 + psrldq xmm1, 8 + + paddd xmm6, xmm7 + paddd xmm0, xmm1 + + mov rsi, arg(5) ;[Sum] + mov rdi, arg(6) ;[SSE] + + movd [rsi], xmm0 + movd [rdi], xmm6 + + ; begin epilog + pop rdi + pop rsi + RESTORE_GOT + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret + + +;void aom_half_vert_variance16x_h_sse2(unsigned char *ref, +; int ref_stride, +; unsigned char *src, +; int src_stride, +; unsigned int height, +; int *sum, +; unsigned int *sumsquared) +global sym(aom_half_vert_variance16x_h_sse2) PRIVATE +sym(aom_half_vert_variance16x_h_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 7 + SAVE_XMM 7 + GET_GOT rbx + push rsi + push rdi + ; end prolog + + pxor xmm6, xmm6 ; error accumulator + pxor xmm7, xmm7 ; sse eaccumulator + mov rsi, arg(0) ;ref + + mov rdi, arg(2) ;src + movsxd rcx, dword ptr arg(4) ;height + movsxd rax, dword ptr arg(1) ;ref_stride + movsxd rdx, dword ptr arg(3) ;src_stride + + movdqu xmm5, XMMWORD PTR [rsi] + lea rsi, [rsi + rax ] + pxor xmm0, xmm0 + +aom_half_vert_variance16x_h_1: + movdqu xmm3, XMMWORD PTR [rsi] + + pavgb xmm5, xmm3 ; xmm5 = avg(xmm1,xmm3) + movdqa xmm4, xmm5 + punpcklbw xmm5, xmm0 + punpckhbw xmm4, xmm0 + + movq xmm2, QWORD PTR [rdi] + punpcklbw xmm2, xmm0 + psubw xmm5, xmm2 + movq xmm2, QWORD PTR [rdi+8] + punpcklbw xmm2, xmm0 + psubw xmm4, xmm2 + + paddw xmm6, xmm5 ; xmm6 += accumulated column differences + paddw xmm6, xmm4 + pmaddwd xmm5, xmm5 ; xmm5 *= xmm5 + pmaddwd xmm4, xmm4 + paddd xmm7, xmm5 ; xmm7 += accumulated square column differences + paddd xmm7, xmm4 + + movdqa xmm5, xmm3 + + lea rsi, [rsi + rax] + lea rdi, [rdi + rdx] + + sub rcx, 1 + jnz aom_half_vert_variance16x_h_1 + + pxor xmm1, xmm1 + pxor xmm5, xmm5 + + punpcklwd xmm0, xmm6 + punpckhwd xmm1, xmm6 + psrad xmm0, 16 + psrad xmm1, 16 + paddd xmm0, xmm1 + movdqa xmm1, xmm0 + + movdqa xmm6, xmm7 + punpckldq xmm6, xmm5 + punpckhdq xmm7, xmm5 + paddd xmm6, xmm7 + + punpckldq xmm0, xmm5 + punpckhdq xmm1, xmm5 + paddd xmm0, xmm1 + + movdqa xmm7, xmm6 + movdqa xmm1, xmm0 + + psrldq xmm7, 8 + psrldq xmm1, 8 + + paddd xmm6, xmm7 + paddd xmm0, xmm1 + + mov rsi, arg(5) ;[Sum] + mov rdi, arg(6) ;[SSE] + + movd [rsi], xmm0 + movd [rdi], xmm6 + + ; begin epilog + pop rdi + pop rsi + RESTORE_GOT + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret + + +;void aom_half_horiz_variance16x_h_sse2(unsigned char *ref, +; int ref_stride +; unsigned char *src, +; int src_stride, +; unsigned int height, +; int *sum, +; unsigned int *sumsquared) +global sym(aom_half_horiz_variance16x_h_sse2) PRIVATE +sym(aom_half_horiz_variance16x_h_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 7 + SAVE_XMM 7 + GET_GOT rbx + push rsi + push rdi + ; end prolog + + pxor xmm6, xmm6 ; error accumulator + pxor xmm7, xmm7 ; sse eaccumulator + mov rsi, arg(0) ;ref + + mov rdi, arg(2) ;src + movsxd rcx, dword ptr arg(4) ;height + movsxd rax, dword ptr arg(1) ;ref_stride + movsxd rdx, dword ptr arg(3) ;src_stride + + pxor xmm0, xmm0 ; + +aom_half_horiz_variance16x_h_1: + movdqu xmm5, XMMWORD PTR [rsi] ; xmm5 = s0,s1,s2..s15 + movdqu xmm3, XMMWORD PTR [rsi+1] ; xmm3 = s1,s2,s3..s16 + + pavgb xmm5, xmm3 ; xmm5 = avg(xmm1,xmm3) + movdqa xmm1, xmm5 + punpcklbw xmm5, xmm0 ; xmm5 = words of above + punpckhbw xmm1, xmm0 + + movq xmm3, QWORD PTR [rdi] ; xmm3 = d0,d1,d2..d7 + punpcklbw xmm3, xmm0 ; xmm3 = words of above + movq xmm2, QWORD PTR [rdi+8] + punpcklbw xmm2, xmm0 + + psubw xmm5, xmm3 ; xmm5 -= xmm3 + psubw xmm1, xmm2 + paddw xmm6, xmm5 ; xmm6 += accumulated column differences + paddw xmm6, xmm1 + pmaddwd xmm5, xmm5 ; xmm5 *= xmm5 + pmaddwd xmm1, xmm1 + paddd xmm7, xmm5 ; xmm7 += accumulated square column differences + paddd xmm7, xmm1 + + lea rsi, [rsi + rax] + lea rdi, [rdi + rdx] + + sub rcx, 1 ; + jnz aom_half_horiz_variance16x_h_1 ; + + pxor xmm1, xmm1 + pxor xmm5, xmm5 + + punpcklwd xmm0, xmm6 + punpckhwd xmm1, xmm6 + psrad xmm0, 16 + psrad xmm1, 16 + paddd xmm0, xmm1 + movdqa xmm1, xmm0 + + movdqa xmm6, xmm7 + punpckldq xmm6, xmm5 + punpckhdq xmm7, xmm5 + paddd xmm6, xmm7 + + punpckldq xmm0, xmm5 + punpckhdq xmm1, xmm5 + paddd xmm0, xmm1 + + movdqa xmm7, xmm6 + movdqa xmm1, xmm0 + + psrldq xmm7, 8 + psrldq xmm1, 8 + + paddd xmm6, xmm7 + paddd xmm0, xmm1 + + mov rsi, arg(5) ;[Sum] + mov rdi, arg(6) ;[SSE] + + movd [rsi], xmm0 + movd [rdi], xmm6 + + ; begin epilog + pop rdi + pop rsi + RESTORE_GOT + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret + +SECTION_RODATA +; short xmm_bi_rd[8] = { 64, 64, 64, 64,64, 64, 64, 64}; +align 16 +xmm_bi_rd: + times 8 dw 64 +align 16 +aom_bilinear_filters_sse2: + dw 128, 128, 128, 128, 128, 128, 128, 128, 0, 0, 0, 0, 0, 0, 0, 0 + dw 112, 112, 112, 112, 112, 112, 112, 112, 16, 16, 16, 16, 16, 16, 16, 16 + dw 96, 96, 96, 96, 96, 96, 96, 96, 32, 32, 32, 32, 32, 32, 32, 32 + dw 80, 80, 80, 80, 80, 80, 80, 80, 48, 48, 48, 48, 48, 48, 48, 48 + dw 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64 + dw 48, 48, 48, 48, 48, 48, 48, 48, 80, 80, 80, 80, 80, 80, 80, 80 + dw 32, 32, 32, 32, 32, 32, 32, 32, 96, 96, 96, 96, 96, 96, 96, 96 + dw 16, 16, 16, 16, 16, 16, 16, 16, 112, 112, 112, 112, 112, 112, 112, 112 diff --git a/third_party/aom/aom_dsp/x86/halfpix_variance_sse2.c b/third_party/aom/aom_dsp/x86/halfpix_variance_sse2.c new file mode 100644 index 000000000..a99c0b40e --- /dev/null +++ b/third_party/aom/aom_dsp/x86/halfpix_variance_sse2.c @@ -0,0 +1,77 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include <assert.h> + +#include "./aom_config.h" +#include "./aom_dsp_rtcd.h" +#include "aom/aom_integer.h" + +void aom_half_horiz_vert_variance16x_h_sse2(const unsigned char *ref, + int ref_stride, + const unsigned char *src, + int src_stride, unsigned int height, + int *sum, unsigned int *sumsquared); +void aom_half_horiz_variance16x_h_sse2(const unsigned char *ref, int ref_stride, + const unsigned char *src, int src_stride, + unsigned int height, int *sum, + unsigned int *sumsquared); +void aom_half_vert_variance16x_h_sse2(const unsigned char *ref, int ref_stride, + const unsigned char *src, int src_stride, + unsigned int height, int *sum, + unsigned int *sumsquared); + +uint32_t aom_variance_halfpixvar16x16_h_sse2(const unsigned char *src, + int src_stride, + const unsigned char *dst, + int dst_stride, uint32_t *sse) { + int xsum0; + unsigned int xxsum0; + + aom_half_horiz_variance16x_h_sse2(src, src_stride, dst, dst_stride, 16, + &xsum0, &xxsum0); + + *sse = xxsum0; + assert(xsum0 <= 255 * 16 * 16); + assert(xsum0 >= -255 * 16 * 16); + return (xxsum0 - ((uint32_t)((int64_t)xsum0 * xsum0) >> 8)); +} + +uint32_t aom_variance_halfpixvar16x16_v_sse2(const unsigned char *src, + int src_stride, + const unsigned char *dst, + int dst_stride, uint32_t *sse) { + int xsum0; + unsigned int xxsum0; + aom_half_vert_variance16x_h_sse2(src, src_stride, dst, dst_stride, 16, &xsum0, + &xxsum0); + + *sse = xxsum0; + assert(xsum0 <= 255 * 16 * 16); + assert(xsum0 >= -255 * 16 * 16); + return (xxsum0 - ((uint32_t)((int64_t)xsum0 * xsum0) >> 8)); +} + +uint32_t aom_variance_halfpixvar16x16_hv_sse2(const unsigned char *src, + int src_stride, + const unsigned char *dst, + int dst_stride, uint32_t *sse) { + int xsum0; + unsigned int xxsum0; + + aom_half_horiz_vert_variance16x_h_sse2(src, src_stride, dst, dst_stride, 16, + &xsum0, &xxsum0); + + *sse = xxsum0; + assert(xsum0 <= 255 * 16 * 16); + assert(xsum0 >= -255 * 16 * 16); + return (xxsum0 - ((uint32_t)((int64_t)xsum0 * xsum0) >> 8)); +} diff --git a/third_party/aom/aom_dsp/x86/highbd_convolve_avx2.c b/third_party/aom/aom_dsp/x86/highbd_convolve_avx2.c new file mode 100644 index 000000000..7d96e26ae --- /dev/null +++ b/third_party/aom/aom_dsp/x86/highbd_convolve_avx2.c @@ -0,0 +1,1151 @@ +/* + * Copyright (c) 2017, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ +#include <immintrin.h> +#include <string.h> + +#include "./aom_dsp_rtcd.h" +#include "aom_dsp/x86/convolve.h" + +#define CONV8_ROUNDING_BITS (7) + +static const uint8_t signal_pattern_0[32] = { 0, 1, 2, 3, 2, 3, 4, 5, 4, 5, 6, + 7, 6, 7, 8, 9, 0, 1, 2, 3, 2, 3, + 4, 5, 4, 5, 6, 7, 6, 7, 8, 9 }; + +static const uint8_t signal_pattern_1[32] = { 4, 5, 6, 7, 6, 7, 8, 9, + 8, 9, 10, 11, 10, 11, 12, 13, + 4, 5, 6, 7, 6, 7, 8, 9, + 8, 9, 10, 11, 10, 11, 12, 13 }; + +static const uint8_t signal_pattern_2[32] = { 6, 7, 8, 9, 8, 9, 10, 11, + 10, 11, 12, 13, 12, 13, 14, 15, + 6, 7, 8, 9, 8, 9, 10, 11, + 10, 11, 12, 13, 12, 13, 14, 15 }; + +static const uint32_t signal_index[8] = { 2, 3, 4, 5, 2, 3, 4, 5 }; + +typedef enum { PACK_8x1, PACK_8x2, PACK_16x1 } PixelPackFormat; + +typedef void (*WritePixels)(const __m256i *y0, const __m256i *y1, + const __m256i *mask, uint16_t *dst, + ptrdiff_t pitch); + +// ----------------------------------------------------------------------------- +// Copy and average + +void aom_highbd_convolve_copy_avx2(const uint8_t *src8, ptrdiff_t src_stride, + uint8_t *dst8, ptrdiff_t dst_stride, + const int16_t *filter_x, int filter_x_stride, + const int16_t *filter_y, int filter_y_stride, + int width, int h, int bd) { + const uint16_t *src = CONVERT_TO_SHORTPTR(src8); + uint16_t *dst = CONVERT_TO_SHORTPTR(dst8); + (void)filter_x; + (void)filter_y; + (void)filter_x_stride; + (void)filter_y_stride; + (void)bd; + + assert(width % 4 == 0); + if (width > 32) { // width = 64 + do { + const __m256i p0 = _mm256_loadu_si256((const __m256i *)src); + const __m256i p1 = _mm256_loadu_si256((const __m256i *)(src + 16)); + const __m256i p2 = _mm256_loadu_si256((const __m256i *)(src + 32)); + const __m256i p3 = _mm256_loadu_si256((const __m256i *)(src + 48)); + src += src_stride; + _mm256_storeu_si256((__m256i *)dst, p0); + _mm256_storeu_si256((__m256i *)(dst + 16), p1); + _mm256_storeu_si256((__m256i *)(dst + 32), p2); + _mm256_storeu_si256((__m256i *)(dst + 48), p3); + dst += dst_stride; + h--; + } while (h > 0); + } else if (width > 16) { // width = 32 + do { + const __m256i p0 = _mm256_loadu_si256((const __m256i *)src); + const __m256i p1 = _mm256_loadu_si256((const __m256i *)(src + 16)); + src += src_stride; + _mm256_storeu_si256((__m256i *)dst, p0); + _mm256_storeu_si256((__m256i *)(dst + 16), p1); + dst += dst_stride; + h--; + } while (h > 0); + } else if (width > 8) { // width = 16 + __m256i p0, p1; + do { + p0 = _mm256_loadu_si256((const __m256i *)src); + src += src_stride; + p1 = _mm256_loadu_si256((const __m256i *)src); + src += src_stride; + + _mm256_storeu_si256((__m256i *)dst, p0); + dst += dst_stride; + _mm256_storeu_si256((__m256i *)dst, p1); + dst += dst_stride; + h -= 2; + } while (h > 0); + } else if (width > 4) { // width = 8 + __m128i p0, p1; + do { + p0 = _mm_loadu_si128((const __m128i *)src); + src += src_stride; + p1 = _mm_loadu_si128((const __m128i *)src); + src += src_stride; + + _mm_storeu_si128((__m128i *)dst, p0); + dst += dst_stride; + _mm_storeu_si128((__m128i *)dst, p1); + dst += dst_stride; + h -= 2; + } while (h > 0); + } else { // width = 4 + __m128i p0, p1; + do { + p0 = _mm_loadl_epi64((const __m128i *)src); + src += src_stride; + p1 = _mm_loadl_epi64((const __m128i *)src); + src += src_stride; + + _mm_storel_epi64((__m128i *)dst, p0); + dst += dst_stride; + _mm_storel_epi64((__m128i *)dst, p1); + dst += dst_stride; + h -= 2; + } while (h > 0); + } +} + +void aom_highbd_convolve_avg_avx2(const uint8_t *src8, ptrdiff_t src_stride, + uint8_t *dst8, ptrdiff_t dst_stride, + const int16_t *filter_x, int filter_x_stride, + const int16_t *filter_y, int filter_y_stride, + int width, int h, int bd) { + uint16_t *src = CONVERT_TO_SHORTPTR(src8); + uint16_t *dst = CONVERT_TO_SHORTPTR(dst8); + (void)filter_x; + (void)filter_y; + (void)filter_x_stride; + (void)filter_y_stride; + (void)bd; + + assert(width % 4 == 0); + if (width > 32) { // width = 64 + __m256i p0, p1, p2, p3, u0, u1, u2, u3; + do { + p0 = _mm256_loadu_si256((const __m256i *)src); + p1 = _mm256_loadu_si256((const __m256i *)(src + 16)); + p2 = _mm256_loadu_si256((const __m256i *)(src + 32)); + p3 = _mm256_loadu_si256((const __m256i *)(src + 48)); + src += src_stride; + u0 = _mm256_loadu_si256((const __m256i *)dst); + u1 = _mm256_loadu_si256((const __m256i *)(dst + 16)); + u2 = _mm256_loadu_si256((const __m256i *)(dst + 32)); + u3 = _mm256_loadu_si256((const __m256i *)(dst + 48)); + _mm256_storeu_si256((__m256i *)dst, _mm256_avg_epu16(p0, u0)); + _mm256_storeu_si256((__m256i *)(dst + 16), _mm256_avg_epu16(p1, u1)); + _mm256_storeu_si256((__m256i *)(dst + 32), _mm256_avg_epu16(p2, u2)); + _mm256_storeu_si256((__m256i *)(dst + 48), _mm256_avg_epu16(p3, u3)); + dst += dst_stride; + h--; + } while (h > 0); + } else if (width > 16) { // width = 32 + __m256i p0, p1, u0, u1; + do { + p0 = _mm256_loadu_si256((const __m256i *)src); + p1 = _mm256_loadu_si256((const __m256i *)(src + 16)); + src += src_stride; + u0 = _mm256_loadu_si256((const __m256i *)dst); + u1 = _mm256_loadu_si256((const __m256i *)(dst + 16)); + _mm256_storeu_si256((__m256i *)dst, _mm256_avg_epu16(p0, u0)); + _mm256_storeu_si256((__m256i *)(dst + 16), _mm256_avg_epu16(p1, u1)); + dst += dst_stride; + h--; + } while (h > 0); + } else if (width > 8) { // width = 16 + __m256i p0, p1, u0, u1; + do { + p0 = _mm256_loadu_si256((const __m256i *)src); + p1 = _mm256_loadu_si256((const __m256i *)(src + src_stride)); + src += src_stride << 1; + u0 = _mm256_loadu_si256((const __m256i *)dst); + u1 = _mm256_loadu_si256((const __m256i *)(dst + dst_stride)); + + _mm256_storeu_si256((__m256i *)dst, _mm256_avg_epu16(p0, u0)); + _mm256_storeu_si256((__m256i *)(dst + dst_stride), + _mm256_avg_epu16(p1, u1)); + dst += dst_stride << 1; + h -= 2; + } while (h > 0); + } else if (width > 4) { // width = 8 + __m128i p0, p1, u0, u1; + do { + p0 = _mm_loadu_si128((const __m128i *)src); + p1 = _mm_loadu_si128((const __m128i *)(src + src_stride)); + src += src_stride << 1; + u0 = _mm_loadu_si128((const __m128i *)dst); + u1 = _mm_loadu_si128((const __m128i *)(dst + dst_stride)); + + _mm_storeu_si128((__m128i *)dst, _mm_avg_epu16(p0, u0)); + _mm_storeu_si128((__m128i *)(dst + dst_stride), _mm_avg_epu16(p1, u1)); + dst += dst_stride << 1; + h -= 2; + } while (h > 0); + } else { // width = 4 + __m128i p0, p1, u0, u1; + do { + p0 = _mm_loadl_epi64((const __m128i *)src); + p1 = _mm_loadl_epi64((const __m128i *)(src + src_stride)); + src += src_stride << 1; + u0 = _mm_loadl_epi64((const __m128i *)dst); + u1 = _mm_loadl_epi64((const __m128i *)(dst + dst_stride)); + + _mm_storel_epi64((__m128i *)dst, _mm_avg_epu16(u0, p0)); + _mm_storel_epi64((__m128i *)(dst + dst_stride), _mm_avg_epu16(u1, p1)); + dst += dst_stride << 1; + h -= 2; + } while (h > 0); + } +} + +// ----------------------------------------------------------------------------- +// Horizontal Filtering + +static INLINE void pack_pixels(const __m256i *s, __m256i *p /*p[4]*/) { + const __m256i idx = _mm256_loadu_si256((const __m256i *)signal_index); + const __m256i sf0 = _mm256_loadu_si256((const __m256i *)signal_pattern_0); + const __m256i sf1 = _mm256_loadu_si256((const __m256i *)signal_pattern_1); + const __m256i c = _mm256_permutevar8x32_epi32(*s, idx); + + p[0] = _mm256_shuffle_epi8(*s, sf0); // x0x6 + p[1] = _mm256_shuffle_epi8(*s, sf1); // x1x7 + p[2] = _mm256_shuffle_epi8(c, sf0); // x2x4 + p[3] = _mm256_shuffle_epi8(c, sf1); // x3x5 +} + +// Note: +// Shared by 8x2 and 16x1 block +static INLINE void pack_16_pixels(const __m256i *s0, const __m256i *s1, + __m256i *x /*x[8]*/) { + __m256i pp[8]; + pack_pixels(s0, pp); + pack_pixels(s1, &pp[4]); + x[0] = _mm256_permute2x128_si256(pp[0], pp[4], 0x20); + x[1] = _mm256_permute2x128_si256(pp[1], pp[5], 0x20); + x[2] = _mm256_permute2x128_si256(pp[2], pp[6], 0x20); + x[3] = _mm256_permute2x128_si256(pp[3], pp[7], 0x20); + x[4] = x[2]; + x[5] = x[3]; + x[6] = _mm256_permute2x128_si256(pp[0], pp[4], 0x31); + x[7] = _mm256_permute2x128_si256(pp[1], pp[5], 0x31); +} + +static INLINE void pack_pixels_with_format(const uint16_t *src, + PixelPackFormat fmt, + ptrdiff_t stride, __m256i *x) { + switch (fmt) { + case PACK_8x1: { + __m256i pp[8]; + __m256i s0; + s0 = _mm256_loadu_si256((const __m256i *)src); + pack_pixels(&s0, pp); + x[0] = _mm256_permute2x128_si256(pp[0], pp[2], 0x30); + x[1] = _mm256_permute2x128_si256(pp[1], pp[3], 0x30); + x[2] = _mm256_permute2x128_si256(pp[2], pp[0], 0x30); + x[3] = _mm256_permute2x128_si256(pp[3], pp[1], 0x30); + break; + } + case PACK_8x2: { + __m256i s0, s1; + s0 = _mm256_loadu_si256((const __m256i *)src); + s1 = _mm256_loadu_si256((const __m256i *)(src + stride)); + pack_16_pixels(&s0, &s1, x); + break; + } + case PACK_16x1: { + __m256i s0, s1; + s0 = _mm256_loadu_si256((const __m256i *)src); + s1 = _mm256_loadu_si256((const __m256i *)(src + 8)); + pack_16_pixels(&s0, &s1, x); + break; + } + default: { assert(0); } + } +} + +static INLINE void pack_8x1_pixels(const uint16_t *src, const ptrdiff_t pitch, + __m256i *x /*x[4]*/) { + pack_pixels_with_format(src, PACK_8x1, pitch, x); +} + +static INLINE void pack_8x2_pixels(const uint16_t *src, const ptrdiff_t pitch, + __m256i *x /*x[8]*/) { + pack_pixels_with_format(src, PACK_8x2, pitch, x); +} + +static INLINE void pack_16x1_pixels(const uint16_t *src, const ptrdiff_t pitch, + __m256i *x /*x[8]*/) { + pack_pixels_with_format(src, PACK_16x1, pitch, x); +} + +// Note: +// Shared by horizontal and vertical filtering +static INLINE void pack_filters(const int16_t *filter, __m256i *f /*f[4]*/) { + const __m128i h = _mm_loadu_si128((const __m128i *)filter); + const __m256i hh = _mm256_insertf128_si256(_mm256_castsi128_si256(h), h, 1); + const __m256i p0 = _mm256_set1_epi32(0x03020100); + const __m256i p1 = _mm256_set1_epi32(0x07060504); + const __m256i p2 = _mm256_set1_epi32(0x0b0a0908); + const __m256i p3 = _mm256_set1_epi32(0x0f0e0d0c); + f[0] = _mm256_shuffle_epi8(hh, p0); + f[1] = _mm256_shuffle_epi8(hh, p1); + f[2] = _mm256_shuffle_epi8(hh, p2); + f[3] = _mm256_shuffle_epi8(hh, p3); +} + +static INLINE void filter_8x1_pixels(const __m256i *sig /*sig[4]*/, + const __m256i *fil /*fil[4]*/, + __m256i *y) { + __m256i a, a0, a1; + + a0 = _mm256_madd_epi16(fil[0], sig[0]); + a1 = _mm256_madd_epi16(fil[3], sig[3]); + a = _mm256_add_epi32(a0, a1); + + a0 = _mm256_madd_epi16(fil[1], sig[1]); + a1 = _mm256_madd_epi16(fil[2], sig[2]); + + const __m256i min = _mm256_min_epi32(a0, a1); + a = _mm256_add_epi32(a, min); + + const __m256i max = _mm256_max_epi32(a0, a1); + a = _mm256_add_epi32(a, max); + + const __m256i rounding = _mm256_set1_epi32(1 << (CONV8_ROUNDING_BITS - 1)); + a = _mm256_add_epi32(a, rounding); + *y = _mm256_srai_epi32(a, CONV8_ROUNDING_BITS); +} + +static void write_8x1_pixels(const __m256i *y, const __m256i *z, + const __m256i *mask, uint16_t *dst, + ptrdiff_t pitch) { + const __m128i a0 = _mm256_castsi256_si128(*y); + const __m128i a1 = _mm256_extractf128_si256(*y, 1); + __m128i res = _mm_packus_epi32(a0, a1); + (void)z; + (void)pitch; + res = _mm_min_epi16(res, _mm256_castsi256_si128(*mask)); + _mm_storeu_si128((__m128i *)dst, res); +} + +static void write_8x2_pixels(const __m256i *y0, const __m256i *y1, + const __m256i *mask, uint16_t *dst, + ptrdiff_t pitch) { + __m256i a = _mm256_packus_epi32(*y0, *y1); + a = _mm256_min_epi16(a, *mask); + _mm_storeu_si128((__m128i *)dst, _mm256_castsi256_si128(a)); + _mm_storeu_si128((__m128i *)(dst + pitch), _mm256_extractf128_si256(a, 1)); +} + +static void write_16x1_pixels(const __m256i *y0, const __m256i *y1, + const __m256i *mask, uint16_t *dst, + ptrdiff_t dst_pitch) { + (void)dst_pitch; + __m256i a = _mm256_packus_epi32(*y0, *y1); + a = _mm256_min_epi16(a, *mask); + _mm256_storeu_si256((__m256i *)dst, a); +} + +static void filter_block_width8_horiz( + const uint16_t *src_ptr, ptrdiff_t src_pitch, const WritePixels write_8x1, + const WritePixels write_8x2, uint16_t *dst_ptr, ptrdiff_t dst_pitch, + uint32_t height, const int16_t *filter, int bd) { + __m256i signal[8], res0, res1; + const __m256i max = _mm256_set1_epi16((1 << bd) - 1); + + __m256i ff[4]; + pack_filters(filter, ff); + + src_ptr -= 3; + do { + pack_8x2_pixels(src_ptr, src_pitch, signal); + filter_8x1_pixels(signal, ff, &res0); + filter_8x1_pixels(&signal[4], ff, &res1); + write_8x2(&res0, &res1, &max, dst_ptr, dst_pitch); + height -= 2; + src_ptr += src_pitch << 1; + dst_ptr += dst_pitch << 1; + } while (height > 1); + + if (height > 0) { + pack_8x1_pixels(src_ptr, src_pitch, signal); + filter_8x1_pixels(signal, ff, &res0); + write_8x1(&res0, &res1, &max, dst_ptr, dst_pitch); + } +} + +static void aom_highbd_filter_block1d8_h8_avx2( + const uint16_t *src, ptrdiff_t src_pitch, uint16_t *dst, + ptrdiff_t dst_pitch, uint32_t height, const int16_t *filter, int bd) { + filter_block_width8_horiz(src, src_pitch, write_8x1_pixels, write_8x2_pixels, + dst, dst_pitch, height, filter, bd); +} + +static void filter_block_width16_horiz(const uint16_t *src_ptr, + ptrdiff_t src_pitch, + const WritePixels write_16x1, + uint16_t *dst_ptr, ptrdiff_t dst_pitch, + uint32_t height, const int16_t *filter, + int bd) { + __m256i signal[8], res0, res1; + const __m256i max = _mm256_set1_epi16((1 << bd) - 1); + + __m256i ff[4]; + pack_filters(filter, ff); + + src_ptr -= 3; + do { + pack_16x1_pixels(src_ptr, src_pitch, signal); + filter_8x1_pixels(signal, ff, &res0); + filter_8x1_pixels(&signal[4], ff, &res1); + write_16x1(&res0, &res1, &max, dst_ptr, dst_pitch); + height -= 1; + src_ptr += src_pitch; + dst_ptr += dst_pitch; + } while (height > 0); +} + +static void aom_highbd_filter_block1d16_h8_avx2( + const uint16_t *src, ptrdiff_t src_pitch, uint16_t *dst, + ptrdiff_t dst_pitch, uint32_t height, const int16_t *filter, int bd) { + filter_block_width16_horiz(src, src_pitch, write_16x1_pixels, dst, dst_pitch, + height, filter, bd); +} + +// 2-tap horizontal filtering + +static INLINE void pack_2t_filter(const int16_t *filter, __m256i *f) { + const __m128i h = _mm_loadu_si128((const __m128i *)filter); + const __m256i hh = _mm256_insertf128_si256(_mm256_castsi128_si256(h), h, 1); + const __m256i p = _mm256_set1_epi32(0x09080706); + f[0] = _mm256_shuffle_epi8(hh, p); +} + +// can be used by pack_8x2_2t_pixels() and pack_16x1_2t_pixels() +// the difference is s0/s1 specifies first and second rows or, +// first 16 samples and 8-sample shifted 16 samples +static INLINE void pack_16_2t_pixels(const __m256i *s0, const __m256i *s1, + __m256i *sig) { + const __m256i idx = _mm256_loadu_si256((const __m256i *)signal_index); + const __m256i sf2 = _mm256_loadu_si256((const __m256i *)signal_pattern_2); + __m256i x0 = _mm256_shuffle_epi8(*s0, sf2); + __m256i x1 = _mm256_shuffle_epi8(*s1, sf2); + __m256i r0 = _mm256_permutevar8x32_epi32(*s0, idx); + __m256i r1 = _mm256_permutevar8x32_epi32(*s1, idx); + r0 = _mm256_shuffle_epi8(r0, sf2); + r1 = _mm256_shuffle_epi8(r1, sf2); + sig[0] = _mm256_permute2x128_si256(x0, x1, 0x20); + sig[1] = _mm256_permute2x128_si256(r0, r1, 0x20); +} + +static INLINE void pack_8x2_2t_pixels(const uint16_t *src, + const ptrdiff_t pitch, __m256i *sig) { + const __m256i r0 = _mm256_loadu_si256((const __m256i *)src); + const __m256i r1 = _mm256_loadu_si256((const __m256i *)(src + pitch)); + pack_16_2t_pixels(&r0, &r1, sig); +} + +static INLINE void pack_16x1_2t_pixels(const uint16_t *src, + __m256i *sig /*sig[2]*/) { + const __m256i r0 = _mm256_loadu_si256((const __m256i *)src); + const __m256i r1 = _mm256_loadu_si256((const __m256i *)(src + 8)); + pack_16_2t_pixels(&r0, &r1, sig); +} + +static INLINE void pack_8x1_2t_pixels(const uint16_t *src, + __m256i *sig /*sig[2]*/) { + const __m256i idx = _mm256_loadu_si256((const __m256i *)signal_index); + const __m256i sf2 = _mm256_loadu_si256((const __m256i *)signal_pattern_2); + __m256i r0 = _mm256_loadu_si256((const __m256i *)src); + __m256i x0 = _mm256_shuffle_epi8(r0, sf2); + r0 = _mm256_permutevar8x32_epi32(r0, idx); + r0 = _mm256_shuffle_epi8(r0, sf2); + sig[0] = _mm256_permute2x128_si256(x0, r0, 0x20); +} + +// can be used by filter_8x2_2t_pixels() and filter_16x1_2t_pixels() +static INLINE void filter_16_2t_pixels(const __m256i *sig, const __m256i *f, + __m256i *y0, __m256i *y1) { + const __m256i rounding = _mm256_set1_epi32(1 << (CONV8_ROUNDING_BITS - 1)); + __m256i x0 = _mm256_madd_epi16(sig[0], *f); + __m256i x1 = _mm256_madd_epi16(sig[1], *f); + x0 = _mm256_add_epi32(x0, rounding); + x1 = _mm256_add_epi32(x1, rounding); + *y0 = _mm256_srai_epi32(x0, CONV8_ROUNDING_BITS); + *y1 = _mm256_srai_epi32(x1, CONV8_ROUNDING_BITS); +} + +static INLINE void filter_8x2_2t_pixels(const __m256i *sig, const __m256i *f, + __m256i *y0, __m256i *y1) { + filter_16_2t_pixels(sig, f, y0, y1); +} + +static INLINE void filter_16x1_2t_pixels(const __m256i *sig, const __m256i *f, + __m256i *y0, __m256i *y1) { + filter_16_2t_pixels(sig, f, y0, y1); +} + +static INLINE void filter_8x1_2t_pixels(const __m256i *sig, const __m256i *f, + __m256i *y0) { + const __m256i rounding = _mm256_set1_epi32(1 << (CONV8_ROUNDING_BITS - 1)); + __m256i x0 = _mm256_madd_epi16(sig[0], *f); + x0 = _mm256_add_epi32(x0, rounding); + *y0 = _mm256_srai_epi32(x0, CONV8_ROUNDING_BITS); +} + +static void filter_block_width8_2t_horiz( + const uint16_t *src_ptr, ptrdiff_t src_pitch, const WritePixels write_8x1, + const WritePixels write_8x2, uint16_t *dst_ptr, ptrdiff_t dst_pitch, + uint32_t height, const int16_t *filter, int bd) { + __m256i signal[2], res0, res1; + const __m256i max = _mm256_set1_epi16((1 << bd) - 1); + + __m256i ff; + pack_2t_filter(filter, &ff); + + src_ptr -= 3; + do { + pack_8x2_2t_pixels(src_ptr, src_pitch, signal); + filter_8x2_2t_pixels(signal, &ff, &res0, &res1); + write_8x2(&res0, &res1, &max, dst_ptr, dst_pitch); + height -= 2; + src_ptr += src_pitch << 1; + dst_ptr += dst_pitch << 1; + } while (height > 1); + + if (height > 0) { + pack_8x1_2t_pixels(src_ptr, signal); + filter_8x1_2t_pixels(signal, &ff, &res0); + write_8x1(&res0, &res1, &max, dst_ptr, dst_pitch); + } +} + +static void aom_highbd_filter_block1d8_h2_avx2( + const uint16_t *src, ptrdiff_t src_pitch, uint16_t *dst, + ptrdiff_t dst_pitch, uint32_t height, const int16_t *filter, int bd) { + filter_block_width8_2t_horiz(src, src_pitch, write_8x1_pixels, + write_8x2_pixels, dst, dst_pitch, height, filter, + bd); +} + +static void filter_block_width16_2t_horiz(const uint16_t *src_ptr, + ptrdiff_t src_pitch, + const WritePixels write_16x1, + uint16_t *dst_ptr, + ptrdiff_t dst_pitch, uint32_t height, + const int16_t *filter, int bd) { + __m256i signal[2], res0, res1; + const __m256i max = _mm256_set1_epi16((1 << bd) - 1); + + __m256i ff; + pack_2t_filter(filter, &ff); + + src_ptr -= 3; + do { + pack_16x1_2t_pixels(src_ptr, signal); + filter_16x1_2t_pixels(signal, &ff, &res0, &res1); + write_16x1(&res0, &res1, &max, dst_ptr, dst_pitch); + height -= 1; + src_ptr += src_pitch; + dst_ptr += dst_pitch; + } while (height > 0); +} + +static void aom_highbd_filter_block1d16_h2_avx2( + const uint16_t *src, ptrdiff_t src_pitch, uint16_t *dst, + ptrdiff_t dst_pitch, uint32_t height, const int16_t *filter, int bd) { + filter_block_width16_2t_horiz(src, src_pitch, write_16x1_pixels, dst, + dst_pitch, height, filter, bd); +} + +// Vertical Filtering + +static void pack_8x9_init(const uint16_t *src, ptrdiff_t pitch, __m256i *sig) { + __m256i s0 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)src)); + __m256i s1 = + _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(src + pitch))); + __m256i s2 = _mm256_castsi128_si256( + _mm_loadu_si128((const __m128i *)(src + 2 * pitch))); + __m256i s3 = _mm256_castsi128_si256( + _mm_loadu_si128((const __m128i *)(src + 3 * pitch))); + __m256i s4 = _mm256_castsi128_si256( + _mm_loadu_si128((const __m128i *)(src + 4 * pitch))); + __m256i s5 = _mm256_castsi128_si256( + _mm_loadu_si128((const __m128i *)(src + 5 * pitch))); + __m256i s6 = _mm256_castsi128_si256( + _mm_loadu_si128((const __m128i *)(src + 6 * pitch))); + + s0 = _mm256_inserti128_si256(s0, _mm256_castsi256_si128(s1), 1); + s1 = _mm256_inserti128_si256(s1, _mm256_castsi256_si128(s2), 1); + s2 = _mm256_inserti128_si256(s2, _mm256_castsi256_si128(s3), 1); + s3 = _mm256_inserti128_si256(s3, _mm256_castsi256_si128(s4), 1); + s4 = _mm256_inserti128_si256(s4, _mm256_castsi256_si128(s5), 1); + s5 = _mm256_inserti128_si256(s5, _mm256_castsi256_si128(s6), 1); + + sig[0] = _mm256_unpacklo_epi16(s0, s1); + sig[4] = _mm256_unpackhi_epi16(s0, s1); + sig[1] = _mm256_unpacklo_epi16(s2, s3); + sig[5] = _mm256_unpackhi_epi16(s2, s3); + sig[2] = _mm256_unpacklo_epi16(s4, s5); + sig[6] = _mm256_unpackhi_epi16(s4, s5); + sig[8] = s6; +} + +static INLINE void pack_8x9_pixels(const uint16_t *src, ptrdiff_t pitch, + __m256i *sig) { + // base + 7th row + __m256i s0 = _mm256_castsi128_si256( + _mm_loadu_si128((const __m128i *)(src + 7 * pitch))); + // base + 8th row + __m256i s1 = _mm256_castsi128_si256( + _mm_loadu_si128((const __m128i *)(src + 8 * pitch))); + __m256i s2 = _mm256_inserti128_si256(sig[8], _mm256_castsi256_si128(s0), 1); + __m256i s3 = _mm256_inserti128_si256(s0, _mm256_castsi256_si128(s1), 1); + sig[3] = _mm256_unpacklo_epi16(s2, s3); + sig[7] = _mm256_unpackhi_epi16(s2, s3); + sig[8] = s1; +} + +static INLINE void filter_8x9_pixels(const __m256i *sig, const __m256i *f, + __m256i *y0, __m256i *y1) { + filter_8x1_pixels(sig, f, y0); + filter_8x1_pixels(&sig[4], f, y1); +} + +static INLINE void update_pixels(__m256i *sig) { + int i; + for (i = 0; i < 3; ++i) { + sig[i] = sig[i + 1]; + sig[i + 4] = sig[i + 5]; + } +} + +static INLINE void write_8x1_pixels_ver(const __m256i *y0, const __m256i *y1, + const __m256i *mask, uint16_t *dst, + ptrdiff_t pitch) { + (void)pitch; + const __m128i v0 = _mm256_castsi256_si128(*y0); + const __m128i v1 = _mm256_castsi256_si128(*y1); + __m128i p = _mm_packus_epi32(v0, v1); + p = _mm_min_epi16(p, _mm256_castsi256_si128(*mask)); + _mm_storeu_si128((__m128i *)dst, p); +} + +static void filter_block_width8_vert(const uint16_t *src_ptr, + ptrdiff_t src_pitch, WritePixels write_8x1, + WritePixels write_8x2, uint16_t *dst_ptr, + ptrdiff_t dst_pitch, uint32_t height, + const int16_t *filter, int bd) { + __m256i signal[9], res0, res1; + const __m256i max = _mm256_set1_epi16((1 << bd) - 1); + + __m256i ff[4]; + pack_filters(filter, ff); + + pack_8x9_init(src_ptr, src_pitch, signal); + + do { + pack_8x9_pixels(src_ptr, src_pitch, signal); + + filter_8x9_pixels(signal, ff, &res0, &res1); + write_8x2(&res0, &res1, &max, dst_ptr, dst_pitch); + update_pixels(signal); + + src_ptr += src_pitch << 1; + dst_ptr += dst_pitch << 1; + height -= 2; + } while (height > 1); + + if (height > 0) { + pack_8x9_pixels(src_ptr, src_pitch, signal); + filter_8x9_pixels(signal, ff, &res0, &res1); + write_8x1(&res0, &res1, &max, dst_ptr, dst_pitch); + } +} + +static void aom_highbd_filter_block1d8_v8_avx2( + const uint16_t *src, ptrdiff_t src_pitch, uint16_t *dst, + ptrdiff_t dst_pitch, uint32_t height, const int16_t *filter, int bd) { + filter_block_width8_vert(src, src_pitch, write_8x1_pixels_ver, + write_8x2_pixels, dst, dst_pitch, height, filter, + bd); +} + +static void pack_16x9_init(const uint16_t *src, ptrdiff_t pitch, __m256i *sig) { + __m256i u0, u1, u2, u3; + // load 0-6 rows + const __m256i s0 = _mm256_loadu_si256((const __m256i *)src); + const __m256i s1 = _mm256_loadu_si256((const __m256i *)(src + pitch)); + const __m256i s2 = _mm256_loadu_si256((const __m256i *)(src + 2 * pitch)); + const __m256i s3 = _mm256_loadu_si256((const __m256i *)(src + 3 * pitch)); + const __m256i s4 = _mm256_loadu_si256((const __m256i *)(src + 4 * pitch)); + const __m256i s5 = _mm256_loadu_si256((const __m256i *)(src + 5 * pitch)); + const __m256i s6 = _mm256_loadu_si256((const __m256i *)(src + 6 * pitch)); + + u0 = _mm256_permute2x128_si256(s0, s1, 0x20); // 0, 1 low + u1 = _mm256_permute2x128_si256(s0, s1, 0x31); // 0, 1 high + + u2 = _mm256_permute2x128_si256(s1, s2, 0x20); // 1, 2 low + u3 = _mm256_permute2x128_si256(s1, s2, 0x31); // 1, 2 high + + sig[0] = _mm256_unpacklo_epi16(u0, u2); + sig[4] = _mm256_unpackhi_epi16(u0, u2); + + sig[8] = _mm256_unpacklo_epi16(u1, u3); + sig[12] = _mm256_unpackhi_epi16(u1, u3); + + u0 = _mm256_permute2x128_si256(s2, s3, 0x20); + u1 = _mm256_permute2x128_si256(s2, s3, 0x31); + + u2 = _mm256_permute2x128_si256(s3, s4, 0x20); + u3 = _mm256_permute2x128_si256(s3, s4, 0x31); + + sig[1] = _mm256_unpacklo_epi16(u0, u2); + sig[5] = _mm256_unpackhi_epi16(u0, u2); + + sig[9] = _mm256_unpacklo_epi16(u1, u3); + sig[13] = _mm256_unpackhi_epi16(u1, u3); + + u0 = _mm256_permute2x128_si256(s4, s5, 0x20); + u1 = _mm256_permute2x128_si256(s4, s5, 0x31); + + u2 = _mm256_permute2x128_si256(s5, s6, 0x20); + u3 = _mm256_permute2x128_si256(s5, s6, 0x31); + + sig[2] = _mm256_unpacklo_epi16(u0, u2); + sig[6] = _mm256_unpackhi_epi16(u0, u2); + + sig[10] = _mm256_unpacklo_epi16(u1, u3); + sig[14] = _mm256_unpackhi_epi16(u1, u3); + + sig[16] = s6; +} + +static void pack_16x9_pixels(const uint16_t *src, ptrdiff_t pitch, + __m256i *sig) { + // base + 7th row + const __m256i s7 = _mm256_loadu_si256((const __m256i *)(src + 7 * pitch)); + // base + 8th row + const __m256i s8 = _mm256_loadu_si256((const __m256i *)(src + 8 * pitch)); + + __m256i u0, u1, u2, u3; + u0 = _mm256_permute2x128_si256(sig[16], s7, 0x20); + u1 = _mm256_permute2x128_si256(sig[16], s7, 0x31); + + u2 = _mm256_permute2x128_si256(s7, s8, 0x20); + u3 = _mm256_permute2x128_si256(s7, s8, 0x31); + + sig[3] = _mm256_unpacklo_epi16(u0, u2); + sig[7] = _mm256_unpackhi_epi16(u0, u2); + + sig[11] = _mm256_unpacklo_epi16(u1, u3); + sig[15] = _mm256_unpackhi_epi16(u1, u3); + + sig[16] = s8; +} + +static INLINE void filter_16x9_pixels(const __m256i *sig, const __m256i *f, + __m256i *y0, __m256i *y1) { + __m256i res[4]; + int i; + for (i = 0; i < 4; ++i) { + filter_8x1_pixels(&sig[i << 2], f, &res[i]); + } + + const __m256i l0l1 = _mm256_packus_epi32(res[0], res[1]); + const __m256i h0h1 = _mm256_packus_epi32(res[2], res[3]); + *y0 = _mm256_permute2x128_si256(l0l1, h0h1, 0x20); + *y1 = _mm256_permute2x128_si256(l0l1, h0h1, 0x31); +} + +static INLINE void write_16x2_pixels(const __m256i *y0, const __m256i *y1, + const __m256i *mask, uint16_t *dst, + ptrdiff_t pitch) { + __m256i p = _mm256_min_epi16(*y0, *mask); + _mm256_storeu_si256((__m256i *)dst, p); + p = _mm256_min_epi16(*y1, *mask); + _mm256_storeu_si256((__m256i *)(dst + pitch), p); +} + +static INLINE void write_16x1_pixels_ver(const __m256i *y0, const __m256i *y1, + const __m256i *mask, uint16_t *dst, + ptrdiff_t pitch) { + (void)y1; + (void)pitch; + const __m256i p = _mm256_min_epi16(*y0, *mask); + _mm256_storeu_si256((__m256i *)dst, p); +} + +static void update_16x9_pixels(__m256i *sig) { + update_pixels(&sig[0]); + update_pixels(&sig[8]); +} + +static void filter_block_width16_vert(const uint16_t *src_ptr, + ptrdiff_t src_pitch, + WritePixels write_16x1, + WritePixels write_16x2, uint16_t *dst_ptr, + ptrdiff_t dst_pitch, uint32_t height, + const int16_t *filter, int bd) { + __m256i signal[17], res0, res1; + const __m256i max = _mm256_set1_epi16((1 << bd) - 1); + + __m256i ff[4]; + pack_filters(filter, ff); + + pack_16x9_init(src_ptr, src_pitch, signal); + + do { + pack_16x9_pixels(src_ptr, src_pitch, signal); + filter_16x9_pixels(signal, ff, &res0, &res1); + write_16x2(&res0, &res1, &max, dst_ptr, dst_pitch); + update_16x9_pixels(signal); + + src_ptr += src_pitch << 1; + dst_ptr += dst_pitch << 1; + height -= 2; + } while (height > 1); + + if (height > 0) { + pack_16x9_pixels(src_ptr, src_pitch, signal); + filter_16x9_pixels(signal, ff, &res0, &res1); + write_16x1(&res0, &res1, &max, dst_ptr, dst_pitch); + } +} + +static void aom_highbd_filter_block1d16_v8_avx2( + const uint16_t *src, ptrdiff_t src_pitch, uint16_t *dst, + ptrdiff_t dst_pitch, uint32_t height, const int16_t *filter, int bd) { + filter_block_width16_vert(src, src_pitch, write_16x1_pixels_ver, + write_16x2_pixels, dst, dst_pitch, height, filter, + bd); +} + +// 2-tap vertical filtering + +static void pack_16x2_init(const uint16_t *src, __m256i *sig) { + sig[2] = _mm256_loadu_si256((const __m256i *)src); +} + +static INLINE void pack_16x2_2t_pixels(const uint16_t *src, ptrdiff_t pitch, + __m256i *sig) { + // load the next row + const __m256i u = _mm256_loadu_si256((const __m256i *)(src + pitch)); + sig[0] = _mm256_unpacklo_epi16(sig[2], u); + sig[1] = _mm256_unpackhi_epi16(sig[2], u); + sig[2] = u; +} + +static INLINE void filter_16x2_2t_pixels(const __m256i *sig, const __m256i *f, + __m256i *y0, __m256i *y1) { + filter_16_2t_pixels(sig, f, y0, y1); +} + +static void filter_block_width16_2t_vert(const uint16_t *src_ptr, + ptrdiff_t src_pitch, + WritePixels write_16x1, + uint16_t *dst_ptr, ptrdiff_t dst_pitch, + uint32_t height, const int16_t *filter, + int bd) { + __m256i signal[3], res0, res1; + const __m256i max = _mm256_set1_epi16((1 << bd) - 1); + __m256i ff; + + pack_2t_filter(filter, &ff); + pack_16x2_init(src_ptr, signal); + + do { + pack_16x2_2t_pixels(src_ptr, src_pitch, signal); + filter_16x2_2t_pixels(signal, &ff, &res0, &res1); + write_16x1(&res0, &res1, &max, dst_ptr, dst_pitch); + + src_ptr += src_pitch; + dst_ptr += dst_pitch; + height -= 1; + } while (height > 0); +} + +static void aom_highbd_filter_block1d16_v2_avx2( + const uint16_t *src, ptrdiff_t src_pitch, uint16_t *dst, + ptrdiff_t dst_pitch, uint32_t height, const int16_t *filter, int bd) { + filter_block_width16_2t_vert(src, src_pitch, write_16x1_pixels, dst, + dst_pitch, height, filter, bd); +} + +static INLINE void pack_8x1_2t_filter(const int16_t *filter, __m128i *f) { + const __m128i h = _mm_loadu_si128((const __m128i *)filter); + const __m128i p = _mm_set1_epi32(0x09080706); + f[0] = _mm_shuffle_epi8(h, p); +} + +static void pack_8x2_init(const uint16_t *src, __m128i *sig) { + sig[2] = _mm_loadu_si128((const __m128i *)src); +} + +static INLINE void pack_8x2_2t_pixels_ver(const uint16_t *src, ptrdiff_t pitch, + __m128i *sig) { + // load the next row + const __m128i u = _mm_loadu_si128((const __m128i *)(src + pitch)); + sig[0] = _mm_unpacklo_epi16(sig[2], u); + sig[1] = _mm_unpackhi_epi16(sig[2], u); + sig[2] = u; +} + +static INLINE void filter_8_2t_pixels(const __m128i *sig, const __m128i *f, + __m128i *y0, __m128i *y1) { + const __m128i rounding = _mm_set1_epi32(1 << (CONV8_ROUNDING_BITS - 1)); + __m128i x0 = _mm_madd_epi16(sig[0], *f); + __m128i x1 = _mm_madd_epi16(sig[1], *f); + x0 = _mm_add_epi32(x0, rounding); + x1 = _mm_add_epi32(x1, rounding); + *y0 = _mm_srai_epi32(x0, CONV8_ROUNDING_BITS); + *y1 = _mm_srai_epi32(x1, CONV8_ROUNDING_BITS); +} + +static void write_8x1_2t_pixels_ver(const __m128i *y0, const __m128i *y1, + const __m128i *mask, uint16_t *dst) { + __m128i res = _mm_packus_epi32(*y0, *y1); + res = _mm_min_epi16(res, *mask); + _mm_storeu_si128((__m128i *)dst, res); +} + +typedef void (*Write8Pixels)(const __m128i *y0, const __m128i *y1, + const __m128i *mask, uint16_t *dst); + +static void filter_block_width8_2t_vert(const uint16_t *src_ptr, + ptrdiff_t src_pitch, + Write8Pixels write_8x1, + uint16_t *dst_ptr, ptrdiff_t dst_pitch, + uint32_t height, const int16_t *filter, + int bd) { + __m128i signal[3], res0, res1; + const __m128i max = _mm_set1_epi16((1 << bd) - 1); + __m128i ff; + + pack_8x1_2t_filter(filter, &ff); + pack_8x2_init(src_ptr, signal); + + do { + pack_8x2_2t_pixels_ver(src_ptr, src_pitch, signal); + filter_8_2t_pixels(signal, &ff, &res0, &res1); + write_8x1(&res0, &res1, &max, dst_ptr); + + src_ptr += src_pitch; + dst_ptr += dst_pitch; + height -= 1; + } while (height > 0); +} + +static void aom_highbd_filter_block1d8_v2_avx2( + const uint16_t *src, ptrdiff_t src_pitch, uint16_t *dst, + ptrdiff_t dst_pitch, uint32_t height, const int16_t *filter, int bd) { + filter_block_width8_2t_vert(src, src_pitch, write_8x1_2t_pixels_ver, dst, + dst_pitch, height, filter, bd); +} + +// Calculation with averaging the input pixels + +static void write_8x1_avg_pixels(const __m256i *y0, const __m256i *y1, + const __m256i *mask, uint16_t *dst, + ptrdiff_t pitch) { + (void)y1; + (void)pitch; + const __m128i a0 = _mm256_castsi256_si128(*y0); + const __m128i a1 = _mm256_extractf128_si256(*y0, 1); + __m128i res = _mm_packus_epi32(a0, a1); + const __m128i pix = _mm_loadu_si128((const __m128i *)dst); + res = _mm_min_epi16(res, _mm256_castsi256_si128(*mask)); + res = _mm_avg_epu16(res, pix); + _mm_storeu_si128((__m128i *)dst, res); +} + +static void write_8x2_avg_pixels(const __m256i *y0, const __m256i *y1, + const __m256i *mask, uint16_t *dst, + ptrdiff_t pitch) { + __m256i a = _mm256_packus_epi32(*y0, *y1); + const __m128i pix0 = _mm_loadu_si128((const __m128i *)dst); + const __m128i pix1 = _mm_loadu_si128((const __m128i *)(dst + pitch)); + const __m256i pix = + _mm256_insertf128_si256(_mm256_castsi128_si256(pix0), pix1, 1); + a = _mm256_min_epi16(a, *mask); + a = _mm256_avg_epu16(a, pix); + _mm_storeu_si128((__m128i *)dst, _mm256_castsi256_si128(a)); + _mm_storeu_si128((__m128i *)(dst + pitch), _mm256_extractf128_si256(a, 1)); +} + +static void write_16x1_avg_pixels(const __m256i *y0, const __m256i *y1, + const __m256i *mask, uint16_t *dst, + ptrdiff_t pitch) { + (void)pitch; + __m256i a = _mm256_packus_epi32(*y0, *y1); + const __m256i pix = _mm256_loadu_si256((const __m256i *)dst); + a = _mm256_min_epi16(a, *mask); + a = _mm256_avg_epu16(a, pix); + _mm256_storeu_si256((__m256i *)dst, a); +} + +static INLINE void write_8x1_avg_pixels_ver(const __m256i *y0, + const __m256i *y1, + const __m256i *mask, uint16_t *dst, + ptrdiff_t pitch) { + (void)pitch; + const __m128i v0 = _mm256_castsi256_si128(*y0); + const __m128i v1 = _mm256_castsi256_si128(*y1); + __m128i p = _mm_packus_epi32(v0, v1); + const __m128i pix = _mm_loadu_si128((const __m128i *)dst); + p = _mm_min_epi16(p, _mm256_castsi256_si128(*mask)); + p = _mm_avg_epu16(p, pix); + _mm_storeu_si128((__m128i *)dst, p); +} + +static INLINE void write_16x2_avg_pixels(const __m256i *y0, const __m256i *y1, + const __m256i *mask, uint16_t *dst, + ptrdiff_t pitch) { + const __m256i pix0 = _mm256_loadu_si256((const __m256i *)dst); + const __m256i pix1 = _mm256_loadu_si256((const __m256i *)(dst + pitch)); + __m256i p = _mm256_min_epi16(*y0, *mask); + p = _mm256_avg_epu16(p, pix0); + _mm256_storeu_si256((__m256i *)dst, p); + + p = _mm256_min_epi16(*y1, *mask); + p = _mm256_avg_epu16(p, pix1); + _mm256_storeu_si256((__m256i *)(dst + pitch), p); +} + +static INLINE void write_16x1_avg_pixels_ver(const __m256i *y0, + const __m256i *y1, + const __m256i *mask, uint16_t *dst, + ptrdiff_t pitch) { + (void)y1; + (void)pitch; + __m256i p = _mm256_min_epi16(*y0, *mask); + const __m256i pix = _mm256_loadu_si256((const __m256i *)dst); + p = _mm256_avg_epu16(p, pix); + _mm256_storeu_si256((__m256i *)dst, p); +} + +static void write_8x1_2t_avg_pixels_ver(const __m128i *y0, const __m128i *y1, + const __m128i *mask, uint16_t *dst) { + __m128i res = _mm_packus_epi32(*y0, *y1); + const __m128i pix = _mm_loadu_si128((const __m128i *)dst); + res = _mm_min_epi16(res, *mask); + res = _mm_avg_epu16(res, pix); + _mm_storeu_si128((__m128i *)dst, res); +} + +static void aom_highbd_filter_block1d8_h8_avg_avx2( + const uint16_t *src, ptrdiff_t src_pitch, uint16_t *dst, + ptrdiff_t dst_pitch, uint32_t height, const int16_t *filter, int bd) { + filter_block_width8_horiz(src, src_pitch, write_8x1_avg_pixels, + write_8x2_avg_pixels, dst, dst_pitch, height, + filter, bd); +} + +static void aom_highbd_filter_block1d16_h8_avg_avx2( + const uint16_t *src, ptrdiff_t src_pitch, uint16_t *dst, + ptrdiff_t dst_pitch, uint32_t height, const int16_t *filter, int bd) { + filter_block_width16_horiz(src, src_pitch, write_16x1_avg_pixels, dst, + dst_pitch, height, filter, bd); +} + +static void aom_highbd_filter_block1d8_v8_avg_avx2( + const uint16_t *src, ptrdiff_t src_pitch, uint16_t *dst, + ptrdiff_t dst_pitch, uint32_t height, const int16_t *filter, int bd) { + filter_block_width8_vert(src, src_pitch, write_8x1_avg_pixels_ver, + write_8x2_avg_pixels, dst, dst_pitch, height, filter, + bd); +} + +static void aom_highbd_filter_block1d16_v8_avg_avx2( + const uint16_t *src, ptrdiff_t src_pitch, uint16_t *dst, + ptrdiff_t dst_pitch, uint32_t height, const int16_t *filter, int bd) { + filter_block_width16_vert(src, src_pitch, write_16x1_avg_pixels_ver, + write_16x2_avg_pixels, dst, dst_pitch, height, + filter, bd); +} + +// 2-tap averaging + +static void aom_highbd_filter_block1d8_h2_avg_avx2( + const uint16_t *src, ptrdiff_t src_pitch, uint16_t *dst, + ptrdiff_t dst_pitch, uint32_t height, const int16_t *filter, int bd) { + filter_block_width8_2t_horiz(src, src_pitch, write_8x1_avg_pixels, + write_8x2_avg_pixels, dst, dst_pitch, height, + filter, bd); +} + +static void aom_highbd_filter_block1d16_h2_avg_avx2( + const uint16_t *src, ptrdiff_t src_pitch, uint16_t *dst, + ptrdiff_t dst_pitch, uint32_t height, const int16_t *filter, int bd) { + filter_block_width16_2t_horiz(src, src_pitch, write_16x1_avg_pixels, dst, + dst_pitch, height, filter, bd); +} + +static void aom_highbd_filter_block1d16_v2_avg_avx2( + const uint16_t *src, ptrdiff_t src_pitch, uint16_t *dst, + ptrdiff_t dst_pitch, uint32_t height, const int16_t *filter, int bd) { + filter_block_width16_2t_vert(src, src_pitch, write_16x1_avg_pixels, dst, + dst_pitch, height, filter, bd); +} + +static void aom_highbd_filter_block1d8_v2_avg_avx2( + const uint16_t *src, ptrdiff_t src_pitch, uint16_t *dst, + ptrdiff_t dst_pitch, uint32_t height, const int16_t *filter, int bd) { + filter_block_width8_2t_vert(src, src_pitch, write_8x1_2t_avg_pixels_ver, dst, + dst_pitch, height, filter, bd); +} + +typedef void HbdFilter1dFunc(const uint16_t *, ptrdiff_t, uint16_t *, ptrdiff_t, + uint32_t, const int16_t *, int); + +#define HIGHBD_FUNC(width, dir, avg, opt) \ + aom_highbd_filter_block1d##width##_##dir##_##avg##opt + +HbdFilter1dFunc HIGHBD_FUNC(4, h8, , sse2); +HbdFilter1dFunc HIGHBD_FUNC(4, h2, , sse2); +HbdFilter1dFunc HIGHBD_FUNC(4, v8, , sse2); +HbdFilter1dFunc HIGHBD_FUNC(4, v2, , sse2); + +#define aom_highbd_filter_block1d4_h8_avx2 HIGHBD_FUNC(4, h8, , sse2) +#define aom_highbd_filter_block1d4_h2_avx2 HIGHBD_FUNC(4, h2, , sse2) +#define aom_highbd_filter_block1d4_v8_avx2 HIGHBD_FUNC(4, v8, , sse2) +#define aom_highbd_filter_block1d4_v2_avx2 HIGHBD_FUNC(4, v2, , sse2) + +HIGH_FUN_CONV_1D(horiz, x_step_q4, filter_x, h, src, , avx2); +HIGH_FUN_CONV_1D(vert, y_step_q4, filter_y, v, src - src_stride * 3, , avx2); +HIGH_FUN_CONV_2D(, avx2); + +HbdFilter1dFunc HIGHBD_FUNC(4, h8, avg_, sse2); +HbdFilter1dFunc HIGHBD_FUNC(4, h2, avg_, sse2); +HbdFilter1dFunc HIGHBD_FUNC(4, v8, avg_, sse2); +HbdFilter1dFunc HIGHBD_FUNC(4, v2, avg_, sse2); + +#define aom_highbd_filter_block1d4_h8_avg_avx2 HIGHBD_FUNC(4, h8, avg_, sse2) +#define aom_highbd_filter_block1d4_h2_avg_avx2 HIGHBD_FUNC(4, h2, avg_, sse2) +#define aom_highbd_filter_block1d4_v8_avg_avx2 HIGHBD_FUNC(4, v8, avg_, sse2) +#define aom_highbd_filter_block1d4_v2_avg_avx2 HIGHBD_FUNC(4, v2, avg_, sse2) + +HIGH_FUN_CONV_1D(avg_horiz, x_step_q4, filter_x, h, src, avg_, avx2); +HIGH_FUN_CONV_1D(avg_vert, y_step_q4, filter_y, v, src - src_stride * 3, avg_, + avx2); +HIGH_FUN_CONV_2D(avg_, avx2); + +#undef HIGHBD_FUNC diff --git a/third_party/aom/aom_dsp/x86/highbd_intrapred_sse2.asm b/third_party/aom/aom_dsp/x86/highbd_intrapred_sse2.asm new file mode 100644 index 000000000..5d84ef8a7 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/highbd_intrapred_sse2.asm @@ -0,0 +1,456 @@ +; +; Copyright (c) 2016, Alliance for Open Media. All rights reserved +; +; This source code is subject to the terms of the BSD 2 Clause License and +; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License +; was not distributed with this source code in the LICENSE file, you can +; obtain it at www.aomedia.org/license/software. If the Alliance for Open +; Media Patent License 1.0 was not distributed with this source code in the +; PATENTS file, you can obtain it at www.aomedia.org/license/patent. +; + +; + +%include "third_party/x86inc/x86inc.asm" + +SECTION_RODATA +pw_4: times 8 dw 4 +pw_8: times 8 dw 8 +pw_16: times 4 dd 16 +pw_32: times 4 dd 32 + +SECTION .text +INIT_XMM sse2 +cglobal highbd_dc_predictor_4x4, 4, 5, 4, dst, stride, above, left, goffset + GET_GOT goffsetq + + movq m0, [aboveq] + movq m2, [leftq] + paddw m0, m2 + pshuflw m1, m0, 0xe + paddw m0, m1 + pshuflw m1, m0, 0x1 + paddw m0, m1 + paddw m0, [GLOBAL(pw_4)] + psraw m0, 3 + pshuflw m0, m0, 0x0 + movq [dstq ], m0 + movq [dstq+strideq*2], m0 + lea dstq, [dstq+strideq*4] + movq [dstq ], m0 + movq [dstq+strideq*2], m0 + + RESTORE_GOT + RET + +INIT_XMM sse2 +cglobal highbd_dc_predictor_8x8, 4, 5, 4, dst, stride, above, left, goffset + GET_GOT goffsetq + + pxor m1, m1 + mova m0, [aboveq] + mova m2, [leftq] + DEFINE_ARGS dst, stride, stride3, one + mov oned, 0x00010001 + lea stride3q, [strideq*3] + movd m3, oned + pshufd m3, m3, 0x0 + paddw m0, m2 + pmaddwd m0, m3 + packssdw m0, m1 + pmaddwd m0, m3 + packssdw m0, m1 + pmaddwd m0, m3 + paddw m0, [GLOBAL(pw_8)] + psrlw m0, 4 + pshuflw m0, m0, 0x0 + punpcklqdq m0, m0 + mova [dstq ], m0 + mova [dstq+strideq*2 ], m0 + mova [dstq+strideq*4 ], m0 + mova [dstq+stride3q*2], m0 + lea dstq, [dstq+strideq*8] + mova [dstq ], m0 + mova [dstq+strideq*2 ], m0 + mova [dstq+strideq*4 ], m0 + mova [dstq+stride3q*2], m0 + + RESTORE_GOT + RET + +INIT_XMM sse2 +cglobal highbd_dc_predictor_16x16, 4, 5, 5, dst, stride, above, left, goffset + GET_GOT goffsetq + + pxor m1, m1 + mova m0, [aboveq] + mova m3, [aboveq+16] + mova m2, [leftq] + mova m4, [leftq+16] + DEFINE_ARGS dst, stride, stride3, lines4 + lea stride3q, [strideq*3] + mov lines4d, 4 + paddw m0, m2 + paddw m0, m3 + paddw m0, m4 + movhlps m2, m0 + paddw m0, m2 + punpcklwd m0, m1 + movhlps m2, m0 + paddd m0, m2 + punpckldq m0, m1 + movhlps m2, m0 + paddd m0, m2 + paddd m0, [GLOBAL(pw_16)] + psrad m0, 5 + pshuflw m0, m0, 0x0 + punpcklqdq m0, m0 +.loop: + mova [dstq ], m0 + mova [dstq +16], m0 + mova [dstq+strideq*2 ], m0 + mova [dstq+strideq*2 +16], m0 + mova [dstq+strideq*4 ], m0 + mova [dstq+strideq*4 +16], m0 + mova [dstq+stride3q*2 ], m0 + mova [dstq+stride3q*2+16], m0 + lea dstq, [dstq+strideq*8] + dec lines4d + jnz .loop + + RESTORE_GOT + REP_RET + +INIT_XMM sse2 +cglobal highbd_dc_predictor_32x32, 4, 5, 7, dst, stride, above, left, goffset + GET_GOT goffsetq + + mova m0, [aboveq] + mova m2, [aboveq+16] + mova m3, [aboveq+32] + mova m4, [aboveq+48] + paddw m0, m2 + paddw m3, m4 + mova m2, [leftq] + mova m4, [leftq+16] + mova m5, [leftq+32] + mova m6, [leftq+48] + paddw m2, m4 + paddw m5, m6 + paddw m0, m3 + paddw m2, m5 + pxor m1, m1 + paddw m0, m2 + DEFINE_ARGS dst, stride, stride3, lines4 + lea stride3q, [strideq*3] + mov lines4d, 8 + movhlps m2, m0 + paddw m0, m2 + punpcklwd m0, m1 + movhlps m2, m0 + paddd m0, m2 + punpckldq m0, m1 + movhlps m2, m0 + paddd m0, m2 + paddd m0, [GLOBAL(pw_32)] + psrad m0, 6 + pshuflw m0, m0, 0x0 + punpcklqdq m0, m0 +.loop: + mova [dstq ], m0 + mova [dstq +16 ], m0 + mova [dstq +32 ], m0 + mova [dstq +48 ], m0 + mova [dstq+strideq*2 ], m0 + mova [dstq+strideq*2+16 ], m0 + mova [dstq+strideq*2+32 ], m0 + mova [dstq+strideq*2+48 ], m0 + mova [dstq+strideq*4 ], m0 + mova [dstq+strideq*4+16 ], m0 + mova [dstq+strideq*4+32 ], m0 + mova [dstq+strideq*4+48 ], m0 + mova [dstq+stride3q*2 ], m0 + mova [dstq+stride3q*2 +16], m0 + mova [dstq+stride3q*2 +32], m0 + mova [dstq+stride3q*2 +48], m0 + lea dstq, [dstq+strideq*8] + dec lines4d + jnz .loop + + RESTORE_GOT + REP_RET + +INIT_XMM sse2 +cglobal highbd_v_predictor_4x4, 3, 3, 1, dst, stride, above + movq m0, [aboveq] + movq [dstq ], m0 + movq [dstq+strideq*2], m0 + lea dstq, [dstq+strideq*4] + movq [dstq ], m0 + movq [dstq+strideq*2], m0 + RET + +INIT_XMM sse2 +cglobal highbd_v_predictor_8x8, 3, 3, 1, dst, stride, above + mova m0, [aboveq] + DEFINE_ARGS dst, stride, stride3 + lea stride3q, [strideq*3] + mova [dstq ], m0 + mova [dstq+strideq*2 ], m0 + mova [dstq+strideq*4 ], m0 + mova [dstq+stride3q*2], m0 + lea dstq, [dstq+strideq*8] + mova [dstq ], m0 + mova [dstq+strideq*2 ], m0 + mova [dstq+strideq*4 ], m0 + mova [dstq+stride3q*2], m0 + RET + +INIT_XMM sse2 +cglobal highbd_v_predictor_16x16, 3, 4, 2, dst, stride, above + mova m0, [aboveq] + mova m1, [aboveq+16] + DEFINE_ARGS dst, stride, stride3, nlines4 + lea stride3q, [strideq*3] + mov nlines4d, 4 +.loop: + mova [dstq ], m0 + mova [dstq +16], m1 + mova [dstq+strideq*2 ], m0 + mova [dstq+strideq*2 +16], m1 + mova [dstq+strideq*4 ], m0 + mova [dstq+strideq*4 +16], m1 + mova [dstq+stride3q*2 ], m0 + mova [dstq+stride3q*2+16], m1 + lea dstq, [dstq+strideq*8] + dec nlines4d + jnz .loop + REP_RET + +INIT_XMM sse2 +cglobal highbd_v_predictor_32x32, 3, 4, 4, dst, stride, above + mova m0, [aboveq] + mova m1, [aboveq+16] + mova m2, [aboveq+32] + mova m3, [aboveq+48] + DEFINE_ARGS dst, stride, stride3, nlines4 + lea stride3q, [strideq*3] + mov nlines4d, 8 +.loop: + mova [dstq ], m0 + mova [dstq +16], m1 + mova [dstq +32], m2 + mova [dstq +48], m3 + mova [dstq+strideq*2 ], m0 + mova [dstq+strideq*2 +16], m1 + mova [dstq+strideq*2 +32], m2 + mova [dstq+strideq*2 +48], m3 + mova [dstq+strideq*4 ], m0 + mova [dstq+strideq*4 +16], m1 + mova [dstq+strideq*4 +32], m2 + mova [dstq+strideq*4 +48], m3 + mova [dstq+stride3q*2 ], m0 + mova [dstq+stride3q*2 +16], m1 + mova [dstq+stride3q*2 +32], m2 + mova [dstq+stride3q*2 +48], m3 + lea dstq, [dstq+strideq*8] + dec nlines4d + jnz .loop + REP_RET + +INIT_XMM sse2 +cglobal highbd_tm_predictor_4x4, 5, 5, 6, dst, stride, above, left, bps + movd m1, [aboveq-2] + movq m0, [aboveq] + pshuflw m1, m1, 0x0 + movlhps m0, m0 ; t1 t2 t3 t4 t1 t2 t3 t4 + movlhps m1, m1 ; tl tl tl tl tl tl tl tl + ; Get the values to compute the maximum value at this bit depth + pcmpeqw m3, m3 + movd m4, bpsd + psubw m0, m1 ; t1-tl t2-tl t3-tl t4-tl + psllw m3, m4 + pcmpeqw m2, m2 + pxor m4, m4 ; min possible value + pxor m3, m2 ; max possible value + mova m1, [leftq] + pshuflw m2, m1, 0x0 + pshuflw m5, m1, 0x55 + movlhps m2, m5 ; l1 l1 l1 l1 l2 l2 l2 l2 + paddw m2, m0 + ;Clamp to the bit-depth + pminsw m2, m3 + pmaxsw m2, m4 + ;Store the values + movq [dstq ], m2 + movhpd [dstq+strideq*2], m2 + lea dstq, [dstq+strideq*4] + pshuflw m2, m1, 0xaa + pshuflw m5, m1, 0xff + movlhps m2, m5 + paddw m2, m0 + ;Clamp to the bit-depth + pminsw m2, m3 + pmaxsw m2, m4 + ;Store the values + movq [dstq ], m2 + movhpd [dstq+strideq*2], m2 + RET + +INIT_XMM sse2 +cglobal highbd_tm_predictor_8x8, 5, 6, 5, dst, stride, above, left, bps, one + movd m1, [aboveq-2] + mova m0, [aboveq] + pshuflw m1, m1, 0x0 + ; Get the values to compute the maximum value at this bit depth + mov oned, 1 + pxor m3, m3 + pxor m4, m4 + pinsrw m3, oned, 0 + pinsrw m4, bpsd, 0 + pshuflw m3, m3, 0x0 + DEFINE_ARGS dst, stride, line, left + punpcklqdq m3, m3 + mov lineq, -4 + mova m2, m3 + punpcklqdq m1, m1 + psllw m3, m4 + add leftq, 16 + psubw m3, m2 ; max possible value + pxor m4, m4 ; min possible value + psubw m0, m1 +.loop: + movd m1, [leftq+lineq*4] + movd m2, [leftq+lineq*4+2] + pshuflw m1, m1, 0x0 + pshuflw m2, m2, 0x0 + punpcklqdq m1, m1 + punpcklqdq m2, m2 + paddw m1, m0 + paddw m2, m0 + ;Clamp to the bit-depth + pminsw m1, m3 + pminsw m2, m3 + pmaxsw m1, m4 + pmaxsw m2, m4 + ;Store the values + mova [dstq ], m1 + mova [dstq+strideq*2], m2 + lea dstq, [dstq+strideq*4] + inc lineq + jnz .loop + REP_RET + +INIT_XMM sse2 +cglobal highbd_tm_predictor_16x16, 5, 5, 8, dst, stride, above, left, bps + movd m2, [aboveq-2] + mova m0, [aboveq] + mova m1, [aboveq+16] + pshuflw m2, m2, 0x0 + ; Get the values to compute the maximum value at this bit depth + pcmpeqw m3, m3 + movd m4, bpsd + punpcklqdq m2, m2 + psllw m3, m4 + pcmpeqw m5, m5 + pxor m4, m4 ; min possible value + pxor m3, m5 ; max possible value + DEFINE_ARGS dst, stride, line, left + mov lineq, -8 + psubw m0, m2 + psubw m1, m2 +.loop: + movd m7, [leftq] + pshuflw m5, m7, 0x0 + pshuflw m2, m7, 0x55 + punpcklqdq m5, m5 ; l1 l1 l1 l1 l1 l1 l1 l1 + punpcklqdq m2, m2 ; l2 l2 l2 l2 l2 l2 l2 l2 + paddw m6, m5, m0 ; t1-tl+l1 to t4-tl+l1 + paddw m5, m1 ; t5-tl+l1 to t8-tl+l1 + pminsw m6, m3 + pminsw m5, m3 + pmaxsw m6, m4 ; Clamp to the bit-depth + pmaxsw m5, m4 + mova [dstq ], m6 + mova [dstq +16], m5 + paddw m6, m2, m0 + paddw m2, m1 + pminsw m6, m3 + pminsw m2, m3 + pmaxsw m6, m4 + pmaxsw m2, m4 + mova [dstq+strideq*2 ], m6 + mova [dstq+strideq*2+16], m2 + lea dstq, [dstq+strideq*4] + inc lineq + lea leftq, [leftq+4] + + jnz .loop + REP_RET + +INIT_XMM sse2 +cglobal highbd_tm_predictor_32x32, 5, 5, 8, dst, stride, above, left, bps + movd m0, [aboveq-2] + mova m1, [aboveq] + mova m2, [aboveq+16] + mova m3, [aboveq+32] + mova m4, [aboveq+48] + pshuflw m0, m0, 0x0 + ; Get the values to compute the maximum value at this bit depth + pcmpeqw m5, m5 + movd m6, bpsd + psllw m5, m6 + pcmpeqw m7, m7 + pxor m6, m6 ; min possible value + pxor m5, m7 ; max possible value + punpcklqdq m0, m0 + DEFINE_ARGS dst, stride, line, left + mov lineq, -16 + psubw m1, m0 + psubw m2, m0 + psubw m3, m0 + psubw m4, m0 +.loop: + movd m7, [leftq] + pshuflw m7, m7, 0x0 + punpcklqdq m7, m7 ; l1 l1 l1 l1 l1 l1 l1 l1 + paddw m0, m7, m1 + pminsw m0, m5 + pmaxsw m0, m6 + mova [dstq ], m0 + paddw m0, m7, m2 + pminsw m0, m5 + pmaxsw m0, m6 + mova [dstq +16], m0 + paddw m0, m7, m3 + pminsw m0, m5 + pmaxsw m0, m6 + mova [dstq +32], m0 + paddw m0, m7, m4 + pminsw m0, m5 + pmaxsw m0, m6 + mova [dstq +48], m0 + movd m7, [leftq+2] + pshuflw m7, m7, 0x0 + punpcklqdq m7, m7 ; l2 l2 l2 l2 l2 l2 l2 l2 + paddw m0, m7, m1 + pminsw m0, m5 + pmaxsw m0, m6 + mova [dstq+strideq*2 ], m0 + paddw m0, m7, m2 + pminsw m0, m5 + pmaxsw m0, m6 + mova [dstq+strideq*2+16], m0 + paddw m0, m7, m3 + pminsw m0, m5 + pmaxsw m0, m6 + mova [dstq+strideq*2+32], m0 + paddw m0, m7, m4 + pminsw m0, m5 + pmaxsw m0, m6 + mova [dstq+strideq*2+48], m0 + lea dstq, [dstq+strideq*4] + lea leftq, [leftq+4] + inc lineq + jnz .loop + REP_RET diff --git a/third_party/aom/aom_dsp/x86/highbd_loopfilter_sse2.c b/third_party/aom/aom_dsp/x86/highbd_loopfilter_sse2.c new file mode 100644 index 000000000..76369871b --- /dev/null +++ b/third_party/aom/aom_dsp/x86/highbd_loopfilter_sse2.c @@ -0,0 +1,1140 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include <emmintrin.h> // SSE2 + +#include "./aom_dsp_rtcd.h" +#include "aom_ports/mem.h" +#include "aom_ports/emmintrin_compat.h" + +static INLINE __m128i signed_char_clamp_bd_sse2(__m128i value, int bd) { + __m128i ubounded; + __m128i lbounded; + __m128i retval; + + const __m128i zero = _mm_set1_epi16(0); + const __m128i one = _mm_set1_epi16(1); + __m128i t80, max, min; + + if (bd == 8) { + t80 = _mm_set1_epi16(0x80); + max = _mm_subs_epi16(_mm_subs_epi16(_mm_slli_epi16(one, 8), one), t80); + } else if (bd == 10) { + t80 = _mm_set1_epi16(0x200); + max = _mm_subs_epi16(_mm_subs_epi16(_mm_slli_epi16(one, 10), one), t80); + } else { // bd == 12 + t80 = _mm_set1_epi16(0x800); + max = _mm_subs_epi16(_mm_subs_epi16(_mm_slli_epi16(one, 12), one), t80); + } + + min = _mm_subs_epi16(zero, t80); + + ubounded = _mm_cmpgt_epi16(value, max); + lbounded = _mm_cmplt_epi16(value, min); + retval = _mm_andnot_si128(_mm_or_si128(ubounded, lbounded), value); + ubounded = _mm_and_si128(ubounded, max); + lbounded = _mm_and_si128(lbounded, min); + retval = _mm_or_si128(retval, ubounded); + retval = _mm_or_si128(retval, lbounded); + return retval; +} + +// TODO(debargha, peter): Break up large functions into smaller ones +// in this file. +void aom_highbd_lpf_horizontal_edge_8_sse2(uint16_t *s, int p, + const uint8_t *_blimit, + const uint8_t *_limit, + const uint8_t *_thresh, int bd) { + const __m128i zero = _mm_set1_epi16(0); + const __m128i one = _mm_set1_epi16(1); + __m128i blimit, limit, thresh; + __m128i q7, p7, q6, p6, q5, p5, q4, p4, q3, p3, q2, p2, q1, p1, q0, p0; + __m128i mask, hev, flat, flat2, abs_p1p0, abs_q1q0; + __m128i ps1, qs1, ps0, qs0; + __m128i abs_p0q0, abs_p1q1, ffff, work; + __m128i filt, work_a, filter1, filter2; + __m128i flat2_q6, flat2_p6, flat2_q5, flat2_p5, flat2_q4, flat2_p4; + __m128i flat2_q3, flat2_p3, flat2_q2, flat2_p2, flat2_q1, flat2_p1; + __m128i flat2_q0, flat2_p0; + __m128i flat_q2, flat_p2, flat_q1, flat_p1, flat_q0, flat_p0; + __m128i pixelFilter_p, pixelFilter_q; + __m128i pixetFilter_p2p1p0, pixetFilter_q2q1q0; + __m128i sum_p7, sum_q7, sum_p3, sum_q3; + __m128i t4, t3, t80, t1; + __m128i eight, four; + + if (bd == 8) { + blimit = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_blimit), zero); + limit = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_limit), zero); + thresh = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_thresh), zero); + } else if (bd == 10) { + blimit = _mm_slli_epi16( + _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_blimit), zero), 2); + limit = _mm_slli_epi16( + _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_limit), zero), 2); + thresh = _mm_slli_epi16( + _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_thresh), zero), 2); + } else { // bd == 12 + blimit = _mm_slli_epi16( + _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_blimit), zero), 4); + limit = _mm_slli_epi16( + _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_limit), zero), 4); + thresh = _mm_slli_epi16( + _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_thresh), zero), 4); + } + + q4 = _mm_load_si128((__m128i *)(s + 4 * p)); + p4 = _mm_load_si128((__m128i *)(s - 5 * p)); + q3 = _mm_load_si128((__m128i *)(s + 3 * p)); + p3 = _mm_load_si128((__m128i *)(s - 4 * p)); + q2 = _mm_load_si128((__m128i *)(s + 2 * p)); + p2 = _mm_load_si128((__m128i *)(s - 3 * p)); + q1 = _mm_load_si128((__m128i *)(s + 1 * p)); + p1 = _mm_load_si128((__m128i *)(s - 2 * p)); + q0 = _mm_load_si128((__m128i *)(s + 0 * p)); + p0 = _mm_load_si128((__m128i *)(s - 1 * p)); + + // highbd_filter_mask + abs_p1p0 = _mm_or_si128(_mm_subs_epu16(p1, p0), _mm_subs_epu16(p0, p1)); + abs_q1q0 = _mm_or_si128(_mm_subs_epu16(q1, q0), _mm_subs_epu16(q0, q1)); + + ffff = _mm_cmpeq_epi16(abs_p1p0, abs_p1p0); + + abs_p0q0 = _mm_or_si128(_mm_subs_epu16(p0, q0), _mm_subs_epu16(q0, p0)); + abs_p1q1 = _mm_or_si128(_mm_subs_epu16(p1, q1), _mm_subs_epu16(q1, p1)); + + // highbd_hev_mask (in C code this is actually called from highbd_filter4) + flat = _mm_max_epi16(abs_p1p0, abs_q1q0); + hev = _mm_subs_epu16(flat, thresh); + hev = _mm_xor_si128(_mm_cmpeq_epi16(hev, zero), ffff); + + abs_p0q0 = _mm_adds_epu16(abs_p0q0, abs_p0q0); // abs(p0 - q0) * 2 + abs_p1q1 = _mm_srli_epi16(abs_p1q1, 1); // abs(p1 - q1) / 2 + mask = _mm_subs_epu16(_mm_adds_epu16(abs_p0q0, abs_p1q1), blimit); + mask = _mm_xor_si128(_mm_cmpeq_epi16(mask, zero), ffff); + mask = _mm_and_si128(mask, _mm_adds_epu16(limit, one)); + work = _mm_max_epi16( + _mm_or_si128(_mm_subs_epu16(p1, p0), _mm_subs_epu16(p0, p1)), + _mm_or_si128(_mm_subs_epu16(q1, q0), _mm_subs_epu16(q0, q1))); + mask = _mm_max_epi16(work, mask); + work = _mm_max_epi16( + _mm_or_si128(_mm_subs_epu16(p2, p1), _mm_subs_epu16(p1, p2)), + _mm_or_si128(_mm_subs_epu16(q2, q1), _mm_subs_epu16(q1, q2))); + mask = _mm_max_epi16(work, mask); + work = _mm_max_epi16( + _mm_or_si128(_mm_subs_epu16(p3, p2), _mm_subs_epu16(p2, p3)), + _mm_or_si128(_mm_subs_epu16(q3, q2), _mm_subs_epu16(q2, q3))); + mask = _mm_max_epi16(work, mask); + + mask = _mm_subs_epu16(mask, limit); + mask = _mm_cmpeq_epi16(mask, zero); // return ~mask + + // lp filter + // highbd_filter4 + t4 = _mm_set1_epi16(4); + t3 = _mm_set1_epi16(3); + if (bd == 8) + t80 = _mm_set1_epi16(0x80); + else if (bd == 10) + t80 = _mm_set1_epi16(0x200); + else // bd == 12 + t80 = _mm_set1_epi16(0x800); + + t1 = _mm_set1_epi16(0x1); + + ps1 = _mm_subs_epi16(p1, t80); + qs1 = _mm_subs_epi16(q1, t80); + ps0 = _mm_subs_epi16(p0, t80); + qs0 = _mm_subs_epi16(q0, t80); + + filt = _mm_and_si128(signed_char_clamp_bd_sse2(_mm_subs_epi16(ps1, qs1), bd), + hev); + work_a = _mm_subs_epi16(qs0, ps0); + filt = _mm_adds_epi16(filt, work_a); + filt = _mm_adds_epi16(filt, work_a); + filt = signed_char_clamp_bd_sse2(_mm_adds_epi16(filt, work_a), bd); + filt = _mm_and_si128(filt, mask); + filter1 = signed_char_clamp_bd_sse2(_mm_adds_epi16(filt, t4), bd); + filter2 = signed_char_clamp_bd_sse2(_mm_adds_epi16(filt, t3), bd); + + // Filter1 >> 3 + filter1 = _mm_srai_epi16(filter1, 0x3); + filter2 = _mm_srai_epi16(filter2, 0x3); + + qs0 = _mm_adds_epi16( + signed_char_clamp_bd_sse2(_mm_subs_epi16(qs0, filter1), bd), t80); + ps0 = _mm_adds_epi16( + signed_char_clamp_bd_sse2(_mm_adds_epi16(ps0, filter2), bd), t80); + filt = _mm_adds_epi16(filter1, t1); + filt = _mm_srai_epi16(filt, 1); + filt = _mm_andnot_si128(hev, filt); + qs1 = _mm_adds_epi16(signed_char_clamp_bd_sse2(_mm_subs_epi16(qs1, filt), bd), + t80); + ps1 = _mm_adds_epi16(signed_char_clamp_bd_sse2(_mm_adds_epi16(ps1, filt), bd), + t80); + + // end highbd_filter4 + // loopfilter done + + // highbd_flat_mask4 + flat = _mm_max_epi16( + _mm_or_si128(_mm_subs_epu16(p2, p0), _mm_subs_epu16(p0, p2)), + _mm_or_si128(_mm_subs_epu16(p3, p0), _mm_subs_epu16(p0, p3))); + work = _mm_max_epi16( + _mm_or_si128(_mm_subs_epu16(q2, q0), _mm_subs_epu16(q0, q2)), + _mm_or_si128(_mm_subs_epu16(q3, q0), _mm_subs_epu16(q0, q3))); + flat = _mm_max_epi16(work, flat); + work = _mm_max_epi16(abs_p1p0, abs_q1q0); + flat = _mm_max_epi16(work, flat); + + if (bd == 8) + flat = _mm_subs_epu16(flat, one); + else if (bd == 10) + flat = _mm_subs_epu16(flat, _mm_slli_epi16(one, 2)); + else // bd == 12 + flat = _mm_subs_epu16(flat, _mm_slli_epi16(one, 4)); + + flat = _mm_cmpeq_epi16(flat, zero); + // end flat_mask4 + + // flat & mask = flat && mask (as used in filter8) + // (because, in both vars, each block of 16 either all 1s or all 0s) + flat = _mm_and_si128(flat, mask); + + p5 = _mm_load_si128((__m128i *)(s - 6 * p)); + q5 = _mm_load_si128((__m128i *)(s + 5 * p)); + p6 = _mm_load_si128((__m128i *)(s - 7 * p)); + q6 = _mm_load_si128((__m128i *)(s + 6 * p)); + p7 = _mm_load_si128((__m128i *)(s - 8 * p)); + q7 = _mm_load_si128((__m128i *)(s + 7 * p)); + + // highbd_flat_mask5 (arguments passed in are p0, q0, p4-p7, q4-q7 + // but referred to as p0-p4 & q0-q4 in fn) + flat2 = _mm_max_epi16( + _mm_or_si128(_mm_subs_epu16(p4, p0), _mm_subs_epu16(p0, p4)), + _mm_or_si128(_mm_subs_epu16(q4, q0), _mm_subs_epu16(q0, q4))); + + work = _mm_max_epi16( + _mm_or_si128(_mm_subs_epu16(p5, p0), _mm_subs_epu16(p0, p5)), + _mm_or_si128(_mm_subs_epu16(q5, q0), _mm_subs_epu16(q0, q5))); + flat2 = _mm_max_epi16(work, flat2); + + work = _mm_max_epi16( + _mm_or_si128(_mm_subs_epu16(p6, p0), _mm_subs_epu16(p0, p6)), + _mm_or_si128(_mm_subs_epu16(q6, q0), _mm_subs_epu16(q0, q6))); + flat2 = _mm_max_epi16(work, flat2); + + work = _mm_max_epi16( + _mm_or_si128(_mm_subs_epu16(p7, p0), _mm_subs_epu16(p0, p7)), + _mm_or_si128(_mm_subs_epu16(q7, q0), _mm_subs_epu16(q0, q7))); + flat2 = _mm_max_epi16(work, flat2); + + if (bd == 8) + flat2 = _mm_subs_epu16(flat2, one); + else if (bd == 10) + flat2 = _mm_subs_epu16(flat2, _mm_slli_epi16(one, 2)); + else // bd == 12 + flat2 = _mm_subs_epu16(flat2, _mm_slli_epi16(one, 4)); + + flat2 = _mm_cmpeq_epi16(flat2, zero); + flat2 = _mm_and_si128(flat2, flat); // flat2 & flat & mask + // end highbd_flat_mask5 + + // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ + // flat and wide flat calculations + eight = _mm_set1_epi16(8); + four = _mm_set1_epi16(4); + + pixelFilter_p = _mm_add_epi16(_mm_add_epi16(p6, p5), _mm_add_epi16(p4, p3)); + pixelFilter_q = _mm_add_epi16(_mm_add_epi16(q6, q5), _mm_add_epi16(q4, q3)); + + pixetFilter_p2p1p0 = _mm_add_epi16(p0, _mm_add_epi16(p2, p1)); + pixelFilter_p = _mm_add_epi16(pixelFilter_p, pixetFilter_p2p1p0); + + pixetFilter_q2q1q0 = _mm_add_epi16(q0, _mm_add_epi16(q2, q1)); + pixelFilter_q = _mm_add_epi16(pixelFilter_q, pixetFilter_q2q1q0); + pixelFilter_p = + _mm_add_epi16(eight, _mm_add_epi16(pixelFilter_p, pixelFilter_q)); + pixetFilter_p2p1p0 = _mm_add_epi16( + four, _mm_add_epi16(pixetFilter_p2p1p0, pixetFilter_q2q1q0)); + flat2_p0 = + _mm_srli_epi16(_mm_add_epi16(pixelFilter_p, _mm_add_epi16(p7, p0)), 4); + flat2_q0 = + _mm_srli_epi16(_mm_add_epi16(pixelFilter_p, _mm_add_epi16(q7, q0)), 4); + flat_p0 = _mm_srli_epi16( + _mm_add_epi16(pixetFilter_p2p1p0, _mm_add_epi16(p3, p0)), 3); + flat_q0 = _mm_srli_epi16( + _mm_add_epi16(pixetFilter_p2p1p0, _mm_add_epi16(q3, q0)), 3); + + sum_p7 = _mm_add_epi16(p7, p7); + sum_q7 = _mm_add_epi16(q7, q7); + sum_p3 = _mm_add_epi16(p3, p3); + sum_q3 = _mm_add_epi16(q3, q3); + + pixelFilter_q = _mm_sub_epi16(pixelFilter_p, p6); + pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q6); + flat2_p1 = _mm_srli_epi16( + _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p1)), 4); + flat2_q1 = _mm_srli_epi16( + _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q1)), 4); + + pixetFilter_q2q1q0 = _mm_sub_epi16(pixetFilter_p2p1p0, p2); + pixetFilter_p2p1p0 = _mm_sub_epi16(pixetFilter_p2p1p0, q2); + flat_p1 = _mm_srli_epi16( + _mm_add_epi16(pixetFilter_p2p1p0, _mm_add_epi16(sum_p3, p1)), 3); + flat_q1 = _mm_srli_epi16( + _mm_add_epi16(pixetFilter_q2q1q0, _mm_add_epi16(sum_q3, q1)), 3); + + sum_p7 = _mm_add_epi16(sum_p7, p7); + sum_q7 = _mm_add_epi16(sum_q7, q7); + sum_p3 = _mm_add_epi16(sum_p3, p3); + sum_q3 = _mm_add_epi16(sum_q3, q3); + + pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q5); + pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p5); + flat2_p2 = _mm_srli_epi16( + _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p2)), 4); + flat2_q2 = _mm_srli_epi16( + _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q2)), 4); + + pixetFilter_p2p1p0 = _mm_sub_epi16(pixetFilter_p2p1p0, q1); + pixetFilter_q2q1q0 = _mm_sub_epi16(pixetFilter_q2q1q0, p1); + flat_p2 = _mm_srli_epi16( + _mm_add_epi16(pixetFilter_p2p1p0, _mm_add_epi16(sum_p3, p2)), 3); + flat_q2 = _mm_srli_epi16( + _mm_add_epi16(pixetFilter_q2q1q0, _mm_add_epi16(sum_q3, q2)), 3); + + sum_p7 = _mm_add_epi16(sum_p7, p7); + sum_q7 = _mm_add_epi16(sum_q7, q7); + pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q4); + pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p4); + flat2_p3 = _mm_srli_epi16( + _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p3)), 4); + flat2_q3 = _mm_srli_epi16( + _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q3)), 4); + + sum_p7 = _mm_add_epi16(sum_p7, p7); + sum_q7 = _mm_add_epi16(sum_q7, q7); + pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q3); + pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p3); + flat2_p4 = _mm_srli_epi16( + _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p4)), 4); + flat2_q4 = _mm_srli_epi16( + _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q4)), 4); + + sum_p7 = _mm_add_epi16(sum_p7, p7); + sum_q7 = _mm_add_epi16(sum_q7, q7); + pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q2); + pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p2); + flat2_p5 = _mm_srli_epi16( + _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p5)), 4); + flat2_q5 = _mm_srli_epi16( + _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q5)), 4); + + sum_p7 = _mm_add_epi16(sum_p7, p7); + sum_q7 = _mm_add_epi16(sum_q7, q7); + pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q1); + pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p1); + flat2_p6 = _mm_srli_epi16( + _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p6)), 4); + flat2_q6 = _mm_srli_epi16( + _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q6)), 4); + + // wide flat + // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ + + // highbd_filter8 + p2 = _mm_andnot_si128(flat, p2); + // p2 remains unchanged if !(flat && mask) + flat_p2 = _mm_and_si128(flat, flat_p2); + // when (flat && mask) + p2 = _mm_or_si128(p2, flat_p2); // full list of p2 values + q2 = _mm_andnot_si128(flat, q2); + flat_q2 = _mm_and_si128(flat, flat_q2); + q2 = _mm_or_si128(q2, flat_q2); // full list of q2 values + + ps1 = _mm_andnot_si128(flat, ps1); + // p1 takes the value assigned to in in filter4 if !(flat && mask) + flat_p1 = _mm_and_si128(flat, flat_p1); + // when (flat && mask) + p1 = _mm_or_si128(ps1, flat_p1); // full list of p1 values + qs1 = _mm_andnot_si128(flat, qs1); + flat_q1 = _mm_and_si128(flat, flat_q1); + q1 = _mm_or_si128(qs1, flat_q1); // full list of q1 values + + ps0 = _mm_andnot_si128(flat, ps0); + // p0 takes the value assigned to in in filter4 if !(flat && mask) + flat_p0 = _mm_and_si128(flat, flat_p0); + // when (flat && mask) + p0 = _mm_or_si128(ps0, flat_p0); // full list of p0 values + qs0 = _mm_andnot_si128(flat, qs0); + flat_q0 = _mm_and_si128(flat, flat_q0); + q0 = _mm_or_si128(qs0, flat_q0); // full list of q0 values + // end highbd_filter8 + + // highbd_filter16 + p6 = _mm_andnot_si128(flat2, p6); + // p6 remains unchanged if !(flat2 && flat && mask) + flat2_p6 = _mm_and_si128(flat2, flat2_p6); + // get values for when (flat2 && flat && mask) + p6 = _mm_or_si128(p6, flat2_p6); // full list of p6 values + q6 = _mm_andnot_si128(flat2, q6); + // q6 remains unchanged if !(flat2 && flat && mask) + flat2_q6 = _mm_and_si128(flat2, flat2_q6); + // get values for when (flat2 && flat && mask) + q6 = _mm_or_si128(q6, flat2_q6); // full list of q6 values + _mm_store_si128((__m128i *)(s - 7 * p), p6); + _mm_store_si128((__m128i *)(s + 6 * p), q6); + + p5 = _mm_andnot_si128(flat2, p5); + // p5 remains unchanged if !(flat2 && flat && mask) + flat2_p5 = _mm_and_si128(flat2, flat2_p5); + // get values for when (flat2 && flat && mask) + p5 = _mm_or_si128(p5, flat2_p5); + // full list of p5 values + q5 = _mm_andnot_si128(flat2, q5); + // q5 remains unchanged if !(flat2 && flat && mask) + flat2_q5 = _mm_and_si128(flat2, flat2_q5); + // get values for when (flat2 && flat && mask) + q5 = _mm_or_si128(q5, flat2_q5); + // full list of q5 values + _mm_store_si128((__m128i *)(s - 6 * p), p5); + _mm_store_si128((__m128i *)(s + 5 * p), q5); + + p4 = _mm_andnot_si128(flat2, p4); + // p4 remains unchanged if !(flat2 && flat && mask) + flat2_p4 = _mm_and_si128(flat2, flat2_p4); + // get values for when (flat2 && flat && mask) + p4 = _mm_or_si128(p4, flat2_p4); // full list of p4 values + q4 = _mm_andnot_si128(flat2, q4); + // q4 remains unchanged if !(flat2 && flat && mask) + flat2_q4 = _mm_and_si128(flat2, flat2_q4); + // get values for when (flat2 && flat && mask) + q4 = _mm_or_si128(q4, flat2_q4); // full list of q4 values + _mm_store_si128((__m128i *)(s - 5 * p), p4); + _mm_store_si128((__m128i *)(s + 4 * p), q4); + + p3 = _mm_andnot_si128(flat2, p3); + // p3 takes value from highbd_filter8 if !(flat2 && flat && mask) + flat2_p3 = _mm_and_si128(flat2, flat2_p3); + // get values for when (flat2 && flat && mask) + p3 = _mm_or_si128(p3, flat2_p3); // full list of p3 values + q3 = _mm_andnot_si128(flat2, q3); + // q3 takes value from highbd_filter8 if !(flat2 && flat && mask) + flat2_q3 = _mm_and_si128(flat2, flat2_q3); + // get values for when (flat2 && flat && mask) + q3 = _mm_or_si128(q3, flat2_q3); // full list of q3 values + _mm_store_si128((__m128i *)(s - 4 * p), p3); + _mm_store_si128((__m128i *)(s + 3 * p), q3); + + p2 = _mm_andnot_si128(flat2, p2); + // p2 takes value from highbd_filter8 if !(flat2 && flat && mask) + flat2_p2 = _mm_and_si128(flat2, flat2_p2); + // get values for when (flat2 && flat && mask) + p2 = _mm_or_si128(p2, flat2_p2); + // full list of p2 values + q2 = _mm_andnot_si128(flat2, q2); + // q2 takes value from highbd_filter8 if !(flat2 && flat && mask) + flat2_q2 = _mm_and_si128(flat2, flat2_q2); + // get values for when (flat2 && flat && mask) + q2 = _mm_or_si128(q2, flat2_q2); // full list of q2 values + _mm_store_si128((__m128i *)(s - 3 * p), p2); + _mm_store_si128((__m128i *)(s + 2 * p), q2); + + p1 = _mm_andnot_si128(flat2, p1); + // p1 takes value from highbd_filter8 if !(flat2 && flat && mask) + flat2_p1 = _mm_and_si128(flat2, flat2_p1); + // get values for when (flat2 && flat && mask) + p1 = _mm_or_si128(p1, flat2_p1); // full list of p1 values + q1 = _mm_andnot_si128(flat2, q1); + // q1 takes value from highbd_filter8 if !(flat2 && flat && mask) + flat2_q1 = _mm_and_si128(flat2, flat2_q1); + // get values for when (flat2 && flat && mask) + q1 = _mm_or_si128(q1, flat2_q1); // full list of q1 values + _mm_store_si128((__m128i *)(s - 2 * p), p1); + _mm_store_si128((__m128i *)(s + 1 * p), q1); + + p0 = _mm_andnot_si128(flat2, p0); + // p0 takes value from highbd_filter8 if !(flat2 && flat && mask) + flat2_p0 = _mm_and_si128(flat2, flat2_p0); + // get values for when (flat2 && flat && mask) + p0 = _mm_or_si128(p0, flat2_p0); // full list of p0 values + q0 = _mm_andnot_si128(flat2, q0); + // q0 takes value from highbd_filter8 if !(flat2 && flat && mask) + flat2_q0 = _mm_and_si128(flat2, flat2_q0); + // get values for when (flat2 && flat && mask) + q0 = _mm_or_si128(q0, flat2_q0); // full list of q0 values + _mm_store_si128((__m128i *)(s - 1 * p), p0); + _mm_store_si128((__m128i *)(s - 0 * p), q0); +} + +void aom_highbd_lpf_horizontal_edge_16_sse2(uint16_t *s, int p, + const uint8_t *_blimit, + const uint8_t *_limit, + const uint8_t *_thresh, int bd) { + aom_highbd_lpf_horizontal_edge_8_sse2(s, p, _blimit, _limit, _thresh, bd); + aom_highbd_lpf_horizontal_edge_8_sse2(s + 8, p, _blimit, _limit, _thresh, bd); +} + +void aom_highbd_lpf_horizontal_8_sse2(uint16_t *s, int p, + const uint8_t *_blimit, + const uint8_t *_limit, + const uint8_t *_thresh, int bd) { + DECLARE_ALIGNED(16, uint16_t, flat_op2[16]); + DECLARE_ALIGNED(16, uint16_t, flat_op1[16]); + DECLARE_ALIGNED(16, uint16_t, flat_op0[16]); + DECLARE_ALIGNED(16, uint16_t, flat_oq2[16]); + DECLARE_ALIGNED(16, uint16_t, flat_oq1[16]); + DECLARE_ALIGNED(16, uint16_t, flat_oq0[16]); + const __m128i zero = _mm_set1_epi16(0); + __m128i blimit, limit, thresh; + __m128i mask, hev, flat; + __m128i p3 = _mm_load_si128((__m128i *)(s - 4 * p)); + __m128i q3 = _mm_load_si128((__m128i *)(s + 3 * p)); + __m128i p2 = _mm_load_si128((__m128i *)(s - 3 * p)); + __m128i q2 = _mm_load_si128((__m128i *)(s + 2 * p)); + __m128i p1 = _mm_load_si128((__m128i *)(s - 2 * p)); + __m128i q1 = _mm_load_si128((__m128i *)(s + 1 * p)); + __m128i p0 = _mm_load_si128((__m128i *)(s - 1 * p)); + __m128i q0 = _mm_load_si128((__m128i *)(s + 0 * p)); + const __m128i one = _mm_set1_epi16(1); + const __m128i ffff = _mm_cmpeq_epi16(one, one); + __m128i abs_p1q1, abs_p0q0, abs_q1q0, abs_p1p0, work; + const __m128i four = _mm_set1_epi16(4); + __m128i workp_a, workp_b, workp_shft; + + const __m128i t4 = _mm_set1_epi16(4); + const __m128i t3 = _mm_set1_epi16(3); + __m128i t80; + const __m128i t1 = _mm_set1_epi16(0x1); + __m128i ps1, ps0, qs0, qs1; + __m128i filt; + __m128i work_a; + __m128i filter1, filter2; + + if (bd == 8) { + blimit = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_blimit), zero); + limit = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_limit), zero); + thresh = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_thresh), zero); + t80 = _mm_set1_epi16(0x80); + } else if (bd == 10) { + blimit = _mm_slli_epi16( + _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_blimit), zero), 2); + limit = _mm_slli_epi16( + _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_limit), zero), 2); + thresh = _mm_slli_epi16( + _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_thresh), zero), 2); + t80 = _mm_set1_epi16(0x200); + } else { // bd == 12 + blimit = _mm_slli_epi16( + _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_blimit), zero), 4); + limit = _mm_slli_epi16( + _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_limit), zero), 4); + thresh = _mm_slli_epi16( + _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_thresh), zero), 4); + t80 = _mm_set1_epi16(0x800); + } + + ps1 = _mm_subs_epi16(p1, t80); + ps0 = _mm_subs_epi16(p0, t80); + qs0 = _mm_subs_epi16(q0, t80); + qs1 = _mm_subs_epi16(q1, t80); + + // filter_mask and hev_mask + abs_p1p0 = _mm_or_si128(_mm_subs_epu16(p1, p0), _mm_subs_epu16(p0, p1)); + abs_q1q0 = _mm_or_si128(_mm_subs_epu16(q1, q0), _mm_subs_epu16(q0, q1)); + + abs_p0q0 = _mm_or_si128(_mm_subs_epu16(p0, q0), _mm_subs_epu16(q0, p0)); + abs_p1q1 = _mm_or_si128(_mm_subs_epu16(p1, q1), _mm_subs_epu16(q1, p1)); + flat = _mm_max_epi16(abs_p1p0, abs_q1q0); + hev = _mm_subs_epu16(flat, thresh); + hev = _mm_xor_si128(_mm_cmpeq_epi16(hev, zero), ffff); + + abs_p0q0 = _mm_adds_epu16(abs_p0q0, abs_p0q0); + abs_p1q1 = _mm_srli_epi16(abs_p1q1, 1); + mask = _mm_subs_epu16(_mm_adds_epu16(abs_p0q0, abs_p1q1), blimit); + mask = _mm_xor_si128(_mm_cmpeq_epi16(mask, zero), ffff); + // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 > blimit) * -1; + // So taking maximums continues to work: + mask = _mm_and_si128(mask, _mm_adds_epu16(limit, one)); + mask = _mm_max_epi16(abs_p1p0, mask); + // mask |= (abs(p1 - p0) > limit) * -1; + mask = _mm_max_epi16(abs_q1q0, mask); + // mask |= (abs(q1 - q0) > limit) * -1; + + work = _mm_max_epi16( + _mm_or_si128(_mm_subs_epu16(p2, p1), _mm_subs_epu16(p1, p2)), + _mm_or_si128(_mm_subs_epu16(q2, q1), _mm_subs_epu16(q1, q2))); + mask = _mm_max_epi16(work, mask); + work = _mm_max_epi16( + _mm_or_si128(_mm_subs_epu16(p3, p2), _mm_subs_epu16(p2, p3)), + _mm_or_si128(_mm_subs_epu16(q3, q2), _mm_subs_epu16(q2, q3))); + mask = _mm_max_epi16(work, mask); + mask = _mm_subs_epu16(mask, limit); + mask = _mm_cmpeq_epi16(mask, zero); + + // flat_mask4 + flat = _mm_max_epi16( + _mm_or_si128(_mm_subs_epu16(p2, p0), _mm_subs_epu16(p0, p2)), + _mm_or_si128(_mm_subs_epu16(q2, q0), _mm_subs_epu16(q0, q2))); + work = _mm_max_epi16( + _mm_or_si128(_mm_subs_epu16(p3, p0), _mm_subs_epu16(p0, p3)), + _mm_or_si128(_mm_subs_epu16(q3, q0), _mm_subs_epu16(q0, q3))); + flat = _mm_max_epi16(work, flat); + flat = _mm_max_epi16(abs_p1p0, flat); + flat = _mm_max_epi16(abs_q1q0, flat); + + if (bd == 8) + flat = _mm_subs_epu16(flat, one); + else if (bd == 10) + flat = _mm_subs_epu16(flat, _mm_slli_epi16(one, 2)); + else // bd == 12 + flat = _mm_subs_epu16(flat, _mm_slli_epi16(one, 4)); + + flat = _mm_cmpeq_epi16(flat, zero); + flat = _mm_and_si128(flat, mask); // flat & mask + + // Added before shift for rounding part of ROUND_POWER_OF_TWO + + workp_a = _mm_add_epi16(_mm_add_epi16(p3, p3), _mm_add_epi16(p2, p1)); + workp_a = _mm_add_epi16(_mm_add_epi16(workp_a, four), p0); + workp_b = _mm_add_epi16(_mm_add_epi16(q0, p2), p3); + workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3); + _mm_store_si128((__m128i *)&flat_op2[0], workp_shft); + + workp_b = _mm_add_epi16(_mm_add_epi16(q0, q1), p1); + workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3); + _mm_store_si128((__m128i *)&flat_op1[0], workp_shft); + + workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p3), q2); + workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, p1), p0); + workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3); + _mm_store_si128((__m128i *)&flat_op0[0], workp_shft); + + workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p3), q3); + workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, p0), q0); + workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3); + _mm_store_si128((__m128i *)&flat_oq0[0], workp_shft); + + workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p2), q3); + workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, q0), q1); + workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3); + _mm_store_si128((__m128i *)&flat_oq1[0], workp_shft); + + workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p1), q3); + workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, q1), q2); + workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3); + _mm_store_si128((__m128i *)&flat_oq2[0], workp_shft); + + // lp filter + filt = signed_char_clamp_bd_sse2(_mm_subs_epi16(ps1, qs1), bd); + filt = _mm_and_si128(filt, hev); + work_a = _mm_subs_epi16(qs0, ps0); + filt = _mm_adds_epi16(filt, work_a); + filt = _mm_adds_epi16(filt, work_a); + filt = _mm_adds_epi16(filt, work_a); + // (aom_filter + 3 * (qs0 - ps0)) & mask + filt = signed_char_clamp_bd_sse2(filt, bd); + filt = _mm_and_si128(filt, mask); + + filter1 = _mm_adds_epi16(filt, t4); + filter2 = _mm_adds_epi16(filt, t3); + + // Filter1 >> 3 + filter1 = signed_char_clamp_bd_sse2(filter1, bd); + filter1 = _mm_srai_epi16(filter1, 3); + + // Filter2 >> 3 + filter2 = signed_char_clamp_bd_sse2(filter2, bd); + filter2 = _mm_srai_epi16(filter2, 3); + + // filt >> 1 + filt = _mm_adds_epi16(filter1, t1); + filt = _mm_srai_epi16(filt, 1); + // filter = ROUND_POWER_OF_TWO(filter1, 1) & ~hev; + filt = _mm_andnot_si128(hev, filt); + + work_a = signed_char_clamp_bd_sse2(_mm_subs_epi16(qs0, filter1), bd); + work_a = _mm_adds_epi16(work_a, t80); + q0 = _mm_load_si128((__m128i *)flat_oq0); + work_a = _mm_andnot_si128(flat, work_a); + q0 = _mm_and_si128(flat, q0); + q0 = _mm_or_si128(work_a, q0); + + work_a = signed_char_clamp_bd_sse2(_mm_subs_epi16(qs1, filt), bd); + work_a = _mm_adds_epi16(work_a, t80); + q1 = _mm_load_si128((__m128i *)flat_oq1); + work_a = _mm_andnot_si128(flat, work_a); + q1 = _mm_and_si128(flat, q1); + q1 = _mm_or_si128(work_a, q1); + + work_a = _mm_loadu_si128((__m128i *)(s + 2 * p)); + q2 = _mm_load_si128((__m128i *)flat_oq2); + work_a = _mm_andnot_si128(flat, work_a); + q2 = _mm_and_si128(flat, q2); + q2 = _mm_or_si128(work_a, q2); + + work_a = signed_char_clamp_bd_sse2(_mm_adds_epi16(ps0, filter2), bd); + work_a = _mm_adds_epi16(work_a, t80); + p0 = _mm_load_si128((__m128i *)flat_op0); + work_a = _mm_andnot_si128(flat, work_a); + p0 = _mm_and_si128(flat, p0); + p0 = _mm_or_si128(work_a, p0); + + work_a = signed_char_clamp_bd_sse2(_mm_adds_epi16(ps1, filt), bd); + work_a = _mm_adds_epi16(work_a, t80); + p1 = _mm_load_si128((__m128i *)flat_op1); + work_a = _mm_andnot_si128(flat, work_a); + p1 = _mm_and_si128(flat, p1); + p1 = _mm_or_si128(work_a, p1); + + work_a = _mm_loadu_si128((__m128i *)(s - 3 * p)); + p2 = _mm_load_si128((__m128i *)flat_op2); + work_a = _mm_andnot_si128(flat, work_a); + p2 = _mm_and_si128(flat, p2); + p2 = _mm_or_si128(work_a, p2); + + _mm_store_si128((__m128i *)(s - 3 * p), p2); + _mm_store_si128((__m128i *)(s - 2 * p), p1); + _mm_store_si128((__m128i *)(s - 1 * p), p0); + _mm_store_si128((__m128i *)(s + 0 * p), q0); + _mm_store_si128((__m128i *)(s + 1 * p), q1); + _mm_store_si128((__m128i *)(s + 2 * p), q2); +} + +void aom_highbd_lpf_horizontal_8_dual_sse2( + uint16_t *s, int p, const uint8_t *_blimit0, const uint8_t *_limit0, + const uint8_t *_thresh0, const uint8_t *_blimit1, const uint8_t *_limit1, + const uint8_t *_thresh1, int bd) { + aom_highbd_lpf_horizontal_8_sse2(s, p, _blimit0, _limit0, _thresh0, bd); + aom_highbd_lpf_horizontal_8_sse2(s + 8, p, _blimit1, _limit1, _thresh1, bd); +} + +void aom_highbd_lpf_horizontal_4_sse2(uint16_t *s, int p, + const uint8_t *_blimit, + const uint8_t *_limit, + const uint8_t *_thresh, int bd) { + const __m128i zero = _mm_set1_epi16(0); + __m128i blimit, limit, thresh; + __m128i mask, hev, flat; + __m128i p3 = _mm_loadu_si128((__m128i *)(s - 4 * p)); + __m128i p2 = _mm_loadu_si128((__m128i *)(s - 3 * p)); + __m128i p1 = _mm_loadu_si128((__m128i *)(s - 2 * p)); + __m128i p0 = _mm_loadu_si128((__m128i *)(s - 1 * p)); + __m128i q0 = _mm_loadu_si128((__m128i *)(s - 0 * p)); + __m128i q1 = _mm_loadu_si128((__m128i *)(s + 1 * p)); + __m128i q2 = _mm_loadu_si128((__m128i *)(s + 2 * p)); + __m128i q3 = _mm_loadu_si128((__m128i *)(s + 3 * p)); + const __m128i abs_p1p0 = + _mm_or_si128(_mm_subs_epu16(p1, p0), _mm_subs_epu16(p0, p1)); + const __m128i abs_q1q0 = + _mm_or_si128(_mm_subs_epu16(q1, q0), _mm_subs_epu16(q0, q1)); + const __m128i ffff = _mm_cmpeq_epi16(abs_p1p0, abs_p1p0); + const __m128i one = _mm_set1_epi16(1); + __m128i abs_p0q0 = + _mm_or_si128(_mm_subs_epu16(p0, q0), _mm_subs_epu16(q0, p0)); + __m128i abs_p1q1 = + _mm_or_si128(_mm_subs_epu16(p1, q1), _mm_subs_epu16(q1, p1)); + __m128i work; + const __m128i t4 = _mm_set1_epi16(4); + const __m128i t3 = _mm_set1_epi16(3); + __m128i t80; + __m128i tff80; + __m128i tffe0; + __m128i t1f; + // equivalent to shifting 0x1f left by bitdepth - 8 + // and setting new bits to 1 + const __m128i t1 = _mm_set1_epi16(0x1); + __m128i t7f; + // equivalent to shifting 0x7f left by bitdepth - 8 + // and setting new bits to 1 + __m128i ps1, ps0, qs0, qs1; + __m128i filt; + __m128i work_a; + __m128i filter1, filter2; + + if (bd == 8) { + blimit = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_blimit), zero); + limit = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_limit), zero); + thresh = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_thresh), zero); + t80 = _mm_set1_epi16(0x80); + tff80 = _mm_set1_epi16(0xff80); + tffe0 = _mm_set1_epi16(0xffe0); + t1f = _mm_srli_epi16(_mm_set1_epi16(0x1fff), 8); + t7f = _mm_srli_epi16(_mm_set1_epi16(0x7fff), 8); + } else if (bd == 10) { + blimit = _mm_slli_epi16( + _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_blimit), zero), 2); + limit = _mm_slli_epi16( + _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_limit), zero), 2); + thresh = _mm_slli_epi16( + _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_thresh), zero), 2); + t80 = _mm_slli_epi16(_mm_set1_epi16(0x80), 2); + tff80 = _mm_slli_epi16(_mm_set1_epi16(0xff80), 2); + tffe0 = _mm_slli_epi16(_mm_set1_epi16(0xffe0), 2); + t1f = _mm_srli_epi16(_mm_set1_epi16(0x1fff), 6); + t7f = _mm_srli_epi16(_mm_set1_epi16(0x7fff), 6); + } else { // bd == 12 + blimit = _mm_slli_epi16( + _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_blimit), zero), 4); + limit = _mm_slli_epi16( + _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_limit), zero), 4); + thresh = _mm_slli_epi16( + _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_thresh), zero), 4); + t80 = _mm_slli_epi16(_mm_set1_epi16(0x80), 4); + tff80 = _mm_slli_epi16(_mm_set1_epi16(0xff80), 4); + tffe0 = _mm_slli_epi16(_mm_set1_epi16(0xffe0), 4); + t1f = _mm_srli_epi16(_mm_set1_epi16(0x1fff), 4); + t7f = _mm_srli_epi16(_mm_set1_epi16(0x7fff), 4); + } + + ps1 = _mm_subs_epi16(_mm_loadu_si128((__m128i *)(s - 2 * p)), t80); + ps0 = _mm_subs_epi16(_mm_loadu_si128((__m128i *)(s - 1 * p)), t80); + qs0 = _mm_subs_epi16(_mm_loadu_si128((__m128i *)(s + 0 * p)), t80); + qs1 = _mm_subs_epi16(_mm_loadu_si128((__m128i *)(s + 1 * p)), t80); + + // filter_mask and hev_mask + flat = _mm_max_epi16(abs_p1p0, abs_q1q0); + hev = _mm_subs_epu16(flat, thresh); + hev = _mm_xor_si128(_mm_cmpeq_epi16(hev, zero), ffff); + + abs_p0q0 = _mm_adds_epu16(abs_p0q0, abs_p0q0); + abs_p1q1 = _mm_srli_epi16(abs_p1q1, 1); + mask = _mm_subs_epu16(_mm_adds_epu16(abs_p0q0, abs_p1q1), blimit); + mask = _mm_xor_si128(_mm_cmpeq_epi16(mask, zero), ffff); + // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 > blimit) * -1; + // So taking maximums continues to work: + mask = _mm_and_si128(mask, _mm_adds_epu16(limit, one)); + mask = _mm_max_epi16(flat, mask); + // mask |= (abs(p1 - p0) > limit) * -1; + // mask |= (abs(q1 - q0) > limit) * -1; + work = _mm_max_epi16( + _mm_or_si128(_mm_subs_epu16(p2, p1), _mm_subs_epu16(p1, p2)), + _mm_or_si128(_mm_subs_epu16(p3, p2), _mm_subs_epu16(p2, p3))); + mask = _mm_max_epi16(work, mask); + work = _mm_max_epi16( + _mm_or_si128(_mm_subs_epu16(q2, q1), _mm_subs_epu16(q1, q2)), + _mm_or_si128(_mm_subs_epu16(q3, q2), _mm_subs_epu16(q2, q3))); + mask = _mm_max_epi16(work, mask); + mask = _mm_subs_epu16(mask, limit); + mask = _mm_cmpeq_epi16(mask, zero); + + // filter4 + filt = signed_char_clamp_bd_sse2(_mm_subs_epi16(ps1, qs1), bd); + filt = _mm_and_si128(filt, hev); + work_a = _mm_subs_epi16(qs0, ps0); + filt = _mm_adds_epi16(filt, work_a); + filt = _mm_adds_epi16(filt, work_a); + filt = signed_char_clamp_bd_sse2(_mm_adds_epi16(filt, work_a), bd); + + // (aom_filter + 3 * (qs0 - ps0)) & mask + filt = _mm_and_si128(filt, mask); + + filter1 = signed_char_clamp_bd_sse2(_mm_adds_epi16(filt, t4), bd); + filter2 = signed_char_clamp_bd_sse2(_mm_adds_epi16(filt, t3), bd); + + // Filter1 >> 3 + work_a = _mm_cmpgt_epi16(zero, filter1); // get the values that are <0 + filter1 = _mm_srli_epi16(filter1, 3); + work_a = _mm_and_si128(work_a, tffe0); // sign bits for the values < 0 + filter1 = _mm_and_si128(filter1, t1f); // clamp the range + filter1 = _mm_or_si128(filter1, work_a); // reinsert the sign bits + + // Filter2 >> 3 + work_a = _mm_cmpgt_epi16(zero, filter2); + filter2 = _mm_srli_epi16(filter2, 3); + work_a = _mm_and_si128(work_a, tffe0); + filter2 = _mm_and_si128(filter2, t1f); + filter2 = _mm_or_si128(filter2, work_a); + + // filt >> 1 + filt = _mm_adds_epi16(filter1, t1); + work_a = _mm_cmpgt_epi16(zero, filt); + filt = _mm_srli_epi16(filt, 1); + work_a = _mm_and_si128(work_a, tff80); + filt = _mm_and_si128(filt, t7f); + filt = _mm_or_si128(filt, work_a); + + filt = _mm_andnot_si128(hev, filt); + + q0 = _mm_adds_epi16( + signed_char_clamp_bd_sse2(_mm_subs_epi16(qs0, filter1), bd), t80); + q1 = _mm_adds_epi16(signed_char_clamp_bd_sse2(_mm_subs_epi16(qs1, filt), bd), + t80); + p0 = _mm_adds_epi16( + signed_char_clamp_bd_sse2(_mm_adds_epi16(ps0, filter2), bd), t80); + p1 = _mm_adds_epi16(signed_char_clamp_bd_sse2(_mm_adds_epi16(ps1, filt), bd), + t80); + + _mm_storeu_si128((__m128i *)(s - 2 * p), p1); + _mm_storeu_si128((__m128i *)(s - 1 * p), p0); + _mm_storeu_si128((__m128i *)(s + 0 * p), q0); + _mm_storeu_si128((__m128i *)(s + 1 * p), q1); +} + +void aom_highbd_lpf_horizontal_4_dual_sse2( + uint16_t *s, int p, const uint8_t *_blimit0, const uint8_t *_limit0, + const uint8_t *_thresh0, const uint8_t *_blimit1, const uint8_t *_limit1, + const uint8_t *_thresh1, int bd) { + aom_highbd_lpf_horizontal_4_sse2(s, p, _blimit0, _limit0, _thresh0, bd); + aom_highbd_lpf_horizontal_4_sse2(s + 8, p, _blimit1, _limit1, _thresh1, bd); +} + +static INLINE void highbd_transpose(uint16_t *src[], int in_p, uint16_t *dst[], + int out_p, int num_8x8_to_transpose) { + int idx8x8 = 0; + __m128i p0, p1, p2, p3, p4, p5, p6, p7, x0, x1, x2, x3, x4, x5, x6, x7; + do { + uint16_t *in = src[idx8x8]; + uint16_t *out = dst[idx8x8]; + + p0 = + _mm_loadu_si128((__m128i *)(in + 0 * in_p)); // 00 01 02 03 04 05 06 07 + p1 = + _mm_loadu_si128((__m128i *)(in + 1 * in_p)); // 10 11 12 13 14 15 16 17 + p2 = + _mm_loadu_si128((__m128i *)(in + 2 * in_p)); // 20 21 22 23 24 25 26 27 + p3 = + _mm_loadu_si128((__m128i *)(in + 3 * in_p)); // 30 31 32 33 34 35 36 37 + p4 = + _mm_loadu_si128((__m128i *)(in + 4 * in_p)); // 40 41 42 43 44 45 46 47 + p5 = + _mm_loadu_si128((__m128i *)(in + 5 * in_p)); // 50 51 52 53 54 55 56 57 + p6 = + _mm_loadu_si128((__m128i *)(in + 6 * in_p)); // 60 61 62 63 64 65 66 67 + p7 = + _mm_loadu_si128((__m128i *)(in + 7 * in_p)); // 70 71 72 73 74 75 76 77 + // 00 10 01 11 02 12 03 13 + x0 = _mm_unpacklo_epi16(p0, p1); + // 20 30 21 31 22 32 23 33 + x1 = _mm_unpacklo_epi16(p2, p3); + // 40 50 41 51 42 52 43 53 + x2 = _mm_unpacklo_epi16(p4, p5); + // 60 70 61 71 62 72 63 73 + x3 = _mm_unpacklo_epi16(p6, p7); + // 00 10 20 30 01 11 21 31 + x4 = _mm_unpacklo_epi32(x0, x1); + // 40 50 60 70 41 51 61 71 + x5 = _mm_unpacklo_epi32(x2, x3); + // 00 10 20 30 40 50 60 70 + x6 = _mm_unpacklo_epi64(x4, x5); + // 01 11 21 31 41 51 61 71 + x7 = _mm_unpackhi_epi64(x4, x5); + + _mm_storeu_si128((__m128i *)(out + 0 * out_p), x6); + // 00 10 20 30 40 50 60 70 + _mm_storeu_si128((__m128i *)(out + 1 * out_p), x7); + // 01 11 21 31 41 51 61 71 + + // 02 12 22 32 03 13 23 33 + x4 = _mm_unpackhi_epi32(x0, x1); + // 42 52 62 72 43 53 63 73 + x5 = _mm_unpackhi_epi32(x2, x3); + // 02 12 22 32 42 52 62 72 + x6 = _mm_unpacklo_epi64(x4, x5); + // 03 13 23 33 43 53 63 73 + x7 = _mm_unpackhi_epi64(x4, x5); + + _mm_storeu_si128((__m128i *)(out + 2 * out_p), x6); + // 02 12 22 32 42 52 62 72 + _mm_storeu_si128((__m128i *)(out + 3 * out_p), x7); + // 03 13 23 33 43 53 63 73 + + // 04 14 05 15 06 16 07 17 + x0 = _mm_unpackhi_epi16(p0, p1); + // 24 34 25 35 26 36 27 37 + x1 = _mm_unpackhi_epi16(p2, p3); + // 44 54 45 55 46 56 47 57 + x2 = _mm_unpackhi_epi16(p4, p5); + // 64 74 65 75 66 76 67 77 + x3 = _mm_unpackhi_epi16(p6, p7); + // 04 14 24 34 05 15 25 35 + x4 = _mm_unpacklo_epi32(x0, x1); + // 44 54 64 74 45 55 65 75 + x5 = _mm_unpacklo_epi32(x2, x3); + // 04 14 24 34 44 54 64 74 + x6 = _mm_unpacklo_epi64(x4, x5); + // 05 15 25 35 45 55 65 75 + x7 = _mm_unpackhi_epi64(x4, x5); + + _mm_storeu_si128((__m128i *)(out + 4 * out_p), x6); + // 04 14 24 34 44 54 64 74 + _mm_storeu_si128((__m128i *)(out + 5 * out_p), x7); + // 05 15 25 35 45 55 65 75 + + // 06 16 26 36 07 17 27 37 + x4 = _mm_unpackhi_epi32(x0, x1); + // 46 56 66 76 47 57 67 77 + x5 = _mm_unpackhi_epi32(x2, x3); + // 06 16 26 36 46 56 66 76 + x6 = _mm_unpacklo_epi64(x4, x5); + // 07 17 27 37 47 57 67 77 + x7 = _mm_unpackhi_epi64(x4, x5); + + _mm_storeu_si128((__m128i *)(out + 6 * out_p), x6); + // 06 16 26 36 46 56 66 76 + _mm_storeu_si128((__m128i *)(out + 7 * out_p), x7); + // 07 17 27 37 47 57 67 77 + } while (++idx8x8 < num_8x8_to_transpose); +} + +static INLINE void highbd_transpose8x16(uint16_t *in0, uint16_t *in1, int in_p, + uint16_t *out, int out_p) { + uint16_t *src0[1]; + uint16_t *src1[1]; + uint16_t *dest0[1]; + uint16_t *dest1[1]; + src0[0] = in0; + src1[0] = in1; + dest0[0] = out; + dest1[0] = out + 8; + highbd_transpose(src0, in_p, dest0, out_p, 1); + highbd_transpose(src1, in_p, dest1, out_p, 1); +} + +void aom_highbd_lpf_vertical_4_sse2(uint16_t *s, int p, const uint8_t *blimit, + const uint8_t *limit, const uint8_t *thresh, + int bd) { + DECLARE_ALIGNED(16, uint16_t, t_dst[8 * 8]); + uint16_t *src[1]; + uint16_t *dst[1]; + + // Transpose 8x8 + src[0] = s - 4; + dst[0] = t_dst; + + highbd_transpose(src, p, dst, 8, 1); + + // Loop filtering + aom_highbd_lpf_horizontal_4_sse2(t_dst + 4 * 8, 8, blimit, limit, thresh, bd); + + src[0] = t_dst; + dst[0] = s - 4; + + // Transpose back + highbd_transpose(src, 8, dst, p, 1); +} + +void aom_highbd_lpf_vertical_4_dual_sse2( + uint16_t *s, int p, const uint8_t *blimit0, const uint8_t *limit0, + const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, + const uint8_t *thresh1, int bd) { + DECLARE_ALIGNED(16, uint16_t, t_dst[16 * 8]); + uint16_t *src[2]; + uint16_t *dst[2]; + + // Transpose 8x16 + highbd_transpose8x16(s - 4, s - 4 + p * 8, p, t_dst, 16); + + // Loop filtering + aom_highbd_lpf_horizontal_4_dual_sse2(t_dst + 4 * 16, 16, blimit0, limit0, + thresh0, blimit1, limit1, thresh1, bd); + src[0] = t_dst; + src[1] = t_dst + 8; + dst[0] = s - 4; + dst[1] = s - 4 + p * 8; + + // Transpose back + highbd_transpose(src, 16, dst, p, 2); +} + +void aom_highbd_lpf_vertical_8_sse2(uint16_t *s, int p, const uint8_t *blimit, + const uint8_t *limit, const uint8_t *thresh, + int bd) { + DECLARE_ALIGNED(16, uint16_t, t_dst[8 * 8]); + uint16_t *src[1]; + uint16_t *dst[1]; + + // Transpose 8x8 + src[0] = s - 4; + dst[0] = t_dst; + + highbd_transpose(src, p, dst, 8, 1); + + // Loop filtering + aom_highbd_lpf_horizontal_8_sse2(t_dst + 4 * 8, 8, blimit, limit, thresh, bd); + + src[0] = t_dst; + dst[0] = s - 4; + + // Transpose back + highbd_transpose(src, 8, dst, p, 1); +} + +void aom_highbd_lpf_vertical_8_dual_sse2( + uint16_t *s, int p, const uint8_t *blimit0, const uint8_t *limit0, + const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, + const uint8_t *thresh1, int bd) { + DECLARE_ALIGNED(16, uint16_t, t_dst[16 * 8]); + uint16_t *src[2]; + uint16_t *dst[2]; + + // Transpose 8x16 + highbd_transpose8x16(s - 4, s - 4 + p * 8, p, t_dst, 16); + + // Loop filtering + aom_highbd_lpf_horizontal_8_dual_sse2(t_dst + 4 * 16, 16, blimit0, limit0, + thresh0, blimit1, limit1, thresh1, bd); + src[0] = t_dst; + src[1] = t_dst + 8; + + dst[0] = s - 4; + dst[1] = s - 4 + p * 8; + + // Transpose back + highbd_transpose(src, 16, dst, p, 2); +} + +void aom_highbd_lpf_vertical_16_sse2(uint16_t *s, int p, const uint8_t *blimit, + const uint8_t *limit, + const uint8_t *thresh, int bd) { + DECLARE_ALIGNED(16, uint16_t, t_dst[8 * 16]); + uint16_t *src[2]; + uint16_t *dst[2]; + + src[0] = s - 8; + src[1] = s; + dst[0] = t_dst; + dst[1] = t_dst + 8 * 8; + + // Transpose 16x8 + highbd_transpose(src, p, dst, 8, 2); + + // Loop filtering + aom_highbd_lpf_horizontal_edge_8_sse2(t_dst + 8 * 8, 8, blimit, limit, thresh, + bd); + src[0] = t_dst; + src[1] = t_dst + 8 * 8; + dst[0] = s - 8; + dst[1] = s; + + // Transpose back + highbd_transpose(src, 8, dst, p, 2); +} + +void aom_highbd_lpf_vertical_16_dual_sse2(uint16_t *s, int p, + const uint8_t *blimit, + const uint8_t *limit, + const uint8_t *thresh, int bd) { + DECLARE_ALIGNED(16, uint16_t, t_dst[256]); + + // Transpose 16x16 + highbd_transpose8x16(s - 8, s - 8 + 8 * p, p, t_dst, 16); + highbd_transpose8x16(s, s + 8 * p, p, t_dst + 8 * 16, 16); + + // Loop filtering + aom_highbd_lpf_horizontal_edge_16_sse2(t_dst + 8 * 16, 16, blimit, limit, + thresh, bd); + + // Transpose back + highbd_transpose8x16(t_dst, t_dst + 8 * 16, 16, s - 8, p); + highbd_transpose8x16(t_dst + 8, t_dst + 8 + 8 * 16, 16, s - 8 + 8 * p, p); +} diff --git a/third_party/aom/aom_dsp/x86/highbd_quantize_intrin_sse2.c b/third_party/aom/aom_dsp/x86/highbd_quantize_intrin_sse2.c new file mode 100644 index 000000000..3ee24ab16 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/highbd_quantize_intrin_sse2.c @@ -0,0 +1,155 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include <emmintrin.h> + +#include "aom_dsp/aom_dsp_common.h" +#include "aom_mem/aom_mem.h" +#include "aom_ports/mem.h" + +#if CONFIG_HIGHBITDEPTH +void aom_highbd_quantize_b_sse2(const tran_low_t *coeff_ptr, intptr_t count, + int skip_block, const int16_t *zbin_ptr, + const int16_t *round_ptr, + const int16_t *quant_ptr, + const int16_t *quant_shift_ptr, + tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, + const int16_t *dequant_ptr, uint16_t *eob_ptr, + const int16_t *scan, const int16_t *iscan) { + int i, j, non_zero_regs = (int)count / 4, eob_i = -1; + __m128i zbins[2]; + __m128i nzbins[2]; + + zbins[0] = _mm_set_epi32((int)zbin_ptr[1], (int)zbin_ptr[1], (int)zbin_ptr[1], + (int)zbin_ptr[0]); + zbins[1] = _mm_set1_epi32((int)zbin_ptr[1]); + + nzbins[0] = _mm_setzero_si128(); + nzbins[1] = _mm_setzero_si128(); + nzbins[0] = _mm_sub_epi32(nzbins[0], zbins[0]); + nzbins[1] = _mm_sub_epi32(nzbins[1], zbins[1]); + + (void)scan; + + memset(qcoeff_ptr, 0, count * sizeof(*qcoeff_ptr)); + memset(dqcoeff_ptr, 0, count * sizeof(*dqcoeff_ptr)); + + if (!skip_block) { + // Pre-scan pass + for (i = ((int)count / 4) - 1; i >= 0; i--) { + __m128i coeffs, cmp1, cmp2; + int test; + coeffs = _mm_load_si128((const __m128i *)(coeff_ptr + i * 4)); + cmp1 = _mm_cmplt_epi32(coeffs, zbins[i != 0]); + cmp2 = _mm_cmpgt_epi32(coeffs, nzbins[i != 0]); + cmp1 = _mm_and_si128(cmp1, cmp2); + test = _mm_movemask_epi8(cmp1); + if (test == 0xffff) + non_zero_regs--; + else + break; + } + + // Quantization pass: + for (i = 0; i < non_zero_regs; i++) { + __m128i coeffs, coeffs_sign, tmp1, tmp2; + int test; + int abs_coeff[4]; + int coeff_sign[4]; + + coeffs = _mm_load_si128((const __m128i *)(coeff_ptr + i * 4)); + coeffs_sign = _mm_srai_epi32(coeffs, 31); + coeffs = _mm_sub_epi32(_mm_xor_si128(coeffs, coeffs_sign), coeffs_sign); + tmp1 = _mm_cmpgt_epi32(coeffs, zbins[i != 0]); + tmp2 = _mm_cmpeq_epi32(coeffs, zbins[i != 0]); + tmp1 = _mm_or_si128(tmp1, tmp2); + test = _mm_movemask_epi8(tmp1); + _mm_storeu_si128((__m128i *)abs_coeff, coeffs); + _mm_storeu_si128((__m128i *)coeff_sign, coeffs_sign); + + for (j = 0; j < 4; j++) { + if (test & (1 << (4 * j))) { + int k = 4 * i + j; + const int64_t tmp3 = abs_coeff[j] + round_ptr[k != 0]; + const int64_t tmp4 = ((tmp3 * quant_ptr[k != 0]) >> 16) + tmp3; + const uint32_t abs_qcoeff = + (uint32_t)((tmp4 * quant_shift_ptr[k != 0]) >> 16); + qcoeff_ptr[k] = (int)(abs_qcoeff ^ coeff_sign[j]) - coeff_sign[j]; + dqcoeff_ptr[k] = qcoeff_ptr[k] * dequant_ptr[k != 0]; + if (abs_qcoeff) eob_i = iscan[k] > eob_i ? iscan[k] : eob_i; + } + } + } + } + *eob_ptr = eob_i + 1; +} + +void aom_highbd_quantize_b_32x32_sse2( + const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, + const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, + const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, + tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, + const int16_t *scan, const int16_t *iscan) { + __m128i zbins[2]; + __m128i nzbins[2]; + int idx = 0; + int idx_arr[1024]; + int i, eob = -1; + const int zbin0_tmp = ROUND_POWER_OF_TWO(zbin_ptr[0], 1); + const int zbin1_tmp = ROUND_POWER_OF_TWO(zbin_ptr[1], 1); + (void)scan; + zbins[0] = _mm_set_epi32(zbin1_tmp, zbin1_tmp, zbin1_tmp, zbin0_tmp); + zbins[1] = _mm_set1_epi32(zbin1_tmp); + + nzbins[0] = _mm_setzero_si128(); + nzbins[1] = _mm_setzero_si128(); + nzbins[0] = _mm_sub_epi32(nzbins[0], zbins[0]); + nzbins[1] = _mm_sub_epi32(nzbins[1], zbins[1]); + + memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr)); + memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr)); + + if (!skip_block) { + // Pre-scan pass + for (i = 0; i < n_coeffs / 4; i++) { + __m128i coeffs, cmp1, cmp2; + int test; + coeffs = _mm_load_si128((const __m128i *)(coeff_ptr + i * 4)); + cmp1 = _mm_cmplt_epi32(coeffs, zbins[i != 0]); + cmp2 = _mm_cmpgt_epi32(coeffs, nzbins[i != 0]); + cmp1 = _mm_and_si128(cmp1, cmp2); + test = _mm_movemask_epi8(cmp1); + if (!(test & 0xf)) idx_arr[idx++] = i * 4; + if (!(test & 0xf0)) idx_arr[idx++] = i * 4 + 1; + if (!(test & 0xf00)) idx_arr[idx++] = i * 4 + 2; + if (!(test & 0xf000)) idx_arr[idx++] = i * 4 + 3; + } + + // Quantization pass: only process the coefficients selected in + // pre-scan pass. Note: idx can be zero. + for (i = 0; i < idx; i++) { + const int rc = idx_arr[i]; + const int coeff = coeff_ptr[rc]; + const int coeff_sign = (coeff >> 31); + const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign; + const int64_t tmp1 = + abs_coeff + ROUND_POWER_OF_TWO(round_ptr[rc != 0], 1); + const int64_t tmp2 = ((tmp1 * quant_ptr[rc != 0]) >> 16) + tmp1; + const uint32_t abs_qcoeff = + (uint32_t)((tmp2 * quant_shift_ptr[rc != 0]) >> 15); + qcoeff_ptr[rc] = (int)(abs_qcoeff ^ coeff_sign) - coeff_sign; + dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr[rc != 0] / 2; + if (abs_qcoeff) eob = iscan[idx_arr[i]] > eob ? iscan[idx_arr[i]] : eob; + } + } + *eob_ptr = eob + 1; +} +#endif diff --git a/third_party/aom/aom_dsp/x86/highbd_sad4d_sse2.asm b/third_party/aom/aom_dsp/x86/highbd_sad4d_sse2.asm new file mode 100644 index 000000000..0c7cb3998 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/highbd_sad4d_sse2.asm @@ -0,0 +1,290 @@ +; +; Copyright (c) 2016, Alliance for Open Media. All rights reserved +; +; This source code is subject to the terms of the BSD 2 Clause License and +; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License +; was not distributed with this source code in the LICENSE file, you can +; obtain it at www.aomedia.org/license/software. If the Alliance for Open +; Media Patent License 1.0 was not distributed with this source code in the +; PATENTS file, you can obtain it at www.aomedia.org/license/patent. +; + +; + +%include "third_party/x86inc/x86inc.asm" + +SECTION .text + +; HIGH_PROCESS_4x2x4 first, off_{first,second}_{src,ref}, advance_at_end +%macro HIGH_PROCESS_4x2x4 5-6 0 + movh m0, [srcq +%2*2] +%if %1 == 1 + movu m4, [ref1q+%3*2] + movu m5, [ref2q+%3*2] + movu m6, [ref3q+%3*2] + movu m7, [ref4q+%3*2] + movhps m0, [srcq +%4*2] + movhps m4, [ref1q+%5*2] + movhps m5, [ref2q+%5*2] + movhps m6, [ref3q+%5*2] + movhps m7, [ref4q+%5*2] + mova m3, m0 + mova m2, m0 + psubusw m3, m4 + psubusw m2, m5 + psubusw m4, m0 + psubusw m5, m0 + por m4, m3 + por m5, m2 + pmaddwd m4, m1 + pmaddwd m5, m1 + mova m3, m0 + mova m2, m0 + psubusw m3, m6 + psubusw m2, m7 + psubusw m6, m0 + psubusw m7, m0 + por m6, m3 + por m7, m2 + pmaddwd m6, m1 + pmaddwd m7, m1 +%else + movu m2, [ref1q+%3*2] + movhps m0, [srcq +%4*2] + movhps m2, [ref1q+%5*2] + mova m3, m0 + psubusw m3, m2 + psubusw m2, m0 + por m2, m3 + pmaddwd m2, m1 + paddd m4, m2 + + movu m2, [ref2q+%3*2] + mova m3, m0 + movhps m2, [ref2q+%5*2] + psubusw m3, m2 + psubusw m2, m0 + por m2, m3 + pmaddwd m2, m1 + paddd m5, m2 + + movu m2, [ref3q+%3*2] + mova m3, m0 + movhps m2, [ref3q+%5*2] + psubusw m3, m2 + psubusw m2, m0 + por m2, m3 + pmaddwd m2, m1 + paddd m6, m2 + + movu m2, [ref4q+%3*2] + mova m3, m0 + movhps m2, [ref4q+%5*2] + psubusw m3, m2 + psubusw m2, m0 + por m2, m3 + pmaddwd m2, m1 + paddd m7, m2 +%endif +%if %6 == 1 + lea srcq, [srcq +src_strideq*4] + lea ref1q, [ref1q+ref_strideq*4] + lea ref2q, [ref2q+ref_strideq*4] + lea ref3q, [ref3q+ref_strideq*4] + lea ref4q, [ref4q+ref_strideq*4] +%endif +%endmacro + +; PROCESS_8x2x4 first, off_{first,second}_{src,ref}, advance_at_end +%macro HIGH_PROCESS_8x2x4 5-6 0 + ; 1st 8 px + mova m0, [srcq +%2*2] +%if %1 == 1 + movu m4, [ref1q+%3*2] + movu m5, [ref2q+%3*2] + movu m6, [ref3q+%3*2] + movu m7, [ref4q+%3*2] + mova m3, m0 + mova m2, m0 + psubusw m3, m4 + psubusw m2, m5 + psubusw m4, m0 + psubusw m5, m0 + por m4, m3 + por m5, m2 + pmaddwd m4, m1 + pmaddwd m5, m1 + mova m3, m0 + mova m2, m0 + psubusw m3, m6 + psubusw m2, m7 + psubusw m6, m0 + psubusw m7, m0 + por m6, m3 + por m7, m2 + pmaddwd m6, m1 + pmaddwd m7, m1 +%else + mova m3, m0 + movu m2, [ref1q+%3*2] + psubusw m3, m2 + psubusw m2, m0 + por m2, m3 + mova m3, m0 + pmaddwd m2, m1 + paddd m4, m2 + movu m2, [ref2q+%3*2] + psubusw m3, m2 + psubusw m2, m0 + por m2, m3 + mova m3, m0 + pmaddwd m2, m1 + paddd m5, m2 + movu m2, [ref3q+%3*2] + psubusw m3, m2 + psubusw m2, m0 + por m2, m3 + mova m3, m0 + pmaddwd m2, m1 + paddd m6, m2 + movu m2, [ref4q+%3*2] + psubusw m3, m2 + psubusw m2, m0 + por m2, m3 + pmaddwd m2, m1 + paddd m7, m2 +%endif + + ; 2nd 8 px + mova m0, [srcq +(%4)*2] + mova m3, m0 + movu m2, [ref1q+(%5)*2] + psubusw m3, m2 + psubusw m2, m0 + por m2, m3 + mova m3, m0 + pmaddwd m2, m1 + paddd m4, m2 + movu m2, [ref2q+(%5)*2] + psubusw m3, m2 + psubusw m2, m0 + por m2, m3 + mova m3, m0 + pmaddwd m2, m1 + paddd m5, m2 + movu m2, [ref3q+(%5)*2] + psubusw m3, m2 + psubusw m2, m0 + por m2, m3 + mova m3, m0 + pmaddwd m2, m1 + paddd m6, m2 + movu m2, [ref4q+(%5)*2] + psubusw m3, m2 + psubusw m2, m0 +%if %6 == 1 + lea srcq, [srcq +src_strideq*4] + lea ref1q, [ref1q+ref_strideq*4] + lea ref2q, [ref2q+ref_strideq*4] + lea ref3q, [ref3q+ref_strideq*4] + lea ref4q, [ref4q+ref_strideq*4] +%endif + por m2, m3 + pmaddwd m2, m1 + paddd m7, m2 +%endmacro + +; HIGH_PROCESS_16x2x4 first, off_{first,second}_{src,ref}, advance_at_end +%macro HIGH_PROCESS_16x2x4 5-6 0 + HIGH_PROCESS_8x2x4 %1, %2, %3, (%2 + 8), (%3 + 8) + HIGH_PROCESS_8x2x4 0, %4, %5, (%4 + 8), (%5 + 8), %6 +%endmacro + +; HIGH_PROCESS_32x2x4 first, off_{first,second}_{src,ref}, advance_at_end +%macro HIGH_PROCESS_32x2x4 5-6 0 + HIGH_PROCESS_16x2x4 %1, %2, %3, (%2 + 16), (%3 + 16) + HIGH_PROCESS_16x2x4 0, %4, %5, (%4 + 16), (%5 + 16), %6 +%endmacro + +; HIGH_PROCESS_64x2x4 first, off_{first,second}_{src,ref}, advance_at_end +%macro HIGH_PROCESS_64x2x4 5-6 0 + HIGH_PROCESS_32x2x4 %1, %2, %3, (%2 + 32), (%3 + 32) + HIGH_PROCESS_32x2x4 0, %4, %5, (%4 + 32), (%5 + 32), %6 +%endmacro + +; void aom_highbd_sadNxNx4d_sse2(uint8_t *src, int src_stride, +; uint8_t *ref[4], int ref_stride, +; uint32_t res[4]); +; where NxN = 64x64, 32x32, 16x16, 16x8, 8x16 or 8x8 +%macro HIGH_SADNXN4D 2 +%if UNIX64 +cglobal highbd_sad%1x%2x4d, 5, 8, 8, src, src_stride, ref1, ref_stride, \ + res, ref2, ref3, ref4 +%else +cglobal highbd_sad%1x%2x4d, 4, 7, 8, src, src_stride, ref1, ref_stride, \ + ref2, ref3, ref4 +%endif + +; set m1 + push srcq + mov srcd, 0x00010001 + movd m1, srcd + pshufd m1, m1, 0x0 + pop srcq + + movsxdifnidn src_strideq, src_strided + movsxdifnidn ref_strideq, ref_strided + mov ref2q, [ref1q+gprsize*1] + mov ref3q, [ref1q+gprsize*2] + mov ref4q, [ref1q+gprsize*3] + mov ref1q, [ref1q+gprsize*0] + +; convert byte pointers to short pointers + shl srcq, 1 + shl ref2q, 1 + shl ref3q, 1 + shl ref4q, 1 + shl ref1q, 1 + + HIGH_PROCESS_%1x2x4 1, 0, 0, src_strideq, ref_strideq, 1 +%rep (%2-4)/2 + HIGH_PROCESS_%1x2x4 0, 0, 0, src_strideq, ref_strideq, 1 +%endrep + HIGH_PROCESS_%1x2x4 0, 0, 0, src_strideq, ref_strideq, 0 + ; N.B. HIGH_PROCESS outputs dwords (32 bits) + ; so in high bit depth even the smallest width (4) needs 128bits i.e. XMM + movhlps m0, m4 + movhlps m1, m5 + movhlps m2, m6 + movhlps m3, m7 + paddd m4, m0 + paddd m5, m1 + paddd m6, m2 + paddd m7, m3 + punpckldq m4, m5 + punpckldq m6, m7 + movhlps m0, m4 + movhlps m1, m6 + paddd m4, m0 + paddd m6, m1 + punpcklqdq m4, m6 + movifnidn r4, r4mp + movu [r4], m4 + RET +%endmacro + + +INIT_XMM sse2 +HIGH_SADNXN4D 64, 64 +HIGH_SADNXN4D 64, 32 +HIGH_SADNXN4D 32, 64 +HIGH_SADNXN4D 32, 32 +HIGH_SADNXN4D 32, 16 +HIGH_SADNXN4D 16, 32 +HIGH_SADNXN4D 16, 16 +HIGH_SADNXN4D 16, 8 +HIGH_SADNXN4D 8, 16 +HIGH_SADNXN4D 8, 8 +HIGH_SADNXN4D 8, 4 +HIGH_SADNXN4D 4, 8 +HIGH_SADNXN4D 4, 4 diff --git a/third_party/aom/aom_dsp/x86/highbd_sad_sse2.asm b/third_party/aom/aom_dsp/x86/highbd_sad_sse2.asm new file mode 100644 index 000000000..8427b891c --- /dev/null +++ b/third_party/aom/aom_dsp/x86/highbd_sad_sse2.asm @@ -0,0 +1,366 @@ +; +; Copyright (c) 2016, Alliance for Open Media. All rights reserved +; +; This source code is subject to the terms of the BSD 2 Clause License and +; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License +; was not distributed with this source code in the LICENSE file, you can +; obtain it at www.aomedia.org/license/software. If the Alliance for Open +; Media Patent License 1.0 was not distributed with this source code in the +; PATENTS file, you can obtain it at www.aomedia.org/license/patent. +; + +; + +%include "third_party/x86inc/x86inc.asm" + +SECTION .text + +%macro HIGH_SAD_FN 4 +%if %4 == 0 +%if %3 == 5 +cglobal highbd_sad%1x%2, 4, %3, 7, src, src_stride, ref, ref_stride, n_rows +%else ; %3 == 7 +cglobal highbd_sad%1x%2, 4, %3, 7, src, src_stride, ref, ref_stride, \ + src_stride3, ref_stride3, n_rows +%endif ; %3 == 5/7 +%else ; avg +%if %3 == 5 +cglobal highbd_sad%1x%2_avg, 5, 1 + %3, 7, src, src_stride, ref, ref_stride, \ + second_pred, n_rows +%else ; %3 == 7 +cglobal highbd_sad%1x%2_avg, 5, ARCH_X86_64 + %3, 7, src, src_stride, \ + ref, ref_stride, \ + second_pred, \ + src_stride3, ref_stride3 +%if ARCH_X86_64 +%define n_rowsd r7d +%else ; x86-32 +%define n_rowsd dword r0m +%endif ; x86-32/64 +%endif ; %3 == 5/7 +%endif ; avg/sad + movsxdifnidn src_strideq, src_strided + movsxdifnidn ref_strideq, ref_strided +%if %3 == 7 + lea src_stride3q, [src_strideq*3] + lea ref_stride3q, [ref_strideq*3] +%endif ; %3 == 7 +; convert src, ref & second_pred to short ptrs (from byte ptrs) + shl srcq, 1 + shl refq, 1 +%if %4 == 1 + shl second_predq, 1 +%endif +%endmacro + +; unsigned int aom_highbd_sad64x{16,32,64}_sse2(uint8_t *src, int src_stride, +; uint8_t *ref, int ref_stride); +%macro HIGH_SAD64XN 1-2 0 + HIGH_SAD_FN 64, %1, 5, %2 + mov n_rowsd, %1 + pxor m0, m0 + pxor m6, m6 + +.loop: + ; first half of each row + movu m1, [refq] + movu m2, [refq+16] + movu m3, [refq+32] + movu m4, [refq+48] +%if %2 == 1 + pavgw m1, [second_predq+mmsize*0] + pavgw m2, [second_predq+mmsize*1] + pavgw m3, [second_predq+mmsize*2] + pavgw m4, [second_predq+mmsize*3] + lea second_predq, [second_predq+mmsize*4] +%endif + mova m5, [srcq] + psubusw m5, m1 + psubusw m1, [srcq] + por m1, m5 + mova m5, [srcq+16] + psubusw m5, m2 + psubusw m2, [srcq+16] + por m2, m5 + mova m5, [srcq+32] + psubusw m5, m3 + psubusw m3, [srcq+32] + por m3, m5 + mova m5, [srcq+48] + psubusw m5, m4 + psubusw m4, [srcq+48] + por m4, m5 + paddw m1, m2 + paddw m3, m4 + movhlps m2, m1 + movhlps m4, m3 + paddw m1, m2 + paddw m3, m4 + punpcklwd m1, m6 + punpcklwd m3, m6 + paddd m0, m1 + paddd m0, m3 + ; second half of each row + movu m1, [refq+64] + movu m2, [refq+80] + movu m3, [refq+96] + movu m4, [refq+112] +%if %2 == 1 + pavgw m1, [second_predq+mmsize*0] + pavgw m2, [second_predq+mmsize*1] + pavgw m3, [second_predq+mmsize*2] + pavgw m4, [second_predq+mmsize*3] + lea second_predq, [second_predq+mmsize*4] +%endif + mova m5, [srcq+64] + psubusw m5, m1 + psubusw m1, [srcq+64] + por m1, m5 + mova m5, [srcq+80] + psubusw m5, m2 + psubusw m2, [srcq+80] + por m2, m5 + mova m5, [srcq+96] + psubusw m5, m3 + psubusw m3, [srcq+96] + por m3, m5 + mova m5, [srcq+112] + psubusw m5, m4 + psubusw m4, [srcq+112] + por m4, m5 + paddw m1, m2 + paddw m3, m4 + movhlps m2, m1 + movhlps m4, m3 + paddw m1, m2 + paddw m3, m4 + punpcklwd m1, m6 + punpcklwd m3, m6 + lea refq, [refq+ref_strideq*2] + paddd m0, m1 + lea srcq, [srcq+src_strideq*2] + paddd m0, m3 + + dec n_rowsd + jg .loop + + movhlps m1, m0 + paddd m0, m1 + punpckldq m0, m6 + movhlps m1, m0 + paddd m0, m1 + movd eax, m0 + RET +%endmacro + +INIT_XMM sse2 +HIGH_SAD64XN 64 ; highbd_sad64x64_sse2 +HIGH_SAD64XN 32 ; highbd_sad64x32_sse2 +HIGH_SAD64XN 64, 1 ; highbd_sad64x64_avg_sse2 +HIGH_SAD64XN 32, 1 ; highbd_sad64x32_avg_sse2 + + +; unsigned int aom_highbd_sad32x{16,32,64}_sse2(uint8_t *src, int src_stride, +; uint8_t *ref, int ref_stride); +%macro HIGH_SAD32XN 1-2 0 + HIGH_SAD_FN 32, %1, 5, %2 + mov n_rowsd, %1 + pxor m0, m0 + pxor m6, m6 + +.loop: + movu m1, [refq] + movu m2, [refq+16] + movu m3, [refq+32] + movu m4, [refq+48] +%if %2 == 1 + pavgw m1, [second_predq+mmsize*0] + pavgw m2, [second_predq+mmsize*1] + pavgw m3, [second_predq+mmsize*2] + pavgw m4, [second_predq+mmsize*3] + lea second_predq, [second_predq+mmsize*4] +%endif + mova m5, [srcq] + psubusw m5, m1 + psubusw m1, [srcq] + por m1, m5 + mova m5, [srcq+16] + psubusw m5, m2 + psubusw m2, [srcq+16] + por m2, m5 + mova m5, [srcq+32] + psubusw m5, m3 + psubusw m3, [srcq+32] + por m3, m5 + mova m5, [srcq+48] + psubusw m5, m4 + psubusw m4, [srcq+48] + por m4, m5 + paddw m1, m2 + paddw m3, m4 + movhlps m2, m1 + movhlps m4, m3 + paddw m1, m2 + paddw m3, m4 + punpcklwd m1, m6 + punpcklwd m3, m6 + lea refq, [refq+ref_strideq*2] + paddd m0, m1 + lea srcq, [srcq+src_strideq*2] + paddd m0, m3 + dec n_rowsd + jg .loop + + movhlps m1, m0 + paddd m0, m1 + punpckldq m0, m6 + movhlps m1, m0 + paddd m0, m1 + movd eax, m0 + RET +%endmacro + +INIT_XMM sse2 +HIGH_SAD32XN 64 ; highbd_sad32x64_sse2 +HIGH_SAD32XN 32 ; highbd_sad32x32_sse2 +HIGH_SAD32XN 16 ; highbd_sad32x16_sse2 +HIGH_SAD32XN 64, 1 ; highbd_sad32x64_avg_sse2 +HIGH_SAD32XN 32, 1 ; highbd_sad32x32_avg_sse2 +HIGH_SAD32XN 16, 1 ; highbd_sad32x16_avg_sse2 + +; unsigned int aom_highbd_sad16x{8,16,32}_sse2(uint8_t *src, int src_stride, +; uint8_t *ref, int ref_stride); +%macro HIGH_SAD16XN 1-2 0 + HIGH_SAD_FN 16, %1, 5, %2 + mov n_rowsd, %1/2 + pxor m0, m0 + pxor m6, m6 + +.loop: + movu m1, [refq] + movu m2, [refq+16] + movu m3, [refq+ref_strideq*2] + movu m4, [refq+ref_strideq*2+16] +%if %2 == 1 + pavgw m1, [second_predq+mmsize*0] + pavgw m2, [second_predq+16] + pavgw m3, [second_predq+mmsize*2] + pavgw m4, [second_predq+mmsize*2+16] + lea second_predq, [second_predq+mmsize*4] +%endif + mova m5, [srcq] + psubusw m5, m1 + psubusw m1, [srcq] + por m1, m5 + mova m5, [srcq+16] + psubusw m5, m2 + psubusw m2, [srcq+16] + por m2, m5 + mova m5, [srcq+src_strideq*2] + psubusw m5, m3 + psubusw m3, [srcq+src_strideq*2] + por m3, m5 + mova m5, [srcq+src_strideq*2+16] + psubusw m5, m4 + psubusw m4, [srcq+src_strideq*2+16] + por m4, m5 + paddw m1, m2 + paddw m3, m4 + movhlps m2, m1 + movhlps m4, m3 + paddw m1, m2 + paddw m3, m4 + punpcklwd m1, m6 + punpcklwd m3, m6 + lea refq, [refq+ref_strideq*4] + paddd m0, m1 + lea srcq, [srcq+src_strideq*4] + paddd m0, m3 + dec n_rowsd + jg .loop + + movhlps m1, m0 + paddd m0, m1 + punpckldq m0, m6 + movhlps m1, m0 + paddd m0, m1 + movd eax, m0 + RET +%endmacro + +INIT_XMM sse2 +HIGH_SAD16XN 32 ; highbd_sad16x32_sse2 +HIGH_SAD16XN 16 ; highbd_sad16x16_sse2 +HIGH_SAD16XN 8 ; highbd_sad16x8_sse2 +HIGH_SAD16XN 32, 1 ; highbd_sad16x32_avg_sse2 +HIGH_SAD16XN 16, 1 ; highbd_sad16x16_avg_sse2 +HIGH_SAD16XN 8, 1 ; highbd_sad16x8_avg_sse2 + + +; unsigned int aom_highbd_sad8x{4,8,16}_sse2(uint8_t *src, int src_stride, +; uint8_t *ref, int ref_stride); +%macro HIGH_SAD8XN 1-2 0 + HIGH_SAD_FN 8, %1, 7, %2 + mov n_rowsd, %1/4 + pxor m0, m0 + pxor m6, m6 + +.loop: + movu m1, [refq] + movu m2, [refq+ref_strideq*2] + movu m3, [refq+ref_strideq*4] + movu m4, [refq+ref_stride3q*2] +%if %2 == 1 + pavgw m1, [second_predq+mmsize*0] + pavgw m2, [second_predq+mmsize*1] + pavgw m3, [second_predq+mmsize*2] + pavgw m4, [second_predq+mmsize*3] + lea second_predq, [second_predq+mmsize*4] +%endif + mova m5, [srcq] + psubusw m5, m1 + psubusw m1, [srcq] + por m1, m5 + mova m5, [srcq+src_strideq*2] + psubusw m5, m2 + psubusw m2, [srcq+src_strideq*2] + por m2, m5 + mova m5, [srcq+src_strideq*4] + psubusw m5, m3 + psubusw m3, [srcq+src_strideq*4] + por m3, m5 + mova m5, [srcq+src_stride3q*2] + psubusw m5, m4 + psubusw m4, [srcq+src_stride3q*2] + por m4, m5 + paddw m1, m2 + paddw m3, m4 + movhlps m2, m1 + movhlps m4, m3 + paddw m1, m2 + paddw m3, m4 + punpcklwd m1, m6 + punpcklwd m3, m6 + lea refq, [refq+ref_strideq*8] + paddd m0, m1 + lea srcq, [srcq+src_strideq*8] + paddd m0, m3 + dec n_rowsd + jg .loop + + movhlps m1, m0 + paddd m0, m1 + punpckldq m0, m6 + movhlps m1, m0 + paddd m0, m1 + movd eax, m0 + RET +%endmacro + +INIT_XMM sse2 +HIGH_SAD8XN 16 ; highbd_sad8x16_sse2 +HIGH_SAD8XN 8 ; highbd_sad8x8_sse2 +HIGH_SAD8XN 4 ; highbd_sad8x4_sse2 +HIGH_SAD8XN 16, 1 ; highbd_sad8x16_avg_sse2 +HIGH_SAD8XN 8, 1 ; highbd_sad8x8_avg_sse2 +HIGH_SAD8XN 4, 1 ; highbd_sad8x4_avg_sse2 diff --git a/third_party/aom/aom_dsp/x86/highbd_subpel_variance_impl_sse2.asm b/third_party/aom/aom_dsp/x86/highbd_subpel_variance_impl_sse2.asm new file mode 100644 index 000000000..797e9c1d4 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/highbd_subpel_variance_impl_sse2.asm @@ -0,0 +1,1040 @@ +; +; Copyright (c) 2016, Alliance for Open Media. All rights reserved +; +; This source code is subject to the terms of the BSD 2 Clause License and +; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License +; was not distributed with this source code in the LICENSE file, you can +; obtain it at www.aomedia.org/license/software. If the Alliance for Open +; Media Patent License 1.0 was not distributed with this source code in the +; PATENTS file, you can obtain it at www.aomedia.org/license/patent. +; + +; + +%include "third_party/x86inc/x86inc.asm" + +SECTION_RODATA +pw_8: times 8 dw 8 +bilin_filter_m_sse2: times 8 dw 16 + times 8 dw 0 + times 8 dw 14 + times 8 dw 2 + times 8 dw 12 + times 8 dw 4 + times 8 dw 10 + times 8 dw 6 + times 16 dw 8 + times 8 dw 6 + times 8 dw 10 + times 8 dw 4 + times 8 dw 12 + times 8 dw 2 + times 8 dw 14 + +SECTION .text + +; int aom_sub_pixel_varianceNxh(const uint8_t *src, ptrdiff_t src_stride, +; int x_offset, int y_offset, +; const uint8_t *dst, ptrdiff_t dst_stride, +; int height, unsigned int *sse); +; +; This function returns the SE and stores SSE in the given pointer. + +%macro SUM_SSE 6 ; src1, dst1, src2, dst2, sum, sse + psubw %3, %4 + psubw %1, %2 + mova %4, %3 ; make copies to manipulate to calc sum + mova %2, %1 ; use originals for calc sse + pmaddwd %3, %3 + paddw %4, %2 + pmaddwd %1, %1 + movhlps %2, %4 + paddd %6, %3 + paddw %4, %2 + pxor %2, %2 + pcmpgtw %2, %4 ; mask for 0 > %4 (sum) + punpcklwd %4, %2 ; sign-extend word to dword + paddd %6, %1 + paddd %5, %4 + +%endmacro + +%macro STORE_AND_RET 0 +%if mmsize == 16 + ; if H=64 and W=16, we have 8 words of each 2(1bit)x64(6bit)x9bit=16bit + ; in m6, i.e. it _exactly_ fits in a signed word per word in the xmm reg. + ; We have to sign-extend it before adding the words within the register + ; and outputing to a dword. + movhlps m3, m7 + movhlps m4, m6 + paddd m7, m3 + paddd m6, m4 + pshufd m3, m7, 0x1 + pshufd m4, m6, 0x1 + paddd m7, m3 + paddd m6, m4 + mov r1, ssem ; r1 = unsigned int *sse + movd [r1], m7 ; store sse + movd rax, m6 ; store sum as return value +%endif + RET +%endmacro + +%macro INC_SRC_BY_SRC_STRIDE 0 +%if ARCH_X86=1 && CONFIG_PIC=1 + add srcq, src_stridemp + add srcq, src_stridemp +%else + lea srcq, [srcq + src_strideq*2] +%endif +%endmacro + +%macro SUBPEL_VARIANCE 1-2 0 ; W +%define bilin_filter_m bilin_filter_m_sse2 +%define filter_idx_shift 5 + + +%ifdef PIC ; 64bit PIC + %if %2 == 1 ; avg + cglobal highbd_sub_pixel_avg_variance%1xh, 9, 10, 13, src, src_stride, \ + x_offset, y_offset, \ + dst, dst_stride, \ + sec, sec_stride, height, sse + %define sec_str sec_strideq + %else + cglobal highbd_sub_pixel_variance%1xh, 7, 8, 13, src, src_stride, x_offset, \ + y_offset, dst, dst_stride, height, sse + %endif + %define block_height heightd + %define bilin_filter sseq +%else + %if ARCH_X86=1 && CONFIG_PIC=1 + %if %2 == 1 ; avg + cglobal highbd_sub_pixel_avg_variance%1xh, 7, 7, 13, src, src_stride, \ + x_offset, y_offset, \ + dst, dst_stride, \ + sec, sec_stride, \ + height, sse, g_bilin_filter, g_pw_8 + %define block_height dword heightm + %define sec_str sec_stridemp + + ; Store bilin_filter and pw_8 location in stack + %if GET_GOT_DEFINED == 1 + GET_GOT eax + add esp, 4 ; restore esp + %endif + + lea ecx, [GLOBAL(bilin_filter_m)] + mov g_bilin_filterm, ecx + + lea ecx, [GLOBAL(pw_8)] + mov g_pw_8m, ecx + + LOAD_IF_USED 0, 1 ; load eax, ecx back + %else + cglobal highbd_sub_pixel_variance%1xh, 7, 7, 13, src, src_stride, \ + x_offset, y_offset, dst, dst_stride, height, \ + sse, g_bilin_filter, g_pw_8 + %define block_height heightd + + ; Store bilin_filter and pw_8 location in stack + %if GET_GOT_DEFINED == 1 + GET_GOT eax + add esp, 4 ; restore esp + %endif + + lea ecx, [GLOBAL(bilin_filter_m)] + mov g_bilin_filterm, ecx + + lea ecx, [GLOBAL(pw_8)] + mov g_pw_8m, ecx + + LOAD_IF_USED 0, 1 ; load eax, ecx back + %endif + %else + %if %2 == 1 ; avg + cglobal highbd_sub_pixel_avg_variance%1xh, 7 + 2 * ARCH_X86_64, \ + 7 + 2 * ARCH_X86_64, 13, src, src_stride, \ + x_offset, y_offset, \ + dst, dst_stride, \ + sec, sec_stride, \ + height, sse + %if ARCH_X86_64 + %define block_height heightd + %define sec_str sec_strideq + %else + %define block_height dword heightm + %define sec_str sec_stridemp + %endif + %else + cglobal highbd_sub_pixel_variance%1xh, 7, 7, 13, src, src_stride, \ + x_offset, y_offset, dst, dst_stride, height, sse + %define block_height heightd + %endif + + %define bilin_filter bilin_filter_m + %endif +%endif + + ASSERT %1 <= 16 ; m6 overflows if w > 16 + pxor m6, m6 ; sum + pxor m7, m7 ; sse + +%if %1 < 16 + sar block_height, 1 +%endif +%if %2 == 1 ; avg + shl sec_str, 1 +%endif + + ; FIXME(rbultje) replace by jumptable? + test x_offsetd, x_offsetd + jnz .x_nonzero + ; x_offset == 0 + test y_offsetd, y_offsetd + jnz .x_zero_y_nonzero + + ; x_offset == 0 && y_offset == 0 +.x_zero_y_zero_loop: +%if %1 == 16 + movu m0, [srcq] + movu m2, [srcq + 16] + mova m1, [dstq] + mova m3, [dstq + 16] +%if %2 == 1 ; avg + pavgw m0, [secq] + pavgw m2, [secq+16] +%endif + SUM_SSE m0, m1, m2, m3, m6, m7 + + lea srcq, [srcq + src_strideq*2] + lea dstq, [dstq + dst_strideq*2] +%if %2 == 1 ; avg + add secq, sec_str +%endif +%else ; %1 < 16 + movu m0, [srcq] + movu m2, [srcq + src_strideq*2] + mova m1, [dstq] + mova m3, [dstq + dst_strideq*2] +%if %2 == 1 ; avg + pavgw m0, [secq] + add secq, sec_str + pavgw m2, [secq] +%endif + SUM_SSE m0, m1, m2, m3, m6, m7 + + lea srcq, [srcq + src_strideq*4] + lea dstq, [dstq + dst_strideq*4] +%if %2 == 1 ; avg + add secq, sec_str +%endif +%endif + dec block_height + jg .x_zero_y_zero_loop + STORE_AND_RET + +.x_zero_y_nonzero: + cmp y_offsetd, 8 + jne .x_zero_y_nonhalf + + ; x_offset == 0 && y_offset == 0.5 +.x_zero_y_half_loop: +%if %1 == 16 + movu m0, [srcq] + movu m1, [srcq+16] + movu m4, [srcq+src_strideq*2] + movu m5, [srcq+src_strideq*2+16] + mova m2, [dstq] + mova m3, [dstq+16] + pavgw m0, m4 + pavgw m1, m5 +%if %2 == 1 ; avg + pavgw m0, [secq] + pavgw m1, [secq+16] +%endif + SUM_SSE m0, m2, m1, m3, m6, m7 + + lea srcq, [srcq + src_strideq*2] + lea dstq, [dstq + dst_strideq*2] +%if %2 == 1 ; avg + add secq, sec_str +%endif +%else ; %1 < 16 + movu m0, [srcq] + movu m1, [srcq+src_strideq*2] + movu m5, [srcq+src_strideq*4] + mova m2, [dstq] + mova m3, [dstq+dst_strideq*2] + pavgw m0, m1 + pavgw m1, m5 +%if %2 == 1 ; avg + pavgw m0, [secq] + add secq, sec_str + pavgw m1, [secq] +%endif + SUM_SSE m0, m2, m1, m3, m6, m7 + + lea srcq, [srcq + src_strideq*4] + lea dstq, [dstq + dst_strideq*4] +%if %2 == 1 ; avg + add secq, sec_str +%endif +%endif + dec block_height + jg .x_zero_y_half_loop + STORE_AND_RET + +.x_zero_y_nonhalf: + ; x_offset == 0 && y_offset == bilin interpolation +%ifdef PIC + lea bilin_filter, [bilin_filter_m] +%endif + shl y_offsetd, filter_idx_shift +%if ARCH_X86_64 && mmsize == 16 + mova m8, [bilin_filter+y_offsetq] + mova m9, [bilin_filter+y_offsetq+16] + mova m10, [pw_8] +%define filter_y_a m8 +%define filter_y_b m9 +%define filter_rnd m10 +%else ; x86-32 or mmx +%if ARCH_X86=1 && CONFIG_PIC=1 +; x_offset == 0, reuse x_offset reg +%define tempq x_offsetq + add y_offsetq, g_bilin_filterm +%define filter_y_a [y_offsetq] +%define filter_y_b [y_offsetq+16] + mov tempq, g_pw_8m +%define filter_rnd [tempq] +%else + add y_offsetq, bilin_filter +%define filter_y_a [y_offsetq] +%define filter_y_b [y_offsetq+16] +%define filter_rnd [pw_8] +%endif +%endif + +.x_zero_y_other_loop: +%if %1 == 16 + movu m0, [srcq] + movu m1, [srcq + 16] + movu m4, [srcq+src_strideq*2] + movu m5, [srcq+src_strideq*2+16] + mova m2, [dstq] + mova m3, [dstq+16] + ; FIXME(rbultje) instead of out=((num-x)*in1+x*in2+rnd)>>log2(num), we can + ; also do out=in1+(((num-x)*(in2-in1)+rnd)>>log2(num)). Total number of + ; instructions is the same (5), but it is 1 mul instead of 2, so might be + ; slightly faster because of pmullw latency. It would also cut our rodata + ; tables in half for this function, and save 1-2 registers on x86-64. + pmullw m1, filter_y_a + pmullw m5, filter_y_b + paddw m1, filter_rnd + pmullw m0, filter_y_a + pmullw m4, filter_y_b + paddw m0, filter_rnd + paddw m1, m5 + paddw m0, m4 + psrlw m1, 4 + psrlw m0, 4 +%if %2 == 1 ; avg + pavgw m0, [secq] + pavgw m1, [secq+16] +%endif + SUM_SSE m0, m2, m1, m3, m6, m7 + + lea srcq, [srcq + src_strideq*2] + lea dstq, [dstq + dst_strideq*2] +%if %2 == 1 ; avg + add secq, sec_str +%endif +%else ; %1 < 16 + movu m0, [srcq] + movu m1, [srcq+src_strideq*2] + movu m5, [srcq+src_strideq*4] + mova m4, m1 + mova m2, [dstq] + mova m3, [dstq+dst_strideq*2] + pmullw m1, filter_y_a + pmullw m5, filter_y_b + paddw m1, filter_rnd + pmullw m0, filter_y_a + pmullw m4, filter_y_b + paddw m0, filter_rnd + paddw m1, m5 + paddw m0, m4 + psrlw m1, 4 + psrlw m0, 4 +%if %2 == 1 ; avg + pavgw m0, [secq] + add secq, sec_str + pavgw m1, [secq] +%endif + SUM_SSE m0, m2, m1, m3, m6, m7 + + lea srcq, [srcq + src_strideq*4] + lea dstq, [dstq + dst_strideq*4] +%if %2 == 1 ; avg + add secq, sec_str +%endif +%endif + dec block_height + jg .x_zero_y_other_loop +%undef filter_y_a +%undef filter_y_b +%undef filter_rnd + STORE_AND_RET + +.x_nonzero: + cmp x_offsetd, 8 + jne .x_nonhalf + ; x_offset == 0.5 + test y_offsetd, y_offsetd + jnz .x_half_y_nonzero + + ; x_offset == 0.5 && y_offset == 0 +.x_half_y_zero_loop: +%if %1 == 16 + movu m0, [srcq] + movu m1, [srcq + 16] + movu m4, [srcq + 2] + movu m5, [srcq + 18] + mova m2, [dstq] + mova m3, [dstq + 16] + pavgw m0, m4 + pavgw m1, m5 +%if %2 == 1 ; avg + pavgw m0, [secq] + pavgw m1, [secq+16] +%endif + SUM_SSE m0, m2, m1, m3, m6, m7 + + lea srcq, [srcq + src_strideq*2] + lea dstq, [dstq + dst_strideq*2] +%if %2 == 1 ; avg + add secq, sec_str +%endif +%else ; %1 < 16 + movu m0, [srcq] + movu m1, [srcq + src_strideq*2] + movu m4, [srcq + 2] + movu m5, [srcq + src_strideq*2 + 2] + mova m2, [dstq] + mova m3, [dstq + dst_strideq*2] + pavgw m0, m4 + pavgw m1, m5 +%if %2 == 1 ; avg + pavgw m0, [secq] + add secq, sec_str + pavgw m1, [secq] +%endif + SUM_SSE m0, m2, m1, m3, m6, m7 + + lea srcq, [srcq + src_strideq*4] + lea dstq, [dstq + dst_strideq*4] +%if %2 == 1 ; avg + add secq, sec_str +%endif +%endif + dec block_height + jg .x_half_y_zero_loop + STORE_AND_RET + +.x_half_y_nonzero: + cmp y_offsetd, 8 + jne .x_half_y_nonhalf + + ; x_offset == 0.5 && y_offset == 0.5 +%if %1 == 16 + movu m0, [srcq] + movu m1, [srcq+16] + movu m2, [srcq+2] + movu m3, [srcq+18] + lea srcq, [srcq + src_strideq*2] + pavgw m0, m2 + pavgw m1, m3 +.x_half_y_half_loop: + movu m2, [srcq] + movu m3, [srcq + 16] + movu m4, [srcq + 2] + movu m5, [srcq + 18] + pavgw m2, m4 + pavgw m3, m5 + pavgw m0, m2 + pavgw m1, m3 + mova m4, [dstq] + mova m5, [dstq + 16] +%if %2 == 1 ; avg + pavgw m0, [secq] + pavgw m1, [secq+16] +%endif + SUM_SSE m0, m4, m1, m5, m6, m7 + mova m0, m2 + mova m1, m3 + + lea srcq, [srcq + src_strideq*2] + lea dstq, [dstq + dst_strideq*2] +%if %2 == 1 ; avg + add secq, sec_str +%endif +%else ; %1 < 16 + movu m0, [srcq] + movu m2, [srcq+2] + lea srcq, [srcq + src_strideq*2] + pavgw m0, m2 +.x_half_y_half_loop: + movu m2, [srcq] + movu m3, [srcq + src_strideq*2] + movu m4, [srcq + 2] + movu m5, [srcq + src_strideq*2 + 2] + pavgw m2, m4 + pavgw m3, m5 + pavgw m0, m2 + pavgw m2, m3 + mova m4, [dstq] + mova m5, [dstq + dst_strideq*2] +%if %2 == 1 ; avg + pavgw m0, [secq] + add secq, sec_str + pavgw m2, [secq] +%endif + SUM_SSE m0, m4, m2, m5, m6, m7 + mova m0, m3 + + lea srcq, [srcq + src_strideq*4] + lea dstq, [dstq + dst_strideq*4] +%if %2 == 1 ; avg + add secq, sec_str +%endif +%endif + dec block_height + jg .x_half_y_half_loop + STORE_AND_RET + +.x_half_y_nonhalf: + ; x_offset == 0.5 && y_offset == bilin interpolation +%ifdef PIC + lea bilin_filter, [bilin_filter_m] +%endif + shl y_offsetd, filter_idx_shift +%if ARCH_X86_64 && mmsize == 16 + mova m8, [bilin_filter+y_offsetq] + mova m9, [bilin_filter+y_offsetq+16] + mova m10, [pw_8] +%define filter_y_a m8 +%define filter_y_b m9 +%define filter_rnd m10 +%else ; x86_32 +%if ARCH_X86=1 && CONFIG_PIC=1 +; x_offset == 0.5. We can reuse x_offset reg +%define tempq x_offsetq + add y_offsetq, g_bilin_filterm +%define filter_y_a [y_offsetq] +%define filter_y_b [y_offsetq+16] + mov tempq, g_pw_8m +%define filter_rnd [tempq] +%else + add y_offsetq, bilin_filter +%define filter_y_a [y_offsetq] +%define filter_y_b [y_offsetq+16] +%define filter_rnd [pw_8] +%endif +%endif + +%if %1 == 16 + movu m0, [srcq] + movu m1, [srcq+16] + movu m2, [srcq+2] + movu m3, [srcq+18] + lea srcq, [srcq + src_strideq*2] + pavgw m0, m2 + pavgw m1, m3 +.x_half_y_other_loop: + movu m2, [srcq] + movu m3, [srcq+16] + movu m4, [srcq+2] + movu m5, [srcq+18] + pavgw m2, m4 + pavgw m3, m5 + mova m4, m2 + mova m5, m3 + pmullw m1, filter_y_a + pmullw m3, filter_y_b + paddw m1, filter_rnd + paddw m1, m3 + pmullw m0, filter_y_a + pmullw m2, filter_y_b + paddw m0, filter_rnd + psrlw m1, 4 + paddw m0, m2 + mova m2, [dstq] + psrlw m0, 4 + mova m3, [dstq+16] +%if %2 == 1 ; avg + pavgw m0, [secq] + pavgw m1, [secq+16] +%endif + SUM_SSE m0, m2, m1, m3, m6, m7 + mova m0, m4 + mova m1, m5 + + lea srcq, [srcq + src_strideq*2] + lea dstq, [dstq + dst_strideq*2] +%if %2 == 1 ; avg + add secq, sec_str +%endif +%else ; %1 < 16 + movu m0, [srcq] + movu m2, [srcq+2] + lea srcq, [srcq + src_strideq*2] + pavgw m0, m2 +.x_half_y_other_loop: + movu m2, [srcq] + movu m3, [srcq+src_strideq*2] + movu m4, [srcq+2] + movu m5, [srcq+src_strideq*2+2] + pavgw m2, m4 + pavgw m3, m5 + mova m4, m2 + mova m5, m3 + pmullw m4, filter_y_a + pmullw m3, filter_y_b + paddw m4, filter_rnd + paddw m4, m3 + pmullw m0, filter_y_a + pmullw m2, filter_y_b + paddw m0, filter_rnd + psrlw m4, 4 + paddw m0, m2 + mova m2, [dstq] + psrlw m0, 4 + mova m3, [dstq+dst_strideq*2] +%if %2 == 1 ; avg + pavgw m0, [secq] + add secq, sec_str + pavgw m4, [secq] +%endif + SUM_SSE m0, m2, m4, m3, m6, m7 + mova m0, m5 + + lea srcq, [srcq + src_strideq*4] + lea dstq, [dstq + dst_strideq*4] +%if %2 == 1 ; avg + add secq, sec_str +%endif +%endif + dec block_height + jg .x_half_y_other_loop +%undef filter_y_a +%undef filter_y_b +%undef filter_rnd + STORE_AND_RET + +.x_nonhalf: + test y_offsetd, y_offsetd + jnz .x_nonhalf_y_nonzero + + ; x_offset == bilin interpolation && y_offset == 0 +%ifdef PIC + lea bilin_filter, [bilin_filter_m] +%endif + shl x_offsetd, filter_idx_shift +%if ARCH_X86_64 && mmsize == 16 + mova m8, [bilin_filter+x_offsetq] + mova m9, [bilin_filter+x_offsetq+16] + mova m10, [pw_8] +%define filter_x_a m8 +%define filter_x_b m9 +%define filter_rnd m10 +%else ; x86-32 +%if ARCH_X86=1 && CONFIG_PIC=1 +; y_offset == 0. We can reuse y_offset reg. +%define tempq y_offsetq + add x_offsetq, g_bilin_filterm +%define filter_x_a [x_offsetq] +%define filter_x_b [x_offsetq+16] + mov tempq, g_pw_8m +%define filter_rnd [tempq] +%else + add x_offsetq, bilin_filter +%define filter_x_a [x_offsetq] +%define filter_x_b [x_offsetq+16] +%define filter_rnd [pw_8] +%endif +%endif + +.x_other_y_zero_loop: +%if %1 == 16 + movu m0, [srcq] + movu m1, [srcq+16] + movu m2, [srcq+2] + movu m3, [srcq+18] + mova m4, [dstq] + mova m5, [dstq+16] + pmullw m1, filter_x_a + pmullw m3, filter_x_b + paddw m1, filter_rnd + pmullw m0, filter_x_a + pmullw m2, filter_x_b + paddw m0, filter_rnd + paddw m1, m3 + paddw m0, m2 + psrlw m1, 4 + psrlw m0, 4 +%if %2 == 1 ; avg + pavgw m0, [secq] + pavgw m1, [secq+16] +%endif + SUM_SSE m0, m4, m1, m5, m6, m7 + + lea srcq, [srcq+src_strideq*2] + lea dstq, [dstq+dst_strideq*2] +%if %2 == 1 ; avg + add secq, sec_str +%endif +%else ; %1 < 16 + movu m0, [srcq] + movu m1, [srcq+src_strideq*2] + movu m2, [srcq+2] + movu m3, [srcq+src_strideq*2+2] + mova m4, [dstq] + mova m5, [dstq+dst_strideq*2] + pmullw m1, filter_x_a + pmullw m3, filter_x_b + paddw m1, filter_rnd + pmullw m0, filter_x_a + pmullw m2, filter_x_b + paddw m0, filter_rnd + paddw m1, m3 + paddw m0, m2 + psrlw m1, 4 + psrlw m0, 4 +%if %2 == 1 ; avg + pavgw m0, [secq] + add secq, sec_str + pavgw m1, [secq] +%endif + SUM_SSE m0, m4, m1, m5, m6, m7 + + lea srcq, [srcq+src_strideq*4] + lea dstq, [dstq+dst_strideq*4] +%if %2 == 1 ; avg + add secq, sec_str +%endif +%endif + dec block_height + jg .x_other_y_zero_loop +%undef filter_x_a +%undef filter_x_b +%undef filter_rnd + STORE_AND_RET + +.x_nonhalf_y_nonzero: + cmp y_offsetd, 8 + jne .x_nonhalf_y_nonhalf + + ; x_offset == bilin interpolation && y_offset == 0.5 +%ifdef PIC + lea bilin_filter, [bilin_filter_m] +%endif + shl x_offsetd, filter_idx_shift +%if ARCH_X86_64 && mmsize == 16 + mova m8, [bilin_filter+x_offsetq] + mova m9, [bilin_filter+x_offsetq+16] + mova m10, [pw_8] +%define filter_x_a m8 +%define filter_x_b m9 +%define filter_rnd m10 +%else ; x86-32 +%if ARCH_X86=1 && CONFIG_PIC=1 +; y_offset == 0.5. We can reuse y_offset reg. +%define tempq y_offsetq + add x_offsetq, g_bilin_filterm +%define filter_x_a [x_offsetq] +%define filter_x_b [x_offsetq+16] + mov tempq, g_pw_8m +%define filter_rnd [tempq] +%else + add x_offsetq, bilin_filter +%define filter_x_a [x_offsetq] +%define filter_x_b [x_offsetq+16] +%define filter_rnd [pw_8] +%endif +%endif + +%if %1 == 16 + movu m0, [srcq] + movu m1, [srcq+16] + movu m2, [srcq+2] + movu m3, [srcq+18] + pmullw m0, filter_x_a + pmullw m2, filter_x_b + paddw m0, filter_rnd + pmullw m1, filter_x_a + pmullw m3, filter_x_b + paddw m1, filter_rnd + paddw m0, m2 + paddw m1, m3 + psrlw m0, 4 + psrlw m1, 4 + lea srcq, [srcq+src_strideq*2] +.x_other_y_half_loop: + movu m2, [srcq] + movu m3, [srcq+16] + movu m4, [srcq+2] + movu m5, [srcq+18] + pmullw m2, filter_x_a + pmullw m4, filter_x_b + paddw m2, filter_rnd + pmullw m3, filter_x_a + pmullw m5, filter_x_b + paddw m3, filter_rnd + paddw m2, m4 + paddw m3, m5 + mova m4, [dstq] + mova m5, [dstq+16] + psrlw m2, 4 + psrlw m3, 4 + pavgw m0, m2 + pavgw m1, m3 +%if %2 == 1 ; avg + pavgw m0, [secq] + pavgw m1, [secq+16] +%endif + SUM_SSE m0, m4, m1, m5, m6, m7 + mova m0, m2 + mova m1, m3 + + lea srcq, [srcq+src_strideq*2] + lea dstq, [dstq+dst_strideq*2] +%if %2 == 1 ; avg + add secq, sec_str +%endif +%else ; %1 < 16 + movu m0, [srcq] + movu m2, [srcq+2] + pmullw m0, filter_x_a + pmullw m2, filter_x_b + paddw m0, filter_rnd + paddw m0, m2 + psrlw m0, 4 + lea srcq, [srcq+src_strideq*2] +.x_other_y_half_loop: + movu m2, [srcq] + movu m3, [srcq+src_strideq*2] + movu m4, [srcq+2] + movu m5, [srcq+src_strideq*2+2] + pmullw m2, filter_x_a + pmullw m4, filter_x_b + paddw m2, filter_rnd + pmullw m3, filter_x_a + pmullw m5, filter_x_b + paddw m3, filter_rnd + paddw m2, m4 + paddw m3, m5 + mova m4, [dstq] + mova m5, [dstq+dst_strideq*2] + psrlw m2, 4 + psrlw m3, 4 + pavgw m0, m2 + pavgw m2, m3 +%if %2 == 1 ; avg + pavgw m0, [secq] + add secq, sec_str + pavgw m2, [secq] +%endif + SUM_SSE m0, m4, m2, m5, m6, m7 + mova m0, m3 + + lea srcq, [srcq+src_strideq*4] + lea dstq, [dstq+dst_strideq*4] +%if %2 == 1 ; avg + add secq, sec_str +%endif +%endif + dec block_height + jg .x_other_y_half_loop +%undef filter_x_a +%undef filter_x_b +%undef filter_rnd + STORE_AND_RET + +.x_nonhalf_y_nonhalf: +; loading filter - this is same as in 8-bit depth +%ifdef PIC + lea bilin_filter, [bilin_filter_m] +%endif + shl x_offsetd, filter_idx_shift ; filter_idx_shift = 5 + shl y_offsetd, filter_idx_shift +%if ARCH_X86_64 && mmsize == 16 + mova m8, [bilin_filter+x_offsetq] + mova m9, [bilin_filter+x_offsetq+16] + mova m10, [bilin_filter+y_offsetq] + mova m11, [bilin_filter+y_offsetq+16] + mova m12, [pw_8] +%define filter_x_a m8 +%define filter_x_b m9 +%define filter_y_a m10 +%define filter_y_b m11 +%define filter_rnd m12 +%else ; x86-32 +%if ARCH_X86=1 && CONFIG_PIC=1 +; In this case, there is NO unused register. Used src_stride register. Later, +; src_stride has to be loaded from stack when it is needed. +%define tempq src_strideq + mov tempq, g_bilin_filterm + add x_offsetq, tempq + add y_offsetq, tempq +%define filter_x_a [x_offsetq] +%define filter_x_b [x_offsetq+16] +%define filter_y_a [y_offsetq] +%define filter_y_b [y_offsetq+16] + + mov tempq, g_pw_8m +%define filter_rnd [tempq] +%else + add x_offsetq, bilin_filter + add y_offsetq, bilin_filter +%define filter_x_a [x_offsetq] +%define filter_x_b [x_offsetq+16] +%define filter_y_a [y_offsetq] +%define filter_y_b [y_offsetq+16] +%define filter_rnd [pw_8] +%endif +%endif +; end of load filter + + ; x_offset == bilin interpolation && y_offset == bilin interpolation +%if %1 == 16 + movu m0, [srcq] + movu m2, [srcq+2] + movu m1, [srcq+16] + movu m3, [srcq+18] + pmullw m0, filter_x_a + pmullw m2, filter_x_b + paddw m0, filter_rnd + pmullw m1, filter_x_a + pmullw m3, filter_x_b + paddw m1, filter_rnd + paddw m0, m2 + paddw m1, m3 + psrlw m0, 4 + psrlw m1, 4 + + INC_SRC_BY_SRC_STRIDE + +.x_other_y_other_loop: + movu m2, [srcq] + movu m4, [srcq+2] + movu m3, [srcq+16] + movu m5, [srcq+18] + pmullw m2, filter_x_a + pmullw m4, filter_x_b + paddw m2, filter_rnd + pmullw m3, filter_x_a + pmullw m5, filter_x_b + paddw m3, filter_rnd + paddw m2, m4 + paddw m3, m5 + psrlw m2, 4 + psrlw m3, 4 + mova m4, m2 + mova m5, m3 + pmullw m0, filter_y_a + pmullw m2, filter_y_b + paddw m0, filter_rnd + pmullw m1, filter_y_a + pmullw m3, filter_y_b + paddw m0, m2 + paddw m1, filter_rnd + mova m2, [dstq] + paddw m1, m3 + psrlw m0, 4 + psrlw m1, 4 + mova m3, [dstq+16] +%if %2 == 1 ; avg + pavgw m0, [secq] + pavgw m1, [secq+16] +%endif + SUM_SSE m0, m2, m1, m3, m6, m7 + mova m0, m4 + mova m1, m5 + + INC_SRC_BY_SRC_STRIDE + lea dstq, [dstq + dst_strideq * 2] +%if %2 == 1 ; avg + add secq, sec_str +%endif +%else ; %1 < 16 + movu m0, [srcq] + movu m2, [srcq+2] + pmullw m0, filter_x_a + pmullw m2, filter_x_b + paddw m0, filter_rnd + paddw m0, m2 + psrlw m0, 4 + + INC_SRC_BY_SRC_STRIDE + +.x_other_y_other_loop: + movu m2, [srcq] + movu m4, [srcq+2] + INC_SRC_BY_SRC_STRIDE + movu m3, [srcq] + movu m5, [srcq+2] + pmullw m2, filter_x_a + pmullw m4, filter_x_b + paddw m2, filter_rnd + pmullw m3, filter_x_a + pmullw m5, filter_x_b + paddw m3, filter_rnd + paddw m2, m4 + paddw m3, m5 + psrlw m2, 4 + psrlw m3, 4 + mova m4, m2 + mova m5, m3 + pmullw m0, filter_y_a + pmullw m2, filter_y_b + paddw m0, filter_rnd + pmullw m4, filter_y_a + pmullw m3, filter_y_b + paddw m0, m2 + paddw m4, filter_rnd + mova m2, [dstq] + paddw m4, m3 + psrlw m0, 4 + psrlw m4, 4 + mova m3, [dstq+dst_strideq*2] +%if %2 == 1 ; avg + pavgw m0, [secq] + add secq, sec_str + pavgw m4, [secq] +%endif + SUM_SSE m0, m2, m4, m3, m6, m7 + mova m0, m5 + + INC_SRC_BY_SRC_STRIDE + lea dstq, [dstq + dst_strideq * 4] +%if %2 == 1 ; avg + add secq, sec_str +%endif +%endif + dec block_height + jg .x_other_y_other_loop +%undef filter_x_a +%undef filter_x_b +%undef filter_y_a +%undef filter_y_b +%undef filter_rnd + STORE_AND_RET +%endmacro + +INIT_XMM sse2 +SUBPEL_VARIANCE 8 +SUBPEL_VARIANCE 16 + +INIT_XMM sse2 +SUBPEL_VARIANCE 8, 1 +SUBPEL_VARIANCE 16, 1 diff --git a/third_party/aom/aom_dsp/x86/highbd_subtract_sse2.c b/third_party/aom/aom_dsp/x86/highbd_subtract_sse2.c new file mode 100644 index 000000000..7bc8a0df3 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/highbd_subtract_sse2.c @@ -0,0 +1,364 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include <assert.h> +#include <emmintrin.h> +#include <stddef.h> + +#include "./aom_config.h" +#include "./aom_dsp_rtcd.h" + +typedef void (*SubtractWxHFuncType)(int16_t *diff, ptrdiff_t diff_stride, + const uint16_t *src, ptrdiff_t src_stride, + const uint16_t *pred, + ptrdiff_t pred_stride); + +static void subtract_4x4(int16_t *diff, ptrdiff_t diff_stride, + const uint16_t *src, ptrdiff_t src_stride, + const uint16_t *pred, ptrdiff_t pred_stride) { + __m128i u0, u1, u2, u3; + __m128i v0, v1, v2, v3; + __m128i x0, x1, x2, x3; + int64_t *store_diff = (int64_t *)(diff + 0 * diff_stride); + + u0 = _mm_loadu_si128((__m128i const *)(src + 0 * src_stride)); + u1 = _mm_loadu_si128((__m128i const *)(src + 1 * src_stride)); + u2 = _mm_loadu_si128((__m128i const *)(src + 2 * src_stride)); + u3 = _mm_loadu_si128((__m128i const *)(src + 3 * src_stride)); + + v0 = _mm_loadu_si128((__m128i const *)(pred + 0 * pred_stride)); + v1 = _mm_loadu_si128((__m128i const *)(pred + 1 * pred_stride)); + v2 = _mm_loadu_si128((__m128i const *)(pred + 2 * pred_stride)); + v3 = _mm_loadu_si128((__m128i const *)(pred + 3 * pred_stride)); + + x0 = _mm_sub_epi16(u0, v0); + x1 = _mm_sub_epi16(u1, v1); + x2 = _mm_sub_epi16(u2, v2); + x3 = _mm_sub_epi16(u3, v3); + + _mm_storel_epi64((__m128i *)store_diff, x0); + store_diff = (int64_t *)(diff + 1 * diff_stride); + _mm_storel_epi64((__m128i *)store_diff, x1); + store_diff = (int64_t *)(diff + 2 * diff_stride); + _mm_storel_epi64((__m128i *)store_diff, x2); + store_diff = (int64_t *)(diff + 3 * diff_stride); + _mm_storel_epi64((__m128i *)store_diff, x3); +} + +static void subtract_4x8(int16_t *diff, ptrdiff_t diff_stride, + const uint16_t *src, ptrdiff_t src_stride, + const uint16_t *pred, ptrdiff_t pred_stride) { + __m128i u0, u1, u2, u3, u4, u5, u6, u7; + __m128i v0, v1, v2, v3, v4, v5, v6, v7; + __m128i x0, x1, x2, x3, x4, x5, x6, x7; + int64_t *store_diff = (int64_t *)(diff + 0 * diff_stride); + + u0 = _mm_loadu_si128((__m128i const *)(src + 0 * src_stride)); + u1 = _mm_loadu_si128((__m128i const *)(src + 1 * src_stride)); + u2 = _mm_loadu_si128((__m128i const *)(src + 2 * src_stride)); + u3 = _mm_loadu_si128((__m128i const *)(src + 3 * src_stride)); + u4 = _mm_loadu_si128((__m128i const *)(src + 4 * src_stride)); + u5 = _mm_loadu_si128((__m128i const *)(src + 5 * src_stride)); + u6 = _mm_loadu_si128((__m128i const *)(src + 6 * src_stride)); + u7 = _mm_loadu_si128((__m128i const *)(src + 7 * src_stride)); + + v0 = _mm_loadu_si128((__m128i const *)(pred + 0 * pred_stride)); + v1 = _mm_loadu_si128((__m128i const *)(pred + 1 * pred_stride)); + v2 = _mm_loadu_si128((__m128i const *)(pred + 2 * pred_stride)); + v3 = _mm_loadu_si128((__m128i const *)(pred + 3 * pred_stride)); + v4 = _mm_loadu_si128((__m128i const *)(pred + 4 * pred_stride)); + v5 = _mm_loadu_si128((__m128i const *)(pred + 5 * pred_stride)); + v6 = _mm_loadu_si128((__m128i const *)(pred + 6 * pred_stride)); + v7 = _mm_loadu_si128((__m128i const *)(pred + 7 * pred_stride)); + + x0 = _mm_sub_epi16(u0, v0); + x1 = _mm_sub_epi16(u1, v1); + x2 = _mm_sub_epi16(u2, v2); + x3 = _mm_sub_epi16(u3, v3); + x4 = _mm_sub_epi16(u4, v4); + x5 = _mm_sub_epi16(u5, v5); + x6 = _mm_sub_epi16(u6, v6); + x7 = _mm_sub_epi16(u7, v7); + + _mm_storel_epi64((__m128i *)store_diff, x0); + store_diff = (int64_t *)(diff + 1 * diff_stride); + _mm_storel_epi64((__m128i *)store_diff, x1); + store_diff = (int64_t *)(diff + 2 * diff_stride); + _mm_storel_epi64((__m128i *)store_diff, x2); + store_diff = (int64_t *)(diff + 3 * diff_stride); + _mm_storel_epi64((__m128i *)store_diff, x3); + store_diff = (int64_t *)(diff + 4 * diff_stride); + _mm_storel_epi64((__m128i *)store_diff, x4); + store_diff = (int64_t *)(diff + 5 * diff_stride); + _mm_storel_epi64((__m128i *)store_diff, x5); + store_diff = (int64_t *)(diff + 6 * diff_stride); + _mm_storel_epi64((__m128i *)store_diff, x6); + store_diff = (int64_t *)(diff + 7 * diff_stride); + _mm_storel_epi64((__m128i *)store_diff, x7); +} + +static void subtract_8x4(int16_t *diff, ptrdiff_t diff_stride, + const uint16_t *src, ptrdiff_t src_stride, + const uint16_t *pred, ptrdiff_t pred_stride) { + __m128i u0, u1, u2, u3; + __m128i v0, v1, v2, v3; + __m128i x0, x1, x2, x3; + + u0 = _mm_loadu_si128((__m128i const *)(src + 0 * src_stride)); + u1 = _mm_loadu_si128((__m128i const *)(src + 1 * src_stride)); + u2 = _mm_loadu_si128((__m128i const *)(src + 2 * src_stride)); + u3 = _mm_loadu_si128((__m128i const *)(src + 3 * src_stride)); + + v0 = _mm_loadu_si128((__m128i const *)(pred + 0 * pred_stride)); + v1 = _mm_loadu_si128((__m128i const *)(pred + 1 * pred_stride)); + v2 = _mm_loadu_si128((__m128i const *)(pred + 2 * pred_stride)); + v3 = _mm_loadu_si128((__m128i const *)(pred + 3 * pred_stride)); + + x0 = _mm_sub_epi16(u0, v0); + x1 = _mm_sub_epi16(u1, v1); + x2 = _mm_sub_epi16(u2, v2); + x3 = _mm_sub_epi16(u3, v3); + + _mm_storeu_si128((__m128i *)(diff + 0 * diff_stride), x0); + _mm_storeu_si128((__m128i *)(diff + 1 * diff_stride), x1); + _mm_storeu_si128((__m128i *)(diff + 2 * diff_stride), x2); + _mm_storeu_si128((__m128i *)(diff + 3 * diff_stride), x3); +} + +static void subtract_8x8(int16_t *diff, ptrdiff_t diff_stride, + const uint16_t *src, ptrdiff_t src_stride, + const uint16_t *pred, ptrdiff_t pred_stride) { + __m128i u0, u1, u2, u3, u4, u5, u6, u7; + __m128i v0, v1, v2, v3, v4, v5, v6, v7; + __m128i x0, x1, x2, x3, x4, x5, x6, x7; + + u0 = _mm_loadu_si128((__m128i const *)(src + 0 * src_stride)); + u1 = _mm_loadu_si128((__m128i const *)(src + 1 * src_stride)); + u2 = _mm_loadu_si128((__m128i const *)(src + 2 * src_stride)); + u3 = _mm_loadu_si128((__m128i const *)(src + 3 * src_stride)); + u4 = _mm_loadu_si128((__m128i const *)(src + 4 * src_stride)); + u5 = _mm_loadu_si128((__m128i const *)(src + 5 * src_stride)); + u6 = _mm_loadu_si128((__m128i const *)(src + 6 * src_stride)); + u7 = _mm_loadu_si128((__m128i const *)(src + 7 * src_stride)); + + v0 = _mm_loadu_si128((__m128i const *)(pred + 0 * pred_stride)); + v1 = _mm_loadu_si128((__m128i const *)(pred + 1 * pred_stride)); + v2 = _mm_loadu_si128((__m128i const *)(pred + 2 * pred_stride)); + v3 = _mm_loadu_si128((__m128i const *)(pred + 3 * pred_stride)); + v4 = _mm_loadu_si128((__m128i const *)(pred + 4 * pred_stride)); + v5 = _mm_loadu_si128((__m128i const *)(pred + 5 * pred_stride)); + v6 = _mm_loadu_si128((__m128i const *)(pred + 6 * pred_stride)); + v7 = _mm_loadu_si128((__m128i const *)(pred + 7 * pred_stride)); + + x0 = _mm_sub_epi16(u0, v0); + x1 = _mm_sub_epi16(u1, v1); + x2 = _mm_sub_epi16(u2, v2); + x3 = _mm_sub_epi16(u3, v3); + x4 = _mm_sub_epi16(u4, v4); + x5 = _mm_sub_epi16(u5, v5); + x6 = _mm_sub_epi16(u6, v6); + x7 = _mm_sub_epi16(u7, v7); + + _mm_storeu_si128((__m128i *)(diff + 0 * diff_stride), x0); + _mm_storeu_si128((__m128i *)(diff + 1 * diff_stride), x1); + _mm_storeu_si128((__m128i *)(diff + 2 * diff_stride), x2); + _mm_storeu_si128((__m128i *)(diff + 3 * diff_stride), x3); + _mm_storeu_si128((__m128i *)(diff + 4 * diff_stride), x4); + _mm_storeu_si128((__m128i *)(diff + 5 * diff_stride), x5); + _mm_storeu_si128((__m128i *)(diff + 6 * diff_stride), x6); + _mm_storeu_si128((__m128i *)(diff + 7 * diff_stride), x7); +} + +static void subtract_8x16(int16_t *diff, ptrdiff_t diff_stride, + const uint16_t *src, ptrdiff_t src_stride, + const uint16_t *pred, ptrdiff_t pred_stride) { + subtract_8x8(diff, diff_stride, src, src_stride, pred, pred_stride); + diff += diff_stride << 3; + src += src_stride << 3; + pred += pred_stride << 3; + subtract_8x8(diff, diff_stride, src, src_stride, pred, pred_stride); +} + +static void subtract_16x8(int16_t *diff, ptrdiff_t diff_stride, + const uint16_t *src, ptrdiff_t src_stride, + const uint16_t *pred, ptrdiff_t pred_stride) { + subtract_8x8(diff, diff_stride, src, src_stride, pred, pred_stride); + diff += 8; + src += 8; + pred += 8; + subtract_8x8(diff, diff_stride, src, src_stride, pred, pred_stride); +} + +static void subtract_16x16(int16_t *diff, ptrdiff_t diff_stride, + const uint16_t *src, ptrdiff_t src_stride, + const uint16_t *pred, ptrdiff_t pred_stride) { + subtract_16x8(diff, diff_stride, src, src_stride, pred, pred_stride); + diff += diff_stride << 3; + src += src_stride << 3; + pred += pred_stride << 3; + subtract_16x8(diff, diff_stride, src, src_stride, pred, pred_stride); +} + +static void subtract_16x32(int16_t *diff, ptrdiff_t diff_stride, + const uint16_t *src, ptrdiff_t src_stride, + const uint16_t *pred, ptrdiff_t pred_stride) { + subtract_16x16(diff, diff_stride, src, src_stride, pred, pred_stride); + diff += diff_stride << 4; + src += src_stride << 4; + pred += pred_stride << 4; + subtract_16x16(diff, diff_stride, src, src_stride, pred, pred_stride); +} + +static void subtract_32x16(int16_t *diff, ptrdiff_t diff_stride, + const uint16_t *src, ptrdiff_t src_stride, + const uint16_t *pred, ptrdiff_t pred_stride) { + subtract_16x16(diff, diff_stride, src, src_stride, pred, pred_stride); + diff += 16; + src += 16; + pred += 16; + subtract_16x16(diff, diff_stride, src, src_stride, pred, pred_stride); +} + +static void subtract_32x32(int16_t *diff, ptrdiff_t diff_stride, + const uint16_t *src, ptrdiff_t src_stride, + const uint16_t *pred, ptrdiff_t pred_stride) { + subtract_32x16(diff, diff_stride, src, src_stride, pred, pred_stride); + diff += diff_stride << 4; + src += src_stride << 4; + pred += pred_stride << 4; + subtract_32x16(diff, diff_stride, src, src_stride, pred, pred_stride); +} + +static void subtract_32x64(int16_t *diff, ptrdiff_t diff_stride, + const uint16_t *src, ptrdiff_t src_stride, + const uint16_t *pred, ptrdiff_t pred_stride) { + subtract_32x32(diff, diff_stride, src, src_stride, pred, pred_stride); + diff += diff_stride << 5; + src += src_stride << 5; + pred += pred_stride << 5; + subtract_32x32(diff, diff_stride, src, src_stride, pred, pred_stride); +} + +static void subtract_64x32(int16_t *diff, ptrdiff_t diff_stride, + const uint16_t *src, ptrdiff_t src_stride, + const uint16_t *pred, ptrdiff_t pred_stride) { + subtract_32x32(diff, diff_stride, src, src_stride, pred, pred_stride); + diff += 32; + src += 32; + pred += 32; + subtract_32x32(diff, diff_stride, src, src_stride, pred, pred_stride); +} + +static void subtract_64x64(int16_t *diff, ptrdiff_t diff_stride, + const uint16_t *src, ptrdiff_t src_stride, + const uint16_t *pred, ptrdiff_t pred_stride) { + subtract_64x32(diff, diff_stride, src, src_stride, pred, pred_stride); + diff += diff_stride << 5; + src += src_stride << 5; + pred += pred_stride << 5; + subtract_64x32(diff, diff_stride, src, src_stride, pred, pred_stride); +} + +static void subtract_64x128(int16_t *diff, ptrdiff_t diff_stride, + const uint16_t *src, ptrdiff_t src_stride, + const uint16_t *pred, ptrdiff_t pred_stride) { + subtract_64x64(diff, diff_stride, src, src_stride, pred, pred_stride); + diff += diff_stride << 6; + src += src_stride << 6; + pred += pred_stride << 6; + subtract_64x64(diff, diff_stride, src, src_stride, pred, pred_stride); +} + +static void subtract_128x64(int16_t *diff, ptrdiff_t diff_stride, + const uint16_t *src, ptrdiff_t src_stride, + const uint16_t *pred, ptrdiff_t pred_stride) { + subtract_64x64(diff, diff_stride, src, src_stride, pred, pred_stride); + diff += 64; + src += 64; + pred += 64; + subtract_64x64(diff, diff_stride, src, src_stride, pred, pred_stride); +} + +static void subtract_128x128(int16_t *diff, ptrdiff_t diff_stride, + const uint16_t *src, ptrdiff_t src_stride, + const uint16_t *pred, ptrdiff_t pred_stride) { + subtract_128x64(diff, diff_stride, src, src_stride, pred, pred_stride); + diff += diff_stride << 6; + src += src_stride << 6; + pred += pred_stride << 6; + subtract_128x64(diff, diff_stride, src, src_stride, pred, pred_stride); +} + +static SubtractWxHFuncType getSubtractFunc(int rows, int cols) { + SubtractWxHFuncType ret_func_ptr = NULL; + if (rows == 4) { + if (cols == 4) { + ret_func_ptr = subtract_4x4; + } else if (cols == 8) { + ret_func_ptr = subtract_8x4; + } + } else if (rows == 8) { + if (cols == 4) { + ret_func_ptr = subtract_4x8; + } else if (cols == 8) { + ret_func_ptr = subtract_8x8; + } else if (cols == 16) { + ret_func_ptr = subtract_16x8; + } + } else if (rows == 16) { + if (cols == 8) { + ret_func_ptr = subtract_8x16; + } else if (cols == 16) { + ret_func_ptr = subtract_16x16; + } else if (cols == 32) { + ret_func_ptr = subtract_32x16; + } + } else if (rows == 32) { + if (cols == 16) { + ret_func_ptr = subtract_16x32; + } else if (cols == 32) { + ret_func_ptr = subtract_32x32; + } else if (cols == 64) { + ret_func_ptr = subtract_64x32; + } + } else if (rows == 64) { + if (cols == 32) { + ret_func_ptr = subtract_32x64; + } else if (cols == 64) { + ret_func_ptr = subtract_64x64; + } else if (cols == 128) { + ret_func_ptr = subtract_128x64; + } + } else if (rows == 128) { + if (cols == 64) { + ret_func_ptr = subtract_64x128; + } else if (cols == 128) { + ret_func_ptr = subtract_128x128; + } + } + if (!ret_func_ptr) { + assert(0); + } + return ret_func_ptr; +} + +void aom_highbd_subtract_block_sse2(int rows, int cols, int16_t *diff, + ptrdiff_t diff_stride, const uint8_t *src8, + ptrdiff_t src_stride, const uint8_t *pred8, + ptrdiff_t pred_stride, int bd) { + uint16_t *src = CONVERT_TO_SHORTPTR(src8); + uint16_t *pred = CONVERT_TO_SHORTPTR(pred8); + SubtractWxHFuncType func; + (void)bd; + + func = getSubtractFunc(rows, cols); + func(diff, diff_stride, src, src_stride, pred, pred_stride); +} diff --git a/third_party/aom/aom_dsp/x86/highbd_variance_impl_sse2.asm b/third_party/aom/aom_dsp/x86/highbd_variance_impl_sse2.asm new file mode 100644 index 000000000..cf8ea498c --- /dev/null +++ b/third_party/aom/aom_dsp/x86/highbd_variance_impl_sse2.asm @@ -0,0 +1,316 @@ +; +; Copyright (c) 2016, Alliance for Open Media. All rights reserved +; +; This source code is subject to the terms of the BSD 2 Clause License and +; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License +; was not distributed with this source code in the LICENSE file, you can +; obtain it at www.aomedia.org/license/software. If the Alliance for Open +; Media Patent License 1.0 was not distributed with this source code in the +; PATENTS file, you can obtain it at www.aomedia.org/license/patent. +; + +; + + +%include "aom_ports/x86_abi_support.asm" + +;unsigned int aom_highbd_calc16x16var_sse2 +;( +; unsigned char * src_ptr, +; int source_stride, +; unsigned char * ref_ptr, +; int recon_stride, +; unsigned int * SSE, +; int * Sum +;) +global sym(aom_highbd_calc16x16var_sse2) PRIVATE +sym(aom_highbd_calc16x16var_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 6 + SAVE_XMM 7 + push rbx + push rsi + push rdi + ; end prolog + + mov rsi, arg(0) ;[src_ptr] + mov rdi, arg(2) ;[ref_ptr] + + movsxd rax, DWORD PTR arg(1) ;[source_stride] + movsxd rdx, DWORD PTR arg(3) ;[recon_stride] + add rax, rax ; source stride in bytes + add rdx, rdx ; recon stride in bytes + + ; Prefetch data + prefetcht0 [rsi] + prefetcht0 [rsi+16] + prefetcht0 [rsi+rax] + prefetcht0 [rsi+rax+16] + lea rbx, [rsi+rax*2] + prefetcht0 [rbx] + prefetcht0 [rbx+16] + prefetcht0 [rbx+rax] + prefetcht0 [rbx+rax+16] + + prefetcht0 [rdi] + prefetcht0 [rdi+16] + prefetcht0 [rdi+rdx] + prefetcht0 [rdi+rdx+16] + lea rbx, [rdi+rdx*2] + prefetcht0 [rbx] + prefetcht0 [rbx+16] + prefetcht0 [rbx+rdx] + prefetcht0 [rbx+rdx+16] + + pxor xmm0, xmm0 ; clear xmm0 for unpack + pxor xmm7, xmm7 ; clear xmm7 for accumulating diffs + + pxor xmm6, xmm6 ; clear xmm6 for accumulating sse + mov rcx, 16 + +.var16loop: + movdqu xmm1, XMMWORD PTR [rsi] + movdqu xmm2, XMMWORD PTR [rdi] + + lea rbx, [rsi+rax*2] + prefetcht0 [rbx] + prefetcht0 [rbx+16] + prefetcht0 [rbx+rax] + prefetcht0 [rbx+rax+16] + lea rbx, [rdi+rdx*2] + prefetcht0 [rbx] + prefetcht0 [rbx+16] + prefetcht0 [rbx+rdx] + prefetcht0 [rbx+rdx+16] + + pxor xmm5, xmm5 + + psubw xmm1, xmm2 + movdqu xmm3, XMMWORD PTR [rsi+16] + paddw xmm5, xmm1 + pmaddwd xmm1, xmm1 + movdqu xmm2, XMMWORD PTR [rdi+16] + paddd xmm6, xmm1 + + psubw xmm3, xmm2 + movdqu xmm1, XMMWORD PTR [rsi+rax] + paddw xmm5, xmm3 + pmaddwd xmm3, xmm3 + movdqu xmm2, XMMWORD PTR [rdi+rdx] + paddd xmm6, xmm3 + + psubw xmm1, xmm2 + movdqu xmm3, XMMWORD PTR [rsi+rax+16] + paddw xmm5, xmm1 + pmaddwd xmm1, xmm1 + movdqu xmm2, XMMWORD PTR [rdi+rdx+16] + paddd xmm6, xmm1 + + psubw xmm3, xmm2 + paddw xmm5, xmm3 + pmaddwd xmm3, xmm3 + paddd xmm6, xmm3 + + movdqa xmm1, xmm5 + movdqa xmm2, xmm5 + pcmpgtw xmm1, xmm0 + pcmpeqw xmm2, xmm0 + por xmm1, xmm2 + pcmpeqw xmm1, xmm0 + movdqa xmm2, xmm5 + punpcklwd xmm5, xmm1 + punpckhwd xmm2, xmm1 + paddd xmm7, xmm5 + paddd xmm7, xmm2 + + lea rsi, [rsi + 2*rax] + lea rdi, [rdi + 2*rdx] + sub rcx, 2 + jnz .var16loop + + movdqa xmm4, xmm6 + punpckldq xmm6, xmm0 + + punpckhdq xmm4, xmm0 + movdqa xmm5, xmm7 + + paddd xmm6, xmm4 + punpckldq xmm7, xmm0 + + punpckhdq xmm5, xmm0 + paddd xmm7, xmm5 + + movdqa xmm4, xmm6 + movdqa xmm5, xmm7 + + psrldq xmm4, 8 + psrldq xmm5, 8 + + paddd xmm6, xmm4 + paddd xmm7, xmm5 + + mov rdi, arg(4) ; [SSE] + mov rax, arg(5) ; [Sum] + + movd DWORD PTR [rdi], xmm6 + movd DWORD PTR [rax], xmm7 + + + ; begin epilog + pop rdi + pop rsi + pop rbx + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret + + +;unsigned int aom_highbd_calc8x8var_sse2 +;( +; unsigned char * src_ptr, +; int source_stride, +; unsigned char * ref_ptr, +; int recon_stride, +; unsigned int * SSE, +; int * Sum +;) +global sym(aom_highbd_calc8x8var_sse2) PRIVATE +sym(aom_highbd_calc8x8var_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 6 + SAVE_XMM 7 + push rbx + push rsi + push rdi + ; end prolog + + mov rsi, arg(0) ;[src_ptr] + mov rdi, arg(2) ;[ref_ptr] + + movsxd rax, DWORD PTR arg(1) ;[source_stride] + movsxd rdx, DWORD PTR arg(3) ;[recon_stride] + add rax, rax ; source stride in bytes + add rdx, rdx ; recon stride in bytes + + ; Prefetch data + prefetcht0 [rsi] + prefetcht0 [rsi+rax] + lea rbx, [rsi+rax*2] + prefetcht0 [rbx] + prefetcht0 [rbx+rax] + + prefetcht0 [rdi] + prefetcht0 [rdi+rdx] + lea rbx, [rdi+rdx*2] + prefetcht0 [rbx] + prefetcht0 [rbx+rdx] + + pxor xmm0, xmm0 ; clear xmm0 for unpack + pxor xmm7, xmm7 ; clear xmm7 for accumulating diffs + + pxor xmm6, xmm6 ; clear xmm6 for accumulating sse + mov rcx, 8 + +.var8loop: + movdqu xmm1, XMMWORD PTR [rsi] + movdqu xmm2, XMMWORD PTR [rdi] + + lea rbx, [rsi+rax*4] + prefetcht0 [rbx] + prefetcht0 [rbx+rax] + lea rbx, [rbx+rax*2] + prefetcht0 [rbx] + prefetcht0 [rbx+rax] + lea rbx, [rdi+rdx*4] + prefetcht0 [rbx] + prefetcht0 [rbx+rdx] + lea rbx, [rbx+rdx*2] + prefetcht0 [rbx] + prefetcht0 [rbx+rdx] + + pxor xmm5, xmm5 + + psubw xmm1, xmm2 + movdqu xmm3, XMMWORD PTR [rsi+rax] + paddw xmm5, xmm1 + pmaddwd xmm1, xmm1 + movdqu xmm2, XMMWORD PTR [rdi+rdx] + paddd xmm6, xmm1 + + lea rsi, [rsi + 2*rax] + lea rdi, [rdi + 2*rdx] + + psubw xmm3, xmm2 + movdqu xmm1, XMMWORD PTR [rsi] + paddw xmm5, xmm3 + pmaddwd xmm3, xmm3 + movdqu xmm2, XMMWORD PTR [rdi] + paddd xmm6, xmm3 + + psubw xmm1, xmm2 + movdqu xmm3, XMMWORD PTR [rsi+rax] + paddw xmm5, xmm1 + pmaddwd xmm1, xmm1 + movdqu xmm2, XMMWORD PTR [rdi+rdx] + paddd xmm6, xmm1 + + psubw xmm3, xmm2 + paddw xmm5, xmm3 + pmaddwd xmm3, xmm3 + paddd xmm6, xmm3 + + movdqa xmm1, xmm5 + movdqa xmm2, xmm5 + pcmpgtw xmm1, xmm0 + pcmpeqw xmm2, xmm0 + por xmm1, xmm2 + pcmpeqw xmm1, xmm0 + movdqa xmm2, xmm5 + punpcklwd xmm5, xmm1 + punpckhwd xmm2, xmm1 + paddd xmm7, xmm5 + paddd xmm7, xmm2 + + lea rsi, [rsi + 2*rax] + lea rdi, [rdi + 2*rdx] + sub rcx, 4 + jnz .var8loop + + movdqa xmm4, xmm6 + punpckldq xmm6, xmm0 + + punpckhdq xmm4, xmm0 + movdqa xmm5, xmm7 + + paddd xmm6, xmm4 + punpckldq xmm7, xmm0 + + punpckhdq xmm5, xmm0 + paddd xmm7, xmm5 + + movdqa xmm4, xmm6 + movdqa xmm5, xmm7 + + psrldq xmm4, 8 + psrldq xmm5, 8 + + paddd xmm6, xmm4 + paddd xmm7, xmm5 + + mov rdi, arg(4) ; [SSE] + mov rax, arg(5) ; [Sum] + + movd DWORD PTR [rdi], xmm6 + movd DWORD PTR [rax], xmm7 + + ; begin epilog + pop rdi + pop rsi + pop rbx + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret diff --git a/third_party/aom/aom_dsp/x86/highbd_variance_sse2.c b/third_party/aom/aom_dsp/x86/highbd_variance_sse2.c new file mode 100644 index 000000000..29f96ce24 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/highbd_variance_sse2.c @@ -0,0 +1,695 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include <emmintrin.h> // SSE2 + +#include "./aom_config.h" +#include "./aom_dsp_rtcd.h" + +#include "aom_ports/mem.h" + +typedef uint32_t (*high_variance_fn_t)(const uint16_t *src, int src_stride, + const uint16_t *ref, int ref_stride, + uint32_t *sse, int *sum); + +uint32_t aom_highbd_calc8x8var_sse2(const uint16_t *src, int src_stride, + const uint16_t *ref, int ref_stride, + uint32_t *sse, int *sum); + +uint32_t aom_highbd_calc16x16var_sse2(const uint16_t *src, int src_stride, + const uint16_t *ref, int ref_stride, + uint32_t *sse, int *sum); + +static void highbd_8_variance_sse2(const uint16_t *src, int src_stride, + const uint16_t *ref, int ref_stride, int w, + int h, uint32_t *sse, int *sum, + high_variance_fn_t var_fn, int block_size) { + int i, j; + + *sse = 0; + *sum = 0; + + for (i = 0; i < h; i += block_size) { + for (j = 0; j < w; j += block_size) { + unsigned int sse0; + int sum0; + var_fn(src + src_stride * i + j, src_stride, ref + ref_stride * i + j, + ref_stride, &sse0, &sum0); + *sse += sse0; + *sum += sum0; + } + } +} + +static void highbd_10_variance_sse2(const uint16_t *src, int src_stride, + const uint16_t *ref, int ref_stride, int w, + int h, uint32_t *sse, int *sum, + high_variance_fn_t var_fn, int block_size) { + int i, j; + uint64_t sse_long = 0; + int32_t sum_long = 0; + + for (i = 0; i < h; i += block_size) { + for (j = 0; j < w; j += block_size) { + unsigned int sse0; + int sum0; + var_fn(src + src_stride * i + j, src_stride, ref + ref_stride * i + j, + ref_stride, &sse0, &sum0); + sse_long += sse0; + sum_long += sum0; + } + } + *sum = ROUND_POWER_OF_TWO(sum_long, 2); + *sse = (uint32_t)ROUND_POWER_OF_TWO(sse_long, 4); +} + +static void highbd_12_variance_sse2(const uint16_t *src, int src_stride, + const uint16_t *ref, int ref_stride, int w, + int h, uint32_t *sse, int *sum, + high_variance_fn_t var_fn, int block_size) { + int i, j; + uint64_t sse_long = 0; + int32_t sum_long = 0; + + for (i = 0; i < h; i += block_size) { + for (j = 0; j < w; j += block_size) { + unsigned int sse0; + int sum0; + var_fn(src + src_stride * i + j, src_stride, ref + ref_stride * i + j, + ref_stride, &sse0, &sum0); + sse_long += sse0; + sum_long += sum0; + } + } + *sum = ROUND_POWER_OF_TWO(sum_long, 4); + *sse = (uint32_t)ROUND_POWER_OF_TWO(sse_long, 8); +} + +#define HIGH_GET_VAR(S) \ + void aom_highbd_get##S##x##S##var_sse2(const uint8_t *src8, int src_stride, \ + const uint8_t *ref8, int ref_stride, \ + uint32_t *sse, int *sum) { \ + uint16_t *src = CONVERT_TO_SHORTPTR(src8); \ + uint16_t *ref = CONVERT_TO_SHORTPTR(ref8); \ + aom_highbd_calc##S##x##S##var_sse2(src, src_stride, ref, ref_stride, sse, \ + sum); \ + } \ + \ + void aom_highbd_10_get##S##x##S##var_sse2( \ + const uint8_t *src8, int src_stride, const uint8_t *ref8, \ + int ref_stride, uint32_t *sse, int *sum) { \ + uint16_t *src = CONVERT_TO_SHORTPTR(src8); \ + uint16_t *ref = CONVERT_TO_SHORTPTR(ref8); \ + aom_highbd_calc##S##x##S##var_sse2(src, src_stride, ref, ref_stride, sse, \ + sum); \ + *sum = ROUND_POWER_OF_TWO(*sum, 2); \ + *sse = ROUND_POWER_OF_TWO(*sse, 4); \ + } \ + \ + void aom_highbd_12_get##S##x##S##var_sse2( \ + const uint8_t *src8, int src_stride, const uint8_t *ref8, \ + int ref_stride, uint32_t *sse, int *sum) { \ + uint16_t *src = CONVERT_TO_SHORTPTR(src8); \ + uint16_t *ref = CONVERT_TO_SHORTPTR(ref8); \ + aom_highbd_calc##S##x##S##var_sse2(src, src_stride, ref, ref_stride, sse, \ + sum); \ + *sum = ROUND_POWER_OF_TWO(*sum, 4); \ + *sse = ROUND_POWER_OF_TWO(*sse, 8); \ + } + +HIGH_GET_VAR(16); +HIGH_GET_VAR(8); + +#undef HIGH_GET_VAR + +#define VAR_FN(w, h, block_size, shift) \ + uint32_t aom_highbd_8_variance##w##x##h##_sse2( \ + const uint8_t *src8, int src_stride, const uint8_t *ref8, \ + int ref_stride, uint32_t *sse) { \ + int sum; \ + uint16_t *src = CONVERT_TO_SHORTPTR(src8); \ + uint16_t *ref = CONVERT_TO_SHORTPTR(ref8); \ + highbd_8_variance_sse2( \ + src, src_stride, ref, ref_stride, w, h, sse, &sum, \ + aom_highbd_calc##block_size##x##block_size##var_sse2, block_size); \ + return *sse - (uint32_t)(((int64_t)sum * sum) >> shift); \ + } \ + \ + uint32_t aom_highbd_10_variance##w##x##h##_sse2( \ + const uint8_t *src8, int src_stride, const uint8_t *ref8, \ + int ref_stride, uint32_t *sse) { \ + int sum; \ + int64_t var; \ + uint16_t *src = CONVERT_TO_SHORTPTR(src8); \ + uint16_t *ref = CONVERT_TO_SHORTPTR(ref8); \ + highbd_10_variance_sse2( \ + src, src_stride, ref, ref_stride, w, h, sse, &sum, \ + aom_highbd_calc##block_size##x##block_size##var_sse2, block_size); \ + var = (int64_t)(*sse) - (((int64_t)sum * sum) >> shift); \ + return (var >= 0) ? (uint32_t)var : 0; \ + } \ + \ + uint32_t aom_highbd_12_variance##w##x##h##_sse2( \ + const uint8_t *src8, int src_stride, const uint8_t *ref8, \ + int ref_stride, uint32_t *sse) { \ + int sum; \ + int64_t var; \ + uint16_t *src = CONVERT_TO_SHORTPTR(src8); \ + uint16_t *ref = CONVERT_TO_SHORTPTR(ref8); \ + highbd_12_variance_sse2( \ + src, src_stride, ref, ref_stride, w, h, sse, &sum, \ + aom_highbd_calc##block_size##x##block_size##var_sse2, block_size); \ + var = (int64_t)(*sse) - (((int64_t)sum * sum) >> shift); \ + return (var >= 0) ? (uint32_t)var : 0; \ + } + +VAR_FN(64, 64, 16, 12); +VAR_FN(64, 32, 16, 11); +VAR_FN(32, 64, 16, 11); +VAR_FN(32, 32, 16, 10); +VAR_FN(32, 16, 16, 9); +VAR_FN(16, 32, 16, 9); +VAR_FN(16, 16, 16, 8); +VAR_FN(16, 8, 8, 7); +VAR_FN(8, 16, 8, 7); +VAR_FN(8, 8, 8, 6); + +#undef VAR_FN + +unsigned int aom_highbd_8_mse16x16_sse2(const uint8_t *src8, int src_stride, + const uint8_t *ref8, int ref_stride, + unsigned int *sse) { + int sum; + uint16_t *src = CONVERT_TO_SHORTPTR(src8); + uint16_t *ref = CONVERT_TO_SHORTPTR(ref8); + highbd_8_variance_sse2(src, src_stride, ref, ref_stride, 16, 16, sse, &sum, + aom_highbd_calc16x16var_sse2, 16); + return *sse; +} + +unsigned int aom_highbd_10_mse16x16_sse2(const uint8_t *src8, int src_stride, + const uint8_t *ref8, int ref_stride, + unsigned int *sse) { + int sum; + uint16_t *src = CONVERT_TO_SHORTPTR(src8); + uint16_t *ref = CONVERT_TO_SHORTPTR(ref8); + highbd_10_variance_sse2(src, src_stride, ref, ref_stride, 16, 16, sse, &sum, + aom_highbd_calc16x16var_sse2, 16); + return *sse; +} + +unsigned int aom_highbd_12_mse16x16_sse2(const uint8_t *src8, int src_stride, + const uint8_t *ref8, int ref_stride, + unsigned int *sse) { + int sum; + uint16_t *src = CONVERT_TO_SHORTPTR(src8); + uint16_t *ref = CONVERT_TO_SHORTPTR(ref8); + highbd_12_variance_sse2(src, src_stride, ref, ref_stride, 16, 16, sse, &sum, + aom_highbd_calc16x16var_sse2, 16); + return *sse; +} + +unsigned int aom_highbd_8_mse8x8_sse2(const uint8_t *src8, int src_stride, + const uint8_t *ref8, int ref_stride, + unsigned int *sse) { + int sum; + uint16_t *src = CONVERT_TO_SHORTPTR(src8); + uint16_t *ref = CONVERT_TO_SHORTPTR(ref8); + highbd_8_variance_sse2(src, src_stride, ref, ref_stride, 8, 8, sse, &sum, + aom_highbd_calc8x8var_sse2, 8); + return *sse; +} + +unsigned int aom_highbd_10_mse8x8_sse2(const uint8_t *src8, int src_stride, + const uint8_t *ref8, int ref_stride, + unsigned int *sse) { + int sum; + uint16_t *src = CONVERT_TO_SHORTPTR(src8); + uint16_t *ref = CONVERT_TO_SHORTPTR(ref8); + highbd_10_variance_sse2(src, src_stride, ref, ref_stride, 8, 8, sse, &sum, + aom_highbd_calc8x8var_sse2, 8); + return *sse; +} + +unsigned int aom_highbd_12_mse8x8_sse2(const uint8_t *src8, int src_stride, + const uint8_t *ref8, int ref_stride, + unsigned int *sse) { + int sum; + uint16_t *src = CONVERT_TO_SHORTPTR(src8); + uint16_t *ref = CONVERT_TO_SHORTPTR(ref8); + highbd_12_variance_sse2(src, src_stride, ref, ref_stride, 8, 8, sse, &sum, + aom_highbd_calc8x8var_sse2, 8); + return *sse; +} + +// The 2 unused parameters are place holders for PIC enabled build. +// These definitions are for functions defined in +// highbd_subpel_variance_impl_sse2.asm +#define DECL(w, opt) \ + int aom_highbd_sub_pixel_variance##w##xh_##opt( \ + const uint16_t *src, ptrdiff_t src_stride, int x_offset, int y_offset, \ + const uint16_t *dst, ptrdiff_t dst_stride, int height, \ + unsigned int *sse, void *unused0, void *unused); +#define DECLS(opt) \ + DECL(8, opt); \ + DECL(16, opt) + +DECLS(sse2); + +#undef DECLS +#undef DECL + +#define FN(w, h, wf, wlog2, hlog2, opt, cast) \ + uint32_t aom_highbd_8_sub_pixel_variance##w##x##h##_##opt( \ + const uint8_t *src8, int src_stride, int x_offset, int y_offset, \ + const uint8_t *dst8, int dst_stride, uint32_t *sse_ptr) { \ + uint32_t sse; \ + uint16_t *src = CONVERT_TO_SHORTPTR(src8); \ + uint16_t *dst = CONVERT_TO_SHORTPTR(dst8); \ + int se = aom_highbd_sub_pixel_variance##wf##xh_##opt( \ + src, src_stride, x_offset, y_offset, dst, dst_stride, h, &sse, NULL, \ + NULL); \ + if (w > wf) { \ + unsigned int sse2; \ + int se2 = aom_highbd_sub_pixel_variance##wf##xh_##opt( \ + src + 16, src_stride, x_offset, y_offset, dst + 16, dst_stride, h, \ + &sse2, NULL, NULL); \ + se += se2; \ + sse += sse2; \ + if (w > wf * 2) { \ + se2 = aom_highbd_sub_pixel_variance##wf##xh_##opt( \ + src + 32, src_stride, x_offset, y_offset, dst + 32, dst_stride, h, \ + &sse2, NULL, NULL); \ + se += se2; \ + sse += sse2; \ + se2 = aom_highbd_sub_pixel_variance##wf##xh_##opt( \ + src + 48, src_stride, x_offset, y_offset, dst + 48, dst_stride, h, \ + &sse2, NULL, NULL); \ + se += se2; \ + sse += sse2; \ + } \ + } \ + *sse_ptr = sse; \ + return sse - (uint32_t)((cast se * se) >> (wlog2 + hlog2)); \ + } \ + \ + uint32_t aom_highbd_10_sub_pixel_variance##w##x##h##_##opt( \ + const uint8_t *src8, int src_stride, int x_offset, int y_offset, \ + const uint8_t *dst8, int dst_stride, uint32_t *sse_ptr) { \ + int64_t var; \ + uint32_t sse; \ + uint16_t *src = CONVERT_TO_SHORTPTR(src8); \ + uint16_t *dst = CONVERT_TO_SHORTPTR(dst8); \ + int se = aom_highbd_sub_pixel_variance##wf##xh_##opt( \ + src, src_stride, x_offset, y_offset, dst, dst_stride, h, &sse, NULL, \ + NULL); \ + if (w > wf) { \ + uint32_t sse2; \ + int se2 = aom_highbd_sub_pixel_variance##wf##xh_##opt( \ + src + 16, src_stride, x_offset, y_offset, dst + 16, dst_stride, h, \ + &sse2, NULL, NULL); \ + se += se2; \ + sse += sse2; \ + if (w > wf * 2) { \ + se2 = aom_highbd_sub_pixel_variance##wf##xh_##opt( \ + src + 32, src_stride, x_offset, y_offset, dst + 32, dst_stride, h, \ + &sse2, NULL, NULL); \ + se += se2; \ + sse += sse2; \ + se2 = aom_highbd_sub_pixel_variance##wf##xh_##opt( \ + src + 48, src_stride, x_offset, y_offset, dst + 48, dst_stride, h, \ + &sse2, NULL, NULL); \ + se += se2; \ + sse += sse2; \ + } \ + } \ + se = ROUND_POWER_OF_TWO(se, 2); \ + sse = ROUND_POWER_OF_TWO(sse, 4); \ + *sse_ptr = sse; \ + var = (int64_t)(sse) - ((cast se * se) >> (wlog2 + hlog2)); \ + return (var >= 0) ? (uint32_t)var : 0; \ + } \ + \ + uint32_t aom_highbd_12_sub_pixel_variance##w##x##h##_##opt( \ + const uint8_t *src8, int src_stride, int x_offset, int y_offset, \ + const uint8_t *dst8, int dst_stride, uint32_t *sse_ptr) { \ + int start_row; \ + uint32_t sse; \ + int se = 0; \ + int64_t var; \ + uint64_t long_sse = 0; \ + uint16_t *src = CONVERT_TO_SHORTPTR(src8); \ + uint16_t *dst = CONVERT_TO_SHORTPTR(dst8); \ + for (start_row = 0; start_row < h; start_row += 16) { \ + uint32_t sse2; \ + int height = h - start_row < 16 ? h - start_row : 16; \ + int se2 = aom_highbd_sub_pixel_variance##wf##xh_##opt( \ + src + (start_row * src_stride), src_stride, x_offset, y_offset, \ + dst + (start_row * dst_stride), dst_stride, height, &sse2, NULL, \ + NULL); \ + se += se2; \ + long_sse += sse2; \ + if (w > wf) { \ + se2 = aom_highbd_sub_pixel_variance##wf##xh_##opt( \ + src + 16 + (start_row * src_stride), src_stride, x_offset, \ + y_offset, dst + 16 + (start_row * dst_stride), dst_stride, height, \ + &sse2, NULL, NULL); \ + se += se2; \ + long_sse += sse2; \ + if (w > wf * 2) { \ + se2 = aom_highbd_sub_pixel_variance##wf##xh_##opt( \ + src + 32 + (start_row * src_stride), src_stride, x_offset, \ + y_offset, dst + 32 + (start_row * dst_stride), dst_stride, \ + height, &sse2, NULL, NULL); \ + se += se2; \ + long_sse += sse2; \ + se2 = aom_highbd_sub_pixel_variance##wf##xh_##opt( \ + src + 48 + (start_row * src_stride), src_stride, x_offset, \ + y_offset, dst + 48 + (start_row * dst_stride), dst_stride, \ + height, &sse2, NULL, NULL); \ + se += se2; \ + long_sse += sse2; \ + } \ + } \ + } \ + se = ROUND_POWER_OF_TWO(se, 4); \ + sse = (uint32_t)ROUND_POWER_OF_TWO(long_sse, 8); \ + *sse_ptr = sse; \ + var = (int64_t)(sse) - ((cast se * se) >> (wlog2 + hlog2)); \ + return (var >= 0) ? (uint32_t)var : 0; \ + } + +#define FNS(opt) \ + FN(64, 64, 16, 6, 6, opt, (int64_t)); \ + FN(64, 32, 16, 6, 5, opt, (int64_t)); \ + FN(32, 64, 16, 5, 6, opt, (int64_t)); \ + FN(32, 32, 16, 5, 5, opt, (int64_t)); \ + FN(32, 16, 16, 5, 4, opt, (int64_t)); \ + FN(16, 32, 16, 4, 5, opt, (int64_t)); \ + FN(16, 16, 16, 4, 4, opt, (int64_t)); \ + FN(16, 8, 16, 4, 3, opt, (int64_t)); \ + FN(8, 16, 8, 3, 4, opt, (int64_t)); \ + FN(8, 8, 8, 3, 3, opt, (int64_t)); \ + FN(8, 4, 8, 3, 2, opt, (int64_t)); + +FNS(sse2); + +#undef FNS +#undef FN + +// The 2 unused parameters are place holders for PIC enabled build. +#define DECL(w, opt) \ + int aom_highbd_sub_pixel_avg_variance##w##xh_##opt( \ + const uint16_t *src, ptrdiff_t src_stride, int x_offset, int y_offset, \ + const uint16_t *dst, ptrdiff_t dst_stride, const uint16_t *sec, \ + ptrdiff_t sec_stride, int height, unsigned int *sse, void *unused0, \ + void *unused); +#define DECLS(opt1) \ + DECL(16, opt1) \ + DECL(8, opt1) + +DECLS(sse2); +#undef DECL +#undef DECLS + +#define FN(w, h, wf, wlog2, hlog2, opt, cast) \ + uint32_t aom_highbd_8_sub_pixel_avg_variance##w##x##h##_##opt( \ + const uint8_t *src8, int src_stride, int x_offset, int y_offset, \ + const uint8_t *dst8, int dst_stride, uint32_t *sse_ptr, \ + const uint8_t *sec8) { \ + uint32_t sse; \ + uint16_t *src = CONVERT_TO_SHORTPTR(src8); \ + uint16_t *dst = CONVERT_TO_SHORTPTR(dst8); \ + uint16_t *sec = CONVERT_TO_SHORTPTR(sec8); \ + int se = aom_highbd_sub_pixel_avg_variance##wf##xh_##opt( \ + src, src_stride, x_offset, y_offset, dst, dst_stride, sec, w, h, &sse, \ + NULL, NULL); \ + if (w > wf) { \ + uint32_t sse2; \ + int se2 = aom_highbd_sub_pixel_avg_variance##wf##xh_##opt( \ + src + 16, src_stride, x_offset, y_offset, dst + 16, dst_stride, \ + sec + 16, w, h, &sse2, NULL, NULL); \ + se += se2; \ + sse += sse2; \ + if (w > wf * 2) { \ + se2 = aom_highbd_sub_pixel_avg_variance##wf##xh_##opt( \ + src + 32, src_stride, x_offset, y_offset, dst + 32, dst_stride, \ + sec + 32, w, h, &sse2, NULL, NULL); \ + se += se2; \ + sse += sse2; \ + se2 = aom_highbd_sub_pixel_avg_variance##wf##xh_##opt( \ + src + 48, src_stride, x_offset, y_offset, dst + 48, dst_stride, \ + sec + 48, w, h, &sse2, NULL, NULL); \ + se += se2; \ + sse += sse2; \ + } \ + } \ + *sse_ptr = sse; \ + return sse - (uint32_t)((cast se * se) >> (wlog2 + hlog2)); \ + } \ + \ + uint32_t aom_highbd_10_sub_pixel_avg_variance##w##x##h##_##opt( \ + const uint8_t *src8, int src_stride, int x_offset, int y_offset, \ + const uint8_t *dst8, int dst_stride, uint32_t *sse_ptr, \ + const uint8_t *sec8) { \ + int64_t var; \ + uint32_t sse; \ + uint16_t *src = CONVERT_TO_SHORTPTR(src8); \ + uint16_t *dst = CONVERT_TO_SHORTPTR(dst8); \ + uint16_t *sec = CONVERT_TO_SHORTPTR(sec8); \ + int se = aom_highbd_sub_pixel_avg_variance##wf##xh_##opt( \ + src, src_stride, x_offset, y_offset, dst, dst_stride, sec, w, h, &sse, \ + NULL, NULL); \ + if (w > wf) { \ + uint32_t sse2; \ + int se2 = aom_highbd_sub_pixel_avg_variance##wf##xh_##opt( \ + src + 16, src_stride, x_offset, y_offset, dst + 16, dst_stride, \ + sec + 16, w, h, &sse2, NULL, NULL); \ + se += se2; \ + sse += sse2; \ + if (w > wf * 2) { \ + se2 = aom_highbd_sub_pixel_avg_variance##wf##xh_##opt( \ + src + 32, src_stride, x_offset, y_offset, dst + 32, dst_stride, \ + sec + 32, w, h, &sse2, NULL, NULL); \ + se += se2; \ + sse += sse2; \ + se2 = aom_highbd_sub_pixel_avg_variance##wf##xh_##opt( \ + src + 48, src_stride, x_offset, y_offset, dst + 48, dst_stride, \ + sec + 48, w, h, &sse2, NULL, NULL); \ + se += se2; \ + sse += sse2; \ + } \ + } \ + se = ROUND_POWER_OF_TWO(se, 2); \ + sse = ROUND_POWER_OF_TWO(sse, 4); \ + *sse_ptr = sse; \ + var = (int64_t)(sse) - ((cast se * se) >> (wlog2 + hlog2)); \ + return (var >= 0) ? (uint32_t)var : 0; \ + } \ + \ + uint32_t aom_highbd_12_sub_pixel_avg_variance##w##x##h##_##opt( \ + const uint8_t *src8, int src_stride, int x_offset, int y_offset, \ + const uint8_t *dst8, int dst_stride, uint32_t *sse_ptr, \ + const uint8_t *sec8) { \ + int start_row; \ + int64_t var; \ + uint32_t sse; \ + int se = 0; \ + uint64_t long_sse = 0; \ + uint16_t *src = CONVERT_TO_SHORTPTR(src8); \ + uint16_t *dst = CONVERT_TO_SHORTPTR(dst8); \ + uint16_t *sec = CONVERT_TO_SHORTPTR(sec8); \ + for (start_row = 0; start_row < h; start_row += 16) { \ + uint32_t sse2; \ + int height = h - start_row < 16 ? h - start_row : 16; \ + int se2 = aom_highbd_sub_pixel_avg_variance##wf##xh_##opt( \ + src + (start_row * src_stride), src_stride, x_offset, y_offset, \ + dst + (start_row * dst_stride), dst_stride, sec + (start_row * w), \ + w, height, &sse2, NULL, NULL); \ + se += se2; \ + long_sse += sse2; \ + if (w > wf) { \ + se2 = aom_highbd_sub_pixel_avg_variance##wf##xh_##opt( \ + src + 16 + (start_row * src_stride), src_stride, x_offset, \ + y_offset, dst + 16 + (start_row * dst_stride), dst_stride, \ + sec + 16 + (start_row * w), w, height, &sse2, NULL, NULL); \ + se += se2; \ + long_sse += sse2; \ + if (w > wf * 2) { \ + se2 = aom_highbd_sub_pixel_avg_variance##wf##xh_##opt( \ + src + 32 + (start_row * src_stride), src_stride, x_offset, \ + y_offset, dst + 32 + (start_row * dst_stride), dst_stride, \ + sec + 32 + (start_row * w), w, height, &sse2, NULL, NULL); \ + se += se2; \ + long_sse += sse2; \ + se2 = aom_highbd_sub_pixel_avg_variance##wf##xh_##opt( \ + src + 48 + (start_row * src_stride), src_stride, x_offset, \ + y_offset, dst + 48 + (start_row * dst_stride), dst_stride, \ + sec + 48 + (start_row * w), w, height, &sse2, NULL, NULL); \ + se += se2; \ + long_sse += sse2; \ + } \ + } \ + } \ + se = ROUND_POWER_OF_TWO(se, 4); \ + sse = (uint32_t)ROUND_POWER_OF_TWO(long_sse, 8); \ + *sse_ptr = sse; \ + var = (int64_t)(sse) - ((cast se * se) >> (wlog2 + hlog2)); \ + return (var >= 0) ? (uint32_t)var : 0; \ + } + +#define FNS(opt1) \ + FN(64, 64, 16, 6, 6, opt1, (int64_t)); \ + FN(64, 32, 16, 6, 5, opt1, (int64_t)); \ + FN(32, 64, 16, 5, 6, opt1, (int64_t)); \ + FN(32, 32, 16, 5, 5, opt1, (int64_t)); \ + FN(32, 16, 16, 5, 4, opt1, (int64_t)); \ + FN(16, 32, 16, 4, 5, opt1, (int64_t)); \ + FN(16, 16, 16, 4, 4, opt1, (int64_t)); \ + FN(16, 8, 16, 4, 3, opt1, (int64_t)); \ + FN(8, 16, 8, 4, 3, opt1, (int64_t)); \ + FN(8, 8, 8, 3, 3, opt1, (int64_t)); \ + FN(8, 4, 8, 3, 2, opt1, (int64_t)); + +FNS(sse2); + +#undef FNS +#undef FN + +void aom_highbd_upsampled_pred_sse2(uint16_t *comp_pred, int width, int height, + const uint8_t *ref8, int ref_stride) { + int i, j; + int stride = ref_stride << 3; + uint16_t *ref = CONVERT_TO_SHORTPTR(ref8); + + if (width >= 8) { + // read 8 points at one time + for (i = 0; i < height; i++) { + for (j = 0; j < width; j += 8) { + __m128i s0 = _mm_cvtsi32_si128(*(const uint32_t *)ref); + __m128i s1 = _mm_cvtsi32_si128(*(const uint32_t *)(ref + 8)); + __m128i s2 = _mm_cvtsi32_si128(*(const uint32_t *)(ref + 16)); + __m128i s3 = _mm_cvtsi32_si128(*(const uint32_t *)(ref + 24)); + __m128i s4 = _mm_cvtsi32_si128(*(const uint32_t *)(ref + 32)); + __m128i s5 = _mm_cvtsi32_si128(*(const uint32_t *)(ref + 40)); + __m128i s6 = _mm_cvtsi32_si128(*(const uint32_t *)(ref + 48)); + __m128i s7 = _mm_cvtsi32_si128(*(const uint32_t *)(ref + 56)); + __m128i t0, t1, t2, t3; + + t0 = _mm_unpacklo_epi16(s0, s1); + t1 = _mm_unpacklo_epi16(s2, s3); + t2 = _mm_unpacklo_epi16(s4, s5); + t3 = _mm_unpacklo_epi16(s6, s7); + t0 = _mm_unpacklo_epi32(t0, t1); + t2 = _mm_unpacklo_epi32(t2, t3); + t0 = _mm_unpacklo_epi64(t0, t2); + + _mm_storeu_si128((__m128i *)(comp_pred), t0); + comp_pred += 8; + ref += 64; // 8 * 8; + } + ref += stride - (width << 3); + } + } else { + // read 4 points at one time + for (i = 0; i < height; i++) { + for (j = 0; j < width; j += 4) { + __m128i s0 = _mm_cvtsi32_si128(*(const uint32_t *)ref); + __m128i s1 = _mm_cvtsi32_si128(*(const uint32_t *)(ref + 8)); + __m128i s2 = _mm_cvtsi32_si128(*(const uint32_t *)(ref + 16)); + __m128i s3 = _mm_cvtsi32_si128(*(const uint32_t *)(ref + 24)); + __m128i t0, t1; + + t0 = _mm_unpacklo_epi16(s0, s1); + t1 = _mm_unpacklo_epi16(s2, s3); + t0 = _mm_unpacklo_epi32(t0, t1); + + _mm_storel_epi64((__m128i *)(comp_pred), t0); + comp_pred += 4; + ref += 4 * 8; + } + ref += stride - (width << 3); + } + } +} + +void aom_highbd_comp_avg_upsampled_pred_sse2(uint16_t *comp_pred, + const uint8_t *pred8, int width, + int height, const uint8_t *ref8, + int ref_stride) { + const __m128i one = _mm_set1_epi16(1); + int i, j; + int stride = ref_stride << 3; + uint16_t *pred = CONVERT_TO_SHORTPTR(pred8); + uint16_t *ref = CONVERT_TO_SHORTPTR(ref8); + + if (width >= 8) { + // read 8 points at one time + for (i = 0; i < height; i++) { + for (j = 0; j < width; j += 8) { + __m128i s0 = _mm_cvtsi32_si128(*(const uint32_t *)ref); + __m128i s1 = _mm_cvtsi32_si128(*(const uint32_t *)(ref + 8)); + __m128i s2 = _mm_cvtsi32_si128(*(const uint32_t *)(ref + 16)); + __m128i s3 = _mm_cvtsi32_si128(*(const uint32_t *)(ref + 24)); + __m128i s4 = _mm_cvtsi32_si128(*(const uint32_t *)(ref + 32)); + __m128i s5 = _mm_cvtsi32_si128(*(const uint32_t *)(ref + 40)); + __m128i s6 = _mm_cvtsi32_si128(*(const uint32_t *)(ref + 48)); + __m128i s7 = _mm_cvtsi32_si128(*(const uint32_t *)(ref + 56)); + __m128i p0 = _mm_loadu_si128((const __m128i *)pred); + __m128i t0, t1, t2, t3; + + t0 = _mm_unpacklo_epi16(s0, s1); + t1 = _mm_unpacklo_epi16(s2, s3); + t2 = _mm_unpacklo_epi16(s4, s5); + t3 = _mm_unpacklo_epi16(s6, s7); + t0 = _mm_unpacklo_epi32(t0, t1); + t2 = _mm_unpacklo_epi32(t2, t3); + t0 = _mm_unpacklo_epi64(t0, t2); + + p0 = _mm_adds_epu16(t0, p0); + p0 = _mm_adds_epu16(p0, one); + p0 = _mm_srli_epi16(p0, 1); + + _mm_storeu_si128((__m128i *)(comp_pred), p0); + comp_pred += 8; + pred += 8; + ref += 8 * 8; + } + ref += stride - (width << 3); + } + } else { + // read 4 points at one time + for (i = 0; i < height; i++) { + for (j = 0; j < width; j += 4) { + __m128i s0 = _mm_cvtsi32_si128(*(const uint32_t *)ref); + __m128i s1 = _mm_cvtsi32_si128(*(const uint32_t *)(ref + 8)); + __m128i s2 = _mm_cvtsi32_si128(*(const uint32_t *)(ref + 16)); + __m128i s3 = _mm_cvtsi32_si128(*(const uint32_t *)(ref + 24)); + __m128i p0 = _mm_loadl_epi64((const __m128i *)pred); + __m128i t0, t1; + + t0 = _mm_unpacklo_epi16(s0, s1); + t1 = _mm_unpacklo_epi16(s2, s3); + t0 = _mm_unpacklo_epi32(t0, t1); + + p0 = _mm_adds_epu16(t0, p0); + p0 = _mm_adds_epu16(p0, one); + p0 = _mm_srli_epi16(p0, 1); + + _mm_storel_epi64((__m128i *)(comp_pred), p0); + comp_pred += 4; + pred += 4; + ref += 4 * 8; + } + ref += stride - (width << 3); + } + } +} diff --git a/third_party/aom/aom_dsp/x86/highbd_variance_sse4.c b/third_party/aom/aom_dsp/x86/highbd_variance_sse4.c new file mode 100644 index 000000000..cc7f52811 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/highbd_variance_sse4.c @@ -0,0 +1,216 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include <smmintrin.h> /* SSE4.1 */ + +#include "./aom_config.h" +#include "./aom_dsp_rtcd.h" + +#include "aom_dsp/variance.h" +#include "aom_dsp/aom_filter.h" + +static INLINE void variance4x4_64_sse4_1(const uint8_t *a8, int a_stride, + const uint8_t *b8, int b_stride, + uint64_t *sse, int64_t *sum) { + __m128i u0, u1, u2, u3; + __m128i s0, s1, s2, s3; + __m128i t0, t1, x0, y0; + __m128i a0, a1, a2, a3; + __m128i b0, b1, b2, b3; + __m128i k_one_epi16 = _mm_set1_epi16((int16_t)1); + + uint16_t *a = CONVERT_TO_SHORTPTR(a8); + uint16_t *b = CONVERT_TO_SHORTPTR(b8); + + a0 = _mm_loadl_epi64((__m128i const *)(a + 0 * a_stride)); + a1 = _mm_loadl_epi64((__m128i const *)(a + 1 * a_stride)); + a2 = _mm_loadl_epi64((__m128i const *)(a + 2 * a_stride)); + a3 = _mm_loadl_epi64((__m128i const *)(a + 3 * a_stride)); + + b0 = _mm_loadl_epi64((__m128i const *)(b + 0 * b_stride)); + b1 = _mm_loadl_epi64((__m128i const *)(b + 1 * b_stride)); + b2 = _mm_loadl_epi64((__m128i const *)(b + 2 * b_stride)); + b3 = _mm_loadl_epi64((__m128i const *)(b + 3 * b_stride)); + + u0 = _mm_unpacklo_epi16(a0, a1); + u1 = _mm_unpacklo_epi16(a2, a3); + u2 = _mm_unpacklo_epi16(b0, b1); + u3 = _mm_unpacklo_epi16(b2, b3); + + s0 = _mm_sub_epi16(u0, u2); + s1 = _mm_sub_epi16(u1, u3); + + t0 = _mm_madd_epi16(s0, k_one_epi16); + t1 = _mm_madd_epi16(s1, k_one_epi16); + + s2 = _mm_hadd_epi32(t0, t1); + s3 = _mm_hadd_epi32(s2, s2); + y0 = _mm_hadd_epi32(s3, s3); + + t0 = _mm_madd_epi16(s0, s0); + t1 = _mm_madd_epi16(s1, s1); + + s2 = _mm_hadd_epi32(t0, t1); + s3 = _mm_hadd_epi32(s2, s2); + x0 = _mm_hadd_epi32(s3, s3); + + *sse = (uint64_t)_mm_extract_epi32(x0, 0); + *sum = (int64_t)_mm_extract_epi32(y0, 0); +} + +uint32_t aom_highbd_8_variance4x4_sse4_1(const uint8_t *a, int a_stride, + const uint8_t *b, int b_stride, + uint32_t *sse) { + int64_t sum, diff; + uint64_t local_sse; + + variance4x4_64_sse4_1(a, a_stride, b, b_stride, &local_sse, &sum); + *sse = (uint32_t)local_sse; + + diff = (int64_t)*sse - ((sum * sum) >> 4); + return (diff >= 0) ? (uint32_t)diff : 0; +} + +uint32_t aom_highbd_10_variance4x4_sse4_1(const uint8_t *a, int a_stride, + const uint8_t *b, int b_stride, + uint32_t *sse) { + int64_t sum, diff; + uint64_t local_sse; + + variance4x4_64_sse4_1(a, a_stride, b, b_stride, &local_sse, &sum); + *sse = (uint32_t)ROUND_POWER_OF_TWO(local_sse, 4); + sum = ROUND_POWER_OF_TWO(sum, 2); + + diff = (int64_t)*sse - ((sum * sum) >> 4); + return (diff >= 0) ? (uint32_t)diff : 0; +} + +uint32_t aom_highbd_12_variance4x4_sse4_1(const uint8_t *a, int a_stride, + const uint8_t *b, int b_stride, + uint32_t *sse) { + int64_t sum, diff; + uint64_t local_sse; + + variance4x4_64_sse4_1(a, a_stride, b, b_stride, &local_sse, &sum); + *sse = (uint32_t)ROUND_POWER_OF_TWO(local_sse, 8); + sum = ROUND_POWER_OF_TWO(sum, 4); + + diff = (int64_t)*sse - ((sum * sum) >> 4); + return diff >= 0 ? (uint32_t)diff : 0; +} + +// Sub-pixel +uint32_t aom_highbd_8_sub_pixel_variance4x4_sse4_1( + const uint8_t *src, int src_stride, int xoffset, int yoffset, + const uint8_t *dst, int dst_stride, uint32_t *sse) { + uint16_t fdata3[(4 + 1) * 4]; + uint16_t temp2[4 * 4]; + + aom_highbd_var_filter_block2d_bil_first_pass( + src, fdata3, src_stride, 1, 4 + 1, 4, bilinear_filters_2t[xoffset]); + aom_highbd_var_filter_block2d_bil_second_pass(fdata3, temp2, 4, 4, 4, 4, + bilinear_filters_2t[yoffset]); + + return aom_highbd_8_variance4x4(CONVERT_TO_BYTEPTR(temp2), 4, dst, dst_stride, + sse); +} + +uint32_t aom_highbd_10_sub_pixel_variance4x4_sse4_1( + const uint8_t *src, int src_stride, int xoffset, int yoffset, + const uint8_t *dst, int dst_stride, uint32_t *sse) { + uint16_t fdata3[(4 + 1) * 4]; + uint16_t temp2[4 * 4]; + + aom_highbd_var_filter_block2d_bil_first_pass( + src, fdata3, src_stride, 1, 4 + 1, 4, bilinear_filters_2t[xoffset]); + aom_highbd_var_filter_block2d_bil_second_pass(fdata3, temp2, 4, 4, 4, 4, + bilinear_filters_2t[yoffset]); + + return aom_highbd_10_variance4x4(CONVERT_TO_BYTEPTR(temp2), 4, dst, + dst_stride, sse); +} + +uint32_t aom_highbd_12_sub_pixel_variance4x4_sse4_1( + const uint8_t *src, int src_stride, int xoffset, int yoffset, + const uint8_t *dst, int dst_stride, uint32_t *sse) { + uint16_t fdata3[(4 + 1) * 4]; + uint16_t temp2[4 * 4]; + + aom_highbd_var_filter_block2d_bil_first_pass( + src, fdata3, src_stride, 1, 4 + 1, 4, bilinear_filters_2t[xoffset]); + aom_highbd_var_filter_block2d_bil_second_pass(fdata3, temp2, 4, 4, 4, 4, + bilinear_filters_2t[yoffset]); + + return aom_highbd_12_variance4x4(CONVERT_TO_BYTEPTR(temp2), 4, dst, + dst_stride, sse); +} + +// Sub-pixel average + +uint32_t aom_highbd_8_sub_pixel_avg_variance4x4_sse4_1( + const uint8_t *src, int src_stride, int xoffset, int yoffset, + const uint8_t *dst, int dst_stride, uint32_t *sse, + const uint8_t *second_pred) { + uint16_t fdata3[(4 + 1) * 4]; + uint16_t temp2[4 * 4]; + DECLARE_ALIGNED(16, uint16_t, temp3[4 * 4]); + + aom_highbd_var_filter_block2d_bil_first_pass( + src, fdata3, src_stride, 1, 4 + 1, 4, bilinear_filters_2t[xoffset]); + aom_highbd_var_filter_block2d_bil_second_pass(fdata3, temp2, 4, 4, 4, 4, + bilinear_filters_2t[yoffset]); + + aom_highbd_comp_avg_pred(temp3, second_pred, 4, 4, CONVERT_TO_BYTEPTR(temp2), + 4); + + return aom_highbd_8_variance4x4(CONVERT_TO_BYTEPTR(temp3), 4, dst, dst_stride, + sse); +} + +uint32_t aom_highbd_10_sub_pixel_avg_variance4x4_sse4_1( + const uint8_t *src, int src_stride, int xoffset, int yoffset, + const uint8_t *dst, int dst_stride, uint32_t *sse, + const uint8_t *second_pred) { + uint16_t fdata3[(4 + 1) * 4]; + uint16_t temp2[4 * 4]; + DECLARE_ALIGNED(16, uint16_t, temp3[4 * 4]); + + aom_highbd_var_filter_block2d_bil_first_pass( + src, fdata3, src_stride, 1, 4 + 1, 4, bilinear_filters_2t[xoffset]); + aom_highbd_var_filter_block2d_bil_second_pass(fdata3, temp2, 4, 4, 4, 4, + bilinear_filters_2t[yoffset]); + + aom_highbd_comp_avg_pred(temp3, second_pred, 4, 4, CONVERT_TO_BYTEPTR(temp2), + 4); + + return aom_highbd_10_variance4x4(CONVERT_TO_BYTEPTR(temp3), 4, dst, + dst_stride, sse); +} + +uint32_t aom_highbd_12_sub_pixel_avg_variance4x4_sse4_1( + const uint8_t *src, int src_stride, int xoffset, int yoffset, + const uint8_t *dst, int dst_stride, uint32_t *sse, + const uint8_t *second_pred) { + uint16_t fdata3[(4 + 1) * 4]; + uint16_t temp2[4 * 4]; + DECLARE_ALIGNED(16, uint16_t, temp3[4 * 4]); + + aom_highbd_var_filter_block2d_bil_first_pass( + src, fdata3, src_stride, 1, 4 + 1, 4, bilinear_filters_2t[xoffset]); + aom_highbd_var_filter_block2d_bil_second_pass(fdata3, temp2, 4, 4, 4, 4, + bilinear_filters_2t[yoffset]); + + aom_highbd_comp_avg_pred(temp3, second_pred, 4, 4, CONVERT_TO_BYTEPTR(temp2), + 4); + + return aom_highbd_12_variance4x4(CONVERT_TO_BYTEPTR(temp3), 4, dst, + dst_stride, sse); +} diff --git a/third_party/aom/aom_dsp/x86/intrapred_sse2.asm b/third_party/aom/aom_dsp/x86/intrapred_sse2.asm new file mode 100644 index 000000000..02567db49 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/intrapred_sse2.asm @@ -0,0 +1,771 @@ +; +; Copyright (c) 2016, Alliance for Open Media. All rights reserved +; +; This source code is subject to the terms of the BSD 2 Clause License and +; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License +; was not distributed with this source code in the LICENSE file, you can +; obtain it at www.aomedia.org/license/software. If the Alliance for Open +; Media Patent License 1.0 was not distributed with this source code in the +; PATENTS file, you can obtain it at www.aomedia.org/license/patent. +; + +; + +%include "third_party/x86inc/x86inc.asm" + +SECTION_RODATA +pb_1: times 16 db 1 +pw_4: times 8 dw 4 +pw_8: times 8 dw 8 +pw_16: times 8 dw 16 +pw_32: times 8 dw 32 +dc_128: times 16 db 128 +pw2_4: times 8 dw 2 +pw2_8: times 8 dw 4 +pw2_16: times 8 dw 8 +pw2_32: times 8 dw 16 + +SECTION .text + +; ------------------------------------------ +; input: x, y, z, result +; +; trick from pascal +; (x+2y+z+2)>>2 can be calculated as: +; result = avg(x,z) +; result -= xor(x,z) & 1 +; result = avg(result,y) +; ------------------------------------------ +%macro X_PLUS_2Y_PLUS_Z_PLUS_2_RSH_2 4 + pavgb %4, %1, %3 + pxor %3, %1 + pand %3, [GLOBAL(pb_1)] + psubb %4, %3 + pavgb %4, %2 +%endmacro + +INIT_XMM sse2 +cglobal dc_predictor_4x4, 4, 5, 3, dst, stride, above, left, goffset + GET_GOT goffsetq + + movd m2, [leftq] + movd m0, [aboveq] + pxor m1, m1 + punpckldq m0, m2 + psadbw m0, m1 + paddw m0, [GLOBAL(pw_4)] + psraw m0, 3 + pshuflw m0, m0, 0x0 + packuswb m0, m0 + movd [dstq ], m0 + movd [dstq+strideq], m0 + lea dstq, [dstq+strideq*2] + movd [dstq ], m0 + movd [dstq+strideq], m0 + + RESTORE_GOT + RET + +INIT_XMM sse2 +cglobal dc_left_predictor_4x4, 2, 5, 2, dst, stride, above, left, goffset + movifnidn leftq, leftmp + GET_GOT goffsetq + + pxor m1, m1 + movd m0, [leftq] + psadbw m0, m1 + paddw m0, [GLOBAL(pw2_4)] + psraw m0, 2 + pshuflw m0, m0, 0x0 + packuswb m0, m0 + movd [dstq ], m0 + movd [dstq+strideq], m0 + lea dstq, [dstq+strideq*2] + movd [dstq ], m0 + movd [dstq+strideq], m0 + + RESTORE_GOT + RET + +INIT_XMM sse2 +cglobal dc_top_predictor_4x4, 3, 5, 2, dst, stride, above, left, goffset + GET_GOT goffsetq + + pxor m1, m1 + movd m0, [aboveq] + psadbw m0, m1 + paddw m0, [GLOBAL(pw2_4)] + psraw m0, 2 + pshuflw m0, m0, 0x0 + packuswb m0, m0 + movd [dstq ], m0 + movd [dstq+strideq], m0 + lea dstq, [dstq+strideq*2] + movd [dstq ], m0 + movd [dstq+strideq], m0 + + RESTORE_GOT + RET + +INIT_XMM sse2 +cglobal dc_predictor_8x8, 4, 5, 3, dst, stride, above, left, goffset + GET_GOT goffsetq + + pxor m1, m1 + movq m0, [aboveq] + movq m2, [leftq] + DEFINE_ARGS dst, stride, stride3 + lea stride3q, [strideq*3] + psadbw m0, m1 + psadbw m2, m1 + paddw m0, m2 + paddw m0, [GLOBAL(pw_8)] + psraw m0, 4 + punpcklbw m0, m0 + pshuflw m0, m0, 0x0 + movq [dstq ], m0 + movq [dstq+strideq ], m0 + movq [dstq+strideq*2], m0 + movq [dstq+stride3q ], m0 + lea dstq, [dstq+strideq*4] + movq [dstq ], m0 + movq [dstq+strideq ], m0 + movq [dstq+strideq*2], m0 + movq [dstq+stride3q ], m0 + + RESTORE_GOT + RET + +INIT_XMM sse2 +cglobal dc_top_predictor_8x8, 3, 5, 2, dst, stride, above, left, goffset + GET_GOT goffsetq + + pxor m1, m1 + movq m0, [aboveq] + DEFINE_ARGS dst, stride, stride3 + lea stride3q, [strideq*3] + psadbw m0, m1 + paddw m0, [GLOBAL(pw2_8)] + psraw m0, 3 + punpcklbw m0, m0 + pshuflw m0, m0, 0x0 + movq [dstq ], m0 + movq [dstq+strideq ], m0 + movq [dstq+strideq*2], m0 + movq [dstq+stride3q ], m0 + lea dstq, [dstq+strideq*4] + movq [dstq ], m0 + movq [dstq+strideq ], m0 + movq [dstq+strideq*2], m0 + movq [dstq+stride3q ], m0 + + RESTORE_GOT + RET + +INIT_XMM sse2 +cglobal dc_left_predictor_8x8, 2, 5, 2, dst, stride, above, left, goffset + movifnidn leftq, leftmp + GET_GOT goffsetq + + pxor m1, m1 + movq m0, [leftq] + DEFINE_ARGS dst, stride, stride3 + lea stride3q, [strideq*3] + psadbw m0, m1 + paddw m0, [GLOBAL(pw2_8)] + psraw m0, 3 + punpcklbw m0, m0 + pshuflw m0, m0, 0x0 + movq [dstq ], m0 + movq [dstq+strideq ], m0 + movq [dstq+strideq*2], m0 + movq [dstq+stride3q ], m0 + lea dstq, [dstq+strideq*4] + movq [dstq ], m0 + movq [dstq+strideq ], m0 + movq [dstq+strideq*2], m0 + movq [dstq+stride3q ], m0 + + RESTORE_GOT + RET + +INIT_XMM sse2 +cglobal dc_128_predictor_4x4, 2, 5, 1, dst, stride, above, left, goffset + GET_GOT goffsetq + + DEFINE_ARGS dst, stride, stride3 + lea stride3q, [strideq*3] + movd m0, [GLOBAL(dc_128)] + movd [dstq ], m0 + movd [dstq+strideq ], m0 + movd [dstq+strideq*2], m0 + movd [dstq+stride3q ], m0 + RESTORE_GOT + RET + +INIT_XMM sse2 +cglobal dc_128_predictor_8x8, 2, 5, 1, dst, stride, above, left, goffset + GET_GOT goffsetq + + DEFINE_ARGS dst, stride, stride3 + lea stride3q, [strideq*3] + movq m0, [GLOBAL(dc_128)] + movq [dstq ], m0 + movq [dstq+strideq ], m0 + movq [dstq+strideq*2], m0 + movq [dstq+stride3q ], m0 + lea dstq, [dstq+strideq*4] + movq [dstq ], m0 + movq [dstq+strideq ], m0 + movq [dstq+strideq*2], m0 + movq [dstq+stride3q ], m0 + RESTORE_GOT + RET + +INIT_XMM sse2 +cglobal dc_predictor_16x16, 4, 5, 3, dst, stride, above, left, goffset + GET_GOT goffsetq + + pxor m1, m1 + mova m0, [aboveq] + mova m2, [leftq] + DEFINE_ARGS dst, stride, stride3, lines4 + lea stride3q, [strideq*3] + mov lines4d, 4 + psadbw m0, m1 + psadbw m2, m1 + paddw m0, m2 + movhlps m2, m0 + paddw m0, m2 + paddw m0, [GLOBAL(pw_16)] + psraw m0, 5 + pshuflw m0, m0, 0x0 + punpcklqdq m0, m0 + packuswb m0, m0 +.loop: + mova [dstq ], m0 + mova [dstq+strideq ], m0 + mova [dstq+strideq*2], m0 + mova [dstq+stride3q ], m0 + lea dstq, [dstq+strideq*4] + dec lines4d + jnz .loop + + RESTORE_GOT + REP_RET + + +INIT_XMM sse2 +cglobal dc_top_predictor_16x16, 4, 5, 3, dst, stride, above, left, goffset + GET_GOT goffsetq + + pxor m1, m1 + mova m0, [aboveq] + DEFINE_ARGS dst, stride, stride3, lines4 + lea stride3q, [strideq*3] + mov lines4d, 4 + psadbw m0, m1 + movhlps m2, m0 + paddw m0, m2 + paddw m0, [GLOBAL(pw2_16)] + psraw m0, 4 + pshuflw m0, m0, 0x0 + punpcklqdq m0, m0 + packuswb m0, m0 +.loop: + mova [dstq ], m0 + mova [dstq+strideq ], m0 + mova [dstq+strideq*2], m0 + mova [dstq+stride3q ], m0 + lea dstq, [dstq+strideq*4] + dec lines4d + jnz .loop + + RESTORE_GOT + REP_RET + +INIT_XMM sse2 +cglobal dc_left_predictor_16x16, 4, 5, 3, dst, stride, above, left, goffset + GET_GOT goffsetq + + pxor m1, m1 + mova m0, [leftq] + DEFINE_ARGS dst, stride, stride3, lines4 + lea stride3q, [strideq*3] + mov lines4d, 4 + psadbw m0, m1 + movhlps m2, m0 + paddw m0, m2 + paddw m0, [GLOBAL(pw2_16)] + psraw m0, 4 + pshuflw m0, m0, 0x0 + punpcklqdq m0, m0 + packuswb m0, m0 +.loop: + mova [dstq ], m0 + mova [dstq+strideq ], m0 + mova [dstq+strideq*2], m0 + mova [dstq+stride3q ], m0 + lea dstq, [dstq+strideq*4] + dec lines4d + jnz .loop + + RESTORE_GOT + REP_RET + +INIT_XMM sse2 +cglobal dc_128_predictor_16x16, 4, 5, 3, dst, stride, above, left, goffset + GET_GOT goffsetq + + DEFINE_ARGS dst, stride, stride3, lines4 + lea stride3q, [strideq*3] + mov lines4d, 4 + mova m0, [GLOBAL(dc_128)] +.loop: + mova [dstq ], m0 + mova [dstq+strideq ], m0 + mova [dstq+strideq*2], m0 + mova [dstq+stride3q ], m0 + lea dstq, [dstq+strideq*4] + dec lines4d + jnz .loop + RESTORE_GOT + RET + + +INIT_XMM sse2 +cglobal dc_predictor_32x32, 4, 5, 5, dst, stride, above, left, goffset + GET_GOT goffsetq + + pxor m1, m1 + mova m0, [aboveq] + mova m2, [aboveq+16] + mova m3, [leftq] + mova m4, [leftq+16] + DEFINE_ARGS dst, stride, stride3, lines4 + lea stride3q, [strideq*3] + mov lines4d, 8 + psadbw m0, m1 + psadbw m2, m1 + psadbw m3, m1 + psadbw m4, m1 + paddw m0, m2 + paddw m0, m3 + paddw m0, m4 + movhlps m2, m0 + paddw m0, m2 + paddw m0, [GLOBAL(pw_32)] + psraw m0, 6 + pshuflw m0, m0, 0x0 + punpcklqdq m0, m0 + packuswb m0, m0 +.loop: + mova [dstq ], m0 + mova [dstq +16], m0 + mova [dstq+strideq ], m0 + mova [dstq+strideq +16], m0 + mova [dstq+strideq*2 ], m0 + mova [dstq+strideq*2+16], m0 + mova [dstq+stride3q ], m0 + mova [dstq+stride3q +16], m0 + lea dstq, [dstq+strideq*4] + dec lines4d + jnz .loop + + RESTORE_GOT + REP_RET + +INIT_XMM sse2 +cglobal dc_top_predictor_32x32, 4, 5, 5, dst, stride, above, left, goffset + GET_GOT goffsetq + + pxor m1, m1 + mova m0, [aboveq] + mova m2, [aboveq+16] + DEFINE_ARGS dst, stride, stride3, lines4 + lea stride3q, [strideq*3] + mov lines4d, 8 + psadbw m0, m1 + psadbw m2, m1 + paddw m0, m2 + movhlps m2, m0 + paddw m0, m2 + paddw m0, [GLOBAL(pw2_32)] + psraw m0, 5 + pshuflw m0, m0, 0x0 + punpcklqdq m0, m0 + packuswb m0, m0 +.loop: + mova [dstq ], m0 + mova [dstq +16], m0 + mova [dstq+strideq ], m0 + mova [dstq+strideq +16], m0 + mova [dstq+strideq*2 ], m0 + mova [dstq+strideq*2+16], m0 + mova [dstq+stride3q ], m0 + mova [dstq+stride3q +16], m0 + lea dstq, [dstq+strideq*4] + dec lines4d + jnz .loop + + RESTORE_GOT + REP_RET + +INIT_XMM sse2 +cglobal dc_left_predictor_32x32, 4, 5, 5, dst, stride, above, left, goffset + GET_GOT goffsetq + + pxor m1, m1 + mova m0, [leftq] + mova m2, [leftq+16] + DEFINE_ARGS dst, stride, stride3, lines4 + lea stride3q, [strideq*3] + mov lines4d, 8 + psadbw m0, m1 + psadbw m2, m1 + paddw m0, m2 + movhlps m2, m0 + paddw m0, m2 + paddw m0, [GLOBAL(pw2_32)] + psraw m0, 5 + pshuflw m0, m0, 0x0 + punpcklqdq m0, m0 + packuswb m0, m0 +.loop: + mova [dstq ], m0 + mova [dstq +16], m0 + mova [dstq+strideq ], m0 + mova [dstq+strideq +16], m0 + mova [dstq+strideq*2 ], m0 + mova [dstq+strideq*2+16], m0 + mova [dstq+stride3q ], m0 + mova [dstq+stride3q +16], m0 + lea dstq, [dstq+strideq*4] + dec lines4d + jnz .loop + + RESTORE_GOT + REP_RET + +INIT_XMM sse2 +cglobal dc_128_predictor_32x32, 4, 5, 3, dst, stride, above, left, goffset + GET_GOT goffsetq + + DEFINE_ARGS dst, stride, stride3, lines4 + lea stride3q, [strideq*3] + mov lines4d, 8 + mova m0, [GLOBAL(dc_128)] +.loop: + mova [dstq ], m0 + mova [dstq +16], m0 + mova [dstq+strideq ], m0 + mova [dstq+strideq +16], m0 + mova [dstq+strideq*2 ], m0 + mova [dstq+strideq*2+16], m0 + mova [dstq+stride3q ], m0 + mova [dstq+stride3q +16], m0 + lea dstq, [dstq+strideq*4] + dec lines4d + jnz .loop + RESTORE_GOT + RET + +INIT_XMM sse2 +cglobal v_predictor_4x4, 3, 3, 1, dst, stride, above + movd m0, [aboveq] + movd [dstq ], m0 + movd [dstq+strideq], m0 + lea dstq, [dstq+strideq*2] + movd [dstq ], m0 + movd [dstq+strideq], m0 + RET + +INIT_XMM sse2 +cglobal v_predictor_8x8, 3, 3, 1, dst, stride, above + movq m0, [aboveq] + DEFINE_ARGS dst, stride, stride3 + lea stride3q, [strideq*3] + movq [dstq ], m0 + movq [dstq+strideq ], m0 + movq [dstq+strideq*2], m0 + movq [dstq+stride3q ], m0 + lea dstq, [dstq+strideq*4] + movq [dstq ], m0 + movq [dstq+strideq ], m0 + movq [dstq+strideq*2], m0 + movq [dstq+stride3q ], m0 + RET + +INIT_XMM sse2 +cglobal v_predictor_16x16, 3, 4, 1, dst, stride, above + mova m0, [aboveq] + DEFINE_ARGS dst, stride, stride3, nlines4 + lea stride3q, [strideq*3] + mov nlines4d, 4 +.loop: + mova [dstq ], m0 + mova [dstq+strideq ], m0 + mova [dstq+strideq*2], m0 + mova [dstq+stride3q ], m0 + lea dstq, [dstq+strideq*4] + dec nlines4d + jnz .loop + REP_RET + +INIT_XMM sse2 +cglobal v_predictor_32x32, 3, 4, 2, dst, stride, above + mova m0, [aboveq] + mova m1, [aboveq+16] + DEFINE_ARGS dst, stride, stride3, nlines4 + lea stride3q, [strideq*3] + mov nlines4d, 8 +.loop: + mova [dstq ], m0 + mova [dstq +16], m1 + mova [dstq+strideq ], m0 + mova [dstq+strideq +16], m1 + mova [dstq+strideq*2 ], m0 + mova [dstq+strideq*2+16], m1 + mova [dstq+stride3q ], m0 + mova [dstq+stride3q +16], m1 + lea dstq, [dstq+strideq*4] + dec nlines4d + jnz .loop + REP_RET + +INIT_XMM sse2 +cglobal h_predictor_4x4, 2, 4, 4, dst, stride, line, left + movifnidn leftq, leftmp + movd m0, [leftq] + punpcklbw m0, m0 + punpcklbw m0, m0 + pshufd m1, m0, 0x1 + movd [dstq ], m0 + movd [dstq+strideq], m1 + pshufd m2, m0, 0x2 + lea dstq, [dstq+strideq*2] + pshufd m3, m0, 0x3 + movd [dstq ], m2 + movd [dstq+strideq], m3 + RET + +INIT_XMM sse2 +cglobal h_predictor_8x8, 2, 5, 3, dst, stride, line, left + movifnidn leftq, leftmp + mov lineq, -2 + DEFINE_ARGS dst, stride, line, left, stride3 + lea stride3q, [strideq*3] + movq m0, [leftq ] + punpcklbw m0, m0 ; l1 l1 l2 l2 ... l8 l8 +.loop: + pshuflw m1, m0, 0x0 ; l1 l1 l1 l1 l1 l1 l1 l1 + pshuflw m2, m0, 0x55 ; l2 l2 l2 l2 l2 l2 l2 l2 + movq [dstq ], m1 + movq [dstq+strideq], m2 + pshuflw m1, m0, 0xaa + pshuflw m2, m0, 0xff + movq [dstq+strideq*2], m1 + movq [dstq+stride3q ], m2 + pshufd m0, m0, 0xe ; [63:0] l5 l5 l6 l6 l7 l7 l8 l8 + inc lineq + lea dstq, [dstq+strideq*4] + jnz .loop + REP_RET + +INIT_XMM sse2 +cglobal h_predictor_16x16, 2, 5, 3, dst, stride, line, left + movifnidn leftq, leftmp + mov lineq, -4 + DEFINE_ARGS dst, stride, line, left, stride3 + lea stride3q, [strideq*3] +.loop: + movd m0, [leftq] + punpcklbw m0, m0 + punpcklbw m0, m0 ; l1 to l4 each repeated 4 times + pshufd m1, m0, 0x0 ; l1 repeated 16 times + pshufd m2, m0, 0x55 ; l2 repeated 16 times + mova [dstq ], m1 + mova [dstq+strideq ], m2 + pshufd m1, m0, 0xaa + pshufd m2, m0, 0xff + mova [dstq+strideq*2], m1 + mova [dstq+stride3q ], m2 + inc lineq + lea leftq, [leftq+4 ] + lea dstq, [dstq+strideq*4] + jnz .loop + REP_RET + +INIT_XMM sse2 +cglobal h_predictor_32x32, 2, 5, 3, dst, stride, line, left + movifnidn leftq, leftmp + mov lineq, -8 + DEFINE_ARGS dst, stride, line, left, stride3 + lea stride3q, [strideq*3] +.loop: + movd m0, [leftq] + punpcklbw m0, m0 + punpcklbw m0, m0 ; l1 to l4 each repeated 4 times + pshufd m1, m0, 0x0 ; l1 repeated 16 times + pshufd m2, m0, 0x55 ; l2 repeated 16 times + mova [dstq ], m1 + mova [dstq+16 ], m1 + mova [dstq+strideq ], m2 + mova [dstq+strideq+16 ], m2 + pshufd m1, m0, 0xaa + pshufd m2, m0, 0xff + mova [dstq+strideq*2 ], m1 + mova [dstq+strideq*2+16], m1 + mova [dstq+stride3q ], m2 + mova [dstq+stride3q+16 ], m2 + inc lineq + lea leftq, [leftq+4 ] + lea dstq, [dstq+strideq*4] + jnz .loop + REP_RET + +INIT_XMM sse2 +cglobal tm_predictor_4x4, 4, 4, 5, dst, stride, above, left + pxor m1, m1 + movq m0, [aboveq-1]; [63:0] tl t1 t2 t3 t4 x x x + punpcklbw m0, m1 + pshuflw m2, m0, 0x0 ; [63:0] tl tl tl tl [word] + psrldq m0, 2 + psubw m0, m2 ; [63:0] t1-tl t2-tl t3-tl t4-tl [word] + movd m2, [leftq] + punpcklbw m2, m1 + pshuflw m4, m2, 0x0 ; [63:0] l1 l1 l1 l1 [word] + pshuflw m3, m2, 0x55 ; [63:0] l2 l2 l2 l2 [word] + paddw m4, m0 + paddw m3, m0 + packuswb m4, m4 + packuswb m3, m3 + movd [dstq ], m4 + movd [dstq+strideq], m3 + lea dstq, [dstq+strideq*2] + pshuflw m4, m2, 0xaa + pshuflw m3, m2, 0xff + paddw m4, m0 + paddw m3, m0 + packuswb m4, m4 + packuswb m3, m3 + movd [dstq ], m4 + movd [dstq+strideq], m3 + RET + +INIT_XMM sse2 +cglobal tm_predictor_8x8, 4, 4, 5, dst, stride, above, left + pxor m1, m1 + movd m2, [aboveq-1] + movq m0, [aboveq] + punpcklbw m2, m1 + punpcklbw m0, m1 ; t1 t2 t3 t4 t5 t6 t7 t8 [word] + pshuflw m2, m2, 0x0 ; [63:0] tl tl tl tl [word] + DEFINE_ARGS dst, stride, line, left + mov lineq, -4 + punpcklqdq m2, m2 ; tl tl tl tl tl tl tl tl [word] + psubw m0, m2 ; t1-tl t2-tl ... t8-tl [word] + movq m2, [leftq] + punpcklbw m2, m1 ; l1 l2 l3 l4 l5 l6 l7 l8 [word] +.loop: + pshuflw m4, m2, 0x0 ; [63:0] l1 l1 l1 l1 [word] + pshuflw m3, m2, 0x55 ; [63:0] l2 l2 l2 l2 [word] + punpcklqdq m4, m4 ; l1 l1 l1 l1 l1 l1 l1 l1 [word] + punpcklqdq m3, m3 ; l2 l2 l2 l2 l2 l2 l2 l2 [word] + paddw m4, m0 + paddw m3, m0 + packuswb m4, m3 + movq [dstq ], m4 + movhps [dstq+strideq], m4 + lea dstq, [dstq+strideq*2] + psrldq m2, 4 + inc lineq + jnz .loop + REP_RET + +INIT_XMM sse2 +cglobal tm_predictor_16x16, 4, 5, 8, dst, stride, above, left + pxor m1, m1 + mova m2, [aboveq-16]; + mova m0, [aboveq] ; t1 t2 ... t16 [byte] + punpckhbw m2, m1 ; [127:112] tl [word] + punpckhbw m4, m0, m1 + punpcklbw m0, m1 ; m0:m4 t1 t2 ... t16 [word] + DEFINE_ARGS dst, stride, line, left, stride8 + mov lineq, -8 + pshufhw m2, m2, 0xff + mova m3, [leftq] ; l1 l2 ... l16 [byte] + punpckhqdq m2, m2 ; tl repeated 8 times [word] + psubw m0, m2 + psubw m4, m2 ; m0:m4 t1-tl t2-tl ... t16-tl [word] + punpckhbw m5, m3, m1 + punpcklbw m3, m1 ; m3:m5 l1 l2 ... l16 [word] + lea stride8q, [strideq*8] +.loop: + pshuflw m6, m3, 0x0 + pshuflw m7, m5, 0x0 + punpcklqdq m6, m6 ; l1 repeated 8 times [word] + punpcklqdq m7, m7 ; l8 repeated 8 times [word] + paddw m1, m6, m0 + paddw m6, m4 ; m1:m6 ti-tl+l1 [i=1,15] [word] + psrldq m5, 2 + packuswb m1, m6 + mova [dstq ], m1 + paddw m1, m7, m0 + paddw m7, m4 ; m1:m7 ti-tl+l8 [i=1,15] [word] + psrldq m3, 2 + packuswb m1, m7 + mova [dstq+stride8q], m1 + inc lineq + lea dstq, [dstq+strideq] + jnz .loop + REP_RET + +INIT_XMM sse2 +cglobal tm_predictor_32x32, 4, 4, 8, dst, stride, above, left + pxor m1, m1 + movd m2, [aboveq-1] + mova m0, [aboveq] + mova m4, [aboveq+16] + punpcklbw m2, m1 + punpckhbw m3, m0, m1 + punpckhbw m5, m4, m1 + punpcklbw m0, m1 + punpcklbw m4, m1 + pshuflw m2, m2, 0x0 + DEFINE_ARGS dst, stride, line, left + mov lineq, -16 + punpcklqdq m2, m2 + add leftq, 32 + psubw m0, m2 + psubw m3, m2 + psubw m4, m2 + psubw m5, m2 +.loop: + movd m2, [leftq+lineq*2] + pxor m1, m1 + punpcklbw m2, m1 + pshuflw m7, m2, 0x55 + pshuflw m2, m2, 0x0 + punpcklqdq m2, m2 + punpcklqdq m7, m7 + paddw m6, m2, m3 + paddw m1, m2, m0 + packuswb m1, m6 + mova [dstq ], m1 + paddw m6, m2, m5 + paddw m1, m2, m4 + packuswb m1, m6 + mova [dstq+16 ], m1 + paddw m6, m7, m3 + paddw m1, m7, m0 + packuswb m1, m6 + mova [dstq+strideq ], m1 + paddw m6, m7, m5 + paddw m1, m7, m4 + packuswb m1, m6 + mova [dstq+strideq+16], m1 + lea dstq, [dstq+strideq*2] + inc lineq + jnz .loop + REP_RET diff --git a/third_party/aom/aom_dsp/x86/intrapred_ssse3.asm b/third_party/aom/aom_dsp/x86/intrapred_ssse3.asm new file mode 100644 index 000000000..bc1bb2ff3 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/intrapred_ssse3.asm @@ -0,0 +1,410 @@ +; +; Copyright (c) 2016, Alliance for Open Media. All rights reserved +; +; This source code is subject to the terms of the BSD 2 Clause License and +; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License +; was not distributed with this source code in the LICENSE file, you can +; obtain it at www.aomedia.org/license/software. If the Alliance for Open +; Media Patent License 1.0 was not distributed with this source code in the +; PATENTS file, you can obtain it at www.aomedia.org/license/patent. +; + +; + +%include "third_party/x86inc/x86inc.asm" + +SECTION_RODATA + +pb_1: times 16 db 1 +sh_b12345677: db 1, 2, 3, 4, 5, 6, 7, 7, 0, 0, 0, 0, 0, 0, 0, 0 +sh_b23456777: db 2, 3, 4, 5, 6, 7, 7, 7, 0, 0, 0, 0, 0, 0, 0, 0 +sh_b0123456777777777: db 0, 1, 2, 3, 4, 5, 6, 7, 7, 7, 7, 7, 7, 7, 7, 7 +sh_b1234567777777777: db 1, 2, 3, 4, 5, 6, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7 +sh_b2345677777777777: db 2, 3, 4, 5, 6, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7 +sh_b123456789abcdeff: db 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 15 +sh_b23456789abcdefff: db 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 15, 15 +sh_b32104567: db 3, 2, 1, 0, 4, 5, 6, 7, 0, 0, 0, 0, 0, 0, 0, 0 +sh_b8091a2b345: db 8, 0, 9, 1, 10, 2, 11, 3, 4, 5, 0, 0, 0, 0, 0, 0 +sh_b76543210: db 7, 6, 5, 4, 3, 2, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0 +sh_b65432108: db 6, 5, 4, 3, 2, 1, 0, 8, 0, 0, 0, 0, 0, 0, 0, 0 +sh_b54321089: db 5, 4, 3, 2, 1, 0, 8, 9, 0, 0, 0, 0, 0, 0, 0, 0 +sh_b89abcdef: db 8, 9, 10, 11, 12, 13, 14, 15, 0, 0, 0, 0, 0, 0, 0, 0 +sh_bfedcba9876543210: db 15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0 + +SECTION .text + +; ------------------------------------------ +; input: x, y, z, result +; +; trick from pascal +; (x+2y+z+2)>>2 can be calculated as: +; result = avg(x,z) +; result -= xor(x,z) & 1 +; result = avg(result,y) +; ------------------------------------------ +%macro X_PLUS_2Y_PLUS_Z_PLUS_2_RSH_2 4 + pavgb %4, %1, %3 + pxor %3, %1 + pand %3, [GLOBAL(pb_1)] + psubb %4, %3 + pavgb %4, %2 +%endmacro + +INIT_XMM ssse3 +cglobal d63e_predictor_4x4, 3, 4, 5, dst, stride, above, goffset + GET_GOT goffsetq + + movq m3, [aboveq] + pshufb m1, m3, [GLOBAL(sh_b23456777)] + pshufb m2, m3, [GLOBAL(sh_b12345677)] + + X_PLUS_2Y_PLUS_Z_PLUS_2_RSH_2 m3, m2, m1, m4 + pavgb m3, m2 + + ; store 4 lines + movd [dstq ], m3 + movd [dstq+strideq], m4 + lea dstq, [dstq+strideq*2] + psrldq m3, 1 + psrldq m4, 1 + movd [dstq ], m3 + movd [dstq+strideq], m4 + RESTORE_GOT + RET + +INIT_XMM ssse3 +cglobal d153_predictor_4x4, 4, 5, 4, dst, stride, above, left, goffset + GET_GOT goffsetq + movd m0, [leftq] ; l1, l2, l3, l4 + movd m1, [aboveq-1] ; tl, t1, t2, t3 + punpckldq m0, m1 ; l1, l2, l3, l4, tl, t1, t2, t3 + pshufb m0, [GLOBAL(sh_b32104567)]; l4, l3, l2, l1, tl, t1, t2, t3 + psrldq m1, m0, 1 ; l3, l2, l1, tl, t1, t2, t3 + psrldq m2, m0, 2 ; l2, l1, tl, t1, t2, t3 + ; comments below are for a predictor like this + ; A1 B1 C1 D1 + ; A2 B2 A1 B1 + ; A3 B3 A2 B2 + ; A4 B4 A3 B3 + X_PLUS_2Y_PLUS_Z_PLUS_2_RSH_2 m0, m1, m2, m3 ; 3-tap avg B4 B3 B2 B1 C1 D1 + pavgb m1, m0 ; 2-tap avg A4 A3 A2 A1 + + punpcklqdq m3, m1 ; B4 B3 B2 B1 C1 D1 x x A4 A3 A2 A1 .. + + DEFINE_ARGS dst, stride, stride3 + lea stride3q, [strideq*3] + pshufb m3, [GLOBAL(sh_b8091a2b345)] ; A4 B4 A3 B3 A2 B2 A1 B1 C1 D1 .. + movd [dstq+stride3q ], m3 + psrldq m3, 2 ; A3 B3 A2 B2 A1 B1 C1 D1 .. + movd [dstq+strideq*2], m3 + psrldq m3, 2 ; A2 B2 A1 B1 C1 D1 .. + movd [dstq+strideq ], m3 + psrldq m3, 2 ; A1 B1 C1 D1 .. + movd [dstq ], m3 + RESTORE_GOT + RET + +INIT_XMM ssse3 +cglobal d153_predictor_8x8, 4, 5, 8, dst, stride, above, left, goffset + GET_GOT goffsetq + movq m0, [leftq] ; [0- 7] l1-8 [byte] + movhps m0, [aboveq-1] ; [8-15] tl, t1-7 [byte] + pshufb m1, m0, [GLOBAL(sh_b76543210)] ; l8-1 [word] + pshufb m2, m0, [GLOBAL(sh_b65432108)] ; l7-1,tl [word] + pshufb m3, m0, [GLOBAL(sh_b54321089)] ; l6-1,tl,t1 [word] + pshufb m0, [GLOBAL(sh_b89abcdef)] ; tl,t1-7 [word] + psrldq m4, m0, 1 ; t1-7 [word] + psrldq m5, m0, 2 ; t2-7 [word] + ; comments below are for a predictor like this + ; A1 B1 C1 D1 E1 F1 G1 H1 + ; A2 B2 A1 B1 C1 D1 E1 F1 + ; A3 B3 A2 B2 A1 B1 C1 D1 + ; A4 B4 A3 B3 A2 B2 A1 B1 + ; A5 B5 A4 B4 A3 B3 A2 B2 + ; A6 B6 A5 B5 A4 B4 A3 B3 + ; A7 B7 A6 B6 A5 B5 A4 B4 + ; A8 B8 A7 B7 A6 B6 A5 B5 + pavgb m6, m1, m2 ; 2-tap avg A8-A1 + + X_PLUS_2Y_PLUS_Z_PLUS_2_RSH_2 m0, m4, m5, m7 ; 3-tap avg C-H1 + + X_PLUS_2Y_PLUS_Z_PLUS_2_RSH_2 m1, m2, m3, m0 ; 3-tap avg B8-1 + + punpcklbw m6, m0 ; A-B8, A-B7 ... A-B2, A-B1 + + DEFINE_ARGS dst, stride, stride3 + lea stride3q, [strideq*3] + + movhps [dstq+stride3q], m6 ; A-B4, A-B3, A-B2, A-B1 + palignr m0, m7, m6, 10 ; A-B3, A-B2, A-B1, C-H1 + movq [dstq+strideq*2], m0 + psrldq m0, 2 ; A-B2, A-B1, C-H1 + movq [dstq+strideq ], m0 + psrldq m0, 2 ; A-H1 + movq [dstq ], m0 + lea dstq, [dstq+strideq*4] + movq [dstq+stride3q ], m6 ; A-B8, A-B7, A-B6, A-B5 + psrldq m6, 2 ; A-B7, A-B6, A-B5, A-B4 + movq [dstq+strideq*2], m6 + psrldq m6, 2 ; A-B6, A-B5, A-B4, A-B3 + movq [dstq+strideq ], m6 + psrldq m6, 2 ; A-B5, A-B4, A-B3, A-B2 + movq [dstq ], m6 + RESTORE_GOT + RET + +INIT_XMM ssse3 +cglobal d153_predictor_16x16, 4, 5, 8, dst, stride, above, left, goffset + GET_GOT goffsetq + mova m0, [leftq] + movu m7, [aboveq-1] + ; comments below are for a predictor like this + ; A1 B1 C1 D1 E1 F1 G1 H1 I1 J1 K1 L1 M1 N1 O1 P1 + ; A2 B2 A1 B1 C1 D1 E1 F1 G1 H1 I1 J1 K1 L1 M1 N1 + ; A3 B3 A2 B2 A1 B1 C1 D1 E1 F1 G1 H1 I1 J1 K1 L1 + ; A4 B4 A3 B3 A2 B2 A1 B1 C1 D1 E1 F1 G1 H1 I1 J1 + ; A5 B5 A4 B4 A3 B3 A2 B2 A1 B1 C1 D1 E1 F1 G1 H1 + ; A6 B6 A5 B5 A4 B4 A3 B3 A2 B2 A1 B1 C1 D1 E1 F1 + ; A7 B7 A6 B6 A5 B5 A4 B4 A3 B3 A2 B2 A1 B1 C1 D1 + ; A8 B8 A7 B7 A6 B6 A5 B5 A4 B4 A3 B3 A2 B2 A1 B1 + ; A9 B9 A8 B8 A7 B7 A6 B6 A5 B5 A4 B4 A3 B3 A2 B2 + ; Aa Ba A9 B9 A8 B8 A7 B7 A6 B6 A5 B5 A4 B4 A3 B3 + ; Ab Bb Aa Ba A9 B9 A8 B8 A7 B7 A6 B6 A5 B5 A4 B4 + ; Ac Bc Ab Bb Aa Ba A9 B9 A8 B8 A7 B7 A6 B6 A5 B5 + ; Ad Bd Ac Bc Ab Bb Aa Ba A9 B9 A8 B8 A7 B7 A6 B6 + ; Ae Be Ad Bd Ac Bc Ab Bb Aa Ba A9 B9 A8 B8 A7 B7 + ; Af Bf Ae Be Ad Bd Ac Bc Ab Bb Aa Ba A9 B9 A8 B8 + ; Ag Bg Af Bf Ae Be Ad Bd Ac Bc Ab Bb Aa Ba A9 B9 + pshufb m6, m7, [GLOBAL(sh_bfedcba9876543210)] + palignr m5, m0, m6, 15 + palignr m3, m0, m6, 14 + + X_PLUS_2Y_PLUS_Z_PLUS_2_RSH_2 m0, m5, m3, m4 ; 3-tap avg B3-Bg + pshufb m1, m0, [GLOBAL(sh_b123456789abcdeff)] + pavgb m5, m0 ; A1 - Ag + + punpcklbw m0, m4, m5 ; A-B8 ... A-B1 + punpckhbw m4, m5 ; A-B9 ... A-Bg + + pshufb m3, m7, [GLOBAL(sh_b123456789abcdeff)] + pshufb m5, m7, [GLOBAL(sh_b23456789abcdefff)] + + X_PLUS_2Y_PLUS_Z_PLUS_2_RSH_2 m7, m3, m5, m1 ; 3-tap avg C1-P1 + + pshufb m6, m0, [GLOBAL(sh_bfedcba9876543210)] + DEFINE_ARGS dst, stride, stride3 + lea stride3q, [strideq*3] + palignr m2, m1, m6, 14 + mova [dstq ], m2 + palignr m2, m1, m6, 12 + mova [dstq+strideq ], m2 + palignr m2, m1, m6, 10 + mova [dstq+strideq*2], m2 + palignr m2, m1, m6, 8 + mova [dstq+stride3q ], m2 + lea dstq, [dstq+strideq*4] + palignr m2, m1, m6, 6 + mova [dstq ], m2 + palignr m2, m1, m6, 4 + mova [dstq+strideq ], m2 + palignr m2, m1, m6, 2 + mova [dstq+strideq*2], m2 + pshufb m4, [GLOBAL(sh_bfedcba9876543210)] + mova [dstq+stride3q ], m6 + lea dstq, [dstq+strideq*4] + + palignr m2, m6, m4, 14 + mova [dstq ], m2 + palignr m2, m6, m4, 12 + mova [dstq+strideq ], m2 + palignr m2, m6, m4, 10 + mova [dstq+strideq*2], m2 + palignr m2, m6, m4, 8 + mova [dstq+stride3q ], m2 + lea dstq, [dstq+strideq*4] + palignr m2, m6, m4, 6 + mova [dstq ], m2 + palignr m2, m6, m4, 4 + mova [dstq+strideq ], m2 + palignr m2, m6, m4, 2 + mova [dstq+strideq*2], m2 + mova [dstq+stride3q ], m4 + RESTORE_GOT + RET + +INIT_XMM ssse3 +cglobal d153_predictor_32x32, 4, 5, 8, dst, stride, above, left, goffset + GET_GOT goffsetq + mova m0, [leftq] + movu m7, [aboveq-1] + movu m1, [aboveq+15] + + pshufb m4, m1, [GLOBAL(sh_b123456789abcdeff)] + pshufb m6, m1, [GLOBAL(sh_b23456789abcdefff)] + + X_PLUS_2Y_PLUS_Z_PLUS_2_RSH_2 m1, m4, m6, m2 ; 3-tap avg above [high] + + palignr m3, m1, m7, 1 + palignr m5, m1, m7, 2 + + X_PLUS_2Y_PLUS_Z_PLUS_2_RSH_2 m7, m3, m5, m1 ; 3-tap avg above [low] + + pshufb m7, [GLOBAL(sh_bfedcba9876543210)] + palignr m5, m0, m7, 15 + palignr m3, m0, m7, 14 + + X_PLUS_2Y_PLUS_Z_PLUS_2_RSH_2 m0, m5, m3, m4 ; 3-tap avg B3-Bg + pavgb m5, m0 ; A1 - Ag + punpcklbw m6, m4, m5 ; A-B8 ... A-B1 + punpckhbw m4, m5 ; A-B9 ... A-Bg + pshufb m6, [GLOBAL(sh_bfedcba9876543210)] + pshufb m4, [GLOBAL(sh_bfedcba9876543210)] + + DEFINE_ARGS dst, stride, stride3, left, line + lea stride3q, [strideq*3] + + palignr m5, m2, m1, 14 + palignr m7, m1, m6, 14 + mova [dstq ], m7 + mova [dstq+16 ], m5 + palignr m5, m2, m1, 12 + palignr m7, m1, m6, 12 + mova [dstq+strideq ], m7 + mova [dstq+strideq+16 ], m5 + palignr m5, m2, m1, 10 + palignr m7, m1, m6, 10 + mova [dstq+strideq*2 ], m7 + mova [dstq+strideq*2+16], m5 + palignr m5, m2, m1, 8 + palignr m7, m1, m6, 8 + mova [dstq+stride3q ], m7 + mova [dstq+stride3q+16 ], m5 + lea dstq, [dstq+strideq*4] + palignr m5, m2, m1, 6 + palignr m7, m1, m6, 6 + mova [dstq ], m7 + mova [dstq+16 ], m5 + palignr m5, m2, m1, 4 + palignr m7, m1, m6, 4 + mova [dstq+strideq ], m7 + mova [dstq+strideq+16 ], m5 + palignr m5, m2, m1, 2 + palignr m7, m1, m6, 2 + mova [dstq+strideq*2 ], m7 + mova [dstq+strideq*2+16], m5 + mova [dstq+stride3q ], m6 + mova [dstq+stride3q+16 ], m1 + lea dstq, [dstq+strideq*4] + + palignr m5, m1, m6, 14 + palignr m3, m6, m4, 14 + mova [dstq ], m3 + mova [dstq+16 ], m5 + palignr m5, m1, m6, 12 + palignr m3, m6, m4, 12 + mova [dstq+strideq ], m3 + mova [dstq+strideq+16 ], m5 + palignr m5, m1, m6, 10 + palignr m3, m6, m4, 10 + mova [dstq+strideq*2 ], m3 + mova [dstq+strideq*2+16], m5 + palignr m5, m1, m6, 8 + palignr m3, m6, m4, 8 + mova [dstq+stride3q ], m3 + mova [dstq+stride3q+16 ], m5 + lea dstq, [dstq+strideq*4] + palignr m5, m1, m6, 6 + palignr m3, m6, m4, 6 + mova [dstq ], m3 + mova [dstq+16 ], m5 + palignr m5, m1, m6, 4 + palignr m3, m6, m4, 4 + mova [dstq+strideq ], m3 + mova [dstq+strideq+16 ], m5 + palignr m5, m1, m6, 2 + palignr m3, m6, m4, 2 + mova [dstq+strideq*2 ], m3 + mova [dstq+strideq*2+16], m5 + mova [dstq+stride3q ], m4 + mova [dstq+stride3q+16 ], m6 + lea dstq, [dstq+strideq*4] + + mova m7, [leftq] + mova m3, [leftq+16] + palignr m5, m3, m7, 15 + palignr m0, m3, m7, 14 + + X_PLUS_2Y_PLUS_Z_PLUS_2_RSH_2 m3, m5, m0, m2 ; 3-tap avg Bh - + pavgb m5, m3 ; Ah - + punpcklbw m3, m2, m5 ; A-B8 ... A-B1 + punpckhbw m2, m5 ; A-B9 ... A-Bg + pshufb m3, [GLOBAL(sh_bfedcba9876543210)] + pshufb m2, [GLOBAL(sh_bfedcba9876543210)] + + palignr m7, m6, m4, 14 + palignr m0, m4, m3, 14 + mova [dstq ], m0 + mova [dstq+16 ], m7 + palignr m7, m6, m4, 12 + palignr m0, m4, m3, 12 + mova [dstq+strideq ], m0 + mova [dstq+strideq+16 ], m7 + palignr m7, m6, m4, 10 + palignr m0, m4, m3, 10 + mova [dstq+strideq*2 ], m0 + mova [dstq+strideq*2+16], m7 + palignr m7, m6, m4, 8 + palignr m0, m4, m3, 8 + mova [dstq+stride3q ], m0 + mova [dstq+stride3q+16 ], m7 + lea dstq, [dstq+strideq*4] + palignr m7, m6, m4, 6 + palignr m0, m4, m3, 6 + mova [dstq ], m0 + mova [dstq+16 ], m7 + palignr m7, m6, m4, 4 + palignr m0, m4, m3, 4 + mova [dstq+strideq ], m0 + mova [dstq+strideq+16 ], m7 + palignr m7, m6, m4, 2 + palignr m0, m4, m3, 2 + mova [dstq+strideq*2 ], m0 + mova [dstq+strideq*2+16], m7 + mova [dstq+stride3q ], m3 + mova [dstq+stride3q+16 ], m4 + lea dstq, [dstq+strideq*4] + + palignr m7, m4, m3, 14 + palignr m0, m3, m2, 14 + mova [dstq ], m0 + mova [dstq+16 ], m7 + palignr m7, m4, m3, 12 + palignr m0, m3, m2, 12 + mova [dstq+strideq ], m0 + mova [dstq+strideq+16 ], m7 + palignr m7, m4, m3, 10 + palignr m0, m3, m2, 10 + mova [dstq+strideq*2 ], m0 + mova [dstq+strideq*2+16], m7 + palignr m7, m4, m3, 8 + palignr m0, m3, m2, 8 + mova [dstq+stride3q ], m0 + mova [dstq+stride3q+16 ], m7 + lea dstq, [dstq+strideq*4] + palignr m7, m4, m3, 6 + palignr m0, m3, m2, 6 + mova [dstq ], m0 + mova [dstq+16 ], m7 + palignr m7, m4, m3, 4 + palignr m0, m3, m2, 4 + mova [dstq+strideq ], m0 + mova [dstq+strideq+16 ], m7 + palignr m7, m4, m3, 2 + palignr m0, m3, m2, 2 + mova [dstq+strideq*2 ], m0 + mova [dstq+strideq*2+16], m7 + mova [dstq+stride3q ], m2 + mova [dstq+stride3q+16 ], m3 + + RESTORE_GOT + RET diff --git a/third_party/aom/aom_dsp/x86/inv_txfm_sse2.c b/third_party/aom/aom_dsp/x86/inv_txfm_sse2.c new file mode 100644 index 000000000..5795a1845 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/inv_txfm_sse2.c @@ -0,0 +1,3631 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include "./aom_dsp_rtcd.h" +#include "aom_dsp/x86/inv_txfm_sse2.h" +#include "aom_dsp/x86/txfm_common_sse2.h" + +#define RECON_AND_STORE4X4(dest, in_x) \ + { \ + __m128i d0 = _mm_cvtsi32_si128(*(const int *)(dest)); \ + d0 = _mm_unpacklo_epi8(d0, zero); \ + d0 = _mm_add_epi16(in_x, d0); \ + d0 = _mm_packus_epi16(d0, d0); \ + *(int *)(dest) = _mm_cvtsi128_si32(d0); \ + } + +void aom_idct4x4_16_add_sse2(const tran_low_t *input, uint8_t *dest, + int stride) { + const __m128i zero = _mm_setzero_si128(); + const __m128i eight = _mm_set1_epi16(8); + const __m128i cst = _mm_setr_epi16( + (int16_t)cospi_16_64, (int16_t)cospi_16_64, (int16_t)cospi_16_64, + (int16_t)-cospi_16_64, (int16_t)cospi_24_64, (int16_t)-cospi_8_64, + (int16_t)cospi_8_64, (int16_t)cospi_24_64); + const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING); + __m128i input0, input1, input2, input3; + + // Rows + input0 = load_input_data(input); + input2 = load_input_data(input + 8); + + // Construct i3, i1, i3, i1, i2, i0, i2, i0 + input0 = _mm_shufflelo_epi16(input0, 0xd8); + input0 = _mm_shufflehi_epi16(input0, 0xd8); + input2 = _mm_shufflelo_epi16(input2, 0xd8); + input2 = _mm_shufflehi_epi16(input2, 0xd8); + + input1 = _mm_unpackhi_epi32(input0, input0); + input0 = _mm_unpacklo_epi32(input0, input0); + input3 = _mm_unpackhi_epi32(input2, input2); + input2 = _mm_unpacklo_epi32(input2, input2); + + // Stage 1 + input0 = _mm_madd_epi16(input0, cst); + input1 = _mm_madd_epi16(input1, cst); + input2 = _mm_madd_epi16(input2, cst); + input3 = _mm_madd_epi16(input3, cst); + + input0 = _mm_add_epi32(input0, rounding); + input1 = _mm_add_epi32(input1, rounding); + input2 = _mm_add_epi32(input2, rounding); + input3 = _mm_add_epi32(input3, rounding); + + input0 = _mm_srai_epi32(input0, DCT_CONST_BITS); + input1 = _mm_srai_epi32(input1, DCT_CONST_BITS); + input2 = _mm_srai_epi32(input2, DCT_CONST_BITS); + input3 = _mm_srai_epi32(input3, DCT_CONST_BITS); + + // Stage 2 + input0 = _mm_packs_epi32(input0, input1); + input1 = _mm_packs_epi32(input2, input3); + + // Transpose + input2 = _mm_unpacklo_epi16(input0, input1); + input3 = _mm_unpackhi_epi16(input0, input1); + input0 = _mm_unpacklo_epi32(input2, input3); + input1 = _mm_unpackhi_epi32(input2, input3); + + // Switch column2, column 3, and then, we got: + // input2: column1, column 0; input3: column2, column 3. + input1 = _mm_shuffle_epi32(input1, 0x4e); + input2 = _mm_add_epi16(input0, input1); + input3 = _mm_sub_epi16(input0, input1); + + // Columns + // Construct i3, i1, i3, i1, i2, i0, i2, i0 + input0 = _mm_unpacklo_epi32(input2, input2); + input1 = _mm_unpackhi_epi32(input2, input2); + input2 = _mm_unpackhi_epi32(input3, input3); + input3 = _mm_unpacklo_epi32(input3, input3); + + // Stage 1 + input0 = _mm_madd_epi16(input0, cst); + input1 = _mm_madd_epi16(input1, cst); + input2 = _mm_madd_epi16(input2, cst); + input3 = _mm_madd_epi16(input3, cst); + + input0 = _mm_add_epi32(input0, rounding); + input1 = _mm_add_epi32(input1, rounding); + input2 = _mm_add_epi32(input2, rounding); + input3 = _mm_add_epi32(input3, rounding); + + input0 = _mm_srai_epi32(input0, DCT_CONST_BITS); + input1 = _mm_srai_epi32(input1, DCT_CONST_BITS); + input2 = _mm_srai_epi32(input2, DCT_CONST_BITS); + input3 = _mm_srai_epi32(input3, DCT_CONST_BITS); + + // Stage 2 + input0 = _mm_packs_epi32(input0, input2); + input1 = _mm_packs_epi32(input1, input3); + + // Transpose + input2 = _mm_unpacklo_epi16(input0, input1); + input3 = _mm_unpackhi_epi16(input0, input1); + input0 = _mm_unpacklo_epi32(input2, input3); + input1 = _mm_unpackhi_epi32(input2, input3); + + // Switch column2, column 3, and then, we got: + // input2: column1, column 0; input3: column2, column 3. + input1 = _mm_shuffle_epi32(input1, 0x4e); + input2 = _mm_add_epi16(input0, input1); + input3 = _mm_sub_epi16(input0, input1); + + // Final round and shift + input2 = _mm_add_epi16(input2, eight); + input3 = _mm_add_epi16(input3, eight); + + input2 = _mm_srai_epi16(input2, 4); + input3 = _mm_srai_epi16(input3, 4); + + // Reconstruction and Store + { + __m128i d0 = _mm_cvtsi32_si128(*(const int *)(dest)); + __m128i d2 = _mm_cvtsi32_si128(*(const int *)(dest + stride * 2)); + d0 = _mm_unpacklo_epi32(d0, + _mm_cvtsi32_si128(*(const int *)(dest + stride))); + d2 = _mm_unpacklo_epi32( + _mm_cvtsi32_si128(*(const int *)(dest + stride * 3)), d2); + d0 = _mm_unpacklo_epi8(d0, zero); + d2 = _mm_unpacklo_epi8(d2, zero); + d0 = _mm_add_epi16(d0, input2); + d2 = _mm_add_epi16(d2, input3); + d0 = _mm_packus_epi16(d0, d2); + // store input0 + *(int *)dest = _mm_cvtsi128_si32(d0); + // store input1 + d0 = _mm_srli_si128(d0, 4); + *(int *)(dest + stride) = _mm_cvtsi128_si32(d0); + // store input2 + d0 = _mm_srli_si128(d0, 4); + *(int *)(dest + stride * 3) = _mm_cvtsi128_si32(d0); + // store input3 + d0 = _mm_srli_si128(d0, 4); + *(int *)(dest + stride * 2) = _mm_cvtsi128_si32(d0); + } +} + +void aom_idct4x4_1_add_sse2(const tran_low_t *input, uint8_t *dest, + int stride) { + __m128i dc_value; + const __m128i zero = _mm_setzero_si128(); + int a; + + a = (int)dct_const_round_shift(input[0] * cospi_16_64); + a = (int)dct_const_round_shift(a * cospi_16_64); + a = ROUND_POWER_OF_TWO(a, 4); + + if (a == 0) return; + + dc_value = _mm_set1_epi16(a); + + RECON_AND_STORE4X4(dest + 0 * stride, dc_value); + RECON_AND_STORE4X4(dest + 1 * stride, dc_value); + RECON_AND_STORE4X4(dest + 2 * stride, dc_value); + RECON_AND_STORE4X4(dest + 3 * stride, dc_value); +} + +void aom_idct4_sse2(__m128i *in) { + const __m128i k__cospi_p16_p16 = pair_set_epi16(cospi_16_64, cospi_16_64); + const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64); + const __m128i k__cospi_p24_m08 = pair_set_epi16(cospi_24_64, -cospi_8_64); + const __m128i k__cospi_p08_p24 = pair_set_epi16(cospi_8_64, cospi_24_64); + const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING); + __m128i u[8], v[8]; + + array_transpose_4x4(in); + // stage 1 + u[0] = _mm_unpacklo_epi16(in[0], in[1]); + u[1] = _mm_unpackhi_epi16(in[0], in[1]); + v[0] = _mm_madd_epi16(u[0], k__cospi_p16_p16); + v[1] = _mm_madd_epi16(u[0], k__cospi_p16_m16); + v[2] = _mm_madd_epi16(u[1], k__cospi_p24_m08); + v[3] = _mm_madd_epi16(u[1], k__cospi_p08_p24); + + u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING); + u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING); + u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING); + u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING); + + v[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS); + v[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS); + v[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS); + v[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS); + + u[0] = _mm_packs_epi32(v[0], v[1]); + u[1] = _mm_packs_epi32(v[3], v[2]); + + // stage 2 + in[0] = _mm_add_epi16(u[0], u[1]); + in[1] = _mm_sub_epi16(u[0], u[1]); + in[1] = _mm_shuffle_epi32(in[1], 0x4E); +} + +void aom_iadst4_sse2(__m128i *in) { + const __m128i k__sinpi_p01_p04 = pair_set_epi16(sinpi_1_9, sinpi_4_9); + const __m128i k__sinpi_p03_p02 = pair_set_epi16(sinpi_3_9, sinpi_2_9); + const __m128i k__sinpi_p02_m01 = pair_set_epi16(sinpi_2_9, -sinpi_1_9); + const __m128i k__sinpi_p03_m04 = pair_set_epi16(sinpi_3_9, -sinpi_4_9); + const __m128i k__sinpi_p03_p03 = _mm_set1_epi16((int16_t)sinpi_3_9); + const __m128i kZero = _mm_set1_epi16(0); + const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING); + __m128i u[8], v[8], in7; + + array_transpose_4x4(in); + in7 = _mm_srli_si128(in[1], 8); + in7 = _mm_add_epi16(in7, in[0]); + in7 = _mm_sub_epi16(in7, in[1]); + + u[0] = _mm_unpacklo_epi16(in[0], in[1]); + u[1] = _mm_unpackhi_epi16(in[0], in[1]); + u[2] = _mm_unpacklo_epi16(in7, kZero); + u[3] = _mm_unpackhi_epi16(in[0], kZero); + + v[0] = _mm_madd_epi16(u[0], k__sinpi_p01_p04); // s0 + s3 + v[1] = _mm_madd_epi16(u[1], k__sinpi_p03_p02); // s2 + s5 + v[2] = _mm_madd_epi16(u[2], k__sinpi_p03_p03); // x2 + v[3] = _mm_madd_epi16(u[0], k__sinpi_p02_m01); // s1 - s4 + v[4] = _mm_madd_epi16(u[1], k__sinpi_p03_m04); // s2 - s6 + v[5] = _mm_madd_epi16(u[3], k__sinpi_p03_p03); // s2 + + u[0] = _mm_add_epi32(v[0], v[1]); + u[1] = _mm_add_epi32(v[3], v[4]); + u[2] = v[2]; + u[3] = _mm_add_epi32(u[0], u[1]); + u[4] = _mm_slli_epi32(v[5], 2); + u[5] = _mm_add_epi32(u[3], v[5]); + u[6] = _mm_sub_epi32(u[5], u[4]); + + v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING); + v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING); + v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING); + v[3] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING); + + u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS); + u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS); + u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS); + u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS); + + in[0] = _mm_packs_epi32(u[0], u[1]); + in[1] = _mm_packs_epi32(u[2], u[3]); +} + +// Define Macro for multiplying elements by constants and adding them together. +#define MULTIPLICATION_AND_ADD(lo_0, hi_0, lo_1, hi_1, cst0, cst1, cst2, cst3, \ + res0, res1, res2, res3) \ + { \ + tmp0 = _mm_madd_epi16(lo_0, cst0); \ + tmp1 = _mm_madd_epi16(hi_0, cst0); \ + tmp2 = _mm_madd_epi16(lo_0, cst1); \ + tmp3 = _mm_madd_epi16(hi_0, cst1); \ + tmp4 = _mm_madd_epi16(lo_1, cst2); \ + tmp5 = _mm_madd_epi16(hi_1, cst2); \ + tmp6 = _mm_madd_epi16(lo_1, cst3); \ + tmp7 = _mm_madd_epi16(hi_1, cst3); \ + \ + tmp0 = _mm_add_epi32(tmp0, rounding); \ + tmp1 = _mm_add_epi32(tmp1, rounding); \ + tmp2 = _mm_add_epi32(tmp2, rounding); \ + tmp3 = _mm_add_epi32(tmp3, rounding); \ + tmp4 = _mm_add_epi32(tmp4, rounding); \ + tmp5 = _mm_add_epi32(tmp5, rounding); \ + tmp6 = _mm_add_epi32(tmp6, rounding); \ + tmp7 = _mm_add_epi32(tmp7, rounding); \ + \ + tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS); \ + tmp1 = _mm_srai_epi32(tmp1, DCT_CONST_BITS); \ + tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS); \ + tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS); \ + tmp4 = _mm_srai_epi32(tmp4, DCT_CONST_BITS); \ + tmp5 = _mm_srai_epi32(tmp5, DCT_CONST_BITS); \ + tmp6 = _mm_srai_epi32(tmp6, DCT_CONST_BITS); \ + tmp7 = _mm_srai_epi32(tmp7, DCT_CONST_BITS); \ + \ + res0 = _mm_packs_epi32(tmp0, tmp1); \ + res1 = _mm_packs_epi32(tmp2, tmp3); \ + res2 = _mm_packs_epi32(tmp4, tmp5); \ + res3 = _mm_packs_epi32(tmp6, tmp7); \ + } + +#define MULTIPLICATION_AND_ADD_2(lo_0, hi_0, cst0, cst1, res0, res1) \ + { \ + tmp0 = _mm_madd_epi16(lo_0, cst0); \ + tmp1 = _mm_madd_epi16(hi_0, cst0); \ + tmp2 = _mm_madd_epi16(lo_0, cst1); \ + tmp3 = _mm_madd_epi16(hi_0, cst1); \ + \ + tmp0 = _mm_add_epi32(tmp0, rounding); \ + tmp1 = _mm_add_epi32(tmp1, rounding); \ + tmp2 = _mm_add_epi32(tmp2, rounding); \ + tmp3 = _mm_add_epi32(tmp3, rounding); \ + \ + tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS); \ + tmp1 = _mm_srai_epi32(tmp1, DCT_CONST_BITS); \ + tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS); \ + tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS); \ + \ + res0 = _mm_packs_epi32(tmp0, tmp1); \ + res1 = _mm_packs_epi32(tmp2, tmp3); \ + } + +#define IDCT8(in0, in1, in2, in3, in4, in5, in6, in7, out0, out1, out2, out3, \ + out4, out5, out6, out7) \ + { \ + /* Stage1 */ \ + { \ + const __m128i lo_17 = _mm_unpacklo_epi16(in1, in7); \ + const __m128i hi_17 = _mm_unpackhi_epi16(in1, in7); \ + const __m128i lo_35 = _mm_unpacklo_epi16(in3, in5); \ + const __m128i hi_35 = _mm_unpackhi_epi16(in3, in5); \ + \ + MULTIPLICATION_AND_ADD(lo_17, hi_17, lo_35, hi_35, stg1_0, stg1_1, \ + stg1_2, stg1_3, stp1_4, stp1_7, stp1_5, stp1_6) \ + } \ + \ + /* Stage2 */ \ + { \ + const __m128i lo_04 = _mm_unpacklo_epi16(in0, in4); \ + const __m128i hi_04 = _mm_unpackhi_epi16(in0, in4); \ + const __m128i lo_26 = _mm_unpacklo_epi16(in2, in6); \ + const __m128i hi_26 = _mm_unpackhi_epi16(in2, in6); \ + \ + MULTIPLICATION_AND_ADD(lo_04, hi_04, lo_26, hi_26, stg2_0, stg2_1, \ + stg2_2, stg2_3, stp2_0, stp2_1, stp2_2, stp2_3) \ + \ + stp2_4 = _mm_adds_epi16(stp1_4, stp1_5); \ + stp2_5 = _mm_subs_epi16(stp1_4, stp1_5); \ + stp2_6 = _mm_subs_epi16(stp1_7, stp1_6); \ + stp2_7 = _mm_adds_epi16(stp1_7, stp1_6); \ + } \ + \ + /* Stage3 */ \ + { \ + const __m128i lo_56 = _mm_unpacklo_epi16(stp2_6, stp2_5); \ + const __m128i hi_56 = _mm_unpackhi_epi16(stp2_6, stp2_5); \ + \ + stp1_0 = _mm_adds_epi16(stp2_0, stp2_3); \ + stp1_1 = _mm_adds_epi16(stp2_1, stp2_2); \ + stp1_2 = _mm_subs_epi16(stp2_1, stp2_2); \ + stp1_3 = _mm_subs_epi16(stp2_0, stp2_3); \ + \ + tmp0 = _mm_madd_epi16(lo_56, stg2_1); \ + tmp1 = _mm_madd_epi16(hi_56, stg2_1); \ + tmp2 = _mm_madd_epi16(lo_56, stg2_0); \ + tmp3 = _mm_madd_epi16(hi_56, stg2_0); \ + \ + tmp0 = _mm_add_epi32(tmp0, rounding); \ + tmp1 = _mm_add_epi32(tmp1, rounding); \ + tmp2 = _mm_add_epi32(tmp2, rounding); \ + tmp3 = _mm_add_epi32(tmp3, rounding); \ + \ + tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS); \ + tmp1 = _mm_srai_epi32(tmp1, DCT_CONST_BITS); \ + tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS); \ + tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS); \ + \ + stp1_5 = _mm_packs_epi32(tmp0, tmp1); \ + stp1_6 = _mm_packs_epi32(tmp2, tmp3); \ + } \ + \ + /* Stage4 */ \ + out0 = _mm_adds_epi16(stp1_0, stp2_7); \ + out1 = _mm_adds_epi16(stp1_1, stp1_6); \ + out2 = _mm_adds_epi16(stp1_2, stp1_5); \ + out3 = _mm_adds_epi16(stp1_3, stp2_4); \ + out4 = _mm_subs_epi16(stp1_3, stp2_4); \ + out5 = _mm_subs_epi16(stp1_2, stp1_5); \ + out6 = _mm_subs_epi16(stp1_1, stp1_6); \ + out7 = _mm_subs_epi16(stp1_0, stp2_7); \ + } + +void aom_idct8x8_64_add_sse2(const tran_low_t *input, uint8_t *dest, + int stride) { + const __m128i zero = _mm_setzero_si128(); + const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING); + const __m128i final_rounding = _mm_set1_epi16(1 << 4); + const __m128i stg1_0 = pair_set_epi16(cospi_28_64, -cospi_4_64); + const __m128i stg1_1 = pair_set_epi16(cospi_4_64, cospi_28_64); + const __m128i stg1_2 = pair_set_epi16(-cospi_20_64, cospi_12_64); + const __m128i stg1_3 = pair_set_epi16(cospi_12_64, cospi_20_64); + const __m128i stg2_0 = pair_set_epi16(cospi_16_64, cospi_16_64); + const __m128i stg2_1 = pair_set_epi16(cospi_16_64, -cospi_16_64); + const __m128i stg2_2 = pair_set_epi16(cospi_24_64, -cospi_8_64); + const __m128i stg2_3 = pair_set_epi16(cospi_8_64, cospi_24_64); + + __m128i in0, in1, in2, in3, in4, in5, in6, in7; + __m128i stp1_0, stp1_1, stp1_2, stp1_3, stp1_4, stp1_5, stp1_6, stp1_7; + __m128i stp2_0, stp2_1, stp2_2, stp2_3, stp2_4, stp2_5, stp2_6, stp2_7; + __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7; + int i; + + // Load input data. + in0 = load_input_data(input); + in1 = load_input_data(input + 8 * 1); + in2 = load_input_data(input + 8 * 2); + in3 = load_input_data(input + 8 * 3); + in4 = load_input_data(input + 8 * 4); + in5 = load_input_data(input + 8 * 5); + in6 = load_input_data(input + 8 * 6); + in7 = load_input_data(input + 8 * 7); + + // 2-D + for (i = 0; i < 2; i++) { + // 8x8 Transpose is copied from aom_fdct8x8_sse2() + TRANSPOSE_8X8(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2, in3, + in4, in5, in6, in7); + + // 4-stage 1D idct8x8 + IDCT8(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2, in3, in4, in5, + in6, in7); + } + + // Final rounding and shift + in0 = _mm_adds_epi16(in0, final_rounding); + in1 = _mm_adds_epi16(in1, final_rounding); + in2 = _mm_adds_epi16(in2, final_rounding); + in3 = _mm_adds_epi16(in3, final_rounding); + in4 = _mm_adds_epi16(in4, final_rounding); + in5 = _mm_adds_epi16(in5, final_rounding); + in6 = _mm_adds_epi16(in6, final_rounding); + in7 = _mm_adds_epi16(in7, final_rounding); + + in0 = _mm_srai_epi16(in0, 5); + in1 = _mm_srai_epi16(in1, 5); + in2 = _mm_srai_epi16(in2, 5); + in3 = _mm_srai_epi16(in3, 5); + in4 = _mm_srai_epi16(in4, 5); + in5 = _mm_srai_epi16(in5, 5); + in6 = _mm_srai_epi16(in6, 5); + in7 = _mm_srai_epi16(in7, 5); + + RECON_AND_STORE(dest + 0 * stride, in0); + RECON_AND_STORE(dest + 1 * stride, in1); + RECON_AND_STORE(dest + 2 * stride, in2); + RECON_AND_STORE(dest + 3 * stride, in3); + RECON_AND_STORE(dest + 4 * stride, in4); + RECON_AND_STORE(dest + 5 * stride, in5); + RECON_AND_STORE(dest + 6 * stride, in6); + RECON_AND_STORE(dest + 7 * stride, in7); +} + +void aom_idct8x8_1_add_sse2(const tran_low_t *input, uint8_t *dest, + int stride) { + __m128i dc_value; + const __m128i zero = _mm_setzero_si128(); + int a; + + a = (int)dct_const_round_shift(input[0] * cospi_16_64); + a = (int)dct_const_round_shift(a * cospi_16_64); + a = ROUND_POWER_OF_TWO(a, 5); + + if (a == 0) return; + + dc_value = _mm_set1_epi16(a); + + RECON_AND_STORE(dest + 0 * stride, dc_value); + RECON_AND_STORE(dest + 1 * stride, dc_value); + RECON_AND_STORE(dest + 2 * stride, dc_value); + RECON_AND_STORE(dest + 3 * stride, dc_value); + RECON_AND_STORE(dest + 4 * stride, dc_value); + RECON_AND_STORE(dest + 5 * stride, dc_value); + RECON_AND_STORE(dest + 6 * stride, dc_value); + RECON_AND_STORE(dest + 7 * stride, dc_value); +} + +void aom_idct8_sse2(__m128i *in) { + const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING); + const __m128i stg1_0 = pair_set_epi16(cospi_28_64, -cospi_4_64); + const __m128i stg1_1 = pair_set_epi16(cospi_4_64, cospi_28_64); + const __m128i stg1_2 = pair_set_epi16(-cospi_20_64, cospi_12_64); + const __m128i stg1_3 = pair_set_epi16(cospi_12_64, cospi_20_64); + const __m128i stg2_0 = pair_set_epi16(cospi_16_64, cospi_16_64); + const __m128i stg2_1 = pair_set_epi16(cospi_16_64, -cospi_16_64); + const __m128i stg2_2 = pair_set_epi16(cospi_24_64, -cospi_8_64); + const __m128i stg2_3 = pair_set_epi16(cospi_8_64, cospi_24_64); + + __m128i in0, in1, in2, in3, in4, in5, in6, in7; + __m128i stp1_0, stp1_1, stp1_2, stp1_3, stp1_4, stp1_5, stp1_6, stp1_7; + __m128i stp2_0, stp2_1, stp2_2, stp2_3, stp2_4, stp2_5, stp2_6, stp2_7; + __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7; + + // 8x8 Transpose is copied from aom_fdct8x8_sse2() + TRANSPOSE_8X8(in[0], in[1], in[2], in[3], in[4], in[5], in[6], in[7], in0, + in1, in2, in3, in4, in5, in6, in7); + + // 4-stage 1D idct8x8 + IDCT8(in0, in1, in2, in3, in4, in5, in6, in7, in[0], in[1], in[2], in[3], + in[4], in[5], in[6], in[7]); +} + +void aom_iadst8_sse2(__m128i *in) { + const __m128i k__cospi_p02_p30 = pair_set_epi16(cospi_2_64, cospi_30_64); + const __m128i k__cospi_p30_m02 = pair_set_epi16(cospi_30_64, -cospi_2_64); + const __m128i k__cospi_p10_p22 = pair_set_epi16(cospi_10_64, cospi_22_64); + const __m128i k__cospi_p22_m10 = pair_set_epi16(cospi_22_64, -cospi_10_64); + const __m128i k__cospi_p18_p14 = pair_set_epi16(cospi_18_64, cospi_14_64); + const __m128i k__cospi_p14_m18 = pair_set_epi16(cospi_14_64, -cospi_18_64); + const __m128i k__cospi_p26_p06 = pair_set_epi16(cospi_26_64, cospi_6_64); + const __m128i k__cospi_p06_m26 = pair_set_epi16(cospi_6_64, -cospi_26_64); + const __m128i k__cospi_p08_p24 = pair_set_epi16(cospi_8_64, cospi_24_64); + const __m128i k__cospi_p24_m08 = pair_set_epi16(cospi_24_64, -cospi_8_64); + const __m128i k__cospi_m24_p08 = pair_set_epi16(-cospi_24_64, cospi_8_64); + const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64); + const __m128i k__cospi_p16_p16 = _mm_set1_epi16((int16_t)cospi_16_64); + const __m128i k__const_0 = _mm_set1_epi16(0); + const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING); + + __m128i u0, u1, u2, u3, u4, u5, u6, u7, u8, u9, u10, u11, u12, u13, u14, u15; + __m128i v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15; + __m128i w0, w1, w2, w3, w4, w5, w6, w7, w8, w9, w10, w11, w12, w13, w14, w15; + __m128i s0, s1, s2, s3, s4, s5, s6, s7; + __m128i in0, in1, in2, in3, in4, in5, in6, in7; + + // transpose + array_transpose_8x8(in, in); + + // properly aligned for butterfly input + in0 = in[7]; + in1 = in[0]; + in2 = in[5]; + in3 = in[2]; + in4 = in[3]; + in5 = in[4]; + in6 = in[1]; + in7 = in[6]; + + // column transformation + // stage 1 + // interleave and multiply/add into 32-bit integer + s0 = _mm_unpacklo_epi16(in0, in1); + s1 = _mm_unpackhi_epi16(in0, in1); + s2 = _mm_unpacklo_epi16(in2, in3); + s3 = _mm_unpackhi_epi16(in2, in3); + s4 = _mm_unpacklo_epi16(in4, in5); + s5 = _mm_unpackhi_epi16(in4, in5); + s6 = _mm_unpacklo_epi16(in6, in7); + s7 = _mm_unpackhi_epi16(in6, in7); + + u0 = _mm_madd_epi16(s0, k__cospi_p02_p30); + u1 = _mm_madd_epi16(s1, k__cospi_p02_p30); + u2 = _mm_madd_epi16(s0, k__cospi_p30_m02); + u3 = _mm_madd_epi16(s1, k__cospi_p30_m02); + u4 = _mm_madd_epi16(s2, k__cospi_p10_p22); + u5 = _mm_madd_epi16(s3, k__cospi_p10_p22); + u6 = _mm_madd_epi16(s2, k__cospi_p22_m10); + u7 = _mm_madd_epi16(s3, k__cospi_p22_m10); + u8 = _mm_madd_epi16(s4, k__cospi_p18_p14); + u9 = _mm_madd_epi16(s5, k__cospi_p18_p14); + u10 = _mm_madd_epi16(s4, k__cospi_p14_m18); + u11 = _mm_madd_epi16(s5, k__cospi_p14_m18); + u12 = _mm_madd_epi16(s6, k__cospi_p26_p06); + u13 = _mm_madd_epi16(s7, k__cospi_p26_p06); + u14 = _mm_madd_epi16(s6, k__cospi_p06_m26); + u15 = _mm_madd_epi16(s7, k__cospi_p06_m26); + + // addition + w0 = _mm_add_epi32(u0, u8); + w1 = _mm_add_epi32(u1, u9); + w2 = _mm_add_epi32(u2, u10); + w3 = _mm_add_epi32(u3, u11); + w4 = _mm_add_epi32(u4, u12); + w5 = _mm_add_epi32(u5, u13); + w6 = _mm_add_epi32(u6, u14); + w7 = _mm_add_epi32(u7, u15); + w8 = _mm_sub_epi32(u0, u8); + w9 = _mm_sub_epi32(u1, u9); + w10 = _mm_sub_epi32(u2, u10); + w11 = _mm_sub_epi32(u3, u11); + w12 = _mm_sub_epi32(u4, u12); + w13 = _mm_sub_epi32(u5, u13); + w14 = _mm_sub_epi32(u6, u14); + w15 = _mm_sub_epi32(u7, u15); + + // shift and rounding + v0 = _mm_add_epi32(w0, k__DCT_CONST_ROUNDING); + v1 = _mm_add_epi32(w1, k__DCT_CONST_ROUNDING); + v2 = _mm_add_epi32(w2, k__DCT_CONST_ROUNDING); + v3 = _mm_add_epi32(w3, k__DCT_CONST_ROUNDING); + v4 = _mm_add_epi32(w4, k__DCT_CONST_ROUNDING); + v5 = _mm_add_epi32(w5, k__DCT_CONST_ROUNDING); + v6 = _mm_add_epi32(w6, k__DCT_CONST_ROUNDING); + v7 = _mm_add_epi32(w7, k__DCT_CONST_ROUNDING); + v8 = _mm_add_epi32(w8, k__DCT_CONST_ROUNDING); + v9 = _mm_add_epi32(w9, k__DCT_CONST_ROUNDING); + v10 = _mm_add_epi32(w10, k__DCT_CONST_ROUNDING); + v11 = _mm_add_epi32(w11, k__DCT_CONST_ROUNDING); + v12 = _mm_add_epi32(w12, k__DCT_CONST_ROUNDING); + v13 = _mm_add_epi32(w13, k__DCT_CONST_ROUNDING); + v14 = _mm_add_epi32(w14, k__DCT_CONST_ROUNDING); + v15 = _mm_add_epi32(w15, k__DCT_CONST_ROUNDING); + + u0 = _mm_srai_epi32(v0, DCT_CONST_BITS); + u1 = _mm_srai_epi32(v1, DCT_CONST_BITS); + u2 = _mm_srai_epi32(v2, DCT_CONST_BITS); + u3 = _mm_srai_epi32(v3, DCT_CONST_BITS); + u4 = _mm_srai_epi32(v4, DCT_CONST_BITS); + u5 = _mm_srai_epi32(v5, DCT_CONST_BITS); + u6 = _mm_srai_epi32(v6, DCT_CONST_BITS); + u7 = _mm_srai_epi32(v7, DCT_CONST_BITS); + u8 = _mm_srai_epi32(v8, DCT_CONST_BITS); + u9 = _mm_srai_epi32(v9, DCT_CONST_BITS); + u10 = _mm_srai_epi32(v10, DCT_CONST_BITS); + u11 = _mm_srai_epi32(v11, DCT_CONST_BITS); + u12 = _mm_srai_epi32(v12, DCT_CONST_BITS); + u13 = _mm_srai_epi32(v13, DCT_CONST_BITS); + u14 = _mm_srai_epi32(v14, DCT_CONST_BITS); + u15 = _mm_srai_epi32(v15, DCT_CONST_BITS); + + // back to 16-bit and pack 8 integers into __m128i + in[0] = _mm_packs_epi32(u0, u1); + in[1] = _mm_packs_epi32(u2, u3); + in[2] = _mm_packs_epi32(u4, u5); + in[3] = _mm_packs_epi32(u6, u7); + in[4] = _mm_packs_epi32(u8, u9); + in[5] = _mm_packs_epi32(u10, u11); + in[6] = _mm_packs_epi32(u12, u13); + in[7] = _mm_packs_epi32(u14, u15); + + // stage 2 + s0 = _mm_add_epi16(in[0], in[2]); + s1 = _mm_add_epi16(in[1], in[3]); + s2 = _mm_sub_epi16(in[0], in[2]); + s3 = _mm_sub_epi16(in[1], in[3]); + u0 = _mm_unpacklo_epi16(in[4], in[5]); + u1 = _mm_unpackhi_epi16(in[4], in[5]); + u2 = _mm_unpacklo_epi16(in[6], in[7]); + u3 = _mm_unpackhi_epi16(in[6], in[7]); + + v0 = _mm_madd_epi16(u0, k__cospi_p08_p24); + v1 = _mm_madd_epi16(u1, k__cospi_p08_p24); + v2 = _mm_madd_epi16(u0, k__cospi_p24_m08); + v3 = _mm_madd_epi16(u1, k__cospi_p24_m08); + v4 = _mm_madd_epi16(u2, k__cospi_m24_p08); + v5 = _mm_madd_epi16(u3, k__cospi_m24_p08); + v6 = _mm_madd_epi16(u2, k__cospi_p08_p24); + v7 = _mm_madd_epi16(u3, k__cospi_p08_p24); + + w0 = _mm_add_epi32(v0, v4); + w1 = _mm_add_epi32(v1, v5); + w2 = _mm_add_epi32(v2, v6); + w3 = _mm_add_epi32(v3, v7); + w4 = _mm_sub_epi32(v0, v4); + w5 = _mm_sub_epi32(v1, v5); + w6 = _mm_sub_epi32(v2, v6); + w7 = _mm_sub_epi32(v3, v7); + + v0 = _mm_add_epi32(w0, k__DCT_CONST_ROUNDING); + v1 = _mm_add_epi32(w1, k__DCT_CONST_ROUNDING); + v2 = _mm_add_epi32(w2, k__DCT_CONST_ROUNDING); + v3 = _mm_add_epi32(w3, k__DCT_CONST_ROUNDING); + v4 = _mm_add_epi32(w4, k__DCT_CONST_ROUNDING); + v5 = _mm_add_epi32(w5, k__DCT_CONST_ROUNDING); + v6 = _mm_add_epi32(w6, k__DCT_CONST_ROUNDING); + v7 = _mm_add_epi32(w7, k__DCT_CONST_ROUNDING); + + u0 = _mm_srai_epi32(v0, DCT_CONST_BITS); + u1 = _mm_srai_epi32(v1, DCT_CONST_BITS); + u2 = _mm_srai_epi32(v2, DCT_CONST_BITS); + u3 = _mm_srai_epi32(v3, DCT_CONST_BITS); + u4 = _mm_srai_epi32(v4, DCT_CONST_BITS); + u5 = _mm_srai_epi32(v5, DCT_CONST_BITS); + u6 = _mm_srai_epi32(v6, DCT_CONST_BITS); + u7 = _mm_srai_epi32(v7, DCT_CONST_BITS); + + // back to 16-bit intergers + s4 = _mm_packs_epi32(u0, u1); + s5 = _mm_packs_epi32(u2, u3); + s6 = _mm_packs_epi32(u4, u5); + s7 = _mm_packs_epi32(u6, u7); + + // stage 3 + u0 = _mm_unpacklo_epi16(s2, s3); + u1 = _mm_unpackhi_epi16(s2, s3); + u2 = _mm_unpacklo_epi16(s6, s7); + u3 = _mm_unpackhi_epi16(s6, s7); + + v0 = _mm_madd_epi16(u0, k__cospi_p16_p16); + v1 = _mm_madd_epi16(u1, k__cospi_p16_p16); + v2 = _mm_madd_epi16(u0, k__cospi_p16_m16); + v3 = _mm_madd_epi16(u1, k__cospi_p16_m16); + v4 = _mm_madd_epi16(u2, k__cospi_p16_p16); + v5 = _mm_madd_epi16(u3, k__cospi_p16_p16); + v6 = _mm_madd_epi16(u2, k__cospi_p16_m16); + v7 = _mm_madd_epi16(u3, k__cospi_p16_m16); + + u0 = _mm_add_epi32(v0, k__DCT_CONST_ROUNDING); + u1 = _mm_add_epi32(v1, k__DCT_CONST_ROUNDING); + u2 = _mm_add_epi32(v2, k__DCT_CONST_ROUNDING); + u3 = _mm_add_epi32(v3, k__DCT_CONST_ROUNDING); + u4 = _mm_add_epi32(v4, k__DCT_CONST_ROUNDING); + u5 = _mm_add_epi32(v5, k__DCT_CONST_ROUNDING); + u6 = _mm_add_epi32(v6, k__DCT_CONST_ROUNDING); + u7 = _mm_add_epi32(v7, k__DCT_CONST_ROUNDING); + + v0 = _mm_srai_epi32(u0, DCT_CONST_BITS); + v1 = _mm_srai_epi32(u1, DCT_CONST_BITS); + v2 = _mm_srai_epi32(u2, DCT_CONST_BITS); + v3 = _mm_srai_epi32(u3, DCT_CONST_BITS); + v4 = _mm_srai_epi32(u4, DCT_CONST_BITS); + v5 = _mm_srai_epi32(u5, DCT_CONST_BITS); + v6 = _mm_srai_epi32(u6, DCT_CONST_BITS); + v7 = _mm_srai_epi32(u7, DCT_CONST_BITS); + + s2 = _mm_packs_epi32(v0, v1); + s3 = _mm_packs_epi32(v2, v3); + s6 = _mm_packs_epi32(v4, v5); + s7 = _mm_packs_epi32(v6, v7); + + in[0] = s0; + in[1] = _mm_sub_epi16(k__const_0, s4); + in[2] = s6; + in[3] = _mm_sub_epi16(k__const_0, s2); + in[4] = s3; + in[5] = _mm_sub_epi16(k__const_0, s7); + in[6] = s5; + in[7] = _mm_sub_epi16(k__const_0, s1); +} + +void aom_idct8x8_12_add_sse2(const tran_low_t *input, uint8_t *dest, + int stride) { + const __m128i zero = _mm_setzero_si128(); + const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING); + const __m128i final_rounding = _mm_set1_epi16(1 << 4); + const __m128i stg1_0 = pair_set_epi16(cospi_28_64, -cospi_4_64); + const __m128i stg1_1 = pair_set_epi16(cospi_4_64, cospi_28_64); + const __m128i stg1_2 = pair_set_epi16(-cospi_20_64, cospi_12_64); + const __m128i stg1_3 = pair_set_epi16(cospi_12_64, cospi_20_64); + const __m128i stg2_0 = pair_set_epi16(cospi_16_64, cospi_16_64); + const __m128i stg2_1 = pair_set_epi16(cospi_16_64, -cospi_16_64); + const __m128i stg2_2 = pair_set_epi16(cospi_24_64, -cospi_8_64); + const __m128i stg2_3 = pair_set_epi16(cospi_8_64, cospi_24_64); + const __m128i stg3_0 = pair_set_epi16(-cospi_16_64, cospi_16_64); + + __m128i in0, in1, in2, in3, in4, in5, in6, in7; + __m128i stp1_0, stp1_1, stp1_2, stp1_3, stp1_4, stp1_5, stp1_6, stp1_7; + __m128i stp2_0, stp2_1, stp2_2, stp2_3, stp2_4, stp2_5, stp2_6, stp2_7; + __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7; + + // Rows. Load 4-row input data. + in0 = load_input_data(input); + in1 = load_input_data(input + 8 * 1); + in2 = load_input_data(input + 8 * 2); + in3 = load_input_data(input + 8 * 3); + + // 8x4 Transpose + TRANSPOSE_8X8_10(in0, in1, in2, in3, in0, in1); + // Stage1 + { + const __m128i lo_17 = _mm_unpackhi_epi16(in0, zero); + const __m128i lo_35 = _mm_unpackhi_epi16(in1, zero); + + tmp0 = _mm_madd_epi16(lo_17, stg1_0); + tmp2 = _mm_madd_epi16(lo_17, stg1_1); + tmp4 = _mm_madd_epi16(lo_35, stg1_2); + tmp6 = _mm_madd_epi16(lo_35, stg1_3); + + tmp0 = _mm_add_epi32(tmp0, rounding); + tmp2 = _mm_add_epi32(tmp2, rounding); + tmp4 = _mm_add_epi32(tmp4, rounding); + tmp6 = _mm_add_epi32(tmp6, rounding); + tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS); + tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS); + tmp4 = _mm_srai_epi32(tmp4, DCT_CONST_BITS); + tmp6 = _mm_srai_epi32(tmp6, DCT_CONST_BITS); + + stp1_4 = _mm_packs_epi32(tmp0, tmp2); + stp1_5 = _mm_packs_epi32(tmp4, tmp6); + } + + // Stage2 + { + const __m128i lo_04 = _mm_unpacklo_epi16(in0, zero); + const __m128i lo_26 = _mm_unpacklo_epi16(in1, zero); + + tmp0 = _mm_madd_epi16(lo_04, stg2_0); + tmp2 = _mm_madd_epi16(lo_04, stg2_1); + tmp4 = _mm_madd_epi16(lo_26, stg2_2); + tmp6 = _mm_madd_epi16(lo_26, stg2_3); + + tmp0 = _mm_add_epi32(tmp0, rounding); + tmp2 = _mm_add_epi32(tmp2, rounding); + tmp4 = _mm_add_epi32(tmp4, rounding); + tmp6 = _mm_add_epi32(tmp6, rounding); + tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS); + tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS); + tmp4 = _mm_srai_epi32(tmp4, DCT_CONST_BITS); + tmp6 = _mm_srai_epi32(tmp6, DCT_CONST_BITS); + + stp2_0 = _mm_packs_epi32(tmp0, tmp2); + stp2_2 = _mm_packs_epi32(tmp6, tmp4); + + tmp0 = _mm_adds_epi16(stp1_4, stp1_5); + tmp1 = _mm_subs_epi16(stp1_4, stp1_5); + + stp2_4 = tmp0; + stp2_5 = _mm_unpacklo_epi64(tmp1, zero); + stp2_6 = _mm_unpackhi_epi64(tmp1, zero); + } + + // Stage3 + { + const __m128i lo_56 = _mm_unpacklo_epi16(stp2_5, stp2_6); + + tmp4 = _mm_adds_epi16(stp2_0, stp2_2); + tmp6 = _mm_subs_epi16(stp2_0, stp2_2); + + stp1_2 = _mm_unpackhi_epi64(tmp6, tmp4); + stp1_3 = _mm_unpacklo_epi64(tmp6, tmp4); + + tmp0 = _mm_madd_epi16(lo_56, stg3_0); + tmp2 = _mm_madd_epi16(lo_56, stg2_0); // stg3_1 = stg2_0 + + tmp0 = _mm_add_epi32(tmp0, rounding); + tmp2 = _mm_add_epi32(tmp2, rounding); + tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS); + tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS); + + stp1_5 = _mm_packs_epi32(tmp0, tmp2); + } + + // Stage4 + tmp0 = _mm_adds_epi16(stp1_3, stp2_4); + tmp1 = _mm_adds_epi16(stp1_2, stp1_5); + tmp2 = _mm_subs_epi16(stp1_3, stp2_4); + tmp3 = _mm_subs_epi16(stp1_2, stp1_5); + + TRANSPOSE_4X8_10(tmp0, tmp1, tmp2, tmp3, in0, in1, in2, in3) + + IDCT8(in0, in1, in2, in3, zero, zero, zero, zero, in0, in1, in2, in3, in4, + in5, in6, in7); + // Final rounding and shift + in0 = _mm_adds_epi16(in0, final_rounding); + in1 = _mm_adds_epi16(in1, final_rounding); + in2 = _mm_adds_epi16(in2, final_rounding); + in3 = _mm_adds_epi16(in3, final_rounding); + in4 = _mm_adds_epi16(in4, final_rounding); + in5 = _mm_adds_epi16(in5, final_rounding); + in6 = _mm_adds_epi16(in6, final_rounding); + in7 = _mm_adds_epi16(in7, final_rounding); + + in0 = _mm_srai_epi16(in0, 5); + in1 = _mm_srai_epi16(in1, 5); + in2 = _mm_srai_epi16(in2, 5); + in3 = _mm_srai_epi16(in3, 5); + in4 = _mm_srai_epi16(in4, 5); + in5 = _mm_srai_epi16(in5, 5); + in6 = _mm_srai_epi16(in6, 5); + in7 = _mm_srai_epi16(in7, 5); + + RECON_AND_STORE(dest + 0 * stride, in0); + RECON_AND_STORE(dest + 1 * stride, in1); + RECON_AND_STORE(dest + 2 * stride, in2); + RECON_AND_STORE(dest + 3 * stride, in3); + RECON_AND_STORE(dest + 4 * stride, in4); + RECON_AND_STORE(dest + 5 * stride, in5); + RECON_AND_STORE(dest + 6 * stride, in6); + RECON_AND_STORE(dest + 7 * stride, in7); +} + +#define IDCT16 \ + /* Stage2 */ \ + { \ + const __m128i lo_1_15 = _mm_unpacklo_epi16(in[1], in[15]); \ + const __m128i hi_1_15 = _mm_unpackhi_epi16(in[1], in[15]); \ + const __m128i lo_9_7 = _mm_unpacklo_epi16(in[9], in[7]); \ + const __m128i hi_9_7 = _mm_unpackhi_epi16(in[9], in[7]); \ + const __m128i lo_5_11 = _mm_unpacklo_epi16(in[5], in[11]); \ + const __m128i hi_5_11 = _mm_unpackhi_epi16(in[5], in[11]); \ + const __m128i lo_13_3 = _mm_unpacklo_epi16(in[13], in[3]); \ + const __m128i hi_13_3 = _mm_unpackhi_epi16(in[13], in[3]); \ + \ + MULTIPLICATION_AND_ADD(lo_1_15, hi_1_15, lo_9_7, hi_9_7, stg2_0, stg2_1, \ + stg2_2, stg2_3, stp2_8, stp2_15, stp2_9, stp2_14) \ + \ + MULTIPLICATION_AND_ADD(lo_5_11, hi_5_11, lo_13_3, hi_13_3, stg2_4, stg2_5, \ + stg2_6, stg2_7, stp2_10, stp2_13, stp2_11, stp2_12) \ + } \ + \ + /* Stage3 */ \ + { \ + const __m128i lo_2_14 = _mm_unpacklo_epi16(in[2], in[14]); \ + const __m128i hi_2_14 = _mm_unpackhi_epi16(in[2], in[14]); \ + const __m128i lo_10_6 = _mm_unpacklo_epi16(in[10], in[6]); \ + const __m128i hi_10_6 = _mm_unpackhi_epi16(in[10], in[6]); \ + \ + MULTIPLICATION_AND_ADD(lo_2_14, hi_2_14, lo_10_6, hi_10_6, stg3_0, stg3_1, \ + stg3_2, stg3_3, stp1_4, stp1_7, stp1_5, stp1_6) \ + \ + stp1_8_0 = _mm_add_epi16(stp2_8, stp2_9); \ + stp1_9 = _mm_sub_epi16(stp2_8, stp2_9); \ + stp1_10 = _mm_sub_epi16(stp2_11, stp2_10); \ + stp1_11 = _mm_add_epi16(stp2_11, stp2_10); \ + \ + stp1_12_0 = _mm_add_epi16(stp2_12, stp2_13); \ + stp1_13 = _mm_sub_epi16(stp2_12, stp2_13); \ + stp1_14 = _mm_sub_epi16(stp2_15, stp2_14); \ + stp1_15 = _mm_add_epi16(stp2_15, stp2_14); \ + } \ + \ + /* Stage4 */ \ + { \ + const __m128i lo_0_8 = _mm_unpacklo_epi16(in[0], in[8]); \ + const __m128i hi_0_8 = _mm_unpackhi_epi16(in[0], in[8]); \ + const __m128i lo_4_12 = _mm_unpacklo_epi16(in[4], in[12]); \ + const __m128i hi_4_12 = _mm_unpackhi_epi16(in[4], in[12]); \ + \ + const __m128i lo_9_14 = _mm_unpacklo_epi16(stp1_9, stp1_14); \ + const __m128i hi_9_14 = _mm_unpackhi_epi16(stp1_9, stp1_14); \ + const __m128i lo_10_13 = _mm_unpacklo_epi16(stp1_10, stp1_13); \ + const __m128i hi_10_13 = _mm_unpackhi_epi16(stp1_10, stp1_13); \ + \ + MULTIPLICATION_AND_ADD(lo_0_8, hi_0_8, lo_4_12, hi_4_12, stg4_0, stg4_1, \ + stg4_2, stg4_3, stp2_0, stp2_1, stp2_2, stp2_3) \ + \ + stp2_4 = _mm_add_epi16(stp1_4, stp1_5); \ + stp2_5 = _mm_sub_epi16(stp1_4, stp1_5); \ + stp2_6 = _mm_sub_epi16(stp1_7, stp1_6); \ + stp2_7 = _mm_add_epi16(stp1_7, stp1_6); \ + \ + MULTIPLICATION_AND_ADD(lo_9_14, hi_9_14, lo_10_13, hi_10_13, stg4_4, \ + stg4_5, stg4_6, stg4_7, stp2_9, stp2_14, stp2_10, \ + stp2_13) \ + } \ + \ + /* Stage5 */ \ + { \ + const __m128i lo_6_5 = _mm_unpacklo_epi16(stp2_6, stp2_5); \ + const __m128i hi_6_5 = _mm_unpackhi_epi16(stp2_6, stp2_5); \ + \ + stp1_0 = _mm_add_epi16(stp2_0, stp2_3); \ + stp1_1 = _mm_add_epi16(stp2_1, stp2_2); \ + stp1_2 = _mm_sub_epi16(stp2_1, stp2_2); \ + stp1_3 = _mm_sub_epi16(stp2_0, stp2_3); \ + \ + tmp0 = _mm_madd_epi16(lo_6_5, stg4_1); \ + tmp1 = _mm_madd_epi16(hi_6_5, stg4_1); \ + tmp2 = _mm_madd_epi16(lo_6_5, stg4_0); \ + tmp3 = _mm_madd_epi16(hi_6_5, stg4_0); \ + \ + tmp0 = _mm_add_epi32(tmp0, rounding); \ + tmp1 = _mm_add_epi32(tmp1, rounding); \ + tmp2 = _mm_add_epi32(tmp2, rounding); \ + tmp3 = _mm_add_epi32(tmp3, rounding); \ + \ + tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS); \ + tmp1 = _mm_srai_epi32(tmp1, DCT_CONST_BITS); \ + tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS); \ + tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS); \ + \ + stp1_5 = _mm_packs_epi32(tmp0, tmp1); \ + stp1_6 = _mm_packs_epi32(tmp2, tmp3); \ + \ + stp1_8 = _mm_add_epi16(stp1_8_0, stp1_11); \ + stp1_9 = _mm_add_epi16(stp2_9, stp2_10); \ + stp1_10 = _mm_sub_epi16(stp2_9, stp2_10); \ + stp1_11 = _mm_sub_epi16(stp1_8_0, stp1_11); \ + \ + stp1_12 = _mm_sub_epi16(stp1_15, stp1_12_0); \ + stp1_13 = _mm_sub_epi16(stp2_14, stp2_13); \ + stp1_14 = _mm_add_epi16(stp2_14, stp2_13); \ + stp1_15 = _mm_add_epi16(stp1_15, stp1_12_0); \ + } \ + \ + /* Stage6 */ \ + { \ + const __m128i lo_10_13 = _mm_unpacklo_epi16(stp1_10, stp1_13); \ + const __m128i hi_10_13 = _mm_unpackhi_epi16(stp1_10, stp1_13); \ + const __m128i lo_11_12 = _mm_unpacklo_epi16(stp1_11, stp1_12); \ + const __m128i hi_11_12 = _mm_unpackhi_epi16(stp1_11, stp1_12); \ + \ + stp2_0 = _mm_add_epi16(stp1_0, stp2_7); \ + stp2_1 = _mm_add_epi16(stp1_1, stp1_6); \ + stp2_2 = _mm_add_epi16(stp1_2, stp1_5); \ + stp2_3 = _mm_add_epi16(stp1_3, stp2_4); \ + stp2_4 = _mm_sub_epi16(stp1_3, stp2_4); \ + stp2_5 = _mm_sub_epi16(stp1_2, stp1_5); \ + stp2_6 = _mm_sub_epi16(stp1_1, stp1_6); \ + stp2_7 = _mm_sub_epi16(stp1_0, stp2_7); \ + \ + MULTIPLICATION_AND_ADD(lo_10_13, hi_10_13, lo_11_12, hi_11_12, stg6_0, \ + stg4_0, stg6_0, stg4_0, stp2_10, stp2_13, stp2_11, \ + stp2_12) \ + } + +#define IDCT16_10 \ + /* Stage2 */ \ + { \ + const __m128i lo_1_15 = _mm_unpacklo_epi16(in[1], zero); \ + const __m128i hi_1_15 = _mm_unpackhi_epi16(in[1], zero); \ + const __m128i lo_13_3 = _mm_unpacklo_epi16(zero, in[3]); \ + const __m128i hi_13_3 = _mm_unpackhi_epi16(zero, in[3]); \ + \ + MULTIPLICATION_AND_ADD(lo_1_15, hi_1_15, lo_13_3, hi_13_3, stg2_0, stg2_1, \ + stg2_6, stg2_7, stp1_8_0, stp1_15, stp1_11, \ + stp1_12_0) \ + } \ + \ + /* Stage3 */ \ + { \ + const __m128i lo_2_14 = _mm_unpacklo_epi16(in[2], zero); \ + const __m128i hi_2_14 = _mm_unpackhi_epi16(in[2], zero); \ + \ + MULTIPLICATION_AND_ADD_2(lo_2_14, hi_2_14, stg3_0, stg3_1, stp2_4, stp2_7) \ + \ + stp1_9 = stp1_8_0; \ + stp1_10 = stp1_11; \ + \ + stp1_13 = stp1_12_0; \ + stp1_14 = stp1_15; \ + } \ + \ + /* Stage4 */ \ + { \ + const __m128i lo_0_8 = _mm_unpacklo_epi16(in[0], zero); \ + const __m128i hi_0_8 = _mm_unpackhi_epi16(in[0], zero); \ + \ + const __m128i lo_9_14 = _mm_unpacklo_epi16(stp1_9, stp1_14); \ + const __m128i hi_9_14 = _mm_unpackhi_epi16(stp1_9, stp1_14); \ + const __m128i lo_10_13 = _mm_unpacklo_epi16(stp1_10, stp1_13); \ + const __m128i hi_10_13 = _mm_unpackhi_epi16(stp1_10, stp1_13); \ + \ + MULTIPLICATION_AND_ADD_2(lo_0_8, hi_0_8, stg4_0, stg4_1, stp1_0, stp1_1) \ + stp2_5 = stp2_4; \ + stp2_6 = stp2_7; \ + \ + MULTIPLICATION_AND_ADD(lo_9_14, hi_9_14, lo_10_13, hi_10_13, stg4_4, \ + stg4_5, stg4_6, stg4_7, stp2_9, stp2_14, stp2_10, \ + stp2_13) \ + } \ + \ + /* Stage5 */ \ + { \ + const __m128i lo_6_5 = _mm_unpacklo_epi16(stp2_6, stp2_5); \ + const __m128i hi_6_5 = _mm_unpackhi_epi16(stp2_6, stp2_5); \ + \ + stp1_2 = stp1_1; \ + stp1_3 = stp1_0; \ + \ + tmp0 = _mm_madd_epi16(lo_6_5, stg4_1); \ + tmp1 = _mm_madd_epi16(hi_6_5, stg4_1); \ + tmp2 = _mm_madd_epi16(lo_6_5, stg4_0); \ + tmp3 = _mm_madd_epi16(hi_6_5, stg4_0); \ + \ + tmp0 = _mm_add_epi32(tmp0, rounding); \ + tmp1 = _mm_add_epi32(tmp1, rounding); \ + tmp2 = _mm_add_epi32(tmp2, rounding); \ + tmp3 = _mm_add_epi32(tmp3, rounding); \ + \ + tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS); \ + tmp1 = _mm_srai_epi32(tmp1, DCT_CONST_BITS); \ + tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS); \ + tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS); \ + \ + stp1_5 = _mm_packs_epi32(tmp0, tmp1); \ + stp1_6 = _mm_packs_epi32(tmp2, tmp3); \ + \ + stp1_8 = _mm_add_epi16(stp1_8_0, stp1_11); \ + stp1_9 = _mm_add_epi16(stp2_9, stp2_10); \ + stp1_10 = _mm_sub_epi16(stp2_9, stp2_10); \ + stp1_11 = _mm_sub_epi16(stp1_8_0, stp1_11); \ + \ + stp1_12 = _mm_sub_epi16(stp1_15, stp1_12_0); \ + stp1_13 = _mm_sub_epi16(stp2_14, stp2_13); \ + stp1_14 = _mm_add_epi16(stp2_14, stp2_13); \ + stp1_15 = _mm_add_epi16(stp1_15, stp1_12_0); \ + } \ + \ + /* Stage6 */ \ + { \ + const __m128i lo_10_13 = _mm_unpacklo_epi16(stp1_10, stp1_13); \ + const __m128i hi_10_13 = _mm_unpackhi_epi16(stp1_10, stp1_13); \ + const __m128i lo_11_12 = _mm_unpacklo_epi16(stp1_11, stp1_12); \ + const __m128i hi_11_12 = _mm_unpackhi_epi16(stp1_11, stp1_12); \ + \ + stp2_0 = _mm_add_epi16(stp1_0, stp2_7); \ + stp2_1 = _mm_add_epi16(stp1_1, stp1_6); \ + stp2_2 = _mm_add_epi16(stp1_2, stp1_5); \ + stp2_3 = _mm_add_epi16(stp1_3, stp2_4); \ + stp2_4 = _mm_sub_epi16(stp1_3, stp2_4); \ + stp2_5 = _mm_sub_epi16(stp1_2, stp1_5); \ + stp2_6 = _mm_sub_epi16(stp1_1, stp1_6); \ + stp2_7 = _mm_sub_epi16(stp1_0, stp2_7); \ + \ + MULTIPLICATION_AND_ADD(lo_10_13, hi_10_13, lo_11_12, hi_11_12, stg6_0, \ + stg4_0, stg6_0, stg4_0, stp2_10, stp2_13, stp2_11, \ + stp2_12) \ + } + +void aom_idct16x16_256_add_sse2(const tran_low_t *input, uint8_t *dest, + int stride) { + const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING); + const __m128i final_rounding = _mm_set1_epi16(1 << 5); + const __m128i zero = _mm_setzero_si128(); + + const __m128i stg2_0 = pair_set_epi16(cospi_30_64, -cospi_2_64); + const __m128i stg2_1 = pair_set_epi16(cospi_2_64, cospi_30_64); + const __m128i stg2_2 = pair_set_epi16(cospi_14_64, -cospi_18_64); + const __m128i stg2_3 = pair_set_epi16(cospi_18_64, cospi_14_64); + const __m128i stg2_4 = pair_set_epi16(cospi_22_64, -cospi_10_64); + const __m128i stg2_5 = pair_set_epi16(cospi_10_64, cospi_22_64); + const __m128i stg2_6 = pair_set_epi16(cospi_6_64, -cospi_26_64); + const __m128i stg2_7 = pair_set_epi16(cospi_26_64, cospi_6_64); + + const __m128i stg3_0 = pair_set_epi16(cospi_28_64, -cospi_4_64); + const __m128i stg3_1 = pair_set_epi16(cospi_4_64, cospi_28_64); + const __m128i stg3_2 = pair_set_epi16(cospi_12_64, -cospi_20_64); + const __m128i stg3_3 = pair_set_epi16(cospi_20_64, cospi_12_64); + + const __m128i stg4_0 = pair_set_epi16(cospi_16_64, cospi_16_64); + const __m128i stg4_1 = pair_set_epi16(cospi_16_64, -cospi_16_64); + const __m128i stg4_2 = pair_set_epi16(cospi_24_64, -cospi_8_64); + const __m128i stg4_3 = pair_set_epi16(cospi_8_64, cospi_24_64); + const __m128i stg4_4 = pair_set_epi16(-cospi_8_64, cospi_24_64); + const __m128i stg4_5 = pair_set_epi16(cospi_24_64, cospi_8_64); + const __m128i stg4_6 = pair_set_epi16(-cospi_24_64, -cospi_8_64); + const __m128i stg4_7 = pair_set_epi16(-cospi_8_64, cospi_24_64); + + const __m128i stg6_0 = pair_set_epi16(-cospi_16_64, cospi_16_64); + + __m128i in[16], l[16], r[16], *curr1; + __m128i stp1_0, stp1_1, stp1_2, stp1_3, stp1_4, stp1_5, stp1_6, stp1_7, + stp1_8, stp1_9, stp1_10, stp1_11, stp1_12, stp1_13, stp1_14, stp1_15, + stp1_8_0, stp1_12_0; + __m128i stp2_0, stp2_1, stp2_2, stp2_3, stp2_4, stp2_5, stp2_6, stp2_7, + stp2_8, stp2_9, stp2_10, stp2_11, stp2_12, stp2_13, stp2_14, stp2_15; + __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7; + int i; + + curr1 = l; + for (i = 0; i < 2; i++) { + // 1-D idct + + // Load input data. + in[0] = load_input_data(input); + in[8] = load_input_data(input + 8 * 1); + in[1] = load_input_data(input + 8 * 2); + in[9] = load_input_data(input + 8 * 3); + in[2] = load_input_data(input + 8 * 4); + in[10] = load_input_data(input + 8 * 5); + in[3] = load_input_data(input + 8 * 6); + in[11] = load_input_data(input + 8 * 7); + in[4] = load_input_data(input + 8 * 8); + in[12] = load_input_data(input + 8 * 9); + in[5] = load_input_data(input + 8 * 10); + in[13] = load_input_data(input + 8 * 11); + in[6] = load_input_data(input + 8 * 12); + in[14] = load_input_data(input + 8 * 13); + in[7] = load_input_data(input + 8 * 14); + in[15] = load_input_data(input + 8 * 15); + + array_transpose_8x8(in, in); + array_transpose_8x8(in + 8, in + 8); + + IDCT16 + + // Stage7 + curr1[0] = _mm_add_epi16(stp2_0, stp1_15); + curr1[1] = _mm_add_epi16(stp2_1, stp1_14); + curr1[2] = _mm_add_epi16(stp2_2, stp2_13); + curr1[3] = _mm_add_epi16(stp2_3, stp2_12); + curr1[4] = _mm_add_epi16(stp2_4, stp2_11); + curr1[5] = _mm_add_epi16(stp2_5, stp2_10); + curr1[6] = _mm_add_epi16(stp2_6, stp1_9); + curr1[7] = _mm_add_epi16(stp2_7, stp1_8); + curr1[8] = _mm_sub_epi16(stp2_7, stp1_8); + curr1[9] = _mm_sub_epi16(stp2_6, stp1_9); + curr1[10] = _mm_sub_epi16(stp2_5, stp2_10); + curr1[11] = _mm_sub_epi16(stp2_4, stp2_11); + curr1[12] = _mm_sub_epi16(stp2_3, stp2_12); + curr1[13] = _mm_sub_epi16(stp2_2, stp2_13); + curr1[14] = _mm_sub_epi16(stp2_1, stp1_14); + curr1[15] = _mm_sub_epi16(stp2_0, stp1_15); + + curr1 = r; + input += 128; + } + for (i = 0; i < 2; i++) { + int j; + // 1-D idct + array_transpose_8x8(l + i * 8, in); + array_transpose_8x8(r + i * 8, in + 8); + + IDCT16 + + // 2-D + in[0] = _mm_add_epi16(stp2_0, stp1_15); + in[1] = _mm_add_epi16(stp2_1, stp1_14); + in[2] = _mm_add_epi16(stp2_2, stp2_13); + in[3] = _mm_add_epi16(stp2_3, stp2_12); + in[4] = _mm_add_epi16(stp2_4, stp2_11); + in[5] = _mm_add_epi16(stp2_5, stp2_10); + in[6] = _mm_add_epi16(stp2_6, stp1_9); + in[7] = _mm_add_epi16(stp2_7, stp1_8); + in[8] = _mm_sub_epi16(stp2_7, stp1_8); + in[9] = _mm_sub_epi16(stp2_6, stp1_9); + in[10] = _mm_sub_epi16(stp2_5, stp2_10); + in[11] = _mm_sub_epi16(stp2_4, stp2_11); + in[12] = _mm_sub_epi16(stp2_3, stp2_12); + in[13] = _mm_sub_epi16(stp2_2, stp2_13); + in[14] = _mm_sub_epi16(stp2_1, stp1_14); + in[15] = _mm_sub_epi16(stp2_0, stp1_15); + + for (j = 0; j < 16; ++j) { + // Final rounding and shift + in[j] = _mm_adds_epi16(in[j], final_rounding); + in[j] = _mm_srai_epi16(in[j], 6); + RECON_AND_STORE(dest + j * stride, in[j]); + } + + dest += 8; + } +} + +void aom_idct16x16_1_add_sse2(const tran_low_t *input, uint8_t *dest, + int stride) { + __m128i dc_value; + const __m128i zero = _mm_setzero_si128(); + int a, i; + + a = (int)dct_const_round_shift(input[0] * cospi_16_64); + a = (int)dct_const_round_shift(a * cospi_16_64); + a = ROUND_POWER_OF_TWO(a, 6); + + if (a == 0) return; + + dc_value = _mm_set1_epi16(a); + + for (i = 0; i < 16; ++i) { + RECON_AND_STORE(dest + 0, dc_value); + RECON_AND_STORE(dest + 8, dc_value); + dest += stride; + } +} + +void iadst16_8col(__m128i *in) { + // perform 16x16 1-D ADST for 8 columns + __m128i s[16], x[16], u[32], v[32]; + const __m128i k__cospi_p01_p31 = pair_set_epi16(cospi_1_64, cospi_31_64); + const __m128i k__cospi_p31_m01 = pair_set_epi16(cospi_31_64, -cospi_1_64); + const __m128i k__cospi_p05_p27 = pair_set_epi16(cospi_5_64, cospi_27_64); + const __m128i k__cospi_p27_m05 = pair_set_epi16(cospi_27_64, -cospi_5_64); + const __m128i k__cospi_p09_p23 = pair_set_epi16(cospi_9_64, cospi_23_64); + const __m128i k__cospi_p23_m09 = pair_set_epi16(cospi_23_64, -cospi_9_64); + const __m128i k__cospi_p13_p19 = pair_set_epi16(cospi_13_64, cospi_19_64); + const __m128i k__cospi_p19_m13 = pair_set_epi16(cospi_19_64, -cospi_13_64); + const __m128i k__cospi_p17_p15 = pair_set_epi16(cospi_17_64, cospi_15_64); + const __m128i k__cospi_p15_m17 = pair_set_epi16(cospi_15_64, -cospi_17_64); + const __m128i k__cospi_p21_p11 = pair_set_epi16(cospi_21_64, cospi_11_64); + const __m128i k__cospi_p11_m21 = pair_set_epi16(cospi_11_64, -cospi_21_64); + const __m128i k__cospi_p25_p07 = pair_set_epi16(cospi_25_64, cospi_7_64); + const __m128i k__cospi_p07_m25 = pair_set_epi16(cospi_7_64, -cospi_25_64); + const __m128i k__cospi_p29_p03 = pair_set_epi16(cospi_29_64, cospi_3_64); + const __m128i k__cospi_p03_m29 = pair_set_epi16(cospi_3_64, -cospi_29_64); + const __m128i k__cospi_p04_p28 = pair_set_epi16(cospi_4_64, cospi_28_64); + const __m128i k__cospi_p28_m04 = pair_set_epi16(cospi_28_64, -cospi_4_64); + const __m128i k__cospi_p20_p12 = pair_set_epi16(cospi_20_64, cospi_12_64); + const __m128i k__cospi_p12_m20 = pair_set_epi16(cospi_12_64, -cospi_20_64); + const __m128i k__cospi_m28_p04 = pair_set_epi16(-cospi_28_64, cospi_4_64); + const __m128i k__cospi_m12_p20 = pair_set_epi16(-cospi_12_64, cospi_20_64); + const __m128i k__cospi_p08_p24 = pair_set_epi16(cospi_8_64, cospi_24_64); + const __m128i k__cospi_p24_m08 = pair_set_epi16(cospi_24_64, -cospi_8_64); + const __m128i k__cospi_m24_p08 = pair_set_epi16(-cospi_24_64, cospi_8_64); + const __m128i k__cospi_m16_m16 = _mm_set1_epi16((int16_t)-cospi_16_64); + const __m128i k__cospi_p16_p16 = _mm_set1_epi16((int16_t)cospi_16_64); + const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64); + const __m128i k__cospi_m16_p16 = pair_set_epi16(-cospi_16_64, cospi_16_64); + const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING); + const __m128i kZero = _mm_set1_epi16(0); + + u[0] = _mm_unpacklo_epi16(in[15], in[0]); + u[1] = _mm_unpackhi_epi16(in[15], in[0]); + u[2] = _mm_unpacklo_epi16(in[13], in[2]); + u[3] = _mm_unpackhi_epi16(in[13], in[2]); + u[4] = _mm_unpacklo_epi16(in[11], in[4]); + u[5] = _mm_unpackhi_epi16(in[11], in[4]); + u[6] = _mm_unpacklo_epi16(in[9], in[6]); + u[7] = _mm_unpackhi_epi16(in[9], in[6]); + u[8] = _mm_unpacklo_epi16(in[7], in[8]); + u[9] = _mm_unpackhi_epi16(in[7], in[8]); + u[10] = _mm_unpacklo_epi16(in[5], in[10]); + u[11] = _mm_unpackhi_epi16(in[5], in[10]); + u[12] = _mm_unpacklo_epi16(in[3], in[12]); + u[13] = _mm_unpackhi_epi16(in[3], in[12]); + u[14] = _mm_unpacklo_epi16(in[1], in[14]); + u[15] = _mm_unpackhi_epi16(in[1], in[14]); + + v[0] = _mm_madd_epi16(u[0], k__cospi_p01_p31); + v[1] = _mm_madd_epi16(u[1], k__cospi_p01_p31); + v[2] = _mm_madd_epi16(u[0], k__cospi_p31_m01); + v[3] = _mm_madd_epi16(u[1], k__cospi_p31_m01); + v[4] = _mm_madd_epi16(u[2], k__cospi_p05_p27); + v[5] = _mm_madd_epi16(u[3], k__cospi_p05_p27); + v[6] = _mm_madd_epi16(u[2], k__cospi_p27_m05); + v[7] = _mm_madd_epi16(u[3], k__cospi_p27_m05); + v[8] = _mm_madd_epi16(u[4], k__cospi_p09_p23); + v[9] = _mm_madd_epi16(u[5], k__cospi_p09_p23); + v[10] = _mm_madd_epi16(u[4], k__cospi_p23_m09); + v[11] = _mm_madd_epi16(u[5], k__cospi_p23_m09); + v[12] = _mm_madd_epi16(u[6], k__cospi_p13_p19); + v[13] = _mm_madd_epi16(u[7], k__cospi_p13_p19); + v[14] = _mm_madd_epi16(u[6], k__cospi_p19_m13); + v[15] = _mm_madd_epi16(u[7], k__cospi_p19_m13); + v[16] = _mm_madd_epi16(u[8], k__cospi_p17_p15); + v[17] = _mm_madd_epi16(u[9], k__cospi_p17_p15); + v[18] = _mm_madd_epi16(u[8], k__cospi_p15_m17); + v[19] = _mm_madd_epi16(u[9], k__cospi_p15_m17); + v[20] = _mm_madd_epi16(u[10], k__cospi_p21_p11); + v[21] = _mm_madd_epi16(u[11], k__cospi_p21_p11); + v[22] = _mm_madd_epi16(u[10], k__cospi_p11_m21); + v[23] = _mm_madd_epi16(u[11], k__cospi_p11_m21); + v[24] = _mm_madd_epi16(u[12], k__cospi_p25_p07); + v[25] = _mm_madd_epi16(u[13], k__cospi_p25_p07); + v[26] = _mm_madd_epi16(u[12], k__cospi_p07_m25); + v[27] = _mm_madd_epi16(u[13], k__cospi_p07_m25); + v[28] = _mm_madd_epi16(u[14], k__cospi_p29_p03); + v[29] = _mm_madd_epi16(u[15], k__cospi_p29_p03); + v[30] = _mm_madd_epi16(u[14], k__cospi_p03_m29); + v[31] = _mm_madd_epi16(u[15], k__cospi_p03_m29); + + u[0] = _mm_add_epi32(v[0], v[16]); + u[1] = _mm_add_epi32(v[1], v[17]); + u[2] = _mm_add_epi32(v[2], v[18]); + u[3] = _mm_add_epi32(v[3], v[19]); + u[4] = _mm_add_epi32(v[4], v[20]); + u[5] = _mm_add_epi32(v[5], v[21]); + u[6] = _mm_add_epi32(v[6], v[22]); + u[7] = _mm_add_epi32(v[7], v[23]); + u[8] = _mm_add_epi32(v[8], v[24]); + u[9] = _mm_add_epi32(v[9], v[25]); + u[10] = _mm_add_epi32(v[10], v[26]); + u[11] = _mm_add_epi32(v[11], v[27]); + u[12] = _mm_add_epi32(v[12], v[28]); + u[13] = _mm_add_epi32(v[13], v[29]); + u[14] = _mm_add_epi32(v[14], v[30]); + u[15] = _mm_add_epi32(v[15], v[31]); + u[16] = _mm_sub_epi32(v[0], v[16]); + u[17] = _mm_sub_epi32(v[1], v[17]); + u[18] = _mm_sub_epi32(v[2], v[18]); + u[19] = _mm_sub_epi32(v[3], v[19]); + u[20] = _mm_sub_epi32(v[4], v[20]); + u[21] = _mm_sub_epi32(v[5], v[21]); + u[22] = _mm_sub_epi32(v[6], v[22]); + u[23] = _mm_sub_epi32(v[7], v[23]); + u[24] = _mm_sub_epi32(v[8], v[24]); + u[25] = _mm_sub_epi32(v[9], v[25]); + u[26] = _mm_sub_epi32(v[10], v[26]); + u[27] = _mm_sub_epi32(v[11], v[27]); + u[28] = _mm_sub_epi32(v[12], v[28]); + u[29] = _mm_sub_epi32(v[13], v[29]); + u[30] = _mm_sub_epi32(v[14], v[30]); + u[31] = _mm_sub_epi32(v[15], v[31]); + + v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING); + v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING); + v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING); + v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING); + v[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING); + v[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING); + v[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING); + v[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING); + v[8] = _mm_add_epi32(u[8], k__DCT_CONST_ROUNDING); + v[9] = _mm_add_epi32(u[9], k__DCT_CONST_ROUNDING); + v[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING); + v[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING); + v[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING); + v[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING); + v[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING); + v[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING); + v[16] = _mm_add_epi32(u[16], k__DCT_CONST_ROUNDING); + v[17] = _mm_add_epi32(u[17], k__DCT_CONST_ROUNDING); + v[18] = _mm_add_epi32(u[18], k__DCT_CONST_ROUNDING); + v[19] = _mm_add_epi32(u[19], k__DCT_CONST_ROUNDING); + v[20] = _mm_add_epi32(u[20], k__DCT_CONST_ROUNDING); + v[21] = _mm_add_epi32(u[21], k__DCT_CONST_ROUNDING); + v[22] = _mm_add_epi32(u[22], k__DCT_CONST_ROUNDING); + v[23] = _mm_add_epi32(u[23], k__DCT_CONST_ROUNDING); + v[24] = _mm_add_epi32(u[24], k__DCT_CONST_ROUNDING); + v[25] = _mm_add_epi32(u[25], k__DCT_CONST_ROUNDING); + v[26] = _mm_add_epi32(u[26], k__DCT_CONST_ROUNDING); + v[27] = _mm_add_epi32(u[27], k__DCT_CONST_ROUNDING); + v[28] = _mm_add_epi32(u[28], k__DCT_CONST_ROUNDING); + v[29] = _mm_add_epi32(u[29], k__DCT_CONST_ROUNDING); + v[30] = _mm_add_epi32(u[30], k__DCT_CONST_ROUNDING); + v[31] = _mm_add_epi32(u[31], k__DCT_CONST_ROUNDING); + + u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS); + u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS); + u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS); + u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS); + u[4] = _mm_srai_epi32(v[4], DCT_CONST_BITS); + u[5] = _mm_srai_epi32(v[5], DCT_CONST_BITS); + u[6] = _mm_srai_epi32(v[6], DCT_CONST_BITS); + u[7] = _mm_srai_epi32(v[7], DCT_CONST_BITS); + u[8] = _mm_srai_epi32(v[8], DCT_CONST_BITS); + u[9] = _mm_srai_epi32(v[9], DCT_CONST_BITS); + u[10] = _mm_srai_epi32(v[10], DCT_CONST_BITS); + u[11] = _mm_srai_epi32(v[11], DCT_CONST_BITS); + u[12] = _mm_srai_epi32(v[12], DCT_CONST_BITS); + u[13] = _mm_srai_epi32(v[13], DCT_CONST_BITS); + u[14] = _mm_srai_epi32(v[14], DCT_CONST_BITS); + u[15] = _mm_srai_epi32(v[15], DCT_CONST_BITS); + u[16] = _mm_srai_epi32(v[16], DCT_CONST_BITS); + u[17] = _mm_srai_epi32(v[17], DCT_CONST_BITS); + u[18] = _mm_srai_epi32(v[18], DCT_CONST_BITS); + u[19] = _mm_srai_epi32(v[19], DCT_CONST_BITS); + u[20] = _mm_srai_epi32(v[20], DCT_CONST_BITS); + u[21] = _mm_srai_epi32(v[21], DCT_CONST_BITS); + u[22] = _mm_srai_epi32(v[22], DCT_CONST_BITS); + u[23] = _mm_srai_epi32(v[23], DCT_CONST_BITS); + u[24] = _mm_srai_epi32(v[24], DCT_CONST_BITS); + u[25] = _mm_srai_epi32(v[25], DCT_CONST_BITS); + u[26] = _mm_srai_epi32(v[26], DCT_CONST_BITS); + u[27] = _mm_srai_epi32(v[27], DCT_CONST_BITS); + u[28] = _mm_srai_epi32(v[28], DCT_CONST_BITS); + u[29] = _mm_srai_epi32(v[29], DCT_CONST_BITS); + u[30] = _mm_srai_epi32(v[30], DCT_CONST_BITS); + u[31] = _mm_srai_epi32(v[31], DCT_CONST_BITS); + + s[0] = _mm_packs_epi32(u[0], u[1]); + s[1] = _mm_packs_epi32(u[2], u[3]); + s[2] = _mm_packs_epi32(u[4], u[5]); + s[3] = _mm_packs_epi32(u[6], u[7]); + s[4] = _mm_packs_epi32(u[8], u[9]); + s[5] = _mm_packs_epi32(u[10], u[11]); + s[6] = _mm_packs_epi32(u[12], u[13]); + s[7] = _mm_packs_epi32(u[14], u[15]); + s[8] = _mm_packs_epi32(u[16], u[17]); + s[9] = _mm_packs_epi32(u[18], u[19]); + s[10] = _mm_packs_epi32(u[20], u[21]); + s[11] = _mm_packs_epi32(u[22], u[23]); + s[12] = _mm_packs_epi32(u[24], u[25]); + s[13] = _mm_packs_epi32(u[26], u[27]); + s[14] = _mm_packs_epi32(u[28], u[29]); + s[15] = _mm_packs_epi32(u[30], u[31]); + + // stage 2 + u[0] = _mm_unpacklo_epi16(s[8], s[9]); + u[1] = _mm_unpackhi_epi16(s[8], s[9]); + u[2] = _mm_unpacklo_epi16(s[10], s[11]); + u[3] = _mm_unpackhi_epi16(s[10], s[11]); + u[4] = _mm_unpacklo_epi16(s[12], s[13]); + u[5] = _mm_unpackhi_epi16(s[12], s[13]); + u[6] = _mm_unpacklo_epi16(s[14], s[15]); + u[7] = _mm_unpackhi_epi16(s[14], s[15]); + + v[0] = _mm_madd_epi16(u[0], k__cospi_p04_p28); + v[1] = _mm_madd_epi16(u[1], k__cospi_p04_p28); + v[2] = _mm_madd_epi16(u[0], k__cospi_p28_m04); + v[3] = _mm_madd_epi16(u[1], k__cospi_p28_m04); + v[4] = _mm_madd_epi16(u[2], k__cospi_p20_p12); + v[5] = _mm_madd_epi16(u[3], k__cospi_p20_p12); + v[6] = _mm_madd_epi16(u[2], k__cospi_p12_m20); + v[7] = _mm_madd_epi16(u[3], k__cospi_p12_m20); + v[8] = _mm_madd_epi16(u[4], k__cospi_m28_p04); + v[9] = _mm_madd_epi16(u[5], k__cospi_m28_p04); + v[10] = _mm_madd_epi16(u[4], k__cospi_p04_p28); + v[11] = _mm_madd_epi16(u[5], k__cospi_p04_p28); + v[12] = _mm_madd_epi16(u[6], k__cospi_m12_p20); + v[13] = _mm_madd_epi16(u[7], k__cospi_m12_p20); + v[14] = _mm_madd_epi16(u[6], k__cospi_p20_p12); + v[15] = _mm_madd_epi16(u[7], k__cospi_p20_p12); + + u[0] = _mm_add_epi32(v[0], v[8]); + u[1] = _mm_add_epi32(v[1], v[9]); + u[2] = _mm_add_epi32(v[2], v[10]); + u[3] = _mm_add_epi32(v[3], v[11]); + u[4] = _mm_add_epi32(v[4], v[12]); + u[5] = _mm_add_epi32(v[5], v[13]); + u[6] = _mm_add_epi32(v[6], v[14]); + u[7] = _mm_add_epi32(v[7], v[15]); + u[8] = _mm_sub_epi32(v[0], v[8]); + u[9] = _mm_sub_epi32(v[1], v[9]); + u[10] = _mm_sub_epi32(v[2], v[10]); + u[11] = _mm_sub_epi32(v[3], v[11]); + u[12] = _mm_sub_epi32(v[4], v[12]); + u[13] = _mm_sub_epi32(v[5], v[13]); + u[14] = _mm_sub_epi32(v[6], v[14]); + u[15] = _mm_sub_epi32(v[7], v[15]); + + v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING); + v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING); + v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING); + v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING); + v[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING); + v[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING); + v[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING); + v[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING); + v[8] = _mm_add_epi32(u[8], k__DCT_CONST_ROUNDING); + v[9] = _mm_add_epi32(u[9], k__DCT_CONST_ROUNDING); + v[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING); + v[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING); + v[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING); + v[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING); + v[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING); + v[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING); + + u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS); + u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS); + u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS); + u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS); + u[4] = _mm_srai_epi32(v[4], DCT_CONST_BITS); + u[5] = _mm_srai_epi32(v[5], DCT_CONST_BITS); + u[6] = _mm_srai_epi32(v[6], DCT_CONST_BITS); + u[7] = _mm_srai_epi32(v[7], DCT_CONST_BITS); + u[8] = _mm_srai_epi32(v[8], DCT_CONST_BITS); + u[9] = _mm_srai_epi32(v[9], DCT_CONST_BITS); + u[10] = _mm_srai_epi32(v[10], DCT_CONST_BITS); + u[11] = _mm_srai_epi32(v[11], DCT_CONST_BITS); + u[12] = _mm_srai_epi32(v[12], DCT_CONST_BITS); + u[13] = _mm_srai_epi32(v[13], DCT_CONST_BITS); + u[14] = _mm_srai_epi32(v[14], DCT_CONST_BITS); + u[15] = _mm_srai_epi32(v[15], DCT_CONST_BITS); + + x[0] = _mm_add_epi16(s[0], s[4]); + x[1] = _mm_add_epi16(s[1], s[5]); + x[2] = _mm_add_epi16(s[2], s[6]); + x[3] = _mm_add_epi16(s[3], s[7]); + x[4] = _mm_sub_epi16(s[0], s[4]); + x[5] = _mm_sub_epi16(s[1], s[5]); + x[6] = _mm_sub_epi16(s[2], s[6]); + x[7] = _mm_sub_epi16(s[3], s[7]); + x[8] = _mm_packs_epi32(u[0], u[1]); + x[9] = _mm_packs_epi32(u[2], u[3]); + x[10] = _mm_packs_epi32(u[4], u[5]); + x[11] = _mm_packs_epi32(u[6], u[7]); + x[12] = _mm_packs_epi32(u[8], u[9]); + x[13] = _mm_packs_epi32(u[10], u[11]); + x[14] = _mm_packs_epi32(u[12], u[13]); + x[15] = _mm_packs_epi32(u[14], u[15]); + + // stage 3 + u[0] = _mm_unpacklo_epi16(x[4], x[5]); + u[1] = _mm_unpackhi_epi16(x[4], x[5]); + u[2] = _mm_unpacklo_epi16(x[6], x[7]); + u[3] = _mm_unpackhi_epi16(x[6], x[7]); + u[4] = _mm_unpacklo_epi16(x[12], x[13]); + u[5] = _mm_unpackhi_epi16(x[12], x[13]); + u[6] = _mm_unpacklo_epi16(x[14], x[15]); + u[7] = _mm_unpackhi_epi16(x[14], x[15]); + + v[0] = _mm_madd_epi16(u[0], k__cospi_p08_p24); + v[1] = _mm_madd_epi16(u[1], k__cospi_p08_p24); + v[2] = _mm_madd_epi16(u[0], k__cospi_p24_m08); + v[3] = _mm_madd_epi16(u[1], k__cospi_p24_m08); + v[4] = _mm_madd_epi16(u[2], k__cospi_m24_p08); + v[5] = _mm_madd_epi16(u[3], k__cospi_m24_p08); + v[6] = _mm_madd_epi16(u[2], k__cospi_p08_p24); + v[7] = _mm_madd_epi16(u[3], k__cospi_p08_p24); + v[8] = _mm_madd_epi16(u[4], k__cospi_p08_p24); + v[9] = _mm_madd_epi16(u[5], k__cospi_p08_p24); + v[10] = _mm_madd_epi16(u[4], k__cospi_p24_m08); + v[11] = _mm_madd_epi16(u[5], k__cospi_p24_m08); + v[12] = _mm_madd_epi16(u[6], k__cospi_m24_p08); + v[13] = _mm_madd_epi16(u[7], k__cospi_m24_p08); + v[14] = _mm_madd_epi16(u[6], k__cospi_p08_p24); + v[15] = _mm_madd_epi16(u[7], k__cospi_p08_p24); + + u[0] = _mm_add_epi32(v[0], v[4]); + u[1] = _mm_add_epi32(v[1], v[5]); + u[2] = _mm_add_epi32(v[2], v[6]); + u[3] = _mm_add_epi32(v[3], v[7]); + u[4] = _mm_sub_epi32(v[0], v[4]); + u[5] = _mm_sub_epi32(v[1], v[5]); + u[6] = _mm_sub_epi32(v[2], v[6]); + u[7] = _mm_sub_epi32(v[3], v[7]); + u[8] = _mm_add_epi32(v[8], v[12]); + u[9] = _mm_add_epi32(v[9], v[13]); + u[10] = _mm_add_epi32(v[10], v[14]); + u[11] = _mm_add_epi32(v[11], v[15]); + u[12] = _mm_sub_epi32(v[8], v[12]); + u[13] = _mm_sub_epi32(v[9], v[13]); + u[14] = _mm_sub_epi32(v[10], v[14]); + u[15] = _mm_sub_epi32(v[11], v[15]); + + u[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING); + u[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING); + u[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING); + u[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING); + u[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING); + u[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING); + u[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING); + u[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING); + u[8] = _mm_add_epi32(u[8], k__DCT_CONST_ROUNDING); + u[9] = _mm_add_epi32(u[9], k__DCT_CONST_ROUNDING); + u[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING); + u[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING); + u[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING); + u[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING); + u[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING); + u[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING); + + v[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS); + v[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS); + v[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS); + v[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS); + v[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS); + v[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS); + v[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS); + v[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS); + v[8] = _mm_srai_epi32(u[8], DCT_CONST_BITS); + v[9] = _mm_srai_epi32(u[9], DCT_CONST_BITS); + v[10] = _mm_srai_epi32(u[10], DCT_CONST_BITS); + v[11] = _mm_srai_epi32(u[11], DCT_CONST_BITS); + v[12] = _mm_srai_epi32(u[12], DCT_CONST_BITS); + v[13] = _mm_srai_epi32(u[13], DCT_CONST_BITS); + v[14] = _mm_srai_epi32(u[14], DCT_CONST_BITS); + v[15] = _mm_srai_epi32(u[15], DCT_CONST_BITS); + + s[0] = _mm_add_epi16(x[0], x[2]); + s[1] = _mm_add_epi16(x[1], x[3]); + s[2] = _mm_sub_epi16(x[0], x[2]); + s[3] = _mm_sub_epi16(x[1], x[3]); + s[4] = _mm_packs_epi32(v[0], v[1]); + s[5] = _mm_packs_epi32(v[2], v[3]); + s[6] = _mm_packs_epi32(v[4], v[5]); + s[7] = _mm_packs_epi32(v[6], v[7]); + s[8] = _mm_add_epi16(x[8], x[10]); + s[9] = _mm_add_epi16(x[9], x[11]); + s[10] = _mm_sub_epi16(x[8], x[10]); + s[11] = _mm_sub_epi16(x[9], x[11]); + s[12] = _mm_packs_epi32(v[8], v[9]); + s[13] = _mm_packs_epi32(v[10], v[11]); + s[14] = _mm_packs_epi32(v[12], v[13]); + s[15] = _mm_packs_epi32(v[14], v[15]); + + // stage 4 + u[0] = _mm_unpacklo_epi16(s[2], s[3]); + u[1] = _mm_unpackhi_epi16(s[2], s[3]); + u[2] = _mm_unpacklo_epi16(s[6], s[7]); + u[3] = _mm_unpackhi_epi16(s[6], s[7]); + u[4] = _mm_unpacklo_epi16(s[10], s[11]); + u[5] = _mm_unpackhi_epi16(s[10], s[11]); + u[6] = _mm_unpacklo_epi16(s[14], s[15]); + u[7] = _mm_unpackhi_epi16(s[14], s[15]); + + v[0] = _mm_madd_epi16(u[0], k__cospi_m16_m16); + v[1] = _mm_madd_epi16(u[1], k__cospi_m16_m16); + v[2] = _mm_madd_epi16(u[0], k__cospi_p16_m16); + v[3] = _mm_madd_epi16(u[1], k__cospi_p16_m16); + v[4] = _mm_madd_epi16(u[2], k__cospi_p16_p16); + v[5] = _mm_madd_epi16(u[3], k__cospi_p16_p16); + v[6] = _mm_madd_epi16(u[2], k__cospi_m16_p16); + v[7] = _mm_madd_epi16(u[3], k__cospi_m16_p16); + v[8] = _mm_madd_epi16(u[4], k__cospi_p16_p16); + v[9] = _mm_madd_epi16(u[5], k__cospi_p16_p16); + v[10] = _mm_madd_epi16(u[4], k__cospi_m16_p16); + v[11] = _mm_madd_epi16(u[5], k__cospi_m16_p16); + v[12] = _mm_madd_epi16(u[6], k__cospi_m16_m16); + v[13] = _mm_madd_epi16(u[7], k__cospi_m16_m16); + v[14] = _mm_madd_epi16(u[6], k__cospi_p16_m16); + v[15] = _mm_madd_epi16(u[7], k__cospi_p16_m16); + + u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING); + u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING); + u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING); + u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING); + u[4] = _mm_add_epi32(v[4], k__DCT_CONST_ROUNDING); + u[5] = _mm_add_epi32(v[5], k__DCT_CONST_ROUNDING); + u[6] = _mm_add_epi32(v[6], k__DCT_CONST_ROUNDING); + u[7] = _mm_add_epi32(v[7], k__DCT_CONST_ROUNDING); + u[8] = _mm_add_epi32(v[8], k__DCT_CONST_ROUNDING); + u[9] = _mm_add_epi32(v[9], k__DCT_CONST_ROUNDING); + u[10] = _mm_add_epi32(v[10], k__DCT_CONST_ROUNDING); + u[11] = _mm_add_epi32(v[11], k__DCT_CONST_ROUNDING); + u[12] = _mm_add_epi32(v[12], k__DCT_CONST_ROUNDING); + u[13] = _mm_add_epi32(v[13], k__DCT_CONST_ROUNDING); + u[14] = _mm_add_epi32(v[14], k__DCT_CONST_ROUNDING); + u[15] = _mm_add_epi32(v[15], k__DCT_CONST_ROUNDING); + + v[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS); + v[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS); + v[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS); + v[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS); + v[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS); + v[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS); + v[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS); + v[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS); + v[8] = _mm_srai_epi32(u[8], DCT_CONST_BITS); + v[9] = _mm_srai_epi32(u[9], DCT_CONST_BITS); + v[10] = _mm_srai_epi32(u[10], DCT_CONST_BITS); + v[11] = _mm_srai_epi32(u[11], DCT_CONST_BITS); + v[12] = _mm_srai_epi32(u[12], DCT_CONST_BITS); + v[13] = _mm_srai_epi32(u[13], DCT_CONST_BITS); + v[14] = _mm_srai_epi32(u[14], DCT_CONST_BITS); + v[15] = _mm_srai_epi32(u[15], DCT_CONST_BITS); + + in[0] = s[0]; + in[1] = _mm_sub_epi16(kZero, s[8]); + in[2] = s[12]; + in[3] = _mm_sub_epi16(kZero, s[4]); + in[4] = _mm_packs_epi32(v[4], v[5]); + in[5] = _mm_packs_epi32(v[12], v[13]); + in[6] = _mm_packs_epi32(v[8], v[9]); + in[7] = _mm_packs_epi32(v[0], v[1]); + in[8] = _mm_packs_epi32(v[2], v[3]); + in[9] = _mm_packs_epi32(v[10], v[11]); + in[10] = _mm_packs_epi32(v[14], v[15]); + in[11] = _mm_packs_epi32(v[6], v[7]); + in[12] = s[5]; + in[13] = _mm_sub_epi16(kZero, s[13]); + in[14] = s[9]; + in[15] = _mm_sub_epi16(kZero, s[1]); +} + +void idct16_8col(__m128i *in) { + const __m128i k__cospi_p30_m02 = pair_set_epi16(cospi_30_64, -cospi_2_64); + const __m128i k__cospi_p02_p30 = pair_set_epi16(cospi_2_64, cospi_30_64); + const __m128i k__cospi_p14_m18 = pair_set_epi16(cospi_14_64, -cospi_18_64); + const __m128i k__cospi_p18_p14 = pair_set_epi16(cospi_18_64, cospi_14_64); + const __m128i k__cospi_p22_m10 = pair_set_epi16(cospi_22_64, -cospi_10_64); + const __m128i k__cospi_p10_p22 = pair_set_epi16(cospi_10_64, cospi_22_64); + const __m128i k__cospi_p06_m26 = pair_set_epi16(cospi_6_64, -cospi_26_64); + const __m128i k__cospi_p26_p06 = pair_set_epi16(cospi_26_64, cospi_6_64); + const __m128i k__cospi_p28_m04 = pair_set_epi16(cospi_28_64, -cospi_4_64); + const __m128i k__cospi_p04_p28 = pair_set_epi16(cospi_4_64, cospi_28_64); + const __m128i k__cospi_p12_m20 = pair_set_epi16(cospi_12_64, -cospi_20_64); + const __m128i k__cospi_p20_p12 = pair_set_epi16(cospi_20_64, cospi_12_64); + const __m128i k__cospi_p16_p16 = _mm_set1_epi16((int16_t)cospi_16_64); + const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64); + const __m128i k__cospi_p24_m08 = pair_set_epi16(cospi_24_64, -cospi_8_64); + const __m128i k__cospi_p08_p24 = pair_set_epi16(cospi_8_64, cospi_24_64); + const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64, cospi_24_64); + const __m128i k__cospi_p24_p08 = pair_set_epi16(cospi_24_64, cospi_8_64); + const __m128i k__cospi_m24_m08 = pair_set_epi16(-cospi_24_64, -cospi_8_64); + const __m128i k__cospi_m16_p16 = pair_set_epi16(-cospi_16_64, cospi_16_64); + const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING); + __m128i v[16], u[16], s[16], t[16]; + + // stage 1 + s[0] = in[0]; + s[1] = in[8]; + s[2] = in[4]; + s[3] = in[12]; + s[4] = in[2]; + s[5] = in[10]; + s[6] = in[6]; + s[7] = in[14]; + s[8] = in[1]; + s[9] = in[9]; + s[10] = in[5]; + s[11] = in[13]; + s[12] = in[3]; + s[13] = in[11]; + s[14] = in[7]; + s[15] = in[15]; + + // stage 2 + u[0] = _mm_unpacklo_epi16(s[8], s[15]); + u[1] = _mm_unpackhi_epi16(s[8], s[15]); + u[2] = _mm_unpacklo_epi16(s[9], s[14]); + u[3] = _mm_unpackhi_epi16(s[9], s[14]); + u[4] = _mm_unpacklo_epi16(s[10], s[13]); + u[5] = _mm_unpackhi_epi16(s[10], s[13]); + u[6] = _mm_unpacklo_epi16(s[11], s[12]); + u[7] = _mm_unpackhi_epi16(s[11], s[12]); + + v[0] = _mm_madd_epi16(u[0], k__cospi_p30_m02); + v[1] = _mm_madd_epi16(u[1], k__cospi_p30_m02); + v[2] = _mm_madd_epi16(u[0], k__cospi_p02_p30); + v[3] = _mm_madd_epi16(u[1], k__cospi_p02_p30); + v[4] = _mm_madd_epi16(u[2], k__cospi_p14_m18); + v[5] = _mm_madd_epi16(u[3], k__cospi_p14_m18); + v[6] = _mm_madd_epi16(u[2], k__cospi_p18_p14); + v[7] = _mm_madd_epi16(u[3], k__cospi_p18_p14); + v[8] = _mm_madd_epi16(u[4], k__cospi_p22_m10); + v[9] = _mm_madd_epi16(u[5], k__cospi_p22_m10); + v[10] = _mm_madd_epi16(u[4], k__cospi_p10_p22); + v[11] = _mm_madd_epi16(u[5], k__cospi_p10_p22); + v[12] = _mm_madd_epi16(u[6], k__cospi_p06_m26); + v[13] = _mm_madd_epi16(u[7], k__cospi_p06_m26); + v[14] = _mm_madd_epi16(u[6], k__cospi_p26_p06); + v[15] = _mm_madd_epi16(u[7], k__cospi_p26_p06); + + u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING); + u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING); + u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING); + u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING); + u[4] = _mm_add_epi32(v[4], k__DCT_CONST_ROUNDING); + u[5] = _mm_add_epi32(v[5], k__DCT_CONST_ROUNDING); + u[6] = _mm_add_epi32(v[6], k__DCT_CONST_ROUNDING); + u[7] = _mm_add_epi32(v[7], k__DCT_CONST_ROUNDING); + u[8] = _mm_add_epi32(v[8], k__DCT_CONST_ROUNDING); + u[9] = _mm_add_epi32(v[9], k__DCT_CONST_ROUNDING); + u[10] = _mm_add_epi32(v[10], k__DCT_CONST_ROUNDING); + u[11] = _mm_add_epi32(v[11], k__DCT_CONST_ROUNDING); + u[12] = _mm_add_epi32(v[12], k__DCT_CONST_ROUNDING); + u[13] = _mm_add_epi32(v[13], k__DCT_CONST_ROUNDING); + u[14] = _mm_add_epi32(v[14], k__DCT_CONST_ROUNDING); + u[15] = _mm_add_epi32(v[15], k__DCT_CONST_ROUNDING); + + u[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS); + u[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS); + u[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS); + u[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS); + u[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS); + u[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS); + u[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS); + u[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS); + u[8] = _mm_srai_epi32(u[8], DCT_CONST_BITS); + u[9] = _mm_srai_epi32(u[9], DCT_CONST_BITS); + u[10] = _mm_srai_epi32(u[10], DCT_CONST_BITS); + u[11] = _mm_srai_epi32(u[11], DCT_CONST_BITS); + u[12] = _mm_srai_epi32(u[12], DCT_CONST_BITS); + u[13] = _mm_srai_epi32(u[13], DCT_CONST_BITS); + u[14] = _mm_srai_epi32(u[14], DCT_CONST_BITS); + u[15] = _mm_srai_epi32(u[15], DCT_CONST_BITS); + + s[8] = _mm_packs_epi32(u[0], u[1]); + s[15] = _mm_packs_epi32(u[2], u[3]); + s[9] = _mm_packs_epi32(u[4], u[5]); + s[14] = _mm_packs_epi32(u[6], u[7]); + s[10] = _mm_packs_epi32(u[8], u[9]); + s[13] = _mm_packs_epi32(u[10], u[11]); + s[11] = _mm_packs_epi32(u[12], u[13]); + s[12] = _mm_packs_epi32(u[14], u[15]); + + // stage 3 + t[0] = s[0]; + t[1] = s[1]; + t[2] = s[2]; + t[3] = s[3]; + u[0] = _mm_unpacklo_epi16(s[4], s[7]); + u[1] = _mm_unpackhi_epi16(s[4], s[7]); + u[2] = _mm_unpacklo_epi16(s[5], s[6]); + u[3] = _mm_unpackhi_epi16(s[5], s[6]); + + v[0] = _mm_madd_epi16(u[0], k__cospi_p28_m04); + v[1] = _mm_madd_epi16(u[1], k__cospi_p28_m04); + v[2] = _mm_madd_epi16(u[0], k__cospi_p04_p28); + v[3] = _mm_madd_epi16(u[1], k__cospi_p04_p28); + v[4] = _mm_madd_epi16(u[2], k__cospi_p12_m20); + v[5] = _mm_madd_epi16(u[3], k__cospi_p12_m20); + v[6] = _mm_madd_epi16(u[2], k__cospi_p20_p12); + v[7] = _mm_madd_epi16(u[3], k__cospi_p20_p12); + + u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING); + u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING); + u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING); + u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING); + u[4] = _mm_add_epi32(v[4], k__DCT_CONST_ROUNDING); + u[5] = _mm_add_epi32(v[5], k__DCT_CONST_ROUNDING); + u[6] = _mm_add_epi32(v[6], k__DCT_CONST_ROUNDING); + u[7] = _mm_add_epi32(v[7], k__DCT_CONST_ROUNDING); + + u[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS); + u[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS); + u[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS); + u[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS); + u[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS); + u[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS); + u[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS); + u[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS); + + t[4] = _mm_packs_epi32(u[0], u[1]); + t[7] = _mm_packs_epi32(u[2], u[3]); + t[5] = _mm_packs_epi32(u[4], u[5]); + t[6] = _mm_packs_epi32(u[6], u[7]); + t[8] = _mm_add_epi16(s[8], s[9]); + t[9] = _mm_sub_epi16(s[8], s[9]); + t[10] = _mm_sub_epi16(s[11], s[10]); + t[11] = _mm_add_epi16(s[10], s[11]); + t[12] = _mm_add_epi16(s[12], s[13]); + t[13] = _mm_sub_epi16(s[12], s[13]); + t[14] = _mm_sub_epi16(s[15], s[14]); + t[15] = _mm_add_epi16(s[14], s[15]); + + // stage 4 + u[0] = _mm_unpacklo_epi16(t[0], t[1]); + u[1] = _mm_unpackhi_epi16(t[0], t[1]); + u[2] = _mm_unpacklo_epi16(t[2], t[3]); + u[3] = _mm_unpackhi_epi16(t[2], t[3]); + u[4] = _mm_unpacklo_epi16(t[9], t[14]); + u[5] = _mm_unpackhi_epi16(t[9], t[14]); + u[6] = _mm_unpacklo_epi16(t[10], t[13]); + u[7] = _mm_unpackhi_epi16(t[10], t[13]); + + v[0] = _mm_madd_epi16(u[0], k__cospi_p16_p16); + v[1] = _mm_madd_epi16(u[1], k__cospi_p16_p16); + v[2] = _mm_madd_epi16(u[0], k__cospi_p16_m16); + v[3] = _mm_madd_epi16(u[1], k__cospi_p16_m16); + v[4] = _mm_madd_epi16(u[2], k__cospi_p24_m08); + v[5] = _mm_madd_epi16(u[3], k__cospi_p24_m08); + v[6] = _mm_madd_epi16(u[2], k__cospi_p08_p24); + v[7] = _mm_madd_epi16(u[3], k__cospi_p08_p24); + v[8] = _mm_madd_epi16(u[4], k__cospi_m08_p24); + v[9] = _mm_madd_epi16(u[5], k__cospi_m08_p24); + v[10] = _mm_madd_epi16(u[4], k__cospi_p24_p08); + v[11] = _mm_madd_epi16(u[5], k__cospi_p24_p08); + v[12] = _mm_madd_epi16(u[6], k__cospi_m24_m08); + v[13] = _mm_madd_epi16(u[7], k__cospi_m24_m08); + v[14] = _mm_madd_epi16(u[6], k__cospi_m08_p24); + v[15] = _mm_madd_epi16(u[7], k__cospi_m08_p24); + + u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING); + u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING); + u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING); + u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING); + u[4] = _mm_add_epi32(v[4], k__DCT_CONST_ROUNDING); + u[5] = _mm_add_epi32(v[5], k__DCT_CONST_ROUNDING); + u[6] = _mm_add_epi32(v[6], k__DCT_CONST_ROUNDING); + u[7] = _mm_add_epi32(v[7], k__DCT_CONST_ROUNDING); + u[8] = _mm_add_epi32(v[8], k__DCT_CONST_ROUNDING); + u[9] = _mm_add_epi32(v[9], k__DCT_CONST_ROUNDING); + u[10] = _mm_add_epi32(v[10], k__DCT_CONST_ROUNDING); + u[11] = _mm_add_epi32(v[11], k__DCT_CONST_ROUNDING); + u[12] = _mm_add_epi32(v[12], k__DCT_CONST_ROUNDING); + u[13] = _mm_add_epi32(v[13], k__DCT_CONST_ROUNDING); + u[14] = _mm_add_epi32(v[14], k__DCT_CONST_ROUNDING); + u[15] = _mm_add_epi32(v[15], k__DCT_CONST_ROUNDING); + + u[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS); + u[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS); + u[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS); + u[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS); + u[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS); + u[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS); + u[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS); + u[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS); + u[8] = _mm_srai_epi32(u[8], DCT_CONST_BITS); + u[9] = _mm_srai_epi32(u[9], DCT_CONST_BITS); + u[10] = _mm_srai_epi32(u[10], DCT_CONST_BITS); + u[11] = _mm_srai_epi32(u[11], DCT_CONST_BITS); + u[12] = _mm_srai_epi32(u[12], DCT_CONST_BITS); + u[13] = _mm_srai_epi32(u[13], DCT_CONST_BITS); + u[14] = _mm_srai_epi32(u[14], DCT_CONST_BITS); + u[15] = _mm_srai_epi32(u[15], DCT_CONST_BITS); + + s[0] = _mm_packs_epi32(u[0], u[1]); + s[1] = _mm_packs_epi32(u[2], u[3]); + s[2] = _mm_packs_epi32(u[4], u[5]); + s[3] = _mm_packs_epi32(u[6], u[7]); + s[4] = _mm_add_epi16(t[4], t[5]); + s[5] = _mm_sub_epi16(t[4], t[5]); + s[6] = _mm_sub_epi16(t[7], t[6]); + s[7] = _mm_add_epi16(t[6], t[7]); + s[8] = t[8]; + s[15] = t[15]; + s[9] = _mm_packs_epi32(u[8], u[9]); + s[14] = _mm_packs_epi32(u[10], u[11]); + s[10] = _mm_packs_epi32(u[12], u[13]); + s[13] = _mm_packs_epi32(u[14], u[15]); + s[11] = t[11]; + s[12] = t[12]; + + // stage 5 + t[0] = _mm_add_epi16(s[0], s[3]); + t[1] = _mm_add_epi16(s[1], s[2]); + t[2] = _mm_sub_epi16(s[1], s[2]); + t[3] = _mm_sub_epi16(s[0], s[3]); + t[4] = s[4]; + t[7] = s[7]; + + u[0] = _mm_unpacklo_epi16(s[5], s[6]); + u[1] = _mm_unpackhi_epi16(s[5], s[6]); + v[0] = _mm_madd_epi16(u[0], k__cospi_m16_p16); + v[1] = _mm_madd_epi16(u[1], k__cospi_m16_p16); + v[2] = _mm_madd_epi16(u[0], k__cospi_p16_p16); + v[3] = _mm_madd_epi16(u[1], k__cospi_p16_p16); + u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING); + u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING); + u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING); + u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING); + u[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS); + u[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS); + u[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS); + u[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS); + t[5] = _mm_packs_epi32(u[0], u[1]); + t[6] = _mm_packs_epi32(u[2], u[3]); + + t[8] = _mm_add_epi16(s[8], s[11]); + t[9] = _mm_add_epi16(s[9], s[10]); + t[10] = _mm_sub_epi16(s[9], s[10]); + t[11] = _mm_sub_epi16(s[8], s[11]); + t[12] = _mm_sub_epi16(s[15], s[12]); + t[13] = _mm_sub_epi16(s[14], s[13]); + t[14] = _mm_add_epi16(s[13], s[14]); + t[15] = _mm_add_epi16(s[12], s[15]); + + // stage 6 + s[0] = _mm_add_epi16(t[0], t[7]); + s[1] = _mm_add_epi16(t[1], t[6]); + s[2] = _mm_add_epi16(t[2], t[5]); + s[3] = _mm_add_epi16(t[3], t[4]); + s[4] = _mm_sub_epi16(t[3], t[4]); + s[5] = _mm_sub_epi16(t[2], t[5]); + s[6] = _mm_sub_epi16(t[1], t[6]); + s[7] = _mm_sub_epi16(t[0], t[7]); + s[8] = t[8]; + s[9] = t[9]; + + u[0] = _mm_unpacklo_epi16(t[10], t[13]); + u[1] = _mm_unpackhi_epi16(t[10], t[13]); + u[2] = _mm_unpacklo_epi16(t[11], t[12]); + u[3] = _mm_unpackhi_epi16(t[11], t[12]); + + v[0] = _mm_madd_epi16(u[0], k__cospi_m16_p16); + v[1] = _mm_madd_epi16(u[1], k__cospi_m16_p16); + v[2] = _mm_madd_epi16(u[0], k__cospi_p16_p16); + v[3] = _mm_madd_epi16(u[1], k__cospi_p16_p16); + v[4] = _mm_madd_epi16(u[2], k__cospi_m16_p16); + v[5] = _mm_madd_epi16(u[3], k__cospi_m16_p16); + v[6] = _mm_madd_epi16(u[2], k__cospi_p16_p16); + v[7] = _mm_madd_epi16(u[3], k__cospi_p16_p16); + + u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING); + u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING); + u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING); + u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING); + u[4] = _mm_add_epi32(v[4], k__DCT_CONST_ROUNDING); + u[5] = _mm_add_epi32(v[5], k__DCT_CONST_ROUNDING); + u[6] = _mm_add_epi32(v[6], k__DCT_CONST_ROUNDING); + u[7] = _mm_add_epi32(v[7], k__DCT_CONST_ROUNDING); + + u[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS); + u[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS); + u[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS); + u[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS); + u[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS); + u[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS); + u[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS); + u[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS); + + s[10] = _mm_packs_epi32(u[0], u[1]); + s[13] = _mm_packs_epi32(u[2], u[3]); + s[11] = _mm_packs_epi32(u[4], u[5]); + s[12] = _mm_packs_epi32(u[6], u[7]); + s[14] = t[14]; + s[15] = t[15]; + + // stage 7 + in[0] = _mm_add_epi16(s[0], s[15]); + in[1] = _mm_add_epi16(s[1], s[14]); + in[2] = _mm_add_epi16(s[2], s[13]); + in[3] = _mm_add_epi16(s[3], s[12]); + in[4] = _mm_add_epi16(s[4], s[11]); + in[5] = _mm_add_epi16(s[5], s[10]); + in[6] = _mm_add_epi16(s[6], s[9]); + in[7] = _mm_add_epi16(s[7], s[8]); + in[8] = _mm_sub_epi16(s[7], s[8]); + in[9] = _mm_sub_epi16(s[6], s[9]); + in[10] = _mm_sub_epi16(s[5], s[10]); + in[11] = _mm_sub_epi16(s[4], s[11]); + in[12] = _mm_sub_epi16(s[3], s[12]); + in[13] = _mm_sub_epi16(s[2], s[13]); + in[14] = _mm_sub_epi16(s[1], s[14]); + in[15] = _mm_sub_epi16(s[0], s[15]); +} + +void aom_idct16_sse2(__m128i *in0, __m128i *in1) { + array_transpose_16x16(in0, in1); + idct16_8col(in0); + idct16_8col(in1); +} + +void aom_iadst16_sse2(__m128i *in0, __m128i *in1) { + array_transpose_16x16(in0, in1); + iadst16_8col(in0); + iadst16_8col(in1); +} + +void aom_idct16x16_10_add_sse2(const tran_low_t *input, uint8_t *dest, + int stride) { + const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING); + const __m128i final_rounding = _mm_set1_epi16(1 << 5); + const __m128i zero = _mm_setzero_si128(); + + const __m128i stg2_0 = pair_set_epi16(cospi_30_64, -cospi_2_64); + const __m128i stg2_1 = pair_set_epi16(cospi_2_64, cospi_30_64); + const __m128i stg2_6 = pair_set_epi16(cospi_6_64, -cospi_26_64); + const __m128i stg2_7 = pair_set_epi16(cospi_26_64, cospi_6_64); + + const __m128i stg3_0 = pair_set_epi16(cospi_28_64, -cospi_4_64); + const __m128i stg3_1 = pair_set_epi16(cospi_4_64, cospi_28_64); + + const __m128i stg4_0 = pair_set_epi16(cospi_16_64, cospi_16_64); + const __m128i stg4_1 = pair_set_epi16(cospi_16_64, -cospi_16_64); + const __m128i stg4_4 = pair_set_epi16(-cospi_8_64, cospi_24_64); + const __m128i stg4_5 = pair_set_epi16(cospi_24_64, cospi_8_64); + const __m128i stg4_6 = pair_set_epi16(-cospi_24_64, -cospi_8_64); + const __m128i stg4_7 = pair_set_epi16(-cospi_8_64, cospi_24_64); + + const __m128i stg6_0 = pair_set_epi16(-cospi_16_64, cospi_16_64); + __m128i in[16], l[16]; + __m128i stp1_0, stp1_1, stp1_2, stp1_3, stp1_4, stp1_5, stp1_6, stp1_8, + stp1_9, stp1_10, stp1_11, stp1_12, stp1_13, stp1_14, stp1_15, stp1_8_0, + stp1_12_0; + __m128i stp2_0, stp2_1, stp2_2, stp2_3, stp2_4, stp2_5, stp2_6, stp2_7, + stp2_8, stp2_9, stp2_10, stp2_11, stp2_12, stp2_13, stp2_14; + __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7; + int i; + // First 1-D inverse DCT + // Load input data. + in[0] = load_input_data(input); + in[1] = load_input_data(input + 8 * 2); + in[2] = load_input_data(input + 8 * 4); + in[3] = load_input_data(input + 8 * 6); + + TRANSPOSE_8X4(in[0], in[1], in[2], in[3], in[0], in[1]); + + // Stage2 + { + const __m128i lo_1_15 = _mm_unpackhi_epi16(in[0], zero); + const __m128i lo_13_3 = _mm_unpackhi_epi16(zero, in[1]); + + tmp0 = _mm_madd_epi16(lo_1_15, stg2_0); + tmp2 = _mm_madd_epi16(lo_1_15, stg2_1); + tmp5 = _mm_madd_epi16(lo_13_3, stg2_6); + tmp7 = _mm_madd_epi16(lo_13_3, stg2_7); + + tmp0 = _mm_add_epi32(tmp0, rounding); + tmp2 = _mm_add_epi32(tmp2, rounding); + tmp5 = _mm_add_epi32(tmp5, rounding); + tmp7 = _mm_add_epi32(tmp7, rounding); + + tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS); + tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS); + tmp5 = _mm_srai_epi32(tmp5, DCT_CONST_BITS); + tmp7 = _mm_srai_epi32(tmp7, DCT_CONST_BITS); + + stp2_8 = _mm_packs_epi32(tmp0, tmp2); + stp2_11 = _mm_packs_epi32(tmp5, tmp7); + } + + // Stage3 + { + const __m128i lo_2_14 = _mm_unpacklo_epi16(in[1], zero); + + tmp0 = _mm_madd_epi16(lo_2_14, stg3_0); + tmp2 = _mm_madd_epi16(lo_2_14, stg3_1); + + tmp0 = _mm_add_epi32(tmp0, rounding); + tmp2 = _mm_add_epi32(tmp2, rounding); + tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS); + tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS); + + stp1_13 = _mm_unpackhi_epi64(stp2_11, zero); + stp1_14 = _mm_unpackhi_epi64(stp2_8, zero); + + stp1_4 = _mm_packs_epi32(tmp0, tmp2); + } + + // Stage4 + { + const __m128i lo_0_8 = _mm_unpacklo_epi16(in[0], zero); + const __m128i lo_9_14 = _mm_unpacklo_epi16(stp2_8, stp1_14); + const __m128i lo_10_13 = _mm_unpacklo_epi16(stp2_11, stp1_13); + + tmp0 = _mm_madd_epi16(lo_0_8, stg4_0); + tmp2 = _mm_madd_epi16(lo_0_8, stg4_1); + tmp1 = _mm_madd_epi16(lo_9_14, stg4_4); + tmp3 = _mm_madd_epi16(lo_9_14, stg4_5); + tmp5 = _mm_madd_epi16(lo_10_13, stg4_6); + tmp7 = _mm_madd_epi16(lo_10_13, stg4_7); + + tmp0 = _mm_add_epi32(tmp0, rounding); + tmp2 = _mm_add_epi32(tmp2, rounding); + tmp1 = _mm_add_epi32(tmp1, rounding); + tmp3 = _mm_add_epi32(tmp3, rounding); + tmp5 = _mm_add_epi32(tmp5, rounding); + tmp7 = _mm_add_epi32(tmp7, rounding); + + tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS); + tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS); + tmp1 = _mm_srai_epi32(tmp1, DCT_CONST_BITS); + tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS); + tmp5 = _mm_srai_epi32(tmp5, DCT_CONST_BITS); + tmp7 = _mm_srai_epi32(tmp7, DCT_CONST_BITS); + + stp1_0 = _mm_packs_epi32(tmp0, tmp0); + stp1_1 = _mm_packs_epi32(tmp2, tmp2); + stp2_9 = _mm_packs_epi32(tmp1, tmp3); + stp2_10 = _mm_packs_epi32(tmp5, tmp7); + + stp2_6 = _mm_unpackhi_epi64(stp1_4, zero); + } + + // Stage5 and Stage6 + { + tmp0 = _mm_add_epi16(stp2_8, stp2_11); + tmp1 = _mm_sub_epi16(stp2_8, stp2_11); + tmp2 = _mm_add_epi16(stp2_9, stp2_10); + tmp3 = _mm_sub_epi16(stp2_9, stp2_10); + + stp1_9 = _mm_unpacklo_epi64(tmp2, zero); + stp1_10 = _mm_unpacklo_epi64(tmp3, zero); + stp1_8 = _mm_unpacklo_epi64(tmp0, zero); + stp1_11 = _mm_unpacklo_epi64(tmp1, zero); + + stp1_13 = _mm_unpackhi_epi64(tmp3, zero); + stp1_14 = _mm_unpackhi_epi64(tmp2, zero); + stp1_12 = _mm_unpackhi_epi64(tmp1, zero); + stp1_15 = _mm_unpackhi_epi64(tmp0, zero); + } + + // Stage6 + { + const __m128i lo_6_5 = _mm_unpacklo_epi16(stp2_6, stp1_4); + const __m128i lo_10_13 = _mm_unpacklo_epi16(stp1_10, stp1_13); + const __m128i lo_11_12 = _mm_unpacklo_epi16(stp1_11, stp1_12); + + tmp1 = _mm_madd_epi16(lo_6_5, stg4_1); + tmp3 = _mm_madd_epi16(lo_6_5, stg4_0); + tmp0 = _mm_madd_epi16(lo_10_13, stg6_0); + tmp2 = _mm_madd_epi16(lo_10_13, stg4_0); + tmp4 = _mm_madd_epi16(lo_11_12, stg6_0); + tmp6 = _mm_madd_epi16(lo_11_12, stg4_0); + + tmp1 = _mm_add_epi32(tmp1, rounding); + tmp3 = _mm_add_epi32(tmp3, rounding); + tmp0 = _mm_add_epi32(tmp0, rounding); + tmp2 = _mm_add_epi32(tmp2, rounding); + tmp4 = _mm_add_epi32(tmp4, rounding); + tmp6 = _mm_add_epi32(tmp6, rounding); + + tmp1 = _mm_srai_epi32(tmp1, DCT_CONST_BITS); + tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS); + tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS); + tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS); + tmp4 = _mm_srai_epi32(tmp4, DCT_CONST_BITS); + tmp6 = _mm_srai_epi32(tmp6, DCT_CONST_BITS); + + stp1_6 = _mm_packs_epi32(tmp3, tmp1); + + stp2_10 = _mm_packs_epi32(tmp0, zero); + stp2_13 = _mm_packs_epi32(tmp2, zero); + stp2_11 = _mm_packs_epi32(tmp4, zero); + stp2_12 = _mm_packs_epi32(tmp6, zero); + + tmp0 = _mm_add_epi16(stp1_0, stp1_4); + tmp1 = _mm_sub_epi16(stp1_0, stp1_4); + tmp2 = _mm_add_epi16(stp1_1, stp1_6); + tmp3 = _mm_sub_epi16(stp1_1, stp1_6); + + stp2_0 = _mm_unpackhi_epi64(tmp0, zero); + stp2_1 = _mm_unpacklo_epi64(tmp2, zero); + stp2_2 = _mm_unpackhi_epi64(tmp2, zero); + stp2_3 = _mm_unpacklo_epi64(tmp0, zero); + stp2_4 = _mm_unpacklo_epi64(tmp1, zero); + stp2_5 = _mm_unpackhi_epi64(tmp3, zero); + stp2_6 = _mm_unpacklo_epi64(tmp3, zero); + stp2_7 = _mm_unpackhi_epi64(tmp1, zero); + } + + // Stage7. Left 8x16 only. + l[0] = _mm_add_epi16(stp2_0, stp1_15); + l[1] = _mm_add_epi16(stp2_1, stp1_14); + l[2] = _mm_add_epi16(stp2_2, stp2_13); + l[3] = _mm_add_epi16(stp2_3, stp2_12); + l[4] = _mm_add_epi16(stp2_4, stp2_11); + l[5] = _mm_add_epi16(stp2_5, stp2_10); + l[6] = _mm_add_epi16(stp2_6, stp1_9); + l[7] = _mm_add_epi16(stp2_7, stp1_8); + l[8] = _mm_sub_epi16(stp2_7, stp1_8); + l[9] = _mm_sub_epi16(stp2_6, stp1_9); + l[10] = _mm_sub_epi16(stp2_5, stp2_10); + l[11] = _mm_sub_epi16(stp2_4, stp2_11); + l[12] = _mm_sub_epi16(stp2_3, stp2_12); + l[13] = _mm_sub_epi16(stp2_2, stp2_13); + l[14] = _mm_sub_epi16(stp2_1, stp1_14); + l[15] = _mm_sub_epi16(stp2_0, stp1_15); + + // Second 1-D inverse transform, performed per 8x16 block + for (i = 0; i < 2; i++) { + int j; + array_transpose_4X8(l + 8 * i, in); + + IDCT16_10 + + // Stage7 + in[0] = _mm_add_epi16(stp2_0, stp1_15); + in[1] = _mm_add_epi16(stp2_1, stp1_14); + in[2] = _mm_add_epi16(stp2_2, stp2_13); + in[3] = _mm_add_epi16(stp2_3, stp2_12); + in[4] = _mm_add_epi16(stp2_4, stp2_11); + in[5] = _mm_add_epi16(stp2_5, stp2_10); + in[6] = _mm_add_epi16(stp2_6, stp1_9); + in[7] = _mm_add_epi16(stp2_7, stp1_8); + in[8] = _mm_sub_epi16(stp2_7, stp1_8); + in[9] = _mm_sub_epi16(stp2_6, stp1_9); + in[10] = _mm_sub_epi16(stp2_5, stp2_10); + in[11] = _mm_sub_epi16(stp2_4, stp2_11); + in[12] = _mm_sub_epi16(stp2_3, stp2_12); + in[13] = _mm_sub_epi16(stp2_2, stp2_13); + in[14] = _mm_sub_epi16(stp2_1, stp1_14); + in[15] = _mm_sub_epi16(stp2_0, stp1_15); + + for (j = 0; j < 16; ++j) { + // Final rounding and shift + in[j] = _mm_adds_epi16(in[j], final_rounding); + in[j] = _mm_srai_epi16(in[j], 6); + RECON_AND_STORE(dest + j * stride, in[j]); + } + + dest += 8; + } +} + +#define LOAD_DQCOEFF(reg, input) \ + { \ + reg = load_input_data(input); \ + input += 8; \ + } + +#define IDCT32_34 \ + /* Stage1 */ \ + { \ + const __m128i lo_1_31 = _mm_unpacklo_epi16(in[1], zero); \ + const __m128i hi_1_31 = _mm_unpackhi_epi16(in[1], zero); \ + \ + const __m128i lo_25_7 = _mm_unpacklo_epi16(zero, in[7]); \ + const __m128i hi_25_7 = _mm_unpackhi_epi16(zero, in[7]); \ + \ + const __m128i lo_5_27 = _mm_unpacklo_epi16(in[5], zero); \ + const __m128i hi_5_27 = _mm_unpackhi_epi16(in[5], zero); \ + \ + const __m128i lo_29_3 = _mm_unpacklo_epi16(zero, in[3]); \ + const __m128i hi_29_3 = _mm_unpackhi_epi16(zero, in[3]); \ + \ + MULTIPLICATION_AND_ADD_2(lo_1_31, hi_1_31, stg1_0, stg1_1, stp1_16, \ + stp1_31); \ + MULTIPLICATION_AND_ADD_2(lo_25_7, hi_25_7, stg1_6, stg1_7, stp1_19, \ + stp1_28); \ + MULTIPLICATION_AND_ADD_2(lo_5_27, hi_5_27, stg1_8, stg1_9, stp1_20, \ + stp1_27); \ + MULTIPLICATION_AND_ADD_2(lo_29_3, hi_29_3, stg1_14, stg1_15, stp1_23, \ + stp1_24); \ + } \ + \ + /* Stage2 */ \ + { \ + const __m128i lo_2_30 = _mm_unpacklo_epi16(in[2], zero); \ + const __m128i hi_2_30 = _mm_unpackhi_epi16(in[2], zero); \ + \ + const __m128i lo_26_6 = _mm_unpacklo_epi16(zero, in[6]); \ + const __m128i hi_26_6 = _mm_unpackhi_epi16(zero, in[6]); \ + \ + MULTIPLICATION_AND_ADD_2(lo_2_30, hi_2_30, stg2_0, stg2_1, stp2_8, \ + stp2_15); \ + MULTIPLICATION_AND_ADD_2(lo_26_6, hi_26_6, stg2_6, stg2_7, stp2_11, \ + stp2_12); \ + \ + stp2_16 = stp1_16; \ + stp2_19 = stp1_19; \ + \ + stp2_20 = stp1_20; \ + stp2_23 = stp1_23; \ + \ + stp2_24 = stp1_24; \ + stp2_27 = stp1_27; \ + \ + stp2_28 = stp1_28; \ + stp2_31 = stp1_31; \ + } \ + \ + /* Stage3 */ \ + { \ + const __m128i lo_4_28 = _mm_unpacklo_epi16(in[4], zero); \ + const __m128i hi_4_28 = _mm_unpackhi_epi16(in[4], zero); \ + \ + const __m128i lo_17_30 = _mm_unpacklo_epi16(stp1_16, stp1_31); \ + const __m128i hi_17_30 = _mm_unpackhi_epi16(stp1_16, stp1_31); \ + const __m128i lo_18_29 = _mm_unpacklo_epi16(stp1_19, stp1_28); \ + const __m128i hi_18_29 = _mm_unpackhi_epi16(stp1_19, stp1_28); \ + \ + const __m128i lo_21_26 = _mm_unpacklo_epi16(stp1_20, stp1_27); \ + const __m128i hi_21_26 = _mm_unpackhi_epi16(stp1_20, stp1_27); \ + const __m128i lo_22_25 = _mm_unpacklo_epi16(stp1_23, stp1_24); \ + const __m128i hi_22_25 = _mm_unpackhi_epi16(stp1_23, stp2_24); \ + \ + MULTIPLICATION_AND_ADD_2(lo_4_28, hi_4_28, stg3_0, stg3_1, stp1_4, \ + stp1_7); \ + \ + stp1_8 = stp2_8; \ + stp1_11 = stp2_11; \ + stp1_12 = stp2_12; \ + stp1_15 = stp2_15; \ + \ + MULTIPLICATION_AND_ADD(lo_17_30, hi_17_30, lo_18_29, hi_18_29, stg3_4, \ + stg3_5, stg3_6, stg3_4, stp1_17, stp1_30, stp1_18, \ + stp1_29) \ + MULTIPLICATION_AND_ADD(lo_21_26, hi_21_26, lo_22_25, hi_22_25, stg3_8, \ + stg3_9, stg3_10, stg3_8, stp1_21, stp1_26, stp1_22, \ + stp1_25) \ + \ + stp1_16 = stp2_16; \ + stp1_31 = stp2_31; \ + stp1_19 = stp2_19; \ + stp1_20 = stp2_20; \ + stp1_23 = stp2_23; \ + stp1_24 = stp2_24; \ + stp1_27 = stp2_27; \ + stp1_28 = stp2_28; \ + } \ + \ + /* Stage4 */ \ + { \ + const __m128i lo_0_16 = _mm_unpacklo_epi16(in[0], zero); \ + const __m128i hi_0_16 = _mm_unpackhi_epi16(in[0], zero); \ + \ + const __m128i lo_9_14 = _mm_unpacklo_epi16(stp2_8, stp2_15); \ + const __m128i hi_9_14 = _mm_unpackhi_epi16(stp2_8, stp2_15); \ + const __m128i lo_10_13 = _mm_unpacklo_epi16(stp2_11, stp2_12); \ + const __m128i hi_10_13 = _mm_unpackhi_epi16(stp2_11, stp2_12); \ + \ + MULTIPLICATION_AND_ADD_2(lo_0_16, hi_0_16, stg4_0, stg4_1, stp2_0, \ + stp2_1); \ + \ + stp2_4 = stp1_4; \ + stp2_5 = stp1_4; \ + stp2_6 = stp1_7; \ + stp2_7 = stp1_7; \ + \ + MULTIPLICATION_AND_ADD(lo_9_14, hi_9_14, lo_10_13, hi_10_13, stg4_4, \ + stg4_5, stg4_6, stg4_4, stp2_9, stp2_14, stp2_10, \ + stp2_13) \ + \ + stp2_8 = stp1_8; \ + stp2_15 = stp1_15; \ + stp2_11 = stp1_11; \ + stp2_12 = stp1_12; \ + \ + stp2_16 = _mm_add_epi16(stp1_16, stp1_19); \ + stp2_17 = _mm_add_epi16(stp1_17, stp1_18); \ + stp2_18 = _mm_sub_epi16(stp1_17, stp1_18); \ + stp2_19 = _mm_sub_epi16(stp1_16, stp1_19); \ + stp2_20 = _mm_sub_epi16(stp1_23, stp1_20); \ + stp2_21 = _mm_sub_epi16(stp1_22, stp1_21); \ + stp2_22 = _mm_add_epi16(stp1_22, stp1_21); \ + stp2_23 = _mm_add_epi16(stp1_23, stp1_20); \ + \ + stp2_24 = _mm_add_epi16(stp1_24, stp1_27); \ + stp2_25 = _mm_add_epi16(stp1_25, stp1_26); \ + stp2_26 = _mm_sub_epi16(stp1_25, stp1_26); \ + stp2_27 = _mm_sub_epi16(stp1_24, stp1_27); \ + stp2_28 = _mm_sub_epi16(stp1_31, stp1_28); \ + stp2_29 = _mm_sub_epi16(stp1_30, stp1_29); \ + stp2_30 = _mm_add_epi16(stp1_29, stp1_30); \ + stp2_31 = _mm_add_epi16(stp1_28, stp1_31); \ + } \ + \ + /* Stage5 */ \ + { \ + const __m128i lo_6_5 = _mm_unpacklo_epi16(stp2_6, stp2_5); \ + const __m128i hi_6_5 = _mm_unpackhi_epi16(stp2_6, stp2_5); \ + const __m128i lo_18_29 = _mm_unpacklo_epi16(stp2_18, stp2_29); \ + const __m128i hi_18_29 = _mm_unpackhi_epi16(stp2_18, stp2_29); \ + \ + const __m128i lo_19_28 = _mm_unpacklo_epi16(stp2_19, stp2_28); \ + const __m128i hi_19_28 = _mm_unpackhi_epi16(stp2_19, stp2_28); \ + const __m128i lo_20_27 = _mm_unpacklo_epi16(stp2_20, stp2_27); \ + const __m128i hi_20_27 = _mm_unpackhi_epi16(stp2_20, stp2_27); \ + \ + const __m128i lo_21_26 = _mm_unpacklo_epi16(stp2_21, stp2_26); \ + const __m128i hi_21_26 = _mm_unpackhi_epi16(stp2_21, stp2_26); \ + \ + stp1_0 = stp2_0; \ + stp1_1 = stp2_1; \ + stp1_2 = stp2_1; \ + stp1_3 = stp2_0; \ + \ + tmp0 = _mm_madd_epi16(lo_6_5, stg4_1); \ + tmp1 = _mm_madd_epi16(hi_6_5, stg4_1); \ + tmp2 = _mm_madd_epi16(lo_6_5, stg4_0); \ + tmp3 = _mm_madd_epi16(hi_6_5, stg4_0); \ + \ + tmp0 = _mm_add_epi32(tmp0, rounding); \ + tmp1 = _mm_add_epi32(tmp1, rounding); \ + tmp2 = _mm_add_epi32(tmp2, rounding); \ + tmp3 = _mm_add_epi32(tmp3, rounding); \ + \ + tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS); \ + tmp1 = _mm_srai_epi32(tmp1, DCT_CONST_BITS); \ + tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS); \ + tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS); \ + \ + stp1_5 = _mm_packs_epi32(tmp0, tmp1); \ + stp1_6 = _mm_packs_epi32(tmp2, tmp3); \ + \ + stp1_4 = stp2_4; \ + stp1_7 = stp2_7; \ + \ + stp1_8 = _mm_add_epi16(stp2_8, stp2_11); \ + stp1_9 = _mm_add_epi16(stp2_9, stp2_10); \ + stp1_10 = _mm_sub_epi16(stp2_9, stp2_10); \ + stp1_11 = _mm_sub_epi16(stp2_8, stp2_11); \ + stp1_12 = _mm_sub_epi16(stp2_15, stp2_12); \ + stp1_13 = _mm_sub_epi16(stp2_14, stp2_13); \ + stp1_14 = _mm_add_epi16(stp2_14, stp2_13); \ + stp1_15 = _mm_add_epi16(stp2_15, stp2_12); \ + \ + stp1_16 = stp2_16; \ + stp1_17 = stp2_17; \ + \ + MULTIPLICATION_AND_ADD(lo_18_29, hi_18_29, lo_19_28, hi_19_28, stg4_4, \ + stg4_5, stg4_4, stg4_5, stp1_18, stp1_29, stp1_19, \ + stp1_28) \ + MULTIPLICATION_AND_ADD(lo_20_27, hi_20_27, lo_21_26, hi_21_26, stg4_6, \ + stg4_4, stg4_6, stg4_4, stp1_20, stp1_27, stp1_21, \ + stp1_26) \ + \ + stp1_22 = stp2_22; \ + stp1_23 = stp2_23; \ + stp1_24 = stp2_24; \ + stp1_25 = stp2_25; \ + stp1_30 = stp2_30; \ + stp1_31 = stp2_31; \ + } \ + \ + /* Stage6 */ \ + { \ + const __m128i lo_10_13 = _mm_unpacklo_epi16(stp1_10, stp1_13); \ + const __m128i hi_10_13 = _mm_unpackhi_epi16(stp1_10, stp1_13); \ + const __m128i lo_11_12 = _mm_unpacklo_epi16(stp1_11, stp1_12); \ + const __m128i hi_11_12 = _mm_unpackhi_epi16(stp1_11, stp1_12); \ + \ + stp2_0 = _mm_add_epi16(stp1_0, stp1_7); \ + stp2_1 = _mm_add_epi16(stp1_1, stp1_6); \ + stp2_2 = _mm_add_epi16(stp1_2, stp1_5); \ + stp2_3 = _mm_add_epi16(stp1_3, stp1_4); \ + stp2_4 = _mm_sub_epi16(stp1_3, stp1_4); \ + stp2_5 = _mm_sub_epi16(stp1_2, stp1_5); \ + stp2_6 = _mm_sub_epi16(stp1_1, stp1_6); \ + stp2_7 = _mm_sub_epi16(stp1_0, stp1_7); \ + \ + stp2_8 = stp1_8; \ + stp2_9 = stp1_9; \ + stp2_14 = stp1_14; \ + stp2_15 = stp1_15; \ + \ + MULTIPLICATION_AND_ADD(lo_10_13, hi_10_13, lo_11_12, hi_11_12, stg6_0, \ + stg4_0, stg6_0, stg4_0, stp2_10, stp2_13, stp2_11, \ + stp2_12) \ + \ + stp2_16 = _mm_add_epi16(stp1_16, stp1_23); \ + stp2_17 = _mm_add_epi16(stp1_17, stp1_22); \ + stp2_18 = _mm_add_epi16(stp1_18, stp1_21); \ + stp2_19 = _mm_add_epi16(stp1_19, stp1_20); \ + stp2_20 = _mm_sub_epi16(stp1_19, stp1_20); \ + stp2_21 = _mm_sub_epi16(stp1_18, stp1_21); \ + stp2_22 = _mm_sub_epi16(stp1_17, stp1_22); \ + stp2_23 = _mm_sub_epi16(stp1_16, stp1_23); \ + \ + stp2_24 = _mm_sub_epi16(stp1_31, stp1_24); \ + stp2_25 = _mm_sub_epi16(stp1_30, stp1_25); \ + stp2_26 = _mm_sub_epi16(stp1_29, stp1_26); \ + stp2_27 = _mm_sub_epi16(stp1_28, stp1_27); \ + stp2_28 = _mm_add_epi16(stp1_27, stp1_28); \ + stp2_29 = _mm_add_epi16(stp1_26, stp1_29); \ + stp2_30 = _mm_add_epi16(stp1_25, stp1_30); \ + stp2_31 = _mm_add_epi16(stp1_24, stp1_31); \ + } \ + \ + /* Stage7 */ \ + { \ + const __m128i lo_20_27 = _mm_unpacklo_epi16(stp2_20, stp2_27); \ + const __m128i hi_20_27 = _mm_unpackhi_epi16(stp2_20, stp2_27); \ + const __m128i lo_21_26 = _mm_unpacklo_epi16(stp2_21, stp2_26); \ + const __m128i hi_21_26 = _mm_unpackhi_epi16(stp2_21, stp2_26); \ + \ + const __m128i lo_22_25 = _mm_unpacklo_epi16(stp2_22, stp2_25); \ + const __m128i hi_22_25 = _mm_unpackhi_epi16(stp2_22, stp2_25); \ + const __m128i lo_23_24 = _mm_unpacklo_epi16(stp2_23, stp2_24); \ + const __m128i hi_23_24 = _mm_unpackhi_epi16(stp2_23, stp2_24); \ + \ + stp1_0 = _mm_add_epi16(stp2_0, stp2_15); \ + stp1_1 = _mm_add_epi16(stp2_1, stp2_14); \ + stp1_2 = _mm_add_epi16(stp2_2, stp2_13); \ + stp1_3 = _mm_add_epi16(stp2_3, stp2_12); \ + stp1_4 = _mm_add_epi16(stp2_4, stp2_11); \ + stp1_5 = _mm_add_epi16(stp2_5, stp2_10); \ + stp1_6 = _mm_add_epi16(stp2_6, stp2_9); \ + stp1_7 = _mm_add_epi16(stp2_7, stp2_8); \ + stp1_8 = _mm_sub_epi16(stp2_7, stp2_8); \ + stp1_9 = _mm_sub_epi16(stp2_6, stp2_9); \ + stp1_10 = _mm_sub_epi16(stp2_5, stp2_10); \ + stp1_11 = _mm_sub_epi16(stp2_4, stp2_11); \ + stp1_12 = _mm_sub_epi16(stp2_3, stp2_12); \ + stp1_13 = _mm_sub_epi16(stp2_2, stp2_13); \ + stp1_14 = _mm_sub_epi16(stp2_1, stp2_14); \ + stp1_15 = _mm_sub_epi16(stp2_0, stp2_15); \ + \ + stp1_16 = stp2_16; \ + stp1_17 = stp2_17; \ + stp1_18 = stp2_18; \ + stp1_19 = stp2_19; \ + \ + MULTIPLICATION_AND_ADD(lo_20_27, hi_20_27, lo_21_26, hi_21_26, stg6_0, \ + stg4_0, stg6_0, stg4_0, stp1_20, stp1_27, stp1_21, \ + stp1_26) \ + MULTIPLICATION_AND_ADD(lo_22_25, hi_22_25, lo_23_24, hi_23_24, stg6_0, \ + stg4_0, stg6_0, stg4_0, stp1_22, stp1_25, stp1_23, \ + stp1_24) \ + \ + stp1_28 = stp2_28; \ + stp1_29 = stp2_29; \ + stp1_30 = stp2_30; \ + stp1_31 = stp2_31; \ + } + +#define IDCT32(in0, in1) \ + /* Stage1 */ \ + { \ + const __m128i lo_1_31 = _mm_unpacklo_epi16((in0)[1], (in1)[15]); \ + const __m128i hi_1_31 = _mm_unpackhi_epi16((in0)[1], (in1)[15]); \ + const __m128i lo_17_15 = _mm_unpacklo_epi16((in1)[1], (in0)[15]); \ + const __m128i hi_17_15 = _mm_unpackhi_epi16((in1)[1], (in0)[15]); \ + \ + const __m128i lo_9_23 = _mm_unpacklo_epi16((in0)[9], (in1)[7]); \ + const __m128i hi_9_23 = _mm_unpackhi_epi16((in0)[9], (in1)[7]); \ + const __m128i lo_25_7 = _mm_unpacklo_epi16((in1)[9], (in0)[7]); \ + const __m128i hi_25_7 = _mm_unpackhi_epi16((in1)[9], (in0)[7]); \ + \ + const __m128i lo_5_27 = _mm_unpacklo_epi16((in0)[5], (in1)[11]); \ + const __m128i hi_5_27 = _mm_unpackhi_epi16((in0)[5], (in1)[11]); \ + const __m128i lo_21_11 = _mm_unpacklo_epi16((in1)[5], (in0)[11]); \ + const __m128i hi_21_11 = _mm_unpackhi_epi16((in1)[5], (in0)[11]); \ + \ + const __m128i lo_13_19 = _mm_unpacklo_epi16((in0)[13], (in1)[3]); \ + const __m128i hi_13_19 = _mm_unpackhi_epi16((in0)[13], (in1)[3]); \ + const __m128i lo_29_3 = _mm_unpacklo_epi16((in1)[13], (in0)[3]); \ + const __m128i hi_29_3 = _mm_unpackhi_epi16((in1)[13], (in0)[3]); \ + \ + MULTIPLICATION_AND_ADD(lo_1_31, hi_1_31, lo_17_15, hi_17_15, stg1_0, \ + stg1_1, stg1_2, stg1_3, stp1_16, stp1_31, stp1_17, \ + stp1_30) \ + MULTIPLICATION_AND_ADD(lo_9_23, hi_9_23, lo_25_7, hi_25_7, stg1_4, stg1_5, \ + stg1_6, stg1_7, stp1_18, stp1_29, stp1_19, stp1_28) \ + MULTIPLICATION_AND_ADD(lo_5_27, hi_5_27, lo_21_11, hi_21_11, stg1_8, \ + stg1_9, stg1_10, stg1_11, stp1_20, stp1_27, \ + stp1_21, stp1_26) \ + MULTIPLICATION_AND_ADD(lo_13_19, hi_13_19, lo_29_3, hi_29_3, stg1_12, \ + stg1_13, stg1_14, stg1_15, stp1_22, stp1_25, \ + stp1_23, stp1_24) \ + } \ + \ + /* Stage2 */ \ + { \ + const __m128i lo_2_30 = _mm_unpacklo_epi16((in0)[2], (in1)[14]); \ + const __m128i hi_2_30 = _mm_unpackhi_epi16((in0)[2], (in1)[14]); \ + const __m128i lo_18_14 = _mm_unpacklo_epi16((in1)[2], (in0)[14]); \ + const __m128i hi_18_14 = _mm_unpackhi_epi16((in1)[2], (in0)[14]); \ + \ + const __m128i lo_10_22 = _mm_unpacklo_epi16((in0)[10], (in1)[6]); \ + const __m128i hi_10_22 = _mm_unpackhi_epi16((in0)[10], (in1)[6]); \ + const __m128i lo_26_6 = _mm_unpacklo_epi16((in1)[10], (in0)[6]); \ + const __m128i hi_26_6 = _mm_unpackhi_epi16((in1)[10], (in0)[6]); \ + \ + MULTIPLICATION_AND_ADD(lo_2_30, hi_2_30, lo_18_14, hi_18_14, stg2_0, \ + stg2_1, stg2_2, stg2_3, stp2_8, stp2_15, stp2_9, \ + stp2_14) \ + MULTIPLICATION_AND_ADD(lo_10_22, hi_10_22, lo_26_6, hi_26_6, stg2_4, \ + stg2_5, stg2_6, stg2_7, stp2_10, stp2_13, stp2_11, \ + stp2_12) \ + \ + stp2_16 = _mm_add_epi16(stp1_16, stp1_17); \ + stp2_17 = _mm_sub_epi16(stp1_16, stp1_17); \ + stp2_18 = _mm_sub_epi16(stp1_19, stp1_18); \ + stp2_19 = _mm_add_epi16(stp1_19, stp1_18); \ + \ + stp2_20 = _mm_add_epi16(stp1_20, stp1_21); \ + stp2_21 = _mm_sub_epi16(stp1_20, stp1_21); \ + stp2_22 = _mm_sub_epi16(stp1_23, stp1_22); \ + stp2_23 = _mm_add_epi16(stp1_23, stp1_22); \ + \ + stp2_24 = _mm_add_epi16(stp1_24, stp1_25); \ + stp2_25 = _mm_sub_epi16(stp1_24, stp1_25); \ + stp2_26 = _mm_sub_epi16(stp1_27, stp1_26); \ + stp2_27 = _mm_add_epi16(stp1_27, stp1_26); \ + \ + stp2_28 = _mm_add_epi16(stp1_28, stp1_29); \ + stp2_29 = _mm_sub_epi16(stp1_28, stp1_29); \ + stp2_30 = _mm_sub_epi16(stp1_31, stp1_30); \ + stp2_31 = _mm_add_epi16(stp1_31, stp1_30); \ + } \ + \ + /* Stage3 */ \ + { \ + const __m128i lo_4_28 = _mm_unpacklo_epi16((in0)[4], (in1)[12]); \ + const __m128i hi_4_28 = _mm_unpackhi_epi16((in0)[4], (in1)[12]); \ + const __m128i lo_20_12 = _mm_unpacklo_epi16((in1)[4], (in0)[12]); \ + const __m128i hi_20_12 = _mm_unpackhi_epi16((in1)[4], (in0)[12]); \ + \ + const __m128i lo_17_30 = _mm_unpacklo_epi16(stp2_17, stp2_30); \ + const __m128i hi_17_30 = _mm_unpackhi_epi16(stp2_17, stp2_30); \ + const __m128i lo_18_29 = _mm_unpacklo_epi16(stp2_18, stp2_29); \ + const __m128i hi_18_29 = _mm_unpackhi_epi16(stp2_18, stp2_29); \ + \ + const __m128i lo_21_26 = _mm_unpacklo_epi16(stp2_21, stp2_26); \ + const __m128i hi_21_26 = _mm_unpackhi_epi16(stp2_21, stp2_26); \ + const __m128i lo_22_25 = _mm_unpacklo_epi16(stp2_22, stp2_25); \ + const __m128i hi_22_25 = _mm_unpackhi_epi16(stp2_22, stp2_25); \ + \ + MULTIPLICATION_AND_ADD(lo_4_28, hi_4_28, lo_20_12, hi_20_12, stg3_0, \ + stg3_1, stg3_2, stg3_3, stp1_4, stp1_7, stp1_5, \ + stp1_6) \ + \ + stp1_8 = _mm_add_epi16(stp2_8, stp2_9); \ + stp1_9 = _mm_sub_epi16(stp2_8, stp2_9); \ + stp1_10 = _mm_sub_epi16(stp2_11, stp2_10); \ + stp1_11 = _mm_add_epi16(stp2_11, stp2_10); \ + stp1_12 = _mm_add_epi16(stp2_12, stp2_13); \ + stp1_13 = _mm_sub_epi16(stp2_12, stp2_13); \ + stp1_14 = _mm_sub_epi16(stp2_15, stp2_14); \ + stp1_15 = _mm_add_epi16(stp2_15, stp2_14); \ + \ + MULTIPLICATION_AND_ADD(lo_17_30, hi_17_30, lo_18_29, hi_18_29, stg3_4, \ + stg3_5, stg3_6, stg3_4, stp1_17, stp1_30, stp1_18, \ + stp1_29) \ + MULTIPLICATION_AND_ADD(lo_21_26, hi_21_26, lo_22_25, hi_22_25, stg3_8, \ + stg3_9, stg3_10, stg3_8, stp1_21, stp1_26, stp1_22, \ + stp1_25) \ + \ + stp1_16 = stp2_16; \ + stp1_31 = stp2_31; \ + stp1_19 = stp2_19; \ + stp1_20 = stp2_20; \ + stp1_23 = stp2_23; \ + stp1_24 = stp2_24; \ + stp1_27 = stp2_27; \ + stp1_28 = stp2_28; \ + } \ + \ + /* Stage4 */ \ + { \ + const __m128i lo_0_16 = _mm_unpacklo_epi16((in0)[0], (in1)[0]); \ + const __m128i hi_0_16 = _mm_unpackhi_epi16((in0)[0], (in1)[0]); \ + const __m128i lo_8_24 = _mm_unpacklo_epi16((in0)[8], (in1)[8]); \ + const __m128i hi_8_24 = _mm_unpackhi_epi16((in0)[8], (in1)[8]); \ + \ + const __m128i lo_9_14 = _mm_unpacklo_epi16(stp1_9, stp1_14); \ + const __m128i hi_9_14 = _mm_unpackhi_epi16(stp1_9, stp1_14); \ + const __m128i lo_10_13 = _mm_unpacklo_epi16(stp1_10, stp1_13); \ + const __m128i hi_10_13 = _mm_unpackhi_epi16(stp1_10, stp1_13); \ + \ + MULTIPLICATION_AND_ADD(lo_0_16, hi_0_16, lo_8_24, hi_8_24, stg4_0, stg4_1, \ + stg4_2, stg4_3, stp2_0, stp2_1, stp2_2, stp2_3) \ + \ + stp2_4 = _mm_add_epi16(stp1_4, stp1_5); \ + stp2_5 = _mm_sub_epi16(stp1_4, stp1_5); \ + stp2_6 = _mm_sub_epi16(stp1_7, stp1_6); \ + stp2_7 = _mm_add_epi16(stp1_7, stp1_6); \ + \ + MULTIPLICATION_AND_ADD(lo_9_14, hi_9_14, lo_10_13, hi_10_13, stg4_4, \ + stg4_5, stg4_6, stg4_4, stp2_9, stp2_14, stp2_10, \ + stp2_13) \ + \ + stp2_8 = stp1_8; \ + stp2_15 = stp1_15; \ + stp2_11 = stp1_11; \ + stp2_12 = stp1_12; \ + \ + stp2_16 = _mm_add_epi16(stp1_16, stp1_19); \ + stp2_17 = _mm_add_epi16(stp1_17, stp1_18); \ + stp2_18 = _mm_sub_epi16(stp1_17, stp1_18); \ + stp2_19 = _mm_sub_epi16(stp1_16, stp1_19); \ + stp2_20 = _mm_sub_epi16(stp1_23, stp1_20); \ + stp2_21 = _mm_sub_epi16(stp1_22, stp1_21); \ + stp2_22 = _mm_add_epi16(stp1_22, stp1_21); \ + stp2_23 = _mm_add_epi16(stp1_23, stp1_20); \ + \ + stp2_24 = _mm_add_epi16(stp1_24, stp1_27); \ + stp2_25 = _mm_add_epi16(stp1_25, stp1_26); \ + stp2_26 = _mm_sub_epi16(stp1_25, stp1_26); \ + stp2_27 = _mm_sub_epi16(stp1_24, stp1_27); \ + stp2_28 = _mm_sub_epi16(stp1_31, stp1_28); \ + stp2_29 = _mm_sub_epi16(stp1_30, stp1_29); \ + stp2_30 = _mm_add_epi16(stp1_29, stp1_30); \ + stp2_31 = _mm_add_epi16(stp1_28, stp1_31); \ + } \ + \ + /* Stage5 */ \ + { \ + const __m128i lo_6_5 = _mm_unpacklo_epi16(stp2_6, stp2_5); \ + const __m128i hi_6_5 = _mm_unpackhi_epi16(stp2_6, stp2_5); \ + const __m128i lo_18_29 = _mm_unpacklo_epi16(stp2_18, stp2_29); \ + const __m128i hi_18_29 = _mm_unpackhi_epi16(stp2_18, stp2_29); \ + \ + const __m128i lo_19_28 = _mm_unpacklo_epi16(stp2_19, stp2_28); \ + const __m128i hi_19_28 = _mm_unpackhi_epi16(stp2_19, stp2_28); \ + const __m128i lo_20_27 = _mm_unpacklo_epi16(stp2_20, stp2_27); \ + const __m128i hi_20_27 = _mm_unpackhi_epi16(stp2_20, stp2_27); \ + \ + const __m128i lo_21_26 = _mm_unpacklo_epi16(stp2_21, stp2_26); \ + const __m128i hi_21_26 = _mm_unpackhi_epi16(stp2_21, stp2_26); \ + \ + stp1_0 = _mm_add_epi16(stp2_0, stp2_3); \ + stp1_1 = _mm_add_epi16(stp2_1, stp2_2); \ + stp1_2 = _mm_sub_epi16(stp2_1, stp2_2); \ + stp1_3 = _mm_sub_epi16(stp2_0, stp2_3); \ + \ + tmp0 = _mm_madd_epi16(lo_6_5, stg4_1); \ + tmp1 = _mm_madd_epi16(hi_6_5, stg4_1); \ + tmp2 = _mm_madd_epi16(lo_6_5, stg4_0); \ + tmp3 = _mm_madd_epi16(hi_6_5, stg4_0); \ + \ + tmp0 = _mm_add_epi32(tmp0, rounding); \ + tmp1 = _mm_add_epi32(tmp1, rounding); \ + tmp2 = _mm_add_epi32(tmp2, rounding); \ + tmp3 = _mm_add_epi32(tmp3, rounding); \ + \ + tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS); \ + tmp1 = _mm_srai_epi32(tmp1, DCT_CONST_BITS); \ + tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS); \ + tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS); \ + \ + stp1_5 = _mm_packs_epi32(tmp0, tmp1); \ + stp1_6 = _mm_packs_epi32(tmp2, tmp3); \ + \ + stp1_4 = stp2_4; \ + stp1_7 = stp2_7; \ + \ + stp1_8 = _mm_add_epi16(stp2_8, stp2_11); \ + stp1_9 = _mm_add_epi16(stp2_9, stp2_10); \ + stp1_10 = _mm_sub_epi16(stp2_9, stp2_10); \ + stp1_11 = _mm_sub_epi16(stp2_8, stp2_11); \ + stp1_12 = _mm_sub_epi16(stp2_15, stp2_12); \ + stp1_13 = _mm_sub_epi16(stp2_14, stp2_13); \ + stp1_14 = _mm_add_epi16(stp2_14, stp2_13); \ + stp1_15 = _mm_add_epi16(stp2_15, stp2_12); \ + \ + stp1_16 = stp2_16; \ + stp1_17 = stp2_17; \ + \ + MULTIPLICATION_AND_ADD(lo_18_29, hi_18_29, lo_19_28, hi_19_28, stg4_4, \ + stg4_5, stg4_4, stg4_5, stp1_18, stp1_29, stp1_19, \ + stp1_28) \ + MULTIPLICATION_AND_ADD(lo_20_27, hi_20_27, lo_21_26, hi_21_26, stg4_6, \ + stg4_4, stg4_6, stg4_4, stp1_20, stp1_27, stp1_21, \ + stp1_26) \ + \ + stp1_22 = stp2_22; \ + stp1_23 = stp2_23; \ + stp1_24 = stp2_24; \ + stp1_25 = stp2_25; \ + stp1_30 = stp2_30; \ + stp1_31 = stp2_31; \ + } \ + \ + /* Stage6 */ \ + { \ + const __m128i lo_10_13 = _mm_unpacklo_epi16(stp1_10, stp1_13); \ + const __m128i hi_10_13 = _mm_unpackhi_epi16(stp1_10, stp1_13); \ + const __m128i lo_11_12 = _mm_unpacklo_epi16(stp1_11, stp1_12); \ + const __m128i hi_11_12 = _mm_unpackhi_epi16(stp1_11, stp1_12); \ + \ + stp2_0 = _mm_add_epi16(stp1_0, stp1_7); \ + stp2_1 = _mm_add_epi16(stp1_1, stp1_6); \ + stp2_2 = _mm_add_epi16(stp1_2, stp1_5); \ + stp2_3 = _mm_add_epi16(stp1_3, stp1_4); \ + stp2_4 = _mm_sub_epi16(stp1_3, stp1_4); \ + stp2_5 = _mm_sub_epi16(stp1_2, stp1_5); \ + stp2_6 = _mm_sub_epi16(stp1_1, stp1_6); \ + stp2_7 = _mm_sub_epi16(stp1_0, stp1_7); \ + \ + stp2_8 = stp1_8; \ + stp2_9 = stp1_9; \ + stp2_14 = stp1_14; \ + stp2_15 = stp1_15; \ + \ + MULTIPLICATION_AND_ADD(lo_10_13, hi_10_13, lo_11_12, hi_11_12, stg6_0, \ + stg4_0, stg6_0, stg4_0, stp2_10, stp2_13, stp2_11, \ + stp2_12) \ + \ + stp2_16 = _mm_add_epi16(stp1_16, stp1_23); \ + stp2_17 = _mm_add_epi16(stp1_17, stp1_22); \ + stp2_18 = _mm_add_epi16(stp1_18, stp1_21); \ + stp2_19 = _mm_add_epi16(stp1_19, stp1_20); \ + stp2_20 = _mm_sub_epi16(stp1_19, stp1_20); \ + stp2_21 = _mm_sub_epi16(stp1_18, stp1_21); \ + stp2_22 = _mm_sub_epi16(stp1_17, stp1_22); \ + stp2_23 = _mm_sub_epi16(stp1_16, stp1_23); \ + \ + stp2_24 = _mm_sub_epi16(stp1_31, stp1_24); \ + stp2_25 = _mm_sub_epi16(stp1_30, stp1_25); \ + stp2_26 = _mm_sub_epi16(stp1_29, stp1_26); \ + stp2_27 = _mm_sub_epi16(stp1_28, stp1_27); \ + stp2_28 = _mm_add_epi16(stp1_27, stp1_28); \ + stp2_29 = _mm_add_epi16(stp1_26, stp1_29); \ + stp2_30 = _mm_add_epi16(stp1_25, stp1_30); \ + stp2_31 = _mm_add_epi16(stp1_24, stp1_31); \ + } \ + \ + /* Stage7 */ \ + { \ + const __m128i lo_20_27 = _mm_unpacklo_epi16(stp2_20, stp2_27); \ + const __m128i hi_20_27 = _mm_unpackhi_epi16(stp2_20, stp2_27); \ + const __m128i lo_21_26 = _mm_unpacklo_epi16(stp2_21, stp2_26); \ + const __m128i hi_21_26 = _mm_unpackhi_epi16(stp2_21, stp2_26); \ + \ + const __m128i lo_22_25 = _mm_unpacklo_epi16(stp2_22, stp2_25); \ + const __m128i hi_22_25 = _mm_unpackhi_epi16(stp2_22, stp2_25); \ + const __m128i lo_23_24 = _mm_unpacklo_epi16(stp2_23, stp2_24); \ + const __m128i hi_23_24 = _mm_unpackhi_epi16(stp2_23, stp2_24); \ + \ + stp1_0 = _mm_add_epi16(stp2_0, stp2_15); \ + stp1_1 = _mm_add_epi16(stp2_1, stp2_14); \ + stp1_2 = _mm_add_epi16(stp2_2, stp2_13); \ + stp1_3 = _mm_add_epi16(stp2_3, stp2_12); \ + stp1_4 = _mm_add_epi16(stp2_4, stp2_11); \ + stp1_5 = _mm_add_epi16(stp2_5, stp2_10); \ + stp1_6 = _mm_add_epi16(stp2_6, stp2_9); \ + stp1_7 = _mm_add_epi16(stp2_7, stp2_8); \ + stp1_8 = _mm_sub_epi16(stp2_7, stp2_8); \ + stp1_9 = _mm_sub_epi16(stp2_6, stp2_9); \ + stp1_10 = _mm_sub_epi16(stp2_5, stp2_10); \ + stp1_11 = _mm_sub_epi16(stp2_4, stp2_11); \ + stp1_12 = _mm_sub_epi16(stp2_3, stp2_12); \ + stp1_13 = _mm_sub_epi16(stp2_2, stp2_13); \ + stp1_14 = _mm_sub_epi16(stp2_1, stp2_14); \ + stp1_15 = _mm_sub_epi16(stp2_0, stp2_15); \ + \ + stp1_16 = stp2_16; \ + stp1_17 = stp2_17; \ + stp1_18 = stp2_18; \ + stp1_19 = stp2_19; \ + \ + MULTIPLICATION_AND_ADD(lo_20_27, hi_20_27, lo_21_26, hi_21_26, stg6_0, \ + stg4_0, stg6_0, stg4_0, stp1_20, stp1_27, stp1_21, \ + stp1_26) \ + MULTIPLICATION_AND_ADD(lo_22_25, hi_22_25, lo_23_24, hi_23_24, stg6_0, \ + stg4_0, stg6_0, stg4_0, stp1_22, stp1_25, stp1_23, \ + stp1_24) \ + \ + stp1_28 = stp2_28; \ + stp1_29 = stp2_29; \ + stp1_30 = stp2_30; \ + stp1_31 = stp2_31; \ + } + +// Only upper-left 8x8 has non-zero coeff +void aom_idct32x32_34_add_sse2(const tran_low_t *input, uint8_t *dest, + int stride) { + const __m128i zero = _mm_setzero_si128(); + const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING); + const __m128i final_rounding = _mm_set1_epi16(1 << 5); + + // idct constants for each stage + const __m128i stg1_0 = pair_set_epi16(cospi_31_64, -cospi_1_64); + const __m128i stg1_1 = pair_set_epi16(cospi_1_64, cospi_31_64); + const __m128i stg1_6 = pair_set_epi16(cospi_7_64, -cospi_25_64); + const __m128i stg1_7 = pair_set_epi16(cospi_25_64, cospi_7_64); + const __m128i stg1_8 = pair_set_epi16(cospi_27_64, -cospi_5_64); + const __m128i stg1_9 = pair_set_epi16(cospi_5_64, cospi_27_64); + const __m128i stg1_14 = pair_set_epi16(cospi_3_64, -cospi_29_64); + const __m128i stg1_15 = pair_set_epi16(cospi_29_64, cospi_3_64); + + const __m128i stg2_0 = pair_set_epi16(cospi_30_64, -cospi_2_64); + const __m128i stg2_1 = pair_set_epi16(cospi_2_64, cospi_30_64); + const __m128i stg2_6 = pair_set_epi16(cospi_6_64, -cospi_26_64); + const __m128i stg2_7 = pair_set_epi16(cospi_26_64, cospi_6_64); + + const __m128i stg3_0 = pair_set_epi16(cospi_28_64, -cospi_4_64); + const __m128i stg3_1 = pair_set_epi16(cospi_4_64, cospi_28_64); + const __m128i stg3_4 = pair_set_epi16(-cospi_4_64, cospi_28_64); + const __m128i stg3_5 = pair_set_epi16(cospi_28_64, cospi_4_64); + const __m128i stg3_6 = pair_set_epi16(-cospi_28_64, -cospi_4_64); + const __m128i stg3_8 = pair_set_epi16(-cospi_20_64, cospi_12_64); + const __m128i stg3_9 = pair_set_epi16(cospi_12_64, cospi_20_64); + const __m128i stg3_10 = pair_set_epi16(-cospi_12_64, -cospi_20_64); + + const __m128i stg4_0 = pair_set_epi16(cospi_16_64, cospi_16_64); + const __m128i stg4_1 = pair_set_epi16(cospi_16_64, -cospi_16_64); + const __m128i stg4_4 = pair_set_epi16(-cospi_8_64, cospi_24_64); + const __m128i stg4_5 = pair_set_epi16(cospi_24_64, cospi_8_64); + const __m128i stg4_6 = pair_set_epi16(-cospi_24_64, -cospi_8_64); + + const __m128i stg6_0 = pair_set_epi16(-cospi_16_64, cospi_16_64); + + __m128i in[32], col[32]; + __m128i stp1_0, stp1_1, stp1_2, stp1_3, stp1_4, stp1_5, stp1_6, stp1_7, + stp1_8, stp1_9, stp1_10, stp1_11, stp1_12, stp1_13, stp1_14, stp1_15, + stp1_16, stp1_17, stp1_18, stp1_19, stp1_20, stp1_21, stp1_22, stp1_23, + stp1_24, stp1_25, stp1_26, stp1_27, stp1_28, stp1_29, stp1_30, stp1_31; + __m128i stp2_0, stp2_1, stp2_2, stp2_3, stp2_4, stp2_5, stp2_6, stp2_7, + stp2_8, stp2_9, stp2_10, stp2_11, stp2_12, stp2_13, stp2_14, stp2_15, + stp2_16, stp2_17, stp2_18, stp2_19, stp2_20, stp2_21, stp2_22, stp2_23, + stp2_24, stp2_25, stp2_26, stp2_27, stp2_28, stp2_29, stp2_30, stp2_31; + __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7; + int i; + + // Load input data. Only need to load the top left 8x8 block. + in[0] = load_input_data(input); + in[1] = load_input_data(input + 32); + in[2] = load_input_data(input + 64); + in[3] = load_input_data(input + 96); + in[4] = load_input_data(input + 128); + in[5] = load_input_data(input + 160); + in[6] = load_input_data(input + 192); + in[7] = load_input_data(input + 224); + + for (i = 8; i < 32; ++i) { + in[i] = _mm_setzero_si128(); + } + + array_transpose_8x8(in, in); + // TODO(hkuang): Following transposes are unnecessary. But remove them will + // lead to performance drop on some devices. + array_transpose_8x8(in + 8, in + 8); + array_transpose_8x8(in + 16, in + 16); + array_transpose_8x8(in + 24, in + 24); + + IDCT32_34 + + // 1_D: Store 32 intermediate results for each 8x32 block. + col[0] = _mm_add_epi16(stp1_0, stp1_31); + col[1] = _mm_add_epi16(stp1_1, stp1_30); + col[2] = _mm_add_epi16(stp1_2, stp1_29); + col[3] = _mm_add_epi16(stp1_3, stp1_28); + col[4] = _mm_add_epi16(stp1_4, stp1_27); + col[5] = _mm_add_epi16(stp1_5, stp1_26); + col[6] = _mm_add_epi16(stp1_6, stp1_25); + col[7] = _mm_add_epi16(stp1_7, stp1_24); + col[8] = _mm_add_epi16(stp1_8, stp1_23); + col[9] = _mm_add_epi16(stp1_9, stp1_22); + col[10] = _mm_add_epi16(stp1_10, stp1_21); + col[11] = _mm_add_epi16(stp1_11, stp1_20); + col[12] = _mm_add_epi16(stp1_12, stp1_19); + col[13] = _mm_add_epi16(stp1_13, stp1_18); + col[14] = _mm_add_epi16(stp1_14, stp1_17); + col[15] = _mm_add_epi16(stp1_15, stp1_16); + col[16] = _mm_sub_epi16(stp1_15, stp1_16); + col[17] = _mm_sub_epi16(stp1_14, stp1_17); + col[18] = _mm_sub_epi16(stp1_13, stp1_18); + col[19] = _mm_sub_epi16(stp1_12, stp1_19); + col[20] = _mm_sub_epi16(stp1_11, stp1_20); + col[21] = _mm_sub_epi16(stp1_10, stp1_21); + col[22] = _mm_sub_epi16(stp1_9, stp1_22); + col[23] = _mm_sub_epi16(stp1_8, stp1_23); + col[24] = _mm_sub_epi16(stp1_7, stp1_24); + col[25] = _mm_sub_epi16(stp1_6, stp1_25); + col[26] = _mm_sub_epi16(stp1_5, stp1_26); + col[27] = _mm_sub_epi16(stp1_4, stp1_27); + col[28] = _mm_sub_epi16(stp1_3, stp1_28); + col[29] = _mm_sub_epi16(stp1_2, stp1_29); + col[30] = _mm_sub_epi16(stp1_1, stp1_30); + col[31] = _mm_sub_epi16(stp1_0, stp1_31); + for (i = 0; i < 4; i++) { + int j; + // Transpose 32x8 block to 8x32 block + array_transpose_8x8(col + i * 8, in); + IDCT32_34 + + // 2_D: Calculate the results and store them to destination. + in[0] = _mm_add_epi16(stp1_0, stp1_31); + in[1] = _mm_add_epi16(stp1_1, stp1_30); + in[2] = _mm_add_epi16(stp1_2, stp1_29); + in[3] = _mm_add_epi16(stp1_3, stp1_28); + in[4] = _mm_add_epi16(stp1_4, stp1_27); + in[5] = _mm_add_epi16(stp1_5, stp1_26); + in[6] = _mm_add_epi16(stp1_6, stp1_25); + in[7] = _mm_add_epi16(stp1_7, stp1_24); + in[8] = _mm_add_epi16(stp1_8, stp1_23); + in[9] = _mm_add_epi16(stp1_9, stp1_22); + in[10] = _mm_add_epi16(stp1_10, stp1_21); + in[11] = _mm_add_epi16(stp1_11, stp1_20); + in[12] = _mm_add_epi16(stp1_12, stp1_19); + in[13] = _mm_add_epi16(stp1_13, stp1_18); + in[14] = _mm_add_epi16(stp1_14, stp1_17); + in[15] = _mm_add_epi16(stp1_15, stp1_16); + in[16] = _mm_sub_epi16(stp1_15, stp1_16); + in[17] = _mm_sub_epi16(stp1_14, stp1_17); + in[18] = _mm_sub_epi16(stp1_13, stp1_18); + in[19] = _mm_sub_epi16(stp1_12, stp1_19); + in[20] = _mm_sub_epi16(stp1_11, stp1_20); + in[21] = _mm_sub_epi16(stp1_10, stp1_21); + in[22] = _mm_sub_epi16(stp1_9, stp1_22); + in[23] = _mm_sub_epi16(stp1_8, stp1_23); + in[24] = _mm_sub_epi16(stp1_7, stp1_24); + in[25] = _mm_sub_epi16(stp1_6, stp1_25); + in[26] = _mm_sub_epi16(stp1_5, stp1_26); + in[27] = _mm_sub_epi16(stp1_4, stp1_27); + in[28] = _mm_sub_epi16(stp1_3, stp1_28); + in[29] = _mm_sub_epi16(stp1_2, stp1_29); + in[30] = _mm_sub_epi16(stp1_1, stp1_30); + in[31] = _mm_sub_epi16(stp1_0, stp1_31); + + for (j = 0; j < 32; ++j) { + // Final rounding and shift + in[j] = _mm_adds_epi16(in[j], final_rounding); + in[j] = _mm_srai_epi16(in[j], 6); + RECON_AND_STORE(dest + j * stride, in[j]); + } + + dest += 8; + } +} + +void aom_idct32x32_1024_add_sse2(const tran_low_t *input, uint8_t *dest, + int stride) { + const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING); + const __m128i final_rounding = _mm_set1_epi16(1 << 5); + const __m128i zero = _mm_setzero_si128(); + + // idct constants for each stage + const __m128i stg1_0 = pair_set_epi16(cospi_31_64, -cospi_1_64); + const __m128i stg1_1 = pair_set_epi16(cospi_1_64, cospi_31_64); + const __m128i stg1_2 = pair_set_epi16(cospi_15_64, -cospi_17_64); + const __m128i stg1_3 = pair_set_epi16(cospi_17_64, cospi_15_64); + const __m128i stg1_4 = pair_set_epi16(cospi_23_64, -cospi_9_64); + const __m128i stg1_5 = pair_set_epi16(cospi_9_64, cospi_23_64); + const __m128i stg1_6 = pair_set_epi16(cospi_7_64, -cospi_25_64); + const __m128i stg1_7 = pair_set_epi16(cospi_25_64, cospi_7_64); + const __m128i stg1_8 = pair_set_epi16(cospi_27_64, -cospi_5_64); + const __m128i stg1_9 = pair_set_epi16(cospi_5_64, cospi_27_64); + const __m128i stg1_10 = pair_set_epi16(cospi_11_64, -cospi_21_64); + const __m128i stg1_11 = pair_set_epi16(cospi_21_64, cospi_11_64); + const __m128i stg1_12 = pair_set_epi16(cospi_19_64, -cospi_13_64); + const __m128i stg1_13 = pair_set_epi16(cospi_13_64, cospi_19_64); + const __m128i stg1_14 = pair_set_epi16(cospi_3_64, -cospi_29_64); + const __m128i stg1_15 = pair_set_epi16(cospi_29_64, cospi_3_64); + + const __m128i stg2_0 = pair_set_epi16(cospi_30_64, -cospi_2_64); + const __m128i stg2_1 = pair_set_epi16(cospi_2_64, cospi_30_64); + const __m128i stg2_2 = pair_set_epi16(cospi_14_64, -cospi_18_64); + const __m128i stg2_3 = pair_set_epi16(cospi_18_64, cospi_14_64); + const __m128i stg2_4 = pair_set_epi16(cospi_22_64, -cospi_10_64); + const __m128i stg2_5 = pair_set_epi16(cospi_10_64, cospi_22_64); + const __m128i stg2_6 = pair_set_epi16(cospi_6_64, -cospi_26_64); + const __m128i stg2_7 = pair_set_epi16(cospi_26_64, cospi_6_64); + + const __m128i stg3_0 = pair_set_epi16(cospi_28_64, -cospi_4_64); + const __m128i stg3_1 = pair_set_epi16(cospi_4_64, cospi_28_64); + const __m128i stg3_2 = pair_set_epi16(cospi_12_64, -cospi_20_64); + const __m128i stg3_3 = pair_set_epi16(cospi_20_64, cospi_12_64); + const __m128i stg3_4 = pair_set_epi16(-cospi_4_64, cospi_28_64); + const __m128i stg3_5 = pair_set_epi16(cospi_28_64, cospi_4_64); + const __m128i stg3_6 = pair_set_epi16(-cospi_28_64, -cospi_4_64); + const __m128i stg3_8 = pair_set_epi16(-cospi_20_64, cospi_12_64); + const __m128i stg3_9 = pair_set_epi16(cospi_12_64, cospi_20_64); + const __m128i stg3_10 = pair_set_epi16(-cospi_12_64, -cospi_20_64); + + const __m128i stg4_0 = pair_set_epi16(cospi_16_64, cospi_16_64); + const __m128i stg4_1 = pair_set_epi16(cospi_16_64, -cospi_16_64); + const __m128i stg4_2 = pair_set_epi16(cospi_24_64, -cospi_8_64); + const __m128i stg4_3 = pair_set_epi16(cospi_8_64, cospi_24_64); + const __m128i stg4_4 = pair_set_epi16(-cospi_8_64, cospi_24_64); + const __m128i stg4_5 = pair_set_epi16(cospi_24_64, cospi_8_64); + const __m128i stg4_6 = pair_set_epi16(-cospi_24_64, -cospi_8_64); + + const __m128i stg6_0 = pair_set_epi16(-cospi_16_64, cospi_16_64); + + __m128i in[32], col[128], zero_idx[16]; + __m128i stp1_0, stp1_1, stp1_2, stp1_3, stp1_4, stp1_5, stp1_6, stp1_7, + stp1_8, stp1_9, stp1_10, stp1_11, stp1_12, stp1_13, stp1_14, stp1_15, + stp1_16, stp1_17, stp1_18, stp1_19, stp1_20, stp1_21, stp1_22, stp1_23, + stp1_24, stp1_25, stp1_26, stp1_27, stp1_28, stp1_29, stp1_30, stp1_31; + __m128i stp2_0, stp2_1, stp2_2, stp2_3, stp2_4, stp2_5, stp2_6, stp2_7, + stp2_8, stp2_9, stp2_10, stp2_11, stp2_12, stp2_13, stp2_14, stp2_15, + stp2_16, stp2_17, stp2_18, stp2_19, stp2_20, stp2_21, stp2_22, stp2_23, + stp2_24, stp2_25, stp2_26, stp2_27, stp2_28, stp2_29, stp2_30, stp2_31; + __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7; + int i, j, i32; + + for (i = 0; i < 4; i++) { + i32 = (i << 5); + // First 1-D idct + // Load input data. + LOAD_DQCOEFF(in[0], input); + LOAD_DQCOEFF(in[8], input); + LOAD_DQCOEFF(in[16], input); + LOAD_DQCOEFF(in[24], input); + LOAD_DQCOEFF(in[1], input); + LOAD_DQCOEFF(in[9], input); + LOAD_DQCOEFF(in[17], input); + LOAD_DQCOEFF(in[25], input); + LOAD_DQCOEFF(in[2], input); + LOAD_DQCOEFF(in[10], input); + LOAD_DQCOEFF(in[18], input); + LOAD_DQCOEFF(in[26], input); + LOAD_DQCOEFF(in[3], input); + LOAD_DQCOEFF(in[11], input); + LOAD_DQCOEFF(in[19], input); + LOAD_DQCOEFF(in[27], input); + + LOAD_DQCOEFF(in[4], input); + LOAD_DQCOEFF(in[12], input); + LOAD_DQCOEFF(in[20], input); + LOAD_DQCOEFF(in[28], input); + LOAD_DQCOEFF(in[5], input); + LOAD_DQCOEFF(in[13], input); + LOAD_DQCOEFF(in[21], input); + LOAD_DQCOEFF(in[29], input); + LOAD_DQCOEFF(in[6], input); + LOAD_DQCOEFF(in[14], input); + LOAD_DQCOEFF(in[22], input); + LOAD_DQCOEFF(in[30], input); + LOAD_DQCOEFF(in[7], input); + LOAD_DQCOEFF(in[15], input); + LOAD_DQCOEFF(in[23], input); + LOAD_DQCOEFF(in[31], input); + + // checking if all entries are zero + zero_idx[0] = _mm_or_si128(in[0], in[1]); + zero_idx[1] = _mm_or_si128(in[2], in[3]); + zero_idx[2] = _mm_or_si128(in[4], in[5]); + zero_idx[3] = _mm_or_si128(in[6], in[7]); + zero_idx[4] = _mm_or_si128(in[8], in[9]); + zero_idx[5] = _mm_or_si128(in[10], in[11]); + zero_idx[6] = _mm_or_si128(in[12], in[13]); + zero_idx[7] = _mm_or_si128(in[14], in[15]); + zero_idx[8] = _mm_or_si128(in[16], in[17]); + zero_idx[9] = _mm_or_si128(in[18], in[19]); + zero_idx[10] = _mm_or_si128(in[20], in[21]); + zero_idx[11] = _mm_or_si128(in[22], in[23]); + zero_idx[12] = _mm_or_si128(in[24], in[25]); + zero_idx[13] = _mm_or_si128(in[26], in[27]); + zero_idx[14] = _mm_or_si128(in[28], in[29]); + zero_idx[15] = _mm_or_si128(in[30], in[31]); + + zero_idx[0] = _mm_or_si128(zero_idx[0], zero_idx[1]); + zero_idx[1] = _mm_or_si128(zero_idx[2], zero_idx[3]); + zero_idx[2] = _mm_or_si128(zero_idx[4], zero_idx[5]); + zero_idx[3] = _mm_or_si128(zero_idx[6], zero_idx[7]); + zero_idx[4] = _mm_or_si128(zero_idx[8], zero_idx[9]); + zero_idx[5] = _mm_or_si128(zero_idx[10], zero_idx[11]); + zero_idx[6] = _mm_or_si128(zero_idx[12], zero_idx[13]); + zero_idx[7] = _mm_or_si128(zero_idx[14], zero_idx[15]); + + zero_idx[8] = _mm_or_si128(zero_idx[0], zero_idx[1]); + zero_idx[9] = _mm_or_si128(zero_idx[2], zero_idx[3]); + zero_idx[10] = _mm_or_si128(zero_idx[4], zero_idx[5]); + zero_idx[11] = _mm_or_si128(zero_idx[6], zero_idx[7]); + zero_idx[12] = _mm_or_si128(zero_idx[8], zero_idx[9]); + zero_idx[13] = _mm_or_si128(zero_idx[10], zero_idx[11]); + zero_idx[14] = _mm_or_si128(zero_idx[12], zero_idx[13]); + + if (_mm_movemask_epi8(_mm_cmpeq_epi32(zero_idx[14], zero)) == 0xFFFF) { + col[i32 + 0] = _mm_setzero_si128(); + col[i32 + 1] = _mm_setzero_si128(); + col[i32 + 2] = _mm_setzero_si128(); + col[i32 + 3] = _mm_setzero_si128(); + col[i32 + 4] = _mm_setzero_si128(); + col[i32 + 5] = _mm_setzero_si128(); + col[i32 + 6] = _mm_setzero_si128(); + col[i32 + 7] = _mm_setzero_si128(); + col[i32 + 8] = _mm_setzero_si128(); + col[i32 + 9] = _mm_setzero_si128(); + col[i32 + 10] = _mm_setzero_si128(); + col[i32 + 11] = _mm_setzero_si128(); + col[i32 + 12] = _mm_setzero_si128(); + col[i32 + 13] = _mm_setzero_si128(); + col[i32 + 14] = _mm_setzero_si128(); + col[i32 + 15] = _mm_setzero_si128(); + col[i32 + 16] = _mm_setzero_si128(); + col[i32 + 17] = _mm_setzero_si128(); + col[i32 + 18] = _mm_setzero_si128(); + col[i32 + 19] = _mm_setzero_si128(); + col[i32 + 20] = _mm_setzero_si128(); + col[i32 + 21] = _mm_setzero_si128(); + col[i32 + 22] = _mm_setzero_si128(); + col[i32 + 23] = _mm_setzero_si128(); + col[i32 + 24] = _mm_setzero_si128(); + col[i32 + 25] = _mm_setzero_si128(); + col[i32 + 26] = _mm_setzero_si128(); + col[i32 + 27] = _mm_setzero_si128(); + col[i32 + 28] = _mm_setzero_si128(); + col[i32 + 29] = _mm_setzero_si128(); + col[i32 + 30] = _mm_setzero_si128(); + col[i32 + 31] = _mm_setzero_si128(); + continue; + } + + // Transpose 32x8 block to 8x32 block + array_transpose_8x8(in, in); + array_transpose_8x8(in + 8, in + 8); + array_transpose_8x8(in + 16, in + 16); + array_transpose_8x8(in + 24, in + 24); + + IDCT32(in, in + 16) + + // 1_D: Store 32 intermediate results for each 8x32 block. + col[i32 + 0] = _mm_add_epi16(stp1_0, stp1_31); + col[i32 + 1] = _mm_add_epi16(stp1_1, stp1_30); + col[i32 + 2] = _mm_add_epi16(stp1_2, stp1_29); + col[i32 + 3] = _mm_add_epi16(stp1_3, stp1_28); + col[i32 + 4] = _mm_add_epi16(stp1_4, stp1_27); + col[i32 + 5] = _mm_add_epi16(stp1_5, stp1_26); + col[i32 + 6] = _mm_add_epi16(stp1_6, stp1_25); + col[i32 + 7] = _mm_add_epi16(stp1_7, stp1_24); + col[i32 + 8] = _mm_add_epi16(stp1_8, stp1_23); + col[i32 + 9] = _mm_add_epi16(stp1_9, stp1_22); + col[i32 + 10] = _mm_add_epi16(stp1_10, stp1_21); + col[i32 + 11] = _mm_add_epi16(stp1_11, stp1_20); + col[i32 + 12] = _mm_add_epi16(stp1_12, stp1_19); + col[i32 + 13] = _mm_add_epi16(stp1_13, stp1_18); + col[i32 + 14] = _mm_add_epi16(stp1_14, stp1_17); + col[i32 + 15] = _mm_add_epi16(stp1_15, stp1_16); + col[i32 + 16] = _mm_sub_epi16(stp1_15, stp1_16); + col[i32 + 17] = _mm_sub_epi16(stp1_14, stp1_17); + col[i32 + 18] = _mm_sub_epi16(stp1_13, stp1_18); + col[i32 + 19] = _mm_sub_epi16(stp1_12, stp1_19); + col[i32 + 20] = _mm_sub_epi16(stp1_11, stp1_20); + col[i32 + 21] = _mm_sub_epi16(stp1_10, stp1_21); + col[i32 + 22] = _mm_sub_epi16(stp1_9, stp1_22); + col[i32 + 23] = _mm_sub_epi16(stp1_8, stp1_23); + col[i32 + 24] = _mm_sub_epi16(stp1_7, stp1_24); + col[i32 + 25] = _mm_sub_epi16(stp1_6, stp1_25); + col[i32 + 26] = _mm_sub_epi16(stp1_5, stp1_26); + col[i32 + 27] = _mm_sub_epi16(stp1_4, stp1_27); + col[i32 + 28] = _mm_sub_epi16(stp1_3, stp1_28); + col[i32 + 29] = _mm_sub_epi16(stp1_2, stp1_29); + col[i32 + 30] = _mm_sub_epi16(stp1_1, stp1_30); + col[i32 + 31] = _mm_sub_epi16(stp1_0, stp1_31); + } + for (i = 0; i < 4; i++) { + // Second 1-D idct + j = i << 3; + + // Transpose 32x8 block to 8x32 block + array_transpose_8x8(col + j, in); + array_transpose_8x8(col + j + 32, in + 8); + array_transpose_8x8(col + j + 64, in + 16); + array_transpose_8x8(col + j + 96, in + 24); + + IDCT32(in, in + 16) + + // 2_D: Calculate the results and store them to destination. + in[0] = _mm_add_epi16(stp1_0, stp1_31); + in[1] = _mm_add_epi16(stp1_1, stp1_30); + in[2] = _mm_add_epi16(stp1_2, stp1_29); + in[3] = _mm_add_epi16(stp1_3, stp1_28); + in[4] = _mm_add_epi16(stp1_4, stp1_27); + in[5] = _mm_add_epi16(stp1_5, stp1_26); + in[6] = _mm_add_epi16(stp1_6, stp1_25); + in[7] = _mm_add_epi16(stp1_7, stp1_24); + in[8] = _mm_add_epi16(stp1_8, stp1_23); + in[9] = _mm_add_epi16(stp1_9, stp1_22); + in[10] = _mm_add_epi16(stp1_10, stp1_21); + in[11] = _mm_add_epi16(stp1_11, stp1_20); + in[12] = _mm_add_epi16(stp1_12, stp1_19); + in[13] = _mm_add_epi16(stp1_13, stp1_18); + in[14] = _mm_add_epi16(stp1_14, stp1_17); + in[15] = _mm_add_epi16(stp1_15, stp1_16); + in[16] = _mm_sub_epi16(stp1_15, stp1_16); + in[17] = _mm_sub_epi16(stp1_14, stp1_17); + in[18] = _mm_sub_epi16(stp1_13, stp1_18); + in[19] = _mm_sub_epi16(stp1_12, stp1_19); + in[20] = _mm_sub_epi16(stp1_11, stp1_20); + in[21] = _mm_sub_epi16(stp1_10, stp1_21); + in[22] = _mm_sub_epi16(stp1_9, stp1_22); + in[23] = _mm_sub_epi16(stp1_8, stp1_23); + in[24] = _mm_sub_epi16(stp1_7, stp1_24); + in[25] = _mm_sub_epi16(stp1_6, stp1_25); + in[26] = _mm_sub_epi16(stp1_5, stp1_26); + in[27] = _mm_sub_epi16(stp1_4, stp1_27); + in[28] = _mm_sub_epi16(stp1_3, stp1_28); + in[29] = _mm_sub_epi16(stp1_2, stp1_29); + in[30] = _mm_sub_epi16(stp1_1, stp1_30); + in[31] = _mm_sub_epi16(stp1_0, stp1_31); + + for (j = 0; j < 32; ++j) { + // Final rounding and shift + in[j] = _mm_adds_epi16(in[j], final_rounding); + in[j] = _mm_srai_epi16(in[j], 6); + RECON_AND_STORE(dest + j * stride, in[j]); + } + + dest += 8; + } +} + +void aom_idct32x32_1_add_sse2(const tran_low_t *input, uint8_t *dest, + int stride) { + __m128i dc_value; + const __m128i zero = _mm_setzero_si128(); + int a, j; + + a = (int)dct_const_round_shift(input[0] * cospi_16_64); + a = (int)dct_const_round_shift(a * cospi_16_64); + a = ROUND_POWER_OF_TWO(a, 6); + + if (a == 0) return; + + dc_value = _mm_set1_epi16(a); + + for (j = 0; j < 32; ++j) { + RECON_AND_STORE(dest + 0 + j * stride, dc_value); + RECON_AND_STORE(dest + 8 + j * stride, dc_value); + RECON_AND_STORE(dest + 16 + j * stride, dc_value); + RECON_AND_STORE(dest + 24 + j * stride, dc_value); + } +} + +// Apply a 32-element IDCT to 8 columns. This does not do any transposition +// of its input - the caller is expected to have done that. +// The input buffers are the top and bottom halves of an 8x32 block. +void idct32_8col(__m128i *in0, __m128i *in1) { + const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING); + + // idct constants for each stage + const __m128i stg1_0 = pair_set_epi16(cospi_31_64, -cospi_1_64); + const __m128i stg1_1 = pair_set_epi16(cospi_1_64, cospi_31_64); + const __m128i stg1_2 = pair_set_epi16(cospi_15_64, -cospi_17_64); + const __m128i stg1_3 = pair_set_epi16(cospi_17_64, cospi_15_64); + const __m128i stg1_4 = pair_set_epi16(cospi_23_64, -cospi_9_64); + const __m128i stg1_5 = pair_set_epi16(cospi_9_64, cospi_23_64); + const __m128i stg1_6 = pair_set_epi16(cospi_7_64, -cospi_25_64); + const __m128i stg1_7 = pair_set_epi16(cospi_25_64, cospi_7_64); + const __m128i stg1_8 = pair_set_epi16(cospi_27_64, -cospi_5_64); + const __m128i stg1_9 = pair_set_epi16(cospi_5_64, cospi_27_64); + const __m128i stg1_10 = pair_set_epi16(cospi_11_64, -cospi_21_64); + const __m128i stg1_11 = pair_set_epi16(cospi_21_64, cospi_11_64); + const __m128i stg1_12 = pair_set_epi16(cospi_19_64, -cospi_13_64); + const __m128i stg1_13 = pair_set_epi16(cospi_13_64, cospi_19_64); + const __m128i stg1_14 = pair_set_epi16(cospi_3_64, -cospi_29_64); + const __m128i stg1_15 = pair_set_epi16(cospi_29_64, cospi_3_64); + + const __m128i stg2_0 = pair_set_epi16(cospi_30_64, -cospi_2_64); + const __m128i stg2_1 = pair_set_epi16(cospi_2_64, cospi_30_64); + const __m128i stg2_2 = pair_set_epi16(cospi_14_64, -cospi_18_64); + const __m128i stg2_3 = pair_set_epi16(cospi_18_64, cospi_14_64); + const __m128i stg2_4 = pair_set_epi16(cospi_22_64, -cospi_10_64); + const __m128i stg2_5 = pair_set_epi16(cospi_10_64, cospi_22_64); + const __m128i stg2_6 = pair_set_epi16(cospi_6_64, -cospi_26_64); + const __m128i stg2_7 = pair_set_epi16(cospi_26_64, cospi_6_64); + + const __m128i stg3_0 = pair_set_epi16(cospi_28_64, -cospi_4_64); + const __m128i stg3_1 = pair_set_epi16(cospi_4_64, cospi_28_64); + const __m128i stg3_2 = pair_set_epi16(cospi_12_64, -cospi_20_64); + const __m128i stg3_3 = pair_set_epi16(cospi_20_64, cospi_12_64); + const __m128i stg3_4 = pair_set_epi16(-cospi_4_64, cospi_28_64); + const __m128i stg3_5 = pair_set_epi16(cospi_28_64, cospi_4_64); + const __m128i stg3_6 = pair_set_epi16(-cospi_28_64, -cospi_4_64); + const __m128i stg3_8 = pair_set_epi16(-cospi_20_64, cospi_12_64); + const __m128i stg3_9 = pair_set_epi16(cospi_12_64, cospi_20_64); + const __m128i stg3_10 = pair_set_epi16(-cospi_12_64, -cospi_20_64); + + const __m128i stg4_0 = pair_set_epi16(cospi_16_64, cospi_16_64); + const __m128i stg4_1 = pair_set_epi16(cospi_16_64, -cospi_16_64); + const __m128i stg4_2 = pair_set_epi16(cospi_24_64, -cospi_8_64); + const __m128i stg4_3 = pair_set_epi16(cospi_8_64, cospi_24_64); + const __m128i stg4_4 = pair_set_epi16(-cospi_8_64, cospi_24_64); + const __m128i stg4_5 = pair_set_epi16(cospi_24_64, cospi_8_64); + const __m128i stg4_6 = pair_set_epi16(-cospi_24_64, -cospi_8_64); + + const __m128i stg6_0 = pair_set_epi16(-cospi_16_64, cospi_16_64); + + __m128i stp1_0, stp1_1, stp1_2, stp1_3, stp1_4, stp1_5, stp1_6, stp1_7, + stp1_8, stp1_9, stp1_10, stp1_11, stp1_12, stp1_13, stp1_14, stp1_15, + stp1_16, stp1_17, stp1_18, stp1_19, stp1_20, stp1_21, stp1_22, stp1_23, + stp1_24, stp1_25, stp1_26, stp1_27, stp1_28, stp1_29, stp1_30, stp1_31; + __m128i stp2_0, stp2_1, stp2_2, stp2_3, stp2_4, stp2_5, stp2_6, stp2_7, + stp2_8, stp2_9, stp2_10, stp2_11, stp2_12, stp2_13, stp2_14, stp2_15, + stp2_16, stp2_17, stp2_18, stp2_19, stp2_20, stp2_21, stp2_22, stp2_23, + stp2_24, stp2_25, stp2_26, stp2_27, stp2_28, stp2_29, stp2_30, stp2_31; + __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7; + + IDCT32(in0, in1) + + // 2_D: Calculate the results and store them to destination. + in0[0] = _mm_add_epi16(stp1_0, stp1_31); + in0[1] = _mm_add_epi16(stp1_1, stp1_30); + in0[2] = _mm_add_epi16(stp1_2, stp1_29); + in0[3] = _mm_add_epi16(stp1_3, stp1_28); + in0[4] = _mm_add_epi16(stp1_4, stp1_27); + in0[5] = _mm_add_epi16(stp1_5, stp1_26); + in0[6] = _mm_add_epi16(stp1_6, stp1_25); + in0[7] = _mm_add_epi16(stp1_7, stp1_24); + in0[8] = _mm_add_epi16(stp1_8, stp1_23); + in0[9] = _mm_add_epi16(stp1_9, stp1_22); + in0[10] = _mm_add_epi16(stp1_10, stp1_21); + in0[11] = _mm_add_epi16(stp1_11, stp1_20); + in0[12] = _mm_add_epi16(stp1_12, stp1_19); + in0[13] = _mm_add_epi16(stp1_13, stp1_18); + in0[14] = _mm_add_epi16(stp1_14, stp1_17); + in0[15] = _mm_add_epi16(stp1_15, stp1_16); + in1[0] = _mm_sub_epi16(stp1_15, stp1_16); + in1[1] = _mm_sub_epi16(stp1_14, stp1_17); + in1[2] = _mm_sub_epi16(stp1_13, stp1_18); + in1[3] = _mm_sub_epi16(stp1_12, stp1_19); + in1[4] = _mm_sub_epi16(stp1_11, stp1_20); + in1[5] = _mm_sub_epi16(stp1_10, stp1_21); + in1[6] = _mm_sub_epi16(stp1_9, stp1_22); + in1[7] = _mm_sub_epi16(stp1_8, stp1_23); + in1[8] = _mm_sub_epi16(stp1_7, stp1_24); + in1[9] = _mm_sub_epi16(stp1_6, stp1_25); + in1[10] = _mm_sub_epi16(stp1_5, stp1_26); + in1[11] = _mm_sub_epi16(stp1_4, stp1_27); + in1[12] = _mm_sub_epi16(stp1_3, stp1_28); + in1[13] = _mm_sub_epi16(stp1_2, stp1_29); + in1[14] = _mm_sub_epi16(stp1_1, stp1_30); + in1[15] = _mm_sub_epi16(stp1_0, stp1_31); +} + +#if CONFIG_HIGHBITDEPTH +static INLINE __m128i clamp_high_sse2(__m128i value, int bd) { + __m128i ubounded, retval; + const __m128i zero = _mm_set1_epi16(0); + const __m128i one = _mm_set1_epi16(1); + const __m128i max = _mm_subs_epi16(_mm_slli_epi16(one, bd), one); + ubounded = _mm_cmpgt_epi16(value, max); + retval = _mm_andnot_si128(ubounded, value); + ubounded = _mm_and_si128(ubounded, max); + retval = _mm_or_si128(retval, ubounded); + retval = _mm_and_si128(retval, _mm_cmpgt_epi16(retval, zero)); + return retval; +} + +void aom_highbd_idct4x4_16_add_sse2(const tran_low_t *input, uint8_t *dest8, + int stride, int bd) { + tran_low_t out[4 * 4]; + tran_low_t *outptr = out; + int i, j; + __m128i inptr[4]; + __m128i sign_bits[2]; + __m128i temp_mm, min_input, max_input; + int test; + uint16_t *dest = CONVERT_TO_SHORTPTR(dest8); + int optimised_cols = 0; + const __m128i zero = _mm_set1_epi16(0); + const __m128i eight = _mm_set1_epi16(8); + const __m128i max = _mm_set1_epi16(12043); + const __m128i min = _mm_set1_epi16(-12043); + // Load input into __m128i + inptr[0] = _mm_loadu_si128((const __m128i *)input); + inptr[1] = _mm_loadu_si128((const __m128i *)(input + 4)); + inptr[2] = _mm_loadu_si128((const __m128i *)(input + 8)); + inptr[3] = _mm_loadu_si128((const __m128i *)(input + 12)); + + // Pack to 16 bits + inptr[0] = _mm_packs_epi32(inptr[0], inptr[1]); + inptr[1] = _mm_packs_epi32(inptr[2], inptr[3]); + + max_input = _mm_max_epi16(inptr[0], inptr[1]); + min_input = _mm_min_epi16(inptr[0], inptr[1]); + max_input = _mm_cmpgt_epi16(max_input, max); + min_input = _mm_cmplt_epi16(min_input, min); + temp_mm = _mm_or_si128(max_input, min_input); + test = _mm_movemask_epi8(temp_mm); + + if (!test) { + // Do the row transform + aom_idct4_sse2(inptr); + + // Check the min & max values + max_input = _mm_max_epi16(inptr[0], inptr[1]); + min_input = _mm_min_epi16(inptr[0], inptr[1]); + max_input = _mm_cmpgt_epi16(max_input, max); + min_input = _mm_cmplt_epi16(min_input, min); + temp_mm = _mm_or_si128(max_input, min_input); + test = _mm_movemask_epi8(temp_mm); + + if (test) { + array_transpose_4x4(inptr); + sign_bits[0] = _mm_cmplt_epi16(inptr[0], zero); + sign_bits[1] = _mm_cmplt_epi16(inptr[1], zero); + inptr[3] = _mm_unpackhi_epi16(inptr[1], sign_bits[1]); + inptr[2] = _mm_unpacklo_epi16(inptr[1], sign_bits[1]); + inptr[1] = _mm_unpackhi_epi16(inptr[0], sign_bits[0]); + inptr[0] = _mm_unpacklo_epi16(inptr[0], sign_bits[0]); + _mm_storeu_si128((__m128i *)outptr, inptr[0]); + _mm_storeu_si128((__m128i *)(outptr + 4), inptr[1]); + _mm_storeu_si128((__m128i *)(outptr + 8), inptr[2]); + _mm_storeu_si128((__m128i *)(outptr + 12), inptr[3]); + } else { + // Set to use the optimised transform for the column + optimised_cols = 1; + } + } else { + // Run the un-optimised row transform + for (i = 0; i < 4; ++i) { + aom_highbd_idct4_c(input, outptr, bd); + input += 4; + outptr += 4; + } + } + + if (optimised_cols) { + aom_idct4_sse2(inptr); + + // Final round and shift + inptr[0] = _mm_add_epi16(inptr[0], eight); + inptr[1] = _mm_add_epi16(inptr[1], eight); + + inptr[0] = _mm_srai_epi16(inptr[0], 4); + inptr[1] = _mm_srai_epi16(inptr[1], 4); + + // Reconstruction and Store + { + __m128i d0 = _mm_loadl_epi64((const __m128i *)dest); + __m128i d2 = _mm_loadl_epi64((const __m128i *)(dest + stride * 2)); + d0 = _mm_unpacklo_epi64( + d0, _mm_loadl_epi64((const __m128i *)(dest + stride))); + d2 = _mm_unpacklo_epi64( + d2, _mm_loadl_epi64((const __m128i *)(dest + stride * 3))); + d0 = clamp_high_sse2(_mm_adds_epi16(d0, inptr[0]), bd); + d2 = clamp_high_sse2(_mm_adds_epi16(d2, inptr[1]), bd); + // store input0 + _mm_storel_epi64((__m128i *)dest, d0); + // store input1 + d0 = _mm_srli_si128(d0, 8); + _mm_storel_epi64((__m128i *)(dest + stride), d0); + // store input2 + _mm_storel_epi64((__m128i *)(dest + stride * 2), d2); + // store input3 + d2 = _mm_srli_si128(d2, 8); + _mm_storel_epi64((__m128i *)(dest + stride * 3), d2); + } + } else { + // Run the un-optimised column transform + tran_low_t temp_in[4], temp_out[4]; + // Columns + for (i = 0; i < 4; ++i) { + for (j = 0; j < 4; ++j) temp_in[j] = out[j * 4 + i]; + aom_highbd_idct4_c(temp_in, temp_out, bd); + for (j = 0; j < 4; ++j) { + dest[j * stride + i] = highbd_clip_pixel_add( + dest[j * stride + i], ROUND_POWER_OF_TWO(temp_out[j], 4), bd); + } + } + } +} + +#endif // CONFIG_HIGHBITDEPTH diff --git a/third_party/aom/aom_dsp/x86/inv_txfm_sse2.h b/third_party/aom/aom_dsp/x86/inv_txfm_sse2.h new file mode 100644 index 000000000..95d246c3c --- /dev/null +++ b/third_party/aom/aom_dsp/x86/inv_txfm_sse2.h @@ -0,0 +1,265 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#ifndef AOM_DSP_X86_INV_TXFM_SSE2_H_ +#define AOM_DSP_X86_INV_TXFM_SSE2_H_ + +#include <emmintrin.h> // SSE2 +#include "./aom_config.h" +#include "aom/aom_integer.h" +#include "aom_dsp/inv_txfm.h" +#include "aom_dsp/x86/txfm_common_sse2.h" + +// perform 8x8 transpose +static INLINE void array_transpose_4x4(__m128i *res) { + const __m128i tr0_0 = _mm_unpacklo_epi16(res[0], res[1]); + const __m128i tr0_1 = _mm_unpackhi_epi16(res[0], res[1]); + + res[0] = _mm_unpacklo_epi16(tr0_0, tr0_1); + res[1] = _mm_unpackhi_epi16(tr0_0, tr0_1); +} + +static INLINE void array_transpose_8x8(__m128i *in, __m128i *res) { + const __m128i tr0_0 = _mm_unpacklo_epi16(in[0], in[1]); + const __m128i tr0_1 = _mm_unpacklo_epi16(in[2], in[3]); + const __m128i tr0_2 = _mm_unpackhi_epi16(in[0], in[1]); + const __m128i tr0_3 = _mm_unpackhi_epi16(in[2], in[3]); + const __m128i tr0_4 = _mm_unpacklo_epi16(in[4], in[5]); + const __m128i tr0_5 = _mm_unpacklo_epi16(in[6], in[7]); + const __m128i tr0_6 = _mm_unpackhi_epi16(in[4], in[5]); + const __m128i tr0_7 = _mm_unpackhi_epi16(in[6], in[7]); + + const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1); + const __m128i tr1_1 = _mm_unpacklo_epi32(tr0_4, tr0_5); + const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1); + const __m128i tr1_3 = _mm_unpackhi_epi32(tr0_4, tr0_5); + const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_2, tr0_3); + const __m128i tr1_5 = _mm_unpacklo_epi32(tr0_6, tr0_7); + const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_2, tr0_3); + const __m128i tr1_7 = _mm_unpackhi_epi32(tr0_6, tr0_7); + + res[0] = _mm_unpacklo_epi64(tr1_0, tr1_1); + res[1] = _mm_unpackhi_epi64(tr1_0, tr1_1); + res[2] = _mm_unpacklo_epi64(tr1_2, tr1_3); + res[3] = _mm_unpackhi_epi64(tr1_2, tr1_3); + res[4] = _mm_unpacklo_epi64(tr1_4, tr1_5); + res[5] = _mm_unpackhi_epi64(tr1_4, tr1_5); + res[6] = _mm_unpacklo_epi64(tr1_6, tr1_7); + res[7] = _mm_unpackhi_epi64(tr1_6, tr1_7); +} + +#define TRANSPOSE_8X8(in0, in1, in2, in3, in4, in5, in6, in7, out0, out1, \ + out2, out3, out4, out5, out6, out7) \ + { \ + const __m128i tr0_0 = _mm_unpacklo_epi16(in0, in1); \ + const __m128i tr0_1 = _mm_unpacklo_epi16(in2, in3); \ + const __m128i tr0_2 = _mm_unpackhi_epi16(in0, in1); \ + const __m128i tr0_3 = _mm_unpackhi_epi16(in2, in3); \ + const __m128i tr0_4 = _mm_unpacklo_epi16(in4, in5); \ + const __m128i tr0_5 = _mm_unpacklo_epi16(in6, in7); \ + const __m128i tr0_6 = _mm_unpackhi_epi16(in4, in5); \ + const __m128i tr0_7 = _mm_unpackhi_epi16(in6, in7); \ + \ + const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1); \ + const __m128i tr1_1 = _mm_unpacklo_epi32(tr0_2, tr0_3); \ + const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1); \ + const __m128i tr1_3 = _mm_unpackhi_epi32(tr0_2, tr0_3); \ + const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_4, tr0_5); \ + const __m128i tr1_5 = _mm_unpacklo_epi32(tr0_6, tr0_7); \ + const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_4, tr0_5); \ + const __m128i tr1_7 = _mm_unpackhi_epi32(tr0_6, tr0_7); \ + \ + out0 = _mm_unpacklo_epi64(tr1_0, tr1_4); \ + out1 = _mm_unpackhi_epi64(tr1_0, tr1_4); \ + out2 = _mm_unpacklo_epi64(tr1_2, tr1_6); \ + out3 = _mm_unpackhi_epi64(tr1_2, tr1_6); \ + out4 = _mm_unpacklo_epi64(tr1_1, tr1_5); \ + out5 = _mm_unpackhi_epi64(tr1_1, tr1_5); \ + out6 = _mm_unpacklo_epi64(tr1_3, tr1_7); \ + out7 = _mm_unpackhi_epi64(tr1_3, tr1_7); \ + } + +#define TRANSPOSE_8X4(in0, in1, in2, in3, out0, out1) \ + { \ + const __m128i tr0_0 = _mm_unpacklo_epi16(in0, in1); \ + const __m128i tr0_1 = _mm_unpacklo_epi16(in2, in3); \ + \ + in0 = _mm_unpacklo_epi32(tr0_0, tr0_1); /* i1 i0 */ \ + in1 = _mm_unpackhi_epi32(tr0_0, tr0_1); /* i3 i2 */ \ + } + +static INLINE void array_transpose_4X8(__m128i *in, __m128i *out) { + const __m128i tr0_0 = _mm_unpacklo_epi16(in[0], in[1]); + const __m128i tr0_1 = _mm_unpacklo_epi16(in[2], in[3]); + const __m128i tr0_4 = _mm_unpacklo_epi16(in[4], in[5]); + const __m128i tr0_5 = _mm_unpacklo_epi16(in[6], in[7]); + + const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1); + const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1); + const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_4, tr0_5); + const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_4, tr0_5); + + out[0] = _mm_unpacklo_epi64(tr1_0, tr1_4); + out[1] = _mm_unpackhi_epi64(tr1_0, tr1_4); + out[2] = _mm_unpacklo_epi64(tr1_2, tr1_6); + out[3] = _mm_unpackhi_epi64(tr1_2, tr1_6); +} + +static INLINE void array_transpose_16x16(__m128i *res0, __m128i *res1) { + __m128i tbuf[8]; + array_transpose_8x8(res0, res0); + array_transpose_8x8(res1, tbuf); + array_transpose_8x8(res0 + 8, res1); + array_transpose_8x8(res1 + 8, res1 + 8); + + res0[8] = tbuf[0]; + res0[9] = tbuf[1]; + res0[10] = tbuf[2]; + res0[11] = tbuf[3]; + res0[12] = tbuf[4]; + res0[13] = tbuf[5]; + res0[14] = tbuf[6]; + res0[15] = tbuf[7]; +} + +// Function to allow 8 bit optimisations to be used when profile 0 is used with +// highbitdepth enabled +static INLINE __m128i load_input_data(const tran_low_t *data) { +#if CONFIG_HIGHBITDEPTH + return octa_set_epi16(data[0], data[1], data[2], data[3], data[4], data[5], + data[6], data[7]); +#else + return _mm_load_si128((const __m128i *)data); +#endif +} + +static INLINE void load_buffer_8x16(const tran_low_t *input, __m128i *in) { + in[0] = load_input_data(input + 0 * 16); + in[1] = load_input_data(input + 1 * 16); + in[2] = load_input_data(input + 2 * 16); + in[3] = load_input_data(input + 3 * 16); + in[4] = load_input_data(input + 4 * 16); + in[5] = load_input_data(input + 5 * 16); + in[6] = load_input_data(input + 6 * 16); + in[7] = load_input_data(input + 7 * 16); + + in[8] = load_input_data(input + 8 * 16); + in[9] = load_input_data(input + 9 * 16); + in[10] = load_input_data(input + 10 * 16); + in[11] = load_input_data(input + 11 * 16); + in[12] = load_input_data(input + 12 * 16); + in[13] = load_input_data(input + 13 * 16); + in[14] = load_input_data(input + 14 * 16); + in[15] = load_input_data(input + 15 * 16); +} + +#define RECON_AND_STORE(dest, in_x) \ + { \ + __m128i d0 = _mm_loadl_epi64((__m128i *)(dest)); \ + d0 = _mm_unpacklo_epi8(d0, zero); \ + d0 = _mm_add_epi16(in_x, d0); \ + d0 = _mm_packus_epi16(d0, d0); \ + _mm_storel_epi64((__m128i *)(dest), d0); \ + } + +static INLINE void write_buffer_8x16(uint8_t *dest, __m128i *in, int stride) { + const __m128i final_rounding = _mm_set1_epi16(1 << 5); + const __m128i zero = _mm_setzero_si128(); + // Final rounding and shift + in[0] = _mm_adds_epi16(in[0], final_rounding); + in[1] = _mm_adds_epi16(in[1], final_rounding); + in[2] = _mm_adds_epi16(in[2], final_rounding); + in[3] = _mm_adds_epi16(in[3], final_rounding); + in[4] = _mm_adds_epi16(in[4], final_rounding); + in[5] = _mm_adds_epi16(in[5], final_rounding); + in[6] = _mm_adds_epi16(in[6], final_rounding); + in[7] = _mm_adds_epi16(in[7], final_rounding); + in[8] = _mm_adds_epi16(in[8], final_rounding); + in[9] = _mm_adds_epi16(in[9], final_rounding); + in[10] = _mm_adds_epi16(in[10], final_rounding); + in[11] = _mm_adds_epi16(in[11], final_rounding); + in[12] = _mm_adds_epi16(in[12], final_rounding); + in[13] = _mm_adds_epi16(in[13], final_rounding); + in[14] = _mm_adds_epi16(in[14], final_rounding); + in[15] = _mm_adds_epi16(in[15], final_rounding); + + in[0] = _mm_srai_epi16(in[0], 6); + in[1] = _mm_srai_epi16(in[1], 6); + in[2] = _mm_srai_epi16(in[2], 6); + in[3] = _mm_srai_epi16(in[3], 6); + in[4] = _mm_srai_epi16(in[4], 6); + in[5] = _mm_srai_epi16(in[5], 6); + in[6] = _mm_srai_epi16(in[6], 6); + in[7] = _mm_srai_epi16(in[7], 6); + in[8] = _mm_srai_epi16(in[8], 6); + in[9] = _mm_srai_epi16(in[9], 6); + in[10] = _mm_srai_epi16(in[10], 6); + in[11] = _mm_srai_epi16(in[11], 6); + in[12] = _mm_srai_epi16(in[12], 6); + in[13] = _mm_srai_epi16(in[13], 6); + in[14] = _mm_srai_epi16(in[14], 6); + in[15] = _mm_srai_epi16(in[15], 6); + + RECON_AND_STORE(dest + 0 * stride, in[0]); + RECON_AND_STORE(dest + 1 * stride, in[1]); + RECON_AND_STORE(dest + 2 * stride, in[2]); + RECON_AND_STORE(dest + 3 * stride, in[3]); + RECON_AND_STORE(dest + 4 * stride, in[4]); + RECON_AND_STORE(dest + 5 * stride, in[5]); + RECON_AND_STORE(dest + 6 * stride, in[6]); + RECON_AND_STORE(dest + 7 * stride, in[7]); + RECON_AND_STORE(dest + 8 * stride, in[8]); + RECON_AND_STORE(dest + 9 * stride, in[9]); + RECON_AND_STORE(dest + 10 * stride, in[10]); + RECON_AND_STORE(dest + 11 * stride, in[11]); + RECON_AND_STORE(dest + 12 * stride, in[12]); + RECON_AND_STORE(dest + 13 * stride, in[13]); + RECON_AND_STORE(dest + 14 * stride, in[14]); + RECON_AND_STORE(dest + 15 * stride, in[15]); +} + +#define TRANSPOSE_4X8_10(tmp0, tmp1, tmp2, tmp3, out0, out1, out2, out3) \ + { \ + const __m128i tr0_0 = _mm_unpackhi_epi16(tmp0, tmp1); \ + const __m128i tr0_1 = _mm_unpacklo_epi16(tmp1, tmp0); \ + const __m128i tr0_4 = _mm_unpacklo_epi16(tmp2, tmp3); \ + const __m128i tr0_5 = _mm_unpackhi_epi16(tmp3, tmp2); \ + \ + const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1); \ + const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1); \ + const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_4, tr0_5); \ + const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_4, tr0_5); \ + \ + out0 = _mm_unpacklo_epi64(tr1_0, tr1_4); \ + out1 = _mm_unpackhi_epi64(tr1_0, tr1_4); \ + out2 = _mm_unpacklo_epi64(tr1_2, tr1_6); \ + out3 = _mm_unpackhi_epi64(tr1_2, tr1_6); \ + } + +#define TRANSPOSE_8X8_10(in0, in1, in2, in3, out0, out1) \ + { \ + const __m128i tr0_0 = _mm_unpacklo_epi16(in0, in1); \ + const __m128i tr0_1 = _mm_unpacklo_epi16(in2, in3); \ + out0 = _mm_unpacklo_epi32(tr0_0, tr0_1); \ + out1 = _mm_unpackhi_epi32(tr0_0, tr0_1); \ + } + +void iadst16_8col(__m128i *in); +void idct16_8col(__m128i *in); +void aom_idct4_sse2(__m128i *in); +void aom_idct8_sse2(__m128i *in); +void aom_idct16_sse2(__m128i *in0, __m128i *in1); +void aom_iadst4_sse2(__m128i *in); +void aom_iadst8_sse2(__m128i *in); +void aom_iadst16_sse2(__m128i *in0, __m128i *in1); +void idct32_8col(__m128i *in0, __m128i *in1); + +#endif // AOM_DSP_X86_INV_TXFM_SSE2_H_ diff --git a/third_party/aom/aom_dsp/x86/inv_txfm_ssse3.c b/third_party/aom/aom_dsp/x86/inv_txfm_ssse3.c new file mode 100644 index 000000000..9d006797b --- /dev/null +++ b/third_party/aom/aom_dsp/x86/inv_txfm_ssse3.c @@ -0,0 +1,1333 @@ +/* + * Copyright (c) 2017 The WebM project authors. All Rights Reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#include <tmmintrin.h> + +#include "./aom_dsp_rtcd.h" +#include "aom_dsp/x86/inv_txfm_sse2.h" +#include "aom_dsp/x86/txfm_common_sse2.h" + +void aom_idct8x8_64_add_ssse3(const tran_low_t *input, uint8_t *dest, + int stride) { + const __m128i zero = _mm_setzero_si128(); + const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING); + const __m128i final_rounding = _mm_set1_epi16(1 << 4); + const __m128i stg1_0 = pair_set_epi16(cospi_28_64, -cospi_4_64); + const __m128i stg1_1 = pair_set_epi16(cospi_4_64, cospi_28_64); + const __m128i stg1_2 = pair_set_epi16(-cospi_20_64, cospi_12_64); + const __m128i stg1_3 = pair_set_epi16(cospi_12_64, cospi_20_64); + const __m128i stk2_0 = pair_set_epi16(cospi_16_64, cospi_16_64); + const __m128i stk2_1 = pair_set_epi16(cospi_16_64, -cospi_16_64); + const __m128i stg2_2 = pair_set_epi16(cospi_24_64, -cospi_8_64); + const __m128i stg2_3 = pair_set_epi16(cospi_8_64, cospi_24_64); + + __m128i in0, in1, in2, in3, in4, in5, in6, in7; + __m128i stp1_0, stp1_1, stp1_2, stp1_3, stp1_4, stp1_5, stp1_6, stp1_7; + __m128i stp2_0, stp2_1, stp2_2, stp2_3, stp2_4, stp2_5, stp2_6, stp2_7; + __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7; + int i; + + // Load input data. + in0 = load_input_data(input); + in1 = load_input_data(input + 8 * 1); + in2 = load_input_data(input + 8 * 2); + in3 = load_input_data(input + 8 * 3); + in4 = load_input_data(input + 8 * 4); + in5 = load_input_data(input + 8 * 5); + in6 = load_input_data(input + 8 * 6); + in7 = load_input_data(input + 8 * 7); + + // 2-D + for (i = 0; i < 2; i++) { + // 8x8 Transpose is copied from vpx_fdct8x8_sse2() + TRANSPOSE_8X8(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2, in3, + in4, in5, in6, in7); + + // 4-stage 1D idct8x8 + { + /* Stage1 */ + { + const __m128i lo_17 = _mm_unpacklo_epi16(in1, in7); + const __m128i hi_17 = _mm_unpackhi_epi16(in1, in7); + const __m128i lo_35 = _mm_unpacklo_epi16(in3, in5); + const __m128i hi_35 = _mm_unpackhi_epi16(in3, in5); + + { + tmp0 = _mm_madd_epi16(lo_17, stg1_0); + tmp1 = _mm_madd_epi16(hi_17, stg1_0); + tmp2 = _mm_madd_epi16(lo_17, stg1_1); + tmp3 = _mm_madd_epi16(hi_17, stg1_1); + tmp4 = _mm_madd_epi16(lo_35, stg1_2); + tmp5 = _mm_madd_epi16(hi_35, stg1_2); + tmp6 = _mm_madd_epi16(lo_35, stg1_3); + tmp7 = _mm_madd_epi16(hi_35, stg1_3); + + tmp0 = _mm_add_epi32(tmp0, rounding); + tmp1 = _mm_add_epi32(tmp1, rounding); + tmp2 = _mm_add_epi32(tmp2, rounding); + tmp3 = _mm_add_epi32(tmp3, rounding); + tmp4 = _mm_add_epi32(tmp4, rounding); + tmp5 = _mm_add_epi32(tmp5, rounding); + tmp6 = _mm_add_epi32(tmp6, rounding); + tmp7 = _mm_add_epi32(tmp7, rounding); + + tmp0 = _mm_srai_epi32(tmp0, 14); + tmp1 = _mm_srai_epi32(tmp1, 14); + tmp2 = _mm_srai_epi32(tmp2, 14); + tmp3 = _mm_srai_epi32(tmp3, 14); + tmp4 = _mm_srai_epi32(tmp4, 14); + tmp5 = _mm_srai_epi32(tmp5, 14); + tmp6 = _mm_srai_epi32(tmp6, 14); + tmp7 = _mm_srai_epi32(tmp7, 14); + + stp1_4 = _mm_packs_epi32(tmp0, tmp1); + stp1_7 = _mm_packs_epi32(tmp2, tmp3); + stp1_5 = _mm_packs_epi32(tmp4, tmp5); + stp1_6 = _mm_packs_epi32(tmp6, tmp7); + } + } + + /* Stage2 */ + { + const __m128i lo_26 = _mm_unpacklo_epi16(in2, in6); + const __m128i hi_26 = _mm_unpackhi_epi16(in2, in6); + + { + tmp0 = _mm_unpacklo_epi16(in0, in4); + tmp1 = _mm_unpackhi_epi16(in0, in4); + + tmp2 = _mm_madd_epi16(tmp0, stk2_0); + tmp3 = _mm_madd_epi16(tmp1, stk2_0); + tmp4 = _mm_madd_epi16(tmp0, stk2_1); + tmp5 = _mm_madd_epi16(tmp1, stk2_1); + + tmp2 = _mm_add_epi32(tmp2, rounding); + tmp3 = _mm_add_epi32(tmp3, rounding); + tmp4 = _mm_add_epi32(tmp4, rounding); + tmp5 = _mm_add_epi32(tmp5, rounding); + + tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS); + tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS); + tmp4 = _mm_srai_epi32(tmp4, DCT_CONST_BITS); + tmp5 = _mm_srai_epi32(tmp5, DCT_CONST_BITS); + + stp2_0 = _mm_packs_epi32(tmp2, tmp3); + stp2_1 = _mm_packs_epi32(tmp4, tmp5); + + tmp0 = _mm_madd_epi16(lo_26, stg2_2); + tmp1 = _mm_madd_epi16(hi_26, stg2_2); + tmp2 = _mm_madd_epi16(lo_26, stg2_3); + tmp3 = _mm_madd_epi16(hi_26, stg2_3); + + tmp0 = _mm_add_epi32(tmp0, rounding); + tmp1 = _mm_add_epi32(tmp1, rounding); + tmp2 = _mm_add_epi32(tmp2, rounding); + tmp3 = _mm_add_epi32(tmp3, rounding); + + tmp0 = _mm_srai_epi32(tmp0, 14); + tmp1 = _mm_srai_epi32(tmp1, 14); + tmp2 = _mm_srai_epi32(tmp2, 14); + tmp3 = _mm_srai_epi32(tmp3, 14); + + stp2_2 = _mm_packs_epi32(tmp0, tmp1); + stp2_3 = _mm_packs_epi32(tmp2, tmp3); + } + + stp2_4 = _mm_add_epi16(stp1_4, stp1_5); + stp2_5 = _mm_sub_epi16(stp1_4, stp1_5); + stp2_6 = _mm_sub_epi16(stp1_7, stp1_6); + stp2_7 = _mm_add_epi16(stp1_7, stp1_6); + } + + /* Stage3 */ + { + stp1_0 = _mm_add_epi16(stp2_0, stp2_3); + stp1_1 = _mm_add_epi16(stp2_1, stp2_2); + stp1_2 = _mm_sub_epi16(stp2_1, stp2_2); + stp1_3 = _mm_sub_epi16(stp2_0, stp2_3); + + tmp0 = _mm_unpacklo_epi16(stp2_6, stp2_5); + tmp1 = _mm_unpackhi_epi16(stp2_6, stp2_5); + + tmp2 = _mm_madd_epi16(tmp0, stk2_1); + tmp3 = _mm_madd_epi16(tmp1, stk2_1); + tmp4 = _mm_madd_epi16(tmp0, stk2_0); + tmp5 = _mm_madd_epi16(tmp1, stk2_0); + + tmp2 = _mm_add_epi32(tmp2, rounding); + tmp3 = _mm_add_epi32(tmp3, rounding); + tmp4 = _mm_add_epi32(tmp4, rounding); + tmp5 = _mm_add_epi32(tmp5, rounding); + + tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS); + tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS); + tmp4 = _mm_srai_epi32(tmp4, DCT_CONST_BITS); + tmp5 = _mm_srai_epi32(tmp5, DCT_CONST_BITS); + + stp1_5 = _mm_packs_epi32(tmp2, tmp3); + stp1_6 = _mm_packs_epi32(tmp4, tmp5); + } + + /* Stage4 */ + in0 = _mm_add_epi16(stp1_0, stp2_7); + in1 = _mm_add_epi16(stp1_1, stp1_6); + in2 = _mm_add_epi16(stp1_2, stp1_5); + in3 = _mm_add_epi16(stp1_3, stp2_4); + in4 = _mm_sub_epi16(stp1_3, stp2_4); + in5 = _mm_sub_epi16(stp1_2, stp1_5); + in6 = _mm_sub_epi16(stp1_1, stp1_6); + in7 = _mm_sub_epi16(stp1_0, stp2_7); + } + } + + // Final rounding and shift + in0 = _mm_adds_epi16(in0, final_rounding); + in1 = _mm_adds_epi16(in1, final_rounding); + in2 = _mm_adds_epi16(in2, final_rounding); + in3 = _mm_adds_epi16(in3, final_rounding); + in4 = _mm_adds_epi16(in4, final_rounding); + in5 = _mm_adds_epi16(in5, final_rounding); + in6 = _mm_adds_epi16(in6, final_rounding); + in7 = _mm_adds_epi16(in7, final_rounding); + + in0 = _mm_srai_epi16(in0, 5); + in1 = _mm_srai_epi16(in1, 5); + in2 = _mm_srai_epi16(in2, 5); + in3 = _mm_srai_epi16(in3, 5); + in4 = _mm_srai_epi16(in4, 5); + in5 = _mm_srai_epi16(in5, 5); + in6 = _mm_srai_epi16(in6, 5); + in7 = _mm_srai_epi16(in7, 5); + + RECON_AND_STORE(dest + 0 * stride, in0); + RECON_AND_STORE(dest + 1 * stride, in1); + RECON_AND_STORE(dest + 2 * stride, in2); + RECON_AND_STORE(dest + 3 * stride, in3); + RECON_AND_STORE(dest + 4 * stride, in4); + RECON_AND_STORE(dest + 5 * stride, in5); + RECON_AND_STORE(dest + 6 * stride, in6); + RECON_AND_STORE(dest + 7 * stride, in7); +} + +void aom_idct8x8_12_add_ssse3(const tran_low_t *input, uint8_t *dest, + int stride) { + const __m128i zero = _mm_setzero_si128(); + const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING); + const __m128i final_rounding = _mm_set1_epi16(1 << 4); + const __m128i stg1_0 = pair_set_epi16(2 * cospi_28_64, 2 * cospi_28_64); + const __m128i stg1_1 = pair_set_epi16(2 * cospi_4_64, 2 * cospi_4_64); + const __m128i stg1_2 = pair_set_epi16(-2 * cospi_20_64, -2 * cospi_20_64); + const __m128i stg1_3 = pair_set_epi16(2 * cospi_12_64, 2 * cospi_12_64); + const __m128i stg2_0 = pair_set_epi16(2 * cospi_16_64, 2 * cospi_16_64); + const __m128i stk2_0 = pair_set_epi16(cospi_16_64, cospi_16_64); + const __m128i stk2_1 = pair_set_epi16(cospi_16_64, -cospi_16_64); + const __m128i stg2_2 = pair_set_epi16(2 * cospi_24_64, 2 * cospi_24_64); + const __m128i stg2_3 = pair_set_epi16(2 * cospi_8_64, 2 * cospi_8_64); + const __m128i stg3_0 = pair_set_epi16(-cospi_16_64, cospi_16_64); + + __m128i in0, in1, in2, in3, in4, in5, in6, in7; + __m128i stp1_0, stp1_1, stp1_2, stp1_3, stp1_4, stp1_5, stp1_6, stp1_7; + __m128i stp2_0, stp2_1, stp2_2, stp2_3, stp2_4, stp2_5, stp2_6, stp2_7; + __m128i tmp0, tmp1, tmp2, tmp3; + + // Rows. Load 4-row input data. + in0 = load_input_data(input); + in1 = load_input_data(input + 8 * 1); + in2 = load_input_data(input + 8 * 2); + in3 = load_input_data(input + 8 * 3); + + // 8x4 Transpose + TRANSPOSE_8X8_10(in0, in1, in2, in3, in0, in1); + + // Stage1 + tmp0 = _mm_mulhrs_epi16(in0, stg1_0); + tmp1 = _mm_mulhrs_epi16(in0, stg1_1); + tmp2 = _mm_mulhrs_epi16(in1, stg1_2); + tmp3 = _mm_mulhrs_epi16(in1, stg1_3); + + stp1_4 = _mm_unpackhi_epi64(tmp0, tmp1); + stp1_5 = _mm_unpackhi_epi64(tmp2, tmp3); + + // Stage2 + tmp0 = _mm_mulhrs_epi16(in0, stg2_0); + stp2_0 = _mm_unpacklo_epi64(tmp0, tmp0); + + tmp1 = _mm_mulhrs_epi16(in1, stg2_2); + tmp2 = _mm_mulhrs_epi16(in1, stg2_3); + stp2_2 = _mm_unpacklo_epi64(tmp2, tmp1); + + tmp0 = _mm_add_epi16(stp1_4, stp1_5); + tmp1 = _mm_sub_epi16(stp1_4, stp1_5); + + stp2_4 = tmp0; + stp2_5 = _mm_unpacklo_epi64(tmp1, zero); + stp2_6 = _mm_unpackhi_epi64(tmp1, zero); + + tmp0 = _mm_unpacklo_epi16(stp2_5, stp2_6); + tmp1 = _mm_madd_epi16(tmp0, stg3_0); + tmp2 = _mm_madd_epi16(tmp0, stk2_0); // stg3_1 = stk2_0 + + tmp1 = _mm_add_epi32(tmp1, rounding); + tmp2 = _mm_add_epi32(tmp2, rounding); + tmp1 = _mm_srai_epi32(tmp1, DCT_CONST_BITS); + tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS); + + stp1_5 = _mm_packs_epi32(tmp1, tmp2); + + // Stage3 + tmp2 = _mm_add_epi16(stp2_0, stp2_2); + tmp3 = _mm_sub_epi16(stp2_0, stp2_2); + + stp1_2 = _mm_unpackhi_epi64(tmp3, tmp2); + stp1_3 = _mm_unpacklo_epi64(tmp3, tmp2); + + // Stage4 + tmp0 = _mm_add_epi16(stp1_3, stp2_4); + tmp1 = _mm_add_epi16(stp1_2, stp1_5); + tmp2 = _mm_sub_epi16(stp1_3, stp2_4); + tmp3 = _mm_sub_epi16(stp1_2, stp1_5); + + TRANSPOSE_4X8_10(tmp0, tmp1, tmp2, tmp3, in0, in1, in2, in3) + + /* Stage1 */ + stp1_4 = _mm_mulhrs_epi16(in1, stg1_0); + stp1_7 = _mm_mulhrs_epi16(in1, stg1_1); + stp1_5 = _mm_mulhrs_epi16(in3, stg1_2); + stp1_6 = _mm_mulhrs_epi16(in3, stg1_3); + + /* Stage2 */ + stp2_0 = _mm_mulhrs_epi16(in0, stg2_0); + stp2_1 = _mm_mulhrs_epi16(in0, stg2_0); + + stp2_2 = _mm_mulhrs_epi16(in2, stg2_2); + stp2_3 = _mm_mulhrs_epi16(in2, stg2_3); + + stp2_4 = _mm_add_epi16(stp1_4, stp1_5); + stp2_5 = _mm_sub_epi16(stp1_4, stp1_5); + stp2_6 = _mm_sub_epi16(stp1_7, stp1_6); + stp2_7 = _mm_add_epi16(stp1_7, stp1_6); + + /* Stage3 */ + stp1_0 = _mm_add_epi16(stp2_0, stp2_3); + stp1_1 = _mm_add_epi16(stp2_1, stp2_2); + stp1_2 = _mm_sub_epi16(stp2_1, stp2_2); + stp1_3 = _mm_sub_epi16(stp2_0, stp2_3); + + tmp0 = _mm_unpacklo_epi16(stp2_6, stp2_5); + tmp1 = _mm_unpackhi_epi16(stp2_6, stp2_5); + + tmp2 = _mm_madd_epi16(tmp0, stk2_0); + tmp3 = _mm_madd_epi16(tmp1, stk2_0); + tmp2 = _mm_add_epi32(tmp2, rounding); + tmp3 = _mm_add_epi32(tmp3, rounding); + tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS); + tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS); + stp1_6 = _mm_packs_epi32(tmp2, tmp3); + + tmp2 = _mm_madd_epi16(tmp0, stk2_1); + tmp3 = _mm_madd_epi16(tmp1, stk2_1); + tmp2 = _mm_add_epi32(tmp2, rounding); + tmp3 = _mm_add_epi32(tmp3, rounding); + tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS); + tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS); + stp1_5 = _mm_packs_epi32(tmp2, tmp3); + + /* Stage4 */ + in0 = _mm_add_epi16(stp1_0, stp2_7); + in1 = _mm_add_epi16(stp1_1, stp1_6); + in2 = _mm_add_epi16(stp1_2, stp1_5); + in3 = _mm_add_epi16(stp1_3, stp2_4); + in4 = _mm_sub_epi16(stp1_3, stp2_4); + in5 = _mm_sub_epi16(stp1_2, stp1_5); + in6 = _mm_sub_epi16(stp1_1, stp1_6); + in7 = _mm_sub_epi16(stp1_0, stp2_7); + + // Final rounding and shift + in0 = _mm_adds_epi16(in0, final_rounding); + in1 = _mm_adds_epi16(in1, final_rounding); + in2 = _mm_adds_epi16(in2, final_rounding); + in3 = _mm_adds_epi16(in3, final_rounding); + in4 = _mm_adds_epi16(in4, final_rounding); + in5 = _mm_adds_epi16(in5, final_rounding); + in6 = _mm_adds_epi16(in6, final_rounding); + in7 = _mm_adds_epi16(in7, final_rounding); + + in0 = _mm_srai_epi16(in0, 5); + in1 = _mm_srai_epi16(in1, 5); + in2 = _mm_srai_epi16(in2, 5); + in3 = _mm_srai_epi16(in3, 5); + in4 = _mm_srai_epi16(in4, 5); + in5 = _mm_srai_epi16(in5, 5); + in6 = _mm_srai_epi16(in6, 5); + in7 = _mm_srai_epi16(in7, 5); + + RECON_AND_STORE(dest + 0 * stride, in0); + RECON_AND_STORE(dest + 1 * stride, in1); + RECON_AND_STORE(dest + 2 * stride, in2); + RECON_AND_STORE(dest + 3 * stride, in3); + RECON_AND_STORE(dest + 4 * stride, in4); + RECON_AND_STORE(dest + 5 * stride, in5); + RECON_AND_STORE(dest + 6 * stride, in6); + RECON_AND_STORE(dest + 7 * stride, in7); +} + +// Only do addition and subtraction butterfly, size = 16, 32 +static INLINE void add_sub_butterfly(const __m128i *in, __m128i *out, + int size) { + int i = 0; + const int num = size >> 1; + const int bound = size - 1; + while (i < num) { + out[i] = _mm_add_epi16(in[i], in[bound - i]); + out[bound - i] = _mm_sub_epi16(in[i], in[bound - i]); + i++; + } +} + +#define BUTTERFLY_PAIR(x0, x1, co0, co1) \ + do { \ + tmp0 = _mm_madd_epi16(x0, co0); \ + tmp1 = _mm_madd_epi16(x1, co0); \ + tmp2 = _mm_madd_epi16(x0, co1); \ + tmp3 = _mm_madd_epi16(x1, co1); \ + tmp0 = _mm_add_epi32(tmp0, rounding); \ + tmp1 = _mm_add_epi32(tmp1, rounding); \ + tmp2 = _mm_add_epi32(tmp2, rounding); \ + tmp3 = _mm_add_epi32(tmp3, rounding); \ + tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS); \ + tmp1 = _mm_srai_epi32(tmp1, DCT_CONST_BITS); \ + tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS); \ + tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS); \ + } while (0) + +static INLINE void butterfly(const __m128i *x0, const __m128i *x1, + const __m128i *c0, const __m128i *c1, __m128i *y0, + __m128i *y1) { + __m128i tmp0, tmp1, tmp2, tmp3, u0, u1; + const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING); + + u0 = _mm_unpacklo_epi16(*x0, *x1); + u1 = _mm_unpackhi_epi16(*x0, *x1); + BUTTERFLY_PAIR(u0, u1, *c0, *c1); + *y0 = _mm_packs_epi32(tmp0, tmp1); + *y1 = _mm_packs_epi32(tmp2, tmp3); +} + +static INLINE void butterfly_self(__m128i *x0, __m128i *x1, const __m128i *c0, + const __m128i *c1) { + __m128i tmp0, tmp1, tmp2, tmp3, u0, u1; + const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING); + + u0 = _mm_unpacklo_epi16(*x0, *x1); + u1 = _mm_unpackhi_epi16(*x0, *x1); + BUTTERFLY_PAIR(u0, u1, *c0, *c1); + *x0 = _mm_packs_epi32(tmp0, tmp1); + *x1 = _mm_packs_epi32(tmp2, tmp3); +} + +static void idct32_34_first_half(const __m128i *in, __m128i *stp1) { + const __m128i stk2_0 = pair_set_epi16(2 * cospi_30_64, 2 * cospi_30_64); + const __m128i stk2_1 = pair_set_epi16(2 * cospi_2_64, 2 * cospi_2_64); + const __m128i stk2_6 = pair_set_epi16(-2 * cospi_26_64, -2 * cospi_26_64); + const __m128i stk2_7 = pair_set_epi16(2 * cospi_6_64, 2 * cospi_6_64); + + const __m128i stk3_0 = pair_set_epi16(2 * cospi_28_64, 2 * cospi_28_64); + const __m128i stk3_1 = pair_set_epi16(2 * cospi_4_64, 2 * cospi_4_64); + + const __m128i stg4_0 = pair_set_epi16(cospi_16_64, cospi_16_64); + const __m128i stk4_0 = pair_set_epi16(2 * cospi_16_64, 2 * cospi_16_64); + const __m128i stg4_1 = pair_set_epi16(cospi_16_64, -cospi_16_64); + const __m128i stg4_4 = pair_set_epi16(-cospi_8_64, cospi_24_64); + const __m128i stg4_5 = pair_set_epi16(cospi_24_64, cospi_8_64); + const __m128i stg4_6 = pair_set_epi16(-cospi_24_64, -cospi_8_64); + + const __m128i stg6_0 = pair_set_epi16(-cospi_16_64, cospi_16_64); + __m128i u0, u1, u2, u3, u4, u5, u6, u7; + __m128i x0, x1, x4, x5, x6, x7; + __m128i v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15; + + // phase 1 + + // 0, 15 + u2 = _mm_mulhrs_epi16(in[2], stk2_1); // stp2_15 + u3 = _mm_mulhrs_epi16(in[6], stk2_7); // stp2_12 + v15 = _mm_add_epi16(u2, u3); + // in[0], in[4] + x0 = _mm_mulhrs_epi16(in[0], stk4_0); // stp1[0] + x7 = _mm_mulhrs_epi16(in[4], stk3_1); // stp1[7] + v0 = _mm_add_epi16(x0, x7); // stp2_0 + stp1[0] = _mm_add_epi16(v0, v15); + stp1[15] = _mm_sub_epi16(v0, v15); + + // in[2], in[6] + u0 = _mm_mulhrs_epi16(in[2], stk2_0); // stp2_8 + u1 = _mm_mulhrs_epi16(in[6], stk2_6); // stp2_11 + butterfly(&u0, &u2, &stg4_4, &stg4_5, &u4, &u5); // stp2_9, stp2_14 + butterfly(&u1, &u3, &stg4_6, &stg4_4, &u6, &u7); // stp2_10, stp2_13 + + v8 = _mm_add_epi16(u0, u1); + v9 = _mm_add_epi16(u4, u6); + v10 = _mm_sub_epi16(u4, u6); + v11 = _mm_sub_epi16(u0, u1); + v12 = _mm_sub_epi16(u2, u3); + v13 = _mm_sub_epi16(u5, u7); + v14 = _mm_add_epi16(u5, u7); + + butterfly_self(&v10, &v13, &stg6_0, &stg4_0); + butterfly_self(&v11, &v12, &stg6_0, &stg4_0); + + // 1, 14 + x1 = _mm_mulhrs_epi16(in[0], stk4_0); // stp1[1], stk4_1 = stk4_0 + // stp1[2] = stp1[0], stp1[3] = stp1[1] + x4 = _mm_mulhrs_epi16(in[4], stk3_0); // stp1[4] + butterfly(&x7, &x4, &stg4_1, &stg4_0, &x5, &x6); + v1 = _mm_add_epi16(x1, x6); // stp2_1 + v2 = _mm_add_epi16(x0, x5); // stp2_2 + stp1[1] = _mm_add_epi16(v1, v14); + stp1[14] = _mm_sub_epi16(v1, v14); + + stp1[2] = _mm_add_epi16(v2, v13); + stp1[13] = _mm_sub_epi16(v2, v13); + + v3 = _mm_add_epi16(x1, x4); // stp2_3 + v4 = _mm_sub_epi16(x1, x4); // stp2_4 + + v5 = _mm_sub_epi16(x0, x5); // stp2_5 + + v6 = _mm_sub_epi16(x1, x6); // stp2_6 + v7 = _mm_sub_epi16(x0, x7); // stp2_7 + stp1[3] = _mm_add_epi16(v3, v12); + stp1[12] = _mm_sub_epi16(v3, v12); + + stp1[6] = _mm_add_epi16(v6, v9); + stp1[9] = _mm_sub_epi16(v6, v9); + + stp1[7] = _mm_add_epi16(v7, v8); + stp1[8] = _mm_sub_epi16(v7, v8); + + stp1[4] = _mm_add_epi16(v4, v11); + stp1[11] = _mm_sub_epi16(v4, v11); + + stp1[5] = _mm_add_epi16(v5, v10); + stp1[10] = _mm_sub_epi16(v5, v10); +} + +static void idct32_34_second_half(const __m128i *in, __m128i *stp1) { + const __m128i stk1_0 = pair_set_epi16(2 * cospi_31_64, 2 * cospi_31_64); + const __m128i stk1_1 = pair_set_epi16(2 * cospi_1_64, 2 * cospi_1_64); + const __m128i stk1_6 = pair_set_epi16(-2 * cospi_25_64, -2 * cospi_25_64); + const __m128i stk1_7 = pair_set_epi16(2 * cospi_7_64, 2 * cospi_7_64); + const __m128i stk1_8 = pair_set_epi16(2 * cospi_27_64, 2 * cospi_27_64); + const __m128i stk1_9 = pair_set_epi16(2 * cospi_5_64, 2 * cospi_5_64); + const __m128i stk1_14 = pair_set_epi16(-2 * cospi_29_64, -2 * cospi_29_64); + const __m128i stk1_15 = pair_set_epi16(2 * cospi_3_64, 2 * cospi_3_64); + const __m128i stg3_4 = pair_set_epi16(-cospi_4_64, cospi_28_64); + const __m128i stg3_5 = pair_set_epi16(cospi_28_64, cospi_4_64); + const __m128i stg3_6 = pair_set_epi16(-cospi_28_64, -cospi_4_64); + const __m128i stg3_8 = pair_set_epi16(-cospi_20_64, cospi_12_64); + const __m128i stg3_9 = pair_set_epi16(cospi_12_64, cospi_20_64); + const __m128i stg3_10 = pair_set_epi16(-cospi_12_64, -cospi_20_64); + + const __m128i stg4_0 = pair_set_epi16(cospi_16_64, cospi_16_64); + const __m128i stg4_4 = pair_set_epi16(-cospi_8_64, cospi_24_64); + const __m128i stg4_5 = pair_set_epi16(cospi_24_64, cospi_8_64); + const __m128i stg4_6 = pair_set_epi16(-cospi_24_64, -cospi_8_64); + + const __m128i stg6_0 = pair_set_epi16(-cospi_16_64, cospi_16_64); + __m128i v16, v17, v18, v19, v20, v21, v22, v23; + __m128i v24, v25, v26, v27, v28, v29, v30, v31; + __m128i u16, u17, u18, u19, u20, u21, u22, u23; + __m128i u24, u25, u26, u27, u28, u29, u30, u31; + + v16 = _mm_mulhrs_epi16(in[1], stk1_0); + v31 = _mm_mulhrs_epi16(in[1], stk1_1); + + v19 = _mm_mulhrs_epi16(in[7], stk1_6); + v28 = _mm_mulhrs_epi16(in[7], stk1_7); + + v20 = _mm_mulhrs_epi16(in[5], stk1_8); + v27 = _mm_mulhrs_epi16(in[5], stk1_9); + + v23 = _mm_mulhrs_epi16(in[3], stk1_14); + v24 = _mm_mulhrs_epi16(in[3], stk1_15); + + butterfly(&v16, &v31, &stg3_4, &stg3_5, &v17, &v30); + butterfly(&v19, &v28, &stg3_6, &stg3_4, &v18, &v29); + butterfly(&v20, &v27, &stg3_8, &stg3_9, &v21, &v26); + butterfly(&v23, &v24, &stg3_10, &stg3_8, &v22, &v25); + + u16 = _mm_add_epi16(v16, v19); + u17 = _mm_add_epi16(v17, v18); + u18 = _mm_sub_epi16(v17, v18); + u19 = _mm_sub_epi16(v16, v19); + u20 = _mm_sub_epi16(v23, v20); + u21 = _mm_sub_epi16(v22, v21); + u22 = _mm_add_epi16(v22, v21); + u23 = _mm_add_epi16(v23, v20); + u24 = _mm_add_epi16(v24, v27); + u27 = _mm_sub_epi16(v24, v27); + u25 = _mm_add_epi16(v25, v26); + u26 = _mm_sub_epi16(v25, v26); + u28 = _mm_sub_epi16(v31, v28); + u31 = _mm_add_epi16(v28, v31); + u29 = _mm_sub_epi16(v30, v29); + u30 = _mm_add_epi16(v29, v30); + + butterfly_self(&u18, &u29, &stg4_4, &stg4_5); + butterfly_self(&u19, &u28, &stg4_4, &stg4_5); + butterfly_self(&u20, &u27, &stg4_6, &stg4_4); + butterfly_self(&u21, &u26, &stg4_6, &stg4_4); + + stp1[16] = _mm_add_epi16(u16, u23); + stp1[23] = _mm_sub_epi16(u16, u23); + + stp1[17] = _mm_add_epi16(u17, u22); + stp1[22] = _mm_sub_epi16(u17, u22); + + stp1[18] = _mm_add_epi16(u18, u21); + stp1[21] = _mm_sub_epi16(u18, u21); + + stp1[19] = _mm_add_epi16(u19, u20); + stp1[20] = _mm_sub_epi16(u19, u20); + + stp1[24] = _mm_sub_epi16(u31, u24); + stp1[31] = _mm_add_epi16(u24, u31); + + stp1[25] = _mm_sub_epi16(u30, u25); + stp1[30] = _mm_add_epi16(u25, u30); + + stp1[26] = _mm_sub_epi16(u29, u26); + stp1[29] = _mm_add_epi16(u26, u29); + + stp1[27] = _mm_sub_epi16(u28, u27); + stp1[28] = _mm_add_epi16(u27, u28); + + butterfly_self(&stp1[20], &stp1[27], &stg6_0, &stg4_0); + butterfly_self(&stp1[21], &stp1[26], &stg6_0, &stg4_0); + butterfly_self(&stp1[22], &stp1[25], &stg6_0, &stg4_0); + butterfly_self(&stp1[23], &stp1[24], &stg6_0, &stg4_0); +} + +// Only upper-left 8x8 has non-zero coeff +void aom_idct32x32_34_add_ssse3(const tran_low_t *input, uint8_t *dest, + int stride) { + const __m128i zero = _mm_setzero_si128(); + const __m128i final_rounding = _mm_set1_epi16(1 << 5); + __m128i in[32], col[32]; + __m128i stp1[32]; + int i; + + // Load input data. Only need to load the top left 8x8 block. + in[0] = load_input_data(input); + in[1] = load_input_data(input + 32); + in[2] = load_input_data(input + 64); + in[3] = load_input_data(input + 96); + in[4] = load_input_data(input + 128); + in[5] = load_input_data(input + 160); + in[6] = load_input_data(input + 192); + in[7] = load_input_data(input + 224); + + array_transpose_8x8(in, in); + idct32_34_first_half(in, stp1); + idct32_34_second_half(in, stp1); + + // 1_D: Store 32 intermediate results for each 8x32 block. + add_sub_butterfly(stp1, col, 32); + for (i = 0; i < 4; i++) { + int j; + // Transpose 32x8 block to 8x32 block + array_transpose_8x8(col + i * 8, in); + idct32_34_first_half(in, stp1); + idct32_34_second_half(in, stp1); + + // 2_D: Calculate the results and store them to destination. + add_sub_butterfly(stp1, in, 32); + for (j = 0; j < 32; ++j) { + // Final rounding and shift + in[j] = _mm_adds_epi16(in[j], final_rounding); + in[j] = _mm_srai_epi16(in[j], 6); + RECON_AND_STORE(dest + j * stride, in[j]); + } + + dest += 8; + } +} + +// in0[16] represents the left 8x16 block +// in1[16] represents the right 8x16 block +static void load_buffer_16x16(const tran_low_t *input, __m128i *in0, + __m128i *in1) { + int i; + for (i = 0; i < 16; i++) { + in0[i] = load_input_data(input); + in1[i] = load_input_data(input + 8); + input += 32; + } +} + +static void array_transpose_16x16_2(__m128i *in0, __m128i *in1, __m128i *out0, + __m128i *out1) { + array_transpose_8x8(in0, out0); + array_transpose_8x8(&in0[8], out1); + array_transpose_8x8(in1, &out0[8]); + array_transpose_8x8(&in1[8], &out1[8]); +} + +// Group the coefficient calculation into smaller functions +// to prevent stack spillover: +// quarter_1: 0-7 +// quarter_2: 8-15 +// quarter_3_4: 16-23, 24-31 +static void idct32_8x32_135_quarter_1(const __m128i *in /*in[16]*/, + __m128i *out /*out[8]*/) { + __m128i u0, u1, u2, u3, u4, u5, u6, u7; + __m128i v0, v1, v2, v3, v4, v5, v6, v7; + + { + const __m128i stk4_0 = pair_set_epi16(2 * cospi_16_64, 2 * cospi_16_64); + const __m128i stk4_2 = pair_set_epi16(2 * cospi_24_64, 2 * cospi_24_64); + const __m128i stk4_3 = pair_set_epi16(2 * cospi_8_64, 2 * cospi_8_64); + u0 = _mm_mulhrs_epi16(in[0], stk4_0); + u2 = _mm_mulhrs_epi16(in[8], stk4_2); + u3 = _mm_mulhrs_epi16(in[8], stk4_3); + u1 = u0; + } + + v0 = _mm_add_epi16(u0, u3); + v1 = _mm_add_epi16(u1, u2); + v2 = _mm_sub_epi16(u1, u2); + v3 = _mm_sub_epi16(u0, u3); + + { + const __m128i stk3_0 = pair_set_epi16(2 * cospi_28_64, 2 * cospi_28_64); + const __m128i stk3_1 = pair_set_epi16(2 * cospi_4_64, 2 * cospi_4_64); + const __m128i stk3_2 = pair_set_epi16(-2 * cospi_20_64, -2 * cospi_20_64); + const __m128i stk3_3 = pair_set_epi16(2 * cospi_12_64, 2 * cospi_12_64); + u4 = _mm_mulhrs_epi16(in[4], stk3_0); + u7 = _mm_mulhrs_epi16(in[4], stk3_1); + u5 = _mm_mulhrs_epi16(in[12], stk3_2); + u6 = _mm_mulhrs_epi16(in[12], stk3_3); + } + + v4 = _mm_add_epi16(u4, u5); + v5 = _mm_sub_epi16(u4, u5); + v6 = _mm_sub_epi16(u7, u6); + v7 = _mm_add_epi16(u7, u6); + + { + const __m128i stg4_0 = pair_set_epi16(cospi_16_64, cospi_16_64); + const __m128i stg4_1 = pair_set_epi16(cospi_16_64, -cospi_16_64); + butterfly(&v6, &v5, &stg4_1, &stg4_0, &v5, &v6); + } + + out[0] = _mm_add_epi16(v0, v7); + out[1] = _mm_add_epi16(v1, v6); + out[2] = _mm_add_epi16(v2, v5); + out[3] = _mm_add_epi16(v3, v4); + out[4] = _mm_sub_epi16(v3, v4); + out[5] = _mm_sub_epi16(v2, v5); + out[6] = _mm_sub_epi16(v1, v6); + out[7] = _mm_sub_epi16(v0, v7); +} + +static void idct32_8x32_135_quarter_2(const __m128i *in /*in[16]*/, + __m128i *out /*out[8]*/) { + __m128i u8, u9, u10, u11, u12, u13, u14, u15; + __m128i v8, v9, v10, v11, v12, v13, v14, v15; + + { + const __m128i stk2_0 = pair_set_epi16(2 * cospi_30_64, 2 * cospi_30_64); + const __m128i stk2_1 = pair_set_epi16(2 * cospi_2_64, 2 * cospi_2_64); + const __m128i stk2_2 = pair_set_epi16(-2 * cospi_18_64, -2 * cospi_18_64); + const __m128i stk2_3 = pair_set_epi16(2 * cospi_14_64, 2 * cospi_14_64); + const __m128i stk2_4 = pair_set_epi16(2 * cospi_22_64, 2 * cospi_22_64); + const __m128i stk2_5 = pair_set_epi16(2 * cospi_10_64, 2 * cospi_10_64); + const __m128i stk2_6 = pair_set_epi16(-2 * cospi_26_64, -2 * cospi_26_64); + const __m128i stk2_7 = pair_set_epi16(2 * cospi_6_64, 2 * cospi_6_64); + u8 = _mm_mulhrs_epi16(in[2], stk2_0); + u15 = _mm_mulhrs_epi16(in[2], stk2_1); + u9 = _mm_mulhrs_epi16(in[14], stk2_2); + u14 = _mm_mulhrs_epi16(in[14], stk2_3); + u10 = _mm_mulhrs_epi16(in[10], stk2_4); + u13 = _mm_mulhrs_epi16(in[10], stk2_5); + u11 = _mm_mulhrs_epi16(in[6], stk2_6); + u12 = _mm_mulhrs_epi16(in[6], stk2_7); + } + + v8 = _mm_add_epi16(u8, u9); + v9 = _mm_sub_epi16(u8, u9); + v10 = _mm_sub_epi16(u11, u10); + v11 = _mm_add_epi16(u11, u10); + v12 = _mm_add_epi16(u12, u13); + v13 = _mm_sub_epi16(u12, u13); + v14 = _mm_sub_epi16(u15, u14); + v15 = _mm_add_epi16(u15, u14); + + { + const __m128i stg4_4 = pair_set_epi16(-cospi_8_64, cospi_24_64); + const __m128i stg4_5 = pair_set_epi16(cospi_24_64, cospi_8_64); + const __m128i stg4_6 = pair_set_epi16(-cospi_24_64, -cospi_8_64); + butterfly_self(&v9, &v14, &stg4_4, &stg4_5); + butterfly_self(&v10, &v13, &stg4_6, &stg4_4); + } + + out[0] = _mm_add_epi16(v8, v11); + out[1] = _mm_add_epi16(v9, v10); + out[2] = _mm_sub_epi16(v9, v10); + out[3] = _mm_sub_epi16(v8, v11); + out[4] = _mm_sub_epi16(v15, v12); + out[5] = _mm_sub_epi16(v14, v13); + out[6] = _mm_add_epi16(v14, v13); + out[7] = _mm_add_epi16(v15, v12); + + { + const __m128i stg4_0 = pair_set_epi16(cospi_16_64, cospi_16_64); + const __m128i stg6_0 = pair_set_epi16(-cospi_16_64, cospi_16_64); + butterfly_self(&out[2], &out[5], &stg6_0, &stg4_0); + butterfly_self(&out[3], &out[4], &stg6_0, &stg4_0); + } +} + +// 8x32 block even indexed 8 inputs of in[16], +// output first half 16 to out[32] +static void idct32_8x32_quarter_1_2(const __m128i *in /*in[16]*/, + __m128i *out /*out[32]*/) { + __m128i temp[16]; + idct32_8x32_135_quarter_1(in, temp); + idct32_8x32_135_quarter_2(in, &temp[8]); + add_sub_butterfly(temp, out, 16); +} + +// 8x32 block odd indexed 8 inputs of in[16], +// output second half 16 to out[32] +static void idct32_8x32_quarter_3_4(const __m128i *in /*in[16]*/, + __m128i *out /*out[32]*/) { + __m128i v16, v17, v18, v19, v20, v21, v22, v23; + __m128i v24, v25, v26, v27, v28, v29, v30, v31; + __m128i u16, u17, u18, u19, u20, u21, u22, u23; + __m128i u24, u25, u26, u27, u28, u29, u30, u31; + + { + const __m128i stk1_0 = pair_set_epi16(2 * cospi_31_64, 2 * cospi_31_64); + const __m128i stk1_1 = pair_set_epi16(2 * cospi_1_64, 2 * cospi_1_64); + const __m128i stk1_2 = pair_set_epi16(-2 * cospi_17_64, -2 * cospi_17_64); + const __m128i stk1_3 = pair_set_epi16(2 * cospi_15_64, 2 * cospi_15_64); + + const __m128i stk1_4 = pair_set_epi16(2 * cospi_23_64, 2 * cospi_23_64); + const __m128i stk1_5 = pair_set_epi16(2 * cospi_9_64, 2 * cospi_9_64); + const __m128i stk1_6 = pair_set_epi16(-2 * cospi_25_64, -2 * cospi_25_64); + const __m128i stk1_7 = pair_set_epi16(2 * cospi_7_64, 2 * cospi_7_64); + const __m128i stk1_8 = pair_set_epi16(2 * cospi_27_64, 2 * cospi_27_64); + const __m128i stk1_9 = pair_set_epi16(2 * cospi_5_64, 2 * cospi_5_64); + const __m128i stk1_10 = pair_set_epi16(-2 * cospi_21_64, -2 * cospi_21_64); + const __m128i stk1_11 = pair_set_epi16(2 * cospi_11_64, 2 * cospi_11_64); + + const __m128i stk1_12 = pair_set_epi16(2 * cospi_19_64, 2 * cospi_19_64); + const __m128i stk1_13 = pair_set_epi16(2 * cospi_13_64, 2 * cospi_13_64); + const __m128i stk1_14 = pair_set_epi16(-2 * cospi_29_64, -2 * cospi_29_64); + const __m128i stk1_15 = pair_set_epi16(2 * cospi_3_64, 2 * cospi_3_64); + u16 = _mm_mulhrs_epi16(in[1], stk1_0); + u31 = _mm_mulhrs_epi16(in[1], stk1_1); + u17 = _mm_mulhrs_epi16(in[15], stk1_2); + u30 = _mm_mulhrs_epi16(in[15], stk1_3); + + u18 = _mm_mulhrs_epi16(in[9], stk1_4); + u29 = _mm_mulhrs_epi16(in[9], stk1_5); + u19 = _mm_mulhrs_epi16(in[7], stk1_6); + u28 = _mm_mulhrs_epi16(in[7], stk1_7); + + u20 = _mm_mulhrs_epi16(in[5], stk1_8); + u27 = _mm_mulhrs_epi16(in[5], stk1_9); + u21 = _mm_mulhrs_epi16(in[11], stk1_10); + u26 = _mm_mulhrs_epi16(in[11], stk1_11); + + u22 = _mm_mulhrs_epi16(in[13], stk1_12); + u25 = _mm_mulhrs_epi16(in[13], stk1_13); + u23 = _mm_mulhrs_epi16(in[3], stk1_14); + u24 = _mm_mulhrs_epi16(in[3], stk1_15); + } + + v16 = _mm_add_epi16(u16, u17); + v17 = _mm_sub_epi16(u16, u17); + v18 = _mm_sub_epi16(u19, u18); + v19 = _mm_add_epi16(u19, u18); + + v20 = _mm_add_epi16(u20, u21); + v21 = _mm_sub_epi16(u20, u21); + v22 = _mm_sub_epi16(u23, u22); + v23 = _mm_add_epi16(u23, u22); + + v24 = _mm_add_epi16(u24, u25); + v25 = _mm_sub_epi16(u24, u25); + v26 = _mm_sub_epi16(u27, u26); + v27 = _mm_add_epi16(u27, u26); + + v28 = _mm_add_epi16(u28, u29); + v29 = _mm_sub_epi16(u28, u29); + v30 = _mm_sub_epi16(u31, u30); + v31 = _mm_add_epi16(u31, u30); + + { + const __m128i stg3_4 = pair_set_epi16(-cospi_4_64, cospi_28_64); + const __m128i stg3_5 = pair_set_epi16(cospi_28_64, cospi_4_64); + const __m128i stg3_6 = pair_set_epi16(-cospi_28_64, -cospi_4_64); + const __m128i stg3_8 = pair_set_epi16(-cospi_20_64, cospi_12_64); + const __m128i stg3_9 = pair_set_epi16(cospi_12_64, cospi_20_64); + const __m128i stg3_10 = pair_set_epi16(-cospi_12_64, -cospi_20_64); + + butterfly_self(&v17, &v30, &stg3_4, &stg3_5); + butterfly_self(&v18, &v29, &stg3_6, &stg3_4); + butterfly_self(&v21, &v26, &stg3_8, &stg3_9); + butterfly_self(&v22, &v25, &stg3_10, &stg3_8); + } + + u16 = _mm_add_epi16(v16, v19); + u17 = _mm_add_epi16(v17, v18); + u18 = _mm_sub_epi16(v17, v18); + u19 = _mm_sub_epi16(v16, v19); + u20 = _mm_sub_epi16(v23, v20); + u21 = _mm_sub_epi16(v22, v21); + u22 = _mm_add_epi16(v22, v21); + u23 = _mm_add_epi16(v23, v20); + + u24 = _mm_add_epi16(v24, v27); + u25 = _mm_add_epi16(v25, v26); + u26 = _mm_sub_epi16(v25, v26); + u27 = _mm_sub_epi16(v24, v27); + u28 = _mm_sub_epi16(v31, v28); + u29 = _mm_sub_epi16(v30, v29); + u30 = _mm_add_epi16(v29, v30); + u31 = _mm_add_epi16(v28, v31); + + { + const __m128i stg4_4 = pair_set_epi16(-cospi_8_64, cospi_24_64); + const __m128i stg4_5 = pair_set_epi16(cospi_24_64, cospi_8_64); + const __m128i stg4_6 = pair_set_epi16(-cospi_24_64, -cospi_8_64); + butterfly_self(&u18, &u29, &stg4_4, &stg4_5); + butterfly_self(&u19, &u28, &stg4_4, &stg4_5); + butterfly_self(&u20, &u27, &stg4_6, &stg4_4); + butterfly_self(&u21, &u26, &stg4_6, &stg4_4); + } + + out[0] = _mm_add_epi16(u16, u23); + out[1] = _mm_add_epi16(u17, u22); + out[2] = _mm_add_epi16(u18, u21); + out[3] = _mm_add_epi16(u19, u20); + v20 = _mm_sub_epi16(u19, u20); + v21 = _mm_sub_epi16(u18, u21); + v22 = _mm_sub_epi16(u17, u22); + v23 = _mm_sub_epi16(u16, u23); + + v24 = _mm_sub_epi16(u31, u24); + v25 = _mm_sub_epi16(u30, u25); + v26 = _mm_sub_epi16(u29, u26); + v27 = _mm_sub_epi16(u28, u27); + out[12] = _mm_add_epi16(u27, u28); + out[13] = _mm_add_epi16(u26, u29); + out[14] = _mm_add_epi16(u25, u30); + out[15] = _mm_add_epi16(u24, u31); + + { + const __m128i stg4_0 = pair_set_epi16(cospi_16_64, cospi_16_64); + const __m128i stg6_0 = pair_set_epi16(-cospi_16_64, cospi_16_64); + butterfly(&v20, &v27, &stg6_0, &stg4_0, &out[4], &out[11]); + butterfly(&v21, &v26, &stg6_0, &stg4_0, &out[5], &out[10]); + butterfly(&v22, &v25, &stg6_0, &stg4_0, &out[6], &out[9]); + butterfly(&v23, &v24, &stg6_0, &stg4_0, &out[7], &out[8]); + } +} + +// 8x16 block, input __m128i in[16], output __m128i in[32] +static void idct32_8x32_135(__m128i *in /*in[32]*/) { + __m128i out[32]; + idct32_8x32_quarter_1_2(in, out); + idct32_8x32_quarter_3_4(in, &out[16]); + add_sub_butterfly(out, in, 32); +} + +static INLINE void store_buffer_8x32(__m128i *in, uint8_t *dst, int stride) { + const __m128i final_rounding = _mm_set1_epi16(1 << 5); + const __m128i zero = _mm_setzero_si128(); + int j = 0; + while (j < 32) { + in[j] = _mm_adds_epi16(in[j], final_rounding); + in[j + 1] = _mm_adds_epi16(in[j + 1], final_rounding); + + in[j] = _mm_srai_epi16(in[j], 6); + in[j + 1] = _mm_srai_epi16(in[j + 1], 6); + + RECON_AND_STORE(dst, in[j]); + dst += stride; + RECON_AND_STORE(dst, in[j + 1]); + dst += stride; + j += 2; + } +} + +static INLINE void recon_and_store(__m128i *in0, __m128i *in1, uint8_t *dest, + int stride) { + store_buffer_8x32(in0, dest, stride); + store_buffer_8x32(in1, dest + 8, stride); +} + +static INLINE void idct32_135(__m128i *col0, __m128i *col1) { + idct32_8x32_135(col0); + idct32_8x32_135(col1); +} + +typedef enum { left_16, right_16 } ColsIndicator; + +static void transpose_and_copy_16x16(__m128i *in0, __m128i *in1, __m128i *store, + ColsIndicator cols) { + switch (cols) { + case left_16: { + int i; + array_transpose_16x16(in0, in1); + for (i = 0; i < 16; ++i) { + store[i] = in0[16 + i]; + store[16 + i] = in1[16 + i]; + } + break; + } + case right_16: { + array_transpose_16x16_2(store, &store[16], in0, in1); + break; + } + default: { assert(0); } + } +} + +// Only upper-left 16x16 has non-zero coeff +void aom_idct32x32_135_add_ssse3(const tran_low_t *input, uint8_t *dest, + int stride) { + // Each array represents an 8x32 block + __m128i col0[32], col1[32]; + // This array represents a 16x16 block + __m128i temp[32]; + + // Load input data. Only need to load the top left 16x16 block. + load_buffer_16x16(input, col0, col1); + + // columns + array_transpose_16x16(col0, col1); + idct32_135(col0, col1); + + // rows + transpose_and_copy_16x16(col0, col1, temp, left_16); + idct32_135(col0, col1); + recon_and_store(col0, col1, dest, stride); + + transpose_and_copy_16x16(col0, col1, temp, right_16); + idct32_135(col0, col1); + recon_and_store(col0, col1, dest + 16, stride); +} + +// For each 8x32 block __m128i in[32], +// Input with index, 2, 6, 10, 14, 18, 22, 26, 30 +// output pixels: 8-15 in __m128i in[32] +static void idct32_full_8x32_quarter_2(const __m128i *in /*in[32]*/, + __m128i *out /*out[16]*/) { + __m128i u8, u9, u10, u11, u12, u13, u14, u15; // stp2_ + __m128i v8, v9, v10, v11, v12, v13, v14, v15; // stp1_ + + { + const __m128i stg2_0 = pair_set_epi16(cospi_30_64, -cospi_2_64); + const __m128i stg2_1 = pair_set_epi16(cospi_2_64, cospi_30_64); + const __m128i stg2_2 = pair_set_epi16(cospi_14_64, -cospi_18_64); + const __m128i stg2_3 = pair_set_epi16(cospi_18_64, cospi_14_64); + butterfly(&in[2], &in[30], &stg2_0, &stg2_1, &u8, &u15); + butterfly(&in[18], &in[14], &stg2_2, &stg2_3, &u9, &u14); + } + + v8 = _mm_add_epi16(u8, u9); + v9 = _mm_sub_epi16(u8, u9); + v14 = _mm_sub_epi16(u15, u14); + v15 = _mm_add_epi16(u15, u14); + + { + const __m128i stg2_4 = pair_set_epi16(cospi_22_64, -cospi_10_64); + const __m128i stg2_5 = pair_set_epi16(cospi_10_64, cospi_22_64); + const __m128i stg2_6 = pair_set_epi16(cospi_6_64, -cospi_26_64); + const __m128i stg2_7 = pair_set_epi16(cospi_26_64, cospi_6_64); + butterfly(&in[10], &in[22], &stg2_4, &stg2_5, &u10, &u13); + butterfly(&in[26], &in[6], &stg2_6, &stg2_7, &u11, &u12); + } + + v10 = _mm_sub_epi16(u11, u10); + v11 = _mm_add_epi16(u11, u10); + v12 = _mm_add_epi16(u12, u13); + v13 = _mm_sub_epi16(u12, u13); + + { + const __m128i stg4_4 = pair_set_epi16(-cospi_8_64, cospi_24_64); + const __m128i stg4_5 = pair_set_epi16(cospi_24_64, cospi_8_64); + const __m128i stg4_6 = pair_set_epi16(-cospi_24_64, -cospi_8_64); + butterfly_self(&v9, &v14, &stg4_4, &stg4_5); + butterfly_self(&v10, &v13, &stg4_6, &stg4_4); + } + + out[0] = _mm_add_epi16(v8, v11); + out[1] = _mm_add_epi16(v9, v10); + out[6] = _mm_add_epi16(v14, v13); + out[7] = _mm_add_epi16(v15, v12); + + out[2] = _mm_sub_epi16(v9, v10); + out[3] = _mm_sub_epi16(v8, v11); + out[4] = _mm_sub_epi16(v15, v12); + out[5] = _mm_sub_epi16(v14, v13); + + { + const __m128i stg4_0 = pair_set_epi16(cospi_16_64, cospi_16_64); + const __m128i stg6_0 = pair_set_epi16(-cospi_16_64, cospi_16_64); + butterfly_self(&out[2], &out[5], &stg6_0, &stg4_0); + butterfly_self(&out[3], &out[4], &stg6_0, &stg4_0); + } +} + +// For each 8x32 block __m128i in[32], +// Input with index, 0, 4, 8, 12, 16, 20, 24, 28 +// output pixels: 0-7 in __m128i in[32] +static void idct32_full_8x32_quarter_1(const __m128i *in /*in[32]*/, + __m128i *out /*out[8]*/) { + __m128i u0, u1, u2, u3, u4, u5, u6, u7; // stp1_ + __m128i v0, v1, v2, v3, v4, v5, v6, v7; // stp2_ + + { + const __m128i stg3_0 = pair_set_epi16(cospi_28_64, -cospi_4_64); + const __m128i stg3_1 = pair_set_epi16(cospi_4_64, cospi_28_64); + const __m128i stg3_2 = pair_set_epi16(cospi_12_64, -cospi_20_64); + const __m128i stg3_3 = pair_set_epi16(cospi_20_64, cospi_12_64); + butterfly(&in[4], &in[28], &stg3_0, &stg3_1, &u4, &u7); + butterfly(&in[20], &in[12], &stg3_2, &stg3_3, &u5, &u6); + } + + v4 = _mm_add_epi16(u4, u5); + v5 = _mm_sub_epi16(u4, u5); + v6 = _mm_sub_epi16(u7, u6); + v7 = _mm_add_epi16(u7, u6); + + { + const __m128i stg4_0 = pair_set_epi16(cospi_16_64, cospi_16_64); + const __m128i stg4_1 = pair_set_epi16(cospi_16_64, -cospi_16_64); + const __m128i stg4_2 = pair_set_epi16(cospi_24_64, -cospi_8_64); + const __m128i stg4_3 = pair_set_epi16(cospi_8_64, cospi_24_64); + butterfly(&v6, &v5, &stg4_1, &stg4_0, &v5, &v6); + + butterfly(&in[0], &in[16], &stg4_0, &stg4_1, &u0, &u1); + butterfly(&in[8], &in[24], &stg4_2, &stg4_3, &u2, &u3); + } + + v0 = _mm_add_epi16(u0, u3); + v1 = _mm_add_epi16(u1, u2); + v2 = _mm_sub_epi16(u1, u2); + v3 = _mm_sub_epi16(u0, u3); + + out[0] = _mm_add_epi16(v0, v7); + out[1] = _mm_add_epi16(v1, v6); + out[2] = _mm_add_epi16(v2, v5); + out[3] = _mm_add_epi16(v3, v4); + out[4] = _mm_sub_epi16(v3, v4); + out[5] = _mm_sub_epi16(v2, v5); + out[6] = _mm_sub_epi16(v1, v6); + out[7] = _mm_sub_epi16(v0, v7); +} + +// For each 8x32 block __m128i in[32], +// Input with odd index, +// 1, 3, 5, 7, 9, 11, 13, 15, 17, 19, 21, 23, 25, 27, 29, 31 +// output pixels: 16-23, 24-31 in __m128i in[32] +// We avoid hide an offset, 16, inside this function. So we output 0-15 into +// array out[16] +static void idct32_full_8x32_quarter_3_4(const __m128i *in /*in[32]*/, + __m128i *out /*out[16]*/) { + __m128i v16, v17, v18, v19, v20, v21, v22, v23; + __m128i v24, v25, v26, v27, v28, v29, v30, v31; + __m128i u16, u17, u18, u19, u20, u21, u22, u23; + __m128i u24, u25, u26, u27, u28, u29, u30, u31; + + { + const __m128i stg1_0 = pair_set_epi16(cospi_31_64, -cospi_1_64); + const __m128i stg1_1 = pair_set_epi16(cospi_1_64, cospi_31_64); + const __m128i stg1_2 = pair_set_epi16(cospi_15_64, -cospi_17_64); + const __m128i stg1_3 = pair_set_epi16(cospi_17_64, cospi_15_64); + const __m128i stg1_4 = pair_set_epi16(cospi_23_64, -cospi_9_64); + const __m128i stg1_5 = pair_set_epi16(cospi_9_64, cospi_23_64); + const __m128i stg1_6 = pair_set_epi16(cospi_7_64, -cospi_25_64); + const __m128i stg1_7 = pair_set_epi16(cospi_25_64, cospi_7_64); + const __m128i stg1_8 = pair_set_epi16(cospi_27_64, -cospi_5_64); + const __m128i stg1_9 = pair_set_epi16(cospi_5_64, cospi_27_64); + const __m128i stg1_10 = pair_set_epi16(cospi_11_64, -cospi_21_64); + const __m128i stg1_11 = pair_set_epi16(cospi_21_64, cospi_11_64); + const __m128i stg1_12 = pair_set_epi16(cospi_19_64, -cospi_13_64); + const __m128i stg1_13 = pair_set_epi16(cospi_13_64, cospi_19_64); + const __m128i stg1_14 = pair_set_epi16(cospi_3_64, -cospi_29_64); + const __m128i stg1_15 = pair_set_epi16(cospi_29_64, cospi_3_64); + butterfly(&in[1], &in[31], &stg1_0, &stg1_1, &u16, &u31); + butterfly(&in[17], &in[15], &stg1_2, &stg1_3, &u17, &u30); + butterfly(&in[9], &in[23], &stg1_4, &stg1_5, &u18, &u29); + butterfly(&in[25], &in[7], &stg1_6, &stg1_7, &u19, &u28); + + butterfly(&in[5], &in[27], &stg1_8, &stg1_9, &u20, &u27); + butterfly(&in[21], &in[11], &stg1_10, &stg1_11, &u21, &u26); + + butterfly(&in[13], &in[19], &stg1_12, &stg1_13, &u22, &u25); + butterfly(&in[29], &in[3], &stg1_14, &stg1_15, &u23, &u24); + } + + v16 = _mm_add_epi16(u16, u17); + v17 = _mm_sub_epi16(u16, u17); + v18 = _mm_sub_epi16(u19, u18); + v19 = _mm_add_epi16(u19, u18); + + v20 = _mm_add_epi16(u20, u21); + v21 = _mm_sub_epi16(u20, u21); + v22 = _mm_sub_epi16(u23, u22); + v23 = _mm_add_epi16(u23, u22); + + v24 = _mm_add_epi16(u24, u25); + v25 = _mm_sub_epi16(u24, u25); + v26 = _mm_sub_epi16(u27, u26); + v27 = _mm_add_epi16(u27, u26); + + v28 = _mm_add_epi16(u28, u29); + v29 = _mm_sub_epi16(u28, u29); + v30 = _mm_sub_epi16(u31, u30); + v31 = _mm_add_epi16(u31, u30); + + { + const __m128i stg3_4 = pair_set_epi16(-cospi_4_64, cospi_28_64); + const __m128i stg3_5 = pair_set_epi16(cospi_28_64, cospi_4_64); + const __m128i stg3_6 = pair_set_epi16(-cospi_28_64, -cospi_4_64); + const __m128i stg3_8 = pair_set_epi16(-cospi_20_64, cospi_12_64); + const __m128i stg3_9 = pair_set_epi16(cospi_12_64, cospi_20_64); + const __m128i stg3_10 = pair_set_epi16(-cospi_12_64, -cospi_20_64); + butterfly_self(&v17, &v30, &stg3_4, &stg3_5); + butterfly_self(&v18, &v29, &stg3_6, &stg3_4); + butterfly_self(&v21, &v26, &stg3_8, &stg3_9); + butterfly_self(&v22, &v25, &stg3_10, &stg3_8); + } + + u16 = _mm_add_epi16(v16, v19); + u17 = _mm_add_epi16(v17, v18); + u18 = _mm_sub_epi16(v17, v18); + u19 = _mm_sub_epi16(v16, v19); + u20 = _mm_sub_epi16(v23, v20); + u21 = _mm_sub_epi16(v22, v21); + u22 = _mm_add_epi16(v22, v21); + u23 = _mm_add_epi16(v23, v20); + + u24 = _mm_add_epi16(v24, v27); + u25 = _mm_add_epi16(v25, v26); + u26 = _mm_sub_epi16(v25, v26); + u27 = _mm_sub_epi16(v24, v27); + + u28 = _mm_sub_epi16(v31, v28); + u29 = _mm_sub_epi16(v30, v29); + u30 = _mm_add_epi16(v29, v30); + u31 = _mm_add_epi16(v28, v31); + + { + const __m128i stg4_4 = pair_set_epi16(-cospi_8_64, cospi_24_64); + const __m128i stg4_5 = pair_set_epi16(cospi_24_64, cospi_8_64); + const __m128i stg4_6 = pair_set_epi16(-cospi_24_64, -cospi_8_64); + butterfly_self(&u18, &u29, &stg4_4, &stg4_5); + butterfly_self(&u19, &u28, &stg4_4, &stg4_5); + butterfly_self(&u20, &u27, &stg4_6, &stg4_4); + butterfly_self(&u21, &u26, &stg4_6, &stg4_4); + } + + out[0] = _mm_add_epi16(u16, u23); + out[1] = _mm_add_epi16(u17, u22); + out[2] = _mm_add_epi16(u18, u21); + out[3] = _mm_add_epi16(u19, u20); + out[4] = _mm_sub_epi16(u19, u20); + out[5] = _mm_sub_epi16(u18, u21); + out[6] = _mm_sub_epi16(u17, u22); + out[7] = _mm_sub_epi16(u16, u23); + + out[8] = _mm_sub_epi16(u31, u24); + out[9] = _mm_sub_epi16(u30, u25); + out[10] = _mm_sub_epi16(u29, u26); + out[11] = _mm_sub_epi16(u28, u27); + out[12] = _mm_add_epi16(u27, u28); + out[13] = _mm_add_epi16(u26, u29); + out[14] = _mm_add_epi16(u25, u30); + out[15] = _mm_add_epi16(u24, u31); + + { + const __m128i stg4_0 = pair_set_epi16(cospi_16_64, cospi_16_64); + const __m128i stg6_0 = pair_set_epi16(-cospi_16_64, cospi_16_64); + butterfly_self(&out[4], &out[11], &stg6_0, &stg4_0); + butterfly_self(&out[5], &out[10], &stg6_0, &stg4_0); + butterfly_self(&out[6], &out[9], &stg6_0, &stg4_0); + butterfly_self(&out[7], &out[8], &stg6_0, &stg4_0); + } +} + +static void idct32_full_8x32_quarter_1_2(const __m128i *in /*in[32]*/, + __m128i *out /*out[32]*/) { + __m128i temp[16]; + idct32_full_8x32_quarter_1(in, temp); + idct32_full_8x32_quarter_2(in, &temp[8]); + add_sub_butterfly(temp, out, 16); +} + +static void idct32_full_8x32(const __m128i *in /*in[32]*/, + __m128i *out /*out[32]*/) { + __m128i temp[32]; + idct32_full_8x32_quarter_1_2(in, temp); + idct32_full_8x32_quarter_3_4(in, &temp[16]); + add_sub_butterfly(temp, out, 32); +} + +static void load_buffer_8x32(const tran_low_t *input, __m128i *in) { + int i; + for (i = 0; i < 8; ++i) { + in[i] = load_input_data(input); + in[i + 8] = load_input_data(input + 8); + in[i + 16] = load_input_data(input + 16); + in[i + 24] = load_input_data(input + 24); + input += 32; + } +} + +void aom_idct32x32_1024_add_ssse3(const tran_low_t *input, uint8_t *dest, + int stride) { + __m128i col[128], in[32]; + int i, j; + + // rows + for (i = 0; i < 4; ++i) { + load_buffer_8x32(input, in); + input += 32 << 3; + + // Transpose 32x8 block to 8x32 block + array_transpose_8x8(in, in); + array_transpose_8x8(in + 8, in + 8); + array_transpose_8x8(in + 16, in + 16); + array_transpose_8x8(in + 24, in + 24); + + idct32_full_8x32(in, col + (i << 5)); + } + + // columns + for (i = 0; i < 4; ++i) { + j = i << 3; + // Transpose 32x8 block to 8x32 block + array_transpose_8x8(col + j, in); + array_transpose_8x8(col + j + 32, in + 8); + array_transpose_8x8(col + j + 64, in + 16); + array_transpose_8x8(col + j + 96, in + 24); + + idct32_full_8x32(in, in); + store_buffer_8x32(in, dest, stride); + dest += 8; + } +} diff --git a/third_party/aom/aom_dsp/x86/inv_wht_sse2.asm b/third_party/aom/aom_dsp/x86/inv_wht_sse2.asm new file mode 100644 index 000000000..f0668e6f3 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/inv_wht_sse2.asm @@ -0,0 +1,112 @@ +; +; Copyright (c) 2016, Alliance for Open Media. All rights reserved +; +; This source code is subject to the terms of the BSD 2 Clause License and +; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License +; was not distributed with this source code in the LICENSE file, you can +; obtain it at www.aomedia.org/license/software. If the Alliance for Open +; Media Patent License 1.0 was not distributed with this source code in the +; PATENTS file, you can obtain it at www.aomedia.org/license/patent. +; + +; + +%include "third_party/x86inc/x86inc.asm" + +SECTION .text + +%macro REORDER_INPUTS 0 + ; a c d b to a b c d + SWAP 1, 3, 2 +%endmacro + +%macro TRANSFORM_COLS 0 + ; input: + ; m0 a + ; m1 b + ; m2 c + ; m3 d + paddw m0, m2 + psubw m3, m1 + + ; wide subtract + punpcklwd m4, m0 + punpcklwd m5, m3 + psrad m4, 16 + psrad m5, 16 + psubd m4, m5 + psrad m4, 1 + packssdw m4, m4 ; e + + psubw m5, m4, m1 ; b + psubw m4, m2 ; c + psubw m0, m5 + paddw m3, m4 + ; m0 a + SWAP 1, 5 ; m1 b + SWAP 2, 4 ; m2 c + ; m3 d +%endmacro + +%macro TRANSPOSE_4X4 0 + punpcklwd m0, m2 + punpcklwd m1, m3 + mova m2, m0 + punpcklwd m0, m1 + punpckhwd m2, m1 + pshufd m1, m0, 0x0e + pshufd m3, m2, 0x0e +%endmacro + +; transpose a 4x4 int16 matrix in xmm0 and xmm1 to the bottom half of xmm0-xmm3 +%macro TRANSPOSE_4X4_WIDE 0 + mova m3, m0 + punpcklwd m0, m1 + punpckhwd m3, m1 + mova m2, m0 + punpcklwd m0, m3 + punpckhwd m2, m3 + pshufd m1, m0, 0x0e + pshufd m3, m2, 0x0e +%endmacro + +%macro ADD_STORE_4P_2X 5 ; src1, src2, tmp1, tmp2, zero + movd m%3, [outputq] + movd m%4, [outputq + strideq] + punpcklbw m%3, m%5 + punpcklbw m%4, m%5 + paddw m%1, m%3 + paddw m%2, m%4 + packuswb m%1, m%5 + packuswb m%2, m%5 + movd [outputq], m%1 + movd [outputq + strideq], m%2 +%endmacro + +INIT_XMM sse2 +cglobal iwht4x4_16_add, 3, 3, 7, input, output, stride +%if CONFIG_HIGHBITDEPTH + mova m0, [inputq + 0] + packssdw m0, [inputq + 16] + mova m1, [inputq + 32] + packssdw m1, [inputq + 48] +%else + mova m0, [inputq + 0] + mova m1, [inputq + 16] +%endif + psraw m0, 2 + psraw m1, 2 + + TRANSPOSE_4X4_WIDE + REORDER_INPUTS + TRANSFORM_COLS + TRANSPOSE_4X4 + REORDER_INPUTS + TRANSFORM_COLS + + pxor m4, m4 + ADD_STORE_4P_2X 0, 1, 5, 6, 4 + lea outputq, [outputq + 2 * strideq] + ADD_STORE_4P_2X 2, 3, 5, 6, 4 + + RET diff --git a/third_party/aom/aom_dsp/x86/loopfilter_avx2.c b/third_party/aom/aom_dsp/x86/loopfilter_avx2.c new file mode 100644 index 000000000..bf8150e2a --- /dev/null +++ b/third_party/aom/aom_dsp/x86/loopfilter_avx2.c @@ -0,0 +1,915 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include <immintrin.h> /* AVX2 */ + +#include "./aom_dsp_rtcd.h" +#include "aom_ports/mem.h" + +void aom_lpf_horizontal_edge_8_avx2(unsigned char *s, int p, + const unsigned char *_blimit, + const unsigned char *_limit, + const unsigned char *_thresh) { + __m128i mask, hev, flat, flat2; + const __m128i zero = _mm_set1_epi16(0); + const __m128i one = _mm_set1_epi8(1); + __m128i q7p7, q6p6, q5p5, q4p4, q3p3, q2p2, q1p1, q0p0, p0q0, p1q1; + __m128i abs_p1p0; + + const __m128i thresh = + _mm_broadcastb_epi8(_mm_cvtsi32_si128((int)_thresh[0])); + const __m128i limit = _mm_broadcastb_epi8(_mm_cvtsi32_si128((int)_limit[0])); + const __m128i blimit = + _mm_broadcastb_epi8(_mm_cvtsi32_si128((int)_blimit[0])); + + q4p4 = _mm_loadl_epi64((__m128i *)(s - 5 * p)); + q4p4 = _mm_castps_si128( + _mm_loadh_pi(_mm_castsi128_ps(q4p4), (__m64 *)(s + 4 * p))); + q3p3 = _mm_loadl_epi64((__m128i *)(s - 4 * p)); + q3p3 = _mm_castps_si128( + _mm_loadh_pi(_mm_castsi128_ps(q3p3), (__m64 *)(s + 3 * p))); + q2p2 = _mm_loadl_epi64((__m128i *)(s - 3 * p)); + q2p2 = _mm_castps_si128( + _mm_loadh_pi(_mm_castsi128_ps(q2p2), (__m64 *)(s + 2 * p))); + q1p1 = _mm_loadl_epi64((__m128i *)(s - 2 * p)); + q1p1 = _mm_castps_si128( + _mm_loadh_pi(_mm_castsi128_ps(q1p1), (__m64 *)(s + 1 * p))); + p1q1 = _mm_shuffle_epi32(q1p1, 78); + q0p0 = _mm_loadl_epi64((__m128i *)(s - 1 * p)); + q0p0 = _mm_castps_si128( + _mm_loadh_pi(_mm_castsi128_ps(q0p0), (__m64 *)(s - 0 * p))); + p0q0 = _mm_shuffle_epi32(q0p0, 78); + + { + __m128i abs_p1q1, abs_p0q0, abs_q1q0, fe, ff, work; + abs_p1p0 = + _mm_or_si128(_mm_subs_epu8(q1p1, q0p0), _mm_subs_epu8(q0p0, q1p1)); + abs_q1q0 = _mm_srli_si128(abs_p1p0, 8); + fe = _mm_set1_epi8(0xfe); + ff = _mm_cmpeq_epi8(abs_p1p0, abs_p1p0); + abs_p0q0 = + _mm_or_si128(_mm_subs_epu8(q0p0, p0q0), _mm_subs_epu8(p0q0, q0p0)); + abs_p1q1 = + _mm_or_si128(_mm_subs_epu8(q1p1, p1q1), _mm_subs_epu8(p1q1, q1p1)); + flat = _mm_max_epu8(abs_p1p0, abs_q1q0); + hev = _mm_subs_epu8(flat, thresh); + hev = _mm_xor_si128(_mm_cmpeq_epi8(hev, zero), ff); + + abs_p0q0 = _mm_adds_epu8(abs_p0q0, abs_p0q0); + abs_p1q1 = _mm_srli_epi16(_mm_and_si128(abs_p1q1, fe), 1); + mask = _mm_subs_epu8(_mm_adds_epu8(abs_p0q0, abs_p1q1), blimit); + mask = _mm_xor_si128(_mm_cmpeq_epi8(mask, zero), ff); + // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 > blimit) * -1; + mask = _mm_max_epu8(abs_p1p0, mask); + // mask |= (abs(p1 - p0) > limit) * -1; + // mask |= (abs(q1 - q0) > limit) * -1; + + work = _mm_max_epu8( + _mm_or_si128(_mm_subs_epu8(q2p2, q1p1), _mm_subs_epu8(q1p1, q2p2)), + _mm_or_si128(_mm_subs_epu8(q3p3, q2p2), _mm_subs_epu8(q2p2, q3p3))); + mask = _mm_max_epu8(work, mask); + mask = _mm_max_epu8(mask, _mm_srli_si128(mask, 8)); + mask = _mm_subs_epu8(mask, limit); + mask = _mm_cmpeq_epi8(mask, zero); + } + + // lp filter + { + const __m128i t4 = _mm_set1_epi8(4); + const __m128i t3 = _mm_set1_epi8(3); + const __m128i t80 = _mm_set1_epi8(0x80); + const __m128i t1 = _mm_set1_epi16(0x1); + __m128i qs1ps1 = _mm_xor_si128(q1p1, t80); + __m128i qs0ps0 = _mm_xor_si128(q0p0, t80); + __m128i qs0 = _mm_xor_si128(p0q0, t80); + __m128i qs1 = _mm_xor_si128(p1q1, t80); + __m128i filt; + __m128i work_a; + __m128i filter1, filter2; + __m128i flat2_q6p6, flat2_q5p5, flat2_q4p4, flat2_q3p3, flat2_q2p2; + __m128i flat2_q1p1, flat2_q0p0, flat_q2p2, flat_q1p1, flat_q0p0; + + filt = _mm_and_si128(_mm_subs_epi8(qs1ps1, qs1), hev); + work_a = _mm_subs_epi8(qs0, qs0ps0); + filt = _mm_adds_epi8(filt, work_a); + filt = _mm_adds_epi8(filt, work_a); + filt = _mm_adds_epi8(filt, work_a); + /* (aom_filter + 3 * (qs0 - ps0)) & mask */ + filt = _mm_and_si128(filt, mask); + + filter1 = _mm_adds_epi8(filt, t4); + filter2 = _mm_adds_epi8(filt, t3); + + filter1 = _mm_unpacklo_epi8(zero, filter1); + filter1 = _mm_srai_epi16(filter1, 0xB); + filter2 = _mm_unpacklo_epi8(zero, filter2); + filter2 = _mm_srai_epi16(filter2, 0xB); + + /* Filter1 >> 3 */ + filt = _mm_packs_epi16(filter2, _mm_subs_epi16(zero, filter1)); + qs0ps0 = _mm_xor_si128(_mm_adds_epi8(qs0ps0, filt), t80); + + /* filt >> 1 */ + filt = _mm_adds_epi16(filter1, t1); + filt = _mm_srai_epi16(filt, 1); + filt = _mm_andnot_si128(_mm_srai_epi16(_mm_unpacklo_epi8(zero, hev), 0x8), + filt); + filt = _mm_packs_epi16(filt, _mm_subs_epi16(zero, filt)); + qs1ps1 = _mm_xor_si128(_mm_adds_epi8(qs1ps1, filt), t80); + // loopfilter done + + { + __m128i work; + flat = _mm_max_epu8( + _mm_or_si128(_mm_subs_epu8(q2p2, q0p0), _mm_subs_epu8(q0p0, q2p2)), + _mm_or_si128(_mm_subs_epu8(q3p3, q0p0), _mm_subs_epu8(q0p0, q3p3))); + flat = _mm_max_epu8(abs_p1p0, flat); + flat = _mm_max_epu8(flat, _mm_srli_si128(flat, 8)); + flat = _mm_subs_epu8(flat, one); + flat = _mm_cmpeq_epi8(flat, zero); + flat = _mm_and_si128(flat, mask); + + q5p5 = _mm_loadl_epi64((__m128i *)(s - 6 * p)); + q5p5 = _mm_castps_si128( + _mm_loadh_pi(_mm_castsi128_ps(q5p5), (__m64 *)(s + 5 * p))); + + q6p6 = _mm_loadl_epi64((__m128i *)(s - 7 * p)); + q6p6 = _mm_castps_si128( + _mm_loadh_pi(_mm_castsi128_ps(q6p6), (__m64 *)(s + 6 * p))); + + flat2 = _mm_max_epu8( + _mm_or_si128(_mm_subs_epu8(q4p4, q0p0), _mm_subs_epu8(q0p0, q4p4)), + _mm_or_si128(_mm_subs_epu8(q5p5, q0p0), _mm_subs_epu8(q0p0, q5p5))); + + q7p7 = _mm_loadl_epi64((__m128i *)(s - 8 * p)); + q7p7 = _mm_castps_si128( + _mm_loadh_pi(_mm_castsi128_ps(q7p7), (__m64 *)(s + 7 * p))); + + work = _mm_max_epu8( + _mm_or_si128(_mm_subs_epu8(q6p6, q0p0), _mm_subs_epu8(q0p0, q6p6)), + _mm_or_si128(_mm_subs_epu8(q7p7, q0p0), _mm_subs_epu8(q0p0, q7p7))); + + flat2 = _mm_max_epu8(work, flat2); + flat2 = _mm_max_epu8(flat2, _mm_srli_si128(flat2, 8)); + flat2 = _mm_subs_epu8(flat2, one); + flat2 = _mm_cmpeq_epi8(flat2, zero); + flat2 = _mm_and_si128(flat2, flat); // flat2 & flat & mask + } + + // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ + // flat and wide flat calculations + { + const __m128i eight = _mm_set1_epi16(8); + const __m128i four = _mm_set1_epi16(4); + __m128i p7_16, p6_16, p5_16, p4_16, p3_16, p2_16, p1_16, p0_16; + __m128i q7_16, q6_16, q5_16, q4_16, q3_16, q2_16, q1_16, q0_16; + __m128i pixelFilter_p, pixelFilter_q; + __m128i pixetFilter_p2p1p0, pixetFilter_q2q1q0; + __m128i sum_p7, sum_q7, sum_p3, sum_q3, res_p, res_q; + + p7_16 = _mm_unpacklo_epi8(q7p7, zero); + p6_16 = _mm_unpacklo_epi8(q6p6, zero); + p5_16 = _mm_unpacklo_epi8(q5p5, zero); + p4_16 = _mm_unpacklo_epi8(q4p4, zero); + p3_16 = _mm_unpacklo_epi8(q3p3, zero); + p2_16 = _mm_unpacklo_epi8(q2p2, zero); + p1_16 = _mm_unpacklo_epi8(q1p1, zero); + p0_16 = _mm_unpacklo_epi8(q0p0, zero); + q0_16 = _mm_unpackhi_epi8(q0p0, zero); + q1_16 = _mm_unpackhi_epi8(q1p1, zero); + q2_16 = _mm_unpackhi_epi8(q2p2, zero); + q3_16 = _mm_unpackhi_epi8(q3p3, zero); + q4_16 = _mm_unpackhi_epi8(q4p4, zero); + q5_16 = _mm_unpackhi_epi8(q5p5, zero); + q6_16 = _mm_unpackhi_epi8(q6p6, zero); + q7_16 = _mm_unpackhi_epi8(q7p7, zero); + + pixelFilter_p = _mm_add_epi16(_mm_add_epi16(p6_16, p5_16), + _mm_add_epi16(p4_16, p3_16)); + pixelFilter_q = _mm_add_epi16(_mm_add_epi16(q6_16, q5_16), + _mm_add_epi16(q4_16, q3_16)); + + pixetFilter_p2p1p0 = _mm_add_epi16(p0_16, _mm_add_epi16(p2_16, p1_16)); + pixelFilter_p = _mm_add_epi16(pixelFilter_p, pixetFilter_p2p1p0); + + pixetFilter_q2q1q0 = _mm_add_epi16(q0_16, _mm_add_epi16(q2_16, q1_16)); + pixelFilter_q = _mm_add_epi16(pixelFilter_q, pixetFilter_q2q1q0); + pixelFilter_p = + _mm_add_epi16(eight, _mm_add_epi16(pixelFilter_p, pixelFilter_q)); + pixetFilter_p2p1p0 = _mm_add_epi16( + four, _mm_add_epi16(pixetFilter_p2p1p0, pixetFilter_q2q1q0)); + res_p = _mm_srli_epi16( + _mm_add_epi16(pixelFilter_p, _mm_add_epi16(p7_16, p0_16)), 4); + res_q = _mm_srli_epi16( + _mm_add_epi16(pixelFilter_p, _mm_add_epi16(q7_16, q0_16)), 4); + flat2_q0p0 = _mm_packus_epi16(res_p, res_q); + res_p = _mm_srli_epi16( + _mm_add_epi16(pixetFilter_p2p1p0, _mm_add_epi16(p3_16, p0_16)), 3); + res_q = _mm_srli_epi16( + _mm_add_epi16(pixetFilter_p2p1p0, _mm_add_epi16(q3_16, q0_16)), 3); + + flat_q0p0 = _mm_packus_epi16(res_p, res_q); + + sum_p7 = _mm_add_epi16(p7_16, p7_16); + sum_q7 = _mm_add_epi16(q7_16, q7_16); + sum_p3 = _mm_add_epi16(p3_16, p3_16); + sum_q3 = _mm_add_epi16(q3_16, q3_16); + + pixelFilter_q = _mm_sub_epi16(pixelFilter_p, p6_16); + pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q6_16); + res_p = _mm_srli_epi16( + _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p1_16)), 4); + res_q = _mm_srli_epi16( + _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q1_16)), 4); + flat2_q1p1 = _mm_packus_epi16(res_p, res_q); + + pixetFilter_q2q1q0 = _mm_sub_epi16(pixetFilter_p2p1p0, p2_16); + pixetFilter_p2p1p0 = _mm_sub_epi16(pixetFilter_p2p1p0, q2_16); + res_p = _mm_srli_epi16( + _mm_add_epi16(pixetFilter_p2p1p0, _mm_add_epi16(sum_p3, p1_16)), 3); + res_q = _mm_srli_epi16( + _mm_add_epi16(pixetFilter_q2q1q0, _mm_add_epi16(sum_q3, q1_16)), 3); + flat_q1p1 = _mm_packus_epi16(res_p, res_q); + + sum_p7 = _mm_add_epi16(sum_p7, p7_16); + sum_q7 = _mm_add_epi16(sum_q7, q7_16); + sum_p3 = _mm_add_epi16(sum_p3, p3_16); + sum_q3 = _mm_add_epi16(sum_q3, q3_16); + + pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q5_16); + pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p5_16); + res_p = _mm_srli_epi16( + _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p2_16)), 4); + res_q = _mm_srli_epi16( + _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q2_16)), 4); + flat2_q2p2 = _mm_packus_epi16(res_p, res_q); + + pixetFilter_p2p1p0 = _mm_sub_epi16(pixetFilter_p2p1p0, q1_16); + pixetFilter_q2q1q0 = _mm_sub_epi16(pixetFilter_q2q1q0, p1_16); + + res_p = _mm_srli_epi16( + _mm_add_epi16(pixetFilter_p2p1p0, _mm_add_epi16(sum_p3, p2_16)), 3); + res_q = _mm_srli_epi16( + _mm_add_epi16(pixetFilter_q2q1q0, _mm_add_epi16(sum_q3, q2_16)), 3); + flat_q2p2 = _mm_packus_epi16(res_p, res_q); + + sum_p7 = _mm_add_epi16(sum_p7, p7_16); + sum_q7 = _mm_add_epi16(sum_q7, q7_16); + pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q4_16); + pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p4_16); + res_p = _mm_srli_epi16( + _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p3_16)), 4); + res_q = _mm_srli_epi16( + _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q3_16)), 4); + flat2_q3p3 = _mm_packus_epi16(res_p, res_q); + + sum_p7 = _mm_add_epi16(sum_p7, p7_16); + sum_q7 = _mm_add_epi16(sum_q7, q7_16); + pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q3_16); + pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p3_16); + res_p = _mm_srli_epi16( + _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p4_16)), 4); + res_q = _mm_srli_epi16( + _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q4_16)), 4); + flat2_q4p4 = _mm_packus_epi16(res_p, res_q); + + sum_p7 = _mm_add_epi16(sum_p7, p7_16); + sum_q7 = _mm_add_epi16(sum_q7, q7_16); + pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q2_16); + pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p2_16); + res_p = _mm_srli_epi16( + _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p5_16)), 4); + res_q = _mm_srli_epi16( + _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q5_16)), 4); + flat2_q5p5 = _mm_packus_epi16(res_p, res_q); + + sum_p7 = _mm_add_epi16(sum_p7, p7_16); + sum_q7 = _mm_add_epi16(sum_q7, q7_16); + pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q1_16); + pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p1_16); + res_p = _mm_srli_epi16( + _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p6_16)), 4); + res_q = _mm_srli_epi16( + _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q6_16)), 4); + flat2_q6p6 = _mm_packus_epi16(res_p, res_q); + } + // wide flat + // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ + + flat = _mm_shuffle_epi32(flat, 68); + flat2 = _mm_shuffle_epi32(flat2, 68); + + q2p2 = _mm_andnot_si128(flat, q2p2); + flat_q2p2 = _mm_and_si128(flat, flat_q2p2); + q2p2 = _mm_or_si128(q2p2, flat_q2p2); + + qs1ps1 = _mm_andnot_si128(flat, qs1ps1); + flat_q1p1 = _mm_and_si128(flat, flat_q1p1); + q1p1 = _mm_or_si128(qs1ps1, flat_q1p1); + + qs0ps0 = _mm_andnot_si128(flat, qs0ps0); + flat_q0p0 = _mm_and_si128(flat, flat_q0p0); + q0p0 = _mm_or_si128(qs0ps0, flat_q0p0); + + q6p6 = _mm_andnot_si128(flat2, q6p6); + flat2_q6p6 = _mm_and_si128(flat2, flat2_q6p6); + q6p6 = _mm_or_si128(q6p6, flat2_q6p6); + _mm_storel_epi64((__m128i *)(s - 7 * p), q6p6); + _mm_storeh_pi((__m64 *)(s + 6 * p), _mm_castsi128_ps(q6p6)); + + q5p5 = _mm_andnot_si128(flat2, q5p5); + flat2_q5p5 = _mm_and_si128(flat2, flat2_q5p5); + q5p5 = _mm_or_si128(q5p5, flat2_q5p5); + _mm_storel_epi64((__m128i *)(s - 6 * p), q5p5); + _mm_storeh_pi((__m64 *)(s + 5 * p), _mm_castsi128_ps(q5p5)); + + q4p4 = _mm_andnot_si128(flat2, q4p4); + flat2_q4p4 = _mm_and_si128(flat2, flat2_q4p4); + q4p4 = _mm_or_si128(q4p4, flat2_q4p4); + _mm_storel_epi64((__m128i *)(s - 5 * p), q4p4); + _mm_storeh_pi((__m64 *)(s + 4 * p), _mm_castsi128_ps(q4p4)); + + q3p3 = _mm_andnot_si128(flat2, q3p3); + flat2_q3p3 = _mm_and_si128(flat2, flat2_q3p3); + q3p3 = _mm_or_si128(q3p3, flat2_q3p3); + _mm_storel_epi64((__m128i *)(s - 4 * p), q3p3); + _mm_storeh_pi((__m64 *)(s + 3 * p), _mm_castsi128_ps(q3p3)); + + q2p2 = _mm_andnot_si128(flat2, q2p2); + flat2_q2p2 = _mm_and_si128(flat2, flat2_q2p2); + q2p2 = _mm_or_si128(q2p2, flat2_q2p2); + _mm_storel_epi64((__m128i *)(s - 3 * p), q2p2); + _mm_storeh_pi((__m64 *)(s + 2 * p), _mm_castsi128_ps(q2p2)); + + q1p1 = _mm_andnot_si128(flat2, q1p1); + flat2_q1p1 = _mm_and_si128(flat2, flat2_q1p1); + q1p1 = _mm_or_si128(q1p1, flat2_q1p1); + _mm_storel_epi64((__m128i *)(s - 2 * p), q1p1); + _mm_storeh_pi((__m64 *)(s + 1 * p), _mm_castsi128_ps(q1p1)); + + q0p0 = _mm_andnot_si128(flat2, q0p0); + flat2_q0p0 = _mm_and_si128(flat2, flat2_q0p0); + q0p0 = _mm_or_si128(q0p0, flat2_q0p0); + _mm_storel_epi64((__m128i *)(s - 1 * p), q0p0); + _mm_storeh_pi((__m64 *)(s - 0 * p), _mm_castsi128_ps(q0p0)); + } +} + +DECLARE_ALIGNED(32, static const uint8_t, filt_loopfilter_avx2[32]) = { + 0, 128, 1, 128, 2, 128, 3, 128, 4, 128, 5, 128, 6, 128, 7, 128, + 8, 128, 9, 128, 10, 128, 11, 128, 12, 128, 13, 128, 14, 128, 15, 128 +}; + +void aom_lpf_horizontal_edge_16_avx2(unsigned char *s, int p, + const unsigned char *_blimit, + const unsigned char *_limit, + const unsigned char *_thresh) { + __m128i mask, hev, flat, flat2; + const __m128i zero = _mm_set1_epi16(0); + const __m128i one = _mm_set1_epi8(1); + __m128i p7, p6, p5; + __m128i p4, p3, p2, p1, p0, q0, q1, q2, q3, q4; + __m128i q5, q6, q7; + __m256i p256_7, q256_7, p256_6, q256_6, p256_5, q256_5, p256_4, q256_4, + p256_3, q256_3, p256_2, q256_2, p256_1, q256_1, p256_0, q256_0; + + const __m128i thresh = + _mm_broadcastb_epi8(_mm_cvtsi32_si128((int)_thresh[0])); + const __m128i limit = _mm_broadcastb_epi8(_mm_cvtsi32_si128((int)_limit[0])); + const __m128i blimit = + _mm_broadcastb_epi8(_mm_cvtsi32_si128((int)_blimit[0])); + + p256_4 = + _mm256_castpd_si256(_mm256_broadcast_pd((__m128d const *)(s - 5 * p))); + p256_3 = + _mm256_castpd_si256(_mm256_broadcast_pd((__m128d const *)(s - 4 * p))); + p256_2 = + _mm256_castpd_si256(_mm256_broadcast_pd((__m128d const *)(s - 3 * p))); + p256_1 = + _mm256_castpd_si256(_mm256_broadcast_pd((__m128d const *)(s - 2 * p))); + p256_0 = + _mm256_castpd_si256(_mm256_broadcast_pd((__m128d const *)(s - 1 * p))); + q256_0 = + _mm256_castpd_si256(_mm256_broadcast_pd((__m128d const *)(s - 0 * p))); + q256_1 = + _mm256_castpd_si256(_mm256_broadcast_pd((__m128d const *)(s + 1 * p))); + q256_2 = + _mm256_castpd_si256(_mm256_broadcast_pd((__m128d const *)(s + 2 * p))); + q256_3 = + _mm256_castpd_si256(_mm256_broadcast_pd((__m128d const *)(s + 3 * p))); + q256_4 = + _mm256_castpd_si256(_mm256_broadcast_pd((__m128d const *)(s + 4 * p))); + + p4 = _mm256_castsi256_si128(p256_4); + p3 = _mm256_castsi256_si128(p256_3); + p2 = _mm256_castsi256_si128(p256_2); + p1 = _mm256_castsi256_si128(p256_1); + p0 = _mm256_castsi256_si128(p256_0); + q0 = _mm256_castsi256_si128(q256_0); + q1 = _mm256_castsi256_si128(q256_1); + q2 = _mm256_castsi256_si128(q256_2); + q3 = _mm256_castsi256_si128(q256_3); + q4 = _mm256_castsi256_si128(q256_4); + + { + const __m128i abs_p1p0 = + _mm_or_si128(_mm_subs_epu8(p1, p0), _mm_subs_epu8(p0, p1)); + const __m128i abs_q1q0 = + _mm_or_si128(_mm_subs_epu8(q1, q0), _mm_subs_epu8(q0, q1)); + const __m128i fe = _mm_set1_epi8(0xfe); + const __m128i ff = _mm_cmpeq_epi8(abs_p1p0, abs_p1p0); + __m128i abs_p0q0 = + _mm_or_si128(_mm_subs_epu8(p0, q0), _mm_subs_epu8(q0, p0)); + __m128i abs_p1q1 = + _mm_or_si128(_mm_subs_epu8(p1, q1), _mm_subs_epu8(q1, p1)); + __m128i work; + flat = _mm_max_epu8(abs_p1p0, abs_q1q0); + hev = _mm_subs_epu8(flat, thresh); + hev = _mm_xor_si128(_mm_cmpeq_epi8(hev, zero), ff); + + abs_p0q0 = _mm_adds_epu8(abs_p0q0, abs_p0q0); + abs_p1q1 = _mm_srli_epi16(_mm_and_si128(abs_p1q1, fe), 1); + mask = _mm_subs_epu8(_mm_adds_epu8(abs_p0q0, abs_p1q1), blimit); + mask = _mm_xor_si128(_mm_cmpeq_epi8(mask, zero), ff); + // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 > blimit) * -1; + mask = _mm_max_epu8(flat, mask); + // mask |= (abs(p1 - p0) > limit) * -1; + // mask |= (abs(q1 - q0) > limit) * -1; + work = _mm_max_epu8( + _mm_or_si128(_mm_subs_epu8(p2, p1), _mm_subs_epu8(p1, p2)), + _mm_or_si128(_mm_subs_epu8(p3, p2), _mm_subs_epu8(p2, p3))); + mask = _mm_max_epu8(work, mask); + work = _mm_max_epu8( + _mm_or_si128(_mm_subs_epu8(q2, q1), _mm_subs_epu8(q1, q2)), + _mm_or_si128(_mm_subs_epu8(q3, q2), _mm_subs_epu8(q2, q3))); + mask = _mm_max_epu8(work, mask); + mask = _mm_subs_epu8(mask, limit); + mask = _mm_cmpeq_epi8(mask, zero); + } + + // lp filter + { + const __m128i t4 = _mm_set1_epi8(4); + const __m128i t3 = _mm_set1_epi8(3); + const __m128i t80 = _mm_set1_epi8(0x80); + const __m128i te0 = _mm_set1_epi8(0xe0); + const __m128i t1f = _mm_set1_epi8(0x1f); + const __m128i t1 = _mm_set1_epi8(0x1); + const __m128i t7f = _mm_set1_epi8(0x7f); + + __m128i ps1 = _mm_xor_si128(p1, t80); + __m128i ps0 = _mm_xor_si128(p0, t80); + __m128i qs0 = _mm_xor_si128(q0, t80); + __m128i qs1 = _mm_xor_si128(q1, t80); + __m128i filt; + __m128i work_a; + __m128i filter1, filter2; + __m128i flat2_p6, flat2_p5, flat2_p4, flat2_p3, flat2_p2, flat2_p1, + flat2_p0, flat2_q0, flat2_q1, flat2_q2, flat2_q3, flat2_q4, flat2_q5, + flat2_q6, flat_p2, flat_p1, flat_p0, flat_q0, flat_q1, flat_q2; + + filt = _mm_and_si128(_mm_subs_epi8(ps1, qs1), hev); + work_a = _mm_subs_epi8(qs0, ps0); + filt = _mm_adds_epi8(filt, work_a); + filt = _mm_adds_epi8(filt, work_a); + filt = _mm_adds_epi8(filt, work_a); + /* (aom_filter + 3 * (qs0 - ps0)) & mask */ + filt = _mm_and_si128(filt, mask); + + filter1 = _mm_adds_epi8(filt, t4); + filter2 = _mm_adds_epi8(filt, t3); + + /* Filter1 >> 3 */ + work_a = _mm_cmpgt_epi8(zero, filter1); + filter1 = _mm_srli_epi16(filter1, 3); + work_a = _mm_and_si128(work_a, te0); + filter1 = _mm_and_si128(filter1, t1f); + filter1 = _mm_or_si128(filter1, work_a); + qs0 = _mm_xor_si128(_mm_subs_epi8(qs0, filter1), t80); + + /* Filter2 >> 3 */ + work_a = _mm_cmpgt_epi8(zero, filter2); + filter2 = _mm_srli_epi16(filter2, 3); + work_a = _mm_and_si128(work_a, te0); + filter2 = _mm_and_si128(filter2, t1f); + filter2 = _mm_or_si128(filter2, work_a); + ps0 = _mm_xor_si128(_mm_adds_epi8(ps0, filter2), t80); + + /* filt >> 1 */ + filt = _mm_adds_epi8(filter1, t1); + work_a = _mm_cmpgt_epi8(zero, filt); + filt = _mm_srli_epi16(filt, 1); + work_a = _mm_and_si128(work_a, t80); + filt = _mm_and_si128(filt, t7f); + filt = _mm_or_si128(filt, work_a); + filt = _mm_andnot_si128(hev, filt); + ps1 = _mm_xor_si128(_mm_adds_epi8(ps1, filt), t80); + qs1 = _mm_xor_si128(_mm_subs_epi8(qs1, filt), t80); + // loopfilter done + + { + __m128i work; + work = _mm_max_epu8( + _mm_or_si128(_mm_subs_epu8(p2, p0), _mm_subs_epu8(p0, p2)), + _mm_or_si128(_mm_subs_epu8(q2, q0), _mm_subs_epu8(q0, q2))); + flat = _mm_max_epu8(work, flat); + work = _mm_max_epu8( + _mm_or_si128(_mm_subs_epu8(p3, p0), _mm_subs_epu8(p0, p3)), + _mm_or_si128(_mm_subs_epu8(q3, q0), _mm_subs_epu8(q0, q3))); + flat = _mm_max_epu8(work, flat); + work = _mm_max_epu8( + _mm_or_si128(_mm_subs_epu8(p4, p0), _mm_subs_epu8(p0, p4)), + _mm_or_si128(_mm_subs_epu8(q4, q0), _mm_subs_epu8(q0, q4))); + flat = _mm_subs_epu8(flat, one); + flat = _mm_cmpeq_epi8(flat, zero); + flat = _mm_and_si128(flat, mask); + + p256_5 = _mm256_castpd_si256( + _mm256_broadcast_pd((__m128d const *)(s - 6 * p))); + q256_5 = _mm256_castpd_si256( + _mm256_broadcast_pd((__m128d const *)(s + 5 * p))); + p5 = _mm256_castsi256_si128(p256_5); + q5 = _mm256_castsi256_si128(q256_5); + flat2 = _mm_max_epu8( + _mm_or_si128(_mm_subs_epu8(p5, p0), _mm_subs_epu8(p0, p5)), + _mm_or_si128(_mm_subs_epu8(q5, q0), _mm_subs_epu8(q0, q5))); + + flat2 = _mm_max_epu8(work, flat2); + p256_6 = _mm256_castpd_si256( + _mm256_broadcast_pd((__m128d const *)(s - 7 * p))); + q256_6 = _mm256_castpd_si256( + _mm256_broadcast_pd((__m128d const *)(s + 6 * p))); + p6 = _mm256_castsi256_si128(p256_6); + q6 = _mm256_castsi256_si128(q256_6); + work = _mm_max_epu8( + _mm_or_si128(_mm_subs_epu8(p6, p0), _mm_subs_epu8(p0, p6)), + _mm_or_si128(_mm_subs_epu8(q6, q0), _mm_subs_epu8(q0, q6))); + + flat2 = _mm_max_epu8(work, flat2); + + p256_7 = _mm256_castpd_si256( + _mm256_broadcast_pd((__m128d const *)(s - 8 * p))); + q256_7 = _mm256_castpd_si256( + _mm256_broadcast_pd((__m128d const *)(s + 7 * p))); + p7 = _mm256_castsi256_si128(p256_7); + q7 = _mm256_castsi256_si128(q256_7); + work = _mm_max_epu8( + _mm_or_si128(_mm_subs_epu8(p7, p0), _mm_subs_epu8(p0, p7)), + _mm_or_si128(_mm_subs_epu8(q7, q0), _mm_subs_epu8(q0, q7))); + + flat2 = _mm_max_epu8(work, flat2); + flat2 = _mm_subs_epu8(flat2, one); + flat2 = _mm_cmpeq_epi8(flat2, zero); + flat2 = _mm_and_si128(flat2, flat); // flat2 & flat & mask + } + + // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ + // flat and wide flat calculations + { + const __m256i eight = _mm256_set1_epi16(8); + const __m256i four = _mm256_set1_epi16(4); + __m256i pixelFilter_p, pixelFilter_q, pixetFilter_p2p1p0, + pixetFilter_q2q1q0, sum_p7, sum_q7, sum_p3, sum_q3, res_p, res_q; + + const __m256i filter = + _mm256_load_si256((__m256i const *)filt_loopfilter_avx2); + p256_7 = _mm256_shuffle_epi8(p256_7, filter); + p256_6 = _mm256_shuffle_epi8(p256_6, filter); + p256_5 = _mm256_shuffle_epi8(p256_5, filter); + p256_4 = _mm256_shuffle_epi8(p256_4, filter); + p256_3 = _mm256_shuffle_epi8(p256_3, filter); + p256_2 = _mm256_shuffle_epi8(p256_2, filter); + p256_1 = _mm256_shuffle_epi8(p256_1, filter); + p256_0 = _mm256_shuffle_epi8(p256_0, filter); + q256_0 = _mm256_shuffle_epi8(q256_0, filter); + q256_1 = _mm256_shuffle_epi8(q256_1, filter); + q256_2 = _mm256_shuffle_epi8(q256_2, filter); + q256_3 = _mm256_shuffle_epi8(q256_3, filter); + q256_4 = _mm256_shuffle_epi8(q256_4, filter); + q256_5 = _mm256_shuffle_epi8(q256_5, filter); + q256_6 = _mm256_shuffle_epi8(q256_6, filter); + q256_7 = _mm256_shuffle_epi8(q256_7, filter); + + pixelFilter_p = _mm256_add_epi16(_mm256_add_epi16(p256_6, p256_5), + _mm256_add_epi16(p256_4, p256_3)); + pixelFilter_q = _mm256_add_epi16(_mm256_add_epi16(q256_6, q256_5), + _mm256_add_epi16(q256_4, q256_3)); + + pixetFilter_p2p1p0 = + _mm256_add_epi16(p256_0, _mm256_add_epi16(p256_2, p256_1)); + pixelFilter_p = _mm256_add_epi16(pixelFilter_p, pixetFilter_p2p1p0); + + pixetFilter_q2q1q0 = + _mm256_add_epi16(q256_0, _mm256_add_epi16(q256_2, q256_1)); + pixelFilter_q = _mm256_add_epi16(pixelFilter_q, pixetFilter_q2q1q0); + + pixelFilter_p = _mm256_add_epi16( + eight, _mm256_add_epi16(pixelFilter_p, pixelFilter_q)); + + pixetFilter_p2p1p0 = _mm256_add_epi16( + four, _mm256_add_epi16(pixetFilter_p2p1p0, pixetFilter_q2q1q0)); + + res_p = _mm256_srli_epi16( + _mm256_add_epi16(pixelFilter_p, _mm256_add_epi16(p256_7, p256_0)), 4); + + flat2_p0 = _mm256_castsi256_si128( + _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p), 168)); + + res_q = _mm256_srli_epi16( + _mm256_add_epi16(pixelFilter_p, _mm256_add_epi16(q256_7, q256_0)), 4); + + flat2_q0 = _mm256_castsi256_si128( + _mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q), 168)); + + res_p = + _mm256_srli_epi16(_mm256_add_epi16(pixetFilter_p2p1p0, + _mm256_add_epi16(p256_3, p256_0)), + 3); + + flat_p0 = _mm256_castsi256_si128( + _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p), 168)); + + res_q = + _mm256_srli_epi16(_mm256_add_epi16(pixetFilter_p2p1p0, + _mm256_add_epi16(q256_3, q256_0)), + 3); + + flat_q0 = _mm256_castsi256_si128( + _mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q), 168)); + + sum_p7 = _mm256_add_epi16(p256_7, p256_7); + + sum_q7 = _mm256_add_epi16(q256_7, q256_7); + + sum_p3 = _mm256_add_epi16(p256_3, p256_3); + + sum_q3 = _mm256_add_epi16(q256_3, q256_3); + + pixelFilter_q = _mm256_sub_epi16(pixelFilter_p, p256_6); + + pixelFilter_p = _mm256_sub_epi16(pixelFilter_p, q256_6); + + res_p = _mm256_srli_epi16( + _mm256_add_epi16(pixelFilter_p, _mm256_add_epi16(sum_p7, p256_1)), 4); + + flat2_p1 = _mm256_castsi256_si128( + _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p), 168)); + + res_q = _mm256_srli_epi16( + _mm256_add_epi16(pixelFilter_q, _mm256_add_epi16(sum_q7, q256_1)), 4); + + flat2_q1 = _mm256_castsi256_si128( + _mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q), 168)); + + pixetFilter_q2q1q0 = _mm256_sub_epi16(pixetFilter_p2p1p0, p256_2); + + pixetFilter_p2p1p0 = _mm256_sub_epi16(pixetFilter_p2p1p0, q256_2); + + res_p = + _mm256_srli_epi16(_mm256_add_epi16(pixetFilter_p2p1p0, + _mm256_add_epi16(sum_p3, p256_1)), + 3); + + flat_p1 = _mm256_castsi256_si128( + _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p), 168)); + + res_q = + _mm256_srli_epi16(_mm256_add_epi16(pixetFilter_q2q1q0, + _mm256_add_epi16(sum_q3, q256_1)), + 3); + + flat_q1 = _mm256_castsi256_si128( + _mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q), 168)); + + sum_p7 = _mm256_add_epi16(sum_p7, p256_7); + + sum_q7 = _mm256_add_epi16(sum_q7, q256_7); + + sum_p3 = _mm256_add_epi16(sum_p3, p256_3); + + sum_q3 = _mm256_add_epi16(sum_q3, q256_3); + + pixelFilter_p = _mm256_sub_epi16(pixelFilter_p, q256_5); + + pixelFilter_q = _mm256_sub_epi16(pixelFilter_q, p256_5); + + res_p = _mm256_srli_epi16( + _mm256_add_epi16(pixelFilter_p, _mm256_add_epi16(sum_p7, p256_2)), 4); + + flat2_p2 = _mm256_castsi256_si128( + _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p), 168)); + + res_q = _mm256_srli_epi16( + _mm256_add_epi16(pixelFilter_q, _mm256_add_epi16(sum_q7, q256_2)), 4); + + flat2_q2 = _mm256_castsi256_si128( + _mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q), 168)); + + pixetFilter_p2p1p0 = _mm256_sub_epi16(pixetFilter_p2p1p0, q256_1); + + pixetFilter_q2q1q0 = _mm256_sub_epi16(pixetFilter_q2q1q0, p256_1); + + res_p = + _mm256_srli_epi16(_mm256_add_epi16(pixetFilter_p2p1p0, + _mm256_add_epi16(sum_p3, p256_2)), + 3); + + flat_p2 = _mm256_castsi256_si128( + _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p), 168)); + + res_q = + _mm256_srli_epi16(_mm256_add_epi16(pixetFilter_q2q1q0, + _mm256_add_epi16(sum_q3, q256_2)), + 3); + + flat_q2 = _mm256_castsi256_si128( + _mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q), 168)); + + sum_p7 = _mm256_add_epi16(sum_p7, p256_7); + + sum_q7 = _mm256_add_epi16(sum_q7, q256_7); + + pixelFilter_p = _mm256_sub_epi16(pixelFilter_p, q256_4); + + pixelFilter_q = _mm256_sub_epi16(pixelFilter_q, p256_4); + + res_p = _mm256_srli_epi16( + _mm256_add_epi16(pixelFilter_p, _mm256_add_epi16(sum_p7, p256_3)), 4); + + flat2_p3 = _mm256_castsi256_si128( + _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p), 168)); + + res_q = _mm256_srli_epi16( + _mm256_add_epi16(pixelFilter_q, _mm256_add_epi16(sum_q7, q256_3)), 4); + + flat2_q3 = _mm256_castsi256_si128( + _mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q), 168)); + + sum_p7 = _mm256_add_epi16(sum_p7, p256_7); + + sum_q7 = _mm256_add_epi16(sum_q7, q256_7); + + pixelFilter_p = _mm256_sub_epi16(pixelFilter_p, q256_3); + + pixelFilter_q = _mm256_sub_epi16(pixelFilter_q, p256_3); + + res_p = _mm256_srli_epi16( + _mm256_add_epi16(pixelFilter_p, _mm256_add_epi16(sum_p7, p256_4)), 4); + + flat2_p4 = _mm256_castsi256_si128( + _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p), 168)); + + res_q = _mm256_srli_epi16( + _mm256_add_epi16(pixelFilter_q, _mm256_add_epi16(sum_q7, q256_4)), 4); + + flat2_q4 = _mm256_castsi256_si128( + _mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q), 168)); + + sum_p7 = _mm256_add_epi16(sum_p7, p256_7); + + sum_q7 = _mm256_add_epi16(sum_q7, q256_7); + + pixelFilter_p = _mm256_sub_epi16(pixelFilter_p, q256_2); + + pixelFilter_q = _mm256_sub_epi16(pixelFilter_q, p256_2); + + res_p = _mm256_srli_epi16( + _mm256_add_epi16(pixelFilter_p, _mm256_add_epi16(sum_p7, p256_5)), 4); + + flat2_p5 = _mm256_castsi256_si128( + _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p), 168)); + + res_q = _mm256_srli_epi16( + _mm256_add_epi16(pixelFilter_q, _mm256_add_epi16(sum_q7, q256_5)), 4); + + flat2_q5 = _mm256_castsi256_si128( + _mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q), 168)); + + sum_p7 = _mm256_add_epi16(sum_p7, p256_7); + + sum_q7 = _mm256_add_epi16(sum_q7, q256_7); + + pixelFilter_p = _mm256_sub_epi16(pixelFilter_p, q256_1); + + pixelFilter_q = _mm256_sub_epi16(pixelFilter_q, p256_1); + + res_p = _mm256_srli_epi16( + _mm256_add_epi16(pixelFilter_p, _mm256_add_epi16(sum_p7, p256_6)), 4); + + flat2_p6 = _mm256_castsi256_si128( + _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p), 168)); + + res_q = _mm256_srli_epi16( + _mm256_add_epi16(pixelFilter_q, _mm256_add_epi16(sum_q7, q256_6)), 4); + + flat2_q6 = _mm256_castsi256_si128( + _mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q), 168)); + } + + // wide flat + // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ + + p2 = _mm_andnot_si128(flat, p2); + flat_p2 = _mm_and_si128(flat, flat_p2); + p2 = _mm_or_si128(flat_p2, p2); + + p1 = _mm_andnot_si128(flat, ps1); + flat_p1 = _mm_and_si128(flat, flat_p1); + p1 = _mm_or_si128(flat_p1, p1); + + p0 = _mm_andnot_si128(flat, ps0); + flat_p0 = _mm_and_si128(flat, flat_p0); + p0 = _mm_or_si128(flat_p0, p0); + + q0 = _mm_andnot_si128(flat, qs0); + flat_q0 = _mm_and_si128(flat, flat_q0); + q0 = _mm_or_si128(flat_q0, q0); + + q1 = _mm_andnot_si128(flat, qs1); + flat_q1 = _mm_and_si128(flat, flat_q1); + q1 = _mm_or_si128(flat_q1, q1); + + q2 = _mm_andnot_si128(flat, q2); + flat_q2 = _mm_and_si128(flat, flat_q2); + q2 = _mm_or_si128(flat_q2, q2); + + p6 = _mm_andnot_si128(flat2, p6); + flat2_p6 = _mm_and_si128(flat2, flat2_p6); + p6 = _mm_or_si128(flat2_p6, p6); + _mm_storeu_si128((__m128i *)(s - 7 * p), p6); + + p5 = _mm_andnot_si128(flat2, p5); + flat2_p5 = _mm_and_si128(flat2, flat2_p5); + p5 = _mm_or_si128(flat2_p5, p5); + _mm_storeu_si128((__m128i *)(s - 6 * p), p5); + + p4 = _mm_andnot_si128(flat2, p4); + flat2_p4 = _mm_and_si128(flat2, flat2_p4); + p4 = _mm_or_si128(flat2_p4, p4); + _mm_storeu_si128((__m128i *)(s - 5 * p), p4); + + p3 = _mm_andnot_si128(flat2, p3); + flat2_p3 = _mm_and_si128(flat2, flat2_p3); + p3 = _mm_or_si128(flat2_p3, p3); + _mm_storeu_si128((__m128i *)(s - 4 * p), p3); + + p2 = _mm_andnot_si128(flat2, p2); + flat2_p2 = _mm_and_si128(flat2, flat2_p2); + p2 = _mm_or_si128(flat2_p2, p2); + _mm_storeu_si128((__m128i *)(s - 3 * p), p2); + + p1 = _mm_andnot_si128(flat2, p1); + flat2_p1 = _mm_and_si128(flat2, flat2_p1); + p1 = _mm_or_si128(flat2_p1, p1); + _mm_storeu_si128((__m128i *)(s - 2 * p), p1); + + p0 = _mm_andnot_si128(flat2, p0); + flat2_p0 = _mm_and_si128(flat2, flat2_p0); + p0 = _mm_or_si128(flat2_p0, p0); + _mm_storeu_si128((__m128i *)(s - 1 * p), p0); + + q0 = _mm_andnot_si128(flat2, q0); + flat2_q0 = _mm_and_si128(flat2, flat2_q0); + q0 = _mm_or_si128(flat2_q0, q0); + _mm_storeu_si128((__m128i *)(s - 0 * p), q0); + + q1 = _mm_andnot_si128(flat2, q1); + flat2_q1 = _mm_and_si128(flat2, flat2_q1); + q1 = _mm_or_si128(flat2_q1, q1); + _mm_storeu_si128((__m128i *)(s + 1 * p), q1); + + q2 = _mm_andnot_si128(flat2, q2); + flat2_q2 = _mm_and_si128(flat2, flat2_q2); + q2 = _mm_or_si128(flat2_q2, q2); + _mm_storeu_si128((__m128i *)(s + 2 * p), q2); + + q3 = _mm_andnot_si128(flat2, q3); + flat2_q3 = _mm_and_si128(flat2, flat2_q3); + q3 = _mm_or_si128(flat2_q3, q3); + _mm_storeu_si128((__m128i *)(s + 3 * p), q3); + + q4 = _mm_andnot_si128(flat2, q4); + flat2_q4 = _mm_and_si128(flat2, flat2_q4); + q4 = _mm_or_si128(flat2_q4, q4); + _mm_storeu_si128((__m128i *)(s + 4 * p), q4); + + q5 = _mm_andnot_si128(flat2, q5); + flat2_q5 = _mm_and_si128(flat2, flat2_q5); + q5 = _mm_or_si128(flat2_q5, q5); + _mm_storeu_si128((__m128i *)(s + 5 * p), q5); + + q6 = _mm_andnot_si128(flat2, q6); + flat2_q6 = _mm_and_si128(flat2, flat2_q6); + q6 = _mm_or_si128(flat2_q6, q6); + _mm_storeu_si128((__m128i *)(s + 6 * p), q6); + } + _mm256_zeroupper(); +} diff --git a/third_party/aom/aom_dsp/x86/loopfilter_sse2.c b/third_party/aom/aom_dsp/x86/loopfilter_sse2.c new file mode 100644 index 000000000..7e134dc63 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/loopfilter_sse2.c @@ -0,0 +1,1892 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include <emmintrin.h> // SSE2 + +#include "./aom_dsp_rtcd.h" +#include "aom_ports/mem.h" +#include "aom_ports/emmintrin_compat.h" + +static INLINE __m128i abs_diff(__m128i a, __m128i b) { + return _mm_or_si128(_mm_subs_epu8(a, b), _mm_subs_epu8(b, a)); +} + +#if CONFIG_PARALLEL_DEBLOCKING +// filter_mask and hev_mask +#define FILTER_HEV_MASK4 \ + do { \ + /* (abs(q1 - q0), abs(p1 - p0) */ \ + __m128i flat = abs_diff(q1p1, q0p0); \ + /* abs(p1 - q1), abs(p0 - q0) */ \ + const __m128i abs_p1q1p0q0 = abs_diff(p1p0, q1q0); \ + __m128i abs_p0q0, abs_p1q1; \ + \ + /* const uint8_t hev = hev_mask(thresh, *op1, *op0, *oq0, *oq1); */ \ + hev = \ + _mm_unpacklo_epi8(_mm_max_epu8(flat, _mm_srli_si128(flat, 8)), zero); \ + hev = _mm_cmpgt_epi16(hev, thresh); \ + hev = _mm_packs_epi16(hev, hev); \ + \ + /* const int8_t mask = filter_mask2(*limit, *blimit, */ \ + /* p1, p0, q0, q1); */ \ + abs_p0q0 = \ + _mm_adds_epu8(abs_p1q1p0q0, abs_p1q1p0q0); /* abs(p0 - q0) * 2 */ \ + abs_p1q1 = \ + _mm_unpackhi_epi8(abs_p1q1p0q0, abs_p1q1p0q0); /* abs(p1 - q1) */ \ + abs_p1q1 = _mm_srli_epi16(abs_p1q1, 9); \ + abs_p1q1 = _mm_packs_epi16(abs_p1q1, abs_p1q1); /* abs(p1 - q1) / 2 */ \ + /* abs(p0 - q0) * 2 + abs(p1 - q1) / 2 */ \ + mask = _mm_adds_epu8(abs_p0q0, abs_p1q1); \ + flat = _mm_max_epu8(flat, _mm_srli_si128(flat, 8)); \ + mask = _mm_unpacklo_epi64(mask, flat); \ + mask = _mm_subs_epu8(mask, limit); \ + mask = _mm_cmpeq_epi8(mask, zero); \ + mask = _mm_and_si128(mask, _mm_srli_si128(mask, 8)); \ + } while (0) +#endif // CONFIG_PARALLEL_DEBLOCKING + +// filter_mask and hev_mask +#define FILTER_HEV_MASK \ + do { \ + /* (abs(q1 - q0), abs(p1 - p0) */ \ + __m128i flat = abs_diff(q1p1, q0p0); \ + /* abs(p1 - q1), abs(p0 - q0) */ \ + const __m128i abs_p1q1p0q0 = abs_diff(p1p0, q1q0); \ + __m128i abs_p0q0, abs_p1q1, work; \ + \ + /* const uint8_t hev = hev_mask(thresh, *op1, *op0, *oq0, *oq1); */ \ + hev = \ + _mm_unpacklo_epi8(_mm_max_epu8(flat, _mm_srli_si128(flat, 8)), zero); \ + hev = _mm_cmpgt_epi16(hev, thresh); \ + hev = _mm_packs_epi16(hev, hev); \ + \ + /* const int8_t mask = filter_mask(*limit, *blimit, */ \ + /* p3, p2, p1, p0, q0, q1, q2, q3); */ \ + abs_p0q0 = \ + _mm_adds_epu8(abs_p1q1p0q0, abs_p1q1p0q0); /* abs(p0 - q0) * 2 */ \ + abs_p1q1 = \ + _mm_unpackhi_epi8(abs_p1q1p0q0, abs_p1q1p0q0); /* abs(p1 - q1) */ \ + abs_p1q1 = _mm_srli_epi16(abs_p1q1, 9); \ + abs_p1q1 = _mm_packs_epi16(abs_p1q1, abs_p1q1); /* abs(p1 - q1) / 2 */ \ + /* abs(p0 - q0) * 2 + abs(p1 - q1) / 2 */ \ + mask = _mm_adds_epu8(abs_p0q0, abs_p1q1); \ + /* abs(p3 - p2), abs(p2 - p1) */ \ + work = abs_diff(p3p2, p2p1); \ + flat = _mm_max_epu8(work, flat); \ + /* abs(q3 - q2), abs(q2 - q1) */ \ + work = abs_diff(q3q2, q2q1); \ + flat = _mm_max_epu8(work, flat); \ + flat = _mm_max_epu8(flat, _mm_srli_si128(flat, 8)); \ + mask = _mm_unpacklo_epi64(mask, flat); \ + mask = _mm_subs_epu8(mask, limit); \ + mask = _mm_cmpeq_epi8(mask, zero); \ + mask = _mm_and_si128(mask, _mm_srli_si128(mask, 8)); \ + } while (0) + +#define FILTER4 \ + do { \ + const __m128i t3t4 = \ + _mm_set_epi8(3, 3, 3, 3, 3, 3, 3, 3, 4, 4, 4, 4, 4, 4, 4, 4); \ + const __m128i t80 = _mm_set1_epi8(0x80); \ + __m128i filter, filter2filter1, work; \ + \ + ps1ps0 = _mm_xor_si128(p1p0, t80); /* ^ 0x80 */ \ + qs1qs0 = _mm_xor_si128(q1q0, t80); \ + \ + /* int8_t filter = signed_char_clamp(ps1 - qs1) & hev; */ \ + work = _mm_subs_epi8(ps1ps0, qs1qs0); \ + filter = _mm_and_si128(_mm_srli_si128(work, 8), hev); \ + /* filter = signed_char_clamp(filter + 3 * (qs0 - ps0)) & mask; */ \ + filter = _mm_subs_epi8(filter, work); \ + filter = _mm_subs_epi8(filter, work); \ + filter = _mm_subs_epi8(filter, work); /* + 3 * (qs0 - ps0) */ \ + filter = _mm_and_si128(filter, mask); /* & mask */ \ + filter = _mm_unpacklo_epi64(filter, filter); \ + \ + /* filter1 = signed_char_clamp(filter + 4) >> 3; */ \ + /* filter2 = signed_char_clamp(filter + 3) >> 3; */ \ + filter2filter1 = _mm_adds_epi8(filter, t3t4); /* signed_char_clamp */ \ + filter = _mm_unpackhi_epi8(filter2filter1, filter2filter1); \ + filter2filter1 = _mm_unpacklo_epi8(filter2filter1, filter2filter1); \ + filter2filter1 = _mm_srai_epi16(filter2filter1, 11); /* >> 3 */ \ + filter = _mm_srai_epi16(filter, 11); /* >> 3 */ \ + filter2filter1 = _mm_packs_epi16(filter2filter1, filter); \ + \ + /* filter = ROUND_POWER_OF_TWO(filter1, 1) & ~hev; */ \ + filter = _mm_subs_epi8(filter2filter1, ff); /* + 1 */ \ + filter = _mm_unpacklo_epi8(filter, filter); \ + filter = _mm_srai_epi16(filter, 9); /* round */ \ + filter = _mm_packs_epi16(filter, filter); \ + filter = _mm_andnot_si128(hev, filter); \ + \ + hev = _mm_unpackhi_epi64(filter2filter1, filter); \ + filter2filter1 = _mm_unpacklo_epi64(filter2filter1, filter); \ + \ + /* signed_char_clamp(qs1 - filter), signed_char_clamp(qs0 - filter1) */ \ + qs1qs0 = _mm_subs_epi8(qs1qs0, filter2filter1); \ + /* signed_char_clamp(ps1 + filter), signed_char_clamp(ps0 + filter2) */ \ + ps1ps0 = _mm_adds_epi8(ps1ps0, hev); \ + qs1qs0 = _mm_xor_si128(qs1qs0, t80); /* ^ 0x80 */ \ + ps1ps0 = _mm_xor_si128(ps1ps0, t80); /* ^ 0x80 */ \ + } while (0) + +void aom_lpf_horizontal_4_sse2(uint8_t *s, int p /* pitch */, + const uint8_t *_blimit, const uint8_t *_limit, + const uint8_t *_thresh) { + const __m128i zero = _mm_set1_epi16(0); + const __m128i limit = + _mm_unpacklo_epi64(_mm_loadl_epi64((const __m128i *)_blimit), + _mm_loadl_epi64((const __m128i *)_limit)); + const __m128i thresh = + _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)_thresh), zero); + const __m128i ff = _mm_cmpeq_epi8(zero, zero); +#if !CONFIG_PARALLEL_DEBLOCKING + __m128i p3p2, p2p1, q3q2, q2q1; +#endif // !CONFIG_PARALLEL_DEBLOCKING + __m128i q1p1, q0p0, p1p0, q1q0, ps1ps0, qs1qs0; + __m128i mask, hev; +#if !CONFIG_PARALLEL_DEBLOCKING + p3p2 = _mm_unpacklo_epi64(_mm_loadl_epi64((__m128i *)(s - 3 * p)), + _mm_loadl_epi64((__m128i *)(s - 4 * p))); +#endif // !CONFIG_PARALLEL_DEBLOCKING + q1p1 = _mm_unpacklo_epi64(_mm_loadl_epi64((__m128i *)(s - 2 * p)), + _mm_loadl_epi64((__m128i *)(s + 1 * p))); + q0p0 = _mm_unpacklo_epi64(_mm_loadl_epi64((__m128i *)(s - 1 * p)), + _mm_loadl_epi64((__m128i *)(s + 0 * p))); +#if !CONFIG_PARALLEL_DEBLOCKING + q3q2 = _mm_unpacklo_epi64(_mm_loadl_epi64((__m128i *)(s + 2 * p)), + _mm_loadl_epi64((__m128i *)(s + 3 * p))); +#endif // !CONFIG_PARALLEL_DEBLOCKING + p1p0 = _mm_unpacklo_epi64(q0p0, q1p1); + q1q0 = _mm_unpackhi_epi64(q0p0, q1p1); +#if !CONFIG_PARALLEL_DEBLOCKING + p2p1 = _mm_unpacklo_epi64(q1p1, p3p2); + q2q1 = _mm_unpacklo_epi64(_mm_srli_si128(q1p1, 8), q3q2); +#endif // !CONFIG_PARALLEL_DEBLOCKING +#if !CONFIG_PARALLEL_DEBLOCKING + FILTER_HEV_MASK; +#else // CONFIG_PARALLEL_DEBLOCKING + FILTER_HEV_MASK4; +#endif // !CONFIG_PARALLEL_DEBLOCKING + FILTER4; + + _mm_storeh_pi((__m64 *)(s - 2 * p), _mm_castsi128_ps(ps1ps0)); // *op1 + _mm_storel_epi64((__m128i *)(s - 1 * p), ps1ps0); // *op0 + _mm_storel_epi64((__m128i *)(s + 0 * p), qs1qs0); // *oq0 + _mm_storeh_pi((__m64 *)(s + 1 * p), _mm_castsi128_ps(qs1qs0)); // *oq1 +} + +void aom_lpf_vertical_4_sse2(uint8_t *s, int p /* pitch */, + const uint8_t *_blimit, const uint8_t *_limit, + const uint8_t *_thresh) { + const __m128i zero = _mm_set1_epi16(0); + const __m128i limit = + _mm_unpacklo_epi64(_mm_loadl_epi64((const __m128i *)_blimit), + _mm_loadl_epi64((const __m128i *)_limit)); + const __m128i thresh = + _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)_thresh), zero); + const __m128i ff = _mm_cmpeq_epi8(zero, zero); + __m128i x0, x1, x2, x3; +#if !CONFIG_PARALLEL_DEBLOCKING + __m128i p3p2, p2p1, q3q2, q2q1; +#endif // !CONFIG_PARALLEL_DEBLOCKING + __m128i q1p1, q0p0, p1p0, q1q0, ps1ps0, qs1qs0; + __m128i mask, hev; + + // 00 10 01 11 02 12 03 13 04 14 05 15 06 16 07 17 + q1q0 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(s + 0 * p - 4)), + _mm_loadl_epi64((__m128i *)(s + 1 * p - 4))); + + // 20 30 21 31 22 32 23 33 24 34 25 35 26 36 27 37 + x1 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(s + 2 * p - 4)), + _mm_loadl_epi64((__m128i *)(s + 3 * p - 4))); + + // 40 50 41 51 42 52 43 53 44 54 45 55 46 56 47 57 + x2 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(s + 4 * p - 4)), + _mm_loadl_epi64((__m128i *)(s + 5 * p - 4))); + + // 60 70 61 71 62 72 63 73 64 74 65 75 66 76 67 77 + x3 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(s + 6 * p - 4)), + _mm_loadl_epi64((__m128i *)(s + 7 * p - 4))); + + // Transpose 8x8 + // 00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33 + p1p0 = _mm_unpacklo_epi16(q1q0, x1); + // 40 50 60 70 41 51 61 71 42 52 62 72 43 53 63 73 + x0 = _mm_unpacklo_epi16(x2, x3); +#if !CONFIG_PARALLEL_DEBLOCKING + // 00 10 20 30 40 50 60 70 01 11 21 31 41 51 61 71 + p3p2 = _mm_unpacklo_epi32(p1p0, x0); +#endif // !CONFIG_PARALLEL_DEBLOCKING + // 02 12 22 32 42 52 62 72 03 13 23 33 43 53 63 73 + p1p0 = _mm_unpackhi_epi32(p1p0, x0); +#if !CONFIG_PARALLEL_DEBLOCKING + p3p2 = _mm_unpackhi_epi64(p3p2, _mm_slli_si128(p3p2, 8)); // swap lo and high +#endif // !CONFIG_PARALLEL_DEBLOCKING + p1p0 = _mm_unpackhi_epi64(p1p0, _mm_slli_si128(p1p0, 8)); // swap lo and high + + // 04 14 24 34 05 15 25 35 06 16 26 36 07 17 27 37 + q1q0 = _mm_unpackhi_epi16(q1q0, x1); + // 44 54 64 74 45 55 65 75 46 56 66 76 47 57 67 77 + x2 = _mm_unpackhi_epi16(x2, x3); +#if !CONFIG_PARALLEL_DEBLOCKING + // 06 16 26 36 46 56 66 76 07 17 27 37 47 57 67 77 + q3q2 = _mm_unpackhi_epi32(q1q0, x2); +#endif // !CONFIG_PARALLEL_DEBLOCKING + // 04 14 24 34 44 54 64 74 05 15 25 35 45 55 65 75 + q1q0 = _mm_unpacklo_epi32(q1q0, x2); + + q0p0 = _mm_unpacklo_epi64(p1p0, q1q0); + q1p1 = _mm_unpackhi_epi64(p1p0, q1q0); + p1p0 = _mm_unpacklo_epi64(q0p0, q1p1); +#if !CONFIG_PARALLEL_DEBLOCKING + p2p1 = _mm_unpacklo_epi64(q1p1, p3p2); + q2q1 = _mm_unpacklo_epi64(_mm_srli_si128(q1p1, 8), q3q2); +#endif // !CONFIG_PARALLEL_DEBLOCKING +#if !CONFIG_PARALLEL_DEBLOCKING + FILTER_HEV_MASK; +#else // CONFIG_PARALLEL_DEBLOCKING + FILTER_HEV_MASK4; +#endif // !CONFIG_PARALLEL_DEBLOCKING + FILTER4; + + // Transpose 8x4 to 4x8 + // qs1qs0: 20 21 22 23 24 25 26 27 30 31 32 33 34 34 36 37 + // ps1ps0: 10 11 12 13 14 15 16 17 00 01 02 03 04 05 06 07 + // 00 01 02 03 04 05 06 07 10 11 12 13 14 15 16 17 + ps1ps0 = _mm_unpackhi_epi64(ps1ps0, _mm_slli_si128(ps1ps0, 8)); + // 10 30 11 31 12 32 13 33 14 34 15 35 16 36 17 37 + x0 = _mm_unpackhi_epi8(ps1ps0, qs1qs0); + // 00 20 01 21 02 22 03 23 04 24 05 25 06 26 07 27 + ps1ps0 = _mm_unpacklo_epi8(ps1ps0, qs1qs0); + // 04 14 24 34 05 15 25 35 06 16 26 36 07 17 27 37 + qs1qs0 = _mm_unpackhi_epi8(ps1ps0, x0); + // 00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33 + ps1ps0 = _mm_unpacklo_epi8(ps1ps0, x0); + + *(int *)(s + 0 * p - 2) = _mm_cvtsi128_si32(ps1ps0); + ps1ps0 = _mm_srli_si128(ps1ps0, 4); + *(int *)(s + 1 * p - 2) = _mm_cvtsi128_si32(ps1ps0); + ps1ps0 = _mm_srli_si128(ps1ps0, 4); + *(int *)(s + 2 * p - 2) = _mm_cvtsi128_si32(ps1ps0); + ps1ps0 = _mm_srli_si128(ps1ps0, 4); + *(int *)(s + 3 * p - 2) = _mm_cvtsi128_si32(ps1ps0); + + *(int *)(s + 4 * p - 2) = _mm_cvtsi128_si32(qs1qs0); + qs1qs0 = _mm_srli_si128(qs1qs0, 4); + *(int *)(s + 5 * p - 2) = _mm_cvtsi128_si32(qs1qs0); + qs1qs0 = _mm_srli_si128(qs1qs0, 4); + *(int *)(s + 6 * p - 2) = _mm_cvtsi128_si32(qs1qs0); + qs1qs0 = _mm_srli_si128(qs1qs0, 4); + *(int *)(s + 7 * p - 2) = _mm_cvtsi128_si32(qs1qs0); +} + +void aom_lpf_horizontal_edge_8_sse2(unsigned char *s, int p, + const unsigned char *_blimit, + const unsigned char *_limit, + const unsigned char *_thresh) { + const __m128i zero = _mm_set1_epi16(0); + const __m128i one = _mm_set1_epi8(1); + const __m128i blimit = _mm_load_si128((const __m128i *)_blimit); + const __m128i limit = _mm_load_si128((const __m128i *)_limit); + const __m128i thresh = _mm_load_si128((const __m128i *)_thresh); + __m128i mask, hev, flat, flat2; + __m128i q7p7, q6p6, q5p5, q4p4, q3p3, q2p2, q1p1, q0p0, p0q0, p1q1; + __m128i abs_p1p0; + + q4p4 = _mm_loadl_epi64((__m128i *)(s - 5 * p)); + q4p4 = _mm_castps_si128( + _mm_loadh_pi(_mm_castsi128_ps(q4p4), (__m64 *)(s + 4 * p))); + q3p3 = _mm_loadl_epi64((__m128i *)(s - 4 * p)); + q3p3 = _mm_castps_si128( + _mm_loadh_pi(_mm_castsi128_ps(q3p3), (__m64 *)(s + 3 * p))); + q2p2 = _mm_loadl_epi64((__m128i *)(s - 3 * p)); + q2p2 = _mm_castps_si128( + _mm_loadh_pi(_mm_castsi128_ps(q2p2), (__m64 *)(s + 2 * p))); + q1p1 = _mm_loadl_epi64((__m128i *)(s - 2 * p)); + q1p1 = _mm_castps_si128( + _mm_loadh_pi(_mm_castsi128_ps(q1p1), (__m64 *)(s + 1 * p))); + p1q1 = _mm_shuffle_epi32(q1p1, 78); + q0p0 = _mm_loadl_epi64((__m128i *)(s - 1 * p)); + q0p0 = _mm_castps_si128( + _mm_loadh_pi(_mm_castsi128_ps(q0p0), (__m64 *)(s - 0 * p))); + p0q0 = _mm_shuffle_epi32(q0p0, 78); + + { + __m128i abs_p1q1, abs_p0q0, abs_q1q0, fe, ff, work; + abs_p1p0 = abs_diff(q1p1, q0p0); + abs_q1q0 = _mm_srli_si128(abs_p1p0, 8); + fe = _mm_set1_epi8(0xfe); + ff = _mm_cmpeq_epi8(abs_p1p0, abs_p1p0); + abs_p0q0 = abs_diff(q0p0, p0q0); + abs_p1q1 = abs_diff(q1p1, p1q1); + flat = _mm_max_epu8(abs_p1p0, abs_q1q0); + hev = _mm_subs_epu8(flat, thresh); + hev = _mm_xor_si128(_mm_cmpeq_epi8(hev, zero), ff); + + abs_p0q0 = _mm_adds_epu8(abs_p0q0, abs_p0q0); + abs_p1q1 = _mm_srli_epi16(_mm_and_si128(abs_p1q1, fe), 1); + mask = _mm_subs_epu8(_mm_adds_epu8(abs_p0q0, abs_p1q1), blimit); + mask = _mm_xor_si128(_mm_cmpeq_epi8(mask, zero), ff); + // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 > blimit) * -1; + mask = _mm_max_epu8(abs_p1p0, mask); + // mask |= (abs(p1 - p0) > limit) * -1; + // mask |= (abs(q1 - q0) > limit) * -1; + + work = _mm_max_epu8(abs_diff(q2p2, q1p1), abs_diff(q3p3, q2p2)); + mask = _mm_max_epu8(work, mask); + mask = _mm_max_epu8(mask, _mm_srli_si128(mask, 8)); + mask = _mm_subs_epu8(mask, limit); + mask = _mm_cmpeq_epi8(mask, zero); + } + + // lp filter + { + const __m128i t4 = _mm_set1_epi8(4); + const __m128i t3 = _mm_set1_epi8(3); + const __m128i t80 = _mm_set1_epi8(0x80); + const __m128i t1 = _mm_set1_epi16(0x1); + __m128i qs1ps1 = _mm_xor_si128(q1p1, t80); + __m128i qs0ps0 = _mm_xor_si128(q0p0, t80); + __m128i qs0 = _mm_xor_si128(p0q0, t80); + __m128i qs1 = _mm_xor_si128(p1q1, t80); + __m128i filt; + __m128i work_a; + __m128i filter1, filter2; + __m128i flat2_q6p6, flat2_q5p5, flat2_q4p4, flat2_q3p3, flat2_q2p2; + __m128i flat2_q1p1, flat2_q0p0, flat_q2p2, flat_q1p1, flat_q0p0; + + filt = _mm_and_si128(_mm_subs_epi8(qs1ps1, qs1), hev); + work_a = _mm_subs_epi8(qs0, qs0ps0); + filt = _mm_adds_epi8(filt, work_a); + filt = _mm_adds_epi8(filt, work_a); + filt = _mm_adds_epi8(filt, work_a); + // (aom_filter + 3 * (qs0 - ps0)) & mask + filt = _mm_and_si128(filt, mask); + + filter1 = _mm_adds_epi8(filt, t4); + filter2 = _mm_adds_epi8(filt, t3); + + filter1 = _mm_unpacklo_epi8(zero, filter1); + filter1 = _mm_srai_epi16(filter1, 0xB); + filter2 = _mm_unpacklo_epi8(zero, filter2); + filter2 = _mm_srai_epi16(filter2, 0xB); + + // Filter1 >> 3 + filt = _mm_packs_epi16(filter2, _mm_subs_epi16(zero, filter1)); + qs0ps0 = _mm_xor_si128(_mm_adds_epi8(qs0ps0, filt), t80); + + // filt >> 1 + filt = _mm_adds_epi16(filter1, t1); + filt = _mm_srai_epi16(filt, 1); + filt = _mm_andnot_si128(_mm_srai_epi16(_mm_unpacklo_epi8(zero, hev), 0x8), + filt); + filt = _mm_packs_epi16(filt, _mm_subs_epi16(zero, filt)); + qs1ps1 = _mm_xor_si128(_mm_adds_epi8(qs1ps1, filt), t80); + // loopfilter done + + { + __m128i work; + flat = _mm_max_epu8(abs_diff(q2p2, q0p0), abs_diff(q3p3, q0p0)); + flat = _mm_max_epu8(abs_p1p0, flat); + flat = _mm_max_epu8(flat, _mm_srli_si128(flat, 8)); + flat = _mm_subs_epu8(flat, one); + flat = _mm_cmpeq_epi8(flat, zero); + flat = _mm_and_si128(flat, mask); + + q5p5 = _mm_loadl_epi64((__m128i *)(s - 6 * p)); + q5p5 = _mm_castps_si128( + _mm_loadh_pi(_mm_castsi128_ps(q5p5), (__m64 *)(s + 5 * p))); + + q6p6 = _mm_loadl_epi64((__m128i *)(s - 7 * p)); + q6p6 = _mm_castps_si128( + _mm_loadh_pi(_mm_castsi128_ps(q6p6), (__m64 *)(s + 6 * p))); + flat2 = _mm_max_epu8(abs_diff(q4p4, q0p0), abs_diff(q5p5, q0p0)); + + q7p7 = _mm_loadl_epi64((__m128i *)(s - 8 * p)); + q7p7 = _mm_castps_si128( + _mm_loadh_pi(_mm_castsi128_ps(q7p7), (__m64 *)(s + 7 * p))); + work = _mm_max_epu8(abs_diff(q6p6, q0p0), abs_diff(q7p7, q0p0)); + flat2 = _mm_max_epu8(work, flat2); + flat2 = _mm_max_epu8(flat2, _mm_srli_si128(flat2, 8)); + flat2 = _mm_subs_epu8(flat2, one); + flat2 = _mm_cmpeq_epi8(flat2, zero); + flat2 = _mm_and_si128(flat2, flat); // flat2 & flat & mask + } + + // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ + // flat and wide flat calculations + { + const __m128i eight = _mm_set1_epi16(8); + const __m128i four = _mm_set1_epi16(4); + __m128i p7_16, p6_16, p5_16, p4_16, p3_16, p2_16, p1_16, p0_16; + __m128i q7_16, q6_16, q5_16, q4_16, q3_16, q2_16, q1_16, q0_16; + __m128i pixelFilter_p, pixelFilter_q; + __m128i pixetFilter_p2p1p0, pixetFilter_q2q1q0; + __m128i sum_p7, sum_q7, sum_p3, sum_q3, res_p, res_q; + + p7_16 = _mm_unpacklo_epi8(q7p7, zero); + p6_16 = _mm_unpacklo_epi8(q6p6, zero); + p5_16 = _mm_unpacklo_epi8(q5p5, zero); + p4_16 = _mm_unpacklo_epi8(q4p4, zero); + p3_16 = _mm_unpacklo_epi8(q3p3, zero); + p2_16 = _mm_unpacklo_epi8(q2p2, zero); + p1_16 = _mm_unpacklo_epi8(q1p1, zero); + p0_16 = _mm_unpacklo_epi8(q0p0, zero); + q0_16 = _mm_unpackhi_epi8(q0p0, zero); + q1_16 = _mm_unpackhi_epi8(q1p1, zero); + q2_16 = _mm_unpackhi_epi8(q2p2, zero); + q3_16 = _mm_unpackhi_epi8(q3p3, zero); + q4_16 = _mm_unpackhi_epi8(q4p4, zero); + q5_16 = _mm_unpackhi_epi8(q5p5, zero); + q6_16 = _mm_unpackhi_epi8(q6p6, zero); + q7_16 = _mm_unpackhi_epi8(q7p7, zero); + + pixelFilter_p = _mm_add_epi16(_mm_add_epi16(p6_16, p5_16), + _mm_add_epi16(p4_16, p3_16)); + pixelFilter_q = _mm_add_epi16(_mm_add_epi16(q6_16, q5_16), + _mm_add_epi16(q4_16, q3_16)); + + pixetFilter_p2p1p0 = _mm_add_epi16(p0_16, _mm_add_epi16(p2_16, p1_16)); + pixelFilter_p = _mm_add_epi16(pixelFilter_p, pixetFilter_p2p1p0); + + pixetFilter_q2q1q0 = _mm_add_epi16(q0_16, _mm_add_epi16(q2_16, q1_16)); + pixelFilter_q = _mm_add_epi16(pixelFilter_q, pixetFilter_q2q1q0); + pixelFilter_p = + _mm_add_epi16(eight, _mm_add_epi16(pixelFilter_p, pixelFilter_q)); + pixetFilter_p2p1p0 = _mm_add_epi16( + four, _mm_add_epi16(pixetFilter_p2p1p0, pixetFilter_q2q1q0)); + res_p = _mm_srli_epi16( + _mm_add_epi16(pixelFilter_p, _mm_add_epi16(p7_16, p0_16)), 4); + res_q = _mm_srli_epi16( + _mm_add_epi16(pixelFilter_p, _mm_add_epi16(q7_16, q0_16)), 4); + flat2_q0p0 = _mm_packus_epi16(res_p, res_q); + res_p = _mm_srli_epi16( + _mm_add_epi16(pixetFilter_p2p1p0, _mm_add_epi16(p3_16, p0_16)), 3); + res_q = _mm_srli_epi16( + _mm_add_epi16(pixetFilter_p2p1p0, _mm_add_epi16(q3_16, q0_16)), 3); + + flat_q0p0 = _mm_packus_epi16(res_p, res_q); + + sum_p7 = _mm_add_epi16(p7_16, p7_16); + sum_q7 = _mm_add_epi16(q7_16, q7_16); + sum_p3 = _mm_add_epi16(p3_16, p3_16); + sum_q3 = _mm_add_epi16(q3_16, q3_16); + + pixelFilter_q = _mm_sub_epi16(pixelFilter_p, p6_16); + pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q6_16); + res_p = _mm_srli_epi16( + _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p1_16)), 4); + res_q = _mm_srli_epi16( + _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q1_16)), 4); + flat2_q1p1 = _mm_packus_epi16(res_p, res_q); + + pixetFilter_q2q1q0 = _mm_sub_epi16(pixetFilter_p2p1p0, p2_16); + pixetFilter_p2p1p0 = _mm_sub_epi16(pixetFilter_p2p1p0, q2_16); + res_p = _mm_srli_epi16( + _mm_add_epi16(pixetFilter_p2p1p0, _mm_add_epi16(sum_p3, p1_16)), 3); + res_q = _mm_srli_epi16( + _mm_add_epi16(pixetFilter_q2q1q0, _mm_add_epi16(sum_q3, q1_16)), 3); + flat_q1p1 = _mm_packus_epi16(res_p, res_q); + + sum_p7 = _mm_add_epi16(sum_p7, p7_16); + sum_q7 = _mm_add_epi16(sum_q7, q7_16); + sum_p3 = _mm_add_epi16(sum_p3, p3_16); + sum_q3 = _mm_add_epi16(sum_q3, q3_16); + + pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q5_16); + pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p5_16); + res_p = _mm_srli_epi16( + _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p2_16)), 4); + res_q = _mm_srli_epi16( + _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q2_16)), 4); + flat2_q2p2 = _mm_packus_epi16(res_p, res_q); + + pixetFilter_p2p1p0 = _mm_sub_epi16(pixetFilter_p2p1p0, q1_16); + pixetFilter_q2q1q0 = _mm_sub_epi16(pixetFilter_q2q1q0, p1_16); + + res_p = _mm_srli_epi16( + _mm_add_epi16(pixetFilter_p2p1p0, _mm_add_epi16(sum_p3, p2_16)), 3); + res_q = _mm_srli_epi16( + _mm_add_epi16(pixetFilter_q2q1q0, _mm_add_epi16(sum_q3, q2_16)), 3); + flat_q2p2 = _mm_packus_epi16(res_p, res_q); + + sum_p7 = _mm_add_epi16(sum_p7, p7_16); + sum_q7 = _mm_add_epi16(sum_q7, q7_16); + pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q4_16); + pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p4_16); + res_p = _mm_srli_epi16( + _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p3_16)), 4); + res_q = _mm_srli_epi16( + _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q3_16)), 4); + flat2_q3p3 = _mm_packus_epi16(res_p, res_q); + + sum_p7 = _mm_add_epi16(sum_p7, p7_16); + sum_q7 = _mm_add_epi16(sum_q7, q7_16); + pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q3_16); + pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p3_16); + res_p = _mm_srli_epi16( + _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p4_16)), 4); + res_q = _mm_srli_epi16( + _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q4_16)), 4); + flat2_q4p4 = _mm_packus_epi16(res_p, res_q); + + sum_p7 = _mm_add_epi16(sum_p7, p7_16); + sum_q7 = _mm_add_epi16(sum_q7, q7_16); + pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q2_16); + pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p2_16); + res_p = _mm_srli_epi16( + _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p5_16)), 4); + res_q = _mm_srli_epi16( + _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q5_16)), 4); + flat2_q5p5 = _mm_packus_epi16(res_p, res_q); + + sum_p7 = _mm_add_epi16(sum_p7, p7_16); + sum_q7 = _mm_add_epi16(sum_q7, q7_16); + pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q1_16); + pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p1_16); + res_p = _mm_srli_epi16( + _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p6_16)), 4); + res_q = _mm_srli_epi16( + _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q6_16)), 4); + flat2_q6p6 = _mm_packus_epi16(res_p, res_q); + } + // wide flat + // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ + + flat = _mm_shuffle_epi32(flat, 68); + flat2 = _mm_shuffle_epi32(flat2, 68); + + q2p2 = _mm_andnot_si128(flat, q2p2); + flat_q2p2 = _mm_and_si128(flat, flat_q2p2); + q2p2 = _mm_or_si128(q2p2, flat_q2p2); + + qs1ps1 = _mm_andnot_si128(flat, qs1ps1); + flat_q1p1 = _mm_and_si128(flat, flat_q1p1); + q1p1 = _mm_or_si128(qs1ps1, flat_q1p1); + + qs0ps0 = _mm_andnot_si128(flat, qs0ps0); + flat_q0p0 = _mm_and_si128(flat, flat_q0p0); + q0p0 = _mm_or_si128(qs0ps0, flat_q0p0); + + q6p6 = _mm_andnot_si128(flat2, q6p6); + flat2_q6p6 = _mm_and_si128(flat2, flat2_q6p6); + q6p6 = _mm_or_si128(q6p6, flat2_q6p6); + _mm_storel_epi64((__m128i *)(s - 7 * p), q6p6); + _mm_storeh_pi((__m64 *)(s + 6 * p), _mm_castsi128_ps(q6p6)); + + q5p5 = _mm_andnot_si128(flat2, q5p5); + flat2_q5p5 = _mm_and_si128(flat2, flat2_q5p5); + q5p5 = _mm_or_si128(q5p5, flat2_q5p5); + _mm_storel_epi64((__m128i *)(s - 6 * p), q5p5); + _mm_storeh_pi((__m64 *)(s + 5 * p), _mm_castsi128_ps(q5p5)); + + q4p4 = _mm_andnot_si128(flat2, q4p4); + flat2_q4p4 = _mm_and_si128(flat2, flat2_q4p4); + q4p4 = _mm_or_si128(q4p4, flat2_q4p4); + _mm_storel_epi64((__m128i *)(s - 5 * p), q4p4); + _mm_storeh_pi((__m64 *)(s + 4 * p), _mm_castsi128_ps(q4p4)); + + q3p3 = _mm_andnot_si128(flat2, q3p3); + flat2_q3p3 = _mm_and_si128(flat2, flat2_q3p3); + q3p3 = _mm_or_si128(q3p3, flat2_q3p3); + _mm_storel_epi64((__m128i *)(s - 4 * p), q3p3); + _mm_storeh_pi((__m64 *)(s + 3 * p), _mm_castsi128_ps(q3p3)); + + q2p2 = _mm_andnot_si128(flat2, q2p2); + flat2_q2p2 = _mm_and_si128(flat2, flat2_q2p2); + q2p2 = _mm_or_si128(q2p2, flat2_q2p2); + _mm_storel_epi64((__m128i *)(s - 3 * p), q2p2); + _mm_storeh_pi((__m64 *)(s + 2 * p), _mm_castsi128_ps(q2p2)); + + q1p1 = _mm_andnot_si128(flat2, q1p1); + flat2_q1p1 = _mm_and_si128(flat2, flat2_q1p1); + q1p1 = _mm_or_si128(q1p1, flat2_q1p1); + _mm_storel_epi64((__m128i *)(s - 2 * p), q1p1); + _mm_storeh_pi((__m64 *)(s + 1 * p), _mm_castsi128_ps(q1p1)); + + q0p0 = _mm_andnot_si128(flat2, q0p0); + flat2_q0p0 = _mm_and_si128(flat2, flat2_q0p0); + q0p0 = _mm_or_si128(q0p0, flat2_q0p0); + _mm_storel_epi64((__m128i *)(s - 1 * p), q0p0); + _mm_storeh_pi((__m64 *)(s - 0 * p), _mm_castsi128_ps(q0p0)); + } +} + +static INLINE __m128i filter_add2_sub2(const __m128i *const total, + const __m128i *const a1, + const __m128i *const a2, + const __m128i *const s1, + const __m128i *const s2) { + __m128i x = _mm_add_epi16(*a1, *total); + x = _mm_add_epi16(_mm_sub_epi16(x, _mm_add_epi16(*s1, *s2)), *a2); + return x; +} + +static INLINE __m128i filter8_mask(const __m128i *const flat, + const __m128i *const other_filt, + const __m128i *const f8_lo, + const __m128i *const f8_hi) { + const __m128i f8 = + _mm_packus_epi16(_mm_srli_epi16(*f8_lo, 3), _mm_srli_epi16(*f8_hi, 3)); + const __m128i result = _mm_and_si128(*flat, f8); + return _mm_or_si128(_mm_andnot_si128(*flat, *other_filt), result); +} + +static INLINE __m128i filter16_mask(const __m128i *const flat, + const __m128i *const other_filt, + const __m128i *const f_lo, + const __m128i *const f_hi) { + const __m128i f = + _mm_packus_epi16(_mm_srli_epi16(*f_lo, 4), _mm_srli_epi16(*f_hi, 4)); + const __m128i result = _mm_and_si128(*flat, f); + return _mm_or_si128(_mm_andnot_si128(*flat, *other_filt), result); +} + +void aom_lpf_horizontal_edge_16_sse2(unsigned char *s, int p, + const unsigned char *_blimit, + const unsigned char *_limit, + const unsigned char *_thresh) { + const __m128i zero = _mm_set1_epi16(0); + const __m128i one = _mm_set1_epi8(1); + const __m128i blimit = _mm_load_si128((const __m128i *)_blimit); + const __m128i limit = _mm_load_si128((const __m128i *)_limit); + const __m128i thresh = _mm_load_si128((const __m128i *)_thresh); + __m128i mask, hev, flat, flat2; + __m128i p7, p6, p5; + __m128i p4, p3, p2, p1, p0, q0, q1, q2, q3, q4; + __m128i q5, q6, q7; + + __m128i op2, op1, op0, oq0, oq1, oq2; + + __m128i max_abs_p1p0q1q0; + + p7 = _mm_loadu_si128((__m128i *)(s - 8 * p)); + p6 = _mm_loadu_si128((__m128i *)(s - 7 * p)); + p5 = _mm_loadu_si128((__m128i *)(s - 6 * p)); + p4 = _mm_loadu_si128((__m128i *)(s - 5 * p)); + p3 = _mm_loadu_si128((__m128i *)(s - 4 * p)); + p2 = _mm_loadu_si128((__m128i *)(s - 3 * p)); + p1 = _mm_loadu_si128((__m128i *)(s - 2 * p)); + p0 = _mm_loadu_si128((__m128i *)(s - 1 * p)); + q0 = _mm_loadu_si128((__m128i *)(s - 0 * p)); + q1 = _mm_loadu_si128((__m128i *)(s + 1 * p)); + q2 = _mm_loadu_si128((__m128i *)(s + 2 * p)); + q3 = _mm_loadu_si128((__m128i *)(s + 3 * p)); + q4 = _mm_loadu_si128((__m128i *)(s + 4 * p)); + q5 = _mm_loadu_si128((__m128i *)(s + 5 * p)); + q6 = _mm_loadu_si128((__m128i *)(s + 6 * p)); + q7 = _mm_loadu_si128((__m128i *)(s + 7 * p)); + + { + const __m128i abs_p1p0 = abs_diff(p1, p0); + const __m128i abs_q1q0 = abs_diff(q1, q0); + const __m128i fe = _mm_set1_epi8(0xfe); + const __m128i ff = _mm_cmpeq_epi8(zero, zero); + __m128i abs_p0q0 = abs_diff(p0, q0); + __m128i abs_p1q1 = abs_diff(p1, q1); + __m128i work; + max_abs_p1p0q1q0 = _mm_max_epu8(abs_p1p0, abs_q1q0); + + abs_p0q0 = _mm_adds_epu8(abs_p0q0, abs_p0q0); + abs_p1q1 = _mm_srli_epi16(_mm_and_si128(abs_p1q1, fe), 1); + mask = _mm_subs_epu8(_mm_adds_epu8(abs_p0q0, abs_p1q1), blimit); + mask = _mm_xor_si128(_mm_cmpeq_epi8(mask, zero), ff); + // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 > blimit) * -1; + mask = _mm_max_epu8(max_abs_p1p0q1q0, mask); + // mask |= (abs(p1 - p0) > limit) * -1; + // mask |= (abs(q1 - q0) > limit) * -1; + work = _mm_max_epu8(abs_diff(p2, p1), abs_diff(p3, p2)); + mask = _mm_max_epu8(work, mask); + work = _mm_max_epu8(abs_diff(q2, q1), abs_diff(q3, q2)); + mask = _mm_max_epu8(work, mask); + mask = _mm_subs_epu8(mask, limit); + mask = _mm_cmpeq_epi8(mask, zero); + } + + { + __m128i work; + work = _mm_max_epu8(abs_diff(p2, p0), abs_diff(q2, q0)); + flat = _mm_max_epu8(work, max_abs_p1p0q1q0); + work = _mm_max_epu8(abs_diff(p3, p0), abs_diff(q3, q0)); + flat = _mm_max_epu8(work, flat); + work = _mm_max_epu8(abs_diff(p4, p0), abs_diff(q4, q0)); + flat = _mm_subs_epu8(flat, one); + flat = _mm_cmpeq_epi8(flat, zero); + flat = _mm_and_si128(flat, mask); + flat2 = _mm_max_epu8(abs_diff(p5, p0), abs_diff(q5, q0)); + flat2 = _mm_max_epu8(work, flat2); + work = _mm_max_epu8(abs_diff(p6, p0), abs_diff(q6, q0)); + flat2 = _mm_max_epu8(work, flat2); + work = _mm_max_epu8(abs_diff(p7, p0), abs_diff(q7, q0)); + flat2 = _mm_max_epu8(work, flat2); + flat2 = _mm_subs_epu8(flat2, one); + flat2 = _mm_cmpeq_epi8(flat2, zero); + flat2 = _mm_and_si128(flat2, flat); // flat2 & flat & mask + } + + // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ + // filter4 + { + const __m128i t4 = _mm_set1_epi8(4); + const __m128i t3 = _mm_set1_epi8(3); + const __m128i t80 = _mm_set1_epi8(0x80); + const __m128i te0 = _mm_set1_epi8(0xe0); + const __m128i t1f = _mm_set1_epi8(0x1f); + const __m128i t1 = _mm_set1_epi8(0x1); + const __m128i t7f = _mm_set1_epi8(0x7f); + const __m128i ff = _mm_cmpeq_epi8(t4, t4); + + __m128i filt; + __m128i work_a; + __m128i filter1, filter2; + + op1 = _mm_xor_si128(p1, t80); + op0 = _mm_xor_si128(p0, t80); + oq0 = _mm_xor_si128(q0, t80); + oq1 = _mm_xor_si128(q1, t80); + + hev = _mm_subs_epu8(max_abs_p1p0q1q0, thresh); + hev = _mm_xor_si128(_mm_cmpeq_epi8(hev, zero), ff); + filt = _mm_and_si128(_mm_subs_epi8(op1, oq1), hev); + + work_a = _mm_subs_epi8(oq0, op0); + filt = _mm_adds_epi8(filt, work_a); + filt = _mm_adds_epi8(filt, work_a); + filt = _mm_adds_epi8(filt, work_a); + // (aom_filter + 3 * (qs0 - ps0)) & mask + filt = _mm_and_si128(filt, mask); + filter1 = _mm_adds_epi8(filt, t4); + filter2 = _mm_adds_epi8(filt, t3); + + // Filter1 >> 3 + work_a = _mm_cmpgt_epi8(zero, filter1); + filter1 = _mm_srli_epi16(filter1, 3); + work_a = _mm_and_si128(work_a, te0); + filter1 = _mm_and_si128(filter1, t1f); + filter1 = _mm_or_si128(filter1, work_a); + oq0 = _mm_xor_si128(_mm_subs_epi8(oq0, filter1), t80); + + // Filter2 >> 3 + work_a = _mm_cmpgt_epi8(zero, filter2); + filter2 = _mm_srli_epi16(filter2, 3); + work_a = _mm_and_si128(work_a, te0); + filter2 = _mm_and_si128(filter2, t1f); + filter2 = _mm_or_si128(filter2, work_a); + op0 = _mm_xor_si128(_mm_adds_epi8(op0, filter2), t80); + + // filt >> 1 + filt = _mm_adds_epi8(filter1, t1); + work_a = _mm_cmpgt_epi8(zero, filt); + filt = _mm_srli_epi16(filt, 1); + work_a = _mm_and_si128(work_a, t80); + filt = _mm_and_si128(filt, t7f); + filt = _mm_or_si128(filt, work_a); + filt = _mm_andnot_si128(hev, filt); + op1 = _mm_xor_si128(_mm_adds_epi8(op1, filt), t80); + oq1 = _mm_xor_si128(_mm_subs_epi8(oq1, filt), t80); + // loopfilter done + + // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ + // filter8 + { + const __m128i four = _mm_set1_epi16(4); + const __m128i p3_lo = _mm_unpacklo_epi8(p3, zero); + const __m128i p2_lo = _mm_unpacklo_epi8(p2, zero); + const __m128i p1_lo = _mm_unpacklo_epi8(p1, zero); + const __m128i p0_lo = _mm_unpacklo_epi8(p0, zero); + const __m128i q0_lo = _mm_unpacklo_epi8(q0, zero); + const __m128i q1_lo = _mm_unpacklo_epi8(q1, zero); + const __m128i q2_lo = _mm_unpacklo_epi8(q2, zero); + const __m128i q3_lo = _mm_unpacklo_epi8(q3, zero); + + const __m128i p3_hi = _mm_unpackhi_epi8(p3, zero); + const __m128i p2_hi = _mm_unpackhi_epi8(p2, zero); + const __m128i p1_hi = _mm_unpackhi_epi8(p1, zero); + const __m128i p0_hi = _mm_unpackhi_epi8(p0, zero); + const __m128i q0_hi = _mm_unpackhi_epi8(q0, zero); + const __m128i q1_hi = _mm_unpackhi_epi8(q1, zero); + const __m128i q2_hi = _mm_unpackhi_epi8(q2, zero); + const __m128i q3_hi = _mm_unpackhi_epi8(q3, zero); + __m128i f8_lo, f8_hi; + + f8_lo = _mm_add_epi16(_mm_add_epi16(p3_lo, four), + _mm_add_epi16(p3_lo, p2_lo)); + f8_lo = _mm_add_epi16(_mm_add_epi16(p3_lo, f8_lo), + _mm_add_epi16(p2_lo, p1_lo)); + f8_lo = _mm_add_epi16(_mm_add_epi16(p0_lo, q0_lo), f8_lo); + + f8_hi = _mm_add_epi16(_mm_add_epi16(p3_hi, four), + _mm_add_epi16(p3_hi, p2_hi)); + f8_hi = _mm_add_epi16(_mm_add_epi16(p3_hi, f8_hi), + _mm_add_epi16(p2_hi, p1_hi)); + f8_hi = _mm_add_epi16(_mm_add_epi16(p0_hi, q0_hi), f8_hi); + + op2 = filter8_mask(&flat, &p2, &f8_lo, &f8_hi); + + f8_lo = filter_add2_sub2(&f8_lo, &q1_lo, &p1_lo, &p2_lo, &p3_lo); + f8_hi = filter_add2_sub2(&f8_hi, &q1_hi, &p1_hi, &p2_hi, &p3_hi); + op1 = filter8_mask(&flat, &op1, &f8_lo, &f8_hi); + + f8_lo = filter_add2_sub2(&f8_lo, &q2_lo, &p0_lo, &p1_lo, &p3_lo); + f8_hi = filter_add2_sub2(&f8_hi, &q2_hi, &p0_hi, &p1_hi, &p3_hi); + op0 = filter8_mask(&flat, &op0, &f8_lo, &f8_hi); + + f8_lo = filter_add2_sub2(&f8_lo, &q3_lo, &q0_lo, &p0_lo, &p3_lo); + f8_hi = filter_add2_sub2(&f8_hi, &q3_hi, &q0_hi, &p0_hi, &p3_hi); + oq0 = filter8_mask(&flat, &oq0, &f8_lo, &f8_hi); + + f8_lo = filter_add2_sub2(&f8_lo, &q3_lo, &q1_lo, &q0_lo, &p2_lo); + f8_hi = filter_add2_sub2(&f8_hi, &q3_hi, &q1_hi, &q0_hi, &p2_hi); + oq1 = filter8_mask(&flat, &oq1, &f8_lo, &f8_hi); + + f8_lo = filter_add2_sub2(&f8_lo, &q3_lo, &q2_lo, &q1_lo, &p1_lo); + f8_hi = filter_add2_sub2(&f8_hi, &q3_hi, &q2_hi, &q1_hi, &p1_hi); + oq2 = filter8_mask(&flat, &q2, &f8_lo, &f8_hi); + } + + // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ + // wide flat calculations + { + const __m128i eight = _mm_set1_epi16(8); + const __m128i p7_lo = _mm_unpacklo_epi8(p7, zero); + const __m128i p6_lo = _mm_unpacklo_epi8(p6, zero); + const __m128i p5_lo = _mm_unpacklo_epi8(p5, zero); + const __m128i p4_lo = _mm_unpacklo_epi8(p4, zero); + const __m128i p3_lo = _mm_unpacklo_epi8(p3, zero); + const __m128i p2_lo = _mm_unpacklo_epi8(p2, zero); + const __m128i p1_lo = _mm_unpacklo_epi8(p1, zero); + const __m128i p0_lo = _mm_unpacklo_epi8(p0, zero); + const __m128i q0_lo = _mm_unpacklo_epi8(q0, zero); + const __m128i q1_lo = _mm_unpacklo_epi8(q1, zero); + const __m128i q2_lo = _mm_unpacklo_epi8(q2, zero); + const __m128i q3_lo = _mm_unpacklo_epi8(q3, zero); + const __m128i q4_lo = _mm_unpacklo_epi8(q4, zero); + const __m128i q5_lo = _mm_unpacklo_epi8(q5, zero); + const __m128i q6_lo = _mm_unpacklo_epi8(q6, zero); + const __m128i q7_lo = _mm_unpacklo_epi8(q7, zero); + + const __m128i p7_hi = _mm_unpackhi_epi8(p7, zero); + const __m128i p6_hi = _mm_unpackhi_epi8(p6, zero); + const __m128i p5_hi = _mm_unpackhi_epi8(p5, zero); + const __m128i p4_hi = _mm_unpackhi_epi8(p4, zero); + const __m128i p3_hi = _mm_unpackhi_epi8(p3, zero); + const __m128i p2_hi = _mm_unpackhi_epi8(p2, zero); + const __m128i p1_hi = _mm_unpackhi_epi8(p1, zero); + const __m128i p0_hi = _mm_unpackhi_epi8(p0, zero); + const __m128i q0_hi = _mm_unpackhi_epi8(q0, zero); + const __m128i q1_hi = _mm_unpackhi_epi8(q1, zero); + const __m128i q2_hi = _mm_unpackhi_epi8(q2, zero); + const __m128i q3_hi = _mm_unpackhi_epi8(q3, zero); + const __m128i q4_hi = _mm_unpackhi_epi8(q4, zero); + const __m128i q5_hi = _mm_unpackhi_epi8(q5, zero); + const __m128i q6_hi = _mm_unpackhi_epi8(q6, zero); + const __m128i q7_hi = _mm_unpackhi_epi8(q7, zero); + + __m128i f_lo; + __m128i f_hi; + + f_lo = _mm_sub_epi16(_mm_slli_epi16(p7_lo, 3), p7_lo); // p7 * 7 + f_lo = + _mm_add_epi16(_mm_slli_epi16(p6_lo, 1), _mm_add_epi16(p4_lo, f_lo)); + f_lo = _mm_add_epi16(_mm_add_epi16(p3_lo, f_lo), + _mm_add_epi16(p2_lo, p1_lo)); + f_lo = _mm_add_epi16(_mm_add_epi16(p0_lo, q0_lo), f_lo); + f_lo = _mm_add_epi16(_mm_add_epi16(p5_lo, eight), f_lo); + + f_hi = _mm_sub_epi16(_mm_slli_epi16(p7_hi, 3), p7_hi); // p7 * 7 + f_hi = + _mm_add_epi16(_mm_slli_epi16(p6_hi, 1), _mm_add_epi16(p4_hi, f_hi)); + f_hi = _mm_add_epi16(_mm_add_epi16(p3_hi, f_hi), + _mm_add_epi16(p2_hi, p1_hi)); + f_hi = _mm_add_epi16(_mm_add_epi16(p0_hi, q0_hi), f_hi); + f_hi = _mm_add_epi16(_mm_add_epi16(p5_hi, eight), f_hi); + + p6 = filter16_mask(&flat2, &p6, &f_lo, &f_hi); + _mm_storeu_si128((__m128i *)(s - 7 * p), p6); + + f_lo = filter_add2_sub2(&f_lo, &q1_lo, &p5_lo, &p6_lo, &p7_lo); + f_hi = filter_add2_sub2(&f_hi, &q1_hi, &p5_hi, &p6_hi, &p7_hi); + p5 = filter16_mask(&flat2, &p5, &f_lo, &f_hi); + _mm_storeu_si128((__m128i *)(s - 6 * p), p5); + + f_lo = filter_add2_sub2(&f_lo, &q2_lo, &p4_lo, &p5_lo, &p7_lo); + f_hi = filter_add2_sub2(&f_hi, &q2_hi, &p4_hi, &p5_hi, &p7_hi); + p4 = filter16_mask(&flat2, &p4, &f_lo, &f_hi); + _mm_storeu_si128((__m128i *)(s - 5 * p), p4); + + f_lo = filter_add2_sub2(&f_lo, &q3_lo, &p3_lo, &p4_lo, &p7_lo); + f_hi = filter_add2_sub2(&f_hi, &q3_hi, &p3_hi, &p4_hi, &p7_hi); + p3 = filter16_mask(&flat2, &p3, &f_lo, &f_hi); + _mm_storeu_si128((__m128i *)(s - 4 * p), p3); + + f_lo = filter_add2_sub2(&f_lo, &q4_lo, &p2_lo, &p3_lo, &p7_lo); + f_hi = filter_add2_sub2(&f_hi, &q4_hi, &p2_hi, &p3_hi, &p7_hi); + op2 = filter16_mask(&flat2, &op2, &f_lo, &f_hi); + _mm_storeu_si128((__m128i *)(s - 3 * p), op2); + + f_lo = filter_add2_sub2(&f_lo, &q5_lo, &p1_lo, &p2_lo, &p7_lo); + f_hi = filter_add2_sub2(&f_hi, &q5_hi, &p1_hi, &p2_hi, &p7_hi); + op1 = filter16_mask(&flat2, &op1, &f_lo, &f_hi); + _mm_storeu_si128((__m128i *)(s - 2 * p), op1); + + f_lo = filter_add2_sub2(&f_lo, &q6_lo, &p0_lo, &p1_lo, &p7_lo); + f_hi = filter_add2_sub2(&f_hi, &q6_hi, &p0_hi, &p1_hi, &p7_hi); + op0 = filter16_mask(&flat2, &op0, &f_lo, &f_hi); + _mm_storeu_si128((__m128i *)(s - 1 * p), op0); + + f_lo = filter_add2_sub2(&f_lo, &q7_lo, &q0_lo, &p0_lo, &p7_lo); + f_hi = filter_add2_sub2(&f_hi, &q7_hi, &q0_hi, &p0_hi, &p7_hi); + oq0 = filter16_mask(&flat2, &oq0, &f_lo, &f_hi); + _mm_storeu_si128((__m128i *)(s - 0 * p), oq0); + + f_lo = filter_add2_sub2(&f_lo, &q7_lo, &q1_lo, &p6_lo, &q0_lo); + f_hi = filter_add2_sub2(&f_hi, &q7_hi, &q1_hi, &p6_hi, &q0_hi); + oq1 = filter16_mask(&flat2, &oq1, &f_lo, &f_hi); + _mm_storeu_si128((__m128i *)(s + 1 * p), oq1); + + f_lo = filter_add2_sub2(&f_lo, &q7_lo, &q2_lo, &p5_lo, &q1_lo); + f_hi = filter_add2_sub2(&f_hi, &q7_hi, &q2_hi, &p5_hi, &q1_hi); + oq2 = filter16_mask(&flat2, &oq2, &f_lo, &f_hi); + _mm_storeu_si128((__m128i *)(s + 2 * p), oq2); + + f_lo = filter_add2_sub2(&f_lo, &q7_lo, &q3_lo, &p4_lo, &q2_lo); + f_hi = filter_add2_sub2(&f_hi, &q7_hi, &q3_hi, &p4_hi, &q2_hi); + q3 = filter16_mask(&flat2, &q3, &f_lo, &f_hi); + _mm_storeu_si128((__m128i *)(s + 3 * p), q3); + + f_lo = filter_add2_sub2(&f_lo, &q7_lo, &q4_lo, &p3_lo, &q3_lo); + f_hi = filter_add2_sub2(&f_hi, &q7_hi, &q4_hi, &p3_hi, &q3_hi); + q4 = filter16_mask(&flat2, &q4, &f_lo, &f_hi); + _mm_storeu_si128((__m128i *)(s + 4 * p), q4); + + f_lo = filter_add2_sub2(&f_lo, &q7_lo, &q5_lo, &p2_lo, &q4_lo); + f_hi = filter_add2_sub2(&f_hi, &q7_hi, &q5_hi, &p2_hi, &q4_hi); + q5 = filter16_mask(&flat2, &q5, &f_lo, &f_hi); + _mm_storeu_si128((__m128i *)(s + 5 * p), q5); + + f_lo = filter_add2_sub2(&f_lo, &q7_lo, &q6_lo, &p1_lo, &q5_lo); + f_hi = filter_add2_sub2(&f_hi, &q7_hi, &q6_hi, &p1_hi, &q5_hi); + q6 = filter16_mask(&flat2, &q6, &f_lo, &f_hi); + _mm_storeu_si128((__m128i *)(s + 6 * p), q6); + } + // wide flat + // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ + } +} + +void aom_lpf_horizontal_8_sse2(unsigned char *s, int p, + const unsigned char *_blimit, + const unsigned char *_limit, + const unsigned char *_thresh) { + DECLARE_ALIGNED(16, unsigned char, flat_op2[16]); + DECLARE_ALIGNED(16, unsigned char, flat_op1[16]); + DECLARE_ALIGNED(16, unsigned char, flat_op0[16]); + DECLARE_ALIGNED(16, unsigned char, flat_oq2[16]); + DECLARE_ALIGNED(16, unsigned char, flat_oq1[16]); + DECLARE_ALIGNED(16, unsigned char, flat_oq0[16]); + const __m128i zero = _mm_set1_epi16(0); + const __m128i blimit = _mm_load_si128((const __m128i *)_blimit); + const __m128i limit = _mm_load_si128((const __m128i *)_limit); + const __m128i thresh = _mm_load_si128((const __m128i *)_thresh); + __m128i mask, hev, flat; + __m128i p3, p2, p1, p0, q0, q1, q2, q3; + __m128i q3p3, q2p2, q1p1, q0p0, p1q1, p0q0; + + q3p3 = _mm_unpacklo_epi64(_mm_loadl_epi64((__m128i *)(s - 4 * p)), + _mm_loadl_epi64((__m128i *)(s + 3 * p))); + q2p2 = _mm_unpacklo_epi64(_mm_loadl_epi64((__m128i *)(s - 3 * p)), + _mm_loadl_epi64((__m128i *)(s + 2 * p))); + q1p1 = _mm_unpacklo_epi64(_mm_loadl_epi64((__m128i *)(s - 2 * p)), + _mm_loadl_epi64((__m128i *)(s + 1 * p))); + q0p0 = _mm_unpacklo_epi64(_mm_loadl_epi64((__m128i *)(s - 1 * p)), + _mm_loadl_epi64((__m128i *)(s - 0 * p))); + p1q1 = _mm_shuffle_epi32(q1p1, 78); + p0q0 = _mm_shuffle_epi32(q0p0, 78); + + { + // filter_mask and hev_mask + const __m128i one = _mm_set1_epi8(1); + const __m128i fe = _mm_set1_epi8(0xfe); + const __m128i ff = _mm_cmpeq_epi8(fe, fe); + __m128i abs_p1q1, abs_p0q0, abs_q1q0, abs_p1p0, work; + abs_p1p0 = abs_diff(q1p1, q0p0); + abs_q1q0 = _mm_srli_si128(abs_p1p0, 8); + + abs_p0q0 = abs_diff(q0p0, p0q0); + abs_p1q1 = abs_diff(q1p1, p1q1); + flat = _mm_max_epu8(abs_p1p0, abs_q1q0); + hev = _mm_subs_epu8(flat, thresh); + hev = _mm_xor_si128(_mm_cmpeq_epi8(hev, zero), ff); + + abs_p0q0 = _mm_adds_epu8(abs_p0q0, abs_p0q0); + abs_p1q1 = _mm_srli_epi16(_mm_and_si128(abs_p1q1, fe), 1); + mask = _mm_subs_epu8(_mm_adds_epu8(abs_p0q0, abs_p1q1), blimit); + mask = _mm_xor_si128(_mm_cmpeq_epi8(mask, zero), ff); + // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 > blimit) * -1; + mask = _mm_max_epu8(abs_p1p0, mask); + // mask |= (abs(p1 - p0) > limit) * -1; + // mask |= (abs(q1 - q0) > limit) * -1; + + work = _mm_max_epu8(abs_diff(q2p2, q1p1), abs_diff(q3p3, q2p2)); + mask = _mm_max_epu8(work, mask); + mask = _mm_max_epu8(mask, _mm_srli_si128(mask, 8)); + mask = _mm_subs_epu8(mask, limit); + mask = _mm_cmpeq_epi8(mask, zero); + + // flat_mask4 + + flat = _mm_max_epu8(abs_diff(q2p2, q0p0), abs_diff(q3p3, q0p0)); + flat = _mm_max_epu8(abs_p1p0, flat); + flat = _mm_max_epu8(flat, _mm_srli_si128(flat, 8)); + flat = _mm_subs_epu8(flat, one); + flat = _mm_cmpeq_epi8(flat, zero); + flat = _mm_and_si128(flat, mask); + } + + { + const __m128i four = _mm_set1_epi16(4); + unsigned char *src = s; + { + __m128i workp_a, workp_b, workp_shft; + p3 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 4 * p)), zero); + p2 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 3 * p)), zero); + p1 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 2 * p)), zero); + p0 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 1 * p)), zero); + q0 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 0 * p)), zero); + q1 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src + 1 * p)), zero); + q2 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src + 2 * p)), zero); + q3 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src + 3 * p)), zero); + + workp_a = _mm_add_epi16(_mm_add_epi16(p3, p3), _mm_add_epi16(p2, p1)); + workp_a = _mm_add_epi16(_mm_add_epi16(workp_a, four), p0); + workp_b = _mm_add_epi16(_mm_add_epi16(q0, p2), p3); + workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3); + _mm_storel_epi64((__m128i *)&flat_op2[0], + _mm_packus_epi16(workp_shft, workp_shft)); + + workp_b = _mm_add_epi16(_mm_add_epi16(q0, q1), p1); + workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3); + _mm_storel_epi64((__m128i *)&flat_op1[0], + _mm_packus_epi16(workp_shft, workp_shft)); + + workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p3), q2); + workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, p1), p0); + workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3); + _mm_storel_epi64((__m128i *)&flat_op0[0], + _mm_packus_epi16(workp_shft, workp_shft)); + + workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p3), q3); + workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, p0), q0); + workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3); + _mm_storel_epi64((__m128i *)&flat_oq0[0], + _mm_packus_epi16(workp_shft, workp_shft)); + + workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p2), q3); + workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, q0), q1); + workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3); + _mm_storel_epi64((__m128i *)&flat_oq1[0], + _mm_packus_epi16(workp_shft, workp_shft)); + + workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p1), q3); + workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, q1), q2); + workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3); + _mm_storel_epi64((__m128i *)&flat_oq2[0], + _mm_packus_epi16(workp_shft, workp_shft)); + } + } + // lp filter + { + const __m128i t4 = _mm_set1_epi8(4); + const __m128i t3 = _mm_set1_epi8(3); + const __m128i t80 = _mm_set1_epi8(0x80); + const __m128i t1 = _mm_set1_epi8(0x1); + const __m128i ps1 = + _mm_xor_si128(_mm_loadl_epi64((__m128i *)(s - 2 * p)), t80); + const __m128i ps0 = + _mm_xor_si128(_mm_loadl_epi64((__m128i *)(s - 1 * p)), t80); + const __m128i qs0 = + _mm_xor_si128(_mm_loadl_epi64((__m128i *)(s + 0 * p)), t80); + const __m128i qs1 = + _mm_xor_si128(_mm_loadl_epi64((__m128i *)(s + 1 * p)), t80); + __m128i filt; + __m128i work_a; + __m128i filter1, filter2; + + filt = _mm_and_si128(_mm_subs_epi8(ps1, qs1), hev); + work_a = _mm_subs_epi8(qs0, ps0); + filt = _mm_adds_epi8(filt, work_a); + filt = _mm_adds_epi8(filt, work_a); + filt = _mm_adds_epi8(filt, work_a); + // (aom_filter + 3 * (qs0 - ps0)) & mask + filt = _mm_and_si128(filt, mask); + + filter1 = _mm_adds_epi8(filt, t4); + filter2 = _mm_adds_epi8(filt, t3); + + // Filter1 >> 3 + filter1 = _mm_unpacklo_epi8(zero, filter1); + filter1 = _mm_srai_epi16(filter1, 11); + filter1 = _mm_packs_epi16(filter1, filter1); + + // Filter2 >> 3 + filter2 = _mm_unpacklo_epi8(zero, filter2); + filter2 = _mm_srai_epi16(filter2, 11); + filter2 = _mm_packs_epi16(filter2, zero); + + // filt >> 1 + filt = _mm_adds_epi8(filter1, t1); + filt = _mm_unpacklo_epi8(zero, filt); + filt = _mm_srai_epi16(filt, 9); + filt = _mm_packs_epi16(filt, zero); + + filt = _mm_andnot_si128(hev, filt); + + work_a = _mm_xor_si128(_mm_subs_epi8(qs0, filter1), t80); + q0 = _mm_loadl_epi64((__m128i *)flat_oq0); + work_a = _mm_andnot_si128(flat, work_a); + q0 = _mm_and_si128(flat, q0); + q0 = _mm_or_si128(work_a, q0); + + work_a = _mm_xor_si128(_mm_subs_epi8(qs1, filt), t80); + q1 = _mm_loadl_epi64((__m128i *)flat_oq1); + work_a = _mm_andnot_si128(flat, work_a); + q1 = _mm_and_si128(flat, q1); + q1 = _mm_or_si128(work_a, q1); + + work_a = _mm_loadu_si128((__m128i *)(s + 2 * p)); + q2 = _mm_loadl_epi64((__m128i *)flat_oq2); + work_a = _mm_andnot_si128(flat, work_a); + q2 = _mm_and_si128(flat, q2); + q2 = _mm_or_si128(work_a, q2); + + work_a = _mm_xor_si128(_mm_adds_epi8(ps0, filter2), t80); + p0 = _mm_loadl_epi64((__m128i *)flat_op0); + work_a = _mm_andnot_si128(flat, work_a); + p0 = _mm_and_si128(flat, p0); + p0 = _mm_or_si128(work_a, p0); + + work_a = _mm_xor_si128(_mm_adds_epi8(ps1, filt), t80); + p1 = _mm_loadl_epi64((__m128i *)flat_op1); + work_a = _mm_andnot_si128(flat, work_a); + p1 = _mm_and_si128(flat, p1); + p1 = _mm_or_si128(work_a, p1); + + work_a = _mm_loadu_si128((__m128i *)(s - 3 * p)); + p2 = _mm_loadl_epi64((__m128i *)flat_op2); + work_a = _mm_andnot_si128(flat, work_a); + p2 = _mm_and_si128(flat, p2); + p2 = _mm_or_si128(work_a, p2); + + _mm_storel_epi64((__m128i *)(s - 3 * p), p2); + _mm_storel_epi64((__m128i *)(s - 2 * p), p1); + _mm_storel_epi64((__m128i *)(s - 1 * p), p0); + _mm_storel_epi64((__m128i *)(s + 0 * p), q0); + _mm_storel_epi64((__m128i *)(s + 1 * p), q1); + _mm_storel_epi64((__m128i *)(s + 2 * p), q2); + } +} + +void aom_lpf_horizontal_8_dual_sse2(uint8_t *s, int p, const uint8_t *_blimit0, + const uint8_t *_limit0, + const uint8_t *_thresh0, + const uint8_t *_blimit1, + const uint8_t *_limit1, + const uint8_t *_thresh1) { + DECLARE_ALIGNED(16, unsigned char, flat_op2[16]); + DECLARE_ALIGNED(16, unsigned char, flat_op1[16]); + DECLARE_ALIGNED(16, unsigned char, flat_op0[16]); + DECLARE_ALIGNED(16, unsigned char, flat_oq2[16]); + DECLARE_ALIGNED(16, unsigned char, flat_oq1[16]); + DECLARE_ALIGNED(16, unsigned char, flat_oq0[16]); + const __m128i zero = _mm_set1_epi16(0); + const __m128i blimit = + _mm_unpacklo_epi64(_mm_load_si128((const __m128i *)_blimit0), + _mm_load_si128((const __m128i *)_blimit1)); + const __m128i limit = + _mm_unpacklo_epi64(_mm_load_si128((const __m128i *)_limit0), + _mm_load_si128((const __m128i *)_limit1)); + const __m128i thresh = + _mm_unpacklo_epi64(_mm_load_si128((const __m128i *)_thresh0), + _mm_load_si128((const __m128i *)_thresh1)); + + __m128i mask, hev, flat; + __m128i p3, p2, p1, p0, q0, q1, q2, q3; + + p3 = _mm_loadu_si128((__m128i *)(s - 4 * p)); + p2 = _mm_loadu_si128((__m128i *)(s - 3 * p)); + p1 = _mm_loadu_si128((__m128i *)(s - 2 * p)); + p0 = _mm_loadu_si128((__m128i *)(s - 1 * p)); + q0 = _mm_loadu_si128((__m128i *)(s - 0 * p)); + q1 = _mm_loadu_si128((__m128i *)(s + 1 * p)); + q2 = _mm_loadu_si128((__m128i *)(s + 2 * p)); + q3 = _mm_loadu_si128((__m128i *)(s + 3 * p)); + { + const __m128i abs_p1p0 = + _mm_or_si128(_mm_subs_epu8(p1, p0), _mm_subs_epu8(p0, p1)); + const __m128i abs_q1q0 = + _mm_or_si128(_mm_subs_epu8(q1, q0), _mm_subs_epu8(q0, q1)); + const __m128i one = _mm_set1_epi8(1); + const __m128i fe = _mm_set1_epi8(0xfe); + const __m128i ff = _mm_cmpeq_epi8(abs_p1p0, abs_p1p0); + __m128i abs_p0q0 = + _mm_or_si128(_mm_subs_epu8(p0, q0), _mm_subs_epu8(q0, p0)); + __m128i abs_p1q1 = + _mm_or_si128(_mm_subs_epu8(p1, q1), _mm_subs_epu8(q1, p1)); + __m128i work; + + // filter_mask and hev_mask + flat = _mm_max_epu8(abs_p1p0, abs_q1q0); + hev = _mm_subs_epu8(flat, thresh); + hev = _mm_xor_si128(_mm_cmpeq_epi8(hev, zero), ff); + + abs_p0q0 = _mm_adds_epu8(abs_p0q0, abs_p0q0); + abs_p1q1 = _mm_srli_epi16(_mm_and_si128(abs_p1q1, fe), 1); + mask = _mm_subs_epu8(_mm_adds_epu8(abs_p0q0, abs_p1q1), blimit); + mask = _mm_xor_si128(_mm_cmpeq_epi8(mask, zero), ff); + // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 > blimit) * -1; + mask = _mm_max_epu8(flat, mask); + // mask |= (abs(p1 - p0) > limit) * -1; + // mask |= (abs(q1 - q0) > limit) * -1; + work = _mm_max_epu8( + _mm_or_si128(_mm_subs_epu8(p2, p1), _mm_subs_epu8(p1, p2)), + _mm_or_si128(_mm_subs_epu8(p3, p2), _mm_subs_epu8(p2, p3))); + mask = _mm_max_epu8(work, mask); + work = _mm_max_epu8( + _mm_or_si128(_mm_subs_epu8(q2, q1), _mm_subs_epu8(q1, q2)), + _mm_or_si128(_mm_subs_epu8(q3, q2), _mm_subs_epu8(q2, q3))); + mask = _mm_max_epu8(work, mask); + mask = _mm_subs_epu8(mask, limit); + mask = _mm_cmpeq_epi8(mask, zero); + + // flat_mask4 + work = _mm_max_epu8( + _mm_or_si128(_mm_subs_epu8(p2, p0), _mm_subs_epu8(p0, p2)), + _mm_or_si128(_mm_subs_epu8(q2, q0), _mm_subs_epu8(q0, q2))); + flat = _mm_max_epu8(work, flat); + work = _mm_max_epu8( + _mm_or_si128(_mm_subs_epu8(p3, p0), _mm_subs_epu8(p0, p3)), + _mm_or_si128(_mm_subs_epu8(q3, q0), _mm_subs_epu8(q0, q3))); + flat = _mm_max_epu8(work, flat); + flat = _mm_subs_epu8(flat, one); + flat = _mm_cmpeq_epi8(flat, zero); + flat = _mm_and_si128(flat, mask); + } + { + const __m128i four = _mm_set1_epi16(4); + unsigned char *src = s; + int i = 0; + + do { + __m128i workp_a, workp_b, workp_shft; + p3 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 4 * p)), zero); + p2 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 3 * p)), zero); + p1 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 2 * p)), zero); + p0 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 1 * p)), zero); + q0 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 0 * p)), zero); + q1 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src + 1 * p)), zero); + q2 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src + 2 * p)), zero); + q3 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src + 3 * p)), zero); + + workp_a = _mm_add_epi16(_mm_add_epi16(p3, p3), _mm_add_epi16(p2, p1)); + workp_a = _mm_add_epi16(_mm_add_epi16(workp_a, four), p0); + workp_b = _mm_add_epi16(_mm_add_epi16(q0, p2), p3); + workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3); + _mm_storel_epi64((__m128i *)&flat_op2[i * 8], + _mm_packus_epi16(workp_shft, workp_shft)); + + workp_b = _mm_add_epi16(_mm_add_epi16(q0, q1), p1); + workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3); + _mm_storel_epi64((__m128i *)&flat_op1[i * 8], + _mm_packus_epi16(workp_shft, workp_shft)); + + workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p3), q2); + workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, p1), p0); + workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3); + _mm_storel_epi64((__m128i *)&flat_op0[i * 8], + _mm_packus_epi16(workp_shft, workp_shft)); + + workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p3), q3); + workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, p0), q0); + workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3); + _mm_storel_epi64((__m128i *)&flat_oq0[i * 8], + _mm_packus_epi16(workp_shft, workp_shft)); + + workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p2), q3); + workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, q0), q1); + workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3); + _mm_storel_epi64((__m128i *)&flat_oq1[i * 8], + _mm_packus_epi16(workp_shft, workp_shft)); + + workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p1), q3); + workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, q1), q2); + workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3); + _mm_storel_epi64((__m128i *)&flat_oq2[i * 8], + _mm_packus_epi16(workp_shft, workp_shft)); + + src += 8; + } while (++i < 2); + } + // lp filter + { + const __m128i t4 = _mm_set1_epi8(4); + const __m128i t3 = _mm_set1_epi8(3); + const __m128i t80 = _mm_set1_epi8(0x80); + const __m128i te0 = _mm_set1_epi8(0xe0); + const __m128i t1f = _mm_set1_epi8(0x1f); + const __m128i t1 = _mm_set1_epi8(0x1); + const __m128i t7f = _mm_set1_epi8(0x7f); + + const __m128i ps1 = + _mm_xor_si128(_mm_loadu_si128((__m128i *)(s - 2 * p)), t80); + const __m128i ps0 = + _mm_xor_si128(_mm_loadu_si128((__m128i *)(s - 1 * p)), t80); + const __m128i qs0 = + _mm_xor_si128(_mm_loadu_si128((__m128i *)(s + 0 * p)), t80); + const __m128i qs1 = + _mm_xor_si128(_mm_loadu_si128((__m128i *)(s + 1 * p)), t80); + __m128i filt; + __m128i work_a; + __m128i filter1, filter2; + + filt = _mm_and_si128(_mm_subs_epi8(ps1, qs1), hev); + work_a = _mm_subs_epi8(qs0, ps0); + filt = _mm_adds_epi8(filt, work_a); + filt = _mm_adds_epi8(filt, work_a); + filt = _mm_adds_epi8(filt, work_a); + // (aom_filter + 3 * (qs0 - ps0)) & mask + filt = _mm_and_si128(filt, mask); + + filter1 = _mm_adds_epi8(filt, t4); + filter2 = _mm_adds_epi8(filt, t3); + + // Filter1 >> 3 + work_a = _mm_cmpgt_epi8(zero, filter1); + filter1 = _mm_srli_epi16(filter1, 3); + work_a = _mm_and_si128(work_a, te0); + filter1 = _mm_and_si128(filter1, t1f); + filter1 = _mm_or_si128(filter1, work_a); + + // Filter2 >> 3 + work_a = _mm_cmpgt_epi8(zero, filter2); + filter2 = _mm_srli_epi16(filter2, 3); + work_a = _mm_and_si128(work_a, te0); + filter2 = _mm_and_si128(filter2, t1f); + filter2 = _mm_or_si128(filter2, work_a); + + // filt >> 1 + filt = _mm_adds_epi8(filter1, t1); + work_a = _mm_cmpgt_epi8(zero, filt); + filt = _mm_srli_epi16(filt, 1); + work_a = _mm_and_si128(work_a, t80); + filt = _mm_and_si128(filt, t7f); + filt = _mm_or_si128(filt, work_a); + + filt = _mm_andnot_si128(hev, filt); + + work_a = _mm_xor_si128(_mm_subs_epi8(qs0, filter1), t80); + q0 = _mm_load_si128((__m128i *)flat_oq0); + work_a = _mm_andnot_si128(flat, work_a); + q0 = _mm_and_si128(flat, q0); + q0 = _mm_or_si128(work_a, q0); + + work_a = _mm_xor_si128(_mm_subs_epi8(qs1, filt), t80); + q1 = _mm_load_si128((__m128i *)flat_oq1); + work_a = _mm_andnot_si128(flat, work_a); + q1 = _mm_and_si128(flat, q1); + q1 = _mm_or_si128(work_a, q1); + + work_a = _mm_loadu_si128((__m128i *)(s + 2 * p)); + q2 = _mm_load_si128((__m128i *)flat_oq2); + work_a = _mm_andnot_si128(flat, work_a); + q2 = _mm_and_si128(flat, q2); + q2 = _mm_or_si128(work_a, q2); + + work_a = _mm_xor_si128(_mm_adds_epi8(ps0, filter2), t80); + p0 = _mm_load_si128((__m128i *)flat_op0); + work_a = _mm_andnot_si128(flat, work_a); + p0 = _mm_and_si128(flat, p0); + p0 = _mm_or_si128(work_a, p0); + + work_a = _mm_xor_si128(_mm_adds_epi8(ps1, filt), t80); + p1 = _mm_load_si128((__m128i *)flat_op1); + work_a = _mm_andnot_si128(flat, work_a); + p1 = _mm_and_si128(flat, p1); + p1 = _mm_or_si128(work_a, p1); + + work_a = _mm_loadu_si128((__m128i *)(s - 3 * p)); + p2 = _mm_load_si128((__m128i *)flat_op2); + work_a = _mm_andnot_si128(flat, work_a); + p2 = _mm_and_si128(flat, p2); + p2 = _mm_or_si128(work_a, p2); + + _mm_storeu_si128((__m128i *)(s - 3 * p), p2); + _mm_storeu_si128((__m128i *)(s - 2 * p), p1); + _mm_storeu_si128((__m128i *)(s - 1 * p), p0); + _mm_storeu_si128((__m128i *)(s + 0 * p), q0); + _mm_storeu_si128((__m128i *)(s + 1 * p), q1); + _mm_storeu_si128((__m128i *)(s + 2 * p), q2); + } +} + +void aom_lpf_horizontal_4_dual_sse2(unsigned char *s, int p, + const unsigned char *_blimit0, + const unsigned char *_limit0, + const unsigned char *_thresh0, + const unsigned char *_blimit1, + const unsigned char *_limit1, + const unsigned char *_thresh1) { + const __m128i blimit = + _mm_unpacklo_epi64(_mm_load_si128((const __m128i *)_blimit0), + _mm_load_si128((const __m128i *)_blimit1)); + const __m128i limit = + _mm_unpacklo_epi64(_mm_load_si128((const __m128i *)_limit0), + _mm_load_si128((const __m128i *)_limit1)); + const __m128i thresh = + _mm_unpacklo_epi64(_mm_load_si128((const __m128i *)_thresh0), + _mm_load_si128((const __m128i *)_thresh1)); + const __m128i zero = _mm_set1_epi16(0); +#if !CONFIG_PARALLEL_DEBLOCKING + __m128i p3, p2, q2, q3; +#endif // !CONFIG_PARALLEL_DEBLOCKING + __m128i p1, p0, q0, q1; + __m128i mask, hev, flat; +#if !CONFIG_PARALLEL_DEBLOCKING + p3 = _mm_loadu_si128((__m128i *)(s - 4 * p)); + p2 = _mm_loadu_si128((__m128i *)(s - 3 * p)); +#endif // !CONFIG_PARALLEL_DEBLOCKING + p1 = _mm_loadu_si128((__m128i *)(s - 2 * p)); + p0 = _mm_loadu_si128((__m128i *)(s - 1 * p)); + q0 = _mm_loadu_si128((__m128i *)(s - 0 * p)); + q1 = _mm_loadu_si128((__m128i *)(s + 1 * p)); +#if !CONFIG_PARALLEL_DEBLOCKING + q2 = _mm_loadu_si128((__m128i *)(s + 2 * p)); + q3 = _mm_loadu_si128((__m128i *)(s + 3 * p)); +#endif // !CONFIG_PARALLEL_DEBLOCKING + // filter_mask and hev_mask + { + const __m128i abs_p1p0 = + _mm_or_si128(_mm_subs_epu8(p1, p0), _mm_subs_epu8(p0, p1)); + const __m128i abs_q1q0 = + _mm_or_si128(_mm_subs_epu8(q1, q0), _mm_subs_epu8(q0, q1)); + const __m128i fe = _mm_set1_epi8(0xfe); + const __m128i ff = _mm_cmpeq_epi8(abs_p1p0, abs_p1p0); + __m128i abs_p0q0 = + _mm_or_si128(_mm_subs_epu8(p0, q0), _mm_subs_epu8(q0, p0)); + __m128i abs_p1q1 = + _mm_or_si128(_mm_subs_epu8(p1, q1), _mm_subs_epu8(q1, p1)); +#if !CONFIG_PARALLEL_DEBLOCKING + __m128i work; +#endif // !CONFIG_PARALLEL_DEBLOCKING + flat = _mm_max_epu8(abs_p1p0, abs_q1q0); + hev = _mm_subs_epu8(flat, thresh); + hev = _mm_xor_si128(_mm_cmpeq_epi8(hev, zero), ff); + + abs_p0q0 = _mm_adds_epu8(abs_p0q0, abs_p0q0); + abs_p1q1 = _mm_srli_epi16(_mm_and_si128(abs_p1q1, fe), 1); + mask = _mm_subs_epu8(_mm_adds_epu8(abs_p0q0, abs_p1q1), blimit); + mask = _mm_xor_si128(_mm_cmpeq_epi8(mask, zero), ff); + // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 > blimit) * -1; + mask = _mm_max_epu8(flat, mask); +#if !CONFIG_PARALLEL_DEBLOCKING + // mask |= (abs(p1 - p0) > limit) * -1; + // mask |= (abs(q1 - q0) > limit) * -1; + work = _mm_max_epu8( + _mm_or_si128(_mm_subs_epu8(p2, p1), _mm_subs_epu8(p1, p2)), + _mm_or_si128(_mm_subs_epu8(p3, p2), _mm_subs_epu8(p2, p3))); + mask = _mm_max_epu8(work, mask); + work = _mm_max_epu8( + _mm_or_si128(_mm_subs_epu8(q2, q1), _mm_subs_epu8(q1, q2)), + _mm_or_si128(_mm_subs_epu8(q3, q2), _mm_subs_epu8(q2, q3))); + mask = _mm_max_epu8(work, mask); +#endif // !CONFIG_PARALLEL_DEBLOCKING + mask = _mm_subs_epu8(mask, limit); + mask = _mm_cmpeq_epi8(mask, zero); + } + + // filter4 + { + const __m128i t4 = _mm_set1_epi8(4); + const __m128i t3 = _mm_set1_epi8(3); + const __m128i t80 = _mm_set1_epi8(0x80); + const __m128i te0 = _mm_set1_epi8(0xe0); + const __m128i t1f = _mm_set1_epi8(0x1f); + const __m128i t1 = _mm_set1_epi8(0x1); + const __m128i t7f = _mm_set1_epi8(0x7f); + + const __m128i ps1 = + _mm_xor_si128(_mm_loadu_si128((__m128i *)(s - 2 * p)), t80); + const __m128i ps0 = + _mm_xor_si128(_mm_loadu_si128((__m128i *)(s - 1 * p)), t80); + const __m128i qs0 = + _mm_xor_si128(_mm_loadu_si128((__m128i *)(s + 0 * p)), t80); + const __m128i qs1 = + _mm_xor_si128(_mm_loadu_si128((__m128i *)(s + 1 * p)), t80); + __m128i filt; + __m128i work_a; + __m128i filter1, filter2; + + filt = _mm_and_si128(_mm_subs_epi8(ps1, qs1), hev); + work_a = _mm_subs_epi8(qs0, ps0); + filt = _mm_adds_epi8(filt, work_a); + filt = _mm_adds_epi8(filt, work_a); + filt = _mm_adds_epi8(filt, work_a); + // (aom_filter + 3 * (qs0 - ps0)) & mask + filt = _mm_and_si128(filt, mask); + + filter1 = _mm_adds_epi8(filt, t4); + filter2 = _mm_adds_epi8(filt, t3); + + // Filter1 >> 3 + work_a = _mm_cmpgt_epi8(zero, filter1); + filter1 = _mm_srli_epi16(filter1, 3); + work_a = _mm_and_si128(work_a, te0); + filter1 = _mm_and_si128(filter1, t1f); + filter1 = _mm_or_si128(filter1, work_a); + + // Filter2 >> 3 + work_a = _mm_cmpgt_epi8(zero, filter2); + filter2 = _mm_srli_epi16(filter2, 3); + work_a = _mm_and_si128(work_a, te0); + filter2 = _mm_and_si128(filter2, t1f); + filter2 = _mm_or_si128(filter2, work_a); + + // filt >> 1 + filt = _mm_adds_epi8(filter1, t1); + work_a = _mm_cmpgt_epi8(zero, filt); + filt = _mm_srli_epi16(filt, 1); + work_a = _mm_and_si128(work_a, t80); + filt = _mm_and_si128(filt, t7f); + filt = _mm_or_si128(filt, work_a); + + filt = _mm_andnot_si128(hev, filt); + + q0 = _mm_xor_si128(_mm_subs_epi8(qs0, filter1), t80); + q1 = _mm_xor_si128(_mm_subs_epi8(qs1, filt), t80); + p0 = _mm_xor_si128(_mm_adds_epi8(ps0, filter2), t80); + p1 = _mm_xor_si128(_mm_adds_epi8(ps1, filt), t80); + + _mm_storeu_si128((__m128i *)(s - 2 * p), p1); + _mm_storeu_si128((__m128i *)(s - 1 * p), p0); + _mm_storeu_si128((__m128i *)(s + 0 * p), q0); + _mm_storeu_si128((__m128i *)(s + 1 * p), q1); + } +} + +static INLINE void transpose8x16(unsigned char *in0, unsigned char *in1, + int in_p, unsigned char *out, int out_p) { + __m128i x0, x1, x2, x3, x4, x5, x6, x7; + __m128i x8, x9, x10, x11, x12, x13, x14, x15; + + // 2-way interleave w/hoisting of unpacks + x0 = _mm_loadl_epi64((__m128i *)in0); // 1 + x1 = _mm_loadl_epi64((__m128i *)(in0 + in_p)); // 3 + x0 = _mm_unpacklo_epi8(x0, x1); // 1 + + x2 = _mm_loadl_epi64((__m128i *)(in0 + 2 * in_p)); // 5 + x3 = _mm_loadl_epi64((__m128i *)(in0 + 3 * in_p)); // 7 + x1 = _mm_unpacklo_epi8(x2, x3); // 2 + + x4 = _mm_loadl_epi64((__m128i *)(in0 + 4 * in_p)); // 9 + x5 = _mm_loadl_epi64((__m128i *)(in0 + 5 * in_p)); // 11 + x2 = _mm_unpacklo_epi8(x4, x5); // 3 + + x6 = _mm_loadl_epi64((__m128i *)(in0 + 6 * in_p)); // 13 + x7 = _mm_loadl_epi64((__m128i *)(in0 + 7 * in_p)); // 15 + x3 = _mm_unpacklo_epi8(x6, x7); // 4 + x4 = _mm_unpacklo_epi16(x0, x1); // 9 + + x8 = _mm_loadl_epi64((__m128i *)in1); // 2 + x9 = _mm_loadl_epi64((__m128i *)(in1 + in_p)); // 4 + x8 = _mm_unpacklo_epi8(x8, x9); // 5 + x5 = _mm_unpacklo_epi16(x2, x3); // 10 + + x10 = _mm_loadl_epi64((__m128i *)(in1 + 2 * in_p)); // 6 + x11 = _mm_loadl_epi64((__m128i *)(in1 + 3 * in_p)); // 8 + x9 = _mm_unpacklo_epi8(x10, x11); // 6 + + x12 = _mm_loadl_epi64((__m128i *)(in1 + 4 * in_p)); // 10 + x13 = _mm_loadl_epi64((__m128i *)(in1 + 5 * in_p)); // 12 + x10 = _mm_unpacklo_epi8(x12, x13); // 7 + x12 = _mm_unpacklo_epi16(x8, x9); // 11 + + x14 = _mm_loadl_epi64((__m128i *)(in1 + 6 * in_p)); // 14 + x15 = _mm_loadl_epi64((__m128i *)(in1 + 7 * in_p)); // 16 + x11 = _mm_unpacklo_epi8(x14, x15); // 8 + x13 = _mm_unpacklo_epi16(x10, x11); // 12 + + x6 = _mm_unpacklo_epi32(x4, x5); // 13 + x7 = _mm_unpackhi_epi32(x4, x5); // 14 + x14 = _mm_unpacklo_epi32(x12, x13); // 15 + x15 = _mm_unpackhi_epi32(x12, x13); // 16 + + // Store first 4-line result + _mm_storeu_si128((__m128i *)out, _mm_unpacklo_epi64(x6, x14)); + _mm_storeu_si128((__m128i *)(out + out_p), _mm_unpackhi_epi64(x6, x14)); + _mm_storeu_si128((__m128i *)(out + 2 * out_p), _mm_unpacklo_epi64(x7, x15)); + _mm_storeu_si128((__m128i *)(out + 3 * out_p), _mm_unpackhi_epi64(x7, x15)); + + x4 = _mm_unpackhi_epi16(x0, x1); + x5 = _mm_unpackhi_epi16(x2, x3); + x12 = _mm_unpackhi_epi16(x8, x9); + x13 = _mm_unpackhi_epi16(x10, x11); + + x6 = _mm_unpacklo_epi32(x4, x5); + x7 = _mm_unpackhi_epi32(x4, x5); + x14 = _mm_unpacklo_epi32(x12, x13); + x15 = _mm_unpackhi_epi32(x12, x13); + + // Store second 4-line result + _mm_storeu_si128((__m128i *)(out + 4 * out_p), _mm_unpacklo_epi64(x6, x14)); + _mm_storeu_si128((__m128i *)(out + 5 * out_p), _mm_unpackhi_epi64(x6, x14)); + _mm_storeu_si128((__m128i *)(out + 6 * out_p), _mm_unpacklo_epi64(x7, x15)); + _mm_storeu_si128((__m128i *)(out + 7 * out_p), _mm_unpackhi_epi64(x7, x15)); +} + +#if CONFIG_PARALLEL_DEBLOCKING +#define movq(p) _mm_loadl_epi64((const __m128i *)(p)) +#define punpcklbw(r0, r1) _mm_unpacklo_epi8(r0, r1) +#define punpcklwd(r0, r1) _mm_unpacklo_epi16(r0, r1) +#define punpckhwd(r0, r1) _mm_unpackhi_epi16(r0, r1) +#define movd(p, r) *((uint32_t *)(p)) = _mm_cvtsi128_si32(r) +#define pshufd(r, imm) _mm_shuffle_epi32(r, imm) +enum { ROTATE_DWORD_RIGHT = 0x39 }; +static INLINE void transpose16x4(uint8_t *pDst, const ptrdiff_t dstStride, + const uint8_t *pSrc, + const ptrdiff_t srcStride) { + for (uint32_t idx = 0; idx < 2; idx += 1) { + __m128i r0, r1, r2, r3; + // load data + r0 = movq(pSrc); + r1 = movq(pSrc + srcStride); + r2 = movq(pSrc + srcStride * 2); + r3 = movq(pSrc + srcStride * 3); + // transpose + r0 = punpcklbw(r0, r1); + r2 = punpcklbw(r2, r3); + r1 = punpckhwd(r0, r2); + r0 = punpcklwd(r0, r2); + // store data + movd(pDst, r0); + r0 = pshufd(r0, ROTATE_DWORD_RIGHT); + movd(pDst + dstStride, r0); + r0 = pshufd(r0, ROTATE_DWORD_RIGHT); + movd(pDst + dstStride * 2, r0); + r0 = pshufd(r0, ROTATE_DWORD_RIGHT); + movd(pDst + dstStride * 3, r0); + movd(pDst + dstStride * 4, r1); + r1 = pshufd(r1, ROTATE_DWORD_RIGHT); + movd(pDst + dstStride * 5, r1); + r1 = pshufd(r1, ROTATE_DWORD_RIGHT); + movd(pDst + dstStride * 6, r1); + r1 = pshufd(r1, ROTATE_DWORD_RIGHT); + movd(pDst + dstStride * 7, r1); + // advance the pointers + pDst += dstStride * 8; + pSrc += 8; + } +} + +#endif // CONFIG_PARALLEL_DEBLOCKING +static INLINE void transpose(unsigned char *src[], int in_p, + unsigned char *dst[], int out_p, + int num_8x8_to_transpose) { + int idx8x8 = 0; + __m128i x0, x1, x2, x3, x4, x5, x6, x7; + do { + unsigned char *in = src[idx8x8]; + unsigned char *out = dst[idx8x8]; + + x0 = + _mm_loadl_epi64((__m128i *)(in + 0 * in_p)); // 00 01 02 03 04 05 06 07 + x1 = + _mm_loadl_epi64((__m128i *)(in + 1 * in_p)); // 10 11 12 13 14 15 16 17 + // 00 10 01 11 02 12 03 13 04 14 05 15 06 16 07 17 + x0 = _mm_unpacklo_epi8(x0, x1); + + x2 = + _mm_loadl_epi64((__m128i *)(in + 2 * in_p)); // 20 21 22 23 24 25 26 27 + x3 = + _mm_loadl_epi64((__m128i *)(in + 3 * in_p)); // 30 31 32 33 34 35 36 37 + // 20 30 21 31 22 32 23 33 24 34 25 35 26 36 27 37 + x1 = _mm_unpacklo_epi8(x2, x3); + + x4 = + _mm_loadl_epi64((__m128i *)(in + 4 * in_p)); // 40 41 42 43 44 45 46 47 + x5 = + _mm_loadl_epi64((__m128i *)(in + 5 * in_p)); // 50 51 52 53 54 55 56 57 + // 40 50 41 51 42 52 43 53 44 54 45 55 46 56 47 57 + x2 = _mm_unpacklo_epi8(x4, x5); + + x6 = + _mm_loadl_epi64((__m128i *)(in + 6 * in_p)); // 60 61 62 63 64 65 66 67 + x7 = + _mm_loadl_epi64((__m128i *)(in + 7 * in_p)); // 70 71 72 73 74 75 76 77 + // 60 70 61 71 62 72 63 73 64 74 65 75 66 76 67 77 + x3 = _mm_unpacklo_epi8(x6, x7); + + // 00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33 + x4 = _mm_unpacklo_epi16(x0, x1); + // 40 50 60 70 41 51 61 71 42 52 62 72 43 53 63 73 + x5 = _mm_unpacklo_epi16(x2, x3); + // 00 10 20 30 40 50 60 70 01 11 21 31 41 51 61 71 + x6 = _mm_unpacklo_epi32(x4, x5); + _mm_storel_pd((double *)(out + 0 * out_p), + _mm_castsi128_pd(x6)); // 00 10 20 30 40 50 60 70 + _mm_storeh_pd((double *)(out + 1 * out_p), + _mm_castsi128_pd(x6)); // 01 11 21 31 41 51 61 71 + // 02 12 22 32 42 52 62 72 03 13 23 33 43 53 63 73 + x7 = _mm_unpackhi_epi32(x4, x5); + _mm_storel_pd((double *)(out + 2 * out_p), + _mm_castsi128_pd(x7)); // 02 12 22 32 42 52 62 72 + _mm_storeh_pd((double *)(out + 3 * out_p), + _mm_castsi128_pd(x7)); // 03 13 23 33 43 53 63 73 + + // 04 14 24 34 05 15 25 35 06 16 26 36 07 17 27 37 + x4 = _mm_unpackhi_epi16(x0, x1); + // 44 54 64 74 45 55 65 75 46 56 66 76 47 57 67 77 + x5 = _mm_unpackhi_epi16(x2, x3); + // 04 14 24 34 44 54 64 74 05 15 25 35 45 55 65 75 + x6 = _mm_unpacklo_epi32(x4, x5); + _mm_storel_pd((double *)(out + 4 * out_p), + _mm_castsi128_pd(x6)); // 04 14 24 34 44 54 64 74 + _mm_storeh_pd((double *)(out + 5 * out_p), + _mm_castsi128_pd(x6)); // 05 15 25 35 45 55 65 75 + // 06 16 26 36 46 56 66 76 07 17 27 37 47 57 67 77 + x7 = _mm_unpackhi_epi32(x4, x5); + + _mm_storel_pd((double *)(out + 6 * out_p), + _mm_castsi128_pd(x7)); // 06 16 26 36 46 56 66 76 + _mm_storeh_pd((double *)(out + 7 * out_p), + _mm_castsi128_pd(x7)); // 07 17 27 37 47 57 67 77 + } while (++idx8x8 < num_8x8_to_transpose); +} + +void aom_lpf_vertical_4_dual_sse2(uint8_t *s, int p, const uint8_t *blimit0, + const uint8_t *limit0, const uint8_t *thresh0, + const uint8_t *blimit1, const uint8_t *limit1, + const uint8_t *thresh1) { + DECLARE_ALIGNED(16, unsigned char, t_dst[16 * 8]); +#if !CONFIG_PARALLEL_DEBLOCKING + unsigned char *src[2]; + unsigned char *dst[2]; +#endif // !CONFIG_PARALLEL_DEBLOCKING + // Transpose 8x16 + transpose8x16(s - 4, s - 4 + p * 8, p, t_dst, 16); + + // Loop filtering + aom_lpf_horizontal_4_dual_sse2(t_dst + 4 * 16, 16, blimit0, limit0, thresh0, + blimit1, limit1, thresh1); +#if !CONFIG_PARALLEL_DEBLOCKING + src[0] = t_dst; + src[1] = t_dst + 8; + dst[0] = s - 4; + dst[1] = s - 4 + p * 8; + + // Transpose back + transpose(src, 16, dst, p, 2); +#else // CONFIG_PARALLEL_DEBLOCKING + transpose16x4(s - 2, p, t_dst + 16 * 2, 16); +#endif // !CONFIG_PARALLEL_DEBLOCKING +} + +void aom_lpf_vertical_8_sse2(unsigned char *s, int p, + const unsigned char *blimit, + const unsigned char *limit, + const unsigned char *thresh) { + DECLARE_ALIGNED(8, unsigned char, t_dst[8 * 8]); + unsigned char *src[1]; + unsigned char *dst[1]; + + // Transpose 8x8 + src[0] = s - 4; + dst[0] = t_dst; + + transpose(src, p, dst, 8, 1); + + // Loop filtering + aom_lpf_horizontal_8_sse2(t_dst + 4 * 8, 8, blimit, limit, thresh); + + src[0] = t_dst; + dst[0] = s - 4; + + // Transpose back + transpose(src, 8, dst, p, 1); +} + +void aom_lpf_vertical_8_dual_sse2(uint8_t *s, int p, const uint8_t *blimit0, + const uint8_t *limit0, const uint8_t *thresh0, + const uint8_t *blimit1, const uint8_t *limit1, + const uint8_t *thresh1) { + DECLARE_ALIGNED(16, unsigned char, t_dst[16 * 8]); + unsigned char *src[2]; + unsigned char *dst[2]; + + // Transpose 8x16 + transpose8x16(s - 4, s - 4 + p * 8, p, t_dst, 16); + + // Loop filtering + aom_lpf_horizontal_8_dual_sse2(t_dst + 4 * 16, 16, blimit0, limit0, thresh0, + blimit1, limit1, thresh1); + src[0] = t_dst; + src[1] = t_dst + 8; + + dst[0] = s - 4; + dst[1] = s - 4 + p * 8; + + // Transpose back + transpose(src, 16, dst, p, 2); +} + +void aom_lpf_vertical_16_sse2(unsigned char *s, int p, + const unsigned char *blimit, + const unsigned char *limit, + const unsigned char *thresh) { + DECLARE_ALIGNED(8, unsigned char, t_dst[8 * 16]); + unsigned char *src[2]; + unsigned char *dst[2]; + + src[0] = s - 8; + src[1] = s; + dst[0] = t_dst; + dst[1] = t_dst + 8 * 8; + + // Transpose 16x8 + transpose(src, p, dst, 8, 2); + + // Loop filtering + aom_lpf_horizontal_edge_8_sse2(t_dst + 8 * 8, 8, blimit, limit, thresh); + + src[0] = t_dst; + src[1] = t_dst + 8 * 8; + dst[0] = s - 8; + dst[1] = s; + + // Transpose back + transpose(src, 8, dst, p, 2); +} + +void aom_lpf_vertical_16_dual_sse2(unsigned char *s, int p, + const uint8_t *blimit, const uint8_t *limit, + const uint8_t *thresh) { + DECLARE_ALIGNED(16, unsigned char, t_dst[256]); + + // Transpose 16x16 + transpose8x16(s - 8, s - 8 + 8 * p, p, t_dst, 16); + transpose8x16(s, s + 8 * p, p, t_dst + 8 * 16, 16); + + // Loop filtering + aom_lpf_horizontal_edge_16_sse2(t_dst + 8 * 16, 16, blimit, limit, thresh); + + // Transpose back + transpose8x16(t_dst, t_dst + 8 * 16, 16, s - 8, p); + transpose8x16(t_dst + 8, t_dst + 8 + 8 * 16, 16, s - 8 + 8 * p, p); +} diff --git a/third_party/aom/aom_dsp/x86/masked_sad_intrin_ssse3.c b/third_party/aom/aom_dsp/x86/masked_sad_intrin_ssse3.c new file mode 100644 index 000000000..5166e9e0a --- /dev/null +++ b/third_party/aom/aom_dsp/x86/masked_sad_intrin_ssse3.c @@ -0,0 +1,334 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include <stdlib.h> +#include <emmintrin.h> +#include <tmmintrin.h> + +#include "aom_ports/mem.h" +#include "./aom_config.h" +#include "aom/aom_integer.h" + +static INLINE __m128i width8_load_2rows(const uint8_t *ptr, int stride) { + __m128i temp1 = _mm_loadl_epi64((const __m128i *)ptr); + __m128i temp2 = _mm_loadl_epi64((const __m128i *)(ptr + stride)); + return _mm_unpacklo_epi64(temp1, temp2); +} + +static INLINE __m128i width4_load_4rows(const uint8_t *ptr, int stride) { + __m128i temp1 = _mm_cvtsi32_si128(*(const uint32_t *)ptr); + __m128i temp2 = _mm_cvtsi32_si128(*(const uint32_t *)(ptr + stride)); + __m128i temp3 = _mm_unpacklo_epi32(temp1, temp2); + temp1 = _mm_cvtsi32_si128(*(const uint32_t *)(ptr + stride * 2)); + temp2 = _mm_cvtsi32_si128(*(const uint32_t *)(ptr + stride * 3)); + temp1 = _mm_unpacklo_epi32(temp1, temp2); + return _mm_unpacklo_epi64(temp3, temp1); +} + +static INLINE unsigned int masked_sad_ssse3(const uint8_t *a_ptr, int a_stride, + const uint8_t *b_ptr, int b_stride, + const uint8_t *m_ptr, int m_stride, + int width, int height); + +static INLINE unsigned int masked_sad8xh_ssse3( + const uint8_t *a_ptr, int a_stride, const uint8_t *b_ptr, int b_stride, + const uint8_t *m_ptr, int m_stride, int height); + +static INLINE unsigned int masked_sad4xh_ssse3( + const uint8_t *a_ptr, int a_stride, const uint8_t *b_ptr, int b_stride, + const uint8_t *m_ptr, int m_stride, int height); + +#define MASKSADMXN_SSSE3(m, n) \ + unsigned int aom_masked_sad##m##x##n##_ssse3( \ + const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, \ + const uint8_t *msk, int msk_stride) { \ + return masked_sad_ssse3(src, src_stride, ref, ref_stride, msk, msk_stride, \ + m, n); \ + } + +#if CONFIG_EXT_PARTITION +MASKSADMXN_SSSE3(128, 128) +MASKSADMXN_SSSE3(128, 64) +MASKSADMXN_SSSE3(64, 128) +#endif // CONFIG_EXT_PARTITION +MASKSADMXN_SSSE3(64, 64) +MASKSADMXN_SSSE3(64, 32) +MASKSADMXN_SSSE3(32, 64) +MASKSADMXN_SSSE3(32, 32) +MASKSADMXN_SSSE3(32, 16) +MASKSADMXN_SSSE3(16, 32) +MASKSADMXN_SSSE3(16, 16) +MASKSADMXN_SSSE3(16, 8) + +#define MASKSAD8XN_SSSE3(n) \ + unsigned int aom_masked_sad8x##n##_ssse3( \ + const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, \ + const uint8_t *msk, int msk_stride) { \ + return masked_sad8xh_ssse3(src, src_stride, ref, ref_stride, msk, \ + msk_stride, n); \ + } + +MASKSAD8XN_SSSE3(16) +MASKSAD8XN_SSSE3(8) +MASKSAD8XN_SSSE3(4) + +#define MASKSAD4XN_SSSE3(n) \ + unsigned int aom_masked_sad4x##n##_ssse3( \ + const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, \ + const uint8_t *msk, int msk_stride) { \ + return masked_sad4xh_ssse3(src, src_stride, ref, ref_stride, msk, \ + msk_stride, n); \ + } + +MASKSAD4XN_SSSE3(8) +MASKSAD4XN_SSSE3(4) + +// For width a multiple of 16 +// Assumes values in m are <=64 +static INLINE unsigned int masked_sad_ssse3(const uint8_t *a_ptr, int a_stride, + const uint8_t *b_ptr, int b_stride, + const uint8_t *m_ptr, int m_stride, + int width, int height) { + int y, x; + __m128i a, b, m, temp1, temp2; + __m128i res = _mm_setzero_si128(); + __m128i one = _mm_set1_epi16(1); + // For each row + for (y = 0; y < height; y++) { + // Covering the full width + for (x = 0; x < width; x += 16) { + // Load a, b, m in xmm registers + a = _mm_loadu_si128((const __m128i *)(a_ptr + x)); + b = _mm_loadu_si128((const __m128i *)(b_ptr + x)); + m = _mm_loadu_si128((const __m128i *)(m_ptr + x)); + + // Calculate the difference between a & b + temp1 = _mm_subs_epu8(a, b); + temp2 = _mm_subs_epu8(b, a); + temp1 = _mm_or_si128(temp1, temp2); + + // Multiply by m and add together + temp2 = _mm_maddubs_epi16(temp1, m); + // Pad out row result to 32 bit integers & add to running total + res = _mm_add_epi32(res, _mm_madd_epi16(temp2, one)); + } + // Move onto the next row + a_ptr += a_stride; + b_ptr += b_stride; + m_ptr += m_stride; + } + res = _mm_hadd_epi32(res, _mm_setzero_si128()); + res = _mm_hadd_epi32(res, _mm_setzero_si128()); + // sad = (sad + 31) >> 6; + return (_mm_cvtsi128_si32(res) + 31) >> 6; +} + +static INLINE unsigned int masked_sad8xh_ssse3( + const uint8_t *a_ptr, int a_stride, const uint8_t *b_ptr, int b_stride, + const uint8_t *m_ptr, int m_stride, int height) { + int y; + __m128i a, b, m, temp1, temp2, row_res; + __m128i res = _mm_setzero_si128(); + __m128i one = _mm_set1_epi16(1); + // Add the masked SAD for 2 rows at a time + for (y = 0; y < height; y += 2) { + // Load a, b, m in xmm registers + a = width8_load_2rows(a_ptr, a_stride); + b = width8_load_2rows(b_ptr, b_stride); + m = width8_load_2rows(m_ptr, m_stride); + + // Calculate the difference between a & b + temp1 = _mm_subs_epu8(a, b); + temp2 = _mm_subs_epu8(b, a); + temp1 = _mm_or_si128(temp1, temp2); + + // Multiply by m and add together + row_res = _mm_maddubs_epi16(temp1, m); + + // Pad out row result to 32 bit integers & add to running total + res = _mm_add_epi32(res, _mm_madd_epi16(row_res, one)); + + // Move onto the next rows + a_ptr += a_stride * 2; + b_ptr += b_stride * 2; + m_ptr += m_stride * 2; + } + res = _mm_hadd_epi32(res, _mm_setzero_si128()); + res = _mm_hadd_epi32(res, _mm_setzero_si128()); + // sad = (sad + 31) >> 6; + return (_mm_cvtsi128_si32(res) + 31) >> 6; +} + +static INLINE unsigned int masked_sad4xh_ssse3( + const uint8_t *a_ptr, int a_stride, const uint8_t *b_ptr, int b_stride, + const uint8_t *m_ptr, int m_stride, int height) { + int y; + __m128i a, b, m, temp1, temp2, row_res; + __m128i res = _mm_setzero_si128(); + __m128i one = _mm_set1_epi16(1); + // Add the masked SAD for 4 rows at a time + for (y = 0; y < height; y += 4) { + // Load a, b, m in xmm registers + a = width4_load_4rows(a_ptr, a_stride); + b = width4_load_4rows(b_ptr, b_stride); + m = width4_load_4rows(m_ptr, m_stride); + + // Calculate the difference between a & b + temp1 = _mm_subs_epu8(a, b); + temp2 = _mm_subs_epu8(b, a); + temp1 = _mm_or_si128(temp1, temp2); + + // Multiply by m and add together + row_res = _mm_maddubs_epi16(temp1, m); + + // Pad out row result to 32 bit integers & add to running total + res = _mm_add_epi32(res, _mm_madd_epi16(row_res, one)); + + // Move onto the next rows + a_ptr += a_stride * 4; + b_ptr += b_stride * 4; + m_ptr += m_stride * 4; + } + // Pad out row result to 32 bit integers & add to running total + res = _mm_hadd_epi32(res, _mm_setzero_si128()); + res = _mm_hadd_epi32(res, _mm_setzero_si128()); + // sad = (sad + 31) >> 6; + return (_mm_cvtsi128_si32(res) + 31) >> 6; +} + +#if CONFIG_HIGHBITDEPTH +static INLINE __m128i highbd_width4_load_2rows(const uint16_t *ptr, + int stride) { + __m128i temp1 = _mm_loadl_epi64((const __m128i *)ptr); + __m128i temp2 = _mm_loadl_epi64((const __m128i *)(ptr + stride)); + return _mm_unpacklo_epi64(temp1, temp2); +} + +static INLINE unsigned int highbd_masked_sad_ssse3( + const uint8_t *a8_ptr, int a_stride, const uint8_t *b8_ptr, int b_stride, + const uint8_t *m_ptr, int m_stride, int width, int height); + +static INLINE unsigned int highbd_masked_sad4xh_ssse3( + const uint8_t *a8_ptr, int a_stride, const uint8_t *b8_ptr, int b_stride, + const uint8_t *m_ptr, int m_stride, int height); + +#define HIGHBD_MASKSADMXN_SSSE3(m, n) \ + unsigned int aom_highbd_masked_sad##m##x##n##_ssse3( \ + const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, \ + const uint8_t *msk, int msk_stride) { \ + return highbd_masked_sad_ssse3(src, src_stride, ref, ref_stride, msk, \ + msk_stride, m, n); \ + } + +#if CONFIG_EXT_PARTITION +HIGHBD_MASKSADMXN_SSSE3(128, 128) +HIGHBD_MASKSADMXN_SSSE3(128, 64) +HIGHBD_MASKSADMXN_SSSE3(64, 128) +#endif // CONFIG_EXT_PARTITION +HIGHBD_MASKSADMXN_SSSE3(64, 64) +HIGHBD_MASKSADMXN_SSSE3(64, 32) +HIGHBD_MASKSADMXN_SSSE3(32, 64) +HIGHBD_MASKSADMXN_SSSE3(32, 32) +HIGHBD_MASKSADMXN_SSSE3(32, 16) +HIGHBD_MASKSADMXN_SSSE3(16, 32) +HIGHBD_MASKSADMXN_SSSE3(16, 16) +HIGHBD_MASKSADMXN_SSSE3(16, 8) +HIGHBD_MASKSADMXN_SSSE3(8, 16) +HIGHBD_MASKSADMXN_SSSE3(8, 8) +HIGHBD_MASKSADMXN_SSSE3(8, 4) + +#define HIGHBD_MASKSAD4XN_SSSE3(n) \ + unsigned int aom_highbd_masked_sad4x##n##_ssse3( \ + const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, \ + const uint8_t *msk, int msk_stride) { \ + return highbd_masked_sad4xh_ssse3(src, src_stride, ref, ref_stride, msk, \ + msk_stride, n); \ + } + +HIGHBD_MASKSAD4XN_SSSE3(8) +HIGHBD_MASKSAD4XN_SSSE3(4) + +// For width a multiple of 8 +// Assumes values in m are <=64 +static INLINE unsigned int highbd_masked_sad_ssse3( + const uint8_t *a8_ptr, int a_stride, const uint8_t *b8_ptr, int b_stride, + const uint8_t *m_ptr, int m_stride, int width, int height) { + int y, x; + __m128i a, b, m, temp1, temp2; + const uint16_t *a_ptr = CONVERT_TO_SHORTPTR(a8_ptr); + const uint16_t *b_ptr = CONVERT_TO_SHORTPTR(b8_ptr); + __m128i res = _mm_setzero_si128(); + // For each row + for (y = 0; y < height; y++) { + // Covering the full width + for (x = 0; x < width; x += 8) { + // Load a, b, m in xmm registers + a = _mm_loadu_si128((const __m128i *)(a_ptr + x)); + b = _mm_loadu_si128((const __m128i *)(b_ptr + x)); + m = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(m_ptr + x)), + _mm_setzero_si128()); + + // Calculate the difference between a & b + temp1 = _mm_subs_epu16(a, b); + temp2 = _mm_subs_epu16(b, a); + temp1 = _mm_or_si128(temp1, temp2); + + // Add result of multiplying by m and add pairs together to running total + res = _mm_add_epi32(res, _mm_madd_epi16(temp1, m)); + } + // Move onto the next row + a_ptr += a_stride; + b_ptr += b_stride; + m_ptr += m_stride; + } + res = _mm_hadd_epi32(res, _mm_setzero_si128()); + res = _mm_hadd_epi32(res, _mm_setzero_si128()); + // sad = (sad + 31) >> 6; + return (_mm_cvtsi128_si32(res) + 31) >> 6; +} + +static INLINE unsigned int highbd_masked_sad4xh_ssse3( + const uint8_t *a8_ptr, int a_stride, const uint8_t *b8_ptr, int b_stride, + const uint8_t *m_ptr, int m_stride, int height) { + int y; + __m128i a, b, m, temp1, temp2; + const uint16_t *a_ptr = CONVERT_TO_SHORTPTR(a8_ptr); + const uint16_t *b_ptr = CONVERT_TO_SHORTPTR(b8_ptr); + __m128i res = _mm_setzero_si128(); + // Add the masked SAD for 2 rows at a time + for (y = 0; y < height; y += 2) { + // Load a, b, m in xmm registers + a = highbd_width4_load_2rows(a_ptr, a_stride); + b = highbd_width4_load_2rows(b_ptr, b_stride); + temp1 = _mm_loadl_epi64((const __m128i *)m_ptr); + temp2 = _mm_loadl_epi64((const __m128i *)(m_ptr + m_stride)); + m = _mm_unpacklo_epi8(_mm_unpacklo_epi32(temp1, temp2), + _mm_setzero_si128()); + + // Calculate the difference between a & b + temp1 = _mm_subs_epu16(a, b); + temp2 = _mm_subs_epu16(b, a); + temp1 = _mm_or_si128(temp1, temp2); + + // Multiply by m and add together + res = _mm_add_epi32(res, _mm_madd_epi16(temp1, m)); + + // Move onto the next rows + a_ptr += a_stride * 2; + b_ptr += b_stride * 2; + m_ptr += m_stride * 2; + } + res = _mm_hadd_epi32(res, _mm_setzero_si128()); + res = _mm_hadd_epi32(res, _mm_setzero_si128()); + // sad = (sad + 31) >> 6; + return (_mm_cvtsi128_si32(res) + 31) >> 6; +} +#endif // CONFIG_HIGHBITDEPTH diff --git a/third_party/aom/aom_dsp/x86/masked_variance_intrin_ssse3.c b/third_party/aom/aom_dsp/x86/masked_variance_intrin_ssse3.c new file mode 100644 index 000000000..fe14597f6 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/masked_variance_intrin_ssse3.c @@ -0,0 +1,1948 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include <assert.h> +#include <stdlib.h> +#include <emmintrin.h> +#include <tmmintrin.h> + +#include "./aom_config.h" +#include "aom/aom_integer.h" +#include "aom_ports/mem.h" +#include "aom_dsp/aom_filter.h" + +// Half pixel shift +#define HALF_PIXEL_OFFSET (BIL_SUBPEL_SHIFTS / 2) + +/***************************************************************************** + * Horizontal additions + *****************************************************************************/ + +static INLINE int32_t hsum_epi32_si32(__m128i v_d) { + v_d = _mm_hadd_epi32(v_d, v_d); + v_d = _mm_hadd_epi32(v_d, v_d); + return _mm_cvtsi128_si32(v_d); +} + +static INLINE int64_t hsum_epi64_si64(__m128i v_q) { + v_q = _mm_add_epi64(v_q, _mm_srli_si128(v_q, 8)); +#if ARCH_X86_64 + return _mm_cvtsi128_si64(v_q); +#else + { + int64_t tmp; + _mm_storel_epi64((__m128i *)&tmp, v_q); + return tmp; + } +#endif +} + +#if CONFIG_HIGHBITDEPTH +static INLINE int64_t hsum_epi32_si64(__m128i v_d) { + const __m128i v_sign_d = _mm_cmplt_epi32(v_d, _mm_setzero_si128()); + const __m128i v_0_q = _mm_unpacklo_epi32(v_d, v_sign_d); + const __m128i v_1_q = _mm_unpackhi_epi32(v_d, v_sign_d); + return hsum_epi64_si64(_mm_add_epi64(v_0_q, v_1_q)); +} +#endif // CONFIG_HIGHBITDEPTH + +static INLINE uint32_t calc_masked_variance(__m128i v_sum_d, __m128i v_sse_q, + uint32_t *sse, int w, int h) { + int64_t sum64; + uint64_t sse64; + + // Horizontal sum + sum64 = hsum_epi32_si32(v_sum_d); + sse64 = hsum_epi64_si64(v_sse_q); + + sum64 = (sum64 >= 0) ? sum64 : -sum64; + + // Round + sum64 = ROUND_POWER_OF_TWO(sum64, 6); + sse64 = ROUND_POWER_OF_TWO(sse64, 12); + + // Store the SSE + *sse = (uint32_t)sse64; + // Compute the variance + return *sse - (uint32_t)((sum64 * sum64) / (w * h)); +} + +/***************************************************************************** + * n*16 Wide versions + *****************************************************************************/ + +static INLINE unsigned int masked_variancewxh_ssse3( + const uint8_t *a, int a_stride, const uint8_t *b, int b_stride, + const uint8_t *m, int m_stride, int w, int h, unsigned int *sse) { + int ii, jj; + + const __m128i v_zero = _mm_setzero_si128(); + + __m128i v_sum_d = _mm_setzero_si128(); + __m128i v_sse_q = _mm_setzero_si128(); + + assert((w % 16) == 0); + + for (ii = 0; ii < h; ii++) { + for (jj = 0; jj < w; jj += 16) { + // Load inputs - 8 bits + const __m128i v_a_b = _mm_loadu_si128((const __m128i *)(a + jj)); + const __m128i v_b_b = _mm_loadu_si128((const __m128i *)(b + jj)); + const __m128i v_m_b = _mm_loadu_si128((const __m128i *)(m + jj)); + + // Unpack to 16 bits - still containing max 8 bits + const __m128i v_a0_w = _mm_unpacklo_epi8(v_a_b, v_zero); + const __m128i v_b0_w = _mm_unpacklo_epi8(v_b_b, v_zero); + const __m128i v_m0_w = _mm_unpacklo_epi8(v_m_b, v_zero); + const __m128i v_a1_w = _mm_unpackhi_epi8(v_a_b, v_zero); + const __m128i v_b1_w = _mm_unpackhi_epi8(v_b_b, v_zero); + const __m128i v_m1_w = _mm_unpackhi_epi8(v_m_b, v_zero); + + // Difference: [-255, 255] + const __m128i v_d0_w = _mm_sub_epi16(v_a0_w, v_b0_w); + const __m128i v_d1_w = _mm_sub_epi16(v_a1_w, v_b1_w); + + // Error - [-255, 255] * [0, 64] = [0xc040, 0x3fc0] => fits in 15 bits + const __m128i v_e0_w = _mm_mullo_epi16(v_d0_w, v_m0_w); + const __m128i v_e0_d = _mm_madd_epi16(v_d0_w, v_m0_w); + const __m128i v_e1_w = _mm_mullo_epi16(v_d1_w, v_m1_w); + const __m128i v_e1_d = _mm_madd_epi16(v_d1_w, v_m1_w); + + // Squared error - using madd it's max (15 bits * 15 bits) * 2 = 31 bits + const __m128i v_se0_d = _mm_madd_epi16(v_e0_w, v_e0_w); + const __m128i v_se1_d = _mm_madd_epi16(v_e1_w, v_e1_w); + + // Sum of v_se{0,1}_d - 31 bits + 31 bits = 32 bits + const __m128i v_se_d = _mm_add_epi32(v_se0_d, v_se1_d); + + // Unpack Squared error to 64 bits + const __m128i v_se_lo_q = _mm_unpacklo_epi32(v_se_d, v_zero); + const __m128i v_se_hi_q = _mm_unpackhi_epi32(v_se_d, v_zero); + + // Accumulate + v_sum_d = _mm_add_epi32(v_sum_d, v_e0_d); + v_sum_d = _mm_add_epi32(v_sum_d, v_e1_d); + v_sse_q = _mm_add_epi64(v_sse_q, v_se_lo_q); + v_sse_q = _mm_add_epi64(v_sse_q, v_se_hi_q); + } + + // Move on to next row + a += a_stride; + b += b_stride; + m += m_stride; + } + + return calc_masked_variance(v_sum_d, v_sse_q, sse, w, h); +} + +#define MASKED_VARWXH(W, H) \ + unsigned int aom_masked_variance##W##x##H##_ssse3( \ + const uint8_t *a, int a_stride, const uint8_t *b, int b_stride, \ + const uint8_t *m, int m_stride, unsigned int *sse) { \ + return masked_variancewxh_ssse3(a, a_stride, b, b_stride, m, m_stride, W, \ + H, sse); \ + } + +MASKED_VARWXH(16, 8) +MASKED_VARWXH(16, 16) +MASKED_VARWXH(16, 32) +MASKED_VARWXH(32, 16) +MASKED_VARWXH(32, 32) +MASKED_VARWXH(32, 64) +MASKED_VARWXH(64, 32) +MASKED_VARWXH(64, 64) +#if CONFIG_EXT_PARTITION +MASKED_VARWXH(64, 128) +MASKED_VARWXH(128, 64) +MASKED_VARWXH(128, 128) +#endif // CONFIG_EXT_PARTITION + +/***************************************************************************** + * 8 Wide versions + *****************************************************************************/ + +static INLINE unsigned int masked_variance8xh_ssse3( + const uint8_t *a, int a_stride, const uint8_t *b, int b_stride, + const uint8_t *m, int m_stride, int h, unsigned int *sse) { + int ii; + + const __m128i v_zero = _mm_setzero_si128(); + + __m128i v_sum_d = _mm_setzero_si128(); + __m128i v_sse_q = _mm_setzero_si128(); + + for (ii = 0; ii < h; ii++) { + // Load inputs - 8 bits + const __m128i v_a_b = _mm_loadl_epi64((const __m128i *)a); + const __m128i v_b_b = _mm_loadl_epi64((const __m128i *)b); + const __m128i v_m_b = _mm_loadl_epi64((const __m128i *)m); + + // Unpack to 16 bits - still containing max 8 bits + const __m128i v_a_w = _mm_unpacklo_epi8(v_a_b, v_zero); + const __m128i v_b_w = _mm_unpacklo_epi8(v_b_b, v_zero); + const __m128i v_m_w = _mm_unpacklo_epi8(v_m_b, v_zero); + + // Difference: [-255, 255] + const __m128i v_d_w = _mm_sub_epi16(v_a_w, v_b_w); + + // Error - [-255, 255] * [0, 64] = [0xc040, 0x3fc0] => fits in 15 bits + const __m128i v_e_w = _mm_mullo_epi16(v_d_w, v_m_w); + const __m128i v_e_d = _mm_madd_epi16(v_d_w, v_m_w); + + // Squared error - using madd it's max (15 bits * 15 bits) * 2 = 31 bits + const __m128i v_se_d = _mm_madd_epi16(v_e_w, v_e_w); + + // Unpack Squared error to 64 bits + const __m128i v_se_lo_q = _mm_unpacklo_epi32(v_se_d, v_zero); + const __m128i v_se_hi_q = _mm_unpackhi_epi32(v_se_d, v_zero); + + // Accumulate + v_sum_d = _mm_add_epi32(v_sum_d, v_e_d); + v_sse_q = _mm_add_epi64(v_sse_q, v_se_lo_q); + v_sse_q = _mm_add_epi64(v_sse_q, v_se_hi_q); + + // Move on to next row + a += a_stride; + b += b_stride; + m += m_stride; + } + + return calc_masked_variance(v_sum_d, v_sse_q, sse, 8, h); +} + +#define MASKED_VAR8XH(H) \ + unsigned int aom_masked_variance8x##H##_ssse3( \ + const uint8_t *a, int a_stride, const uint8_t *b, int b_stride, \ + const uint8_t *m, int m_stride, unsigned int *sse) { \ + return masked_variance8xh_ssse3(a, a_stride, b, b_stride, m, m_stride, H, \ + sse); \ + } + +MASKED_VAR8XH(4) +MASKED_VAR8XH(8) +MASKED_VAR8XH(16) + +/***************************************************************************** + * 4 Wide versions + *****************************************************************************/ + +static INLINE unsigned int masked_variance4xh_ssse3( + const uint8_t *a, int a_stride, const uint8_t *b, int b_stride, + const uint8_t *m, int m_stride, int h, unsigned int *sse) { + int ii; + + const __m128i v_zero = _mm_setzero_si128(); + + __m128i v_sum_d = _mm_setzero_si128(); + __m128i v_sse_q = _mm_setzero_si128(); + + assert((h % 2) == 0); + + for (ii = 0; ii < h / 2; ii++) { + // Load 2 input rows - 8 bits + const __m128i v_a0_b = _mm_cvtsi32_si128(*(const uint32_t *)a); + const __m128i v_b0_b = _mm_cvtsi32_si128(*(const uint32_t *)b); + const __m128i v_m0_b = _mm_cvtsi32_si128(*(const uint32_t *)m); + const __m128i v_a1_b = _mm_cvtsi32_si128(*(const uint32_t *)(a + a_stride)); + const __m128i v_b1_b = _mm_cvtsi32_si128(*(const uint32_t *)(b + b_stride)); + const __m128i v_m1_b = _mm_cvtsi32_si128(*(const uint32_t *)(m + m_stride)); + + // Interleave 2 rows into a single register + const __m128i v_a_b = _mm_unpacklo_epi32(v_a0_b, v_a1_b); + const __m128i v_b_b = _mm_unpacklo_epi32(v_b0_b, v_b1_b); + const __m128i v_m_b = _mm_unpacklo_epi32(v_m0_b, v_m1_b); + + // Unpack to 16 bits - still containing max 8 bits + const __m128i v_a_w = _mm_unpacklo_epi8(v_a_b, v_zero); + const __m128i v_b_w = _mm_unpacklo_epi8(v_b_b, v_zero); + const __m128i v_m_w = _mm_unpacklo_epi8(v_m_b, v_zero); + + // Difference: [-255, 255] + const __m128i v_d_w = _mm_sub_epi16(v_a_w, v_b_w); + + // Error - [-255, 255] * [0, 64] = [0xc040, 0x3fc0] => fits in 15 bits + const __m128i v_e_w = _mm_mullo_epi16(v_d_w, v_m_w); + const __m128i v_e_d = _mm_madd_epi16(v_d_w, v_m_w); + + // Squared error - using madd it's max (15 bits * 15 bits) * 2 = 31 bits + const __m128i v_se_d = _mm_madd_epi16(v_e_w, v_e_w); + + // Unpack Squared error to 64 bits + const __m128i v_se_lo_q = _mm_unpacklo_epi32(v_se_d, v_zero); + const __m128i v_se_hi_q = _mm_unpackhi_epi32(v_se_d, v_zero); + + // Accumulate + v_sum_d = _mm_add_epi32(v_sum_d, v_e_d); + v_sse_q = _mm_add_epi64(v_sse_q, v_se_lo_q); + v_sse_q = _mm_add_epi64(v_sse_q, v_se_hi_q); + + // Move on to next 2 row + a += a_stride * 2; + b += b_stride * 2; + m += m_stride * 2; + } + + return calc_masked_variance(v_sum_d, v_sse_q, sse, 4, h); +} + +#define MASKED_VAR4XH(H) \ + unsigned int aom_masked_variance4x##H##_ssse3( \ + const uint8_t *a, int a_stride, const uint8_t *b, int b_stride, \ + const uint8_t *m, int m_stride, unsigned int *sse) { \ + return masked_variance4xh_ssse3(a, a_stride, b, b_stride, m, m_stride, H, \ + sse); \ + } + +MASKED_VAR4XH(4) +MASKED_VAR4XH(8) + +#if CONFIG_HIGHBITDEPTH + +// Main calculation for n*8 wide blocks +static INLINE void highbd_masked_variance64_ssse3( + const uint16_t *a, int a_stride, const uint16_t *b, int b_stride, + const uint8_t *m, int m_stride, int w, int h, int64_t *sum, uint64_t *sse) { + int ii, jj; + + const __m128i v_zero = _mm_setzero_si128(); + + __m128i v_sum_d = _mm_setzero_si128(); + __m128i v_sse_q = _mm_setzero_si128(); + + assert((w % 8) == 0); + + for (ii = 0; ii < h; ii++) { + for (jj = 0; jj < w; jj += 8) { + // Load inputs - 8 bits + const __m128i v_a_w = _mm_loadu_si128((const __m128i *)(a + jj)); + const __m128i v_b_w = _mm_loadu_si128((const __m128i *)(b + jj)); + const __m128i v_m_b = _mm_loadl_epi64((const __m128i *)(m + jj)); + + // Unpack m to 16 bits - still containing max 8 bits + const __m128i v_m_w = _mm_unpacklo_epi8(v_m_b, v_zero); + + // Difference: [-4095, 4095] + const __m128i v_d_w = _mm_sub_epi16(v_a_w, v_b_w); + + // Error - [-4095, 4095] * [0, 64] => sum of 2 of these fits in 19 bits + const __m128i v_e_d = _mm_madd_epi16(v_d_w, v_m_w); + + // Squared error - max (18 bits * 18 bits) = 36 bits (no sign bit) + const __m128i v_absd_w = _mm_abs_epi16(v_d_w); + const __m128i v_dlo_d = _mm_unpacklo_epi16(v_absd_w, v_zero); + const __m128i v_mlo_d = _mm_unpacklo_epi16(v_m_w, v_zero); + const __m128i v_elo_d = _mm_madd_epi16(v_dlo_d, v_mlo_d); + const __m128i v_dhi_d = _mm_unpackhi_epi16(v_absd_w, v_zero); + const __m128i v_mhi_d = _mm_unpackhi_epi16(v_m_w, v_zero); + const __m128i v_ehi_d = _mm_madd_epi16(v_dhi_d, v_mhi_d); + // Square and sum the errors -> 36bits * 4 = 38bits + __m128i v_se0_q, v_se1_q, v_se2_q, v_se3_q, v_se_q, v_elo1_d, v_ehi3_d; + v_se0_q = _mm_mul_epu32(v_elo_d, v_elo_d); + v_elo1_d = _mm_srli_si128(v_elo_d, 4); + v_se1_q = _mm_mul_epu32(v_elo1_d, v_elo1_d); + v_se0_q = _mm_add_epi64(v_se0_q, v_se1_q); + v_se2_q = _mm_mul_epu32(v_ehi_d, v_ehi_d); + v_ehi3_d = _mm_srli_si128(v_ehi_d, 4); + v_se3_q = _mm_mul_epu32(v_ehi3_d, v_ehi3_d); + v_se1_q = _mm_add_epi64(v_se2_q, v_se3_q); + v_se_q = _mm_add_epi64(v_se0_q, v_se1_q); + + // Accumulate + v_sum_d = _mm_add_epi32(v_sum_d, v_e_d); + v_sse_q = _mm_add_epi64(v_sse_q, v_se_q); + } + + // Move on to next row + a += a_stride; + b += b_stride; + m += m_stride; + } + + // Horizontal sum + *sum = hsum_epi32_si64(v_sum_d); + *sse = hsum_epi64_si64(v_sse_q); + + // Round + *sum = (*sum >= 0) ? *sum : -*sum; + *sum = ROUND_POWER_OF_TWO(*sum, 6); + *sse = ROUND_POWER_OF_TWO(*sse, 12); +} + +// Main calculation for 4 wide blocks +static INLINE void highbd_masked_variance64_4wide_ssse3( + const uint16_t *a, int a_stride, const uint16_t *b, int b_stride, + const uint8_t *m, int m_stride, int h, int64_t *sum, uint64_t *sse) { + int ii; + + const __m128i v_zero = _mm_setzero_si128(); + + __m128i v_sum_d = _mm_setzero_si128(); + __m128i v_sse_q = _mm_setzero_si128(); + + assert((h % 2) == 0); + + for (ii = 0; ii < h / 2; ii++) { + // Load 2 input rows - 8 bits + const __m128i v_a0_w = _mm_loadl_epi64((const __m128i *)a); + const __m128i v_b0_w = _mm_loadl_epi64((const __m128i *)b); + const __m128i v_m0_b = _mm_cvtsi32_si128(*(const uint32_t *)m); + const __m128i v_a1_w = _mm_loadl_epi64((const __m128i *)(a + a_stride)); + const __m128i v_b1_w = _mm_loadl_epi64((const __m128i *)(b + b_stride)); + const __m128i v_m1_b = _mm_cvtsi32_si128(*(const uint32_t *)(m + m_stride)); + + // Interleave 2 rows into a single register + const __m128i v_a_w = _mm_unpacklo_epi64(v_a0_w, v_a1_w); + const __m128i v_b_w = _mm_unpacklo_epi64(v_b0_w, v_b1_w); + const __m128i v_m_b = _mm_unpacklo_epi32(v_m0_b, v_m1_b); + + // Unpack to 16 bits - still containing max 8 bits + const __m128i v_m_w = _mm_unpacklo_epi8(v_m_b, v_zero); + + // Difference: [-4095, 4095] + const __m128i v_d_w = _mm_sub_epi16(v_a_w, v_b_w); + + // Error - [-4095, 4095] * [0, 64] => fits in 19 bits (incld sign bit) + const __m128i v_e_d = _mm_madd_epi16(v_d_w, v_m_w); + + // Squared error - max (18 bits * 18 bits) = 36 bits (no sign bit) + const __m128i v_absd_w = _mm_abs_epi16(v_d_w); + const __m128i v_dlo_d = _mm_unpacklo_epi16(v_absd_w, v_zero); + const __m128i v_mlo_d = _mm_unpacklo_epi16(v_m_w, v_zero); + const __m128i v_elo_d = _mm_madd_epi16(v_dlo_d, v_mlo_d); + const __m128i v_dhi_d = _mm_unpackhi_epi16(v_absd_w, v_zero); + const __m128i v_mhi_d = _mm_unpackhi_epi16(v_m_w, v_zero); + const __m128i v_ehi_d = _mm_madd_epi16(v_dhi_d, v_mhi_d); + // Square and sum the errors -> 36bits * 4 = 38bits + __m128i v_se0_q, v_se1_q, v_se2_q, v_se3_q, v_se_q, v_elo1_d, v_ehi3_d; + v_se0_q = _mm_mul_epu32(v_elo_d, v_elo_d); + v_elo1_d = _mm_srli_si128(v_elo_d, 4); + v_se1_q = _mm_mul_epu32(v_elo1_d, v_elo1_d); + v_se0_q = _mm_add_epi64(v_se0_q, v_se1_q); + v_se2_q = _mm_mul_epu32(v_ehi_d, v_ehi_d); + v_ehi3_d = _mm_srli_si128(v_ehi_d, 4); + v_se3_q = _mm_mul_epu32(v_ehi3_d, v_ehi3_d); + v_se1_q = _mm_add_epi64(v_se2_q, v_se3_q); + v_se_q = _mm_add_epi64(v_se0_q, v_se1_q); + + // Accumulate + v_sum_d = _mm_add_epi32(v_sum_d, v_e_d); + v_sse_q = _mm_add_epi64(v_sse_q, v_se_q); + + // Move on to next row + a += a_stride * 2; + b += b_stride * 2; + m += m_stride * 2; + } + + // Horizontal sum + *sum = hsum_epi32_si32(v_sum_d); + *sse = hsum_epi64_si64(v_sse_q); + + // Round + *sum = (*sum >= 0) ? *sum : -*sum; + *sum = ROUND_POWER_OF_TWO(*sum, 6); + *sse = ROUND_POWER_OF_TWO(*sse, 12); +} + +static INLINE unsigned int highbd_masked_variancewxh_ssse3( + const uint16_t *a, int a_stride, const uint16_t *b, int b_stride, + const uint8_t *m, int m_stride, int w, int h, unsigned int *sse) { + uint64_t sse64; + int64_t sum64; + + if (w == 4) + highbd_masked_variance64_4wide_ssse3(a, a_stride, b, b_stride, m, m_stride, + h, &sum64, &sse64); + else + highbd_masked_variance64_ssse3(a, a_stride, b, b_stride, m, m_stride, w, h, + &sum64, &sse64); + + // Store the SSE + *sse = (uint32_t)sse64; + // Compute and return variance + return *sse - (uint32_t)((sum64 * sum64) / (w * h)); +} + +static INLINE unsigned int highbd_10_masked_variancewxh_ssse3( + const uint16_t *a, int a_stride, const uint16_t *b, int b_stride, + const uint8_t *m, int m_stride, int w, int h, unsigned int *sse) { + uint64_t sse64; + int64_t sum64; + + if (w == 4) + highbd_masked_variance64_4wide_ssse3(a, a_stride, b, b_stride, m, m_stride, + h, &sum64, &sse64); + else + highbd_masked_variance64_ssse3(a, a_stride, b, b_stride, m, m_stride, w, h, + &sum64, &sse64); + + // Normalise + sum64 = ROUND_POWER_OF_TWO(sum64, 2); + sse64 = ROUND_POWER_OF_TWO(sse64, 4); + + // Store the SSE + *sse = (uint32_t)sse64; + // Compute and return variance + return *sse - (uint32_t)((sum64 * sum64) / (w * h)); +} + +static INLINE unsigned int highbd_12_masked_variancewxh_ssse3( + const uint16_t *a, int a_stride, const uint16_t *b, int b_stride, + const uint8_t *m, int m_stride, int w, int h, unsigned int *sse) { + uint64_t sse64; + int64_t sum64; + + if (w == 4) + highbd_masked_variance64_4wide_ssse3(a, a_stride, b, b_stride, m, m_stride, + h, &sum64, &sse64); + else + highbd_masked_variance64_ssse3(a, a_stride, b, b_stride, m, m_stride, w, h, + &sum64, &sse64); + + sum64 = ROUND_POWER_OF_TWO(sum64, 4); + sse64 = ROUND_POWER_OF_TWO(sse64, 8); + + // Store the SSE + *sse = (uint32_t)sse64; + // Compute and return variance + return *sse - (uint32_t)((sum64 * sum64) / (w * h)); +} + +#define HIGHBD_MASKED_VARWXH(W, H) \ + unsigned int aom_highbd_masked_variance##W##x##H##_ssse3( \ + const uint8_t *a8, int a_stride, const uint8_t *b8, int b_stride, \ + const uint8_t *m, int m_stride, unsigned int *sse) { \ + uint16_t *a = CONVERT_TO_SHORTPTR(a8); \ + uint16_t *b = CONVERT_TO_SHORTPTR(b8); \ + return highbd_masked_variancewxh_ssse3(a, a_stride, b, b_stride, m, \ + m_stride, W, H, sse); \ + } \ + \ + unsigned int aom_highbd_10_masked_variance##W##x##H##_ssse3( \ + const uint8_t *a8, int a_stride, const uint8_t *b8, int b_stride, \ + const uint8_t *m, int m_stride, unsigned int *sse) { \ + uint16_t *a = CONVERT_TO_SHORTPTR(a8); \ + uint16_t *b = CONVERT_TO_SHORTPTR(b8); \ + return highbd_10_masked_variancewxh_ssse3(a, a_stride, b, b_stride, m, \ + m_stride, W, H, sse); \ + } \ + \ + unsigned int aom_highbd_12_masked_variance##W##x##H##_ssse3( \ + const uint8_t *a8, int a_stride, const uint8_t *b8, int b_stride, \ + const uint8_t *m, int m_stride, unsigned int *sse) { \ + uint16_t *a = CONVERT_TO_SHORTPTR(a8); \ + uint16_t *b = CONVERT_TO_SHORTPTR(b8); \ + return highbd_12_masked_variancewxh_ssse3(a, a_stride, b, b_stride, m, \ + m_stride, W, H, sse); \ + } + +HIGHBD_MASKED_VARWXH(4, 4) +HIGHBD_MASKED_VARWXH(4, 8) +HIGHBD_MASKED_VARWXH(8, 4) +HIGHBD_MASKED_VARWXH(8, 8) +HIGHBD_MASKED_VARWXH(8, 16) +HIGHBD_MASKED_VARWXH(16, 8) +HIGHBD_MASKED_VARWXH(16, 16) +HIGHBD_MASKED_VARWXH(16, 32) +HIGHBD_MASKED_VARWXH(32, 16) +HIGHBD_MASKED_VARWXH(32, 32) +HIGHBD_MASKED_VARWXH(32, 64) +HIGHBD_MASKED_VARWXH(64, 32) +HIGHBD_MASKED_VARWXH(64, 64) +#if CONFIG_EXT_PARTITION +HIGHBD_MASKED_VARWXH(64, 128) +HIGHBD_MASKED_VARWXH(128, 64) +HIGHBD_MASKED_VARWXH(128, 128) +#endif // CONFIG_EXT_PARTITION + +#endif + +////////////////////////////////////////////////////////////////////////////// +// Sub pixel versions +////////////////////////////////////////////////////////////////////////////// + +typedef __m128i (*filter_fn_t)(__m128i v_a_b, __m128i v_b_b, + __m128i v_filter_b); + +static INLINE __m128i apply_filter_avg(const __m128i v_a_b, const __m128i v_b_b, + const __m128i v_filter_b) { + (void)v_filter_b; + return _mm_avg_epu8(v_a_b, v_b_b); +} + +static INLINE __m128i apply_filter(const __m128i v_a_b, const __m128i v_b_b, + const __m128i v_filter_b) { + const __m128i v_rounding_w = _mm_set1_epi16(1 << (FILTER_BITS - 1)); + __m128i v_input_lo_b = _mm_unpacklo_epi8(v_a_b, v_b_b); + __m128i v_input_hi_b = _mm_unpackhi_epi8(v_a_b, v_b_b); + __m128i v_temp0_w = _mm_maddubs_epi16(v_input_lo_b, v_filter_b); + __m128i v_temp1_w = _mm_maddubs_epi16(v_input_hi_b, v_filter_b); + __m128i v_res_lo_w = + _mm_srai_epi16(_mm_add_epi16(v_temp0_w, v_rounding_w), FILTER_BITS); + __m128i v_res_hi_w = + _mm_srai_epi16(_mm_add_epi16(v_temp1_w, v_rounding_w), FILTER_BITS); + return _mm_packus_epi16(v_res_lo_w, v_res_hi_w); +} + +// Apply the filter to the contents of the lower half of a and b +static INLINE void apply_filter_lo(const __m128i v_a_lo_b, + const __m128i v_b_lo_b, + const __m128i v_filter_b, __m128i *v_res_w) { + const __m128i v_rounding_w = _mm_set1_epi16(1 << (FILTER_BITS - 1)); + __m128i v_input_b = _mm_unpacklo_epi8(v_a_lo_b, v_b_lo_b); + __m128i v_temp0_w = _mm_maddubs_epi16(v_input_b, v_filter_b); + *v_res_w = + _mm_srai_epi16(_mm_add_epi16(v_temp0_w, v_rounding_w), FILTER_BITS); +} + +static void sum_and_sse(const __m128i v_a_b, const __m128i v_b_b, + const __m128i v_m_b, __m128i *v_sum_d, + __m128i *v_sse_q) { + const __m128i v_zero = _mm_setzero_si128(); + // Unpack to 16 bits - still containing max 8 bits + const __m128i v_a0_w = _mm_unpacklo_epi8(v_a_b, v_zero); + const __m128i v_b0_w = _mm_unpacklo_epi8(v_b_b, v_zero); + const __m128i v_m0_w = _mm_unpacklo_epi8(v_m_b, v_zero); + const __m128i v_a1_w = _mm_unpackhi_epi8(v_a_b, v_zero); + const __m128i v_b1_w = _mm_unpackhi_epi8(v_b_b, v_zero); + const __m128i v_m1_w = _mm_unpackhi_epi8(v_m_b, v_zero); + + // Difference: [-255, 255] + const __m128i v_d0_w = _mm_sub_epi16(v_a0_w, v_b0_w); + const __m128i v_d1_w = _mm_sub_epi16(v_a1_w, v_b1_w); + + // Error - [-255, 255] * [0, 64] = [0xc040, 0x3fc0] => fits in 15 bits + const __m128i v_e0_w = _mm_mullo_epi16(v_d0_w, v_m0_w); + const __m128i v_e0_d = _mm_madd_epi16(v_d0_w, v_m0_w); + const __m128i v_e1_w = _mm_mullo_epi16(v_d1_w, v_m1_w); + const __m128i v_e1_d = _mm_madd_epi16(v_d1_w, v_m1_w); + + // Squared error - using madd it's max (15 bits * 15 bits) * 2 = 31 bits + const __m128i v_se0_d = _mm_madd_epi16(v_e0_w, v_e0_w); + const __m128i v_se1_d = _mm_madd_epi16(v_e1_w, v_e1_w); + + // Sum of v_se{0,1}_d - 31 bits + 31 bits = 32 bits + const __m128i v_se_d = _mm_add_epi32(v_se0_d, v_se1_d); + + // Unpack Squared error to 64 bits + const __m128i v_se_lo_q = _mm_unpacklo_epi32(v_se_d, v_zero); + const __m128i v_se_hi_q = _mm_unpackhi_epi32(v_se_d, v_zero); + + // Accumulate + *v_sum_d = _mm_add_epi32(*v_sum_d, v_e0_d); + *v_sum_d = _mm_add_epi32(*v_sum_d, v_e1_d); + *v_sse_q = _mm_add_epi64(*v_sse_q, v_se_lo_q); + *v_sse_q = _mm_add_epi64(*v_sse_q, v_se_hi_q); +} + +// Functions for width (W) >= 16 +unsigned int aom_masked_subpel_varWxH_xzero(const uint8_t *src, int src_stride, + int yoffset, const uint8_t *dst, + int dst_stride, const uint8_t *msk, + int msk_stride, unsigned int *sse, + int w, int h, + filter_fn_t filter_fn) { + int i, j; + __m128i v_src0_b, v_src1_b, v_res_b, v_dst_b, v_msk_b; + __m128i v_sum_d = _mm_setzero_si128(); + __m128i v_sse_q = _mm_setzero_si128(); + const __m128i v_filter_b = _mm_set1_epi16( + (bilinear_filters_2t[yoffset][1] << 8) + bilinear_filters_2t[yoffset][0]); + assert(yoffset < BIL_SUBPEL_SHIFTS); + for (j = 0; j < w; j += 16) { + // Load the first row ready + v_src0_b = _mm_loadu_si128((const __m128i *)(src + j)); + // Process 2 rows at a time + for (i = 0; i < h; i += 2) { + // Load the next row apply the filter + v_src1_b = _mm_loadu_si128((const __m128i *)(src + j + src_stride)); + v_res_b = filter_fn(v_src0_b, v_src1_b, v_filter_b); + // Load the dst and msk for the variance calculation + v_dst_b = _mm_loadu_si128((const __m128i *)(dst + j)); + v_msk_b = _mm_loadu_si128((const __m128i *)(msk + j)); + sum_and_sse(v_res_b, v_dst_b, v_msk_b, &v_sum_d, &v_sse_q); + + // Load the next row apply the filter + v_src0_b = _mm_loadu_si128((const __m128i *)(src + j + src_stride * 2)); + v_res_b = filter_fn(v_src1_b, v_src0_b, v_filter_b); + // Load the dst and msk for the variance calculation + v_dst_b = _mm_loadu_si128((const __m128i *)(dst + j + dst_stride)); + v_msk_b = _mm_loadu_si128((const __m128i *)(msk + j + msk_stride)); + sum_and_sse(v_res_b, v_dst_b, v_msk_b, &v_sum_d, &v_sse_q); + // Move onto the next block of rows + src += src_stride * 2; + dst += dst_stride * 2; + msk += msk_stride * 2; + } + // Reset to the top of the block + src -= src_stride * h; + dst -= dst_stride * h; + msk -= msk_stride * h; + } + return calc_masked_variance(v_sum_d, v_sse_q, sse, w, h); +} +unsigned int aom_masked_subpel_varWxH_yzero(const uint8_t *src, int src_stride, + int xoffset, const uint8_t *dst, + int dst_stride, const uint8_t *msk, + int msk_stride, unsigned int *sse, + int w, int h, + filter_fn_t filter_fn) { + int i, j; + __m128i v_src0_b, v_src1_b, v_res_b, v_dst_b, v_msk_b; + __m128i v_sum_d = _mm_setzero_si128(); + __m128i v_sse_q = _mm_setzero_si128(); + const __m128i v_filter_b = _mm_set1_epi16( + (bilinear_filters_2t[xoffset][1] << 8) + bilinear_filters_2t[xoffset][0]); + assert(xoffset < BIL_SUBPEL_SHIFTS); + for (i = 0; i < h; i++) { + for (j = 0; j < w; j += 16) { + // Load this row and one below & apply the filter to them + v_src0_b = _mm_loadu_si128((const __m128i *)(src + j)); + v_src1_b = _mm_loadu_si128((const __m128i *)(src + j + 1)); + v_res_b = filter_fn(v_src0_b, v_src1_b, v_filter_b); + + // Load the dst and msk for the variance calculation + v_dst_b = _mm_loadu_si128((const __m128i *)(dst + j)); + v_msk_b = _mm_loadu_si128((const __m128i *)(msk + j)); + sum_and_sse(v_res_b, v_dst_b, v_msk_b, &v_sum_d, &v_sse_q); + } + src += src_stride; + dst += dst_stride; + msk += msk_stride; + } + return calc_masked_variance(v_sum_d, v_sse_q, sse, w, h); +} +unsigned int aom_masked_subpel_varWxH_xnonzero_ynonzero( + const uint8_t *src, int src_stride, int xoffset, int yoffset, + const uint8_t *dst, int dst_stride, const uint8_t *msk, int msk_stride, + unsigned int *sse, int w, int h, filter_fn_t xfilter_fn, + filter_fn_t yfilter_fn) { + int i, j; + __m128i v_src0_b, v_src1_b, v_src2_b, v_src3_b; + __m128i v_filtered0_b, v_filtered1_b, v_res_b, v_dst_b, v_msk_b; + __m128i v_sum_d = _mm_setzero_si128(); + __m128i v_sse_q = _mm_setzero_si128(); + const __m128i v_filterx_b = _mm_set1_epi16( + (bilinear_filters_2t[xoffset][1] << 8) + bilinear_filters_2t[xoffset][0]); + const __m128i v_filtery_b = _mm_set1_epi16( + (bilinear_filters_2t[yoffset][1] << 8) + bilinear_filters_2t[yoffset][0]); + assert(yoffset < BIL_SUBPEL_SHIFTS); + assert(xoffset < BIL_SUBPEL_SHIFTS); + for (j = 0; j < w; j += 16) { + // Load the first row ready + v_src0_b = _mm_loadu_si128((const __m128i *)(src + j)); + v_src1_b = _mm_loadu_si128((const __m128i *)(src + j + 1)); + v_filtered0_b = xfilter_fn(v_src0_b, v_src1_b, v_filterx_b); + // Process 2 rows at a time + for (i = 0; i < h; i += 2) { + // Load the next row & apply the filter + v_src2_b = _mm_loadu_si128((const __m128i *)(src + src_stride + j)); + v_src3_b = _mm_loadu_si128((const __m128i *)(src + src_stride + j + 1)); + v_filtered1_b = xfilter_fn(v_src2_b, v_src3_b, v_filterx_b); + // Load the dst and msk for the variance calculation + v_dst_b = _mm_loadu_si128((const __m128i *)(dst + j)); + v_msk_b = _mm_loadu_si128((const __m128i *)(msk + j)); + // Complete the calculation for this row and add it to the running total + v_res_b = yfilter_fn(v_filtered0_b, v_filtered1_b, v_filtery_b); + sum_and_sse(v_res_b, v_dst_b, v_msk_b, &v_sum_d, &v_sse_q); + + // Load the next row & apply the filter + v_src0_b = _mm_loadu_si128((const __m128i *)(src + src_stride * 2 + j)); + v_src1_b = + _mm_loadu_si128((const __m128i *)(src + src_stride * 2 + j + 1)); + v_filtered0_b = xfilter_fn(v_src0_b, v_src1_b, v_filterx_b); + // Load the dst and msk for the variance calculation + v_dst_b = _mm_loadu_si128((const __m128i *)(dst + dst_stride + j)); + v_msk_b = _mm_loadu_si128((const __m128i *)(msk + msk_stride + j)); + // Complete the calculation for this row and add it to the running total + v_res_b = yfilter_fn(v_filtered1_b, v_filtered0_b, v_filtery_b); + sum_and_sse(v_res_b, v_dst_b, v_msk_b, &v_sum_d, &v_sse_q); + // Move onto the next block of rows + src += src_stride * 2; + dst += dst_stride * 2; + msk += msk_stride * 2; + } + // Reset to the top of the block + src -= src_stride * h; + dst -= dst_stride * h; + msk -= msk_stride * h; + } + return calc_masked_variance(v_sum_d, v_sse_q, sse, w, h); +} + +// Note order in which rows loaded xmm[127:96] = row 1, xmm[95:64] = row 2, +// xmm[63:32] = row 3, xmm[31:0] = row 4 +unsigned int aom_masked_subpel_var4xH_xzero(const uint8_t *src, int src_stride, + int yoffset, const uint8_t *dst, + int dst_stride, const uint8_t *msk, + int msk_stride, unsigned int *sse, + int h) { + int i; + __m128i v_src0_b, v_src1_b, v_src2_b, v_src3_b, v_filtered1_w, v_filtered2_w; + __m128i v_dst0_b, v_dst1_b, v_dst2_b, v_dst3_b; + __m128i v_msk0_b, v_msk1_b, v_msk2_b, v_msk3_b, v_res_b; + __m128i v_sum_d = _mm_setzero_si128(); + __m128i v_sse_q = _mm_setzero_si128(); + __m128i v_filter_b = _mm_set1_epi16((bilinear_filters_2t[yoffset][1] << 8) + + bilinear_filters_2t[yoffset][0]); + assert(yoffset < BIL_SUBPEL_SHIFTS); + // Load the first row of src data ready + v_src0_b = _mm_loadl_epi64((const __m128i *)src); + for (i = 0; i < h; i += 4) { + // Load the rest of the source data for these rows + v_src1_b = _mm_loadl_epi64((const __m128i *)(src + src_stride * 1)); + v_src1_b = _mm_unpacklo_epi32(v_src1_b, v_src0_b); + v_src2_b = _mm_loadl_epi64((const __m128i *)(src + src_stride * 2)); + v_src3_b = _mm_loadl_epi64((const __m128i *)(src + src_stride * 3)); + v_src3_b = _mm_unpacklo_epi32(v_src3_b, v_src2_b); + v_src0_b = _mm_loadl_epi64((const __m128i *)(src + src_stride * 4)); + // Load the dst data + v_dst0_b = _mm_cvtsi32_si128(*(const uint32_t *)(dst + dst_stride * 0)); + v_dst1_b = _mm_cvtsi32_si128(*(const uint32_t *)(dst + dst_stride * 1)); + v_dst0_b = _mm_unpacklo_epi32(v_dst1_b, v_dst0_b); + v_dst2_b = _mm_cvtsi32_si128(*(const uint32_t *)(dst + dst_stride * 2)); + v_dst3_b = _mm_cvtsi32_si128(*(const uint32_t *)(dst + dst_stride * 3)); + v_dst2_b = _mm_unpacklo_epi32(v_dst3_b, v_dst2_b); + v_dst0_b = _mm_unpacklo_epi64(v_dst2_b, v_dst0_b); + // Load the mask data + v_msk0_b = _mm_cvtsi32_si128(*(const uint32_t *)(msk + msk_stride * 0)); + v_msk1_b = _mm_cvtsi32_si128(*(const uint32_t *)(msk + msk_stride * 1)); + v_msk0_b = _mm_unpacklo_epi32(v_msk1_b, v_msk0_b); + v_msk2_b = _mm_cvtsi32_si128(*(const uint32_t *)(msk + msk_stride * 2)); + v_msk3_b = _mm_cvtsi32_si128(*(const uint32_t *)(msk + msk_stride * 3)); + v_msk2_b = _mm_unpacklo_epi32(v_msk3_b, v_msk2_b); + v_msk0_b = _mm_unpacklo_epi64(v_msk2_b, v_msk0_b); + // Apply the y filter + if (yoffset == HALF_PIXEL_OFFSET) { + v_src1_b = _mm_unpacklo_epi64(v_src3_b, v_src1_b); + v_src2_b = + _mm_or_si128(_mm_slli_si128(v_src1_b, 4), + _mm_and_si128(v_src0_b, _mm_setr_epi32(-1, 0, 0, 0))); + v_res_b = _mm_avg_epu8(v_src1_b, v_src2_b); + } else { + v_src2_b = + _mm_or_si128(_mm_slli_si128(v_src1_b, 4), + _mm_and_si128(v_src2_b, _mm_setr_epi32(-1, 0, 0, 0))); + apply_filter_lo(v_src1_b, v_src2_b, v_filter_b, &v_filtered1_w); + v_src2_b = + _mm_or_si128(_mm_slli_si128(v_src3_b, 4), + _mm_and_si128(v_src0_b, _mm_setr_epi32(-1, 0, 0, 0))); + apply_filter_lo(v_src3_b, v_src2_b, v_filter_b, &v_filtered2_w); + v_res_b = _mm_packus_epi16(v_filtered2_w, v_filtered1_w); + } + // Compute the sum and SSE + sum_and_sse(v_res_b, v_dst0_b, v_msk0_b, &v_sum_d, &v_sse_q); + // Move onto the next set of rows + src += src_stride * 4; + dst += dst_stride * 4; + msk += msk_stride * 4; + } + return calc_masked_variance(v_sum_d, v_sse_q, sse, 4, h); +} + +// Note order in which rows loaded xmm[127:64] = row 1, xmm[63:0] = row 2 +unsigned int aom_masked_subpel_var8xH_xzero(const uint8_t *src, int src_stride, + int yoffset, const uint8_t *dst, + int dst_stride, const uint8_t *msk, + int msk_stride, unsigned int *sse, + int h) { + int i; + __m128i v_src0_b, v_src1_b, v_filtered0_w, v_filtered1_w, v_res_b; + __m128i v_dst_b = _mm_setzero_si128(); + __m128i v_msk_b = _mm_setzero_si128(); + __m128i v_sum_d = _mm_setzero_si128(); + __m128i v_sse_q = _mm_setzero_si128(); + __m128i v_filter_b = _mm_set1_epi16((bilinear_filters_2t[yoffset][1] << 8) + + bilinear_filters_2t[yoffset][0]); + assert(yoffset < BIL_SUBPEL_SHIFTS); + // Load the first row of src data ready + v_src0_b = _mm_loadl_epi64((const __m128i *)src); + for (i = 0; i < h; i += 2) { + if (yoffset == HALF_PIXEL_OFFSET) { + // Load the rest of the source data for these rows + v_src1_b = _mm_or_si128( + _mm_slli_si128(v_src0_b, 8), + _mm_loadl_epi64((const __m128i *)(src + src_stride * 1))); + v_src0_b = _mm_or_si128( + _mm_slli_si128(v_src1_b, 8), + _mm_loadl_epi64((const __m128i *)(src + src_stride * 2))); + // Apply the y filter + v_res_b = _mm_avg_epu8(v_src1_b, v_src0_b); + } else { + // Load the data and apply the y filter + v_src1_b = _mm_loadl_epi64((const __m128i *)(src + src_stride * 1)); + apply_filter_lo(v_src0_b, v_src1_b, v_filter_b, &v_filtered0_w); + v_src0_b = _mm_loadl_epi64((const __m128i *)(src + src_stride * 2)); + apply_filter_lo(v_src1_b, v_src0_b, v_filter_b, &v_filtered1_w); + v_res_b = _mm_packus_epi16(v_filtered1_w, v_filtered0_w); + } + // Load the dst data + v_dst_b = _mm_unpacklo_epi64( + _mm_loadl_epi64((const __m128i *)(dst + dst_stride * 1)), + _mm_loadl_epi64((const __m128i *)(dst + dst_stride * 0))); + // Load the mask data + v_msk_b = _mm_unpacklo_epi64( + _mm_loadl_epi64((const __m128i *)(msk + msk_stride * 1)), + _mm_loadl_epi64((const __m128i *)(msk + msk_stride * 0))); + // Compute the sum and SSE + sum_and_sse(v_res_b, v_dst_b, v_msk_b, &v_sum_d, &v_sse_q); + // Move onto the next set of rows + src += src_stride * 2; + dst += dst_stride * 2; + msk += msk_stride * 2; + } + return calc_masked_variance(v_sum_d, v_sse_q, sse, 8, h); +} + +// Note order in which rows loaded xmm[127:96] = row 1, xmm[95:64] = row 2, +// xmm[63:32] = row 3, xmm[31:0] = row 4 +unsigned int aom_masked_subpel_var4xH_yzero(const uint8_t *src, int src_stride, + int xoffset, const uint8_t *dst, + int dst_stride, const uint8_t *msk, + int msk_stride, unsigned int *sse, + int h) { + int i; + __m128i v_src0_b, v_src1_b, v_src2_b, v_src3_b, v_filtered0_w, v_filtered2_w; + __m128i v_src0_shift_b, v_src1_shift_b, v_src2_shift_b, v_src3_shift_b; + __m128i v_dst0_b, v_dst1_b, v_dst2_b, v_dst3_b; + __m128i v_msk0_b, v_msk1_b, v_msk2_b, v_msk3_b, v_res_b; + __m128i v_sum_d = _mm_setzero_si128(); + __m128i v_sse_q = _mm_setzero_si128(); + __m128i v_filter_b = _mm_set1_epi16((bilinear_filters_2t[xoffset][1] << 8) + + bilinear_filters_2t[xoffset][0]); + assert(xoffset < BIL_SUBPEL_SHIFTS); + for (i = 0; i < h; i += 4) { + // Load the src data + v_src0_b = _mm_loadl_epi64((const __m128i *)src); + v_src0_shift_b = _mm_srli_si128(v_src0_b, 1); + v_src1_b = _mm_loadl_epi64((const __m128i *)(src + src_stride * 1)); + v_src0_b = _mm_unpacklo_epi32(v_src1_b, v_src0_b); + v_src1_shift_b = _mm_srli_si128(v_src1_b, 1); + v_src2_b = _mm_loadl_epi64((const __m128i *)(src + src_stride * 2)); + v_src0_shift_b = _mm_unpacklo_epi32(v_src1_shift_b, v_src0_shift_b); + v_src2_shift_b = _mm_srli_si128(v_src2_b, 1); + v_src3_b = _mm_loadl_epi64((const __m128i *)(src + src_stride * 3)); + v_src2_b = _mm_unpacklo_epi32(v_src3_b, v_src2_b); + v_src3_shift_b = _mm_srli_si128(v_src3_b, 1); + v_src2_shift_b = _mm_unpacklo_epi32(v_src3_shift_b, v_src2_shift_b); + // Load the dst data + v_dst0_b = _mm_cvtsi32_si128(*(const uint32_t *)(dst + dst_stride * 0)); + v_dst1_b = _mm_cvtsi32_si128(*(const uint32_t *)(dst + dst_stride * 1)); + v_dst0_b = _mm_unpacklo_epi32(v_dst1_b, v_dst0_b); + v_dst2_b = _mm_cvtsi32_si128(*(const uint32_t *)(dst + dst_stride * 2)); + v_dst3_b = _mm_cvtsi32_si128(*(const uint32_t *)(dst + dst_stride * 3)); + v_dst2_b = _mm_unpacklo_epi32(v_dst3_b, v_dst2_b); + v_dst0_b = _mm_unpacklo_epi64(v_dst2_b, v_dst0_b); + // Load the mask data + v_msk0_b = _mm_cvtsi32_si128(*(const uint32_t *)(msk + msk_stride * 0)); + v_msk1_b = _mm_cvtsi32_si128(*(const uint32_t *)(msk + msk_stride * 1)); + v_msk0_b = _mm_unpacklo_epi32(v_msk1_b, v_msk0_b); + v_msk2_b = _mm_cvtsi32_si128(*(const uint32_t *)(msk + msk_stride * 2)); + v_msk3_b = _mm_cvtsi32_si128(*(const uint32_t *)(msk + msk_stride * 3)); + v_msk2_b = _mm_unpacklo_epi32(v_msk3_b, v_msk2_b); + v_msk0_b = _mm_unpacklo_epi64(v_msk2_b, v_msk0_b); + // Apply the x filter + if (xoffset == HALF_PIXEL_OFFSET) { + v_src0_b = _mm_unpacklo_epi64(v_src2_b, v_src0_b); + v_src0_shift_b = _mm_unpacklo_epi64(v_src2_shift_b, v_src0_shift_b); + v_res_b = _mm_avg_epu8(v_src0_b, v_src0_shift_b); + } else { + apply_filter_lo(v_src0_b, v_src0_shift_b, v_filter_b, &v_filtered0_w); + apply_filter_lo(v_src2_b, v_src2_shift_b, v_filter_b, &v_filtered2_w); + v_res_b = _mm_packus_epi16(v_filtered2_w, v_filtered0_w); + } + // Compute the sum and SSE + sum_and_sse(v_res_b, v_dst0_b, v_msk0_b, &v_sum_d, &v_sse_q); + // Move onto the next set of rows + src += src_stride * 4; + dst += dst_stride * 4; + msk += msk_stride * 4; + } + return calc_masked_variance(v_sum_d, v_sse_q, sse, 4, h); +} + +unsigned int aom_masked_subpel_var8xH_yzero(const uint8_t *src, int src_stride, + int xoffset, const uint8_t *dst, + int dst_stride, const uint8_t *msk, + int msk_stride, unsigned int *sse, + int h) { + int i; + __m128i v_src0_b, v_src1_b, v_filtered0_w, v_filtered1_w; + __m128i v_src0_shift_b, v_src1_shift_b, v_res_b, v_dst_b, v_msk_b; + __m128i v_sum_d = _mm_setzero_si128(); + __m128i v_sse_q = _mm_setzero_si128(); + __m128i v_filter_b = _mm_set1_epi16((bilinear_filters_2t[xoffset][1] << 8) + + bilinear_filters_2t[xoffset][0]); + assert(xoffset < BIL_SUBPEL_SHIFTS); + for (i = 0; i < h; i += 2) { + // Load the src data + v_src0_b = _mm_loadu_si128((const __m128i *)(src)); + v_src0_shift_b = _mm_srli_si128(v_src0_b, 1); + v_src1_b = _mm_loadu_si128((const __m128i *)(src + src_stride)); + v_src1_shift_b = _mm_srli_si128(v_src1_b, 1); + // Apply the x filter + if (xoffset == HALF_PIXEL_OFFSET) { + v_src1_b = _mm_unpacklo_epi64(v_src0_b, v_src1_b); + v_src1_shift_b = _mm_unpacklo_epi64(v_src0_shift_b, v_src1_shift_b); + v_res_b = _mm_avg_epu8(v_src1_b, v_src1_shift_b); + } else { + apply_filter_lo(v_src0_b, v_src0_shift_b, v_filter_b, &v_filtered0_w); + apply_filter_lo(v_src1_b, v_src1_shift_b, v_filter_b, &v_filtered1_w); + v_res_b = _mm_packus_epi16(v_filtered0_w, v_filtered1_w); + } + // Load the dst data + v_dst_b = _mm_unpacklo_epi64( + _mm_loadl_epi64((const __m128i *)(dst + dst_stride * 0)), + _mm_loadl_epi64((const __m128i *)(dst + dst_stride * 1))); + // Load the mask data + v_msk_b = _mm_unpacklo_epi64( + _mm_loadl_epi64((const __m128i *)(msk + msk_stride * 0)), + _mm_loadl_epi64((const __m128i *)(msk + msk_stride * 1))); + // Compute the sum and SSE + sum_and_sse(v_res_b, v_dst_b, v_msk_b, &v_sum_d, &v_sse_q); + // Move onto the next set of rows + src += src_stride * 2; + dst += dst_stride * 2; + msk += msk_stride * 2; + } + return calc_masked_variance(v_sum_d, v_sse_q, sse, 8, h); +} + +// Note order in which rows loaded xmm[127:96] = row 1, xmm[95:64] = row 2, +// xmm[63:32] = row 3, xmm[31:0] = row 4 +unsigned int aom_masked_subpel_var4xH_xnonzero_ynonzero( + const uint8_t *src, int src_stride, int xoffset, int yoffset, + const uint8_t *dst, int dst_stride, const uint8_t *msk, int msk_stride, + unsigned int *sse, int h) { + int i; + __m128i v_src0_b, v_src1_b, v_src2_b, v_src3_b, v_filtered0_w, v_filtered2_w; + __m128i v_src0_shift_b, v_src1_shift_b, v_src2_shift_b, v_src3_shift_b; + __m128i v_dst0_b, v_dst1_b, v_dst2_b, v_dst3_b, v_temp_b; + __m128i v_msk0_b, v_msk1_b, v_msk2_b, v_msk3_b, v_extra_row_b, v_res_b; + __m128i v_xres_b[2]; + __m128i v_sum_d = _mm_setzero_si128(); + __m128i v_sse_q = _mm_setzero_si128(); + __m128i v_filterx_b = _mm_set1_epi16((bilinear_filters_2t[xoffset][1] << 8) + + bilinear_filters_2t[xoffset][0]); + __m128i v_filtery_b = _mm_set1_epi16((bilinear_filters_2t[yoffset][1] << 8) + + bilinear_filters_2t[yoffset][0]); + assert(xoffset < BIL_SUBPEL_SHIFTS); + assert(yoffset < BIL_SUBPEL_SHIFTS); + for (i = 0; i < h; i += 4) { + // Load the src data + v_src0_b = _mm_loadl_epi64((const __m128i *)src); + v_src0_shift_b = _mm_srli_si128(v_src0_b, 1); + v_src1_b = _mm_loadl_epi64((const __m128i *)(src + src_stride * 1)); + v_src0_b = _mm_unpacklo_epi32(v_src1_b, v_src0_b); + v_src1_shift_b = _mm_srli_si128(v_src1_b, 1); + v_src2_b = _mm_loadl_epi64((const __m128i *)(src + src_stride * 2)); + v_src0_shift_b = _mm_unpacklo_epi32(v_src1_shift_b, v_src0_shift_b); + v_src2_shift_b = _mm_srli_si128(v_src2_b, 1); + v_src3_b = _mm_loadl_epi64((const __m128i *)(src + src_stride * 3)); + v_src2_b = _mm_unpacklo_epi32(v_src3_b, v_src2_b); + v_src3_shift_b = _mm_srli_si128(v_src3_b, 1); + v_src2_shift_b = _mm_unpacklo_epi32(v_src3_shift_b, v_src2_shift_b); + // Apply the x filter + if (xoffset == HALF_PIXEL_OFFSET) { + v_src0_b = _mm_unpacklo_epi64(v_src2_b, v_src0_b); + v_src0_shift_b = _mm_unpacklo_epi64(v_src2_shift_b, v_src0_shift_b); + v_xres_b[i == 0 ? 0 : 1] = _mm_avg_epu8(v_src0_b, v_src0_shift_b); + } else { + apply_filter_lo(v_src0_b, v_src0_shift_b, v_filterx_b, &v_filtered0_w); + apply_filter_lo(v_src2_b, v_src2_shift_b, v_filterx_b, &v_filtered2_w); + v_xres_b[i == 0 ? 0 : 1] = _mm_packus_epi16(v_filtered2_w, v_filtered0_w); + } + // Move onto the next set of rows + src += src_stride * 4; + } + // Load one more row to be used in the y filter + v_src0_b = _mm_loadl_epi64((const __m128i *)src); + v_src0_shift_b = _mm_srli_si128(v_src0_b, 1); + // Apply the x filter + if (xoffset == HALF_PIXEL_OFFSET) { + v_extra_row_b = _mm_and_si128(_mm_avg_epu8(v_src0_b, v_src0_shift_b), + _mm_setr_epi32(-1, 0, 0, 0)); + } else { + apply_filter_lo(v_src0_b, v_src0_shift_b, v_filterx_b, &v_filtered0_w); + v_extra_row_b = + _mm_and_si128(_mm_packus_epi16(v_filtered0_w, _mm_setzero_si128()), + _mm_setr_epi32(-1, 0, 0, 0)); + } + + for (i = 0; i < h; i += 4) { + if (h == 8 && i == 0) { + v_temp_b = _mm_or_si128(_mm_slli_si128(v_xres_b[0], 4), + _mm_srli_si128(v_xres_b[1], 12)); + } else { + v_temp_b = _mm_or_si128(_mm_slli_si128(v_xres_b[i == 0 ? 0 : 1], 4), + v_extra_row_b); + } + // Apply the y filter + if (yoffset == HALF_PIXEL_OFFSET) { + v_res_b = _mm_avg_epu8(v_xres_b[i == 0 ? 0 : 1], v_temp_b); + } else { + v_res_b = apply_filter(v_xres_b[i == 0 ? 0 : 1], v_temp_b, v_filtery_b); + } + + // Load the dst data + v_dst0_b = _mm_cvtsi32_si128(*(const uint32_t *)(dst + dst_stride * 0)); + v_dst1_b = _mm_cvtsi32_si128(*(const uint32_t *)(dst + dst_stride * 1)); + v_dst0_b = _mm_unpacklo_epi32(v_dst1_b, v_dst0_b); + v_dst2_b = _mm_cvtsi32_si128(*(const uint32_t *)(dst + dst_stride * 2)); + v_dst3_b = _mm_cvtsi32_si128(*(const uint32_t *)(dst + dst_stride * 3)); + v_dst2_b = _mm_unpacklo_epi32(v_dst3_b, v_dst2_b); + v_dst0_b = _mm_unpacklo_epi64(v_dst2_b, v_dst0_b); + // Load the mask data + v_msk0_b = _mm_cvtsi32_si128(*(const uint32_t *)(msk + msk_stride * 0)); + v_msk1_b = _mm_cvtsi32_si128(*(const uint32_t *)(msk + msk_stride * 1)); + v_msk0_b = _mm_unpacklo_epi32(v_msk1_b, v_msk0_b); + v_msk2_b = _mm_cvtsi32_si128(*(const uint32_t *)(msk + msk_stride * 2)); + v_msk3_b = _mm_cvtsi32_si128(*(const uint32_t *)(msk + msk_stride * 3)); + v_msk2_b = _mm_unpacklo_epi32(v_msk3_b, v_msk2_b); + v_msk0_b = _mm_unpacklo_epi64(v_msk2_b, v_msk0_b); + // Compute the sum and SSE + sum_and_sse(v_res_b, v_dst0_b, v_msk0_b, &v_sum_d, &v_sse_q); + // Move onto the next set of rows + dst += dst_stride * 4; + msk += msk_stride * 4; + } + return calc_masked_variance(v_sum_d, v_sse_q, sse, 4, h); +} + +unsigned int aom_masked_subpel_var8xH_xnonzero_ynonzero( + const uint8_t *src, int src_stride, int xoffset, int yoffset, + const uint8_t *dst, int dst_stride, const uint8_t *msk, int msk_stride, + unsigned int *sse, int h) { + int i; + __m128i v_src0_b, v_src1_b, v_filtered0_w, v_filtered1_w, v_dst_b, v_msk_b; + __m128i v_src0_shift_b, v_src1_shift_b; + __m128i v_xres0_b, v_xres1_b, v_res_b, v_temp_b; + __m128i v_sum_d = _mm_setzero_si128(); + __m128i v_sse_q = _mm_setzero_si128(); + __m128i v_filterx_b = _mm_set1_epi16((bilinear_filters_2t[xoffset][1] << 8) + + bilinear_filters_2t[xoffset][0]); + __m128i v_filtery_b = _mm_set1_epi16((bilinear_filters_2t[yoffset][1] << 8) + + bilinear_filters_2t[yoffset][0]); + assert(xoffset < BIL_SUBPEL_SHIFTS); + assert(yoffset < BIL_SUBPEL_SHIFTS); + // Load the first block of src data + v_src0_b = _mm_loadu_si128((const __m128i *)(src)); + v_src0_shift_b = _mm_srli_si128(v_src0_b, 1); + v_src1_b = _mm_loadu_si128((const __m128i *)(src + src_stride)); + v_src1_shift_b = _mm_srli_si128(v_src1_b, 1); + // Apply the x filter + if (xoffset == HALF_PIXEL_OFFSET) { + v_src1_b = _mm_unpacklo_epi64(v_src0_b, v_src1_b); + v_src1_shift_b = _mm_unpacklo_epi64(v_src0_shift_b, v_src1_shift_b); + v_xres0_b = _mm_avg_epu8(v_src1_b, v_src1_shift_b); + } else { + apply_filter_lo(v_src0_b, v_src0_shift_b, v_filterx_b, &v_filtered0_w); + apply_filter_lo(v_src1_b, v_src1_shift_b, v_filterx_b, &v_filtered1_w); + v_xres0_b = _mm_packus_epi16(v_filtered0_w, v_filtered1_w); + } + for (i = 0; i < h; i += 4) { + // Load the next block of src data + v_src0_b = _mm_loadu_si128((const __m128i *)(src + src_stride * 2)); + v_src0_shift_b = _mm_srli_si128(v_src0_b, 1); + v_src1_b = _mm_loadu_si128((const __m128i *)(src + src_stride * 3)); + v_src1_shift_b = _mm_srli_si128(v_src1_b, 1); + // Apply the x filter + if (xoffset == HALF_PIXEL_OFFSET) { + v_src1_b = _mm_unpacklo_epi64(v_src0_b, v_src1_b); + v_src1_shift_b = _mm_unpacklo_epi64(v_src0_shift_b, v_src1_shift_b); + v_xres1_b = _mm_avg_epu8(v_src1_b, v_src1_shift_b); + } else { + apply_filter_lo(v_src0_b, v_src0_shift_b, v_filterx_b, &v_filtered0_w); + apply_filter_lo(v_src1_b, v_src1_shift_b, v_filterx_b, &v_filtered1_w); + v_xres1_b = _mm_packus_epi16(v_filtered0_w, v_filtered1_w); + } + // Apply the y filter to the previous block + v_temp_b = _mm_or_si128(_mm_srli_si128(v_xres0_b, 8), + _mm_slli_si128(v_xres1_b, 8)); + if (yoffset == HALF_PIXEL_OFFSET) { + v_res_b = _mm_avg_epu8(v_xres0_b, v_temp_b); + } else { + v_res_b = apply_filter(v_xres0_b, v_temp_b, v_filtery_b); + } + // Load the dst data + v_dst_b = _mm_unpacklo_epi64( + _mm_loadl_epi64((const __m128i *)(dst + dst_stride * 0)), + _mm_loadl_epi64((const __m128i *)(dst + dst_stride * 1))); + // Load the mask data + v_msk_b = _mm_unpacklo_epi64( + _mm_loadl_epi64((const __m128i *)(msk + msk_stride * 0)), + _mm_loadl_epi64((const __m128i *)(msk + msk_stride * 1))); + // Compute the sum and SSE + sum_and_sse(v_res_b, v_dst_b, v_msk_b, &v_sum_d, &v_sse_q); + + // Load the next block of src data + v_src0_b = _mm_loadu_si128((const __m128i *)(src + src_stride * 4)); + v_src0_shift_b = _mm_srli_si128(v_src0_b, 1); + v_src1_b = _mm_loadu_si128((const __m128i *)(src + src_stride * 5)); + v_src1_shift_b = _mm_srli_si128(v_src1_b, 1); + // Apply the x filter + if (xoffset == HALF_PIXEL_OFFSET) { + v_src1_b = _mm_unpacklo_epi64(v_src0_b, v_src1_b); + v_src1_shift_b = _mm_unpacklo_epi64(v_src0_shift_b, v_src1_shift_b); + v_xres0_b = _mm_avg_epu8(v_src1_b, v_src1_shift_b); + } else { + apply_filter_lo(v_src0_b, v_src0_shift_b, v_filterx_b, &v_filtered0_w); + apply_filter_lo(v_src1_b, v_src1_shift_b, v_filterx_b, &v_filtered1_w); + v_xres0_b = _mm_packus_epi16(v_filtered0_w, v_filtered1_w); + } + // Apply the y filter to the previous block + v_temp_b = _mm_or_si128(_mm_srli_si128(v_xres1_b, 8), + _mm_slli_si128(v_xres0_b, 8)); + if (yoffset == HALF_PIXEL_OFFSET) { + v_res_b = _mm_avg_epu8(v_xres1_b, v_temp_b); + } else { + v_res_b = apply_filter(v_xres1_b, v_temp_b, v_filtery_b); + } + // Load the dst data + v_dst_b = _mm_unpacklo_epi64( + _mm_loadl_epi64((const __m128i *)(dst + dst_stride * 2)), + _mm_loadl_epi64((const __m128i *)(dst + dst_stride * 3))); + // Load the mask data + v_msk_b = _mm_unpacklo_epi64( + _mm_loadl_epi64((const __m128i *)(msk + msk_stride * 2)), + _mm_loadl_epi64((const __m128i *)(msk + msk_stride * 3))); + // Compute the sum and SSE + sum_and_sse(v_res_b, v_dst_b, v_msk_b, &v_sum_d, &v_sse_q); + // Move onto the next set of rows + src += src_stride * 4; + dst += dst_stride * 4; + msk += msk_stride * 4; + } + return calc_masked_variance(v_sum_d, v_sse_q, sse, 8, h); +} + +// For W >=16 +#define MASK_SUBPIX_VAR_LARGE(W, H) \ + unsigned int aom_masked_sub_pixel_variance##W##x##H##_ssse3( \ + const uint8_t *src, int src_stride, int xoffset, int yoffset, \ + const uint8_t *dst, int dst_stride, const uint8_t *msk, int msk_stride, \ + unsigned int *sse) { \ + assert(W % 16 == 0); \ + if (xoffset == 0) { \ + if (yoffset == 0) \ + return aom_masked_variance##W##x##H##_ssse3( \ + src, src_stride, dst, dst_stride, msk, msk_stride, sse); \ + else if (yoffset == HALF_PIXEL_OFFSET) \ + return aom_masked_subpel_varWxH_xzero( \ + src, src_stride, HALF_PIXEL_OFFSET, dst, dst_stride, msk, \ + msk_stride, sse, W, H, apply_filter_avg); \ + else \ + return aom_masked_subpel_varWxH_xzero(src, src_stride, yoffset, dst, \ + dst_stride, msk, msk_stride, \ + sse, W, H, apply_filter); \ + } else if (yoffset == 0) { \ + if (xoffset == HALF_PIXEL_OFFSET) \ + return aom_masked_subpel_varWxH_yzero( \ + src, src_stride, HALF_PIXEL_OFFSET, dst, dst_stride, msk, \ + msk_stride, sse, W, H, apply_filter_avg); \ + else \ + return aom_masked_subpel_varWxH_yzero(src, src_stride, xoffset, dst, \ + dst_stride, msk, msk_stride, \ + sse, W, H, apply_filter); \ + } else if (xoffset == HALF_PIXEL_OFFSET) { \ + if (yoffset == HALF_PIXEL_OFFSET) \ + return aom_masked_subpel_varWxH_xnonzero_ynonzero( \ + src, src_stride, HALF_PIXEL_OFFSET, HALF_PIXEL_OFFSET, dst, \ + dst_stride, msk, msk_stride, sse, W, H, apply_filter_avg, \ + apply_filter_avg); \ + else \ + return aom_masked_subpel_varWxH_xnonzero_ynonzero( \ + src, src_stride, HALF_PIXEL_OFFSET, yoffset, dst, dst_stride, msk, \ + msk_stride, sse, W, H, apply_filter_avg, apply_filter); \ + } else { \ + if (yoffset == HALF_PIXEL_OFFSET) \ + return aom_masked_subpel_varWxH_xnonzero_ynonzero( \ + src, src_stride, xoffset, HALF_PIXEL_OFFSET, dst, dst_stride, msk, \ + msk_stride, sse, W, H, apply_filter, apply_filter_avg); \ + else \ + return aom_masked_subpel_varWxH_xnonzero_ynonzero( \ + src, src_stride, xoffset, yoffset, dst, dst_stride, msk, \ + msk_stride, sse, W, H, apply_filter, apply_filter); \ + } \ + } + +// For W < 16 +#define MASK_SUBPIX_VAR_SMALL(W, H) \ + unsigned int aom_masked_sub_pixel_variance##W##x##H##_ssse3( \ + const uint8_t *src, int src_stride, int xoffset, int yoffset, \ + const uint8_t *dst, int dst_stride, const uint8_t *msk, int msk_stride, \ + unsigned int *sse) { \ + assert(W == 4 || W == 8); \ + if (xoffset == 0 && yoffset == 0) \ + return aom_masked_variance##W##x##H##_ssse3( \ + src, src_stride, dst, dst_stride, msk, msk_stride, sse); \ + else if (xoffset == 0) \ + return aom_masked_subpel_var##W##xH_xzero( \ + src, src_stride, yoffset, dst, dst_stride, msk, msk_stride, sse, H); \ + else if (yoffset == 0) \ + return aom_masked_subpel_var##W##xH_yzero( \ + src, src_stride, xoffset, dst, dst_stride, msk, msk_stride, sse, H); \ + else \ + return aom_masked_subpel_var##W##xH_xnonzero_ynonzero( \ + src, src_stride, xoffset, yoffset, dst, dst_stride, msk, msk_stride, \ + sse, H); \ + } + +MASK_SUBPIX_VAR_SMALL(4, 4) +MASK_SUBPIX_VAR_SMALL(4, 8) +MASK_SUBPIX_VAR_SMALL(8, 4) +MASK_SUBPIX_VAR_SMALL(8, 8) +MASK_SUBPIX_VAR_SMALL(8, 16) +MASK_SUBPIX_VAR_LARGE(16, 8) +MASK_SUBPIX_VAR_LARGE(16, 16) +MASK_SUBPIX_VAR_LARGE(16, 32) +MASK_SUBPIX_VAR_LARGE(32, 16) +MASK_SUBPIX_VAR_LARGE(32, 32) +MASK_SUBPIX_VAR_LARGE(32, 64) +MASK_SUBPIX_VAR_LARGE(64, 32) +MASK_SUBPIX_VAR_LARGE(64, 64) +#if CONFIG_EXT_PARTITION +MASK_SUBPIX_VAR_LARGE(64, 128) +MASK_SUBPIX_VAR_LARGE(128, 64) +MASK_SUBPIX_VAR_LARGE(128, 128) +#endif // CONFIG_EXT_PARTITION + +#if CONFIG_HIGHBITDEPTH +typedef uint32_t (*highbd_calc_masked_var_t)(__m128i v_sum_d, __m128i v_sse_q, + uint32_t *sse, int w, int h); +typedef unsigned int (*highbd_variance_fn_t)(const uint8_t *a8, int a_stride, + const uint8_t *b8, int b_stride, + const uint8_t *m, int m_stride, + unsigned int *sse); +typedef __m128i (*highbd_filter_fn_t)(__m128i v_a_w, __m128i v_b_w, + __m128i v_filter_w); + +static INLINE __m128i highbd_apply_filter_avg(const __m128i v_a_w, + const __m128i v_b_w, + const __m128i v_filter_w) { + (void)v_filter_w; + return _mm_avg_epu16(v_a_w, v_b_w); +} + +static INLINE __m128i highbd_apply_filter(const __m128i v_a_w, + const __m128i v_b_w, + const __m128i v_filter_w) { + const __m128i v_rounding_d = _mm_set1_epi32(1 << (FILTER_BITS - 1)); + __m128i v_input_lo_w = _mm_unpacklo_epi16(v_a_w, v_b_w); + __m128i v_input_hi_w = _mm_unpackhi_epi16(v_a_w, v_b_w); + __m128i v_temp0_d = _mm_madd_epi16(v_input_lo_w, v_filter_w); + __m128i v_temp1_d = _mm_madd_epi16(v_input_hi_w, v_filter_w); + __m128i v_res_lo_d = + _mm_srai_epi32(_mm_add_epi32(v_temp0_d, v_rounding_d), FILTER_BITS); + __m128i v_res_hi_d = + _mm_srai_epi32(_mm_add_epi32(v_temp1_d, v_rounding_d), FILTER_BITS); + return _mm_packs_epi32(v_res_lo_d, v_res_hi_d); +} +// Apply the filter to the contents of the lower half of a and b +static INLINE void highbd_apply_filter_lo(const __m128i v_a_lo_w, + const __m128i v_b_lo_w, + const __m128i v_filter_w, + __m128i *v_res_d) { + const __m128i v_rounding_d = _mm_set1_epi32(1 << (FILTER_BITS - 1)); + __m128i v_input_w = _mm_unpacklo_epi16(v_a_lo_w, v_b_lo_w); + __m128i v_temp0_d = _mm_madd_epi16(v_input_w, v_filter_w); + *v_res_d = + _mm_srai_epi32(_mm_add_epi32(v_temp0_d, v_rounding_d), FILTER_BITS); +} + +static void highbd_sum_and_sse(const __m128i v_a_w, const __m128i v_b_w, + const __m128i v_m_b, __m128i *v_sum_d, + __m128i *v_sse_q) { + const __m128i v_zero = _mm_setzero_si128(); + const __m128i v_m_w = _mm_unpacklo_epi8(v_m_b, v_zero); + + // Difference: [-2^12, 2^12] => 13 bits (incld sign bit) + const __m128i v_d_w = _mm_sub_epi16(v_a_w, v_b_w); + + // Error - [-4095, 4095] * [0, 64] & sum pairs => fits in 19 + 1 bits + const __m128i v_e_d = _mm_madd_epi16(v_d_w, v_m_w); + + // Squared error - max (18 bits * 18 bits) = 36 bits (no sign bit) + const __m128i v_absd_w = _mm_abs_epi16(v_d_w); + const __m128i v_dlo_d = _mm_unpacklo_epi16(v_absd_w, v_zero); + const __m128i v_mlo_d = _mm_unpacklo_epi16(v_m_w, v_zero); + const __m128i v_elo_d = _mm_madd_epi16(v_dlo_d, v_mlo_d); + const __m128i v_dhi_d = _mm_unpackhi_epi16(v_absd_w, v_zero); + const __m128i v_mhi_d = _mm_unpackhi_epi16(v_m_w, v_zero); + const __m128i v_ehi_d = _mm_madd_epi16(v_dhi_d, v_mhi_d); + // Square and sum the errors -> 36bits * 4 = 38bits + __m128i v_se0_q, v_se1_q, v_se2_q, v_se3_q, v_se_q, v_elo1_d, v_ehi3_d; + v_se0_q = _mm_mul_epu32(v_elo_d, v_elo_d); + v_elo1_d = _mm_srli_si128(v_elo_d, 4); + v_se1_q = _mm_mul_epu32(v_elo1_d, v_elo1_d); + v_se0_q = _mm_add_epi64(v_se0_q, v_se1_q); + v_se2_q = _mm_mul_epu32(v_ehi_d, v_ehi_d); + v_ehi3_d = _mm_srli_si128(v_ehi_d, 4); + v_se3_q = _mm_mul_epu32(v_ehi3_d, v_ehi3_d); + v_se1_q = _mm_add_epi64(v_se2_q, v_se3_q); + v_se_q = _mm_add_epi64(v_se0_q, v_se1_q); + + // Accumulate + *v_sum_d = _mm_add_epi32(*v_sum_d, v_e_d); + *v_sse_q = _mm_add_epi64(*v_sse_q, v_se_q); +} + +static INLINE uint32_t highbd_10_calc_masked_variance(__m128i v_sum_d, + __m128i v_sse_q, + uint32_t *sse, int w, + int h) { + int64_t sum64; + uint64_t sse64; + + // Horizontal sum + sum64 = hsum_epi32_si32(v_sum_d); + sse64 = hsum_epi64_si64(v_sse_q); + + sum64 = (sum64 >= 0) ? sum64 : -sum64; + + // Round + sum64 = ROUND_POWER_OF_TWO(sum64, 6); + sse64 = ROUND_POWER_OF_TWO(sse64, 12); + + // Normalise + sum64 = ROUND_POWER_OF_TWO(sum64, 2); + sse64 = ROUND_POWER_OF_TWO(sse64, 4); + + // Store the SSE + *sse = (uint32_t)sse64; + // Compute the variance + return *sse - (uint32_t)((sum64 * sum64) / (w * h)); +} +static INLINE uint32_t highbd_12_calc_masked_variance(__m128i v_sum_d, + __m128i v_sse_q, + uint32_t *sse, int w, + int h) { + int64_t sum64; + uint64_t sse64; + + // Horizontal sum + sum64 = hsum_epi32_si64(v_sum_d); + sse64 = hsum_epi64_si64(v_sse_q); + + sum64 = (sum64 >= 0) ? sum64 : -sum64; + + // Round + sum64 = ROUND_POWER_OF_TWO(sum64, 6); + sse64 = ROUND_POWER_OF_TWO(sse64, 12); + + // Normalise + sum64 = ROUND_POWER_OF_TWO(sum64, 4); + sse64 = ROUND_POWER_OF_TWO(sse64, 8); + + // Store the SSE + *sse = (uint32_t)sse64; + // Compute the variance + return *sse - (uint32_t)((sum64 * sum64) / (w * h)); +} + +// High bit depth functions for width (W) >= 8 +unsigned int aom_highbd_masked_subpel_varWxH_xzero( + const uint16_t *src, int src_stride, int yoffset, const uint16_t *dst, + int dst_stride, const uint8_t *msk, int msk_stride, unsigned int *sse, + int w, int h, highbd_filter_fn_t filter_fn, + highbd_calc_masked_var_t calc_var) { + int i, j; + __m128i v_src0_w, v_src1_w, v_res_w, v_dst_w, v_msk_b; + __m128i v_sum_d = _mm_setzero_si128(); + __m128i v_sse_q = _mm_setzero_si128(); + const __m128i v_filter_w = + _mm_set1_epi32((bilinear_filters_2t[yoffset][1] << 16) + + bilinear_filters_2t[yoffset][0]); + assert(yoffset < BIL_SUBPEL_SHIFTS); + for (j = 0; j < w; j += 8) { + // Load the first row ready + v_src0_w = _mm_loadu_si128((const __m128i *)(src + j)); + // Process 2 rows at a time + for (i = 0; i < h; i += 2) { + // Load the next row apply the filter + v_src1_w = _mm_loadu_si128((const __m128i *)(src + j + src_stride)); + v_res_w = filter_fn(v_src0_w, v_src1_w, v_filter_w); + // Load the dst and msk for the variance calculation + v_dst_w = _mm_loadu_si128((const __m128i *)(dst + j)); + v_msk_b = _mm_loadl_epi64((const __m128i *)(msk + j)); + highbd_sum_and_sse(v_res_w, v_dst_w, v_msk_b, &v_sum_d, &v_sse_q); + + // Load the next row apply the filter + v_src0_w = _mm_loadu_si128((const __m128i *)(src + j + src_stride * 2)); + v_res_w = filter_fn(v_src1_w, v_src0_w, v_filter_w); + // Load the dst and msk for the variance calculation + v_dst_w = _mm_loadu_si128((const __m128i *)(dst + j + dst_stride)); + v_msk_b = _mm_loadl_epi64((const __m128i *)(msk + j + msk_stride)); + highbd_sum_and_sse(v_res_w, v_dst_w, v_msk_b, &v_sum_d, &v_sse_q); + // Move onto the next block of rows + src += src_stride * 2; + dst += dst_stride * 2; + msk += msk_stride * 2; + } + // Reset to the top of the block + src -= src_stride * h; + dst -= dst_stride * h; + msk -= msk_stride * h; + } + return calc_var(v_sum_d, v_sse_q, sse, w, h); +} +unsigned int aom_highbd_masked_subpel_varWxH_yzero( + const uint16_t *src, int src_stride, int xoffset, const uint16_t *dst, + int dst_stride, const uint8_t *msk, int msk_stride, unsigned int *sse, + int w, int h, highbd_filter_fn_t filter_fn, + highbd_calc_masked_var_t calc_var) { + int i, j; + __m128i v_src0_w, v_src1_w, v_res_w, v_dst_w, v_msk_b; + __m128i v_sum_d = _mm_setzero_si128(); + __m128i v_sse_q = _mm_setzero_si128(); + const __m128i v_filter_w = + _mm_set1_epi32((bilinear_filters_2t[xoffset][1] << 16) + + bilinear_filters_2t[xoffset][0]); + assert(xoffset < BIL_SUBPEL_SHIFTS); + for (i = 0; i < h; i++) { + for (j = 0; j < w; j += 8) { + // Load this row & apply the filter to them + v_src0_w = _mm_loadu_si128((const __m128i *)(src + j)); + v_src1_w = _mm_loadu_si128((const __m128i *)(src + j + 1)); + v_res_w = filter_fn(v_src0_w, v_src1_w, v_filter_w); + + // Load the dst and msk for the variance calculation + v_dst_w = _mm_loadu_si128((const __m128i *)(dst + j)); + v_msk_b = _mm_loadl_epi64((const __m128i *)(msk + j)); + highbd_sum_and_sse(v_res_w, v_dst_w, v_msk_b, &v_sum_d, &v_sse_q); + } + src += src_stride; + dst += dst_stride; + msk += msk_stride; + } + return calc_var(v_sum_d, v_sse_q, sse, w, h); +} + +unsigned int aom_highbd_masked_subpel_varWxH_xnonzero_ynonzero( + const uint16_t *src, int src_stride, int xoffset, int yoffset, + const uint16_t *dst, int dst_stride, const uint8_t *msk, int msk_stride, + unsigned int *sse, int w, int h, highbd_filter_fn_t xfilter_fn, + highbd_filter_fn_t yfilter_fn, highbd_calc_masked_var_t calc_var) { + int i, j; + __m128i v_src0_w, v_src1_w, v_src2_w, v_src3_w; + __m128i v_filtered0_w, v_filtered1_w, v_res_w, v_dst_w, v_msk_b; + __m128i v_sum_d = _mm_setzero_si128(); + __m128i v_sse_q = _mm_setzero_si128(); + const __m128i v_filterx_w = + _mm_set1_epi32((bilinear_filters_2t[xoffset][1] << 16) + + bilinear_filters_2t[xoffset][0]); + const __m128i v_filtery_w = + _mm_set1_epi32((bilinear_filters_2t[yoffset][1] << 16) + + bilinear_filters_2t[yoffset][0]); + assert(xoffset < BIL_SUBPEL_SHIFTS); + assert(yoffset < BIL_SUBPEL_SHIFTS); + for (j = 0; j < w; j += 8) { + // Load the first row ready + v_src0_w = _mm_loadu_si128((const __m128i *)(src + j)); + v_src1_w = _mm_loadu_si128((const __m128i *)(src + j + 1)); + v_filtered0_w = xfilter_fn(v_src0_w, v_src1_w, v_filterx_w); + // Process 2 rows at a time + for (i = 0; i < h; i += 2) { + // Load the next row & apply the filter + v_src2_w = _mm_loadu_si128((const __m128i *)(src + src_stride + j)); + v_src3_w = _mm_loadu_si128((const __m128i *)(src + src_stride + j + 1)); + v_filtered1_w = xfilter_fn(v_src2_w, v_src3_w, v_filterx_w); + // Load the dst and msk for the variance calculation + v_dst_w = _mm_loadu_si128((const __m128i *)(dst + j)); + v_msk_b = _mm_loadl_epi64((const __m128i *)(msk + j)); + // Complete the calculation for this row and add it to the running total + v_res_w = yfilter_fn(v_filtered0_w, v_filtered1_w, v_filtery_w); + highbd_sum_and_sse(v_res_w, v_dst_w, v_msk_b, &v_sum_d, &v_sse_q); + + // Load the next row & apply the filter + v_src0_w = _mm_loadu_si128((const __m128i *)(src + src_stride * 2 + j)); + v_src1_w = + _mm_loadu_si128((const __m128i *)(src + src_stride * 2 + j + 1)); + v_filtered0_w = xfilter_fn(v_src0_w, v_src1_w, v_filterx_w); + // Load the dst and msk for the variance calculation + v_dst_w = _mm_loadu_si128((const __m128i *)(dst + dst_stride + j)); + v_msk_b = _mm_loadl_epi64((const __m128i *)(msk + msk_stride + j)); + // Complete the calculation for this row and add it to the running total + v_res_w = yfilter_fn(v_filtered1_w, v_filtered0_w, v_filtery_w); + highbd_sum_and_sse(v_res_w, v_dst_w, v_msk_b, &v_sum_d, &v_sse_q); + // Move onto the next block of rows + src += src_stride * 2; + dst += dst_stride * 2; + msk += msk_stride * 2; + } + // Reset to the top of the block + src -= src_stride * h; + dst -= dst_stride * h; + msk -= msk_stride * h; + } + return calc_var(v_sum_d, v_sse_q, sse, w, h); +} + +// Note order in which rows loaded xmm[127:64] = row 1, xmm[63:0] = row 2 +unsigned int aom_highbd_masked_subpel_var4xH_xzero( + const uint16_t *src, int src_stride, int yoffset, const uint16_t *dst, + int dst_stride, const uint8_t *msk, int msk_stride, unsigned int *sse, + int h, highbd_calc_masked_var_t calc_var) { + int i; + __m128i v_src0_w, v_src1_w, v_filtered0_d, v_filtered1_d, v_res_w; + __m128i v_dst_w, v_msk_b; + __m128i v_sum_d = _mm_setzero_si128(); + __m128i v_sse_q = _mm_setzero_si128(); + __m128i v_filter_w = _mm_set1_epi32((bilinear_filters_2t[yoffset][1] << 16) + + bilinear_filters_2t[yoffset][0]); + assert(yoffset < BIL_SUBPEL_SHIFTS); + // Load the first row of src data ready + v_src0_w = _mm_loadl_epi64((const __m128i *)src); + for (i = 0; i < h; i += 2) { + if (yoffset == HALF_PIXEL_OFFSET) { + // Load the rest of the source data for these rows + v_src1_w = _mm_or_si128( + _mm_slli_si128(v_src0_w, 8), + _mm_loadl_epi64((const __m128i *)(src + src_stride * 1))); + v_src0_w = _mm_or_si128( + _mm_slli_si128(v_src1_w, 8), + _mm_loadl_epi64((const __m128i *)(src + src_stride * 2))); + // Apply the y filter + v_res_w = _mm_avg_epu16(v_src1_w, v_src0_w); + } else { + // Load the data and apply the y filter + v_src1_w = _mm_loadl_epi64((const __m128i *)(src + src_stride * 1)); + highbd_apply_filter_lo(v_src0_w, v_src1_w, v_filter_w, &v_filtered0_d); + v_src0_w = _mm_loadl_epi64((const __m128i *)(src + src_stride * 2)); + highbd_apply_filter_lo(v_src1_w, v_src0_w, v_filter_w, &v_filtered1_d); + v_res_w = _mm_packs_epi32(v_filtered1_d, v_filtered0_d); + } + // Load the dst data + v_dst_w = _mm_unpacklo_epi64( + _mm_loadl_epi64((const __m128i *)(dst + dst_stride * 1)), + _mm_loadl_epi64((const __m128i *)(dst + dst_stride * 0))); + // Load the mask data + v_msk_b = _mm_unpacklo_epi32( + _mm_loadl_epi64((const __m128i *)(msk + msk_stride * 1)), + _mm_loadl_epi64((const __m128i *)(msk + msk_stride * 0))); + // Compute the sum and SSE + highbd_sum_and_sse(v_res_w, v_dst_w, v_msk_b, &v_sum_d, &v_sse_q); + // Move onto the next set of rows + src += src_stride * 2; + dst += dst_stride * 2; + msk += msk_stride * 2; + } + return calc_var(v_sum_d, v_sse_q, sse, 4, h); +} + +unsigned int aom_highbd_masked_subpel_var4xH_yzero( + const uint16_t *src, int src_stride, int xoffset, const uint16_t *dst, + int dst_stride, const uint8_t *msk, int msk_stride, unsigned int *sse, + int h, highbd_calc_masked_var_t calc_var) { + int i; + __m128i v_src0_w, v_src1_w, v_filtered0_d, v_filtered1_d; + __m128i v_src0_shift_w, v_src1_shift_w, v_res_w, v_dst_w, v_msk_b; + __m128i v_sum_d = _mm_setzero_si128(); + __m128i v_sse_q = _mm_setzero_si128(); + __m128i v_filter_w = _mm_set1_epi32((bilinear_filters_2t[xoffset][1] << 16) + + bilinear_filters_2t[xoffset][0]); + assert(xoffset < BIL_SUBPEL_SHIFTS); + for (i = 0; i < h; i += 2) { + // Load the src data + v_src0_w = _mm_loadu_si128((const __m128i *)(src)); + v_src0_shift_w = _mm_srli_si128(v_src0_w, 2); + v_src1_w = _mm_loadu_si128((const __m128i *)(src + src_stride)); + v_src1_shift_w = _mm_srli_si128(v_src1_w, 2); + // Apply the x filter + if (xoffset == HALF_PIXEL_OFFSET) { + v_src1_w = _mm_unpacklo_epi64(v_src0_w, v_src1_w); + v_src1_shift_w = _mm_unpacklo_epi64(v_src0_shift_w, v_src1_shift_w); + v_res_w = _mm_avg_epu16(v_src1_w, v_src1_shift_w); + } else { + highbd_apply_filter_lo(v_src0_w, v_src0_shift_w, v_filter_w, + &v_filtered0_d); + highbd_apply_filter_lo(v_src1_w, v_src1_shift_w, v_filter_w, + &v_filtered1_d); + v_res_w = _mm_packs_epi32(v_filtered0_d, v_filtered1_d); + } + // Load the dst data + v_dst_w = _mm_unpacklo_epi64( + _mm_loadl_epi64((const __m128i *)(dst + dst_stride * 0)), + _mm_loadl_epi64((const __m128i *)(dst + dst_stride * 1))); + // Load the mask data + v_msk_b = _mm_unpacklo_epi32( + _mm_loadl_epi64((const __m128i *)(msk + msk_stride * 0)), + _mm_loadl_epi64((const __m128i *)(msk + msk_stride * 1))); + // Compute the sum and SSE + highbd_sum_and_sse(v_res_w, v_dst_w, v_msk_b, &v_sum_d, &v_sse_q); + // Move onto the next set of rows + src += src_stride * 2; + dst += dst_stride * 2; + msk += msk_stride * 2; + } + return calc_var(v_sum_d, v_sse_q, sse, 4, h); +} + +unsigned int aom_highbd_masked_subpel_var4xH_xnonzero_ynonzero( + const uint16_t *src, int src_stride, int xoffset, int yoffset, + const uint16_t *dst, int dst_stride, const uint8_t *msk, int msk_stride, + unsigned int *sse, int h, highbd_calc_masked_var_t calc_var) { + int i; + __m128i v_src0_w, v_src1_w, v_filtered0_d, v_filtered1_d, v_dst_w, v_msk_b; + __m128i v_src0_shift_w, v_src1_shift_w; + __m128i v_xres0_w, v_xres1_w, v_res_w, v_temp_w; + __m128i v_sum_d = _mm_setzero_si128(); + __m128i v_sse_q = _mm_setzero_si128(); + __m128i v_filterx_w = _mm_set1_epi32((bilinear_filters_2t[xoffset][1] << 16) + + bilinear_filters_2t[xoffset][0]); + __m128i v_filtery_w = _mm_set1_epi32((bilinear_filters_2t[yoffset][1] << 16) + + bilinear_filters_2t[yoffset][0]); + assert(xoffset < BIL_SUBPEL_SHIFTS); + assert(yoffset < BIL_SUBPEL_SHIFTS); + // Load the first block of src data + v_src0_w = _mm_loadu_si128((const __m128i *)(src)); + v_src0_shift_w = _mm_srli_si128(v_src0_w, 2); + v_src1_w = _mm_loadu_si128((const __m128i *)(src + src_stride)); + v_src1_shift_w = _mm_srli_si128(v_src1_w, 2); + // Apply the x filter + if (xoffset == HALF_PIXEL_OFFSET) { + v_src1_w = _mm_unpacklo_epi64(v_src0_w, v_src1_w); + v_src1_shift_w = _mm_unpacklo_epi64(v_src0_shift_w, v_src1_shift_w); + v_xres0_w = _mm_avg_epu16(v_src1_w, v_src1_shift_w); + } else { + highbd_apply_filter_lo(v_src0_w, v_src0_shift_w, v_filterx_w, + &v_filtered0_d); + highbd_apply_filter_lo(v_src1_w, v_src1_shift_w, v_filterx_w, + &v_filtered1_d); + v_xres0_w = _mm_packs_epi32(v_filtered0_d, v_filtered1_d); + } + for (i = 0; i < h; i += 4) { + // Load the next block of src data + v_src0_w = _mm_loadu_si128((const __m128i *)(src + src_stride * 2)); + v_src0_shift_w = _mm_srli_si128(v_src0_w, 2); + v_src1_w = _mm_loadu_si128((const __m128i *)(src + src_stride * 3)); + v_src1_shift_w = _mm_srli_si128(v_src1_w, 2); + // Apply the x filter + if (xoffset == HALF_PIXEL_OFFSET) { + v_src1_w = _mm_unpacklo_epi64(v_src0_w, v_src1_w); + v_src1_shift_w = _mm_unpacklo_epi64(v_src0_shift_w, v_src1_shift_w); + v_xres1_w = _mm_avg_epu16(v_src1_w, v_src1_shift_w); + } else { + highbd_apply_filter_lo(v_src0_w, v_src0_shift_w, v_filterx_w, + &v_filtered0_d); + highbd_apply_filter_lo(v_src1_w, v_src1_shift_w, v_filterx_w, + &v_filtered1_d); + v_xres1_w = _mm_packs_epi32(v_filtered0_d, v_filtered1_d); + } + // Apply the y filter to the previous block + v_temp_w = _mm_or_si128(_mm_srli_si128(v_xres0_w, 8), + _mm_slli_si128(v_xres1_w, 8)); + if (yoffset == HALF_PIXEL_OFFSET) { + v_res_w = _mm_avg_epu16(v_xres0_w, v_temp_w); + } else { + v_res_w = highbd_apply_filter(v_xres0_w, v_temp_w, v_filtery_w); + } + // Load the dst data + v_dst_w = _mm_unpacklo_epi64( + _mm_loadl_epi64((const __m128i *)(dst + dst_stride * 0)), + _mm_loadl_epi64((const __m128i *)(dst + dst_stride * 1))); + // Load the mask data + v_msk_b = _mm_unpacklo_epi32( + _mm_loadl_epi64((const __m128i *)(msk + msk_stride * 0)), + _mm_loadl_epi64((const __m128i *)(msk + msk_stride * 1))); + // Compute the sum and SSE + highbd_sum_and_sse(v_res_w, v_dst_w, v_msk_b, &v_sum_d, &v_sse_q); + + // Load the next block of src data + v_src0_w = _mm_loadu_si128((const __m128i *)(src + src_stride * 4)); + v_src0_shift_w = _mm_srli_si128(v_src0_w, 2); + v_src1_w = _mm_loadu_si128((const __m128i *)(src + src_stride * 5)); + v_src1_shift_w = _mm_srli_si128(v_src1_w, 2); + // Apply the x filter + if (xoffset == HALF_PIXEL_OFFSET) { + v_src1_w = _mm_unpacklo_epi64(v_src0_w, v_src1_w); + v_src1_shift_w = _mm_unpacklo_epi64(v_src0_shift_w, v_src1_shift_w); + v_xres0_w = _mm_avg_epu16(v_src1_w, v_src1_shift_w); + } else { + highbd_apply_filter_lo(v_src0_w, v_src0_shift_w, v_filterx_w, + &v_filtered0_d); + highbd_apply_filter_lo(v_src1_w, v_src1_shift_w, v_filterx_w, + &v_filtered1_d); + v_xres0_w = _mm_packs_epi32(v_filtered0_d, v_filtered1_d); + } + // Apply the y filter to the previous block + v_temp_w = _mm_or_si128(_mm_srli_si128(v_xres1_w, 8), + _mm_slli_si128(v_xres0_w, 8)); + if (yoffset == HALF_PIXEL_OFFSET) { + v_res_w = _mm_avg_epu16(v_xres1_w, v_temp_w); + } else { + v_res_w = highbd_apply_filter(v_xres1_w, v_temp_w, v_filtery_w); + } + // Load the dst data + v_dst_w = _mm_unpacklo_epi64( + _mm_loadl_epi64((const __m128i *)(dst + dst_stride * 2)), + _mm_loadl_epi64((const __m128i *)(dst + dst_stride * 3))); + // Load the mask data + v_msk_b = _mm_unpacklo_epi32( + _mm_loadl_epi64((const __m128i *)(msk + msk_stride * 2)), + _mm_loadl_epi64((const __m128i *)(msk + msk_stride * 3))); + // Compute the sum and SSE + highbd_sum_and_sse(v_res_w, v_dst_w, v_msk_b, &v_sum_d, &v_sse_q); + // Move onto the next set of rows + src += src_stride * 4; + dst += dst_stride * 4; + msk += msk_stride * 4; + } + return calc_var(v_sum_d, v_sse_q, sse, 4, h); +} + +// For W >=8 +#define HIGHBD_MASK_SUBPIX_VAR_LARGE(W, H) \ + unsigned int highbd_masked_sub_pixel_variance##W##x##H##_ssse3( \ + const uint8_t *src8, int src_stride, int xoffset, int yoffset, \ + const uint8_t *dst8, int dst_stride, const uint8_t *msk, int msk_stride, \ + unsigned int *sse, highbd_calc_masked_var_t calc_var, \ + highbd_variance_fn_t full_variance_function) { \ + uint16_t *src = CONVERT_TO_SHORTPTR(src8); \ + uint16_t *dst = CONVERT_TO_SHORTPTR(dst8); \ + assert(W % 8 == 0); \ + if (xoffset == 0) { \ + if (yoffset == 0) \ + return full_variance_function(src8, src_stride, dst8, dst_stride, msk, \ + msk_stride, sse); \ + else if (yoffset == HALF_PIXEL_OFFSET) \ + return aom_highbd_masked_subpel_varWxH_xzero( \ + src, src_stride, HALF_PIXEL_OFFSET, dst, dst_stride, msk, \ + msk_stride, sse, W, H, highbd_apply_filter_avg, calc_var); \ + else \ + return aom_highbd_masked_subpel_varWxH_xzero( \ + src, src_stride, yoffset, dst, dst_stride, msk, msk_stride, sse, \ + W, H, highbd_apply_filter, calc_var); \ + } else if (yoffset == 0) { \ + if (xoffset == HALF_PIXEL_OFFSET) \ + return aom_highbd_masked_subpel_varWxH_yzero( \ + src, src_stride, HALF_PIXEL_OFFSET, dst, dst_stride, msk, \ + msk_stride, sse, W, H, highbd_apply_filter_avg, calc_var); \ + else \ + return aom_highbd_masked_subpel_varWxH_yzero( \ + src, src_stride, xoffset, dst, dst_stride, msk, msk_stride, sse, \ + W, H, highbd_apply_filter, calc_var); \ + } else if (xoffset == HALF_PIXEL_OFFSET) { \ + if (yoffset == HALF_PIXEL_OFFSET) \ + return aom_highbd_masked_subpel_varWxH_xnonzero_ynonzero( \ + src, src_stride, HALF_PIXEL_OFFSET, HALF_PIXEL_OFFSET, dst, \ + dst_stride, msk, msk_stride, sse, W, H, highbd_apply_filter_avg, \ + highbd_apply_filter_avg, calc_var); \ + else \ + return aom_highbd_masked_subpel_varWxH_xnonzero_ynonzero( \ + src, src_stride, HALF_PIXEL_OFFSET, yoffset, dst, dst_stride, msk, \ + msk_stride, sse, W, H, highbd_apply_filter_avg, \ + highbd_apply_filter, calc_var); \ + } else { \ + if (yoffset == HALF_PIXEL_OFFSET) \ + return aom_highbd_masked_subpel_varWxH_xnonzero_ynonzero( \ + src, src_stride, xoffset, HALF_PIXEL_OFFSET, dst, dst_stride, msk, \ + msk_stride, sse, W, H, highbd_apply_filter, \ + highbd_apply_filter_avg, calc_var); \ + else \ + return aom_highbd_masked_subpel_varWxH_xnonzero_ynonzero( \ + src, src_stride, xoffset, yoffset, dst, dst_stride, msk, \ + msk_stride, sse, W, H, highbd_apply_filter, highbd_apply_filter, \ + calc_var); \ + } \ + } + +// For W < 8 +#define HIGHBD_MASK_SUBPIX_VAR_SMALL(W, H) \ + unsigned int highbd_masked_sub_pixel_variance##W##x##H##_ssse3( \ + const uint8_t *src8, int src_stride, int xoffset, int yoffset, \ + const uint8_t *dst8, int dst_stride, const uint8_t *msk, int msk_stride, \ + unsigned int *sse, highbd_calc_masked_var_t calc_var, \ + highbd_variance_fn_t full_variance_function) { \ + uint16_t *src = CONVERT_TO_SHORTPTR(src8); \ + uint16_t *dst = CONVERT_TO_SHORTPTR(dst8); \ + assert(W == 4); \ + if (xoffset == 0 && yoffset == 0) \ + return full_variance_function(src8, src_stride, dst8, dst_stride, msk, \ + msk_stride, sse); \ + else if (xoffset == 0) \ + return aom_highbd_masked_subpel_var4xH_xzero( \ + src, src_stride, yoffset, dst, dst_stride, msk, msk_stride, sse, H, \ + calc_var); \ + else if (yoffset == 0) \ + return aom_highbd_masked_subpel_var4xH_yzero( \ + src, src_stride, xoffset, dst, dst_stride, msk, msk_stride, sse, H, \ + calc_var); \ + else \ + return aom_highbd_masked_subpel_var4xH_xnonzero_ynonzero( \ + src, src_stride, xoffset, yoffset, dst, dst_stride, msk, msk_stride, \ + sse, H, calc_var); \ + } + +#define HIGHBD_MASK_SUBPIX_VAR_WRAPPERS(W, H) \ + unsigned int aom_highbd_masked_sub_pixel_variance##W##x##H##_ssse3( \ + const uint8_t *src8, int src_stride, int xoffset, int yoffset, \ + const uint8_t *dst8, int dst_stride, const uint8_t *msk, int msk_stride, \ + unsigned int *sse) { \ + return highbd_masked_sub_pixel_variance##W##x##H##_ssse3( \ + src8, src_stride, xoffset, yoffset, dst8, dst_stride, msk, msk_stride, \ + sse, calc_masked_variance, \ + aom_highbd_masked_variance##W##x##H##_ssse3); \ + } \ + unsigned int aom_highbd_10_masked_sub_pixel_variance##W##x##H##_ssse3( \ + const uint8_t *src8, int src_stride, int xoffset, int yoffset, \ + const uint8_t *dst8, int dst_stride, const uint8_t *msk, int msk_stride, \ + unsigned int *sse) { \ + return highbd_masked_sub_pixel_variance##W##x##H##_ssse3( \ + src8, src_stride, xoffset, yoffset, dst8, dst_stride, msk, msk_stride, \ + sse, highbd_10_calc_masked_variance, \ + aom_highbd_10_masked_variance##W##x##H##_ssse3); \ + } \ + unsigned int aom_highbd_12_masked_sub_pixel_variance##W##x##H##_ssse3( \ + const uint8_t *src8, int src_stride, int xoffset, int yoffset, \ + const uint8_t *dst8, int dst_stride, const uint8_t *msk, int msk_stride, \ + unsigned int *sse) { \ + return highbd_masked_sub_pixel_variance##W##x##H##_ssse3( \ + src8, src_stride, xoffset, yoffset, dst8, dst_stride, msk, msk_stride, \ + sse, highbd_12_calc_masked_variance, \ + aom_highbd_12_masked_variance##W##x##H##_ssse3); \ + } + +HIGHBD_MASK_SUBPIX_VAR_SMALL(4, 4) +HIGHBD_MASK_SUBPIX_VAR_WRAPPERS(4, 4) +HIGHBD_MASK_SUBPIX_VAR_SMALL(4, 8) +HIGHBD_MASK_SUBPIX_VAR_WRAPPERS(4, 8) +HIGHBD_MASK_SUBPIX_VAR_LARGE(8, 4) +HIGHBD_MASK_SUBPIX_VAR_WRAPPERS(8, 4) +HIGHBD_MASK_SUBPIX_VAR_LARGE(8, 8) +HIGHBD_MASK_SUBPIX_VAR_WRAPPERS(8, 8) +HIGHBD_MASK_SUBPIX_VAR_LARGE(8, 16) +HIGHBD_MASK_SUBPIX_VAR_WRAPPERS(8, 16) +HIGHBD_MASK_SUBPIX_VAR_LARGE(16, 8) +HIGHBD_MASK_SUBPIX_VAR_WRAPPERS(16, 8) +HIGHBD_MASK_SUBPIX_VAR_LARGE(16, 16) +HIGHBD_MASK_SUBPIX_VAR_WRAPPERS(16, 16) +HIGHBD_MASK_SUBPIX_VAR_LARGE(16, 32) +HIGHBD_MASK_SUBPIX_VAR_WRAPPERS(16, 32) +HIGHBD_MASK_SUBPIX_VAR_LARGE(32, 16) +HIGHBD_MASK_SUBPIX_VAR_WRAPPERS(32, 16) +HIGHBD_MASK_SUBPIX_VAR_LARGE(32, 32) +HIGHBD_MASK_SUBPIX_VAR_WRAPPERS(32, 32) +HIGHBD_MASK_SUBPIX_VAR_LARGE(32, 64) +HIGHBD_MASK_SUBPIX_VAR_WRAPPERS(32, 64) +HIGHBD_MASK_SUBPIX_VAR_LARGE(64, 32) +HIGHBD_MASK_SUBPIX_VAR_WRAPPERS(64, 32) +HIGHBD_MASK_SUBPIX_VAR_LARGE(64, 64) +HIGHBD_MASK_SUBPIX_VAR_WRAPPERS(64, 64) +#if CONFIG_EXT_PARTITION +HIGHBD_MASK_SUBPIX_VAR_LARGE(64, 128) +HIGHBD_MASK_SUBPIX_VAR_WRAPPERS(64, 128) +HIGHBD_MASK_SUBPIX_VAR_LARGE(128, 64) +HIGHBD_MASK_SUBPIX_VAR_WRAPPERS(128, 64) +HIGHBD_MASK_SUBPIX_VAR_LARGE(128, 128) +HIGHBD_MASK_SUBPIX_VAR_WRAPPERS(128, 128) +#endif // CONFIG_EXT_PARTITION +#endif diff --git a/third_party/aom/aom_dsp/x86/obmc_sad_sse4.c b/third_party/aom/aom_dsp/x86/obmc_sad_sse4.c new file mode 100644 index 000000000..ad77f974c --- /dev/null +++ b/third_party/aom/aom_dsp/x86/obmc_sad_sse4.c @@ -0,0 +1,262 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include <assert.h> +#include <immintrin.h> + +#include "./aom_config.h" +#include "aom_ports/mem.h" +#include "aom/aom_integer.h" + +#include "aom_dsp/aom_dsp_common.h" +#include "aom_dsp/x86/synonyms.h" + +//////////////////////////////////////////////////////////////////////////////// +// 8 bit +//////////////////////////////////////////////////////////////////////////////// + +static INLINE unsigned int obmc_sad_w4(const uint8_t *pre, const int pre_stride, + const int32_t *wsrc, const int32_t *mask, + const int height) { + const int pre_step = pre_stride - 4; + int n = 0; + __m128i v_sad_d = _mm_setzero_si128(); + + do { + const __m128i v_p_b = xx_loadl_32(pre + n); + const __m128i v_m_d = xx_load_128(mask + n); + const __m128i v_w_d = xx_load_128(wsrc + n); + + const __m128i v_p_d = _mm_cvtepu8_epi32(v_p_b); + + // Values in both pre and mask fit in 15 bits, and are packed at 32 bit + // boundaries. We use pmaddwd, as it has lower latency on Haswell + // than pmulld but produces the same result with these inputs. + const __m128i v_pm_d = _mm_madd_epi16(v_p_d, v_m_d); + + const __m128i v_diff_d = _mm_sub_epi32(v_w_d, v_pm_d); + const __m128i v_absdiff_d = _mm_abs_epi32(v_diff_d); + + // Rounded absolute difference + const __m128i v_rad_d = xx_roundn_epu32(v_absdiff_d, 12); + + v_sad_d = _mm_add_epi32(v_sad_d, v_rad_d); + + n += 4; + + if (n % 4 == 0) pre += pre_step; + } while (n < 4 * height); + + return xx_hsum_epi32_si32(v_sad_d); +} + +static INLINE unsigned int obmc_sad_w8n(const uint8_t *pre, + const int pre_stride, + const int32_t *wsrc, + const int32_t *mask, const int width, + const int height) { + const int pre_step = pre_stride - width; + int n = 0; + __m128i v_sad_d = _mm_setzero_si128(); + + assert(width >= 8); + assert(IS_POWER_OF_TWO(width)); + + do { + const __m128i v_p1_b = xx_loadl_32(pre + n + 4); + const __m128i v_m1_d = xx_load_128(mask + n + 4); + const __m128i v_w1_d = xx_load_128(wsrc + n + 4); + const __m128i v_p0_b = xx_loadl_32(pre + n); + const __m128i v_m0_d = xx_load_128(mask + n); + const __m128i v_w0_d = xx_load_128(wsrc + n); + + const __m128i v_p0_d = _mm_cvtepu8_epi32(v_p0_b); + const __m128i v_p1_d = _mm_cvtepu8_epi32(v_p1_b); + + // Values in both pre and mask fit in 15 bits, and are packed at 32 bit + // boundaries. We use pmaddwd, as it has lower latency on Haswell + // than pmulld but produces the same result with these inputs. + const __m128i v_pm0_d = _mm_madd_epi16(v_p0_d, v_m0_d); + const __m128i v_pm1_d = _mm_madd_epi16(v_p1_d, v_m1_d); + + const __m128i v_diff0_d = _mm_sub_epi32(v_w0_d, v_pm0_d); + const __m128i v_diff1_d = _mm_sub_epi32(v_w1_d, v_pm1_d); + const __m128i v_absdiff0_d = _mm_abs_epi32(v_diff0_d); + const __m128i v_absdiff1_d = _mm_abs_epi32(v_diff1_d); + + // Rounded absolute difference + const __m128i v_rad0_d = xx_roundn_epu32(v_absdiff0_d, 12); + const __m128i v_rad1_d = xx_roundn_epu32(v_absdiff1_d, 12); + + v_sad_d = _mm_add_epi32(v_sad_d, v_rad0_d); + v_sad_d = _mm_add_epi32(v_sad_d, v_rad1_d); + + n += 8; + + if (n % width == 0) pre += pre_step; + } while (n < width * height); + + return xx_hsum_epi32_si32(v_sad_d); +} + +#define OBMCSADWXH(w, h) \ + unsigned int aom_obmc_sad##w##x##h##_sse4_1( \ + const uint8_t *pre, int pre_stride, const int32_t *wsrc, \ + const int32_t *msk) { \ + if (w == 4) { \ + return obmc_sad_w4(pre, pre_stride, wsrc, msk, h); \ + } else { \ + return obmc_sad_w8n(pre, pre_stride, wsrc, msk, w, h); \ + } \ + } + +#if CONFIG_EXT_PARTITION +OBMCSADWXH(128, 128) +OBMCSADWXH(128, 64) +OBMCSADWXH(64, 128) +#endif // CONFIG_EXT_PARTITION +OBMCSADWXH(64, 64) +OBMCSADWXH(64, 32) +OBMCSADWXH(32, 64) +OBMCSADWXH(32, 32) +OBMCSADWXH(32, 16) +OBMCSADWXH(16, 32) +OBMCSADWXH(16, 16) +OBMCSADWXH(16, 8) +OBMCSADWXH(8, 16) +OBMCSADWXH(8, 8) +OBMCSADWXH(8, 4) +OBMCSADWXH(4, 8) +OBMCSADWXH(4, 4) + +//////////////////////////////////////////////////////////////////////////////// +// High bit-depth +//////////////////////////////////////////////////////////////////////////////// + +#if CONFIG_HIGHBITDEPTH +static INLINE unsigned int hbd_obmc_sad_w4(const uint8_t *pre8, + const int pre_stride, + const int32_t *wsrc, + const int32_t *mask, + const int height) { + const uint16_t *pre = CONVERT_TO_SHORTPTR(pre8); + const int pre_step = pre_stride - 4; + int n = 0; + __m128i v_sad_d = _mm_setzero_si128(); + + do { + const __m128i v_p_w = xx_loadl_64(pre + n); + const __m128i v_m_d = xx_load_128(mask + n); + const __m128i v_w_d = xx_load_128(wsrc + n); + + const __m128i v_p_d = _mm_cvtepu16_epi32(v_p_w); + + // Values in both pre and mask fit in 15 bits, and are packed at 32 bit + // boundaries. We use pmaddwd, as it has lower latency on Haswell + // than pmulld but produces the same result with these inputs. + const __m128i v_pm_d = _mm_madd_epi16(v_p_d, v_m_d); + + const __m128i v_diff_d = _mm_sub_epi32(v_w_d, v_pm_d); + const __m128i v_absdiff_d = _mm_abs_epi32(v_diff_d); + + // Rounded absolute difference + const __m128i v_rad_d = xx_roundn_epu32(v_absdiff_d, 12); + + v_sad_d = _mm_add_epi32(v_sad_d, v_rad_d); + + n += 4; + + if (n % 4 == 0) pre += pre_step; + } while (n < 4 * height); + + return xx_hsum_epi32_si32(v_sad_d); +} + +static INLINE unsigned int hbd_obmc_sad_w8n(const uint8_t *pre8, + const int pre_stride, + const int32_t *wsrc, + const int32_t *mask, + const int width, const int height) { + const uint16_t *pre = CONVERT_TO_SHORTPTR(pre8); + const int pre_step = pre_stride - width; + int n = 0; + __m128i v_sad_d = _mm_setzero_si128(); + + assert(width >= 8); + assert(IS_POWER_OF_TWO(width)); + + do { + const __m128i v_p1_w = xx_loadl_64(pre + n + 4); + const __m128i v_m1_d = xx_load_128(mask + n + 4); + const __m128i v_w1_d = xx_load_128(wsrc + n + 4); + const __m128i v_p0_w = xx_loadl_64(pre + n); + const __m128i v_m0_d = xx_load_128(mask + n); + const __m128i v_w0_d = xx_load_128(wsrc + n); + + const __m128i v_p0_d = _mm_cvtepu16_epi32(v_p0_w); + const __m128i v_p1_d = _mm_cvtepu16_epi32(v_p1_w); + + // Values in both pre and mask fit in 15 bits, and are packed at 32 bit + // boundaries. We use pmaddwd, as it has lower latency on Haswell + // than pmulld but produces the same result with these inputs. + const __m128i v_pm0_d = _mm_madd_epi16(v_p0_d, v_m0_d); + const __m128i v_pm1_d = _mm_madd_epi16(v_p1_d, v_m1_d); + + const __m128i v_diff0_d = _mm_sub_epi32(v_w0_d, v_pm0_d); + const __m128i v_diff1_d = _mm_sub_epi32(v_w1_d, v_pm1_d); + const __m128i v_absdiff0_d = _mm_abs_epi32(v_diff0_d); + const __m128i v_absdiff1_d = _mm_abs_epi32(v_diff1_d); + + // Rounded absolute difference + const __m128i v_rad0_d = xx_roundn_epu32(v_absdiff0_d, 12); + const __m128i v_rad1_d = xx_roundn_epu32(v_absdiff1_d, 12); + + v_sad_d = _mm_add_epi32(v_sad_d, v_rad0_d); + v_sad_d = _mm_add_epi32(v_sad_d, v_rad1_d); + + n += 8; + + if (n % width == 0) pre += pre_step; + } while (n < width * height); + + return xx_hsum_epi32_si32(v_sad_d); +} + +#define HBD_OBMCSADWXH(w, h) \ + unsigned int aom_highbd_obmc_sad##w##x##h##_sse4_1( \ + const uint8_t *pre, int pre_stride, const int32_t *wsrc, \ + const int32_t *mask) { \ + if (w == 4) { \ + return hbd_obmc_sad_w4(pre, pre_stride, wsrc, mask, h); \ + } else { \ + return hbd_obmc_sad_w8n(pre, pre_stride, wsrc, mask, w, h); \ + } \ + } + +#if CONFIG_EXT_PARTITION +HBD_OBMCSADWXH(128, 128) +HBD_OBMCSADWXH(128, 64) +HBD_OBMCSADWXH(64, 128) +#endif // CONFIG_EXT_PARTITION +HBD_OBMCSADWXH(64, 64) +HBD_OBMCSADWXH(64, 32) +HBD_OBMCSADWXH(32, 64) +HBD_OBMCSADWXH(32, 32) +HBD_OBMCSADWXH(32, 16) +HBD_OBMCSADWXH(16, 32) +HBD_OBMCSADWXH(16, 16) +HBD_OBMCSADWXH(16, 8) +HBD_OBMCSADWXH(8, 16) +HBD_OBMCSADWXH(8, 8) +HBD_OBMCSADWXH(8, 4) +HBD_OBMCSADWXH(4, 8) +HBD_OBMCSADWXH(4, 4) +#endif // CONFIG_HIGHBITDEPTH diff --git a/third_party/aom/aom_dsp/x86/obmc_variance_sse4.c b/third_party/aom/aom_dsp/x86/obmc_variance_sse4.c new file mode 100644 index 000000000..efb3659cf --- /dev/null +++ b/third_party/aom/aom_dsp/x86/obmc_variance_sse4.c @@ -0,0 +1,355 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include <assert.h> +#include <immintrin.h> + +#include "./aom_config.h" +#include "aom_ports/mem.h" +#include "aom/aom_integer.h" + +#include "aom_dsp/aom_dsp_common.h" +#include "aom_dsp/x86/synonyms.h" +#include "aom_dsp/aom_filter.h" + +//////////////////////////////////////////////////////////////////////////////// +// 8 bit +//////////////////////////////////////////////////////////////////////////////// + +static INLINE void obmc_variance_w4(const uint8_t *pre, const int pre_stride, + const int32_t *wsrc, const int32_t *mask, + unsigned int *const sse, int *const sum, + const int h) { + const int pre_step = pre_stride - 4; + int n = 0; + __m128i v_sum_d = _mm_setzero_si128(); + __m128i v_sse_d = _mm_setzero_si128(); + + assert(IS_POWER_OF_TWO(h)); + + do { + const __m128i v_p_b = xx_loadl_32(pre + n); + const __m128i v_m_d = xx_load_128(mask + n); + const __m128i v_w_d = xx_load_128(wsrc + n); + + const __m128i v_p_d = _mm_cvtepu8_epi32(v_p_b); + + // Values in both pre and mask fit in 15 bits, and are packed at 32 bit + // boundaries. We use pmaddwd, as it has lower latency on Haswell + // than pmulld but produces the same result with these inputs. + const __m128i v_pm_d = _mm_madd_epi16(v_p_d, v_m_d); + + const __m128i v_diff_d = _mm_sub_epi32(v_w_d, v_pm_d); + const __m128i v_rdiff_d = xx_roundn_epi32(v_diff_d, 12); + const __m128i v_sqrdiff_d = _mm_mullo_epi32(v_rdiff_d, v_rdiff_d); + + v_sum_d = _mm_add_epi32(v_sum_d, v_rdiff_d); + v_sse_d = _mm_add_epi32(v_sse_d, v_sqrdiff_d); + + n += 4; + + if (n % 4 == 0) pre += pre_step; + } while (n < 4 * h); + + *sum = xx_hsum_epi32_si32(v_sum_d); + *sse = xx_hsum_epi32_si32(v_sse_d); +} + +static INLINE void obmc_variance_w8n(const uint8_t *pre, const int pre_stride, + const int32_t *wsrc, const int32_t *mask, + unsigned int *const sse, int *const sum, + const int w, const int h) { + const int pre_step = pre_stride - w; + int n = 0; + __m128i v_sum_d = _mm_setzero_si128(); + __m128i v_sse_d = _mm_setzero_si128(); + + assert(w >= 8); + assert(IS_POWER_OF_TWO(w)); + assert(IS_POWER_OF_TWO(h)); + + do { + const __m128i v_p1_b = xx_loadl_32(pre + n + 4); + const __m128i v_m1_d = xx_load_128(mask + n + 4); + const __m128i v_w1_d = xx_load_128(wsrc + n + 4); + const __m128i v_p0_b = xx_loadl_32(pre + n); + const __m128i v_m0_d = xx_load_128(mask + n); + const __m128i v_w0_d = xx_load_128(wsrc + n); + + const __m128i v_p0_d = _mm_cvtepu8_epi32(v_p0_b); + const __m128i v_p1_d = _mm_cvtepu8_epi32(v_p1_b); + + // Values in both pre and mask fit in 15 bits, and are packed at 32 bit + // boundaries. We use pmaddwd, as it has lower latency on Haswell + // than pmulld but produces the same result with these inputs. + const __m128i v_pm0_d = _mm_madd_epi16(v_p0_d, v_m0_d); + const __m128i v_pm1_d = _mm_madd_epi16(v_p1_d, v_m1_d); + + const __m128i v_diff0_d = _mm_sub_epi32(v_w0_d, v_pm0_d); + const __m128i v_diff1_d = _mm_sub_epi32(v_w1_d, v_pm1_d); + + const __m128i v_rdiff0_d = xx_roundn_epi32(v_diff0_d, 12); + const __m128i v_rdiff1_d = xx_roundn_epi32(v_diff1_d, 12); + const __m128i v_rdiff01_w = _mm_packs_epi32(v_rdiff0_d, v_rdiff1_d); + const __m128i v_sqrdiff_d = _mm_madd_epi16(v_rdiff01_w, v_rdiff01_w); + + v_sum_d = _mm_add_epi32(v_sum_d, v_rdiff0_d); + v_sum_d = _mm_add_epi32(v_sum_d, v_rdiff1_d); + v_sse_d = _mm_add_epi32(v_sse_d, v_sqrdiff_d); + + n += 8; + + if (n % w == 0) pre += pre_step; + } while (n < w * h); + + *sum = xx_hsum_epi32_si32(v_sum_d); + *sse = xx_hsum_epi32_si32(v_sse_d); +} + +#define OBMCVARWXH(W, H) \ + unsigned int aom_obmc_variance##W##x##H##_sse4_1( \ + const uint8_t *pre, int pre_stride, const int32_t *wsrc, \ + const int32_t *mask, unsigned int *sse) { \ + int sum; \ + if (W == 4) { \ + obmc_variance_w4(pre, pre_stride, wsrc, mask, sse, &sum, H); \ + } else { \ + obmc_variance_w8n(pre, pre_stride, wsrc, mask, sse, &sum, W, H); \ + } \ + return *sse - (unsigned int)(((int64_t)sum * sum) / (W * H)); \ + } + +#if CONFIG_EXT_PARTITION +OBMCVARWXH(128, 128) +OBMCVARWXH(128, 64) +OBMCVARWXH(64, 128) +#endif // CONFIG_EXT_PARTITION +OBMCVARWXH(64, 64) +OBMCVARWXH(64, 32) +OBMCVARWXH(32, 64) +OBMCVARWXH(32, 32) +OBMCVARWXH(32, 16) +OBMCVARWXH(16, 32) +OBMCVARWXH(16, 16) +OBMCVARWXH(16, 8) +OBMCVARWXH(8, 16) +OBMCVARWXH(8, 8) +OBMCVARWXH(8, 4) +OBMCVARWXH(4, 8) +OBMCVARWXH(4, 4) + +//////////////////////////////////////////////////////////////////////////////// +// High bit-depth +//////////////////////////////////////////////////////////////////////////////// + +#if CONFIG_HIGHBITDEPTH +static INLINE void hbd_obmc_variance_w4( + const uint8_t *pre8, const int pre_stride, const int32_t *wsrc, + const int32_t *mask, uint64_t *const sse, int64_t *const sum, const int h) { + const uint16_t *pre = CONVERT_TO_SHORTPTR(pre8); + const int pre_step = pre_stride - 4; + int n = 0; + __m128i v_sum_d = _mm_setzero_si128(); + __m128i v_sse_d = _mm_setzero_si128(); + + assert(IS_POWER_OF_TWO(h)); + + do { + const __m128i v_p_w = xx_loadl_64(pre + n); + const __m128i v_m_d = xx_load_128(mask + n); + const __m128i v_w_d = xx_load_128(wsrc + n); + + const __m128i v_p_d = _mm_cvtepu16_epi32(v_p_w); + + // Values in both pre and mask fit in 15 bits, and are packed at 32 bit + // boundaries. We use pmaddwd, as it has lower latency on Haswell + // than pmulld but produces the same result with these inputs. + const __m128i v_pm_d = _mm_madd_epi16(v_p_d, v_m_d); + + const __m128i v_diff_d = _mm_sub_epi32(v_w_d, v_pm_d); + const __m128i v_rdiff_d = xx_roundn_epi32(v_diff_d, 12); + const __m128i v_sqrdiff_d = _mm_mullo_epi32(v_rdiff_d, v_rdiff_d); + + v_sum_d = _mm_add_epi32(v_sum_d, v_rdiff_d); + v_sse_d = _mm_add_epi32(v_sse_d, v_sqrdiff_d); + + n += 4; + + if (n % 4 == 0) pre += pre_step; + } while (n < 4 * h); + + *sum = xx_hsum_epi32_si32(v_sum_d); + *sse = xx_hsum_epi32_si32(v_sse_d); +} + +static INLINE void hbd_obmc_variance_w8n( + const uint8_t *pre8, const int pre_stride, const int32_t *wsrc, + const int32_t *mask, uint64_t *const sse, int64_t *const sum, const int w, + const int h) { + const uint16_t *pre = CONVERT_TO_SHORTPTR(pre8); + const int pre_step = pre_stride - w; + int n = 0; + __m128i v_sum_d = _mm_setzero_si128(); + __m128i v_sse_d = _mm_setzero_si128(); + + assert(w >= 8); + assert(IS_POWER_OF_TWO(w)); + assert(IS_POWER_OF_TWO(h)); + + do { + const __m128i v_p1_w = xx_loadl_64(pre + n + 4); + const __m128i v_m1_d = xx_load_128(mask + n + 4); + const __m128i v_w1_d = xx_load_128(wsrc + n + 4); + const __m128i v_p0_w = xx_loadl_64(pre + n); + const __m128i v_m0_d = xx_load_128(mask + n); + const __m128i v_w0_d = xx_load_128(wsrc + n); + + const __m128i v_p0_d = _mm_cvtepu16_epi32(v_p0_w); + const __m128i v_p1_d = _mm_cvtepu16_epi32(v_p1_w); + + // Values in both pre and mask fit in 15 bits, and are packed at 32 bit + // boundaries. We use pmaddwd, as it has lower latency on Haswell + // than pmulld but produces the same result with these inputs. + const __m128i v_pm0_d = _mm_madd_epi16(v_p0_d, v_m0_d); + const __m128i v_pm1_d = _mm_madd_epi16(v_p1_d, v_m1_d); + + const __m128i v_diff0_d = _mm_sub_epi32(v_w0_d, v_pm0_d); + const __m128i v_diff1_d = _mm_sub_epi32(v_w1_d, v_pm1_d); + + const __m128i v_rdiff0_d = xx_roundn_epi32(v_diff0_d, 12); + const __m128i v_rdiff1_d = xx_roundn_epi32(v_diff1_d, 12); + const __m128i v_rdiff01_w = _mm_packs_epi32(v_rdiff0_d, v_rdiff1_d); + const __m128i v_sqrdiff_d = _mm_madd_epi16(v_rdiff01_w, v_rdiff01_w); + + v_sum_d = _mm_add_epi32(v_sum_d, v_rdiff0_d); + v_sum_d = _mm_add_epi32(v_sum_d, v_rdiff1_d); + v_sse_d = _mm_add_epi32(v_sse_d, v_sqrdiff_d); + + n += 8; + + if (n % w == 0) pre += pre_step; + } while (n < w * h); + + *sum += xx_hsum_epi32_si64(v_sum_d); + *sse += xx_hsum_epi32_si64(v_sse_d); +} + +static INLINE void highbd_obmc_variance(const uint8_t *pre8, int pre_stride, + const int32_t *wsrc, + const int32_t *mask, int w, int h, + unsigned int *sse, int *sum) { + int64_t sum64 = 0; + uint64_t sse64 = 0; + if (w == 4) { + hbd_obmc_variance_w4(pre8, pre_stride, wsrc, mask, &sse64, &sum64, h); + } else { + hbd_obmc_variance_w8n(pre8, pre_stride, wsrc, mask, &sse64, &sum64, w, h); + } + *sum = (int)sum64; + *sse = (unsigned int)sse64; +} + +static INLINE void highbd_10_obmc_variance(const uint8_t *pre8, int pre_stride, + const int32_t *wsrc, + const int32_t *mask, int w, int h, + unsigned int *sse, int *sum) { + int64_t sum64 = 0; + uint64_t sse64 = 0; + if (w == 4) { + hbd_obmc_variance_w4(pre8, pre_stride, wsrc, mask, &sse64, &sum64, h); + } else { + hbd_obmc_variance_w8n(pre8, pre_stride, wsrc, mask, &sse64, &sum64, w, h); + } + *sum = (int)ROUND_POWER_OF_TWO(sum64, 2); + *sse = (unsigned int)ROUND_POWER_OF_TWO(sse64, 4); +} + +static INLINE void highbd_12_obmc_variance(const uint8_t *pre8, int pre_stride, + const int32_t *wsrc, + const int32_t *mask, int w, int h, + unsigned int *sse, int *sum) { + int64_t sum64 = 0; + uint64_t sse64 = 0; + if (w == 128) { + do { + hbd_obmc_variance_w8n(pre8, pre_stride, wsrc, mask, &sse64, &sum64, 128, + 32); + pre8 += 32 * pre_stride; + wsrc += 32 * 128; + mask += 32 * 128; + h -= 32; + } while (h > 0); + } else if (w == 64 && h >= 128) { + do { + hbd_obmc_variance_w8n(pre8, pre_stride, wsrc, mask, &sse64, &sum64, 64, + 64); + pre8 += 64 * pre_stride; + wsrc += 64 * 64; + mask += 64 * 64; + h -= 64; + } while (h > 0); + } else if (w == 4) { + hbd_obmc_variance_w4(pre8, pre_stride, wsrc, mask, &sse64, &sum64, h); + } else { + hbd_obmc_variance_w8n(pre8, pre_stride, wsrc, mask, &sse64, &sum64, w, h); + } + *sum = (int)ROUND_POWER_OF_TWO(sum64, 4); + *sse = (unsigned int)ROUND_POWER_OF_TWO(sse64, 8); +} + +#define HBD_OBMCVARWXH(W, H) \ + unsigned int aom_highbd_obmc_variance##W##x##H##_sse4_1( \ + const uint8_t *pre, int pre_stride, const int32_t *wsrc, \ + const int32_t *mask, unsigned int *sse) { \ + int sum; \ + highbd_obmc_variance(pre, pre_stride, wsrc, mask, W, H, sse, &sum); \ + return *sse - (unsigned int)(((int64_t)sum * sum) / (W * H)); \ + } \ + \ + unsigned int aom_highbd_10_obmc_variance##W##x##H##_sse4_1( \ + const uint8_t *pre, int pre_stride, const int32_t *wsrc, \ + const int32_t *mask, unsigned int *sse) { \ + int sum; \ + int64_t var; \ + highbd_10_obmc_variance(pre, pre_stride, wsrc, mask, W, H, sse, &sum); \ + var = (int64_t)(*sse) - (((int64_t)sum * sum) / (W * H)); \ + return (var >= 0) ? (uint32_t)var : 0; \ + } \ + \ + unsigned int aom_highbd_12_obmc_variance##W##x##H##_sse4_1( \ + const uint8_t *pre, int pre_stride, const int32_t *wsrc, \ + const int32_t *mask, unsigned int *sse) { \ + int sum; \ + int64_t var; \ + highbd_12_obmc_variance(pre, pre_stride, wsrc, mask, W, H, sse, &sum); \ + var = (int64_t)(*sse) - (((int64_t)sum * sum) / (W * H)); \ + return (var >= 0) ? (uint32_t)var : 0; \ + } + +#if CONFIG_EXT_PARTITION +HBD_OBMCVARWXH(128, 128) +HBD_OBMCVARWXH(128, 64) +HBD_OBMCVARWXH(64, 128) +#endif // CONFIG_EXT_PARTITION +HBD_OBMCVARWXH(64, 64) +HBD_OBMCVARWXH(64, 32) +HBD_OBMCVARWXH(32, 64) +HBD_OBMCVARWXH(32, 32) +HBD_OBMCVARWXH(32, 16) +HBD_OBMCVARWXH(16, 32) +HBD_OBMCVARWXH(16, 16) +HBD_OBMCVARWXH(16, 8) +HBD_OBMCVARWXH(8, 16) +HBD_OBMCVARWXH(8, 8) +HBD_OBMCVARWXH(8, 4) +HBD_OBMCVARWXH(4, 8) +HBD_OBMCVARWXH(4, 4) +#endif // CONFIG_HIGHBITDEPTH diff --git a/third_party/aom/aom_dsp/x86/quantize_avx_x86_64.asm b/third_party/aom/aom_dsp/x86/quantize_avx_x86_64.asm new file mode 100644 index 000000000..954a95b98 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/quantize_avx_x86_64.asm @@ -0,0 +1,547 @@ +; +; Copyright (c) 2016, Alliance for Open Media. All rights reserved +; +; This source code is subject to the terms of the BSD 2 Clause License and +; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License +; was not distributed with this source code in the LICENSE file, you can +; obtain it at www.aomedia.org/license/software. If the Alliance for Open +; Media Patent License 1.0 was not distributed with this source code in the +; PATENTS file, you can obtain it at www.aomedia.org/license/patent. +; + +; + +%include "third_party/x86inc/x86inc.asm" + +SECTION .text + +%macro QUANTIZE_FN 2 +cglobal quantize_%1, 0, %2, 15, coeff, ncoeff, skip, zbin, round, quant, \ + shift, qcoeff, dqcoeff, dequant, \ + eob, scan, iscan + + vzeroupper + + ; If we can skip this block, then just zero the output + cmp skipmp, 0 + jne .blank + +%ifnidn %1, b_32x32 + + ; Special case for ncoeff == 16, as it is frequent and we can save on + ; not setting up a loop. + cmp ncoeffmp, 16 + jne .generic + + ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;; + ;; Special case of ncoeff == 16 + ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;; + +.single: + + movifnidn coeffq, coeffmp + movifnidn zbinq, zbinmp + mova m0, [zbinq] ; m0 = zbin + + ; Get DC and first 15 AC coeffs - in this special case, that is all. +%if CONFIG_HIGHBITDEPTH + ; coeff stored as 32bit numbers but we process them as 16 bit numbers + mova m9, [coeffq] + packssdw m9, [coeffq+16] ; m9 = c[i] + mova m10, [coeffq+32] + packssdw m10, [coeffq+48] ; m10 = c[i] +%else + mova m9, [coeffq] ; m9 = c[i] + mova m10, [coeffq+16] ; m10 = c[i] +%endif + + mov r0, eobmp ; Output pointer + mov r1, qcoeffmp ; Output pointer + mov r2, dqcoeffmp ; Output pointer + + pxor m5, m5 ; m5 = dedicated zero + + pcmpeqw m4, m4 ; All word lanes -1 + paddw m0, m4 ; m0 = zbin - 1 + + pabsw m6, m9 ; m6 = abs(m9) + pabsw m11, m10 ; m11 = abs(m10) + pcmpgtw m7, m6, m0 ; m7 = c[i] >= zbin + punpckhqdq m0, m0 + pcmpgtw m12, m11, m0 ; m12 = c[i] >= zbin + + ; Check if all coeffs are less than zbin. If yes, we just write zeros + ; to the outputs and we are done. + por m14, m7, m12 + ptest m14, m14 + jnz .single_nonzero + +%if CONFIG_HIGHBITDEPTH + mova [r1 ], ymm5 + mova [r1+32], ymm5 + mova [r2 ], ymm5 + mova [r2+32], ymm5 +%else + mova [r1], ymm5 + mova [r2], ymm5 +%endif + mov [r0], word 0 + + vzeroupper + RET + +.single_nonzero: + + ; Actual quantization of size 16 block - setup pointers, rounders, etc. + movifnidn r4, roundmp + movifnidn r5, quantmp + mov r3, dequantmp + mov r6, shiftmp + mova m1, [r4] ; m1 = round + mova m2, [r5] ; m2 = quant + mova m3, [r3] ; m3 = dequant + mova m4, [r6] ; m4 = shift + + mov r3, iscanmp + + DEFINE_ARGS eob, qcoeff, dqcoeff, iscan + + ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;; + + paddsw m6, m1 ; m6 += round + punpckhqdq m1, m1 + paddsw m11, m1 ; m11 += round + pmulhw m8, m6, m2 ; m8 = m6*q>>16 + punpckhqdq m2, m2 + pmulhw m13, m11, m2 ; m13 = m11*q>>16 + paddw m8, m6 ; m8 += m6 + paddw m13, m11 ; m13 += m11 + pmulhw m8, m4 ; m8 = m8*qsh>>16 + punpckhqdq m4, m4 + pmulhw m13, m4 ; m13 = m13*qsh>>16 + psignw m8, m9 ; m8 = reinsert sign + psignw m13, m10 ; m13 = reinsert sign + pand m8, m7 + pand m13, m12 + +%if CONFIG_HIGHBITDEPTH + ; Store 16bit numbers as 32bit numbers in array pointed to by qcoeff + pcmpgtw m6, m5, m8 + punpckhwd m6, m8, m6 + pmovsxwd m11, m8 + mova [qcoeffq ], m11 + mova [qcoeffq+16], m6 + pcmpgtw m6, m5, m13 + punpckhwd m6, m13, m6 + pmovsxwd m11, m13 + mova [qcoeffq+32], m11 + mova [qcoeffq+48], m6 +%else + mova [qcoeffq ], m8 + mova [qcoeffq+16], m13 +%endif + + pmullw m8, m3 ; dqc[i] = qc[i] * q + punpckhqdq m3, m3 + pmullw m13, m3 ; dqc[i] = qc[i] * q + +%if CONFIG_HIGHBITDEPTH + ; Store 16bit numbers as 32bit numbers in array pointed to by qcoeff + pcmpgtw m6, m5, m8 + punpckhwd m6, m8, m6 + pmovsxwd m11, m8 + mova [dqcoeffq ], m11 + mova [dqcoeffq+16], m6 + pcmpgtw m6, m5, m13 + punpckhwd m6, m13, m6 + pmovsxwd m11, m13 + mova [dqcoeffq+32], m11 + mova [dqcoeffq+48], m6 +%else + mova [dqcoeffq ], m8 + mova [dqcoeffq+16], m13 +%endif + + mova m6, [iscanq] ; m6 = scan[i] + mova m11, [iscanq+16] ; m11 = scan[i] + + pcmpeqw m8, m8, m5 ; m8 = c[i] == 0 + pcmpeqw m13, m13, m5 ; m13 = c[i] == 0 + psubw m6, m6, m7 ; m6 = scan[i] + 1 + psubw m11, m11, m12 ; m11 = scan[i] + 1 + pandn m8, m8, m6 ; m8 = max(eob) + pandn m13, m13, m11 ; m13 = max(eob) + pmaxsw m8, m8, m13 + + ; Horizontally accumulate/max eobs and write into [eob] memory pointer + pshufd m7, m8, 0xe + pmaxsw m8, m7 + pshuflw m7, m8, 0xe + pmaxsw m8, m7 + pshuflw m7, m8, 0x1 + pmaxsw m8, m7 + movq rax, m8 + mov [eobq], ax + + vzeroupper + RET + + ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;; + ;; Generic case of ncoeff != 16 + ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;; + +.generic: + +%endif ; %ifnidn %1, b_32x32 + +DEFINE_ARGS coeff, ncoeff, skip, zbin, round, quant, shift, \ + qcoeff, dqcoeff, dequant, eob, scan, iscan + + ; Actual quantization loop - setup pointers, rounders, etc. + movifnidn coeffq, coeffmp + movifnidn ncoeffq, ncoeffmp + mov r2, dequantmp + movifnidn zbinq, zbinmp + movifnidn roundq, roundmp + movifnidn quantq, quantmp + mova m0, [zbinq] ; m0 = zbin + mova m1, [roundq] ; m1 = round + mova m2, [quantq] ; m2 = quant + mova m3, [r2] ; m3 = dequant + pcmpeqw m4, m4 ; All lanes -1 +%ifidn %1, b_32x32 + psubw m0, m4 + psubw m1, m4 + psrlw m0, 1 ; m0 = (m0 + 1) / 2 + psrlw m1, 1 ; m1 = (m1 + 1) / 2 +%endif + paddw m0, m4 ; m0 = m0 + 1 + + mov r2, shiftmp + mov r3, qcoeffmp + mova m4, [r2] ; m4 = shift + mov r4, dqcoeffmp + mov r5, iscanmp +%ifidn %1, b_32x32 + psllw m4, 1 +%endif + pxor m5, m5 ; m5 = dedicated zero + + DEFINE_ARGS coeff, ncoeff, d1, qcoeff, dqcoeff, iscan, d2, d3, d4, d5, eob + +%if CONFIG_HIGHBITDEPTH + lea coeffq, [ coeffq+ncoeffq*4] + lea qcoeffq, [ qcoeffq+ncoeffq*4] + lea dqcoeffq, [dqcoeffq+ncoeffq*4] +%else + lea coeffq, [ coeffq+ncoeffq*2] + lea qcoeffq, [ qcoeffq+ncoeffq*2] + lea dqcoeffq, [dqcoeffq+ncoeffq*2] +%endif + lea iscanq, [ iscanq+ncoeffq*2] + neg ncoeffq + + ; get DC and first 15 AC coeffs +%if CONFIG_HIGHBITDEPTH + ; coeff stored as 32bit numbers & require 16bit numbers + mova m9, [coeffq+ncoeffq*4+ 0] + packssdw m9, [coeffq+ncoeffq*4+16] + mova m10, [coeffq+ncoeffq*4+32] + packssdw m10, [coeffq+ncoeffq*4+48] +%else + mova m9, [coeffq+ncoeffq*2+ 0] ; m9 = c[i] + mova m10, [coeffq+ncoeffq*2+16] ; m10 = c[i] +%endif + + pabsw m6, m9 ; m6 = abs(m9) + pabsw m11, m10 ; m11 = abs(m10) + pcmpgtw m7, m6, m0 ; m7 = c[i] >= zbin + punpckhqdq m0, m0 + pcmpgtw m12, m11, m0 ; m12 = c[i] >= zbin + + ; Check if all coeffs are less than zbin. If yes, skip forward quickly. + por m14, m7, m12 + ptest m14, m14 + jnz .first_nonzero + +%if CONFIG_HIGHBITDEPTH + mova [qcoeffq+ncoeffq*4 ], ymm5 + mova [qcoeffq+ncoeffq*4+32], ymm5 + mova [dqcoeffq+ncoeffq*4 ], ymm5 + mova [dqcoeffq+ncoeffq*4+32], ymm5 +%else + mova [qcoeffq+ncoeffq*2], ymm5 + mova [dqcoeffq+ncoeffq*2], ymm5 +%endif + + add ncoeffq, mmsize + + punpckhqdq m1, m1 + punpckhqdq m2, m2 + punpckhqdq m3, m3 + punpckhqdq m4, m4 + pxor m8, m8 + + jmp .ac_only_loop + +.first_nonzero: + + paddsw m6, m1 ; m6 += round + punpckhqdq m1, m1 + paddsw m11, m1 ; m11 += round + pmulhw m8, m6, m2 ; m8 = m6*q>>16 + punpckhqdq m2, m2 + pmulhw m13, m11, m2 ; m13 = m11*q>>16 + paddw m8, m6 ; m8 += m6 + paddw m13, m11 ; m13 += m11 + pmulhw m8, m4 ; m8 = m8*qsh>>16 + punpckhqdq m4, m4 + pmulhw m13, m4 ; m13 = m13*qsh>>16 + psignw m8, m9 ; m8 = reinsert sign + psignw m13, m10 ; m13 = reinsert sign + pand m8, m7 + pand m13, m12 + +%if CONFIG_HIGHBITDEPTH + ; store 16bit numbers as 32bit numbers in array pointed to by qcoeff + pcmpgtw m6, m5, m8 + punpckhwd m6, m8, m6 + pmovsxwd m11, m8 + mova [qcoeffq+ncoeffq*4+ 0], m11 + mova [qcoeffq+ncoeffq*4+16], m6 + pcmpgtw m6, m5, m13 + punpckhwd m6, m13, m6 + pmovsxwd m11, m13 + mova [qcoeffq+ncoeffq*4+32], m11 + mova [qcoeffq+ncoeffq*4+48], m6 +%else + mova [qcoeffq+ncoeffq*2+ 0], m8 + mova [qcoeffq+ncoeffq*2+16], m13 +%endif + +%ifidn %1, b_32x32 + pabsw m8, m8 + pabsw m13, m13 +%endif + pmullw m8, m3 ; dqc[i] = qc[i] * q + punpckhqdq m3, m3 + pmullw m13, m3 ; dqc[i] = qc[i] * q +%ifidn %1, b_32x32 + psrlw m8, 1 + psrlw m13, 1 + psignw m8, m9 + psignw m13, m10 +%endif + +%if CONFIG_HIGHBITDEPTH + ; store 16bit numbers as 32bit numbers in array pointed to by qcoeff + pcmpgtw m6, m5, m8 + punpckhwd m6, m8, m6 + pmovsxwd m11, m8 + mova [dqcoeffq+ncoeffq*4+ 0], m11 + mova [dqcoeffq+ncoeffq*4+16], m6 + pcmpgtw m6, m5, m13 + punpckhwd m6, m13, m6 + pmovsxwd m11, m13 + mova [dqcoeffq+ncoeffq*4+32], m11 + mova [dqcoeffq+ncoeffq*4+48], m6 +%else + mova [dqcoeffq+ncoeffq*2+ 0], m8 + mova [dqcoeffq+ncoeffq*2+16], m13 +%endif + + pcmpeqw m8, m5 ; m8 = c[i] == 0 + pcmpeqw m13, m5 ; m13 = c[i] == 0 + mova m6, [iscanq+ncoeffq*2] ; m6 = scan[i] + mova m11, [iscanq+ncoeffq*2+16] ; m11 = scan[i] + psubw m6, m7 ; m6 = scan[i] + 1 + psubw m11, m12 ; m11 = scan[i] + 1 + pandn m8, m6 ; m8 = max(eob) + pandn m13, m11 ; m13 = max(eob) + pmaxsw m8, m13 + add ncoeffq, mmsize + +.ac_only_loop: + +%if CONFIG_HIGHBITDEPTH + ; pack coeff from 32bit to 16bit array + mova m9, [coeffq+ncoeffq*4+ 0] + packssdw m9, [coeffq+ncoeffq*4+16] + mova m10, [coeffq+ncoeffq*4+32] + packssdw m10, [coeffq+ncoeffq*4+48] +%else + mova m9, [coeffq+ncoeffq*2+ 0] ; m9 = c[i] + mova m10, [coeffq+ncoeffq*2+16] ; m10 = c[i] +%endif + + pabsw m6, m9 ; m6 = abs(m9) + pabsw m11, m10 ; m11 = abs(m10) + pcmpgtw m7, m6, m0 ; m7 = c[i] >= zbin + pcmpgtw m12, m11, m0 ; m12 = c[i] >= zbin + + ; Check if all coeffs are less than zbin. If yes, skip this itertion. + ; And just write zeros as the result would be. + por m14, m7, m12 + ptest m14, m14 + jnz .rest_nonzero + +%if CONFIG_HIGHBITDEPTH + mova [qcoeffq+ncoeffq*4+ 0], ymm5 + mova [qcoeffq+ncoeffq*4+32], ymm5 + mova [dqcoeffq+ncoeffq*4+ 0], ymm5 + mova [dqcoeffq+ncoeffq*4+32], ymm5 +%else + mova [qcoeffq+ncoeffq*2+ 0], ymm5 + mova [dqcoeffq+ncoeffq*2+ 0], ymm5 +%endif + add ncoeffq, mmsize + jnz .ac_only_loop + + ; Horizontally accumulate/max eobs and write into [eob] memory pointer + mov r2, eobmp + pshufd m7, m8, 0xe + pmaxsw m8, m7 + pshuflw m7, m8, 0xe + pmaxsw m8, m7 + pshuflw m7, m8, 0x1 + pmaxsw m8, m7 + movq rax, m8 + mov [r2], ax + vzeroupper + RET + +.rest_nonzero: + paddsw m6, m1 ; m6 += round + paddsw m11, m1 ; m11 += round + pmulhw m14, m6, m2 ; m14 = m6*q>>16 + pmulhw m13, m11, m2 ; m13 = m11*q>>16 + paddw m14, m6 ; m14 += m6 + paddw m13, m11 ; m13 += m11 + pmulhw m14, m4 ; m14 = m14*qsh>>16 + pmulhw m13, m4 ; m13 = m13*qsh>>16 + psignw m14, m9 ; m14 = reinsert sign + psignw m13, m10 ; m13 = reinsert sign + pand m14, m7 + pand m13, m12 + +%if CONFIG_HIGHBITDEPTH + ; store 16bit numbers as 32bit numbers in array pointed to by qcoeff + pcmpgtw m6, m5, m14 + punpckhwd m6, m14, m6 + pmovsxwd m11, m14 + mova [qcoeffq+ncoeffq*4+ 0], m11 + mova [qcoeffq+ncoeffq*4+16], m6 + pcmpgtw m6, m5, m13 + punpckhwd m6, m13, m6 + pmovsxwd m11, m13 + mova [qcoeffq+ncoeffq*4+32], m11 + mova [qcoeffq+ncoeffq*4+48], m6 +%else + mova [qcoeffq+ncoeffq*2+ 0], m14 + mova [qcoeffq+ncoeffq*2+16], m13 +%endif + +%ifidn %1, b_32x32 + pabsw m14, m14 + pabsw m13, m13 +%endif + pmullw m14, m3 ; dqc[i] = qc[i] * q + pmullw m13, m3 ; dqc[i] = qc[i] * q +%ifidn %1, b_32x32 + psrlw m14, 1 + psrlw m13, 1 + psignw m14, m9 + psignw m13, m10 +%endif + +%if CONFIG_HIGHBITDEPTH + ; store 16bit numbers as 32bit numbers in array pointed to by qcoeff + pcmpgtw m6, m5, m14 + punpckhwd m6, m14, m6 + pmovsxwd m11, m14 + mova [dqcoeffq+ncoeffq*4+ 0], m11 + mova [dqcoeffq+ncoeffq*4+16], m6 + pcmpgtw m6, m5, m13 + punpckhwd m6, m13, m6 + pmovsxwd m11, m13 + mova [dqcoeffq+ncoeffq*4+32], m11 + mova [dqcoeffq+ncoeffq*4+48], m6 +%else + mova [dqcoeffq+ncoeffq*2+ 0], m14 + mova [dqcoeffq+ncoeffq*2+16], m13 +%endif + + pcmpeqw m14, m5 ; m14 = c[i] == 0 + pcmpeqw m13, m5 ; m13 = c[i] == 0 + mova m6, [iscanq+ncoeffq*2+ 0] ; m6 = scan[i] + mova m11, [iscanq+ncoeffq*2+16] ; m11 = scan[i] + psubw m6, m7 ; m6 = scan[i] + 1 + psubw m11, m12 ; m11 = scan[i] + 1 + pandn m14, m6 ; m14 = max(eob) + pandn m13, m11 ; m13 = max(eob) + pmaxsw m8, m14 + pmaxsw m8, m13 + add ncoeffq, mmsize + jnz .ac_only_loop + + ; Horizontally accumulate/max eobs and write into [eob] memory pointer + mov r2, eobmp + pshufd m7, m8, 0xe + pmaxsw m8, m7 + pshuflw m7, m8, 0xe + pmaxsw m8, m7 + pshuflw m7, m8, 0x1 + pmaxsw m8, m7 + movq rax, m8 + mov [r2], ax + vzeroupper + RET + + ; Skip-block, i.e. just write all zeroes +.blank: + +DEFINE_ARGS coeff, ncoeff, skip, zbin, round, quant, shift, \ + qcoeff, dqcoeff, dequant, eob, scan, iscan + + mov r0, dqcoeffmp + movifnidn ncoeffq, ncoeffmp + mov r2, qcoeffmp + mov r3, eobmp + +DEFINE_ARGS dqcoeff, ncoeff, qcoeff, eob + +%if CONFIG_HIGHBITDEPTH + lea dqcoeffq, [dqcoeffq+ncoeffq*4] + lea qcoeffq, [ qcoeffq+ncoeffq*4] +%else + lea dqcoeffq, [dqcoeffq+ncoeffq*2] + lea qcoeffq, [ qcoeffq+ncoeffq*2] +%endif + + neg ncoeffq + pxor m7, m7 + +.blank_loop: +%if CONFIG_HIGHBITDEPTH + mova [dqcoeffq+ncoeffq*4+ 0], ymm7 + mova [dqcoeffq+ncoeffq*4+32], ymm7 + mova [qcoeffq+ncoeffq*4+ 0], ymm7 + mova [qcoeffq+ncoeffq*4+32], ymm7 +%else + mova [dqcoeffq+ncoeffq*2+ 0], ymm7 + mova [qcoeffq+ncoeffq*2+ 0], ymm7 +%endif + add ncoeffq, mmsize + jl .blank_loop + + mov [eobq], word 0 + + vzeroupper + RET +%endmacro + +INIT_XMM avx +QUANTIZE_FN b, 7 +QUANTIZE_FN b_32x32, 7 + +END diff --git a/third_party/aom/aom_dsp/x86/quantize_sse2.c b/third_party/aom/aom_dsp/x86/quantize_sse2.c new file mode 100644 index 000000000..890c1f01e --- /dev/null +++ b/third_party/aom/aom_dsp/x86/quantize_sse2.c @@ -0,0 +1,249 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include <emmintrin.h> +#include <xmmintrin.h> + +#include "./aom_dsp_rtcd.h" +#include "aom/aom_integer.h" + +static INLINE __m128i load_coefficients(const tran_low_t *coeff_ptr) { +#if CONFIG_HIGHBITDEPTH + return _mm_setr_epi16((int16_t)coeff_ptr[0], (int16_t)coeff_ptr[1], + (int16_t)coeff_ptr[2], (int16_t)coeff_ptr[3], + (int16_t)coeff_ptr[4], (int16_t)coeff_ptr[5], + (int16_t)coeff_ptr[6], (int16_t)coeff_ptr[7]); +#else + return _mm_load_si128((const __m128i *)coeff_ptr); +#endif +} + +static INLINE void store_coefficients(__m128i coeff_vals, + tran_low_t *coeff_ptr) { +#if CONFIG_HIGHBITDEPTH + __m128i one = _mm_set1_epi16(1); + __m128i coeff_vals_hi = _mm_mulhi_epi16(coeff_vals, one); + __m128i coeff_vals_lo = _mm_mullo_epi16(coeff_vals, one); + __m128i coeff_vals_1 = _mm_unpacklo_epi16(coeff_vals_lo, coeff_vals_hi); + __m128i coeff_vals_2 = _mm_unpackhi_epi16(coeff_vals_lo, coeff_vals_hi); + _mm_store_si128((__m128i *)(coeff_ptr), coeff_vals_1); + _mm_store_si128((__m128i *)(coeff_ptr + 4), coeff_vals_2); +#else + _mm_store_si128((__m128i *)(coeff_ptr), coeff_vals); +#endif +} + +void aom_quantize_b_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, + int skip_block, const int16_t *zbin_ptr, + const int16_t *round_ptr, const int16_t *quant_ptr, + const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, + tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, + uint16_t *eob_ptr, const int16_t *scan_ptr, + const int16_t *iscan_ptr) { + __m128i zero; + (void)scan_ptr; + + coeff_ptr += n_coeffs; + iscan_ptr += n_coeffs; + qcoeff_ptr += n_coeffs; + dqcoeff_ptr += n_coeffs; + n_coeffs = -n_coeffs; + zero = _mm_setzero_si128(); + if (!skip_block) { + __m128i eob; + __m128i zbin; + __m128i round, quant, dequant, shift; + { + __m128i coeff0, coeff1; + + // Setup global values + { + __m128i pw_1; + zbin = _mm_load_si128((const __m128i *)zbin_ptr); + round = _mm_load_si128((const __m128i *)round_ptr); + quant = _mm_load_si128((const __m128i *)quant_ptr); + pw_1 = _mm_set1_epi16(1); + zbin = _mm_sub_epi16(zbin, pw_1); + dequant = _mm_load_si128((const __m128i *)dequant_ptr); + shift = _mm_load_si128((const __m128i *)quant_shift_ptr); + } + + { + __m128i coeff0_sign, coeff1_sign; + __m128i qcoeff0, qcoeff1; + __m128i qtmp0, qtmp1; + __m128i cmp_mask0, cmp_mask1; + // Do DC and first 15 AC + coeff0 = load_coefficients(coeff_ptr + n_coeffs); + coeff1 = load_coefficients(coeff_ptr + n_coeffs + 8); + + // Poor man's sign extract + coeff0_sign = _mm_srai_epi16(coeff0, 15); + coeff1_sign = _mm_srai_epi16(coeff1, 15); + qcoeff0 = _mm_xor_si128(coeff0, coeff0_sign); + qcoeff1 = _mm_xor_si128(coeff1, coeff1_sign); + qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign); + qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign); + + cmp_mask0 = _mm_cmpgt_epi16(qcoeff0, zbin); + zbin = _mm_unpackhi_epi64(zbin, zbin); // Switch DC to AC + cmp_mask1 = _mm_cmpgt_epi16(qcoeff1, zbin); + qcoeff0 = _mm_adds_epi16(qcoeff0, round); + round = _mm_unpackhi_epi64(round, round); + qcoeff1 = _mm_adds_epi16(qcoeff1, round); + qtmp0 = _mm_mulhi_epi16(qcoeff0, quant); + quant = _mm_unpackhi_epi64(quant, quant); + qtmp1 = _mm_mulhi_epi16(qcoeff1, quant); + qtmp0 = _mm_add_epi16(qtmp0, qcoeff0); + qtmp1 = _mm_add_epi16(qtmp1, qcoeff1); + qcoeff0 = _mm_mulhi_epi16(qtmp0, shift); + shift = _mm_unpackhi_epi64(shift, shift); + qcoeff1 = _mm_mulhi_epi16(qtmp1, shift); + + // Reinsert signs + qcoeff0 = _mm_xor_si128(qcoeff0, coeff0_sign); + qcoeff1 = _mm_xor_si128(qcoeff1, coeff1_sign); + qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign); + qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign); + + // Mask out zbin threshold coeffs + qcoeff0 = _mm_and_si128(qcoeff0, cmp_mask0); + qcoeff1 = _mm_and_si128(qcoeff1, cmp_mask1); + + store_coefficients(qcoeff0, qcoeff_ptr + n_coeffs); + store_coefficients(qcoeff1, qcoeff_ptr + n_coeffs + 8); + + coeff0 = _mm_mullo_epi16(qcoeff0, dequant); + dequant = _mm_unpackhi_epi64(dequant, dequant); + coeff1 = _mm_mullo_epi16(qcoeff1, dequant); + + store_coefficients(coeff0, dqcoeff_ptr + n_coeffs); + store_coefficients(coeff1, dqcoeff_ptr + n_coeffs + 8); + } + + { + // Scan for eob + __m128i zero_coeff0, zero_coeff1; + __m128i nzero_coeff0, nzero_coeff1; + __m128i iscan0, iscan1; + __m128i eob1; + zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero); + zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero); + nzero_coeff0 = _mm_cmpeq_epi16(zero_coeff0, zero); + nzero_coeff1 = _mm_cmpeq_epi16(zero_coeff1, zero); + iscan0 = _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs)); + iscan1 = _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs) + 1); + // Add one to convert from indices to counts + iscan0 = _mm_sub_epi16(iscan0, nzero_coeff0); + iscan1 = _mm_sub_epi16(iscan1, nzero_coeff1); + eob = _mm_and_si128(iscan0, nzero_coeff0); + eob1 = _mm_and_si128(iscan1, nzero_coeff1); + eob = _mm_max_epi16(eob, eob1); + } + n_coeffs += 8 * 2; + } + + // AC only loop + while (n_coeffs < 0) { + __m128i coeff0, coeff1; + { + __m128i coeff0_sign, coeff1_sign; + __m128i qcoeff0, qcoeff1; + __m128i qtmp0, qtmp1; + __m128i cmp_mask0, cmp_mask1; + + coeff0 = load_coefficients(coeff_ptr + n_coeffs); + coeff1 = load_coefficients(coeff_ptr + n_coeffs + 8); + + // Poor man's sign extract + coeff0_sign = _mm_srai_epi16(coeff0, 15); + coeff1_sign = _mm_srai_epi16(coeff1, 15); + qcoeff0 = _mm_xor_si128(coeff0, coeff0_sign); + qcoeff1 = _mm_xor_si128(coeff1, coeff1_sign); + qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign); + qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign); + + cmp_mask0 = _mm_cmpgt_epi16(qcoeff0, zbin); + cmp_mask1 = _mm_cmpgt_epi16(qcoeff1, zbin); + qcoeff0 = _mm_adds_epi16(qcoeff0, round); + qcoeff1 = _mm_adds_epi16(qcoeff1, round); + qtmp0 = _mm_mulhi_epi16(qcoeff0, quant); + qtmp1 = _mm_mulhi_epi16(qcoeff1, quant); + qtmp0 = _mm_add_epi16(qtmp0, qcoeff0); + qtmp1 = _mm_add_epi16(qtmp1, qcoeff1); + qcoeff0 = _mm_mulhi_epi16(qtmp0, shift); + qcoeff1 = _mm_mulhi_epi16(qtmp1, shift); + + // Reinsert signs + qcoeff0 = _mm_xor_si128(qcoeff0, coeff0_sign); + qcoeff1 = _mm_xor_si128(qcoeff1, coeff1_sign); + qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign); + qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign); + + // Mask out zbin threshold coeffs + qcoeff0 = _mm_and_si128(qcoeff0, cmp_mask0); + qcoeff1 = _mm_and_si128(qcoeff1, cmp_mask1); + + store_coefficients(qcoeff0, qcoeff_ptr + n_coeffs); + store_coefficients(qcoeff1, qcoeff_ptr + n_coeffs + 8); + + coeff0 = _mm_mullo_epi16(qcoeff0, dequant); + coeff1 = _mm_mullo_epi16(qcoeff1, dequant); + + store_coefficients(coeff0, dqcoeff_ptr + n_coeffs); + store_coefficients(coeff1, dqcoeff_ptr + n_coeffs + 8); + } + + { + // Scan for eob + __m128i zero_coeff0, zero_coeff1; + __m128i nzero_coeff0, nzero_coeff1; + __m128i iscan0, iscan1; + __m128i eob0, eob1; + zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero); + zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero); + nzero_coeff0 = _mm_cmpeq_epi16(zero_coeff0, zero); + nzero_coeff1 = _mm_cmpeq_epi16(zero_coeff1, zero); + iscan0 = _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs)); + iscan1 = _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs) + 1); + // Add one to convert from indices to counts + iscan0 = _mm_sub_epi16(iscan0, nzero_coeff0); + iscan1 = _mm_sub_epi16(iscan1, nzero_coeff1); + eob0 = _mm_and_si128(iscan0, nzero_coeff0); + eob1 = _mm_and_si128(iscan1, nzero_coeff1); + eob0 = _mm_max_epi16(eob0, eob1); + eob = _mm_max_epi16(eob, eob0); + } + n_coeffs += 8 * 2; + } + + // Accumulate EOB + { + __m128i eob_shuffled; + eob_shuffled = _mm_shuffle_epi32(eob, 0xe); + eob = _mm_max_epi16(eob, eob_shuffled); + eob_shuffled = _mm_shufflelo_epi16(eob, 0xe); + eob = _mm_max_epi16(eob, eob_shuffled); + eob_shuffled = _mm_shufflelo_epi16(eob, 0x1); + eob = _mm_max_epi16(eob, eob_shuffled); + *eob_ptr = _mm_extract_epi16(eob, 1); + } + } else { + do { + store_coefficients(zero, dqcoeff_ptr + n_coeffs); + store_coefficients(zero, dqcoeff_ptr + n_coeffs + 8); + store_coefficients(zero, qcoeff_ptr + n_coeffs); + store_coefficients(zero, qcoeff_ptr + n_coeffs + 8); + n_coeffs += 8 * 2; + } while (n_coeffs < 0); + *eob_ptr = 0; + } +} diff --git a/third_party/aom/aom_dsp/x86/quantize_ssse3_x86_64.asm b/third_party/aom/aom_dsp/x86/quantize_ssse3_x86_64.asm new file mode 100644 index 000000000..36b4dddbd --- /dev/null +++ b/third_party/aom/aom_dsp/x86/quantize_ssse3_x86_64.asm @@ -0,0 +1,349 @@ +; +; Copyright (c) 2016, Alliance for Open Media. All rights reserved +; +; This source code is subject to the terms of the BSD 2 Clause License and +; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License +; was not distributed with this source code in the LICENSE file, you can +; obtain it at www.aomedia.org/license/software. If the Alliance for Open +; Media Patent License 1.0 was not distributed with this source code in the +; PATENTS file, you can obtain it at www.aomedia.org/license/patent. +; + +; + +%include "third_party/x86inc/x86inc.asm" + +SECTION_RODATA +pw_1: times 8 dw 1 + +SECTION .text + +; TODO(yunqingwang)fix quantize_b code for skip=1 case. +%macro QUANTIZE_FN 2 +cglobal quantize_%1, 0, %2, 15, coeff, ncoeff, skip, zbin, round, quant, \ + shift, qcoeff, dqcoeff, dequant, \ + eob, scan, iscan + cmp dword skipm, 0 + jne .blank + + ; actual quantize loop - setup pointers, rounders, etc. + movifnidn coeffq, coeffmp + movifnidn ncoeffq, ncoeffmp + mov r2, dequantmp + movifnidn zbinq, zbinmp + movifnidn roundq, roundmp + movifnidn quantq, quantmp + mova m0, [zbinq] ; m0 = zbin + mova m1, [roundq] ; m1 = round + mova m2, [quantq] ; m2 = quant +%ifidn %1, b_32x32 + pcmpeqw m5, m5 + psrlw m5, 15 + paddw m0, m5 + paddw m1, m5 + psrlw m0, 1 ; m0 = (m0 + 1) / 2 + psrlw m1, 1 ; m1 = (m1 + 1) / 2 +%endif + mova m3, [r2q] ; m3 = dequant + psubw m0, [pw_1] + mov r2, shiftmp + mov r3, qcoeffmp + mova m4, [r2] ; m4 = shift + mov r4, dqcoeffmp + mov r5, iscanmp +%ifidn %1, b_32x32 + psllw m4, 1 +%endif + pxor m5, m5 ; m5 = dedicated zero + DEFINE_ARGS coeff, ncoeff, d1, qcoeff, dqcoeff, iscan, d2, d3, d4, d5, eob +%if CONFIG_HIGHBITDEPTH + lea coeffq, [ coeffq+ncoeffq*4] + lea qcoeffq, [ qcoeffq+ncoeffq*4] + lea dqcoeffq, [dqcoeffq+ncoeffq*4] +%else + lea coeffq, [ coeffq+ncoeffq*2] + lea qcoeffq, [ qcoeffq+ncoeffq*2] + lea dqcoeffq, [dqcoeffq+ncoeffq*2] +%endif + lea iscanq, [ iscanq+ncoeffq*2] + neg ncoeffq + + ; get DC and first 15 AC coeffs +%if CONFIG_HIGHBITDEPTH + ; coeff stored as 32bit numbers & require 16bit numbers + mova m9, [ coeffq+ncoeffq*4+ 0] + packssdw m9, [ coeffq+ncoeffq*4+16] + mova m10, [ coeffq+ncoeffq*4+32] + packssdw m10, [ coeffq+ncoeffq*4+48] +%else + mova m9, [ coeffq+ncoeffq*2+ 0] ; m9 = c[i] + mova m10, [ coeffq+ncoeffq*2+16] ; m10 = c[i] +%endif + pabsw m6, m9 ; m6 = abs(m9) + pabsw m11, m10 ; m11 = abs(m10) + pcmpgtw m7, m6, m0 ; m7 = c[i] >= zbin + punpckhqdq m0, m0 + pcmpgtw m12, m11, m0 ; m12 = c[i] >= zbin + paddsw m6, m1 ; m6 += round + punpckhqdq m1, m1 + paddsw m11, m1 ; m11 += round + pmulhw m8, m6, m2 ; m8 = m6*q>>16 + punpckhqdq m2, m2 + pmulhw m13, m11, m2 ; m13 = m11*q>>16 + paddw m8, m6 ; m8 += m6 + paddw m13, m11 ; m13 += m11 + pmulhw m8, m4 ; m8 = m8*qsh>>16 + punpckhqdq m4, m4 + pmulhw m13, m4 ; m13 = m13*qsh>>16 + psignw m8, m9 ; m8 = reinsert sign + psignw m13, m10 ; m13 = reinsert sign + pand m8, m7 + pand m13, m12 +%if CONFIG_HIGHBITDEPTH + ; store 16bit numbers as 32bit numbers in array pointed to by qcoeff + mova m11, m8 + mova m6, m8 + pcmpgtw m5, m8 + punpcklwd m11, m5 + punpckhwd m6, m5 + mova [qcoeffq+ncoeffq*4+ 0], m11 + mova [qcoeffq+ncoeffq*4+16], m6 + pxor m5, m5 + mova m11, m13 + mova m6, m13 + pcmpgtw m5, m13 + punpcklwd m11, m5 + punpckhwd m6, m5 + mova [qcoeffq+ncoeffq*4+32], m11 + mova [qcoeffq+ncoeffq*4+48], m6 + pxor m5, m5 ; reset m5 to zero register +%else + mova [qcoeffq+ncoeffq*2+ 0], m8 + mova [qcoeffq+ncoeffq*2+16], m13 +%endif +%ifidn %1, b_32x32 + pabsw m8, m8 + pabsw m13, m13 +%endif + pmullw m8, m3 ; dqc[i] = qc[i] * q + punpckhqdq m3, m3 + pmullw m13, m3 ; dqc[i] = qc[i] * q +%ifidn %1, b_32x32 + psrlw m8, 1 + psrlw m13, 1 + psignw m8, m9 + psignw m13, m10 +%endif +%if CONFIG_HIGHBITDEPTH + ; store 16bit numbers as 32bit numbers in array pointed to by qcoeff + mova m11, m8 + mova m6, m8 + pcmpgtw m5, m8 + punpcklwd m11, m5 + punpckhwd m6, m5 + mova [dqcoeffq+ncoeffq*4+ 0], m11 + mova [dqcoeffq+ncoeffq*4+16], m6 + pxor m5, m5 + mova m11, m13 + mova m6, m13 + pcmpgtw m5, m13 + punpcklwd m11, m5 + punpckhwd m6, m5 + mova [dqcoeffq+ncoeffq*4+32], m11 + mova [dqcoeffq+ncoeffq*4+48], m6 + pxor m5, m5 ; reset m5 to zero register +%else + mova [dqcoeffq+ncoeffq*2+ 0], m8 + mova [dqcoeffq+ncoeffq*2+16], m13 +%endif + pcmpeqw m8, m5 ; m8 = c[i] == 0 + pcmpeqw m13, m5 ; m13 = c[i] == 0 + mova m6, [ iscanq+ncoeffq*2+ 0] ; m6 = scan[i] + mova m11, [ iscanq+ncoeffq*2+16] ; m11 = scan[i] + psubw m6, m7 ; m6 = scan[i] + 1 + psubw m11, m12 ; m11 = scan[i] + 1 + pandn m8, m6 ; m8 = max(eob) + pandn m13, m11 ; m13 = max(eob) + pmaxsw m8, m13 + add ncoeffq, mmsize + jz .accumulate_eob + +.ac_only_loop: +%if CONFIG_HIGHBITDEPTH + ; pack coeff from 32bit to 16bit array + mova m9, [ coeffq+ncoeffq*4+ 0] + packssdw m9, [ coeffq+ncoeffq*4+16] + mova m10, [ coeffq+ncoeffq*4+32] + packssdw m10, [ coeffq+ncoeffq*4+48] +%else + mova m9, [ coeffq+ncoeffq*2+ 0] ; m9 = c[i] + mova m10, [ coeffq+ncoeffq*2+16] ; m10 = c[i] +%endif + pabsw m6, m9 ; m6 = abs(m9) + pabsw m11, m10 ; m11 = abs(m10) + pcmpgtw m7, m6, m0 ; m7 = c[i] >= zbin + pcmpgtw m12, m11, m0 ; m12 = c[i] >= zbin +%ifidn %1, b_32x32 + pmovmskb r6d, m7 + pmovmskb r2d, m12 + or r6, r2 + jz .skip_iter +%endif + paddsw m6, m1 ; m6 += round + paddsw m11, m1 ; m11 += round + pmulhw m14, m6, m2 ; m14 = m6*q>>16 + pmulhw m13, m11, m2 ; m13 = m11*q>>16 + paddw m14, m6 ; m14 += m6 + paddw m13, m11 ; m13 += m11 + pmulhw m14, m4 ; m14 = m14*qsh>>16 + pmulhw m13, m4 ; m13 = m13*qsh>>16 + psignw m14, m9 ; m14 = reinsert sign + psignw m13, m10 ; m13 = reinsert sign + pand m14, m7 + pand m13, m12 +%if CONFIG_HIGHBITDEPTH + ; store 16bit numbers as 32bit numbers in array pointed to by qcoeff + pxor m11, m11 + mova m11, m14 + mova m6, m14 + pcmpgtw m5, m14 + punpcklwd m11, m5 + punpckhwd m6, m5 + mova [qcoeffq+ncoeffq*4+ 0], m11 + mova [qcoeffq+ncoeffq*4+16], m6 + pxor m5, m5 + mova m11, m13 + mova m6, m13 + pcmpgtw m5, m13 + punpcklwd m11, m5 + punpckhwd m6, m5 + mova [qcoeffq+ncoeffq*4+32], m11 + mova [qcoeffq+ncoeffq*4+48], m6 + pxor m5, m5 ; reset m5 to zero register +%else + mova [qcoeffq+ncoeffq*2+ 0], m14 + mova [qcoeffq+ncoeffq*2+16], m13 +%endif +%ifidn %1, b_32x32 + pabsw m14, m14 + pabsw m13, m13 +%endif + pmullw m14, m3 ; dqc[i] = qc[i] * q + pmullw m13, m3 ; dqc[i] = qc[i] * q +%ifidn %1, b_32x32 + psrlw m14, 1 + psrlw m13, 1 + psignw m14, m9 + psignw m13, m10 +%endif +%if CONFIG_HIGHBITDEPTH + ; store 16bit numbers as 32bit numbers in array pointed to by qcoeff + mova m11, m14 + mova m6, m14 + pcmpgtw m5, m14 + punpcklwd m11, m5 + punpckhwd m6, m5 + mova [dqcoeffq+ncoeffq*4+ 0], m11 + mova [dqcoeffq+ncoeffq*4+16], m6 + pxor m5, m5 + mova m11, m13 + mova m6, m13 + pcmpgtw m5, m13 + punpcklwd m11, m5 + punpckhwd m6, m5 + mova [dqcoeffq+ncoeffq*4+32], m11 + mova [dqcoeffq+ncoeffq*4+48], m6 + pxor m5, m5 +%else + mova [dqcoeffq+ncoeffq*2+ 0], m14 + mova [dqcoeffq+ncoeffq*2+16], m13 +%endif + pcmpeqw m14, m5 ; m14 = c[i] == 0 + pcmpeqw m13, m5 ; m13 = c[i] == 0 + mova m6, [ iscanq+ncoeffq*2+ 0] ; m6 = scan[i] + mova m11, [ iscanq+ncoeffq*2+16] ; m11 = scan[i] + psubw m6, m7 ; m6 = scan[i] + 1 + psubw m11, m12 ; m11 = scan[i] + 1 + pandn m14, m6 ; m14 = max(eob) + pandn m13, m11 ; m13 = max(eob) + pmaxsw m8, m14 + pmaxsw m8, m13 + add ncoeffq, mmsize + jl .ac_only_loop + +%ifidn %1, b_32x32 + jmp .accumulate_eob +.skip_iter: +%if CONFIG_HIGHBITDEPTH + mova [qcoeffq+ncoeffq*4+ 0], m5 + mova [qcoeffq+ncoeffq*4+16], m5 + mova [qcoeffq+ncoeffq*4+32], m5 + mova [qcoeffq+ncoeffq*4+48], m5 + mova [dqcoeffq+ncoeffq*4+ 0], m5 + mova [dqcoeffq+ncoeffq*4+16], m5 + mova [dqcoeffq+ncoeffq*4+32], m5 + mova [dqcoeffq+ncoeffq*4+48], m5 +%else + mova [qcoeffq+ncoeffq*2+ 0], m5 + mova [qcoeffq+ncoeffq*2+16], m5 + mova [dqcoeffq+ncoeffq*2+ 0], m5 + mova [dqcoeffq+ncoeffq*2+16], m5 +%endif + add ncoeffq, mmsize + jl .ac_only_loop +%endif + +.accumulate_eob: + ; horizontally accumulate/max eobs and write into [eob] memory pointer + mov r2, eobmp + pshufd m7, m8, 0xe + pmaxsw m8, m7 + pshuflw m7, m8, 0xe + pmaxsw m8, m7 + pshuflw m7, m8, 0x1 + pmaxsw m8, m7 + pextrw r6, m8, 0 + mov [r2], r6 + RET + + ; skip-block, i.e. just write all zeroes +.blank: + mov r0, dqcoeffmp + movifnidn ncoeffq, ncoeffmp + mov r2, qcoeffmp + mov r3, eobmp + DEFINE_ARGS dqcoeff, ncoeff, qcoeff, eob +%if CONFIG_HIGHBITDEPTH + lea dqcoeffq, [dqcoeffq+ncoeffq*4] + lea qcoeffq, [ qcoeffq+ncoeffq*4] +%else + lea dqcoeffq, [dqcoeffq+ncoeffq*2] + lea qcoeffq, [ qcoeffq+ncoeffq*2] +%endif + neg ncoeffq + pxor m7, m7 +.blank_loop: +%if CONFIG_HIGHBITDEPTH + mova [dqcoeffq+ncoeffq*4+ 0], m7 + mova [dqcoeffq+ncoeffq*4+16], m7 + mova [dqcoeffq+ncoeffq*4+32], m7 + mova [dqcoeffq+ncoeffq*4+48], m7 + mova [qcoeffq+ncoeffq*4+ 0], m7 + mova [qcoeffq+ncoeffq*4+16], m7 + mova [qcoeffq+ncoeffq*4+32], m7 + mova [qcoeffq+ncoeffq*4+48], m7 +%else + mova [dqcoeffq+ncoeffq*2+ 0], m7 + mova [dqcoeffq+ncoeffq*2+16], m7 + mova [qcoeffq+ncoeffq*2+ 0], m7 + mova [qcoeffq+ncoeffq*2+16], m7 +%endif + add ncoeffq, mmsize + jl .blank_loop + mov word [eobq], 0 + RET +%endmacro + +INIT_XMM ssse3 +QUANTIZE_FN b, 7 +QUANTIZE_FN b_32x32, 7 diff --git a/third_party/aom/aom_dsp/x86/sad4d_avx2.c b/third_party/aom/aom_dsp/x86/sad4d_avx2.c new file mode 100644 index 000000000..e60f518b4 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/sad4d_avx2.c @@ -0,0 +1,216 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ +#include <immintrin.h> // AVX2 +#include "./aom_dsp_rtcd.h" +#include "aom/aom_integer.h" + +void aom_sad32x32x4d_avx2(const uint8_t *src, int src_stride, + const uint8_t *const ref[4], int ref_stride, + uint32_t res[4]) { + __m256i src_reg, ref0_reg, ref1_reg, ref2_reg, ref3_reg; + __m256i sum_ref0, sum_ref1, sum_ref2, sum_ref3; + __m256i sum_mlow, sum_mhigh; + int i; + const uint8_t *ref0, *ref1, *ref2, *ref3; + + ref0 = ref[0]; + ref1 = ref[1]; + ref2 = ref[2]; + ref3 = ref[3]; + sum_ref0 = _mm256_set1_epi16(0); + sum_ref1 = _mm256_set1_epi16(0); + sum_ref2 = _mm256_set1_epi16(0); + sum_ref3 = _mm256_set1_epi16(0); + for (i = 0; i < 32; i++) { + // load src and all refs + src_reg = _mm256_loadu_si256((const __m256i *)src); + ref0_reg = _mm256_loadu_si256((const __m256i *)ref0); + ref1_reg = _mm256_loadu_si256((const __m256i *)ref1); + ref2_reg = _mm256_loadu_si256((const __m256i *)ref2); + ref3_reg = _mm256_loadu_si256((const __m256i *)ref3); + // sum of the absolute differences between every ref-i to src + ref0_reg = _mm256_sad_epu8(ref0_reg, src_reg); + ref1_reg = _mm256_sad_epu8(ref1_reg, src_reg); + ref2_reg = _mm256_sad_epu8(ref2_reg, src_reg); + ref3_reg = _mm256_sad_epu8(ref3_reg, src_reg); + // sum every ref-i + sum_ref0 = _mm256_add_epi32(sum_ref0, ref0_reg); + sum_ref1 = _mm256_add_epi32(sum_ref1, ref1_reg); + sum_ref2 = _mm256_add_epi32(sum_ref2, ref2_reg); + sum_ref3 = _mm256_add_epi32(sum_ref3, ref3_reg); + + src += src_stride; + ref0 += ref_stride; + ref1 += ref_stride; + ref2 += ref_stride; + ref3 += ref_stride; + } + { + __m128i sum; + // in sum_ref-i the result is saved in the first 4 bytes + // the other 4 bytes are zeroed. + // sum_ref1 and sum_ref3 are shifted left by 4 bytes + sum_ref1 = _mm256_slli_si256(sum_ref1, 4); + sum_ref3 = _mm256_slli_si256(sum_ref3, 4); + + // merge sum_ref0 and sum_ref1 also sum_ref2 and sum_ref3 + sum_ref0 = _mm256_or_si256(sum_ref0, sum_ref1); + sum_ref2 = _mm256_or_si256(sum_ref2, sum_ref3); + + // merge every 64 bit from each sum_ref-i + sum_mlow = _mm256_unpacklo_epi64(sum_ref0, sum_ref2); + sum_mhigh = _mm256_unpackhi_epi64(sum_ref0, sum_ref2); + + // add the low 64 bit to the high 64 bit + sum_mlow = _mm256_add_epi32(sum_mlow, sum_mhigh); + + // add the low 128 bit to the high 128 bit + sum = _mm_add_epi32(_mm256_castsi256_si128(sum_mlow), + _mm256_extractf128_si256(sum_mlow, 1)); + + _mm_storeu_si128((__m128i *)(res), sum); + } + _mm256_zeroupper(); +} + +void aom_sad64x64x4d_avx2(const uint8_t *src, int src_stride, + const uint8_t *const ref[4], int ref_stride, + uint32_t res[4]) { + __m256i src_reg, srcnext_reg, ref0_reg, ref0next_reg; + __m256i ref1_reg, ref1next_reg, ref2_reg, ref2next_reg; + __m256i ref3_reg, ref3next_reg; + __m256i sum_ref0, sum_ref1, sum_ref2, sum_ref3; + __m256i sum_mlow, sum_mhigh; + int i; + const uint8_t *ref0, *ref1, *ref2, *ref3; + + ref0 = ref[0]; + ref1 = ref[1]; + ref2 = ref[2]; + ref3 = ref[3]; + sum_ref0 = _mm256_set1_epi16(0); + sum_ref1 = _mm256_set1_epi16(0); + sum_ref2 = _mm256_set1_epi16(0); + sum_ref3 = _mm256_set1_epi16(0); + for (i = 0; i < 64; i++) { + // load 64 bytes from src and all refs + src_reg = _mm256_loadu_si256((const __m256i *)src); + srcnext_reg = _mm256_loadu_si256((const __m256i *)(src + 32)); + ref0_reg = _mm256_loadu_si256((const __m256i *)ref0); + ref0next_reg = _mm256_loadu_si256((const __m256i *)(ref0 + 32)); + ref1_reg = _mm256_loadu_si256((const __m256i *)ref1); + ref1next_reg = _mm256_loadu_si256((const __m256i *)(ref1 + 32)); + ref2_reg = _mm256_loadu_si256((const __m256i *)ref2); + ref2next_reg = _mm256_loadu_si256((const __m256i *)(ref2 + 32)); + ref3_reg = _mm256_loadu_si256((const __m256i *)ref3); + ref3next_reg = _mm256_loadu_si256((const __m256i *)(ref3 + 32)); + // sum of the absolute differences between every ref-i to src + ref0_reg = _mm256_sad_epu8(ref0_reg, src_reg); + ref1_reg = _mm256_sad_epu8(ref1_reg, src_reg); + ref2_reg = _mm256_sad_epu8(ref2_reg, src_reg); + ref3_reg = _mm256_sad_epu8(ref3_reg, src_reg); + ref0next_reg = _mm256_sad_epu8(ref0next_reg, srcnext_reg); + ref1next_reg = _mm256_sad_epu8(ref1next_reg, srcnext_reg); + ref2next_reg = _mm256_sad_epu8(ref2next_reg, srcnext_reg); + ref3next_reg = _mm256_sad_epu8(ref3next_reg, srcnext_reg); + + // sum every ref-i + sum_ref0 = _mm256_add_epi32(sum_ref0, ref0_reg); + sum_ref1 = _mm256_add_epi32(sum_ref1, ref1_reg); + sum_ref2 = _mm256_add_epi32(sum_ref2, ref2_reg); + sum_ref3 = _mm256_add_epi32(sum_ref3, ref3_reg); + sum_ref0 = _mm256_add_epi32(sum_ref0, ref0next_reg); + sum_ref1 = _mm256_add_epi32(sum_ref1, ref1next_reg); + sum_ref2 = _mm256_add_epi32(sum_ref2, ref2next_reg); + sum_ref3 = _mm256_add_epi32(sum_ref3, ref3next_reg); + src += src_stride; + ref0 += ref_stride; + ref1 += ref_stride; + ref2 += ref_stride; + ref3 += ref_stride; + } + { + __m128i sum; + + // in sum_ref-i the result is saved in the first 4 bytes + // the other 4 bytes are zeroed. + // sum_ref1 and sum_ref3 are shifted left by 4 bytes + sum_ref1 = _mm256_slli_si256(sum_ref1, 4); + sum_ref3 = _mm256_slli_si256(sum_ref3, 4); + + // merge sum_ref0 and sum_ref1 also sum_ref2 and sum_ref3 + sum_ref0 = _mm256_or_si256(sum_ref0, sum_ref1); + sum_ref2 = _mm256_or_si256(sum_ref2, sum_ref3); + + // merge every 64 bit from each sum_ref-i + sum_mlow = _mm256_unpacklo_epi64(sum_ref0, sum_ref2); + sum_mhigh = _mm256_unpackhi_epi64(sum_ref0, sum_ref2); + + // add the low 64 bit to the high 64 bit + sum_mlow = _mm256_add_epi32(sum_mlow, sum_mhigh); + + // add the low 128 bit to the high 128 bit + sum = _mm_add_epi32(_mm256_castsi256_si128(sum_mlow), + _mm256_extractf128_si256(sum_mlow, 1)); + + _mm_storeu_si128((__m128i *)(res), sum); + } + _mm256_zeroupper(); +} + +void aom_sad32x64x4d_avx2(const uint8_t *src, int src_stride, + const uint8_t *const ref[4], int ref_stride, + uint32_t res[4]) { + const uint8_t *rf[4]; + uint32_t sum0[4]; + uint32_t sum1[4]; + + rf[0] = ref[0]; + rf[1] = ref[1]; + rf[2] = ref[2]; + rf[3] = ref[3]; + aom_sad32x32x4d_avx2(src, src_stride, rf, ref_stride, sum0); + src += src_stride << 5; + rf[0] += ref_stride << 5; + rf[1] += ref_stride << 5; + rf[2] += ref_stride << 5; + rf[3] += ref_stride << 5; + aom_sad32x32x4d_avx2(src, src_stride, rf, ref_stride, sum1); + res[0] = sum0[0] + sum1[0]; + res[1] = sum0[1] + sum1[1]; + res[2] = sum0[2] + sum1[2]; + res[3] = sum0[3] + sum1[3]; +} + +void aom_sad64x32x4d_avx2(const uint8_t *src, int src_stride, + const uint8_t *const ref[4], int ref_stride, + uint32_t res[4]) { + const uint8_t *rf[4]; + uint32_t sum0[4]; + uint32_t sum1[4]; + unsigned int half_width = 32; + + rf[0] = ref[0]; + rf[1] = ref[1]; + rf[2] = ref[2]; + rf[3] = ref[3]; + aom_sad32x32x4d_avx2(src, src_stride, rf, ref_stride, sum0); + src += half_width; + rf[0] += half_width; + rf[1] += half_width; + rf[2] += half_width; + rf[3] += half_width; + aom_sad32x32x4d_avx2(src, src_stride, rf, ref_stride, sum1); + res[0] = sum0[0] + sum1[0]; + res[1] = sum0[1] + sum1[1]; + res[2] = sum0[2] + sum1[2]; + res[3] = sum0[3] + sum1[3]; +} diff --git a/third_party/aom/aom_dsp/x86/sad4d_sse2.asm b/third_party/aom/aom_dsp/x86/sad4d_sse2.asm new file mode 100644 index 000000000..8f04ef2f3 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/sad4d_sse2.asm @@ -0,0 +1,253 @@ +; +; Copyright (c) 2016, Alliance for Open Media. All rights reserved +; +; This source code is subject to the terms of the BSD 2 Clause License and +; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License +; was not distributed with this source code in the LICENSE file, you can +; obtain it at www.aomedia.org/license/software. If the Alliance for Open +; Media Patent License 1.0 was not distributed with this source code in the +; PATENTS file, you can obtain it at www.aomedia.org/license/patent. +; + +; + +%include "third_party/x86inc/x86inc.asm" + +SECTION .text + +; PROCESS_4x2x4 first, off_{first,second}_{src,ref}, advance_at_end +%macro PROCESS_4x2x4 5-6 0 + movd m0, [srcq +%2] +%if %1 == 1 + movd m6, [ref1q+%3] + movd m4, [ref2q+%3] + movd m7, [ref3q+%3] + movd m5, [ref4q+%3] + movd m1, [srcq +%4] + movd m2, [ref1q+%5] + punpckldq m0, m1 + punpckldq m6, m2 + movd m1, [ref2q+%5] + movd m2, [ref3q+%5] + movd m3, [ref4q+%5] + punpckldq m4, m1 + punpckldq m7, m2 + punpckldq m5, m3 + movlhps m0, m0 + movlhps m6, m4 + movlhps m7, m5 + psadbw m6, m0 + psadbw m7, m0 +%else + movd m1, [ref1q+%3] + movd m5, [ref1q+%5] + movd m2, [ref2q+%3] + movd m4, [ref2q+%5] + punpckldq m1, m5 + punpckldq m2, m4 + movd m3, [ref3q+%3] + movd m5, [ref3q+%5] + punpckldq m3, m5 + movd m4, [ref4q+%3] + movd m5, [ref4q+%5] + punpckldq m4, m5 + movd m5, [srcq +%4] + punpckldq m0, m5 + movlhps m0, m0 + movlhps m1, m2 + movlhps m3, m4 + psadbw m1, m0 + psadbw m3, m0 + paddd m6, m1 + paddd m7, m3 +%endif +%if %6 == 1 + lea srcq, [srcq +src_strideq*2] + lea ref1q, [ref1q+ref_strideq*2] + lea ref2q, [ref2q+ref_strideq*2] + lea ref3q, [ref3q+ref_strideq*2] + lea ref4q, [ref4q+ref_strideq*2] +%endif +%endmacro + +; PROCESS_8x2x4 first, off_{first,second}_{src,ref}, advance_at_end +%macro PROCESS_8x2x4 5-6 0 + movh m0, [srcq +%2] +%if %1 == 1 + movh m4, [ref1q+%3] + movh m5, [ref2q+%3] + movh m6, [ref3q+%3] + movh m7, [ref4q+%3] + movhps m0, [srcq +%4] + movhps m4, [ref1q+%5] + movhps m5, [ref2q+%5] + movhps m6, [ref3q+%5] + movhps m7, [ref4q+%5] + psadbw m4, m0 + psadbw m5, m0 + psadbw m6, m0 + psadbw m7, m0 +%else + movh m1, [ref1q+%3] + movh m2, [ref2q+%3] + movh m3, [ref3q+%3] + movhps m0, [srcq +%4] + movhps m1, [ref1q+%5] + movhps m2, [ref2q+%5] + movhps m3, [ref3q+%5] + psadbw m1, m0 + psadbw m2, m0 + psadbw m3, m0 + paddd m4, m1 + movh m1, [ref4q+%3] + movhps m1, [ref4q+%5] + paddd m5, m2 + paddd m6, m3 + psadbw m1, m0 + paddd m7, m1 +%endif +%if %6 == 1 + lea srcq, [srcq +src_strideq*2] + lea ref1q, [ref1q+ref_strideq*2] + lea ref2q, [ref2q+ref_strideq*2] + lea ref3q, [ref3q+ref_strideq*2] + lea ref4q, [ref4q+ref_strideq*2] +%endif +%endmacro + +; PROCESS_16x2x4 first, off_{first,second}_{src,ref}, advance_at_end +%macro PROCESS_16x2x4 5-6 0 + ; 1st 16 px + mova m0, [srcq +%2] +%if %1 == 1 + movu m4, [ref1q+%3] + movu m5, [ref2q+%3] + movu m6, [ref3q+%3] + movu m7, [ref4q+%3] + psadbw m4, m0 + psadbw m5, m0 + psadbw m6, m0 + psadbw m7, m0 +%else + movu m1, [ref1q+%3] + movu m2, [ref2q+%3] + movu m3, [ref3q+%3] + psadbw m1, m0 + psadbw m2, m0 + psadbw m3, m0 + paddd m4, m1 + movu m1, [ref4q+%3] + paddd m5, m2 + paddd m6, m3 + psadbw m1, m0 + paddd m7, m1 +%endif + + ; 2nd 16 px + mova m0, [srcq +%4] + movu m1, [ref1q+%5] + movu m2, [ref2q+%5] + movu m3, [ref3q+%5] + psadbw m1, m0 + psadbw m2, m0 + psadbw m3, m0 + paddd m4, m1 + movu m1, [ref4q+%5] + paddd m5, m2 + paddd m6, m3 +%if %6 == 1 + lea srcq, [srcq +src_strideq*2] + lea ref1q, [ref1q+ref_strideq*2] + lea ref2q, [ref2q+ref_strideq*2] + lea ref3q, [ref3q+ref_strideq*2] + lea ref4q, [ref4q+ref_strideq*2] +%endif + psadbw m1, m0 + paddd m7, m1 +%endmacro + +; PROCESS_32x2x4 first, off_{first,second}_{src,ref}, advance_at_end +%macro PROCESS_32x2x4 5-6 0 + PROCESS_16x2x4 %1, %2, %3, %2 + 16, %3 + 16 + PROCESS_16x2x4 0, %4, %5, %4 + 16, %5 + 16, %6 +%endmacro + +; PROCESS_64x2x4 first, off_{first,second}_{src,ref}, advance_at_end +%macro PROCESS_64x2x4 5-6 0 + PROCESS_32x2x4 %1, %2, %3, %2 + 32, %3 + 32 + PROCESS_32x2x4 0, %4, %5, %4 + 32, %5 + 32, %6 +%endmacro + +; PROCESS_128x2x4 first, off_{first,second}_{src,ref}, advance_at_end +%macro PROCESS_128x2x4 5-6 0 + PROCESS_64x2x4 %1, %2, %3, %2 + 64, %3 + 64 + PROCESS_64x2x4 0, %4, %5, %4 + 64, %5 + 64, %6 +%endmacro + +; void aom_sadNxNx4d_sse2(uint8_t *src, int src_stride, +; uint8_t *ref[4], int ref_stride, +; uint32_t res[4]); +; where NxN = 64x64, 32x32, 16x16, 16x8, 8x16, 8x8, 8x4, 4x8 and 4x4 +%macro SADNXN4D 2 +%if UNIX64 +cglobal sad%1x%2x4d, 5, 8, 8, src, src_stride, ref1, ref_stride, \ + res, ref2, ref3, ref4 +%else +cglobal sad%1x%2x4d, 4, 7, 8, src, src_stride, ref1, ref_stride, \ + ref2, ref3, ref4 +%endif + movsxdifnidn src_strideq, src_strided + movsxdifnidn ref_strideq, ref_strided + mov ref2q, [ref1q+gprsize*1] + mov ref3q, [ref1q+gprsize*2] + mov ref4q, [ref1q+gprsize*3] + mov ref1q, [ref1q+gprsize*0] + + PROCESS_%1x2x4 1, 0, 0, src_strideq, ref_strideq, 1 +%rep (%2-4)/2 + PROCESS_%1x2x4 0, 0, 0, src_strideq, ref_strideq, 1 +%endrep + PROCESS_%1x2x4 0, 0, 0, src_strideq, ref_strideq, 0 + +%if %1 > 4 + pslldq m5, 4 + pslldq m7, 4 + por m4, m5 + por m6, m7 + mova m5, m4 + mova m7, m6 + punpcklqdq m4, m6 + punpckhqdq m5, m7 + movifnidn r4, r4mp + paddd m4, m5 + movu [r4], m4 + RET +%else + movifnidn r4, r4mp + pshufd m6, m6, 0x08 + pshufd m7, m7, 0x08 + movq [r4+0], m6 + movq [r4+8], m7 + RET +%endif +%endmacro + +INIT_XMM sse2 +%if CONFIG_EXT_PARTITION +SADNXN4D 128, 128 +SADNXN4D 128, 64 +SADNXN4D 64, 128 +%endif +SADNXN4D 64, 64 +SADNXN4D 64, 32 +SADNXN4D 32, 64 +SADNXN4D 32, 32 +SADNXN4D 32, 16 +SADNXN4D 16, 32 +SADNXN4D 16, 16 +SADNXN4D 16, 8 +SADNXN4D 8, 16 +SADNXN4D 8, 8 +SADNXN4D 8, 4 +SADNXN4D 4, 8 +SADNXN4D 4, 4 diff --git a/third_party/aom/aom_dsp/x86/sad_avx2.c b/third_party/aom/aom_dsp/x86/sad_avx2.c new file mode 100644 index 000000000..efba61289 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/sad_avx2.c @@ -0,0 +1,187 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ +#include <immintrin.h> +#include "./aom_dsp_rtcd.h" +#include "aom_ports/mem.h" + +#define FSAD64_H(h) \ + unsigned int aom_sad64x##h##_avx2(const uint8_t *src_ptr, int src_stride, \ + const uint8_t *ref_ptr, int ref_stride) { \ + int i, res; \ + __m256i sad1_reg, sad2_reg, ref1_reg, ref2_reg; \ + __m256i sum_sad = _mm256_setzero_si256(); \ + __m256i sum_sad_h; \ + __m128i sum_sad128; \ + for (i = 0; i < h; i++) { \ + ref1_reg = _mm256_loadu_si256((__m256i const *)ref_ptr); \ + ref2_reg = _mm256_loadu_si256((__m256i const *)(ref_ptr + 32)); \ + sad1_reg = _mm256_sad_epu8( \ + ref1_reg, _mm256_loadu_si256((__m256i const *)src_ptr)); \ + sad2_reg = _mm256_sad_epu8( \ + ref2_reg, _mm256_loadu_si256((__m256i const *)(src_ptr + 32))); \ + sum_sad = \ + _mm256_add_epi32(sum_sad, _mm256_add_epi32(sad1_reg, sad2_reg)); \ + ref_ptr += ref_stride; \ + src_ptr += src_stride; \ + } \ + sum_sad_h = _mm256_srli_si256(sum_sad, 8); \ + sum_sad = _mm256_add_epi32(sum_sad, sum_sad_h); \ + sum_sad128 = _mm256_extracti128_si256(sum_sad, 1); \ + sum_sad128 = _mm_add_epi32(_mm256_castsi256_si128(sum_sad), sum_sad128); \ + res = _mm_cvtsi128_si32(sum_sad128); \ + _mm256_zeroupper(); \ + return res; \ + } + +#define FSAD32_H(h) \ + unsigned int aom_sad32x##h##_avx2(const uint8_t *src_ptr, int src_stride, \ + const uint8_t *ref_ptr, int ref_stride) { \ + int i, res; \ + __m256i sad1_reg, sad2_reg, ref1_reg, ref2_reg; \ + __m256i sum_sad = _mm256_setzero_si256(); \ + __m256i sum_sad_h; \ + __m128i sum_sad128; \ + int ref2_stride = ref_stride << 1; \ + int src2_stride = src_stride << 1; \ + int max = h >> 1; \ + for (i = 0; i < max; i++) { \ + ref1_reg = _mm256_loadu_si256((__m256i const *)ref_ptr); \ + ref2_reg = _mm256_loadu_si256((__m256i const *)(ref_ptr + ref_stride)); \ + sad1_reg = _mm256_sad_epu8( \ + ref1_reg, _mm256_loadu_si256((__m256i const *)src_ptr)); \ + sad2_reg = _mm256_sad_epu8( \ + ref2_reg, \ + _mm256_loadu_si256((__m256i const *)(src_ptr + src_stride))); \ + sum_sad = \ + _mm256_add_epi32(sum_sad, _mm256_add_epi32(sad1_reg, sad2_reg)); \ + ref_ptr += ref2_stride; \ + src_ptr += src2_stride; \ + } \ + sum_sad_h = _mm256_srli_si256(sum_sad, 8); \ + sum_sad = _mm256_add_epi32(sum_sad, sum_sad_h); \ + sum_sad128 = _mm256_extracti128_si256(sum_sad, 1); \ + sum_sad128 = _mm_add_epi32(_mm256_castsi256_si128(sum_sad), sum_sad128); \ + res = _mm_cvtsi128_si32(sum_sad128); \ + _mm256_zeroupper(); \ + return res; \ + } + +#define FSAD64 \ + FSAD64_H(64); \ + FSAD64_H(32); + +#define FSAD32 \ + FSAD32_H(64); \ + FSAD32_H(32); \ + FSAD32_H(16); + +/* clang-format off */ +FSAD64 +FSAD32 +/* clang-format on */ + +#undef FSAD64 +#undef FSAD32 +#undef FSAD64_H +#undef FSAD32_H + +#define FSADAVG64_H(h) \ + unsigned int aom_sad64x##h##_avg_avx2( \ + const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, \ + int ref_stride, const uint8_t *second_pred) { \ + int i, res; \ + __m256i sad1_reg, sad2_reg, ref1_reg, ref2_reg; \ + __m256i sum_sad = _mm256_setzero_si256(); \ + __m256i sum_sad_h; \ + __m128i sum_sad128; \ + for (i = 0; i < h; i++) { \ + ref1_reg = _mm256_loadu_si256((__m256i const *)ref_ptr); \ + ref2_reg = _mm256_loadu_si256((__m256i const *)(ref_ptr + 32)); \ + ref1_reg = _mm256_avg_epu8( \ + ref1_reg, _mm256_loadu_si256((__m256i const *)second_pred)); \ + ref2_reg = _mm256_avg_epu8( \ + ref2_reg, _mm256_loadu_si256((__m256i const *)(second_pred + 32))); \ + sad1_reg = _mm256_sad_epu8( \ + ref1_reg, _mm256_loadu_si256((__m256i const *)src_ptr)); \ + sad2_reg = _mm256_sad_epu8( \ + ref2_reg, _mm256_loadu_si256((__m256i const *)(src_ptr + 32))); \ + sum_sad = \ + _mm256_add_epi32(sum_sad, _mm256_add_epi32(sad1_reg, sad2_reg)); \ + ref_ptr += ref_stride; \ + src_ptr += src_stride; \ + second_pred += 64; \ + } \ + sum_sad_h = _mm256_srli_si256(sum_sad, 8); \ + sum_sad = _mm256_add_epi32(sum_sad, sum_sad_h); \ + sum_sad128 = _mm256_extracti128_si256(sum_sad, 1); \ + sum_sad128 = _mm_add_epi32(_mm256_castsi256_si128(sum_sad), sum_sad128); \ + res = _mm_cvtsi128_si32(sum_sad128); \ + _mm256_zeroupper(); \ + return res; \ + } + +#define FSADAVG32_H(h) \ + unsigned int aom_sad32x##h##_avg_avx2( \ + const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, \ + int ref_stride, const uint8_t *second_pred) { \ + int i, res; \ + __m256i sad1_reg, sad2_reg, ref1_reg, ref2_reg; \ + __m256i sum_sad = _mm256_setzero_si256(); \ + __m256i sum_sad_h; \ + __m128i sum_sad128; \ + int ref2_stride = ref_stride << 1; \ + int src2_stride = src_stride << 1; \ + int max = h >> 1; \ + for (i = 0; i < max; i++) { \ + ref1_reg = _mm256_loadu_si256((__m256i const *)ref_ptr); \ + ref2_reg = _mm256_loadu_si256((__m256i const *)(ref_ptr + ref_stride)); \ + ref1_reg = _mm256_avg_epu8( \ + ref1_reg, _mm256_loadu_si256((__m256i const *)second_pred)); \ + ref2_reg = _mm256_avg_epu8( \ + ref2_reg, _mm256_loadu_si256((__m256i const *)(second_pred + 32))); \ + sad1_reg = _mm256_sad_epu8( \ + ref1_reg, _mm256_loadu_si256((__m256i const *)src_ptr)); \ + sad2_reg = _mm256_sad_epu8( \ + ref2_reg, \ + _mm256_loadu_si256((__m256i const *)(src_ptr + src_stride))); \ + sum_sad = \ + _mm256_add_epi32(sum_sad, _mm256_add_epi32(sad1_reg, sad2_reg)); \ + ref_ptr += ref2_stride; \ + src_ptr += src2_stride; \ + second_pred += 64; \ + } \ + sum_sad_h = _mm256_srli_si256(sum_sad, 8); \ + sum_sad = _mm256_add_epi32(sum_sad, sum_sad_h); \ + sum_sad128 = _mm256_extracti128_si256(sum_sad, 1); \ + sum_sad128 = _mm_add_epi32(_mm256_castsi256_si128(sum_sad), sum_sad128); \ + res = _mm_cvtsi128_si32(sum_sad128); \ + _mm256_zeroupper(); \ + return res; \ + } + +#define FSADAVG64 \ + FSADAVG64_H(64); \ + FSADAVG64_H(32); + +#define FSADAVG32 \ + FSADAVG32_H(64); \ + FSADAVG32_H(32); \ + FSADAVG32_H(16); + +/* clang-format off */ +FSADAVG64 +FSADAVG32 +/* clang-format on */ + +#undef FSADAVG64 +#undef FSADAVG32 +#undef FSADAVG64_H +#undef FSADAVG32_H diff --git a/third_party/aom/aom_dsp/x86/sad_highbd_avx2.c b/third_party/aom/aom_dsp/x86/sad_highbd_avx2.c new file mode 100644 index 000000000..196394379 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/sad_highbd_avx2.c @@ -0,0 +1,1043 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include <immintrin.h> + +#include "./aom_config.h" +#include "./aom_dsp_rtcd.h" + +#include "aom/aom_integer.h" +#include "aom_ports/mem.h" + +// SAD +static INLINE unsigned int get_sad_from_mm256_epi32(const __m256i *v) { + // input 8 32-bit summation + __m128i lo128, hi128; + __m256i u = _mm256_srli_si256(*v, 8); + u = _mm256_add_epi32(u, *v); + + // 4 32-bit summation + hi128 = _mm256_extracti128_si256(u, 1); + lo128 = _mm256_castsi256_si128(u); + lo128 = _mm_add_epi32(hi128, lo128); + + // 2 32-bit summation + hi128 = _mm_srli_si128(lo128, 4); + lo128 = _mm_add_epi32(lo128, hi128); + + return (unsigned int)_mm_cvtsi128_si32(lo128); +} + +unsigned int aom_highbd_sad16x8_avx2(const uint8_t *src, int src_stride, + const uint8_t *ref, int ref_stride) { + const uint16_t *src_ptr = CONVERT_TO_SHORTPTR(src); + const uint16_t *ref_ptr = CONVERT_TO_SHORTPTR(ref); + + // first 4 rows + __m256i s0 = _mm256_loadu_si256((const __m256i *)src_ptr); + __m256i s1 = _mm256_loadu_si256((const __m256i *)(src_ptr + src_stride)); + __m256i s2 = _mm256_loadu_si256((const __m256i *)(src_ptr + 2 * src_stride)); + __m256i s3 = _mm256_loadu_si256((const __m256i *)(src_ptr + 3 * src_stride)); + + __m256i r0 = _mm256_loadu_si256((const __m256i *)ref_ptr); + __m256i r1 = _mm256_loadu_si256((const __m256i *)(ref_ptr + ref_stride)); + __m256i r2 = _mm256_loadu_si256((const __m256i *)(ref_ptr + 2 * ref_stride)); + __m256i r3 = _mm256_loadu_si256((const __m256i *)(ref_ptr + 3 * ref_stride)); + + __m256i u0 = _mm256_sub_epi16(s0, r0); + __m256i u1 = _mm256_sub_epi16(s1, r1); + __m256i u2 = _mm256_sub_epi16(s2, r2); + __m256i u3 = _mm256_sub_epi16(s3, r3); + __m256i zero = _mm256_setzero_si256(); + __m256i sum0, sum1; + + u0 = _mm256_abs_epi16(u0); + u1 = _mm256_abs_epi16(u1); + u2 = _mm256_abs_epi16(u2); + u3 = _mm256_abs_epi16(u3); + + sum0 = _mm256_add_epi16(u0, u1); + sum0 = _mm256_add_epi16(sum0, u2); + sum0 = _mm256_add_epi16(sum0, u3); + + // second 4 rows + src_ptr += src_stride << 2; + ref_ptr += ref_stride << 2; + s0 = _mm256_loadu_si256((const __m256i *)src_ptr); + s1 = _mm256_loadu_si256((const __m256i *)(src_ptr + src_stride)); + s2 = _mm256_loadu_si256((const __m256i *)(src_ptr + 2 * src_stride)); + s3 = _mm256_loadu_si256((const __m256i *)(src_ptr + 3 * src_stride)); + + r0 = _mm256_loadu_si256((const __m256i *)ref_ptr); + r1 = _mm256_loadu_si256((const __m256i *)(ref_ptr + ref_stride)); + r2 = _mm256_loadu_si256((const __m256i *)(ref_ptr + 2 * ref_stride)); + r3 = _mm256_loadu_si256((const __m256i *)(ref_ptr + 3 * ref_stride)); + + u0 = _mm256_sub_epi16(s0, r0); + u1 = _mm256_sub_epi16(s1, r1); + u2 = _mm256_sub_epi16(s2, r2); + u3 = _mm256_sub_epi16(s3, r3); + + u0 = _mm256_abs_epi16(u0); + u1 = _mm256_abs_epi16(u1); + u2 = _mm256_abs_epi16(u2); + u3 = _mm256_abs_epi16(u3); + + sum1 = _mm256_add_epi16(u0, u1); + sum1 = _mm256_add_epi16(sum1, u2); + sum1 = _mm256_add_epi16(sum1, u3); + + // find out the SAD + s0 = _mm256_unpacklo_epi16(sum0, zero); + s1 = _mm256_unpackhi_epi16(sum0, zero); + r0 = _mm256_unpacklo_epi16(sum1, zero); + r1 = _mm256_unpackhi_epi16(sum1, zero); + s0 = _mm256_add_epi32(s0, s1); + r0 = _mm256_add_epi32(r0, r1); + sum0 = _mm256_add_epi32(s0, r0); + // 8 32-bit summation + + return (unsigned int)get_sad_from_mm256_epi32(&sum0); +} + +unsigned int aom_highbd_sad16x16_avx2(const uint8_t *src, int src_stride, + const uint8_t *ref, int ref_stride) { + const uint16_t *src_ptr = CONVERT_TO_SHORTPTR(src); + const uint16_t *ref_ptr = CONVERT_TO_SHORTPTR(ref); + __m256i s0, s1, s2, s3, r0, r1, r2, r3, u0, u1, u2, u3; + __m256i sum0; + __m256i sum = _mm256_setzero_si256(); + const __m256i zero = _mm256_setzero_si256(); + int row = 0; + + // Loop for every 4 rows + while (row < 16) { + s0 = _mm256_loadu_si256((const __m256i *)src_ptr); + s1 = _mm256_loadu_si256((const __m256i *)(src_ptr + src_stride)); + s2 = _mm256_loadu_si256((const __m256i *)(src_ptr + 2 * src_stride)); + s3 = _mm256_loadu_si256((const __m256i *)(src_ptr + 3 * src_stride)); + + r0 = _mm256_loadu_si256((const __m256i *)ref_ptr); + r1 = _mm256_loadu_si256((const __m256i *)(ref_ptr + ref_stride)); + r2 = _mm256_loadu_si256((const __m256i *)(ref_ptr + 2 * ref_stride)); + r3 = _mm256_loadu_si256((const __m256i *)(ref_ptr + 3 * ref_stride)); + + u0 = _mm256_sub_epi16(s0, r0); + u1 = _mm256_sub_epi16(s1, r1); + u2 = _mm256_sub_epi16(s2, r2); + u3 = _mm256_sub_epi16(s3, r3); + + u0 = _mm256_abs_epi16(u0); + u1 = _mm256_abs_epi16(u1); + u2 = _mm256_abs_epi16(u2); + u3 = _mm256_abs_epi16(u3); + + sum0 = _mm256_add_epi16(u0, u1); + sum0 = _mm256_add_epi16(sum0, u2); + sum0 = _mm256_add_epi16(sum0, u3); + + s0 = _mm256_unpacklo_epi16(sum0, zero); + s1 = _mm256_unpackhi_epi16(sum0, zero); + sum = _mm256_add_epi32(sum, s0); + sum = _mm256_add_epi32(sum, s1); + // 8 32-bit summation + + row += 4; + src_ptr += src_stride << 2; + ref_ptr += ref_stride << 2; + } + return get_sad_from_mm256_epi32(&sum); +} + +static void sad32x4(const uint16_t *src_ptr, int src_stride, + const uint16_t *ref_ptr, int ref_stride, + const uint16_t *sec_ptr, __m256i *sad_acc) { + __m256i s0, s1, s2, s3, r0, r1, r2, r3; + const __m256i zero = _mm256_setzero_si256(); + int row_sections = 0; + + while (row_sections < 2) { + s0 = _mm256_loadu_si256((const __m256i *)src_ptr); + s1 = _mm256_loadu_si256((const __m256i *)(src_ptr + 16)); + s2 = _mm256_loadu_si256((const __m256i *)(src_ptr + src_stride)); + s3 = _mm256_loadu_si256((const __m256i *)(src_ptr + src_stride + 16)); + + r0 = _mm256_loadu_si256((const __m256i *)ref_ptr); + r1 = _mm256_loadu_si256((const __m256i *)(ref_ptr + 16)); + r2 = _mm256_loadu_si256((const __m256i *)(ref_ptr + ref_stride)); + r3 = _mm256_loadu_si256((const __m256i *)(ref_ptr + ref_stride + 16)); + + if (sec_ptr) { + r0 = _mm256_avg_epu16(r0, _mm256_loadu_si256((const __m256i *)sec_ptr)); + r1 = _mm256_avg_epu16( + r1, _mm256_loadu_si256((const __m256i *)(sec_ptr + 16))); + r2 = _mm256_avg_epu16( + r2, _mm256_loadu_si256((const __m256i *)(sec_ptr + 32))); + r3 = _mm256_avg_epu16( + r3, _mm256_loadu_si256((const __m256i *)(sec_ptr + 48))); + } + s0 = _mm256_sub_epi16(s0, r0); + s1 = _mm256_sub_epi16(s1, r1); + s2 = _mm256_sub_epi16(s2, r2); + s3 = _mm256_sub_epi16(s3, r3); + + s0 = _mm256_abs_epi16(s0); + s1 = _mm256_abs_epi16(s1); + s2 = _mm256_abs_epi16(s2); + s3 = _mm256_abs_epi16(s3); + + s0 = _mm256_add_epi16(s0, s1); + s0 = _mm256_add_epi16(s0, s2); + s0 = _mm256_add_epi16(s0, s3); + + r0 = _mm256_unpacklo_epi16(s0, zero); + r1 = _mm256_unpackhi_epi16(s0, zero); + + r0 = _mm256_add_epi32(r0, r1); + *sad_acc = _mm256_add_epi32(*sad_acc, r0); + + row_sections += 1; + src_ptr += src_stride << 1; + ref_ptr += ref_stride << 1; + if (sec_ptr) sec_ptr += 32 << 1; + } +} + +unsigned int aom_highbd_sad32x16_avx2(const uint8_t *src, int src_stride, + const uint8_t *ref, int ref_stride) { + __m256i sad = _mm256_setzero_si256(); + uint16_t *srcp = CONVERT_TO_SHORTPTR(src); + uint16_t *refp = CONVERT_TO_SHORTPTR(ref); + const int left_shift = 2; + int row_section = 0; + + while (row_section < 4) { + sad32x4(srcp, src_stride, refp, ref_stride, NULL, &sad); + srcp += src_stride << left_shift; + refp += ref_stride << left_shift; + row_section += 1; + } + return get_sad_from_mm256_epi32(&sad); +} + +unsigned int aom_highbd_sad16x32_avx2(const uint8_t *src, int src_stride, + const uint8_t *ref, int ref_stride) { + uint32_t sum = aom_highbd_sad16x16_avx2(src, src_stride, ref, ref_stride); + src += src_stride << 4; + ref += ref_stride << 4; + sum += aom_highbd_sad16x16_avx2(src, src_stride, ref, ref_stride); + return sum; +} + +unsigned int aom_highbd_sad32x32_avx2(const uint8_t *src, int src_stride, + const uint8_t *ref, int ref_stride) { + uint32_t sum = aom_highbd_sad32x16_avx2(src, src_stride, ref, ref_stride); + src += src_stride << 4; + ref += ref_stride << 4; + sum += aom_highbd_sad32x16_avx2(src, src_stride, ref, ref_stride); + return sum; +} + +unsigned int aom_highbd_sad32x64_avx2(const uint8_t *src, int src_stride, + const uint8_t *ref, int ref_stride) { + uint32_t sum = aom_highbd_sad32x32_avx2(src, src_stride, ref, ref_stride); + src += src_stride << 5; + ref += ref_stride << 5; + sum += aom_highbd_sad32x32_avx2(src, src_stride, ref, ref_stride); + return sum; +} + +static void sad64x2(const uint16_t *src_ptr, int src_stride, + const uint16_t *ref_ptr, int ref_stride, + const uint16_t *sec_ptr, __m256i *sad_acc) { + __m256i s[8], r[8]; + const __m256i zero = _mm256_setzero_si256(); + + s[0] = _mm256_loadu_si256((const __m256i *)src_ptr); + s[1] = _mm256_loadu_si256((const __m256i *)(src_ptr + 16)); + s[2] = _mm256_loadu_si256((const __m256i *)(src_ptr + 32)); + s[3] = _mm256_loadu_si256((const __m256i *)(src_ptr + 48)); + s[4] = _mm256_loadu_si256((const __m256i *)(src_ptr + src_stride)); + s[5] = _mm256_loadu_si256((const __m256i *)(src_ptr + src_stride + 16)); + s[6] = _mm256_loadu_si256((const __m256i *)(src_ptr + src_stride + 32)); + s[7] = _mm256_loadu_si256((const __m256i *)(src_ptr + src_stride + 48)); + + r[0] = _mm256_loadu_si256((const __m256i *)ref_ptr); + r[1] = _mm256_loadu_si256((const __m256i *)(ref_ptr + 16)); + r[2] = _mm256_loadu_si256((const __m256i *)(ref_ptr + 32)); + r[3] = _mm256_loadu_si256((const __m256i *)(ref_ptr + 48)); + r[4] = _mm256_loadu_si256((const __m256i *)(ref_ptr + ref_stride)); + r[5] = _mm256_loadu_si256((const __m256i *)(ref_ptr + ref_stride + 16)); + r[6] = _mm256_loadu_si256((const __m256i *)(ref_ptr + ref_stride + 32)); + r[7] = _mm256_loadu_si256((const __m256i *)(ref_ptr + ref_stride + 48)); + + if (sec_ptr) { + r[0] = _mm256_avg_epu16(r[0], _mm256_loadu_si256((const __m256i *)sec_ptr)); + r[1] = _mm256_avg_epu16( + r[1], _mm256_loadu_si256((const __m256i *)(sec_ptr + 16))); + r[2] = _mm256_avg_epu16( + r[2], _mm256_loadu_si256((const __m256i *)(sec_ptr + 32))); + r[3] = _mm256_avg_epu16( + r[3], _mm256_loadu_si256((const __m256i *)(sec_ptr + 48))); + r[4] = _mm256_avg_epu16( + r[4], _mm256_loadu_si256((const __m256i *)(sec_ptr + 64))); + r[5] = _mm256_avg_epu16( + r[5], _mm256_loadu_si256((const __m256i *)(sec_ptr + 80))); + r[6] = _mm256_avg_epu16( + r[6], _mm256_loadu_si256((const __m256i *)(sec_ptr + 96))); + r[7] = _mm256_avg_epu16( + r[7], _mm256_loadu_si256((const __m256i *)(sec_ptr + 112))); + } + + s[0] = _mm256_sub_epi16(s[0], r[0]); + s[1] = _mm256_sub_epi16(s[1], r[1]); + s[2] = _mm256_sub_epi16(s[2], r[2]); + s[3] = _mm256_sub_epi16(s[3], r[3]); + s[4] = _mm256_sub_epi16(s[4], r[4]); + s[5] = _mm256_sub_epi16(s[5], r[5]); + s[6] = _mm256_sub_epi16(s[6], r[6]); + s[7] = _mm256_sub_epi16(s[7], r[7]); + + s[0] = _mm256_abs_epi16(s[0]); + s[1] = _mm256_abs_epi16(s[1]); + s[2] = _mm256_abs_epi16(s[2]); + s[3] = _mm256_abs_epi16(s[3]); + s[4] = _mm256_abs_epi16(s[4]); + s[5] = _mm256_abs_epi16(s[5]); + s[6] = _mm256_abs_epi16(s[6]); + s[7] = _mm256_abs_epi16(s[7]); + + s[0] = _mm256_add_epi16(s[0], s[1]); + s[0] = _mm256_add_epi16(s[0], s[2]); + s[0] = _mm256_add_epi16(s[0], s[3]); + + s[4] = _mm256_add_epi16(s[4], s[5]); + s[4] = _mm256_add_epi16(s[4], s[6]); + s[4] = _mm256_add_epi16(s[4], s[7]); + + r[0] = _mm256_unpacklo_epi16(s[0], zero); + r[1] = _mm256_unpackhi_epi16(s[0], zero); + r[2] = _mm256_unpacklo_epi16(s[4], zero); + r[3] = _mm256_unpackhi_epi16(s[4], zero); + + r[0] = _mm256_add_epi32(r[0], r[1]); + r[0] = _mm256_add_epi32(r[0], r[2]); + r[0] = _mm256_add_epi32(r[0], r[3]); + *sad_acc = _mm256_add_epi32(*sad_acc, r[0]); +} + +unsigned int aom_highbd_sad64x32_avx2(const uint8_t *src, int src_stride, + const uint8_t *ref, int ref_stride) { + __m256i sad = _mm256_setzero_si256(); + uint16_t *srcp = CONVERT_TO_SHORTPTR(src); + uint16_t *refp = CONVERT_TO_SHORTPTR(ref); + const int left_shift = 1; + int row_section = 0; + + while (row_section < 16) { + sad64x2(srcp, src_stride, refp, ref_stride, NULL, &sad); + srcp += src_stride << left_shift; + refp += ref_stride << left_shift; + row_section += 1; + } + return get_sad_from_mm256_epi32(&sad); +} + +unsigned int aom_highbd_sad64x64_avx2(const uint8_t *src, int src_stride, + const uint8_t *ref, int ref_stride) { + uint32_t sum = aom_highbd_sad64x32_avx2(src, src_stride, ref, ref_stride); + src += src_stride << 5; + ref += ref_stride << 5; + sum += aom_highbd_sad64x32_avx2(src, src_stride, ref, ref_stride); + return sum; +} + +#if CONFIG_EXT_PARTITION +static void sad128x1(const uint16_t *src_ptr, const uint16_t *ref_ptr, + const uint16_t *sec_ptr, __m256i *sad_acc) { + __m256i s[8], r[8]; + const __m256i zero = _mm256_setzero_si256(); + + s[0] = _mm256_loadu_si256((const __m256i *)src_ptr); + s[1] = _mm256_loadu_si256((const __m256i *)(src_ptr + 16)); + s[2] = _mm256_loadu_si256((const __m256i *)(src_ptr + 32)); + s[3] = _mm256_loadu_si256((const __m256i *)(src_ptr + 48)); + s[4] = _mm256_loadu_si256((const __m256i *)(src_ptr + 64)); + s[5] = _mm256_loadu_si256((const __m256i *)(src_ptr + 80)); + s[6] = _mm256_loadu_si256((const __m256i *)(src_ptr + 96)); + s[7] = _mm256_loadu_si256((const __m256i *)(src_ptr + 112)); + + r[0] = _mm256_loadu_si256((const __m256i *)ref_ptr); + r[1] = _mm256_loadu_si256((const __m256i *)(ref_ptr + 16)); + r[2] = _mm256_loadu_si256((const __m256i *)(ref_ptr + 32)); + r[3] = _mm256_loadu_si256((const __m256i *)(ref_ptr + 48)); + r[4] = _mm256_loadu_si256((const __m256i *)(ref_ptr + 64)); + r[5] = _mm256_loadu_si256((const __m256i *)(ref_ptr + 80)); + r[6] = _mm256_loadu_si256((const __m256i *)(ref_ptr + 96)); + r[7] = _mm256_loadu_si256((const __m256i *)(ref_ptr + 112)); + + if (sec_ptr) { + r[0] = _mm256_avg_epu16(r[0], _mm256_loadu_si256((const __m256i *)sec_ptr)); + r[1] = _mm256_avg_epu16( + r[1], _mm256_loadu_si256((const __m256i *)(sec_ptr + 16))); + r[2] = _mm256_avg_epu16( + r[2], _mm256_loadu_si256((const __m256i *)(sec_ptr + 32))); + r[3] = _mm256_avg_epu16( + r[3], _mm256_loadu_si256((const __m256i *)(sec_ptr + 48))); + r[4] = _mm256_avg_epu16( + r[4], _mm256_loadu_si256((const __m256i *)(sec_ptr + 64))); + r[5] = _mm256_avg_epu16( + r[5], _mm256_loadu_si256((const __m256i *)(sec_ptr + 80))); + r[6] = _mm256_avg_epu16( + r[6], _mm256_loadu_si256((const __m256i *)(sec_ptr + 96))); + r[7] = _mm256_avg_epu16( + r[7], _mm256_loadu_si256((const __m256i *)(sec_ptr + 112))); + } + + s[0] = _mm256_sub_epi16(s[0], r[0]); + s[1] = _mm256_sub_epi16(s[1], r[1]); + s[2] = _mm256_sub_epi16(s[2], r[2]); + s[3] = _mm256_sub_epi16(s[3], r[3]); + s[4] = _mm256_sub_epi16(s[4], r[4]); + s[5] = _mm256_sub_epi16(s[5], r[5]); + s[6] = _mm256_sub_epi16(s[6], r[6]); + s[7] = _mm256_sub_epi16(s[7], r[7]); + + s[0] = _mm256_abs_epi16(s[0]); + s[1] = _mm256_abs_epi16(s[1]); + s[2] = _mm256_abs_epi16(s[2]); + s[3] = _mm256_abs_epi16(s[3]); + s[4] = _mm256_abs_epi16(s[4]); + s[5] = _mm256_abs_epi16(s[5]); + s[6] = _mm256_abs_epi16(s[6]); + s[7] = _mm256_abs_epi16(s[7]); + + s[0] = _mm256_add_epi16(s[0], s[1]); + s[0] = _mm256_add_epi16(s[0], s[2]); + s[0] = _mm256_add_epi16(s[0], s[3]); + + s[4] = _mm256_add_epi16(s[4], s[5]); + s[4] = _mm256_add_epi16(s[4], s[6]); + s[4] = _mm256_add_epi16(s[4], s[7]); + + r[0] = _mm256_unpacklo_epi16(s[0], zero); + r[1] = _mm256_unpackhi_epi16(s[0], zero); + r[2] = _mm256_unpacklo_epi16(s[4], zero); + r[3] = _mm256_unpackhi_epi16(s[4], zero); + + r[0] = _mm256_add_epi32(r[0], r[1]); + r[0] = _mm256_add_epi32(r[0], r[2]); + r[0] = _mm256_add_epi32(r[0], r[3]); + *sad_acc = _mm256_add_epi32(*sad_acc, r[0]); +} + +unsigned int aom_highbd_sad128x64_avx2(const uint8_t *src, int src_stride, + const uint8_t *ref, int ref_stride) { + __m256i sad = _mm256_setzero_si256(); + uint16_t *srcp = CONVERT_TO_SHORTPTR(src); + uint16_t *refp = CONVERT_TO_SHORTPTR(ref); + int row = 0; + while (row < 64) { + sad128x1(srcp, refp, NULL, &sad); + srcp += src_stride; + refp += ref_stride; + row += 1; + } + return get_sad_from_mm256_epi32(&sad); +} + +unsigned int aom_highbd_sad64x128_avx2(const uint8_t *src, int src_stride, + const uint8_t *ref, int ref_stride) { + uint32_t sum = aom_highbd_sad64x64_avx2(src, src_stride, ref, ref_stride); + src += src_stride << 6; + ref += ref_stride << 6; + sum += aom_highbd_sad64x64_avx2(src, src_stride, ref, ref_stride); + return sum; +} + +unsigned int aom_highbd_sad128x128_avx2(const uint8_t *src, int src_stride, + const uint8_t *ref, int ref_stride) { + uint32_t sum = aom_highbd_sad128x64_avx2(src, src_stride, ref, ref_stride); + src += src_stride << 6; + ref += ref_stride << 6; + sum += aom_highbd_sad128x64_avx2(src, src_stride, ref, ref_stride); + return sum; +} +#endif // CONFIG_EXT_PARTITION + +// If sec_ptr = 0, calculate regular SAD. Otherwise, calculate average SAD. +static INLINE void sad16x4(const uint16_t *src_ptr, int src_stride, + const uint16_t *ref_ptr, int ref_stride, + const uint16_t *sec_ptr, __m256i *sad_acc) { + __m256i s0, s1, s2, s3, r0, r1, r2, r3; + const __m256i zero = _mm256_setzero_si256(); + + s0 = _mm256_loadu_si256((const __m256i *)src_ptr); + s1 = _mm256_loadu_si256((const __m256i *)(src_ptr + src_stride)); + s2 = _mm256_loadu_si256((const __m256i *)(src_ptr + 2 * src_stride)); + s3 = _mm256_loadu_si256((const __m256i *)(src_ptr + 3 * src_stride)); + + r0 = _mm256_loadu_si256((const __m256i *)ref_ptr); + r1 = _mm256_loadu_si256((const __m256i *)(ref_ptr + ref_stride)); + r2 = _mm256_loadu_si256((const __m256i *)(ref_ptr + 2 * ref_stride)); + r3 = _mm256_loadu_si256((const __m256i *)(ref_ptr + 3 * ref_stride)); + + if (sec_ptr) { + r0 = _mm256_avg_epu16(r0, _mm256_loadu_si256((const __m256i *)sec_ptr)); + r1 = _mm256_avg_epu16(r1, + _mm256_loadu_si256((const __m256i *)(sec_ptr + 16))); + r2 = _mm256_avg_epu16(r2, + _mm256_loadu_si256((const __m256i *)(sec_ptr + 32))); + r3 = _mm256_avg_epu16(r3, + _mm256_loadu_si256((const __m256i *)(sec_ptr + 48))); + } + + s0 = _mm256_sub_epi16(s0, r0); + s1 = _mm256_sub_epi16(s1, r1); + s2 = _mm256_sub_epi16(s2, r2); + s3 = _mm256_sub_epi16(s3, r3); + + s0 = _mm256_abs_epi16(s0); + s1 = _mm256_abs_epi16(s1); + s2 = _mm256_abs_epi16(s2); + s3 = _mm256_abs_epi16(s3); + + s0 = _mm256_add_epi16(s0, s1); + s0 = _mm256_add_epi16(s0, s2); + s0 = _mm256_add_epi16(s0, s3); + + r0 = _mm256_unpacklo_epi16(s0, zero); + r1 = _mm256_unpackhi_epi16(s0, zero); + + r0 = _mm256_add_epi32(r0, r1); + *sad_acc = _mm256_add_epi32(*sad_acc, r0); +} + +unsigned int aom_highbd_sad16x8_avg_avx2(const uint8_t *src, int src_stride, + const uint8_t *ref, int ref_stride, + const uint8_t *second_pred) { + __m256i sad = _mm256_setzero_si256(); + uint16_t *srcp = CONVERT_TO_SHORTPTR(src); + uint16_t *refp = CONVERT_TO_SHORTPTR(ref); + uint16_t *secp = CONVERT_TO_SHORTPTR(second_pred); + + sad16x4(srcp, src_stride, refp, ref_stride, secp, &sad); + + // Next 4 rows + srcp += src_stride << 2; + refp += ref_stride << 2; + secp += 64; + sad16x4(srcp, src_stride, refp, ref_stride, secp, &sad); + return get_sad_from_mm256_epi32(&sad); +} + +unsigned int aom_highbd_sad16x16_avg_avx2(const uint8_t *src, int src_stride, + const uint8_t *ref, int ref_stride, + const uint8_t *second_pred) { + const int left_shift = 3; + uint32_t sum = aom_highbd_sad16x8_avg_avx2(src, src_stride, ref, ref_stride, + second_pred); + src += src_stride << left_shift; + ref += ref_stride << left_shift; + second_pred += 16 << left_shift; + sum += aom_highbd_sad16x8_avg_avx2(src, src_stride, ref, ref_stride, + second_pred); + return sum; +} + +unsigned int aom_highbd_sad16x32_avg_avx2(const uint8_t *src, int src_stride, + const uint8_t *ref, int ref_stride, + const uint8_t *second_pred) { + const int left_shift = 4; + uint32_t sum = aom_highbd_sad16x16_avg_avx2(src, src_stride, ref, ref_stride, + second_pred); + src += src_stride << left_shift; + ref += ref_stride << left_shift; + second_pred += 16 << left_shift; + sum += aom_highbd_sad16x16_avg_avx2(src, src_stride, ref, ref_stride, + second_pred); + return sum; +} + +unsigned int aom_highbd_sad32x16_avg_avx2(const uint8_t *src, int src_stride, + const uint8_t *ref, int ref_stride, + const uint8_t *second_pred) { + __m256i sad = _mm256_setzero_si256(); + uint16_t *srcp = CONVERT_TO_SHORTPTR(src); + uint16_t *refp = CONVERT_TO_SHORTPTR(ref); + uint16_t *secp = CONVERT_TO_SHORTPTR(second_pred); + const int left_shift = 2; + int row_section = 0; + + while (row_section < 4) { + sad32x4(srcp, src_stride, refp, ref_stride, secp, &sad); + srcp += src_stride << left_shift; + refp += ref_stride << left_shift; + secp += 32 << left_shift; + row_section += 1; + } + return get_sad_from_mm256_epi32(&sad); +} + +unsigned int aom_highbd_sad32x32_avg_avx2(const uint8_t *src, int src_stride, + const uint8_t *ref, int ref_stride, + const uint8_t *second_pred) { + const int left_shift = 4; + uint32_t sum = aom_highbd_sad32x16_avg_avx2(src, src_stride, ref, ref_stride, + second_pred); + src += src_stride << left_shift; + ref += ref_stride << left_shift; + second_pred += 32 << left_shift; + sum += aom_highbd_sad32x16_avg_avx2(src, src_stride, ref, ref_stride, + second_pred); + return sum; +} + +unsigned int aom_highbd_sad32x64_avg_avx2(const uint8_t *src, int src_stride, + const uint8_t *ref, int ref_stride, + const uint8_t *second_pred) { + const int left_shift = 5; + uint32_t sum = aom_highbd_sad32x32_avg_avx2(src, src_stride, ref, ref_stride, + second_pred); + src += src_stride << left_shift; + ref += ref_stride << left_shift; + second_pred += 32 << left_shift; + sum += aom_highbd_sad32x32_avg_avx2(src, src_stride, ref, ref_stride, + second_pred); + return sum; +} + +unsigned int aom_highbd_sad64x32_avg_avx2(const uint8_t *src, int src_stride, + const uint8_t *ref, int ref_stride, + const uint8_t *second_pred) { + __m256i sad = _mm256_setzero_si256(); + uint16_t *srcp = CONVERT_TO_SHORTPTR(src); + uint16_t *refp = CONVERT_TO_SHORTPTR(ref); + uint16_t *secp = CONVERT_TO_SHORTPTR(second_pred); + const int left_shift = 1; + int row_section = 0; + + while (row_section < 16) { + sad64x2(srcp, src_stride, refp, ref_stride, secp, &sad); + srcp += src_stride << left_shift; + refp += ref_stride << left_shift; + secp += 64 << left_shift; + row_section += 1; + } + return get_sad_from_mm256_epi32(&sad); +} + +unsigned int aom_highbd_sad64x64_avg_avx2(const uint8_t *src, int src_stride, + const uint8_t *ref, int ref_stride, + const uint8_t *second_pred) { + const int left_shift = 5; + uint32_t sum = aom_highbd_sad64x32_avg_avx2(src, src_stride, ref, ref_stride, + second_pred); + src += src_stride << left_shift; + ref += ref_stride << left_shift; + second_pred += 64 << left_shift; + sum += aom_highbd_sad64x32_avg_avx2(src, src_stride, ref, ref_stride, + second_pred); + return sum; +} + +#if CONFIG_EXT_PARTITION +unsigned int aom_highbd_sad64x128_avg_avx2(const uint8_t *src, int src_stride, + const uint8_t *ref, int ref_stride, + const uint8_t *second_pred) { + const int left_shift = 6; + uint32_t sum = aom_highbd_sad64x64_avg_avx2(src, src_stride, ref, ref_stride, + second_pred); + src += src_stride << left_shift; + ref += ref_stride << left_shift; + second_pred += 64 << left_shift; + sum += aom_highbd_sad64x64_avg_avx2(src, src_stride, ref, ref_stride, + second_pred); + return sum; +} + +unsigned int aom_highbd_sad128x64_avg_avx2(const uint8_t *src, int src_stride, + const uint8_t *ref, int ref_stride, + const uint8_t *second_pred) { + __m256i sad = _mm256_setzero_si256(); + uint16_t *srcp = CONVERT_TO_SHORTPTR(src); + uint16_t *refp = CONVERT_TO_SHORTPTR(ref); + uint16_t *secp = CONVERT_TO_SHORTPTR(second_pred); + int row = 0; + while (row < 64) { + sad128x1(srcp, refp, secp, &sad); + srcp += src_stride; + refp += ref_stride; + secp += 16 << 3; + row += 1; + } + return get_sad_from_mm256_epi32(&sad); +} + +unsigned int aom_highbd_sad128x128_avg_avx2(const uint8_t *src, int src_stride, + const uint8_t *ref, int ref_stride, + const uint8_t *second_pred) { + unsigned int sum; + const int left_shift = 6; + + sum = aom_highbd_sad128x64_avg_avx2(src, src_stride, ref, ref_stride, + second_pred); + src += src_stride << left_shift; + ref += ref_stride << left_shift; + second_pred += 128 << left_shift; + sum += aom_highbd_sad128x64_avg_avx2(src, src_stride, ref, ref_stride, + second_pred); + return sum; +} +#endif // CONFIG_EXT_PARTITION + +// SAD 4D +// Combine 4 __m256i vectors to uint32_t result[4] +static INLINE void get_4d_sad_from_mm256_epi32(const __m256i *v, + uint32_t *res) { + __m256i u0, u1, u2, u3; + const __m256i mask = _mm256_set1_epi64x(UINT32_MAX); + __m128i sad; + + // 8 32-bit summation + u0 = _mm256_srli_si256(v[0], 4); + u1 = _mm256_srli_si256(v[1], 4); + u2 = _mm256_srli_si256(v[2], 4); + u3 = _mm256_srli_si256(v[3], 4); + + u0 = _mm256_add_epi32(u0, v[0]); + u1 = _mm256_add_epi32(u1, v[1]); + u2 = _mm256_add_epi32(u2, v[2]); + u3 = _mm256_add_epi32(u3, v[3]); + + u0 = _mm256_and_si256(u0, mask); + u1 = _mm256_and_si256(u1, mask); + u2 = _mm256_and_si256(u2, mask); + u3 = _mm256_and_si256(u3, mask); + // 4 32-bit summation, evenly positioned + + u1 = _mm256_slli_si256(u1, 4); + u3 = _mm256_slli_si256(u3, 4); + + u0 = _mm256_or_si256(u0, u1); + u2 = _mm256_or_si256(u2, u3); + // 8 32-bit summation, interleaved + + u1 = _mm256_unpacklo_epi64(u0, u2); + u3 = _mm256_unpackhi_epi64(u0, u2); + + u0 = _mm256_add_epi32(u1, u3); + sad = _mm_add_epi32(_mm256_extractf128_si256(u0, 1), + _mm256_castsi256_si128(u0)); + _mm_storeu_si128((__m128i *)res, sad); +} + +static void convert_pointers(const uint8_t *const ref8[], + const uint16_t *ref[]) { + ref[0] = CONVERT_TO_SHORTPTR(ref8[0]); + ref[1] = CONVERT_TO_SHORTPTR(ref8[1]); + ref[2] = CONVERT_TO_SHORTPTR(ref8[2]); + ref[3] = CONVERT_TO_SHORTPTR(ref8[3]); +} + +static void init_sad(__m256i *s) { + s[0] = _mm256_setzero_si256(); + s[1] = _mm256_setzero_si256(); + s[2] = _mm256_setzero_si256(); + s[3] = _mm256_setzero_si256(); +} + +void aom_highbd_sad16x8x4d_avx2(const uint8_t *src, int src_stride, + const uint8_t *const ref_array[], + int ref_stride, uint32_t *sad_array) { + __m256i sad_vec[4]; + const uint16_t *refp[4]; + const uint16_t *keep = CONVERT_TO_SHORTPTR(src); + const uint16_t *srcp; + const int shift_for_4_rows = 2; + int i; + + init_sad(sad_vec); + convert_pointers(ref_array, refp); + + for (i = 0; i < 4; ++i) { + srcp = keep; + sad16x4(srcp, src_stride, refp[i], ref_stride, 0, &sad_vec[i]); + srcp += src_stride << shift_for_4_rows; + refp[i] += ref_stride << shift_for_4_rows; + sad16x4(srcp, src_stride, refp[i], ref_stride, 0, &sad_vec[i]); + } + get_4d_sad_from_mm256_epi32(sad_vec, sad_array); +} + +void aom_highbd_sad16x16x4d_avx2(const uint8_t *src, int src_stride, + const uint8_t *const ref_array[], + int ref_stride, uint32_t *sad_array) { + uint32_t first8rows[4]; + uint32_t second8rows[4]; + const uint8_t *ref[4]; + const int shift_for_8_rows = 3; + + ref[0] = ref_array[0]; + ref[1] = ref_array[1]; + ref[2] = ref_array[2]; + ref[3] = ref_array[3]; + + aom_highbd_sad16x8x4d_avx2(src, src_stride, ref, ref_stride, first8rows); + src += src_stride << shift_for_8_rows; + ref[0] += ref_stride << shift_for_8_rows; + ref[1] += ref_stride << shift_for_8_rows; + ref[2] += ref_stride << shift_for_8_rows; + ref[3] += ref_stride << shift_for_8_rows; + aom_highbd_sad16x8x4d_avx2(src, src_stride, ref, ref_stride, second8rows); + sad_array[0] = first8rows[0] + second8rows[0]; + sad_array[1] = first8rows[1] + second8rows[1]; + sad_array[2] = first8rows[2] + second8rows[2]; + sad_array[3] = first8rows[3] + second8rows[3]; +} + +void aom_highbd_sad16x32x4d_avx2(const uint8_t *src, int src_stride, + const uint8_t *const ref_array[], + int ref_stride, uint32_t *sad_array) { + uint32_t first_half[4]; + uint32_t second_half[4]; + const uint8_t *ref[4]; + const int shift_for_rows = 4; + + ref[0] = ref_array[0]; + ref[1] = ref_array[1]; + ref[2] = ref_array[2]; + ref[3] = ref_array[3]; + + aom_highbd_sad16x16x4d_avx2(src, src_stride, ref, ref_stride, first_half); + src += src_stride << shift_for_rows; + ref[0] += ref_stride << shift_for_rows; + ref[1] += ref_stride << shift_for_rows; + ref[2] += ref_stride << shift_for_rows; + ref[3] += ref_stride << shift_for_rows; + aom_highbd_sad16x16x4d_avx2(src, src_stride, ref, ref_stride, second_half); + sad_array[0] = first_half[0] + second_half[0]; + sad_array[1] = first_half[1] + second_half[1]; + sad_array[2] = first_half[2] + second_half[2]; + sad_array[3] = first_half[3] + second_half[3]; +} + +void aom_highbd_sad32x16x4d_avx2(const uint8_t *src, int src_stride, + const uint8_t *const ref_array[], + int ref_stride, uint32_t *sad_array) { + __m256i sad_vec[4]; + const uint16_t *refp[4]; + const uint16_t *keep = CONVERT_TO_SHORTPTR(src); + const uint16_t *srcp; + const int shift_for_4_rows = 2; + int i; + int rows_section; + + init_sad(sad_vec); + convert_pointers(ref_array, refp); + + for (i = 0; i < 4; ++i) { + srcp = keep; + rows_section = 0; + while (rows_section < 4) { + sad32x4(srcp, src_stride, refp[i], ref_stride, 0, &sad_vec[i]); + srcp += src_stride << shift_for_4_rows; + refp[i] += ref_stride << shift_for_4_rows; + rows_section++; + } + } + get_4d_sad_from_mm256_epi32(sad_vec, sad_array); +} + +void aom_highbd_sad32x32x4d_avx2(const uint8_t *src, int src_stride, + const uint8_t *const ref_array[], + int ref_stride, uint32_t *sad_array) { + uint32_t first_half[4]; + uint32_t second_half[4]; + const uint8_t *ref[4]; + const int shift_for_rows = 4; + + ref[0] = ref_array[0]; + ref[1] = ref_array[1]; + ref[2] = ref_array[2]; + ref[3] = ref_array[3]; + + aom_highbd_sad32x16x4d_avx2(src, src_stride, ref, ref_stride, first_half); + src += src_stride << shift_for_rows; + ref[0] += ref_stride << shift_for_rows; + ref[1] += ref_stride << shift_for_rows; + ref[2] += ref_stride << shift_for_rows; + ref[3] += ref_stride << shift_for_rows; + aom_highbd_sad32x16x4d_avx2(src, src_stride, ref, ref_stride, second_half); + sad_array[0] = first_half[0] + second_half[0]; + sad_array[1] = first_half[1] + second_half[1]; + sad_array[2] = first_half[2] + second_half[2]; + sad_array[3] = first_half[3] + second_half[3]; +} + +void aom_highbd_sad32x64x4d_avx2(const uint8_t *src, int src_stride, + const uint8_t *const ref_array[], + int ref_stride, uint32_t *sad_array) { + uint32_t first_half[4]; + uint32_t second_half[4]; + const uint8_t *ref[4]; + const int shift_for_rows = 5; + + ref[0] = ref_array[0]; + ref[1] = ref_array[1]; + ref[2] = ref_array[2]; + ref[3] = ref_array[3]; + + aom_highbd_sad32x32x4d_avx2(src, src_stride, ref, ref_stride, first_half); + src += src_stride << shift_for_rows; + ref[0] += ref_stride << shift_for_rows; + ref[1] += ref_stride << shift_for_rows; + ref[2] += ref_stride << shift_for_rows; + ref[3] += ref_stride << shift_for_rows; + aom_highbd_sad32x32x4d_avx2(src, src_stride, ref, ref_stride, second_half); + sad_array[0] = first_half[0] + second_half[0]; + sad_array[1] = first_half[1] + second_half[1]; + sad_array[2] = first_half[2] + second_half[2]; + sad_array[3] = first_half[3] + second_half[3]; +} + +void aom_highbd_sad64x32x4d_avx2(const uint8_t *src, int src_stride, + const uint8_t *const ref_array[], + int ref_stride, uint32_t *sad_array) { + __m256i sad_vec[4]; + const uint16_t *refp[4]; + const uint16_t *keep = CONVERT_TO_SHORTPTR(src); + const uint16_t *srcp; + const int shift_for_rows = 1; + int i; + int rows_section; + + init_sad(sad_vec); + convert_pointers(ref_array, refp); + + for (i = 0; i < 4; ++i) { + srcp = keep; + rows_section = 0; + while (rows_section < 16) { + sad64x2(srcp, src_stride, refp[i], ref_stride, NULL, &sad_vec[i]); + srcp += src_stride << shift_for_rows; + refp[i] += ref_stride << shift_for_rows; + rows_section++; + } + } + get_4d_sad_from_mm256_epi32(sad_vec, sad_array); +} + +void aom_highbd_sad64x64x4d_avx2(const uint8_t *src, int src_stride, + const uint8_t *const ref_array[], + int ref_stride, uint32_t *sad_array) { + uint32_t first_half[4]; + uint32_t second_half[4]; + const uint8_t *ref[4]; + const int shift_for_rows = 5; + + ref[0] = ref_array[0]; + ref[1] = ref_array[1]; + ref[2] = ref_array[2]; + ref[3] = ref_array[3]; + + aom_highbd_sad64x32x4d_avx2(src, src_stride, ref, ref_stride, first_half); + src += src_stride << shift_for_rows; + ref[0] += ref_stride << shift_for_rows; + ref[1] += ref_stride << shift_for_rows; + ref[2] += ref_stride << shift_for_rows; + ref[3] += ref_stride << shift_for_rows; + aom_highbd_sad64x32x4d_avx2(src, src_stride, ref, ref_stride, second_half); + sad_array[0] = first_half[0] + second_half[0]; + sad_array[1] = first_half[1] + second_half[1]; + sad_array[2] = first_half[2] + second_half[2]; + sad_array[3] = first_half[3] + second_half[3]; +} + +#if CONFIG_EXT_PARTITION +void aom_highbd_sad64x128x4d_avx2(const uint8_t *src, int src_stride, + const uint8_t *const ref_array[], + int ref_stride, uint32_t *sad_array) { + uint32_t first_half[4]; + uint32_t second_half[4]; + const uint8_t *ref[4]; + const int shift_for_rows = 6; + + ref[0] = ref_array[0]; + ref[1] = ref_array[1]; + ref[2] = ref_array[2]; + ref[3] = ref_array[3]; + + aom_highbd_sad64x64x4d_avx2(src, src_stride, ref, ref_stride, first_half); + src += src_stride << shift_for_rows; + ref[0] += ref_stride << shift_for_rows; + ref[1] += ref_stride << shift_for_rows; + ref[2] += ref_stride << shift_for_rows; + ref[3] += ref_stride << shift_for_rows; + aom_highbd_sad64x64x4d_avx2(src, src_stride, ref, ref_stride, second_half); + sad_array[0] = first_half[0] + second_half[0]; + sad_array[1] = first_half[1] + second_half[1]; + sad_array[2] = first_half[2] + second_half[2]; + sad_array[3] = first_half[3] + second_half[3]; +} + +void aom_highbd_sad128x64x4d_avx2(const uint8_t *src, int src_stride, + const uint8_t *const ref_array[], + int ref_stride, uint32_t *sad_array) { + __m256i sad_vec[4]; + const uint16_t *refp[4]; + const uint16_t *keep = CONVERT_TO_SHORTPTR(src); + const uint16_t *srcp; + int i; + int rows_section; + + init_sad(sad_vec); + convert_pointers(ref_array, refp); + + for (i = 0; i < 4; ++i) { + srcp = keep; + rows_section = 0; + while (rows_section < 64) { + sad128x1(srcp, refp[i], NULL, &sad_vec[i]); + srcp += src_stride; + refp[i] += ref_stride; + rows_section++; + } + } + get_4d_sad_from_mm256_epi32(sad_vec, sad_array); +} + +void aom_highbd_sad128x128x4d_avx2(const uint8_t *src, int src_stride, + const uint8_t *const ref_array[], + int ref_stride, uint32_t *sad_array) { + uint32_t first_half[4]; + uint32_t second_half[4]; + const uint8_t *ref[4]; + const int shift_for_rows = 6; + + ref[0] = ref_array[0]; + ref[1] = ref_array[1]; + ref[2] = ref_array[2]; + ref[3] = ref_array[3]; + + aom_highbd_sad128x64x4d_avx2(src, src_stride, ref, ref_stride, first_half); + src += src_stride << shift_for_rows; + ref[0] += ref_stride << shift_for_rows; + ref[1] += ref_stride << shift_for_rows; + ref[2] += ref_stride << shift_for_rows; + ref[3] += ref_stride << shift_for_rows; + aom_highbd_sad128x64x4d_avx2(src, src_stride, ref, ref_stride, second_half); + sad_array[0] = first_half[0] + second_half[0]; + sad_array[1] = first_half[1] + second_half[1]; + sad_array[2] = first_half[2] + second_half[2]; + sad_array[3] = first_half[3] + second_half[3]; +} +#endif // CONFIG_EXT_PARTITION diff --git a/third_party/aom/aom_dsp/x86/sad_impl_avx2.c b/third_party/aom/aom_dsp/x86/sad_impl_avx2.c new file mode 100644 index 000000000..4419c65b2 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/sad_impl_avx2.c @@ -0,0 +1,233 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include <immintrin.h> +#include "./aom_dsp_rtcd.h" + +static unsigned int sad32x32(const uint8_t *src_ptr, int src_stride, + const uint8_t *ref_ptr, int ref_stride) { + __m256i s1, s2, r1, r2; + __m256i sum = _mm256_setzero_si256(); + __m128i sum_i128; + int i; + + for (i = 0; i < 16; ++i) { + r1 = _mm256_loadu_si256((__m256i const *)ref_ptr); + r2 = _mm256_loadu_si256((__m256i const *)(ref_ptr + ref_stride)); + s1 = _mm256_sad_epu8(r1, _mm256_loadu_si256((__m256i const *)src_ptr)); + s2 = _mm256_sad_epu8( + r2, _mm256_loadu_si256((__m256i const *)(src_ptr + src_stride))); + sum = _mm256_add_epi32(sum, _mm256_add_epi32(s1, s2)); + ref_ptr += ref_stride << 1; + src_ptr += src_stride << 1; + } + + sum = _mm256_add_epi32(sum, _mm256_srli_si256(sum, 8)); + sum_i128 = _mm_add_epi32(_mm256_extracti128_si256(sum, 1), + _mm256_castsi256_si128(sum)); + return _mm_cvtsi128_si32(sum_i128); +} + +static unsigned int sad64x32(const uint8_t *src_ptr, int src_stride, + const uint8_t *ref_ptr, int ref_stride) { + unsigned int half_width = 32; + uint32_t sum = sad32x32(src_ptr, src_stride, ref_ptr, ref_stride); + src_ptr += half_width; + ref_ptr += half_width; + sum += sad32x32(src_ptr, src_stride, ref_ptr, ref_stride); + return sum; +} + +static unsigned int sad64x64(const uint8_t *src_ptr, int src_stride, + const uint8_t *ref_ptr, int ref_stride) { + uint32_t sum = sad64x32(src_ptr, src_stride, ref_ptr, ref_stride); + src_ptr += src_stride << 5; + ref_ptr += ref_stride << 5; + sum += sad64x32(src_ptr, src_stride, ref_ptr, ref_stride); + return sum; +} + +unsigned int aom_sad128x64_avx2(const uint8_t *src_ptr, int src_stride, + const uint8_t *ref_ptr, int ref_stride) { + unsigned int half_width = 64; + uint32_t sum = sad64x64(src_ptr, src_stride, ref_ptr, ref_stride); + src_ptr += half_width; + ref_ptr += half_width; + sum += sad64x64(src_ptr, src_stride, ref_ptr, ref_stride); + return sum; +} + +unsigned int aom_sad64x128_avx2(const uint8_t *src_ptr, int src_stride, + const uint8_t *ref_ptr, int ref_stride) { + uint32_t sum = sad64x64(src_ptr, src_stride, ref_ptr, ref_stride); + src_ptr += src_stride << 6; + ref_ptr += ref_stride << 6; + sum += sad64x64(src_ptr, src_stride, ref_ptr, ref_stride); + return sum; +} + +unsigned int aom_sad128x128_avx2(const uint8_t *src_ptr, int src_stride, + const uint8_t *ref_ptr, int ref_stride) { + uint32_t sum = aom_sad128x64_avx2(src_ptr, src_stride, ref_ptr, ref_stride); + src_ptr += src_stride << 6; + ref_ptr += ref_stride << 6; + sum += aom_sad128x64_avx2(src_ptr, src_stride, ref_ptr, ref_stride); + return sum; +} + +static void sad64x64x4d(const uint8_t *src, int src_stride, + const uint8_t *const ref[4], int ref_stride, + __m128i *res) { + uint32_t sum[4]; + aom_sad64x64x4d_avx2(src, src_stride, ref, ref_stride, sum); + *res = _mm_loadu_si128((const __m128i *)sum); +} + +void aom_sad64x128x4d_avx2(const uint8_t *src, int src_stride, + const uint8_t *const ref[4], int ref_stride, + uint32_t res[4]) { + __m128i sum0, sum1; + const uint8_t *rf[4]; + + rf[0] = ref[0]; + rf[1] = ref[1]; + rf[2] = ref[2]; + rf[3] = ref[3]; + sad64x64x4d(src, src_stride, rf, ref_stride, &sum0); + src += src_stride << 6; + rf[0] += ref_stride << 6; + rf[1] += ref_stride << 6; + rf[2] += ref_stride << 6; + rf[3] += ref_stride << 6; + sad64x64x4d(src, src_stride, rf, ref_stride, &sum1); + sum0 = _mm_add_epi32(sum0, sum1); + _mm_storeu_si128((__m128i *)res, sum0); +} + +void aom_sad128x64x4d_avx2(const uint8_t *src, int src_stride, + const uint8_t *const ref[4], int ref_stride, + uint32_t res[4]) { + __m128i sum0, sum1; + unsigned int half_width = 64; + const uint8_t *rf[4]; + + rf[0] = ref[0]; + rf[1] = ref[1]; + rf[2] = ref[2]; + rf[3] = ref[3]; + sad64x64x4d(src, src_stride, rf, ref_stride, &sum0); + src += half_width; + rf[0] += half_width; + rf[1] += half_width; + rf[2] += half_width; + rf[3] += half_width; + sad64x64x4d(src, src_stride, rf, ref_stride, &sum1); + sum0 = _mm_add_epi32(sum0, sum1); + _mm_storeu_si128((__m128i *)res, sum0); +} + +void aom_sad128x128x4d_avx2(const uint8_t *src, int src_stride, + const uint8_t *const ref[4], int ref_stride, + uint32_t res[4]) { + const uint8_t *rf[4]; + uint32_t sum0[4]; + uint32_t sum1[4]; + + rf[0] = ref[0]; + rf[1] = ref[1]; + rf[2] = ref[2]; + rf[3] = ref[3]; + aom_sad128x64x4d_avx2(src, src_stride, rf, ref_stride, sum0); + src += src_stride << 6; + rf[0] += ref_stride << 6; + rf[1] += ref_stride << 6; + rf[2] += ref_stride << 6; + rf[3] += ref_stride << 6; + aom_sad128x64x4d_avx2(src, src_stride, rf, ref_stride, sum1); + res[0] = sum0[0] + sum1[0]; + res[1] = sum0[1] + sum1[1]; + res[2] = sum0[2] + sum1[2]; + res[3] = sum0[3] + sum1[3]; +} + +static unsigned int sad_w64_avg_avx2(const uint8_t *src_ptr, int src_stride, + const uint8_t *ref_ptr, int ref_stride, + const int h, const uint8_t *second_pred, + const int second_pred_stride) { + int i, res; + __m256i sad1_reg, sad2_reg, ref1_reg, ref2_reg; + __m256i sum_sad = _mm256_setzero_si256(); + __m256i sum_sad_h; + __m128i sum_sad128; + for (i = 0; i < h; i++) { + ref1_reg = _mm256_loadu_si256((__m256i const *)ref_ptr); + ref2_reg = _mm256_loadu_si256((__m256i const *)(ref_ptr + 32)); + ref1_reg = _mm256_avg_epu8( + ref1_reg, _mm256_loadu_si256((__m256i const *)second_pred)); + ref2_reg = _mm256_avg_epu8( + ref2_reg, _mm256_loadu_si256((__m256i const *)(second_pred + 32))); + sad1_reg = + _mm256_sad_epu8(ref1_reg, _mm256_loadu_si256((__m256i const *)src_ptr)); + sad2_reg = _mm256_sad_epu8( + ref2_reg, _mm256_loadu_si256((__m256i const *)(src_ptr + 32))); + sum_sad = _mm256_add_epi32(sum_sad, _mm256_add_epi32(sad1_reg, sad2_reg)); + ref_ptr += ref_stride; + src_ptr += src_stride; + second_pred += second_pred_stride; + } + sum_sad_h = _mm256_srli_si256(sum_sad, 8); + sum_sad = _mm256_add_epi32(sum_sad, sum_sad_h); + sum_sad128 = _mm256_extracti128_si256(sum_sad, 1); + sum_sad128 = _mm_add_epi32(_mm256_castsi256_si128(sum_sad), sum_sad128); + res = _mm_cvtsi128_si32(sum_sad128); + + return res; +} + +unsigned int aom_sad64x128_avg_avx2(const uint8_t *src_ptr, int src_stride, + const uint8_t *ref_ptr, int ref_stride, + const uint8_t *second_pred) { + uint32_t sum = sad_w64_avg_avx2(src_ptr, src_stride, ref_ptr, ref_stride, 64, + second_pred, 64); + src_ptr += src_stride << 6; + ref_ptr += ref_stride << 6; + second_pred += 64 << 6; + sum += sad_w64_avg_avx2(src_ptr, src_stride, ref_ptr, ref_stride, 64, + second_pred, 64); + return sum; +} + +unsigned int aom_sad128x64_avg_avx2(const uint8_t *src_ptr, int src_stride, + const uint8_t *ref_ptr, int ref_stride, + const uint8_t *second_pred) { + unsigned int half_width = 64; + uint32_t sum = sad_w64_avg_avx2(src_ptr, src_stride, ref_ptr, ref_stride, 64, + second_pred, 128); + src_ptr += half_width; + ref_ptr += half_width; + second_pred += half_width; + sum += sad_w64_avg_avx2(src_ptr, src_stride, ref_ptr, ref_stride, 64, + second_pred, 128); + return sum; +} + +unsigned int aom_sad128x128_avg_avx2(const uint8_t *src_ptr, int src_stride, + const uint8_t *ref_ptr, int ref_stride, + const uint8_t *second_pred) { + uint32_t sum = aom_sad128x64_avg_avx2(src_ptr, src_stride, ref_ptr, + ref_stride, second_pred); + src_ptr += src_stride << 6; + ref_ptr += ref_stride << 6; + second_pred += 128 << 6; + sum += aom_sad128x64_avg_avx2(src_ptr, src_stride, ref_ptr, ref_stride, + second_pred); + return sum; +} diff --git a/third_party/aom/aom_dsp/x86/sad_sse2.asm b/third_party/aom/aom_dsp/x86/sad_sse2.asm new file mode 100644 index 000000000..e45457a57 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/sad_sse2.asm @@ -0,0 +1,345 @@ +; +; Copyright (c) 2016, Alliance for Open Media. All rights reserved +; +; This source code is subject to the terms of the BSD 2 Clause License and +; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License +; was not distributed with this source code in the LICENSE file, you can +; obtain it at www.aomedia.org/license/software. If the Alliance for Open +; Media Patent License 1.0 was not distributed with this source code in the +; PATENTS file, you can obtain it at www.aomedia.org/license/patent. +; + +; + +%include "third_party/x86inc/x86inc.asm" + +SECTION .text + +%macro SAD_FN 4 +%if %4 == 0 +%if %3 == 5 +cglobal sad%1x%2, 4, %3, 5, src, src_stride, ref, ref_stride, n_rows +%else ; %3 == 7 +cglobal sad%1x%2, 4, %3, 6, src, src_stride, ref, ref_stride, \ + src_stride3, ref_stride3, n_rows +%endif ; %3 == 5/7 +%else ; avg +%if %3 == 5 +cglobal sad%1x%2_avg, 5, 1 + %3, 5, src, src_stride, ref, ref_stride, \ + second_pred, n_rows +%else ; %3 == 7 +cglobal sad%1x%2_avg, 5, ARCH_X86_64 + %3, 6, src, src_stride, \ + ref, ref_stride, \ + second_pred, \ + src_stride3, ref_stride3 +%if ARCH_X86_64 +%define n_rowsd r7d +%else ; x86-32 +%define n_rowsd dword r0m +%endif ; x86-32/64 +%endif ; %3 == 5/7 +%endif ; avg/sad + movsxdifnidn src_strideq, src_strided + movsxdifnidn ref_strideq, ref_strided +%if %3 == 7 + lea src_stride3q, [src_strideq*3] + lea ref_stride3q, [ref_strideq*3] +%endif ; %3 == 7 +%endmacro + +%if CONFIG_EXT_PARTITION +; unsigned int aom_sad128x128_sse2(uint8_t *src, int src_stride, +; uint8_t *ref, int ref_stride); +%macro SAD128XN 1-2 0 + SAD_FN 128, %1, 5, %2 + mov n_rowsd, %1 + pxor m0, m0 + +.loop: + movu m1, [refq] + movu m2, [refq+16] + movu m3, [refq+32] + movu m4, [refq+48] +%if %2 == 1 + pavgb m1, [second_predq+mmsize*0] + pavgb m2, [second_predq+mmsize*1] + pavgb m3, [second_predq+mmsize*2] + pavgb m4, [second_predq+mmsize*3] +%endif + psadbw m1, [srcq] + psadbw m2, [srcq+16] + psadbw m3, [srcq+32] + psadbw m4, [srcq+48] + + paddd m1, m2 + paddd m3, m4 + paddd m0, m1 + paddd m0, m3 + + movu m1, [refq+64] + movu m2, [refq+80] + movu m3, [refq+96] + movu m4, [refq+112] +%if %2 == 1 + pavgb m1, [second_predq+mmsize*4] + pavgb m2, [second_predq+mmsize*5] + pavgb m3, [second_predq+mmsize*6] + pavgb m4, [second_predq+mmsize*7] + lea second_predq, [second_predq+mmsize*8] +%endif + psadbw m1, [srcq+64] + psadbw m2, [srcq+80] + psadbw m3, [srcq+96] + psadbw m4, [srcq+112] + + add refq, ref_strideq + add srcq, src_strideq + + paddd m1, m2 + paddd m3, m4 + paddd m0, m1 + paddd m0, m3 + + sub n_rowsd, 1 + jg .loop + + movhlps m1, m0 + paddd m0, m1 + movd eax, m0 + RET +%endmacro + +INIT_XMM sse2 +SAD128XN 128 ; sad128x128_sse2 +SAD128XN 128, 1 ; sad128x128_avg_sse2 +SAD128XN 64 ; sad128x64_sse2 +SAD128XN 64, 1 ; sad128x64_avg_sse2 +%endif + + +; unsigned int aom_sad64x64_sse2(uint8_t *src, int src_stride, +; uint8_t *ref, int ref_stride); +%macro SAD64XN 1-2 0 + SAD_FN 64, %1, 5, %2 + mov n_rowsd, %1 + pxor m0, m0 +.loop: + movu m1, [refq] + movu m2, [refq+16] + movu m3, [refq+32] + movu m4, [refq+48] +%if %2 == 1 + pavgb m1, [second_predq+mmsize*0] + pavgb m2, [second_predq+mmsize*1] + pavgb m3, [second_predq+mmsize*2] + pavgb m4, [second_predq+mmsize*3] + lea second_predq, [second_predq+mmsize*4] +%endif + psadbw m1, [srcq] + psadbw m2, [srcq+16] + psadbw m3, [srcq+32] + psadbw m4, [srcq+48] + paddd m1, m2 + paddd m3, m4 + add refq, ref_strideq + paddd m0, m1 + add srcq, src_strideq + paddd m0, m3 + dec n_rowsd + jg .loop + + movhlps m1, m0 + paddd m0, m1 + movd eax, m0 + RET +%endmacro + +INIT_XMM sse2 +%if CONFIG_EXT_PARTITION +SAD64XN 128 ; sad64x128_sse2 +SAD64XN 128, 1 ; sad64x128_avg_sse2 +%endif +SAD64XN 64 ; sad64x64_sse2 +SAD64XN 32 ; sad64x32_sse2 +SAD64XN 64, 1 ; sad64x64_avg_sse2 +SAD64XN 32, 1 ; sad64x32_avg_sse2 + +; unsigned int aom_sad32x32_sse2(uint8_t *src, int src_stride, +; uint8_t *ref, int ref_stride); +%macro SAD32XN 1-2 0 + SAD_FN 32, %1, 5, %2 + mov n_rowsd, %1/2 + pxor m0, m0 +.loop: + movu m1, [refq] + movu m2, [refq+16] + movu m3, [refq+ref_strideq] + movu m4, [refq+ref_strideq+16] +%if %2 == 1 + pavgb m1, [second_predq+mmsize*0] + pavgb m2, [second_predq+mmsize*1] + pavgb m3, [second_predq+mmsize*2] + pavgb m4, [second_predq+mmsize*3] + lea second_predq, [second_predq+mmsize*4] +%endif + psadbw m1, [srcq] + psadbw m2, [srcq+16] + psadbw m3, [srcq+src_strideq] + psadbw m4, [srcq+src_strideq+16] + paddd m1, m2 + paddd m3, m4 + lea refq, [refq+ref_strideq*2] + paddd m0, m1 + lea srcq, [srcq+src_strideq*2] + paddd m0, m3 + dec n_rowsd + jg .loop + + movhlps m1, m0 + paddd m0, m1 + movd eax, m0 + RET +%endmacro + +INIT_XMM sse2 +SAD32XN 64 ; sad32x64_sse2 +SAD32XN 32 ; sad32x32_sse2 +SAD32XN 16 ; sad32x16_sse2 +SAD32XN 64, 1 ; sad32x64_avg_sse2 +SAD32XN 32, 1 ; sad32x32_avg_sse2 +SAD32XN 16, 1 ; sad32x16_avg_sse2 + +; unsigned int aom_sad16x{8,16}_sse2(uint8_t *src, int src_stride, +; uint8_t *ref, int ref_stride); +%macro SAD16XN 1-2 0 + SAD_FN 16, %1, 7, %2 + mov n_rowsd, %1/4 + pxor m0, m0 + +.loop: + movu m1, [refq] + movu m2, [refq+ref_strideq] + movu m3, [refq+ref_strideq*2] + movu m4, [refq+ref_stride3q] +%if %2 == 1 + pavgb m1, [second_predq+mmsize*0] + pavgb m2, [second_predq+mmsize*1] + pavgb m3, [second_predq+mmsize*2] + pavgb m4, [second_predq+mmsize*3] + lea second_predq, [second_predq+mmsize*4] +%endif + psadbw m1, [srcq] + psadbw m2, [srcq+src_strideq] + psadbw m3, [srcq+src_strideq*2] + psadbw m4, [srcq+src_stride3q] + paddd m1, m2 + paddd m3, m4 + lea refq, [refq+ref_strideq*4] + paddd m0, m1 + lea srcq, [srcq+src_strideq*4] + paddd m0, m3 + dec n_rowsd + jg .loop + + movhlps m1, m0 + paddd m0, m1 + movd eax, m0 + RET +%endmacro + +INIT_XMM sse2 +SAD16XN 32 ; sad16x32_sse2 +SAD16XN 16 ; sad16x16_sse2 +SAD16XN 8 ; sad16x8_sse2 +SAD16XN 32, 1 ; sad16x32_avg_sse2 +SAD16XN 16, 1 ; sad16x16_avg_sse2 +SAD16XN 8, 1 ; sad16x8_avg_sse2 + +; unsigned int aom_sad8x{8,16}_sse2(uint8_t *src, int src_stride, +; uint8_t *ref, int ref_stride); +%macro SAD8XN 1-2 0 + SAD_FN 8, %1, 7, %2 + mov n_rowsd, %1/4 + pxor m0, m0 + +.loop: + movh m1, [refq] + movhps m1, [refq+ref_strideq] + movh m2, [refq+ref_strideq*2] + movhps m2, [refq+ref_stride3q] +%if %2 == 1 + pavgb m1, [second_predq+mmsize*0] + pavgb m2, [second_predq+mmsize*1] + lea second_predq, [second_predq+mmsize*2] +%endif + movh m3, [srcq] + movhps m3, [srcq+src_strideq] + movh m4, [srcq+src_strideq*2] + movhps m4, [srcq+src_stride3q] + psadbw m1, m3 + psadbw m2, m4 + lea refq, [refq+ref_strideq*4] + paddd m0, m1 + lea srcq, [srcq+src_strideq*4] + paddd m0, m2 + dec n_rowsd + jg .loop + + movhlps m1, m0 + paddd m0, m1 + movd eax, m0 + RET +%endmacro + +INIT_XMM sse2 +SAD8XN 16 ; sad8x16_sse2 +SAD8XN 8 ; sad8x8_sse2 +SAD8XN 4 ; sad8x4_sse2 +SAD8XN 16, 1 ; sad8x16_avg_sse2 +SAD8XN 8, 1 ; sad8x8_avg_sse2 +SAD8XN 4, 1 ; sad8x4_avg_sse2 + +; unsigned int aom_sad4x{4, 8}_sse2(uint8_t *src, int src_stride, +; uint8_t *ref, int ref_stride); +%macro SAD4XN 1-2 0 + SAD_FN 4, %1, 7, %2 + mov n_rowsd, %1/4 + pxor m0, m0 + +.loop: + movd m1, [refq] + movd m2, [refq+ref_strideq] + movd m3, [refq+ref_strideq*2] + movd m4, [refq+ref_stride3q] + punpckldq m1, m2 + punpckldq m3, m4 + movlhps m1, m3 +%if %2 == 1 + pavgb m1, [second_predq+mmsize*0] + lea second_predq, [second_predq+mmsize*1] +%endif + movd m2, [srcq] + movd m5, [srcq+src_strideq] + movd m4, [srcq+src_strideq*2] + movd m3, [srcq+src_stride3q] + punpckldq m2, m5 + punpckldq m4, m3 + movlhps m2, m4 + psadbw m1, m2 + lea refq, [refq+ref_strideq*4] + paddd m0, m1 + lea srcq, [srcq+src_strideq*4] + dec n_rowsd + jg .loop + + movhlps m1, m0 + paddd m0, m1 + movd eax, m0 + RET +%endmacro + +INIT_XMM sse2 +SAD4XN 8 ; sad4x8_sse +SAD4XN 4 ; sad4x4_sse +SAD4XN 8, 1 ; sad4x8_avg_sse +SAD4XN 4, 1 ; sad4x4_avg_sse diff --git a/third_party/aom/aom_dsp/x86/sad_sse3.asm b/third_party/aom/aom_dsp/x86/sad_sse3.asm new file mode 100644 index 000000000..f6c27c855 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/sad_sse3.asm @@ -0,0 +1,377 @@ +; +; Copyright (c) 2016, Alliance for Open Media. All rights reserved +; +; This source code is subject to the terms of the BSD 2 Clause License and +; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License +; was not distributed with this source code in the LICENSE file, you can +; obtain it at www.aomedia.org/license/software. If the Alliance for Open +; Media Patent License 1.0 was not distributed with this source code in the +; PATENTS file, you can obtain it at www.aomedia.org/license/patent. +; + +; + +%include "aom_ports/x86_abi_support.asm" + +%macro STACK_FRAME_CREATE_X3 0 +%if ABI_IS_32BIT + %define src_ptr rsi + %define src_stride rax + %define ref_ptr rdi + %define ref_stride rdx + %define end_ptr rcx + %define ret_var rbx + %define result_ptr arg(4) + %define height dword ptr arg(4) + push rbp + mov rbp, rsp + push rsi + push rdi + push rbx + + mov rsi, arg(0) ; src_ptr + mov rdi, arg(2) ; ref_ptr + + movsxd rax, dword ptr arg(1) ; src_stride + movsxd rdx, dword ptr arg(3) ; ref_stride +%else + %if LIBAOM_YASM_WIN64 + SAVE_XMM 7, u + %define src_ptr rcx + %define src_stride rdx + %define ref_ptr r8 + %define ref_stride r9 + %define end_ptr r10 + %define ret_var r11 + %define result_ptr [rsp+xmm_stack_space+8+4*8] + %define height dword ptr [rsp+xmm_stack_space+8+4*8] + %else + %define src_ptr rdi + %define src_stride rsi + %define ref_ptr rdx + %define ref_stride rcx + %define end_ptr r9 + %define ret_var r10 + %define result_ptr r8 + %define height r8 + %endif +%endif + +%endmacro + +%macro STACK_FRAME_DESTROY_X3 0 + %define src_ptr + %define src_stride + %define ref_ptr + %define ref_stride + %define end_ptr + %define ret_var + %define result_ptr + %define height + +%if ABI_IS_32BIT + pop rbx + pop rdi + pop rsi + pop rbp +%else + %if LIBAOM_YASM_WIN64 + RESTORE_XMM + %endif +%endif + ret +%endmacro + +%macro PROCESS_16X2X3 5 +%if %1==0 + movdqa xmm0, XMMWORD PTR [%2] + lddqu xmm5, XMMWORD PTR [%3] + lddqu xmm6, XMMWORD PTR [%3+1] + lddqu xmm7, XMMWORD PTR [%3+2] + + psadbw xmm5, xmm0 + psadbw xmm6, xmm0 + psadbw xmm7, xmm0 +%else + movdqa xmm0, XMMWORD PTR [%2] + lddqu xmm1, XMMWORD PTR [%3] + lddqu xmm2, XMMWORD PTR [%3+1] + lddqu xmm3, XMMWORD PTR [%3+2] + + psadbw xmm1, xmm0 + psadbw xmm2, xmm0 + psadbw xmm3, xmm0 + + paddw xmm5, xmm1 + paddw xmm6, xmm2 + paddw xmm7, xmm3 +%endif + movdqa xmm0, XMMWORD PTR [%2+%4] + lddqu xmm1, XMMWORD PTR [%3+%5] + lddqu xmm2, XMMWORD PTR [%3+%5+1] + lddqu xmm3, XMMWORD PTR [%3+%5+2] + +%if %1==0 || %1==1 + lea %2, [%2+%4*2] + lea %3, [%3+%5*2] +%endif + + psadbw xmm1, xmm0 + psadbw xmm2, xmm0 + psadbw xmm3, xmm0 + + paddw xmm5, xmm1 + paddw xmm6, xmm2 + paddw xmm7, xmm3 +%endmacro + +%macro PROCESS_8X2X3 5 +%if %1==0 + movq mm0, QWORD PTR [%2] + movq mm5, QWORD PTR [%3] + movq mm6, QWORD PTR [%3+1] + movq mm7, QWORD PTR [%3+2] + + psadbw mm5, mm0 + psadbw mm6, mm0 + psadbw mm7, mm0 +%else + movq mm0, QWORD PTR [%2] + movq mm1, QWORD PTR [%3] + movq mm2, QWORD PTR [%3+1] + movq mm3, QWORD PTR [%3+2] + + psadbw mm1, mm0 + psadbw mm2, mm0 + psadbw mm3, mm0 + + paddw mm5, mm1 + paddw mm6, mm2 + paddw mm7, mm3 +%endif + movq mm0, QWORD PTR [%2+%4] + movq mm1, QWORD PTR [%3+%5] + movq mm2, QWORD PTR [%3+%5+1] + movq mm3, QWORD PTR [%3+%5+2] + +%if %1==0 || %1==1 + lea %2, [%2+%4*2] + lea %3, [%3+%5*2] +%endif + + psadbw mm1, mm0 + psadbw mm2, mm0 + psadbw mm3, mm0 + + paddw mm5, mm1 + paddw mm6, mm2 + paddw mm7, mm3 +%endmacro + +;void int aom_sad16x16x3_sse3( +; unsigned char *src_ptr, +; int src_stride, +; unsigned char *ref_ptr, +; int ref_stride, +; int *results) +global sym(aom_sad16x16x3_sse3) PRIVATE +sym(aom_sad16x16x3_sse3): + + STACK_FRAME_CREATE_X3 + + PROCESS_16X2X3 0, src_ptr, ref_ptr, src_stride, ref_stride + PROCESS_16X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride + PROCESS_16X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride + PROCESS_16X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride + PROCESS_16X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride + PROCESS_16X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride + PROCESS_16X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride + PROCESS_16X2X3 2, src_ptr, ref_ptr, src_stride, ref_stride + + mov rcx, result_ptr + + movq xmm0, xmm5 + psrldq xmm5, 8 + + paddw xmm0, xmm5 + movd [rcx], xmm0 +;- + movq xmm0, xmm6 + psrldq xmm6, 8 + + paddw xmm0, xmm6 + movd [rcx+4], xmm0 +;- + movq xmm0, xmm7 + psrldq xmm7, 8 + + paddw xmm0, xmm7 + movd [rcx+8], xmm0 + + STACK_FRAME_DESTROY_X3 + +;void int aom_sad16x8x3_sse3( +; unsigned char *src_ptr, +; int src_stride, +; unsigned char *ref_ptr, +; int ref_stride, +; int *results) +global sym(aom_sad16x8x3_sse3) PRIVATE +sym(aom_sad16x8x3_sse3): + + STACK_FRAME_CREATE_X3 + + PROCESS_16X2X3 0, src_ptr, ref_ptr, src_stride, ref_stride + PROCESS_16X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride + PROCESS_16X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride + PROCESS_16X2X3 2, src_ptr, ref_ptr, src_stride, ref_stride + + mov rcx, result_ptr + + movq xmm0, xmm5 + psrldq xmm5, 8 + + paddw xmm0, xmm5 + movd [rcx], xmm0 +;- + movq xmm0, xmm6 + psrldq xmm6, 8 + + paddw xmm0, xmm6 + movd [rcx+4], xmm0 +;- + movq xmm0, xmm7 + psrldq xmm7, 8 + + paddw xmm0, xmm7 + movd [rcx+8], xmm0 + + STACK_FRAME_DESTROY_X3 + +;void int aom_sad8x16x3_sse3( +; unsigned char *src_ptr, +; int src_stride, +; unsigned char *ref_ptr, +; int ref_stride, +; int *results) +global sym(aom_sad8x16x3_sse3) PRIVATE +sym(aom_sad8x16x3_sse3): + + STACK_FRAME_CREATE_X3 + + PROCESS_8X2X3 0, src_ptr, ref_ptr, src_stride, ref_stride + PROCESS_8X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride + PROCESS_8X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride + PROCESS_8X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride + PROCESS_8X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride + PROCESS_8X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride + PROCESS_8X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride + PROCESS_8X2X3 2, src_ptr, ref_ptr, src_stride, ref_stride + + mov rcx, result_ptr + + punpckldq mm5, mm6 + + movq [rcx], mm5 + movd [rcx+8], mm7 + + STACK_FRAME_DESTROY_X3 + +;void int aom_sad8x8x3_sse3( +; unsigned char *src_ptr, +; int src_stride, +; unsigned char *ref_ptr, +; int ref_stride, +; int *results) +global sym(aom_sad8x8x3_sse3) PRIVATE +sym(aom_sad8x8x3_sse3): + + STACK_FRAME_CREATE_X3 + + PROCESS_8X2X3 0, src_ptr, ref_ptr, src_stride, ref_stride + PROCESS_8X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride + PROCESS_8X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride + PROCESS_8X2X3 2, src_ptr, ref_ptr, src_stride, ref_stride + + mov rcx, result_ptr + + punpckldq mm5, mm6 + + movq [rcx], mm5 + movd [rcx+8], mm7 + + STACK_FRAME_DESTROY_X3 + +;void int aom_sad4x4x3_sse3( +; unsigned char *src_ptr, +; int src_stride, +; unsigned char *ref_ptr, +; int ref_stride, +; int *results) +global sym(aom_sad4x4x3_sse3) PRIVATE +sym(aom_sad4x4x3_sse3): + + STACK_FRAME_CREATE_X3 + + movd mm0, DWORD PTR [src_ptr] + movd mm1, DWORD PTR [ref_ptr] + + movd mm2, DWORD PTR [src_ptr+src_stride] + movd mm3, DWORD PTR [ref_ptr+ref_stride] + + punpcklbw mm0, mm2 + punpcklbw mm1, mm3 + + movd mm4, DWORD PTR [ref_ptr+1] + movd mm5, DWORD PTR [ref_ptr+2] + + movd mm2, DWORD PTR [ref_ptr+ref_stride+1] + movd mm3, DWORD PTR [ref_ptr+ref_stride+2] + + psadbw mm1, mm0 + + punpcklbw mm4, mm2 + punpcklbw mm5, mm3 + + psadbw mm4, mm0 + psadbw mm5, mm0 + + lea src_ptr, [src_ptr+src_stride*2] + lea ref_ptr, [ref_ptr+ref_stride*2] + + movd mm0, DWORD PTR [src_ptr] + movd mm2, DWORD PTR [ref_ptr] + + movd mm3, DWORD PTR [src_ptr+src_stride] + movd mm6, DWORD PTR [ref_ptr+ref_stride] + + punpcklbw mm0, mm3 + punpcklbw mm2, mm6 + + movd mm3, DWORD PTR [ref_ptr+1] + movd mm7, DWORD PTR [ref_ptr+2] + + psadbw mm2, mm0 + + paddw mm1, mm2 + + movd mm2, DWORD PTR [ref_ptr+ref_stride+1] + movd mm6, DWORD PTR [ref_ptr+ref_stride+2] + + punpcklbw mm3, mm2 + punpcklbw mm7, mm6 + + psadbw mm3, mm0 + psadbw mm7, mm0 + + paddw mm3, mm4 + paddw mm7, mm5 + + mov rcx, result_ptr + + punpckldq mm1, mm3 + + movq [rcx], mm1 + movd [rcx+8], mm7 + + STACK_FRAME_DESTROY_X3 diff --git a/third_party/aom/aom_dsp/x86/sad_sse4.asm b/third_party/aom/aom_dsp/x86/sad_sse4.asm new file mode 100644 index 000000000..5e9c75845 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/sad_sse4.asm @@ -0,0 +1,362 @@ +; +; Copyright (c) 2016, Alliance for Open Media. All rights reserved +; +; This source code is subject to the terms of the BSD 2 Clause License and +; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License +; was not distributed with this source code in the LICENSE file, you can +; obtain it at www.aomedia.org/license/software. If the Alliance for Open +; Media Patent License 1.0 was not distributed with this source code in the +; PATENTS file, you can obtain it at www.aomedia.org/license/patent. +; + +; + + +%include "aom_ports/x86_abi_support.asm" + +%macro PROCESS_16X2X8 1 +%if %1 + movdqa xmm0, XMMWORD PTR [rsi] + movq xmm1, MMWORD PTR [rdi] + movq xmm3, MMWORD PTR [rdi+8] + movq xmm2, MMWORD PTR [rdi+16] + punpcklqdq xmm1, xmm3 + punpcklqdq xmm3, xmm2 + + movdqa xmm2, xmm1 + mpsadbw xmm1, xmm0, 0x0 + mpsadbw xmm2, xmm0, 0x5 + + psrldq xmm0, 8 + + movdqa xmm4, xmm3 + mpsadbw xmm3, xmm0, 0x0 + mpsadbw xmm4, xmm0, 0x5 + + paddw xmm1, xmm2 + paddw xmm1, xmm3 + paddw xmm1, xmm4 +%else + movdqa xmm0, XMMWORD PTR [rsi] + movq xmm5, MMWORD PTR [rdi] + movq xmm3, MMWORD PTR [rdi+8] + movq xmm2, MMWORD PTR [rdi+16] + punpcklqdq xmm5, xmm3 + punpcklqdq xmm3, xmm2 + + movdqa xmm2, xmm5 + mpsadbw xmm5, xmm0, 0x0 + mpsadbw xmm2, xmm0, 0x5 + + psrldq xmm0, 8 + + movdqa xmm4, xmm3 + mpsadbw xmm3, xmm0, 0x0 + mpsadbw xmm4, xmm0, 0x5 + + paddw xmm5, xmm2 + paddw xmm5, xmm3 + paddw xmm5, xmm4 + + paddw xmm1, xmm5 +%endif + movdqa xmm0, XMMWORD PTR [rsi + rax] + movq xmm5, MMWORD PTR [rdi+ rdx] + movq xmm3, MMWORD PTR [rdi+ rdx+8] + movq xmm2, MMWORD PTR [rdi+ rdx+16] + punpcklqdq xmm5, xmm3 + punpcklqdq xmm3, xmm2 + + lea rsi, [rsi+rax*2] + lea rdi, [rdi+rdx*2] + + movdqa xmm2, xmm5 + mpsadbw xmm5, xmm0, 0x0 + mpsadbw xmm2, xmm0, 0x5 + + psrldq xmm0, 8 + movdqa xmm4, xmm3 + mpsadbw xmm3, xmm0, 0x0 + mpsadbw xmm4, xmm0, 0x5 + + paddw xmm5, xmm2 + paddw xmm5, xmm3 + paddw xmm5, xmm4 + + paddw xmm1, xmm5 +%endmacro + +%macro PROCESS_8X2X8 1 +%if %1 + movq xmm0, MMWORD PTR [rsi] + movq xmm1, MMWORD PTR [rdi] + movq xmm3, MMWORD PTR [rdi+8] + punpcklqdq xmm1, xmm3 + + movdqa xmm2, xmm1 + mpsadbw xmm1, xmm0, 0x0 + mpsadbw xmm2, xmm0, 0x5 + paddw xmm1, xmm2 +%else + movq xmm0, MMWORD PTR [rsi] + movq xmm5, MMWORD PTR [rdi] + movq xmm3, MMWORD PTR [rdi+8] + punpcklqdq xmm5, xmm3 + + movdqa xmm2, xmm5 + mpsadbw xmm5, xmm0, 0x0 + mpsadbw xmm2, xmm0, 0x5 + paddw xmm5, xmm2 + + paddw xmm1, xmm5 +%endif + movq xmm0, MMWORD PTR [rsi + rax] + movq xmm5, MMWORD PTR [rdi+ rdx] + movq xmm3, MMWORD PTR [rdi+ rdx+8] + punpcklqdq xmm5, xmm3 + + lea rsi, [rsi+rax*2] + lea rdi, [rdi+rdx*2] + + movdqa xmm2, xmm5 + mpsadbw xmm5, xmm0, 0x0 + mpsadbw xmm2, xmm0, 0x5 + paddw xmm5, xmm2 + + paddw xmm1, xmm5 +%endmacro + +%macro PROCESS_4X2X8 1 +%if %1 + movd xmm0, [rsi] + movq xmm1, MMWORD PTR [rdi] + movq xmm3, MMWORD PTR [rdi+8] + punpcklqdq xmm1, xmm3 + + mpsadbw xmm1, xmm0, 0x0 +%else + movd xmm0, [rsi] + movq xmm5, MMWORD PTR [rdi] + movq xmm3, MMWORD PTR [rdi+8] + punpcklqdq xmm5, xmm3 + + mpsadbw xmm5, xmm0, 0x0 + + paddw xmm1, xmm5 +%endif + movd xmm0, [rsi + rax] + movq xmm5, MMWORD PTR [rdi+ rdx] + movq xmm3, MMWORD PTR [rdi+ rdx+8] + punpcklqdq xmm5, xmm3 + + lea rsi, [rsi+rax*2] + lea rdi, [rdi+rdx*2] + + mpsadbw xmm5, xmm0, 0x0 + + paddw xmm1, xmm5 +%endmacro + +%macro WRITE_AS_INTS 0 + mov rdi, arg(4) ;Results + pxor xmm0, xmm0 + movdqa xmm2, xmm1 + punpcklwd xmm1, xmm0 + punpckhwd xmm2, xmm0 + + movdqa [rdi], xmm1 + movdqa [rdi + 16], xmm2 +%endmacro + +;void aom_sad16x16x8_sse4_1( +; const unsigned char *src_ptr, +; int src_stride, +; const unsigned char *ref_ptr, +; int ref_stride, +; unsigned short *sad_array); +global sym(aom_sad16x16x8_sse4_1) PRIVATE +sym(aom_sad16x16x8_sse4_1): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 5 + push rsi + push rdi + ; end prolog + + mov rsi, arg(0) ;src_ptr + mov rdi, arg(2) ;ref_ptr + + movsxd rax, dword ptr arg(1) ;src_stride + movsxd rdx, dword ptr arg(3) ;ref_stride + + PROCESS_16X2X8 1 + PROCESS_16X2X8 0 + PROCESS_16X2X8 0 + PROCESS_16X2X8 0 + PROCESS_16X2X8 0 + PROCESS_16X2X8 0 + PROCESS_16X2X8 0 + PROCESS_16X2X8 0 + + WRITE_AS_INTS + + ; begin epilog + pop rdi + pop rsi + UNSHADOW_ARGS + pop rbp + ret + + +;void aom_sad16x8x8_sse4_1( +; const unsigned char *src_ptr, +; int src_stride, +; const unsigned char *ref_ptr, +; int ref_stride, +; unsigned short *sad_array +;); +global sym(aom_sad16x8x8_sse4_1) PRIVATE +sym(aom_sad16x8x8_sse4_1): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 5 + push rsi + push rdi + ; end prolog + + mov rsi, arg(0) ;src_ptr + mov rdi, arg(2) ;ref_ptr + + movsxd rax, dword ptr arg(1) ;src_stride + movsxd rdx, dword ptr arg(3) ;ref_stride + + PROCESS_16X2X8 1 + PROCESS_16X2X8 0 + PROCESS_16X2X8 0 + PROCESS_16X2X8 0 + + WRITE_AS_INTS + + ; begin epilog + pop rdi + pop rsi + UNSHADOW_ARGS + pop rbp + ret + + +;void aom_sad8x8x8_sse4_1( +; const unsigned char *src_ptr, +; int src_stride, +; const unsigned char *ref_ptr, +; int ref_stride, +; unsigned short *sad_array +;); +global sym(aom_sad8x8x8_sse4_1) PRIVATE +sym(aom_sad8x8x8_sse4_1): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 5 + push rsi + push rdi + ; end prolog + + mov rsi, arg(0) ;src_ptr + mov rdi, arg(2) ;ref_ptr + + movsxd rax, dword ptr arg(1) ;src_stride + movsxd rdx, dword ptr arg(3) ;ref_stride + + PROCESS_8X2X8 1 + PROCESS_8X2X8 0 + PROCESS_8X2X8 0 + PROCESS_8X2X8 0 + + WRITE_AS_INTS + + ; begin epilog + pop rdi + pop rsi + UNSHADOW_ARGS + pop rbp + ret + + +;void aom_sad8x16x8_sse4_1( +; const unsigned char *src_ptr, +; int src_stride, +; const unsigned char *ref_ptr, +; int ref_stride, +; unsigned short *sad_array +;); +global sym(aom_sad8x16x8_sse4_1) PRIVATE +sym(aom_sad8x16x8_sse4_1): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 5 + push rsi + push rdi + ; end prolog + + mov rsi, arg(0) ;src_ptr + mov rdi, arg(2) ;ref_ptr + + movsxd rax, dword ptr arg(1) ;src_stride + movsxd rdx, dword ptr arg(3) ;ref_stride + + PROCESS_8X2X8 1 + PROCESS_8X2X8 0 + PROCESS_8X2X8 0 + PROCESS_8X2X8 0 + PROCESS_8X2X8 0 + PROCESS_8X2X8 0 + PROCESS_8X2X8 0 + PROCESS_8X2X8 0 + + WRITE_AS_INTS + + ; begin epilog + pop rdi + pop rsi + UNSHADOW_ARGS + pop rbp + ret + + +;void aom_sad4x4x8_sse4_1( +; const unsigned char *src_ptr, +; int src_stride, +; const unsigned char *ref_ptr, +; int ref_stride, +; unsigned short *sad_array +;); +global sym(aom_sad4x4x8_sse4_1) PRIVATE +sym(aom_sad4x4x8_sse4_1): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 5 + push rsi + push rdi + ; end prolog + + mov rsi, arg(0) ;src_ptr + mov rdi, arg(2) ;ref_ptr + + movsxd rax, dword ptr arg(1) ;src_stride + movsxd rdx, dword ptr arg(3) ;ref_stride + + PROCESS_4X2X8 1 + PROCESS_4X2X8 0 + + WRITE_AS_INTS + + ; begin epilog + pop rdi + pop rsi + UNSHADOW_ARGS + pop rbp + ret + + + + diff --git a/third_party/aom/aom_dsp/x86/sad_ssse3.asm b/third_party/aom/aom_dsp/x86/sad_ssse3.asm new file mode 100644 index 000000000..96b64b040 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/sad_ssse3.asm @@ -0,0 +1,373 @@ +; +; Copyright (c) 2016, Alliance for Open Media. All rights reserved +; +; This source code is subject to the terms of the BSD 2 Clause License and +; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License +; was not distributed with this source code in the LICENSE file, you can +; obtain it at www.aomedia.org/license/software. If the Alliance for Open +; Media Patent License 1.0 was not distributed with this source code in the +; PATENTS file, you can obtain it at www.aomedia.org/license/patent. +; + +; + + +%include "aom_ports/x86_abi_support.asm" + +%macro PROCESS_16X2X3 1 +%if %1 + movdqa xmm0, XMMWORD PTR [rsi] + lddqu xmm5, XMMWORD PTR [rdi] + lddqu xmm6, XMMWORD PTR [rdi+1] + lddqu xmm7, XMMWORD PTR [rdi+2] + + psadbw xmm5, xmm0 + psadbw xmm6, xmm0 + psadbw xmm7, xmm0 +%else + movdqa xmm0, XMMWORD PTR [rsi] + lddqu xmm1, XMMWORD PTR [rdi] + lddqu xmm2, XMMWORD PTR [rdi+1] + lddqu xmm3, XMMWORD PTR [rdi+2] + + psadbw xmm1, xmm0 + psadbw xmm2, xmm0 + psadbw xmm3, xmm0 + + paddw xmm5, xmm1 + paddw xmm6, xmm2 + paddw xmm7, xmm3 +%endif + movdqa xmm0, XMMWORD PTR [rsi+rax] + lddqu xmm1, XMMWORD PTR [rdi+rdx] + lddqu xmm2, XMMWORD PTR [rdi+rdx+1] + lddqu xmm3, XMMWORD PTR [rdi+rdx+2] + + lea rsi, [rsi+rax*2] + lea rdi, [rdi+rdx*2] + + psadbw xmm1, xmm0 + psadbw xmm2, xmm0 + psadbw xmm3, xmm0 + + paddw xmm5, xmm1 + paddw xmm6, xmm2 + paddw xmm7, xmm3 +%endmacro + +%macro PROCESS_16X2X3_OFFSET 2 +%if %1 + movdqa xmm0, XMMWORD PTR [rsi] + movdqa xmm4, XMMWORD PTR [rdi] + movdqa xmm7, XMMWORD PTR [rdi+16] + + movdqa xmm5, xmm7 + palignr xmm5, xmm4, %2 + + movdqa xmm6, xmm7 + palignr xmm6, xmm4, (%2+1) + + palignr xmm7, xmm4, (%2+2) + + psadbw xmm5, xmm0 + psadbw xmm6, xmm0 + psadbw xmm7, xmm0 +%else + movdqa xmm0, XMMWORD PTR [rsi] + movdqa xmm4, XMMWORD PTR [rdi] + movdqa xmm3, XMMWORD PTR [rdi+16] + + movdqa xmm1, xmm3 + palignr xmm1, xmm4, %2 + + movdqa xmm2, xmm3 + palignr xmm2, xmm4, (%2+1) + + palignr xmm3, xmm4, (%2+2) + + psadbw xmm1, xmm0 + psadbw xmm2, xmm0 + psadbw xmm3, xmm0 + + paddw xmm5, xmm1 + paddw xmm6, xmm2 + paddw xmm7, xmm3 +%endif + movdqa xmm0, XMMWORD PTR [rsi+rax] + movdqa xmm4, XMMWORD PTR [rdi+rdx] + movdqa xmm3, XMMWORD PTR [rdi+rdx+16] + + movdqa xmm1, xmm3 + palignr xmm1, xmm4, %2 + + movdqa xmm2, xmm3 + palignr xmm2, xmm4, (%2+1) + + palignr xmm3, xmm4, (%2+2) + + lea rsi, [rsi+rax*2] + lea rdi, [rdi+rdx*2] + + psadbw xmm1, xmm0 + psadbw xmm2, xmm0 + psadbw xmm3, xmm0 + + paddw xmm5, xmm1 + paddw xmm6, xmm2 + paddw xmm7, xmm3 +%endmacro + +%macro PROCESS_16X16X3_OFFSET 2 +%2_aligned_by_%1: + + sub rdi, %1 + + PROCESS_16X2X3_OFFSET 1, %1 + PROCESS_16X2X3_OFFSET 0, %1 + PROCESS_16X2X3_OFFSET 0, %1 + PROCESS_16X2X3_OFFSET 0, %1 + PROCESS_16X2X3_OFFSET 0, %1 + PROCESS_16X2X3_OFFSET 0, %1 + PROCESS_16X2X3_OFFSET 0, %1 + PROCESS_16X2X3_OFFSET 0, %1 + + jmp %2_store_off + +%endmacro + +%macro PROCESS_16X8X3_OFFSET 2 +%2_aligned_by_%1: + + sub rdi, %1 + + PROCESS_16X2X3_OFFSET 1, %1 + PROCESS_16X2X3_OFFSET 0, %1 + PROCESS_16X2X3_OFFSET 0, %1 + PROCESS_16X2X3_OFFSET 0, %1 + + jmp %2_store_off + +%endmacro + +;void int aom_sad16x16x3_ssse3( +; unsigned char *src_ptr, +; int src_stride, +; unsigned char *ref_ptr, +; int ref_stride, +; int *results) +global sym(aom_sad16x16x3_ssse3) PRIVATE +sym(aom_sad16x16x3_ssse3): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 5 + SAVE_XMM 7 + push rsi + push rdi + push rcx + ; end prolog + + mov rsi, arg(0) ;src_ptr + mov rdi, arg(2) ;ref_ptr + + mov rdx, 0xf + and rdx, rdi + + jmp .aom_sad16x16x3_ssse3_skiptable +.aom_sad16x16x3_ssse3_jumptable: + dd .aom_sad16x16x3_ssse3_aligned_by_0 - .aom_sad16x16x3_ssse3_do_jump + dd .aom_sad16x16x3_ssse3_aligned_by_1 - .aom_sad16x16x3_ssse3_do_jump + dd .aom_sad16x16x3_ssse3_aligned_by_2 - .aom_sad16x16x3_ssse3_do_jump + dd .aom_sad16x16x3_ssse3_aligned_by_3 - .aom_sad16x16x3_ssse3_do_jump + dd .aom_sad16x16x3_ssse3_aligned_by_4 - .aom_sad16x16x3_ssse3_do_jump + dd .aom_sad16x16x3_ssse3_aligned_by_5 - .aom_sad16x16x3_ssse3_do_jump + dd .aom_sad16x16x3_ssse3_aligned_by_6 - .aom_sad16x16x3_ssse3_do_jump + dd .aom_sad16x16x3_ssse3_aligned_by_7 - .aom_sad16x16x3_ssse3_do_jump + dd .aom_sad16x16x3_ssse3_aligned_by_8 - .aom_sad16x16x3_ssse3_do_jump + dd .aom_sad16x16x3_ssse3_aligned_by_9 - .aom_sad16x16x3_ssse3_do_jump + dd .aom_sad16x16x3_ssse3_aligned_by_10 - .aom_sad16x16x3_ssse3_do_jump + dd .aom_sad16x16x3_ssse3_aligned_by_11 - .aom_sad16x16x3_ssse3_do_jump + dd .aom_sad16x16x3_ssse3_aligned_by_12 - .aom_sad16x16x3_ssse3_do_jump + dd .aom_sad16x16x3_ssse3_aligned_by_13 - .aom_sad16x16x3_ssse3_do_jump + dd .aom_sad16x16x3_ssse3_aligned_by_14 - .aom_sad16x16x3_ssse3_do_jump + dd .aom_sad16x16x3_ssse3_aligned_by_15 - .aom_sad16x16x3_ssse3_do_jump +.aom_sad16x16x3_ssse3_skiptable: + + call .aom_sad16x16x3_ssse3_do_jump +.aom_sad16x16x3_ssse3_do_jump: + pop rcx ; get the address of do_jump + mov rax, .aom_sad16x16x3_ssse3_jumptable - .aom_sad16x16x3_ssse3_do_jump + add rax, rcx ; get the absolute address of aom_sad16x16x3_ssse3_jumptable + + movsxd rax, dword [rax + 4*rdx] ; get the 32 bit offset from the jumptable + add rcx, rax + + movsxd rax, dword ptr arg(1) ;src_stride + movsxd rdx, dword ptr arg(3) ;ref_stride + + jmp rcx + + PROCESS_16X16X3_OFFSET 0, .aom_sad16x16x3_ssse3 + PROCESS_16X16X3_OFFSET 1, .aom_sad16x16x3_ssse3 + PROCESS_16X16X3_OFFSET 2, .aom_sad16x16x3_ssse3 + PROCESS_16X16X3_OFFSET 3, .aom_sad16x16x3_ssse3 + PROCESS_16X16X3_OFFSET 4, .aom_sad16x16x3_ssse3 + PROCESS_16X16X3_OFFSET 5, .aom_sad16x16x3_ssse3 + PROCESS_16X16X3_OFFSET 6, .aom_sad16x16x3_ssse3 + PROCESS_16X16X3_OFFSET 7, .aom_sad16x16x3_ssse3 + PROCESS_16X16X3_OFFSET 8, .aom_sad16x16x3_ssse3 + PROCESS_16X16X3_OFFSET 9, .aom_sad16x16x3_ssse3 + PROCESS_16X16X3_OFFSET 10, .aom_sad16x16x3_ssse3 + PROCESS_16X16X3_OFFSET 11, .aom_sad16x16x3_ssse3 + PROCESS_16X16X3_OFFSET 12, .aom_sad16x16x3_ssse3 + PROCESS_16X16X3_OFFSET 13, .aom_sad16x16x3_ssse3 + PROCESS_16X16X3_OFFSET 14, .aom_sad16x16x3_ssse3 + +.aom_sad16x16x3_ssse3_aligned_by_15: + PROCESS_16X2X3 1 + PROCESS_16X2X3 0 + PROCESS_16X2X3 0 + PROCESS_16X2X3 0 + PROCESS_16X2X3 0 + PROCESS_16X2X3 0 + PROCESS_16X2X3 0 + PROCESS_16X2X3 0 + +.aom_sad16x16x3_ssse3_store_off: + mov rdi, arg(4) ;Results + + movq xmm0, xmm5 + psrldq xmm5, 8 + + paddw xmm0, xmm5 + movd [rdi], xmm0 +;- + movq xmm0, xmm6 + psrldq xmm6, 8 + + paddw xmm0, xmm6 + movd [rdi+4], xmm0 +;- + movq xmm0, xmm7 + psrldq xmm7, 8 + + paddw xmm0, xmm7 + movd [rdi+8], xmm0 + + ; begin epilog + pop rcx + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret + +;void int aom_sad16x8x3_ssse3( +; unsigned char *src_ptr, +; int src_stride, +; unsigned char *ref_ptr, +; int ref_stride, +; int *results) +global sym(aom_sad16x8x3_ssse3) PRIVATE +sym(aom_sad16x8x3_ssse3): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 5 + SAVE_XMM 7 + push rsi + push rdi + push rcx + ; end prolog + + mov rsi, arg(0) ;src_ptr + mov rdi, arg(2) ;ref_ptr + + mov rdx, 0xf + and rdx, rdi + + jmp .aom_sad16x8x3_ssse3_skiptable +.aom_sad16x8x3_ssse3_jumptable: + dd .aom_sad16x8x3_ssse3_aligned_by_0 - .aom_sad16x8x3_ssse3_do_jump + dd .aom_sad16x8x3_ssse3_aligned_by_1 - .aom_sad16x8x3_ssse3_do_jump + dd .aom_sad16x8x3_ssse3_aligned_by_2 - .aom_sad16x8x3_ssse3_do_jump + dd .aom_sad16x8x3_ssse3_aligned_by_3 - .aom_sad16x8x3_ssse3_do_jump + dd .aom_sad16x8x3_ssse3_aligned_by_4 - .aom_sad16x8x3_ssse3_do_jump + dd .aom_sad16x8x3_ssse3_aligned_by_5 - .aom_sad16x8x3_ssse3_do_jump + dd .aom_sad16x8x3_ssse3_aligned_by_6 - .aom_sad16x8x3_ssse3_do_jump + dd .aom_sad16x8x3_ssse3_aligned_by_7 - .aom_sad16x8x3_ssse3_do_jump + dd .aom_sad16x8x3_ssse3_aligned_by_8 - .aom_sad16x8x3_ssse3_do_jump + dd .aom_sad16x8x3_ssse3_aligned_by_9 - .aom_sad16x8x3_ssse3_do_jump + dd .aom_sad16x8x3_ssse3_aligned_by_10 - .aom_sad16x8x3_ssse3_do_jump + dd .aom_sad16x8x3_ssse3_aligned_by_11 - .aom_sad16x8x3_ssse3_do_jump + dd .aom_sad16x8x3_ssse3_aligned_by_12 - .aom_sad16x8x3_ssse3_do_jump + dd .aom_sad16x8x3_ssse3_aligned_by_13 - .aom_sad16x8x3_ssse3_do_jump + dd .aom_sad16x8x3_ssse3_aligned_by_14 - .aom_sad16x8x3_ssse3_do_jump + dd .aom_sad16x8x3_ssse3_aligned_by_15 - .aom_sad16x8x3_ssse3_do_jump +.aom_sad16x8x3_ssse3_skiptable: + + call .aom_sad16x8x3_ssse3_do_jump +.aom_sad16x8x3_ssse3_do_jump: + pop rcx ; get the address of do_jump + mov rax, .aom_sad16x8x3_ssse3_jumptable - .aom_sad16x8x3_ssse3_do_jump + add rax, rcx ; get the absolute address of aom_sad16x8x3_ssse3_jumptable + + movsxd rax, dword [rax + 4*rdx] ; get the 32 bit offset from the jumptable + add rcx, rax + + movsxd rax, dword ptr arg(1) ;src_stride + movsxd rdx, dword ptr arg(3) ;ref_stride + + jmp rcx + + PROCESS_16X8X3_OFFSET 0, .aom_sad16x8x3_ssse3 + PROCESS_16X8X3_OFFSET 1, .aom_sad16x8x3_ssse3 + PROCESS_16X8X3_OFFSET 2, .aom_sad16x8x3_ssse3 + PROCESS_16X8X3_OFFSET 3, .aom_sad16x8x3_ssse3 + PROCESS_16X8X3_OFFSET 4, .aom_sad16x8x3_ssse3 + PROCESS_16X8X3_OFFSET 5, .aom_sad16x8x3_ssse3 + PROCESS_16X8X3_OFFSET 6, .aom_sad16x8x3_ssse3 + PROCESS_16X8X3_OFFSET 7, .aom_sad16x8x3_ssse3 + PROCESS_16X8X3_OFFSET 8, .aom_sad16x8x3_ssse3 + PROCESS_16X8X3_OFFSET 9, .aom_sad16x8x3_ssse3 + PROCESS_16X8X3_OFFSET 10, .aom_sad16x8x3_ssse3 + PROCESS_16X8X3_OFFSET 11, .aom_sad16x8x3_ssse3 + PROCESS_16X8X3_OFFSET 12, .aom_sad16x8x3_ssse3 + PROCESS_16X8X3_OFFSET 13, .aom_sad16x8x3_ssse3 + PROCESS_16X8X3_OFFSET 14, .aom_sad16x8x3_ssse3 + +.aom_sad16x8x3_ssse3_aligned_by_15: + + PROCESS_16X2X3 1 + PROCESS_16X2X3 0 + PROCESS_16X2X3 0 + PROCESS_16X2X3 0 + +.aom_sad16x8x3_ssse3_store_off: + mov rdi, arg(4) ;Results + + movq xmm0, xmm5 + psrldq xmm5, 8 + + paddw xmm0, xmm5 + movd [rdi], xmm0 +;- + movq xmm0, xmm6 + psrldq xmm6, 8 + + paddw xmm0, xmm6 + movd [rdi+4], xmm0 +;- + movq xmm0, xmm7 + psrldq xmm7, 8 + + paddw xmm0, xmm7 + movd [rdi+8], xmm0 + + ; begin epilog + pop rcx + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret diff --git a/third_party/aom/aom_dsp/x86/ssim_opt_x86_64.asm b/third_party/aom/aom_dsp/x86/ssim_opt_x86_64.asm new file mode 100644 index 000000000..aa70106c8 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/ssim_opt_x86_64.asm @@ -0,0 +1,219 @@ +; +; Copyright (c) 2016, Alliance for Open Media. All rights reserved +; +; This source code is subject to the terms of the BSD 2 Clause License and +; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License +; was not distributed with this source code in the LICENSE file, you can +; obtain it at www.aomedia.org/license/software. If the Alliance for Open +; Media Patent License 1.0 was not distributed with this source code in the +; PATENTS file, you can obtain it at www.aomedia.org/license/patent. +; + +; + +%include "aom_ports/x86_abi_support.asm" + +; tabulate_ssim - sums sum_s,sum_r,sum_sq_s,sum_sq_r, sum_sxr +%macro TABULATE_SSIM 0 + paddusw xmm15, xmm3 ; sum_s + paddusw xmm14, xmm4 ; sum_r + movdqa xmm1, xmm3 + pmaddwd xmm1, xmm1 + paddd xmm13, xmm1 ; sum_sq_s + movdqa xmm2, xmm4 + pmaddwd xmm2, xmm2 + paddd xmm12, xmm2 ; sum_sq_r + pmaddwd xmm3, xmm4 + paddd xmm11, xmm3 ; sum_sxr +%endmacro + +; Sum across the register %1 starting with q words +%macro SUM_ACROSS_Q 1 + movdqa xmm2,%1 + punpckldq %1,xmm0 + punpckhdq xmm2,xmm0 + paddq %1,xmm2 + movdqa xmm2,%1 + punpcklqdq %1,xmm0 + punpckhqdq xmm2,xmm0 + paddq %1,xmm2 +%endmacro + +; Sum across the register %1 starting with q words +%macro SUM_ACROSS_W 1 + movdqa xmm1, %1 + punpcklwd %1,xmm0 + punpckhwd xmm1,xmm0 + paddd %1, xmm1 + SUM_ACROSS_Q %1 +%endmacro +;void ssim_parms_sse2( +; unsigned char *s, +; int sp, +; unsigned char *r, +; int rp +; uint32_t *sum_s, +; uint32_t *sum_r, +; uint32_t *sum_sq_s, +; uint32_t *sum_sq_r, +; uint32_t *sum_sxr); +; +; TODO: Use parm passing through structure, probably don't need the pxors +; ( calling app will initialize to 0 ) could easily fit everything in sse2 +; without too much hastle, and can probably do better estimates with psadw +; or pavgb At this point this is just meant to be first pass for calculating +; all the parms needed for 16x16 ssim so we can play with dssim as distortion +; in mode selection code. +global sym(aom_ssim_parms_16x16_sse2) PRIVATE +sym(aom_ssim_parms_16x16_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 9 + SAVE_XMM 15 + push rsi + push rdi + ; end prolog + + mov rsi, arg(0) ;s + mov rcx, arg(1) ;sp + mov rdi, arg(2) ;r + mov rax, arg(3) ;rp + + pxor xmm0, xmm0 + pxor xmm15,xmm15 ;sum_s + pxor xmm14,xmm14 ;sum_r + pxor xmm13,xmm13 ;sum_sq_s + pxor xmm12,xmm12 ;sum_sq_r + pxor xmm11,xmm11 ;sum_sxr + + mov rdx, 16 ;row counter +.NextRow: + + ;grab source and reference pixels + movdqu xmm5, [rsi] + movdqu xmm6, [rdi] + movdqa xmm3, xmm5 + movdqa xmm4, xmm6 + punpckhbw xmm3, xmm0 ; high_s + punpckhbw xmm4, xmm0 ; high_r + + TABULATE_SSIM + + movdqa xmm3, xmm5 + movdqa xmm4, xmm6 + punpcklbw xmm3, xmm0 ; low_s + punpcklbw xmm4, xmm0 ; low_r + + TABULATE_SSIM + + add rsi, rcx ; next s row + add rdi, rax ; next r row + + dec rdx ; counter + jnz .NextRow + + SUM_ACROSS_W xmm15 + SUM_ACROSS_W xmm14 + SUM_ACROSS_Q xmm13 + SUM_ACROSS_Q xmm12 + SUM_ACROSS_Q xmm11 + + mov rdi,arg(4) + movd [rdi], xmm15; + mov rdi,arg(5) + movd [rdi], xmm14; + mov rdi,arg(6) + movd [rdi], xmm13; + mov rdi,arg(7) + movd [rdi], xmm12; + mov rdi,arg(8) + movd [rdi], xmm11; + + ; begin epilog + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret + +;void ssim_parms_sse2( +; unsigned char *s, +; int sp, +; unsigned char *r, +; int rp +; uint32_t *sum_s, +; uint32_t *sum_r, +; uint32_t *sum_sq_s, +; uint32_t *sum_sq_r, +; uint32_t *sum_sxr); +; +; TODO: Use parm passing through structure, probably don't need the pxors +; ( calling app will initialize to 0 ) could easily fit everything in sse2 +; without too much hastle, and can probably do better estimates with psadw +; or pavgb At this point this is just meant to be first pass for calculating +; all the parms needed for 16x16 ssim so we can play with dssim as distortion +; in mode selection code. +global sym(aom_ssim_parms_8x8_sse2) PRIVATE +sym(aom_ssim_parms_8x8_sse2): + push rbp + mov rbp, rsp + SHADOW_ARGS_TO_STACK 9 + SAVE_XMM 15 + push rsi + push rdi + ; end prolog + + mov rsi, arg(0) ;s + mov rcx, arg(1) ;sp + mov rdi, arg(2) ;r + mov rax, arg(3) ;rp + + pxor xmm0, xmm0 + pxor xmm15,xmm15 ;sum_s + pxor xmm14,xmm14 ;sum_r + pxor xmm13,xmm13 ;sum_sq_s + pxor xmm12,xmm12 ;sum_sq_r + pxor xmm11,xmm11 ;sum_sxr + + mov rdx, 8 ;row counter +.NextRow: + + ;grab source and reference pixels + movq xmm3, [rsi] + movq xmm4, [rdi] + punpcklbw xmm3, xmm0 ; low_s + punpcklbw xmm4, xmm0 ; low_r + + TABULATE_SSIM + + add rsi, rcx ; next s row + add rdi, rax ; next r row + + dec rdx ; counter + jnz .NextRow + + SUM_ACROSS_W xmm15 + SUM_ACROSS_W xmm14 + SUM_ACROSS_Q xmm13 + SUM_ACROSS_Q xmm12 + SUM_ACROSS_Q xmm11 + + mov rdi,arg(4) + movd [rdi], xmm15; + mov rdi,arg(5) + movd [rdi], xmm14; + mov rdi,arg(6) + movd [rdi], xmm13; + mov rdi,arg(7) + movd [rdi], xmm12; + mov rdi,arg(8) + movd [rdi], xmm11; + + ; begin epilog + pop rdi + pop rsi + RESTORE_XMM + UNSHADOW_ARGS + pop rbp + ret diff --git a/third_party/aom/aom_dsp/x86/subpel_variance_sse2.asm b/third_party/aom/aom_dsp/x86/subpel_variance_sse2.asm new file mode 100644 index 000000000..d3feb7ec0 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/subpel_variance_sse2.asm @@ -0,0 +1,1489 @@ +; +; Copyright (c) 2016, Alliance for Open Media. All rights reserved +; +; This source code is subject to the terms of the BSD 2 Clause License and +; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License +; was not distributed with this source code in the LICENSE file, you can +; obtain it at www.aomedia.org/license/software. If the Alliance for Open +; Media Patent License 1.0 was not distributed with this source code in the +; PATENTS file, you can obtain it at www.aomedia.org/license/patent. +; + +; + +%include "third_party/x86inc/x86inc.asm" + +SECTION_RODATA +pw_8: times 8 dw 8 +bilin_filter_m_sse2: times 8 dw 16 + times 8 dw 0 + times 8 dw 14 + times 8 dw 2 + times 8 dw 12 + times 8 dw 4 + times 8 dw 10 + times 8 dw 6 + times 16 dw 8 + times 8 dw 6 + times 8 dw 10 + times 8 dw 4 + times 8 dw 12 + times 8 dw 2 + times 8 dw 14 + +bilin_filter_m_ssse3: times 8 db 16, 0 + times 8 db 14, 2 + times 8 db 12, 4 + times 8 db 10, 6 + times 16 db 8 + times 8 db 6, 10 + times 8 db 4, 12 + times 8 db 2, 14 + +SECTION .text + +; int aom_sub_pixel_varianceNxh(const uint8_t *src, ptrdiff_t src_stride, +; int x_offset, int y_offset, +; const uint8_t *dst, ptrdiff_t dst_stride, +; int height, unsigned int *sse); +; +; This function returns the SE and stores SSE in the given pointer. + +%macro SUM_SSE 6 ; src1, dst1, src2, dst2, sum, sse + psubw %3, %4 + psubw %1, %2 + paddw %5, %3 + pmaddwd %3, %3 + paddw %5, %1 + pmaddwd %1, %1 + paddd %6, %3 + paddd %6, %1 +%endmacro + +%macro STORE_AND_RET 1 +%if %1 > 4 + ; if H=64 and W=16, we have 8 words of each 2(1bit)x64(6bit)x9bit=16bit + ; in m6, i.e. it _exactly_ fits in a signed word per word in the xmm reg. + ; We have to sign-extend it before adding the words within the register + ; and outputing to a dword. + pcmpgtw m5, m6 ; mask for 0 > x + movhlps m3, m7 + punpcklwd m4, m6, m5 + punpckhwd m6, m5 ; sign-extend m6 word->dword + paddd m7, m3 + paddd m6, m4 + pshufd m3, m7, 0x1 + movhlps m4, m6 + paddd m7, m3 + paddd m6, m4 + mov r1, ssem ; r1 = unsigned int *sse + pshufd m4, m6, 0x1 + movd [r1], m7 ; store sse + paddd m6, m4 + movd raxd, m6 ; store sum as return value +%else ; 4xh + pshuflw m4, m6, 0xe + pshuflw m3, m7, 0xe + paddw m6, m4 + paddd m7, m3 + pcmpgtw m5, m6 ; mask for 0 > x + mov r1, ssem ; r1 = unsigned int *sse + punpcklwd m6, m5 ; sign-extend m6 word->dword + movd [r1], m7 ; store sse + pshuflw m4, m6, 0xe + paddd m6, m4 + movd raxd, m6 ; store sum as return value +%endif + RET +%endmacro + +%macro INC_SRC_BY_SRC_STRIDE 0 +%if ARCH_X86=1 && CONFIG_PIC=1 + add srcq, src_stridemp +%else + add srcq, src_strideq +%endif +%endmacro + +%macro SUBPEL_VARIANCE 1-2 0 ; W +%if cpuflag(ssse3) +%define bilin_filter_m bilin_filter_m_ssse3 +%define filter_idx_shift 4 +%else +%define bilin_filter_m bilin_filter_m_sse2 +%define filter_idx_shift 5 +%endif +; FIXME(rbultje) only bilinear filters use >8 registers, and ssse3 only uses +; 11, not 13, if the registers are ordered correctly. May make a minor speed +; difference on Win64 + +%ifdef PIC ; 64bit PIC + %if %2 == 1 ; avg + cglobal sub_pixel_avg_variance%1xh, 9, 10, 13, src, src_stride, \ + x_offset, y_offset, \ + dst, dst_stride, \ + sec, sec_stride, height, sse + %define sec_str sec_strideq + %else + cglobal sub_pixel_variance%1xh, 7, 8, 13, src, src_stride, x_offset, \ + y_offset, dst, dst_stride, height, sse + %endif + %define block_height heightd + %define bilin_filter sseq +%else + %if ARCH_X86=1 && CONFIG_PIC=1 + %if %2 == 1 ; avg + cglobal sub_pixel_avg_variance%1xh, 7, 7, 13, src, src_stride, \ + x_offset, y_offset, \ + dst, dst_stride, \ + sec, sec_stride, \ + height, sse, g_bilin_filter, g_pw_8 + %define block_height dword heightm + %define sec_str sec_stridemp + + ;Store bilin_filter and pw_8 location in stack + %if GET_GOT_DEFINED == 1 + GET_GOT eax + add esp, 4 ; restore esp + %endif + + lea ecx, [GLOBAL(bilin_filter_m)] + mov g_bilin_filterm, ecx + + lea ecx, [GLOBAL(pw_8)] + mov g_pw_8m, ecx + + LOAD_IF_USED 0, 1 ; load eax, ecx back + %else + cglobal sub_pixel_variance%1xh, 7, 7, 13, src, src_stride, x_offset, \ + y_offset, dst, dst_stride, height, sse, \ + g_bilin_filter, g_pw_8 + %define block_height heightd + + ;Store bilin_filter and pw_8 location in stack + %if GET_GOT_DEFINED == 1 + GET_GOT eax + add esp, 4 ; restore esp + %endif + + lea ecx, [GLOBAL(bilin_filter_m)] + mov g_bilin_filterm, ecx + + lea ecx, [GLOBAL(pw_8)] + mov g_pw_8m, ecx + + LOAD_IF_USED 0, 1 ; load eax, ecx back + %endif + %else + %if %2 == 1 ; avg + cglobal sub_pixel_avg_variance%1xh, 7 + 2 * ARCH_X86_64, \ + 7 + 2 * ARCH_X86_64, 13, src, src_stride, \ + x_offset, y_offset, \ + dst, dst_stride, \ + sec, sec_stride, \ + height, sse + %if ARCH_X86_64 + %define block_height heightd + %define sec_str sec_strideq + %else + %define block_height dword heightm + %define sec_str sec_stridemp + %endif + %else + cglobal sub_pixel_variance%1xh, 7, 7, 13, src, src_stride, x_offset, \ + y_offset, dst, dst_stride, height, sse + %define block_height heightd + %endif + + %define bilin_filter bilin_filter_m + %endif +%endif + +%if %1 == 4 + %define movx movd +%else + %define movx movh +%endif + + ASSERT %1 <= 16 ; m6 overflows if w > 16 + pxor m6, m6 ; sum + pxor m7, m7 ; sse + ; FIXME(rbultje) if both filters are bilinear, we don't actually use m5; we + ; could perhaps use it for something more productive then + pxor m5, m5 ; dedicated zero register +%if %1 < 16 + sar block_height, 1 +%if %2 == 1 ; avg + shl sec_str, 1 +%endif +%endif + + ; FIXME(rbultje) replace by jumptable? + test x_offsetd, x_offsetd + jnz .x_nonzero + ; x_offset == 0 + test y_offsetd, y_offsetd + jnz .x_zero_y_nonzero + + ; x_offset == 0 && y_offset == 0 +.x_zero_y_zero_loop: +%if %1 == 16 + movu m0, [srcq] + mova m1, [dstq] +%if %2 == 1 ; avg + pavgb m0, [secq] + punpckhbw m3, m1, m5 + punpcklbw m1, m5 +%endif + punpckhbw m2, m0, m5 + punpcklbw m0, m5 + +%if %2 == 0 ; !avg + punpckhbw m3, m1, m5 + punpcklbw m1, m5 +%endif + SUM_SSE m0, m1, m2, m3, m6, m7 + + add srcq, src_strideq + add dstq, dst_strideq +%else ; %1 < 16 + movx m0, [srcq] +%if %2 == 1 ; avg +%if %1 > 4 + movhps m0, [srcq+src_strideq] +%else ; 4xh + movx m1, [srcq+src_strideq] + punpckldq m0, m1 +%endif +%else ; !avg + movx m2, [srcq+src_strideq] +%endif + + movx m1, [dstq] + movx m3, [dstq+dst_strideq] + +%if %2 == 1 ; avg +%if %1 > 4 + pavgb m0, [secq] +%else + movh m2, [secq] + pavgb m0, m2 +%endif + punpcklbw m3, m5 + punpcklbw m1, m5 +%if %1 > 4 + punpckhbw m2, m0, m5 + punpcklbw m0, m5 +%else ; 4xh + punpcklbw m0, m5 + movhlps m2, m0 +%endif +%else ; !avg + punpcklbw m0, m5 + punpcklbw m2, m5 + punpcklbw m3, m5 + punpcklbw m1, m5 +%endif + SUM_SSE m0, m1, m2, m3, m6, m7 + + lea srcq, [srcq+src_strideq*2] + lea dstq, [dstq+dst_strideq*2] +%endif +%if %2 == 1 ; avg + add secq, sec_str +%endif + dec block_height + jg .x_zero_y_zero_loop + STORE_AND_RET %1 + +.x_zero_y_nonzero: + cmp y_offsetd, 4 + jne .x_zero_y_nonhalf + + ; x_offset == 0 && y_offset == 0.5 +.x_zero_y_half_loop: +%if %1 == 16 + movu m0, [srcq] + movu m4, [srcq+src_strideq] + mova m1, [dstq] + pavgb m0, m4 + punpckhbw m3, m1, m5 +%if %2 == 1 ; avg + pavgb m0, [secq] +%endif + punpcklbw m1, m5 + punpckhbw m2, m0, m5 + punpcklbw m0, m5 + SUM_SSE m0, m1, m2, m3, m6, m7 + + add srcq, src_strideq + add dstq, dst_strideq +%else ; %1 < 16 + movx m0, [srcq] + movx m2, [srcq+src_strideq] +%if %2 == 1 ; avg +%if %1 > 4 + movhps m2, [srcq+src_strideq*2] +%else ; 4xh + movx m1, [srcq+src_strideq*2] + punpckldq m2, m1 +%endif + movx m1, [dstq] +%if %1 > 4 + movlhps m0, m2 +%else ; 4xh + punpckldq m0, m2 +%endif + movx m3, [dstq+dst_strideq] + pavgb m0, m2 + punpcklbw m1, m5 +%if %1 > 4 + pavgb m0, [secq] + punpcklbw m3, m5 + punpckhbw m2, m0, m5 + punpcklbw m0, m5 +%else ; 4xh + movh m4, [secq] + pavgb m0, m4 + punpcklbw m3, m5 + punpcklbw m0, m5 + movhlps m2, m0 +%endif +%else ; !avg + movx m4, [srcq+src_strideq*2] + movx m1, [dstq] + pavgb m0, m2 + movx m3, [dstq+dst_strideq] + pavgb m2, m4 + punpcklbw m0, m5 + punpcklbw m2, m5 + punpcklbw m3, m5 + punpcklbw m1, m5 +%endif + SUM_SSE m0, m1, m2, m3, m6, m7 + + lea srcq, [srcq+src_strideq*2] + lea dstq, [dstq+dst_strideq*2] +%endif +%if %2 == 1 ; avg + add secq, sec_str +%endif + dec block_height + jg .x_zero_y_half_loop + STORE_AND_RET %1 + +.x_zero_y_nonhalf: + ; x_offset == 0 && y_offset == bilin interpolation +%ifdef PIC + lea bilin_filter, [bilin_filter_m] +%endif + shl y_offsetd, filter_idx_shift +%if ARCH_X86_64 && %1 > 4 + mova m8, [bilin_filter+y_offsetq] +%if notcpuflag(ssse3) ; FIXME(rbultje) don't scatter registers on x86-64 + mova m9, [bilin_filter+y_offsetq+16] +%endif + mova m10, [pw_8] +%define filter_y_a m8 +%define filter_y_b m9 +%define filter_rnd m10 +%else ; x86-32 or mmx +%if ARCH_X86=1 && CONFIG_PIC=1 +; x_offset == 0, reuse x_offset reg +%define tempq x_offsetq + add y_offsetq, g_bilin_filterm +%define filter_y_a [y_offsetq] +%define filter_y_b [y_offsetq+16] + mov tempq, g_pw_8m +%define filter_rnd [tempq] +%else + add y_offsetq, bilin_filter +%define filter_y_a [y_offsetq] +%define filter_y_b [y_offsetq+16] +%define filter_rnd [pw_8] +%endif +%endif + +.x_zero_y_other_loop: +%if %1 == 16 + movu m0, [srcq] + movu m4, [srcq+src_strideq] + mova m1, [dstq] +%if cpuflag(ssse3) + punpckhbw m2, m0, m4 + punpcklbw m0, m4 + pmaddubsw m2, filter_y_a + pmaddubsw m0, filter_y_a + paddw m2, filter_rnd + paddw m0, filter_rnd +%else + punpckhbw m2, m0, m5 + punpckhbw m3, m4, m5 + punpcklbw m0, m5 + punpcklbw m4, m5 + ; FIXME(rbultje) instead of out=((num-x)*in1+x*in2+rnd)>>log2(num), we can + ; also do out=in1+(((num-x)*(in2-in1)+rnd)>>log2(num)). Total number of + ; instructions is the same (5), but it is 1 mul instead of 2, so might be + ; slightly faster because of pmullw latency. It would also cut our rodata + ; tables in half for this function, and save 1-2 registers on x86-64. + pmullw m2, filter_y_a + pmullw m3, filter_y_b + paddw m2, filter_rnd + pmullw m0, filter_y_a + pmullw m4, filter_y_b + paddw m0, filter_rnd + paddw m2, m3 + paddw m0, m4 +%endif + psraw m2, 4 + psraw m0, 4 +%if %2 == 1 ; avg + ; FIXME(rbultje) pipeline + packuswb m0, m2 + pavgb m0, [secq] + punpckhbw m2, m0, m5 + punpcklbw m0, m5 +%endif + punpckhbw m3, m1, m5 + punpcklbw m1, m5 + SUM_SSE m0, m1, m2, m3, m6, m7 + + add srcq, src_strideq + add dstq, dst_strideq +%else ; %1 < 16 + movx m0, [srcq] + movx m2, [srcq+src_strideq] + movx m4, [srcq+src_strideq*2] + movx m3, [dstq+dst_strideq] +%if cpuflag(ssse3) + movx m1, [dstq] + punpcklbw m0, m2 + punpcklbw m2, m4 + pmaddubsw m0, filter_y_a + pmaddubsw m2, filter_y_a + punpcklbw m3, m5 + paddw m2, filter_rnd + paddw m0, filter_rnd +%else + punpcklbw m0, m5 + punpcklbw m2, m5 + punpcklbw m4, m5 + pmullw m0, filter_y_a + pmullw m1, m2, filter_y_b + punpcklbw m3, m5 + paddw m0, filter_rnd + pmullw m2, filter_y_a + pmullw m4, filter_y_b + paddw m0, m1 + paddw m2, filter_rnd + movx m1, [dstq] + paddw m2, m4 +%endif + psraw m0, 4 + psraw m2, 4 +%if %2 == 1 ; avg + ; FIXME(rbultje) pipeline +%if %1 == 4 + movlhps m0, m2 +%endif + packuswb m0, m2 +%if %1 > 4 + pavgb m0, [secq] + punpckhbw m2, m0, m5 + punpcklbw m0, m5 +%else ; 4xh + movh m2, [secq] + pavgb m0, m2 + punpcklbw m0, m5 + movhlps m2, m0 +%endif +%endif + punpcklbw m1, m5 + SUM_SSE m0, m1, m2, m3, m6, m7 + + lea srcq, [srcq+src_strideq*2] + lea dstq, [dstq+dst_strideq*2] +%endif +%if %2 == 1 ; avg + add secq, sec_str +%endif + dec block_height + jg .x_zero_y_other_loop +%undef filter_y_a +%undef filter_y_b +%undef filter_rnd + STORE_AND_RET %1 + +.x_nonzero: + cmp x_offsetd, 4 + jne .x_nonhalf + ; x_offset == 0.5 + test y_offsetd, y_offsetd + jnz .x_half_y_nonzero + + ; x_offset == 0.5 && y_offset == 0 +.x_half_y_zero_loop: +%if %1 == 16 + movu m0, [srcq] + movu m4, [srcq+1] + mova m1, [dstq] + pavgb m0, m4 + punpckhbw m3, m1, m5 +%if %2 == 1 ; avg + pavgb m0, [secq] +%endif + punpcklbw m1, m5 + punpckhbw m2, m0, m5 + punpcklbw m0, m5 + SUM_SSE m0, m1, m2, m3, m6, m7 + + add srcq, src_strideq + add dstq, dst_strideq +%else ; %1 < 16 + movx m0, [srcq] + movx m4, [srcq+1] +%if %2 == 1 ; avg +%if %1 > 4 + movhps m0, [srcq+src_strideq] + movhps m4, [srcq+src_strideq+1] +%else ; 4xh + movx m1, [srcq+src_strideq] + punpckldq m0, m1 + movx m2, [srcq+src_strideq+1] + punpckldq m4, m2 +%endif + movx m1, [dstq] + movx m3, [dstq+dst_strideq] + pavgb m0, m4 + punpcklbw m3, m5 +%if %1 > 4 + pavgb m0, [secq] + punpcklbw m1, m5 + punpckhbw m2, m0, m5 + punpcklbw m0, m5 +%else ; 4xh + movh m2, [secq] + pavgb m0, m2 + punpcklbw m1, m5 + punpcklbw m0, m5 + movhlps m2, m0 +%endif +%else ; !avg + movx m2, [srcq+src_strideq] + movx m1, [dstq] + pavgb m0, m4 + movx m4, [srcq+src_strideq+1] + movx m3, [dstq+dst_strideq] + pavgb m2, m4 + punpcklbw m0, m5 + punpcklbw m2, m5 + punpcklbw m3, m5 + punpcklbw m1, m5 +%endif + SUM_SSE m0, m1, m2, m3, m6, m7 + + lea srcq, [srcq+src_strideq*2] + lea dstq, [dstq+dst_strideq*2] +%endif +%if %2 == 1 ; avg + add secq, sec_str +%endif + dec block_height + jg .x_half_y_zero_loop + STORE_AND_RET %1 + +.x_half_y_nonzero: + cmp y_offsetd, 4 + jne .x_half_y_nonhalf + + ; x_offset == 0.5 && y_offset == 0.5 +%if %1 == 16 + movu m0, [srcq] + movu m3, [srcq+1] + add srcq, src_strideq + pavgb m0, m3 +.x_half_y_half_loop: + movu m4, [srcq] + movu m3, [srcq+1] + mova m1, [dstq] + pavgb m4, m3 + punpckhbw m3, m1, m5 + pavgb m0, m4 +%if %2 == 1 ; avg + punpcklbw m1, m5 + pavgb m0, [secq] + punpckhbw m2, m0, m5 + punpcklbw m0, m5 +%else + punpckhbw m2, m0, m5 + punpcklbw m0, m5 + punpcklbw m1, m5 +%endif + SUM_SSE m0, m1, m2, m3, m6, m7 + mova m0, m4 + + add srcq, src_strideq + add dstq, dst_strideq +%else ; %1 < 16 + movx m0, [srcq] + movx m3, [srcq+1] + add srcq, src_strideq + pavgb m0, m3 +.x_half_y_half_loop: + movx m2, [srcq] + movx m3, [srcq+1] +%if %2 == 1 ; avg +%if %1 > 4 + movhps m2, [srcq+src_strideq] + movhps m3, [srcq+src_strideq+1] +%else + movx m1, [srcq+src_strideq] + punpckldq m2, m1 + movx m1, [srcq+src_strideq+1] + punpckldq m3, m1 +%endif + pavgb m2, m3 +%if %1 > 4 + movlhps m0, m2 + movhlps m4, m2 +%else ; 4xh + punpckldq m0, m2 + pshuflw m4, m2, 0xe +%endif + movx m1, [dstq] + pavgb m0, m2 + movx m3, [dstq+dst_strideq] +%if %1 > 4 + pavgb m0, [secq] +%else + movh m2, [secq] + pavgb m0, m2 +%endif + punpcklbw m3, m5 + punpcklbw m1, m5 +%if %1 > 4 + punpckhbw m2, m0, m5 + punpcklbw m0, m5 +%else + punpcklbw m0, m5 + movhlps m2, m0 +%endif +%else ; !avg + movx m4, [srcq+src_strideq] + movx m1, [srcq+src_strideq+1] + pavgb m2, m3 + pavgb m4, m1 + pavgb m0, m2 + pavgb m2, m4 + movx m1, [dstq] + movx m3, [dstq+dst_strideq] + punpcklbw m0, m5 + punpcklbw m2, m5 + punpcklbw m3, m5 + punpcklbw m1, m5 +%endif + SUM_SSE m0, m1, m2, m3, m6, m7 + mova m0, m4 + + lea srcq, [srcq+src_strideq*2] + lea dstq, [dstq+dst_strideq*2] +%endif +%if %2 == 1 ; avg + add secq, sec_str +%endif + dec block_height + jg .x_half_y_half_loop + STORE_AND_RET %1 + +.x_half_y_nonhalf: + ; x_offset == 0.5 && y_offset == bilin interpolation +%ifdef PIC + lea bilin_filter, [bilin_filter_m] +%endif + shl y_offsetd, filter_idx_shift +%if ARCH_X86_64 && %1 > 4 + mova m8, [bilin_filter+y_offsetq] +%if notcpuflag(ssse3) ; FIXME(rbultje) don't scatter registers on x86-64 + mova m9, [bilin_filter+y_offsetq+16] +%endif + mova m10, [pw_8] +%define filter_y_a m8 +%define filter_y_b m9 +%define filter_rnd m10 +%else ;x86_32 +%if ARCH_X86=1 && CONFIG_PIC=1 +; x_offset == 0.5. We can reuse x_offset reg +%define tempq x_offsetq + add y_offsetq, g_bilin_filterm +%define filter_y_a [y_offsetq] +%define filter_y_b [y_offsetq+16] + mov tempq, g_pw_8m +%define filter_rnd [tempq] +%else + add y_offsetq, bilin_filter +%define filter_y_a [y_offsetq] +%define filter_y_b [y_offsetq+16] +%define filter_rnd [pw_8] +%endif +%endif + +%if %1 == 16 + movu m0, [srcq] + movu m3, [srcq+1] + add srcq, src_strideq + pavgb m0, m3 +.x_half_y_other_loop: + movu m4, [srcq] + movu m2, [srcq+1] + mova m1, [dstq] + pavgb m4, m2 +%if cpuflag(ssse3) + punpckhbw m2, m0, m4 + punpcklbw m0, m4 + pmaddubsw m2, filter_y_a + pmaddubsw m0, filter_y_a + paddw m2, filter_rnd + paddw m0, filter_rnd + psraw m2, 4 +%else + punpckhbw m2, m0, m5 + punpckhbw m3, m4, m5 + pmullw m2, filter_y_a + pmullw m3, filter_y_b + paddw m2, filter_rnd + punpcklbw m0, m5 + paddw m2, m3 + punpcklbw m3, m4, m5 + pmullw m0, filter_y_a + pmullw m3, filter_y_b + paddw m0, filter_rnd + psraw m2, 4 + paddw m0, m3 +%endif + punpckhbw m3, m1, m5 + psraw m0, 4 +%if %2 == 1 ; avg + ; FIXME(rbultje) pipeline + packuswb m0, m2 + pavgb m0, [secq] + punpckhbw m2, m0, m5 + punpcklbw m0, m5 +%endif + punpcklbw m1, m5 + SUM_SSE m0, m1, m2, m3, m6, m7 + mova m0, m4 + + add srcq, src_strideq + add dstq, dst_strideq +%else ; %1 < 16 + movx m0, [srcq] + movx m3, [srcq+1] + add srcq, src_strideq + pavgb m0, m3 +%if notcpuflag(ssse3) + punpcklbw m0, m5 +%endif +.x_half_y_other_loop: + movx m2, [srcq] + movx m1, [srcq+1] + movx m4, [srcq+src_strideq] + movx m3, [srcq+src_strideq+1] + pavgb m2, m1 + pavgb m4, m3 + movx m3, [dstq+dst_strideq] +%if cpuflag(ssse3) + movx m1, [dstq] + punpcklbw m0, m2 + punpcklbw m2, m4 + pmaddubsw m0, filter_y_a + pmaddubsw m2, filter_y_a + punpcklbw m3, m5 + paddw m0, filter_rnd + paddw m2, filter_rnd +%else + punpcklbw m2, m5 + punpcklbw m4, m5 + pmullw m0, filter_y_a + pmullw m1, m2, filter_y_b + punpcklbw m3, m5 + paddw m0, filter_rnd + pmullw m2, filter_y_a + paddw m0, m1 + pmullw m1, m4, filter_y_b + paddw m2, filter_rnd + paddw m2, m1 + movx m1, [dstq] +%endif + psraw m0, 4 + psraw m2, 4 +%if %2 == 1 ; avg + ; FIXME(rbultje) pipeline +%if %1 == 4 + movlhps m0, m2 +%endif + packuswb m0, m2 +%if %1 > 4 + pavgb m0, [secq] + punpckhbw m2, m0, m5 + punpcklbw m0, m5 +%else + movh m2, [secq] + pavgb m0, m2 + punpcklbw m0, m5 + movhlps m2, m0 +%endif +%endif + punpcklbw m1, m5 + SUM_SSE m0, m1, m2, m3, m6, m7 + mova m0, m4 + + lea srcq, [srcq+src_strideq*2] + lea dstq, [dstq+dst_strideq*2] +%endif +%if %2 == 1 ; avg + add secq, sec_str +%endif + dec block_height + jg .x_half_y_other_loop +%undef filter_y_a +%undef filter_y_b +%undef filter_rnd + STORE_AND_RET %1 + +.x_nonhalf: + test y_offsetd, y_offsetd + jnz .x_nonhalf_y_nonzero + + ; x_offset == bilin interpolation && y_offset == 0 +%ifdef PIC + lea bilin_filter, [bilin_filter_m] +%endif + shl x_offsetd, filter_idx_shift +%if ARCH_X86_64 && %1 > 4 + mova m8, [bilin_filter+x_offsetq] +%if notcpuflag(ssse3) ; FIXME(rbultje) don't scatter registers on x86-64 + mova m9, [bilin_filter+x_offsetq+16] +%endif + mova m10, [pw_8] +%define filter_x_a m8 +%define filter_x_b m9 +%define filter_rnd m10 +%else ; x86-32 +%if ARCH_X86=1 && CONFIG_PIC=1 +;y_offset == 0. We can reuse y_offset reg. +%define tempq y_offsetq + add x_offsetq, g_bilin_filterm +%define filter_x_a [x_offsetq] +%define filter_x_b [x_offsetq+16] + mov tempq, g_pw_8m +%define filter_rnd [tempq] +%else + add x_offsetq, bilin_filter +%define filter_x_a [x_offsetq] +%define filter_x_b [x_offsetq+16] +%define filter_rnd [pw_8] +%endif +%endif + +.x_other_y_zero_loop: +%if %1 == 16 + movu m0, [srcq] + movu m4, [srcq+1] + mova m1, [dstq] +%if cpuflag(ssse3) + punpckhbw m2, m0, m4 + punpcklbw m0, m4 + pmaddubsw m2, filter_x_a + pmaddubsw m0, filter_x_a + paddw m2, filter_rnd + paddw m0, filter_rnd +%else + punpckhbw m2, m0, m5 + punpckhbw m3, m4, m5 + punpcklbw m0, m5 + punpcklbw m4, m5 + pmullw m2, filter_x_a + pmullw m3, filter_x_b + paddw m2, filter_rnd + pmullw m0, filter_x_a + pmullw m4, filter_x_b + paddw m0, filter_rnd + paddw m2, m3 + paddw m0, m4 +%endif + psraw m2, 4 + psraw m0, 4 +%if %2 == 1 ; avg + ; FIXME(rbultje) pipeline + packuswb m0, m2 + pavgb m0, [secq] + punpckhbw m2, m0, m5 + punpcklbw m0, m5 +%endif + punpckhbw m3, m1, m5 + punpcklbw m1, m5 + SUM_SSE m0, m1, m2, m3, m6, m7 + + add srcq, src_strideq + add dstq, dst_strideq +%else ; %1 < 16 + movx m0, [srcq] + movx m1, [srcq+1] + movx m2, [srcq+src_strideq] + movx m4, [srcq+src_strideq+1] + movx m3, [dstq+dst_strideq] +%if cpuflag(ssse3) + punpcklbw m0, m1 + movx m1, [dstq] + punpcklbw m2, m4 + pmaddubsw m0, filter_x_a + pmaddubsw m2, filter_x_a + punpcklbw m3, m5 + paddw m0, filter_rnd + paddw m2, filter_rnd +%else + punpcklbw m0, m5 + punpcklbw m1, m5 + punpcklbw m2, m5 + punpcklbw m4, m5 + pmullw m0, filter_x_a + pmullw m1, filter_x_b + punpcklbw m3, m5 + paddw m0, filter_rnd + pmullw m2, filter_x_a + pmullw m4, filter_x_b + paddw m0, m1 + paddw m2, filter_rnd + movx m1, [dstq] + paddw m2, m4 +%endif + psraw m0, 4 + psraw m2, 4 +%if %2 == 1 ; avg + ; FIXME(rbultje) pipeline +%if %1 == 4 + movlhps m0, m2 +%endif + packuswb m0, m2 +%if %1 > 4 + pavgb m0, [secq] + punpckhbw m2, m0, m5 + punpcklbw m0, m5 +%else + movh m2, [secq] + pavgb m0, m2 + punpcklbw m0, m5 + movhlps m2, m0 +%endif +%endif + punpcklbw m1, m5 + SUM_SSE m0, m1, m2, m3, m6, m7 + + lea srcq, [srcq+src_strideq*2] + lea dstq, [dstq+dst_strideq*2] +%endif +%if %2 == 1 ; avg + add secq, sec_str +%endif + dec block_height + jg .x_other_y_zero_loop +%undef filter_x_a +%undef filter_x_b +%undef filter_rnd + STORE_AND_RET %1 + +.x_nonhalf_y_nonzero: + cmp y_offsetd, 4 + jne .x_nonhalf_y_nonhalf + + ; x_offset == bilin interpolation && y_offset == 0.5 +%ifdef PIC + lea bilin_filter, [bilin_filter_m] +%endif + shl x_offsetd, filter_idx_shift +%if ARCH_X86_64 && %1 > 4 + mova m8, [bilin_filter+x_offsetq] +%if notcpuflag(ssse3) ; FIXME(rbultje) don't scatter registers on x86-64 + mova m9, [bilin_filter+x_offsetq+16] +%endif + mova m10, [pw_8] +%define filter_x_a m8 +%define filter_x_b m9 +%define filter_rnd m10 +%else ; x86-32 +%if ARCH_X86=1 && CONFIG_PIC=1 +; y_offset == 0.5. We can reuse y_offset reg. +%define tempq y_offsetq + add x_offsetq, g_bilin_filterm +%define filter_x_a [x_offsetq] +%define filter_x_b [x_offsetq+16] + mov tempq, g_pw_8m +%define filter_rnd [tempq] +%else + add x_offsetq, bilin_filter +%define filter_x_a [x_offsetq] +%define filter_x_b [x_offsetq+16] +%define filter_rnd [pw_8] +%endif +%endif + +%if %1 == 16 + movu m0, [srcq] + movu m1, [srcq+1] +%if cpuflag(ssse3) + punpckhbw m2, m0, m1 + punpcklbw m0, m1 + pmaddubsw m2, filter_x_a + pmaddubsw m0, filter_x_a + paddw m2, filter_rnd + paddw m0, filter_rnd +%else + punpckhbw m2, m0, m5 + punpckhbw m3, m1, m5 + punpcklbw m0, m5 + punpcklbw m1, m5 + pmullw m0, filter_x_a + pmullw m1, filter_x_b + paddw m0, filter_rnd + pmullw m2, filter_x_a + pmullw m3, filter_x_b + paddw m2, filter_rnd + paddw m0, m1 + paddw m2, m3 +%endif + psraw m0, 4 + psraw m2, 4 + add srcq, src_strideq + packuswb m0, m2 +.x_other_y_half_loop: + movu m4, [srcq] + movu m3, [srcq+1] +%if cpuflag(ssse3) + mova m1, [dstq] + punpckhbw m2, m4, m3 + punpcklbw m4, m3 + pmaddubsw m2, filter_x_a + pmaddubsw m4, filter_x_a + paddw m2, filter_rnd + paddw m4, filter_rnd + psraw m2, 4 + psraw m4, 4 + packuswb m4, m2 + pavgb m0, m4 + punpckhbw m3, m1, m5 + punpcklbw m1, m5 +%else + punpckhbw m2, m4, m5 + punpckhbw m1, m3, m5 + punpcklbw m4, m5 + punpcklbw m3, m5 + pmullw m4, filter_x_a + pmullw m3, filter_x_b + paddw m4, filter_rnd + pmullw m2, filter_x_a + pmullw m1, filter_x_b + paddw m2, filter_rnd + paddw m4, m3 + paddw m2, m1 + mova m1, [dstq] + psraw m4, 4 + psraw m2, 4 + punpckhbw m3, m1, m5 + ; FIXME(rbultje) the repeated pack/unpack here around m0/m2 is because we + ; have a 1-register shortage to be able to store the backup of the bilin + ; filtered second line as words as cache for the next line. Packing into + ; a byte costs 1 pack and 2 unpacks, but saves a register. + packuswb m4, m2 + punpcklbw m1, m5 + pavgb m0, m4 +%endif +%if %2 == 1 ; avg + ; FIXME(rbultje) pipeline + pavgb m0, [secq] +%endif + punpckhbw m2, m0, m5 + punpcklbw m0, m5 + SUM_SSE m0, m1, m2, m3, m6, m7 + mova m0, m4 + + add srcq, src_strideq + add dstq, dst_strideq +%else ; %1 < 16 + movx m0, [srcq] + movx m1, [srcq+1] +%if cpuflag(ssse3) + punpcklbw m0, m1 + pmaddubsw m0, filter_x_a + paddw m0, filter_rnd +%else + punpcklbw m0, m5 + punpcklbw m1, m5 + pmullw m0, filter_x_a + pmullw m1, filter_x_b + paddw m0, filter_rnd + paddw m0, m1 +%endif + add srcq, src_strideq + psraw m0, 4 +.x_other_y_half_loop: + movx m2, [srcq] + movx m1, [srcq+1] + movx m4, [srcq+src_strideq] + movx m3, [srcq+src_strideq+1] +%if cpuflag(ssse3) + punpcklbw m2, m1 + punpcklbw m4, m3 + pmaddubsw m2, filter_x_a + pmaddubsw m4, filter_x_a + movx m1, [dstq] + movx m3, [dstq+dst_strideq] + paddw m2, filter_rnd + paddw m4, filter_rnd +%else + punpcklbw m2, m5 + punpcklbw m1, m5 + punpcklbw m4, m5 + punpcklbw m3, m5 + pmullw m2, filter_x_a + pmullw m1, filter_x_b + paddw m2, filter_rnd + pmullw m4, filter_x_a + pmullw m3, filter_x_b + paddw m4, filter_rnd + paddw m2, m1 + movx m1, [dstq] + paddw m4, m3 + movx m3, [dstq+dst_strideq] +%endif + psraw m2, 4 + psraw m4, 4 + pavgw m0, m2 + pavgw m2, m4 +%if %2 == 1 ; avg + ; FIXME(rbultje) pipeline - also consider going to bytes here +%if %1 == 4 + movlhps m0, m2 +%endif + packuswb m0, m2 +%if %1 > 4 + pavgb m0, [secq] + punpckhbw m2, m0, m5 + punpcklbw m0, m5 +%else + movh m2, [secq] + pavgb m0, m2 + punpcklbw m0, m5 + movhlps m2, m0 +%endif +%endif + punpcklbw m3, m5 + punpcklbw m1, m5 + SUM_SSE m0, m1, m2, m3, m6, m7 + mova m0, m4 + + lea srcq, [srcq+src_strideq*2] + lea dstq, [dstq+dst_strideq*2] +%endif +%if %2 == 1 ; avg + add secq, sec_str +%endif + dec block_height + jg .x_other_y_half_loop +%undef filter_x_a +%undef filter_x_b +%undef filter_rnd + STORE_AND_RET %1 + +.x_nonhalf_y_nonhalf: +%ifdef PIC + lea bilin_filter, [bilin_filter_m] +%endif + shl x_offsetd, filter_idx_shift + shl y_offsetd, filter_idx_shift +%if ARCH_X86_64 && %1 > 4 + mova m8, [bilin_filter+x_offsetq] +%if notcpuflag(ssse3) ; FIXME(rbultje) don't scatter registers on x86-64 + mova m9, [bilin_filter+x_offsetq+16] +%endif + mova m10, [bilin_filter+y_offsetq] +%if notcpuflag(ssse3) ; FIXME(rbultje) don't scatter registers on x86-64 + mova m11, [bilin_filter+y_offsetq+16] +%endif + mova m12, [pw_8] +%define filter_x_a m8 +%define filter_x_b m9 +%define filter_y_a m10 +%define filter_y_b m11 +%define filter_rnd m12 +%else ; x86-32 +%if ARCH_X86=1 && CONFIG_PIC=1 +; In this case, there is NO unused register. Used src_stride register. Later, +; src_stride has to be loaded from stack when it is needed. +%define tempq src_strideq + mov tempq, g_bilin_filterm + add x_offsetq, tempq + add y_offsetq, tempq +%define filter_x_a [x_offsetq] +%define filter_x_b [x_offsetq+16] +%define filter_y_a [y_offsetq] +%define filter_y_b [y_offsetq+16] + + mov tempq, g_pw_8m +%define filter_rnd [tempq] +%else + add x_offsetq, bilin_filter + add y_offsetq, bilin_filter +%define filter_x_a [x_offsetq] +%define filter_x_b [x_offsetq+16] +%define filter_y_a [y_offsetq] +%define filter_y_b [y_offsetq+16] +%define filter_rnd [pw_8] +%endif +%endif + + ; x_offset == bilin interpolation && y_offset == bilin interpolation +%if %1 == 16 + movu m0, [srcq] + movu m1, [srcq+1] +%if cpuflag(ssse3) + punpckhbw m2, m0, m1 + punpcklbw m0, m1 + pmaddubsw m2, filter_x_a + pmaddubsw m0, filter_x_a + paddw m2, filter_rnd + paddw m0, filter_rnd +%else + punpckhbw m2, m0, m5 + punpckhbw m3, m1, m5 + punpcklbw m0, m5 + punpcklbw m1, m5 + pmullw m0, filter_x_a + pmullw m1, filter_x_b + paddw m0, filter_rnd + pmullw m2, filter_x_a + pmullw m3, filter_x_b + paddw m2, filter_rnd + paddw m0, m1 + paddw m2, m3 +%endif + psraw m0, 4 + psraw m2, 4 + + INC_SRC_BY_SRC_STRIDE + + packuswb m0, m2 +.x_other_y_other_loop: +%if cpuflag(ssse3) + movu m4, [srcq] + movu m3, [srcq+1] + mova m1, [dstq] + punpckhbw m2, m4, m3 + punpcklbw m4, m3 + pmaddubsw m2, filter_x_a + pmaddubsw m4, filter_x_a + punpckhbw m3, m1, m5 + paddw m2, filter_rnd + paddw m4, filter_rnd + psraw m2, 4 + psraw m4, 4 + packuswb m4, m2 + punpckhbw m2, m0, m4 + punpcklbw m0, m4 + pmaddubsw m2, filter_y_a + pmaddubsw m0, filter_y_a + punpcklbw m1, m5 + paddw m2, filter_rnd + paddw m0, filter_rnd + psraw m2, 4 + psraw m0, 4 +%else + movu m3, [srcq] + movu m4, [srcq+1] + punpckhbw m1, m3, m5 + punpckhbw m2, m4, m5 + punpcklbw m3, m5 + punpcklbw m4, m5 + pmullw m3, filter_x_a + pmullw m4, filter_x_b + paddw m3, filter_rnd + pmullw m1, filter_x_a + pmullw m2, filter_x_b + paddw m1, filter_rnd + paddw m3, m4 + paddw m1, m2 + psraw m3, 4 + psraw m1, 4 + packuswb m4, m3, m1 + punpckhbw m2, m0, m5 + punpcklbw m0, m5 + pmullw m2, filter_y_a + pmullw m1, filter_y_b + paddw m2, filter_rnd + pmullw m0, filter_y_a + pmullw m3, filter_y_b + paddw m2, m1 + mova m1, [dstq] + paddw m0, filter_rnd + psraw m2, 4 + paddw m0, m3 + punpckhbw m3, m1, m5 + psraw m0, 4 + punpcklbw m1, m5 +%endif +%if %2 == 1 ; avg + ; FIXME(rbultje) pipeline + packuswb m0, m2 + pavgb m0, [secq] + punpckhbw m2, m0, m5 + punpcklbw m0, m5 +%endif + SUM_SSE m0, m1, m2, m3, m6, m7 + mova m0, m4 + + INC_SRC_BY_SRC_STRIDE + add dstq, dst_strideq +%else ; %1 < 16 + movx m0, [srcq] + movx m1, [srcq+1] +%if cpuflag(ssse3) + punpcklbw m0, m1 + pmaddubsw m0, filter_x_a + paddw m0, filter_rnd +%else + punpcklbw m0, m5 + punpcklbw m1, m5 + pmullw m0, filter_x_a + pmullw m1, filter_x_b + paddw m0, filter_rnd + paddw m0, m1 +%endif + psraw m0, 4 +%if cpuflag(ssse3) + packuswb m0, m0 +%endif + + INC_SRC_BY_SRC_STRIDE + +.x_other_y_other_loop: + movx m2, [srcq] + movx m1, [srcq+1] + + INC_SRC_BY_SRC_STRIDE + movx m4, [srcq] + movx m3, [srcq+1] + +%if cpuflag(ssse3) + punpcklbw m2, m1 + punpcklbw m4, m3 + pmaddubsw m2, filter_x_a + pmaddubsw m4, filter_x_a + movx m3, [dstq+dst_strideq] + movx m1, [dstq] + paddw m2, filter_rnd + paddw m4, filter_rnd + psraw m2, 4 + psraw m4, 4 + packuswb m2, m2 + packuswb m4, m4 + punpcklbw m0, m2 + punpcklbw m2, m4 + pmaddubsw m0, filter_y_a + pmaddubsw m2, filter_y_a + punpcklbw m3, m5 + paddw m0, filter_rnd + paddw m2, filter_rnd + psraw m0, 4 + psraw m2, 4 + punpcklbw m1, m5 +%else + punpcklbw m2, m5 + punpcklbw m1, m5 + punpcklbw m4, m5 + punpcklbw m3, m5 + pmullw m2, filter_x_a + pmullw m1, filter_x_b + paddw m2, filter_rnd + pmullw m4, filter_x_a + pmullw m3, filter_x_b + paddw m4, filter_rnd + paddw m2, m1 + paddw m4, m3 + psraw m2, 4 + psraw m4, 4 + pmullw m0, filter_y_a + pmullw m3, m2, filter_y_b + paddw m0, filter_rnd + pmullw m2, filter_y_a + pmullw m1, m4, filter_y_b + paddw m2, filter_rnd + paddw m0, m3 + movx m3, [dstq+dst_strideq] + paddw m2, m1 + movx m1, [dstq] + psraw m0, 4 + psraw m2, 4 + punpcklbw m3, m5 + punpcklbw m1, m5 +%endif +%if %2 == 1 ; avg + ; FIXME(rbultje) pipeline +%if %1 == 4 + movlhps m0, m2 +%endif + packuswb m0, m2 +%if %1 > 4 + pavgb m0, [secq] + punpckhbw m2, m0, m5 + punpcklbw m0, m5 +%else + movh m2, [secq] + pavgb m0, m2 + punpcklbw m0, m5 + movhlps m2, m0 +%endif +%endif + SUM_SSE m0, m1, m2, m3, m6, m7 + mova m0, m4 + + INC_SRC_BY_SRC_STRIDE + lea dstq, [dstq+dst_strideq*2] +%endif +%if %2 == 1 ; avg + add secq, sec_str +%endif + dec block_height + jg .x_other_y_other_loop +%undef filter_x_a +%undef filter_x_b +%undef filter_y_a +%undef filter_y_b +%undef filter_rnd +%undef movx + STORE_AND_RET %1 +%endmacro + +; FIXME(rbultje) the non-bilinear versions (i.e. x=0,8&&y=0,8) are identical +; between the ssse3 and non-ssse3 version. It may make sense to merge their +; code in the sense that the ssse3 version would jump to the appropriate +; location in the sse/2 version, rather than duplicating that code in the +; binary. + +INIT_XMM sse2 +SUBPEL_VARIANCE 4 +SUBPEL_VARIANCE 8 +SUBPEL_VARIANCE 16 + +INIT_XMM ssse3 +SUBPEL_VARIANCE 4 +SUBPEL_VARIANCE 8 +SUBPEL_VARIANCE 16 + +INIT_XMM sse2 +SUBPEL_VARIANCE 4, 1 +SUBPEL_VARIANCE 8, 1 +SUBPEL_VARIANCE 16, 1 + +INIT_XMM ssse3 +SUBPEL_VARIANCE 4, 1 +SUBPEL_VARIANCE 8, 1 +SUBPEL_VARIANCE 16, 1 diff --git a/third_party/aom/aom_dsp/x86/subtract_sse2.asm b/third_party/aom/aom_dsp/x86/subtract_sse2.asm new file mode 100644 index 000000000..7bd5b23ad --- /dev/null +++ b/third_party/aom/aom_dsp/x86/subtract_sse2.asm @@ -0,0 +1,150 @@ +; +; Copyright (c) 2016, Alliance for Open Media. All rights reserved +; +; This source code is subject to the terms of the BSD 2 Clause License and +; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License +; was not distributed with this source code in the LICENSE file, you can +; obtain it at www.aomedia.org/license/software. If the Alliance for Open +; Media Patent License 1.0 was not distributed with this source code in the +; PATENTS file, you can obtain it at www.aomedia.org/license/patent. +; + +; + +%include "third_party/x86inc/x86inc.asm" + +SECTION .text + +; void aom_subtract_block(int rows, int cols, +; int16_t *diff, ptrdiff_t diff_stride, +; const uint8_t *src, ptrdiff_t src_stride, +; const uint8_t *pred, ptrdiff_t pred_stride) + +INIT_XMM sse2 +cglobal subtract_block, 7, 7, 8, \ + rows, cols, diff, diff_stride, src, src_stride, \ + pred, pred_stride +%define pred_str colsq + pxor m7, m7 ; dedicated zero register + cmp colsd, 4 + je .case_4 + cmp colsd, 8 + je .case_8 + cmp colsd, 16 + je .case_16 + cmp colsd, 32 + je .case_32 +%if CONFIG_EXT_PARTITION + cmp colsd, 64 + je .case_64 +%endif + +%macro loop16 6 + mova m0, [srcq+%1] + mova m4, [srcq+%2] + mova m1, [predq+%3] + mova m5, [predq+%4] + punpckhbw m2, m0, m7 + punpckhbw m3, m1, m7 + punpcklbw m0, m7 + punpcklbw m1, m7 + psubw m2, m3 + psubw m0, m1 + punpckhbw m1, m4, m7 + punpckhbw m3, m5, m7 + punpcklbw m4, m7 + punpcklbw m5, m7 + psubw m1, m3 + psubw m4, m5 + mova [diffq+mmsize*0+%5], m0 + mova [diffq+mmsize*1+%5], m2 + mova [diffq+mmsize*0+%6], m4 + mova [diffq+mmsize*1+%6], m1 +%endmacro + +%if CONFIG_EXT_PARTITION + mov pred_str, pred_stridemp +.loop_128: + loop16 0*mmsize, 1*mmsize, 0*mmsize, 1*mmsize, 0*mmsize, 2*mmsize + loop16 2*mmsize, 3*mmsize, 2*mmsize, 3*mmsize, 4*mmsize, 6*mmsize + loop16 4*mmsize, 5*mmsize, 4*mmsize, 5*mmsize, 8*mmsize, 10*mmsize + loop16 6*mmsize, 7*mmsize, 6*mmsize, 7*mmsize, 12*mmsize, 14*mmsize + lea diffq, [diffq+diff_strideq*2] + add predq, pred_str + add srcq, src_strideq + sub rowsd, 1 + jnz .loop_128 + RET + +.case_64: +%endif + mov pred_str, pred_stridemp +.loop_64: + loop16 0*mmsize, 1*mmsize, 0*mmsize, 1*mmsize, 0*mmsize, 2*mmsize + loop16 2*mmsize, 3*mmsize, 2*mmsize, 3*mmsize, 4*mmsize, 6*mmsize + lea diffq, [diffq+diff_strideq*2] + add predq, pred_str + add srcq, src_strideq + dec rowsd + jg .loop_64 + RET + +.case_32: + mov pred_str, pred_stridemp +.loop_32: + loop16 0, mmsize, 0, mmsize, 0, 2*mmsize + lea diffq, [diffq+diff_strideq*2] + add predq, pred_str + add srcq, src_strideq + dec rowsd + jg .loop_32 + RET + +.case_16: + mov pred_str, pred_stridemp +.loop_16: + loop16 0, src_strideq, 0, pred_str, 0, diff_strideq*2 + lea diffq, [diffq+diff_strideq*4] + lea predq, [predq+pred_str*2] + lea srcq, [srcq+src_strideq*2] + sub rowsd, 2 + jg .loop_16 + RET + +%macro loop_h 0 + movh m0, [srcq] + movh m2, [srcq+src_strideq] + movh m1, [predq] + movh m3, [predq+pred_str] + punpcklbw m0, m7 + punpcklbw m1, m7 + punpcklbw m2, m7 + punpcklbw m3, m7 + psubw m0, m1 + psubw m2, m3 + mova [diffq], m0 + mova [diffq+diff_strideq*2], m2 +%endmacro + +.case_8: + mov pred_str, pred_stridemp +.loop_8: + loop_h + lea diffq, [diffq+diff_strideq*4] + lea srcq, [srcq+src_strideq*2] + lea predq, [predq+pred_str*2] + sub rowsd, 2 + jg .loop_8 + RET + +INIT_MMX +.case_4: + mov pred_str, pred_stridemp +.loop_4: + loop_h + lea diffq, [diffq+diff_strideq*4] + lea srcq, [srcq+src_strideq*2] + lea predq, [predq+pred_str*2] + sub rowsd, 2 + jg .loop_4 + RET diff --git a/third_party/aom/aom_dsp/x86/sum_squares_sse2.c b/third_party/aom/aom_dsp/x86/sum_squares_sse2.c new file mode 100644 index 000000000..6be99fbca --- /dev/null +++ b/third_party/aom/aom_dsp/x86/sum_squares_sse2.c @@ -0,0 +1,210 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include <assert.h> +#include <emmintrin.h> +#include <stdio.h> + +#include "aom_dsp/x86/synonyms.h" + +#include "./aom_dsp_rtcd.h" + +static uint64_t aom_sum_squares_2d_i16_4x4_sse2(const int16_t *src, + int stride) { + const __m128i v_val_0_w = + _mm_loadl_epi64((const __m128i *)(src + 0 * stride)); + const __m128i v_val_1_w = + _mm_loadl_epi64((const __m128i *)(src + 1 * stride)); + const __m128i v_val_2_w = + _mm_loadl_epi64((const __m128i *)(src + 2 * stride)); + const __m128i v_val_3_w = + _mm_loadl_epi64((const __m128i *)(src + 3 * stride)); + + const __m128i v_sq_0_d = _mm_madd_epi16(v_val_0_w, v_val_0_w); + const __m128i v_sq_1_d = _mm_madd_epi16(v_val_1_w, v_val_1_w); + const __m128i v_sq_2_d = _mm_madd_epi16(v_val_2_w, v_val_2_w); + const __m128i v_sq_3_d = _mm_madd_epi16(v_val_3_w, v_val_3_w); + + const __m128i v_sum_01_d = _mm_add_epi32(v_sq_0_d, v_sq_1_d); + const __m128i v_sum_23_d = _mm_add_epi32(v_sq_2_d, v_sq_3_d); + const __m128i v_sum_0123_d = _mm_add_epi32(v_sum_01_d, v_sum_23_d); + + const __m128i v_sum_d = + _mm_add_epi32(v_sum_0123_d, _mm_srli_epi64(v_sum_0123_d, 32)); + + return (uint64_t)_mm_cvtsi128_si32(v_sum_d); +} + +#ifdef __GNUC__ +// This prevents GCC/Clang from inlining this function into +// aom_sum_squares_2d_i16_sse2, which in turn saves some stack +// maintenance instructions in the common case of 4x4. +__attribute__((noinline)) +#endif +static uint64_t +aom_sum_squares_2d_i16_nxn_sse2(const int16_t *src, int stride, int width, + int height) { + int r, c; + + const __m128i v_zext_mask_q = _mm_set_epi32(0, 0xffffffff, 0, 0xffffffff); + __m128i v_acc_q = _mm_setzero_si128(); + + for (r = 0; r < height; r += 8) { + __m128i v_acc_d = _mm_setzero_si128(); + + for (c = 0; c < width; c += 8) { + const int16_t *b = src + c; + + const __m128i v_val_0_w = + _mm_load_si128((const __m128i *)(b + 0 * stride)); + const __m128i v_val_1_w = + _mm_load_si128((const __m128i *)(b + 1 * stride)); + const __m128i v_val_2_w = + _mm_load_si128((const __m128i *)(b + 2 * stride)); + const __m128i v_val_3_w = + _mm_load_si128((const __m128i *)(b + 3 * stride)); + const __m128i v_val_4_w = + _mm_load_si128((const __m128i *)(b + 4 * stride)); + const __m128i v_val_5_w = + _mm_load_si128((const __m128i *)(b + 5 * stride)); + const __m128i v_val_6_w = + _mm_load_si128((const __m128i *)(b + 6 * stride)); + const __m128i v_val_7_w = + _mm_load_si128((const __m128i *)(b + 7 * stride)); + + const __m128i v_sq_0_d = _mm_madd_epi16(v_val_0_w, v_val_0_w); + const __m128i v_sq_1_d = _mm_madd_epi16(v_val_1_w, v_val_1_w); + const __m128i v_sq_2_d = _mm_madd_epi16(v_val_2_w, v_val_2_w); + const __m128i v_sq_3_d = _mm_madd_epi16(v_val_3_w, v_val_3_w); + const __m128i v_sq_4_d = _mm_madd_epi16(v_val_4_w, v_val_4_w); + const __m128i v_sq_5_d = _mm_madd_epi16(v_val_5_w, v_val_5_w); + const __m128i v_sq_6_d = _mm_madd_epi16(v_val_6_w, v_val_6_w); + const __m128i v_sq_7_d = _mm_madd_epi16(v_val_7_w, v_val_7_w); + + const __m128i v_sum_01_d = _mm_add_epi32(v_sq_0_d, v_sq_1_d); + const __m128i v_sum_23_d = _mm_add_epi32(v_sq_2_d, v_sq_3_d); + const __m128i v_sum_45_d = _mm_add_epi32(v_sq_4_d, v_sq_5_d); + const __m128i v_sum_67_d = _mm_add_epi32(v_sq_6_d, v_sq_7_d); + + const __m128i v_sum_0123_d = _mm_add_epi32(v_sum_01_d, v_sum_23_d); + const __m128i v_sum_4567_d = _mm_add_epi32(v_sum_45_d, v_sum_67_d); + + v_acc_d = _mm_add_epi32(v_acc_d, v_sum_0123_d); + v_acc_d = _mm_add_epi32(v_acc_d, v_sum_4567_d); + } + + v_acc_q = _mm_add_epi64(v_acc_q, _mm_and_si128(v_acc_d, v_zext_mask_q)); + v_acc_q = _mm_add_epi64(v_acc_q, _mm_srli_epi64(v_acc_d, 32)); + + src += 8 * stride; + } + + v_acc_q = _mm_add_epi64(v_acc_q, _mm_srli_si128(v_acc_q, 8)); + +#if ARCH_X86_64 + return (uint64_t)_mm_cvtsi128_si64(v_acc_q); +#else + { + uint64_t tmp; + _mm_storel_epi64((__m128i *)&tmp, v_acc_q); + return tmp; + } +#endif +} + +uint64_t aom_sum_squares_2d_i16_sse2(const int16_t *src, int stride, int width, + int height) { + // 4 elements per row only requires half an XMM register, so this + // must be a special case, but also note that over 75% of all calls + // are with size == 4, so it is also the common case. + if (LIKELY(width == 4 && height == 4)) { + return aom_sum_squares_2d_i16_4x4_sse2(src, stride); + } else if (LIKELY(width % 8 == 0 && height % 8 == 0)) { + // Generic case + return aom_sum_squares_2d_i16_nxn_sse2(src, stride, width, height); + } else { + return aom_sum_squares_2d_i16_c(src, stride, width, height); + } +} + +////////////////////////////////////////////////////////////////////////////// +// 1D version +////////////////////////////////////////////////////////////////////////////// + +static uint64_t aom_sum_squares_i16_64n_sse2(const int16_t *src, uint32_t n) { + const __m128i v_zext_mask_q = _mm_set_epi32(0, 0xffffffff, 0, 0xffffffff); + __m128i v_acc0_q = _mm_setzero_si128(); + __m128i v_acc1_q = _mm_setzero_si128(); + + const int16_t *const end = src + n; + + assert(n % 64 == 0); + + while (src < end) { + const __m128i v_val_0_w = xx_load_128(src); + const __m128i v_val_1_w = xx_load_128(src + 8); + const __m128i v_val_2_w = xx_load_128(src + 16); + const __m128i v_val_3_w = xx_load_128(src + 24); + const __m128i v_val_4_w = xx_load_128(src + 32); + const __m128i v_val_5_w = xx_load_128(src + 40); + const __m128i v_val_6_w = xx_load_128(src + 48); + const __m128i v_val_7_w = xx_load_128(src + 56); + + const __m128i v_sq_0_d = _mm_madd_epi16(v_val_0_w, v_val_0_w); + const __m128i v_sq_1_d = _mm_madd_epi16(v_val_1_w, v_val_1_w); + const __m128i v_sq_2_d = _mm_madd_epi16(v_val_2_w, v_val_2_w); + const __m128i v_sq_3_d = _mm_madd_epi16(v_val_3_w, v_val_3_w); + const __m128i v_sq_4_d = _mm_madd_epi16(v_val_4_w, v_val_4_w); + const __m128i v_sq_5_d = _mm_madd_epi16(v_val_5_w, v_val_5_w); + const __m128i v_sq_6_d = _mm_madd_epi16(v_val_6_w, v_val_6_w); + const __m128i v_sq_7_d = _mm_madd_epi16(v_val_7_w, v_val_7_w); + + const __m128i v_sum_01_d = _mm_add_epi32(v_sq_0_d, v_sq_1_d); + const __m128i v_sum_23_d = _mm_add_epi32(v_sq_2_d, v_sq_3_d); + const __m128i v_sum_45_d = _mm_add_epi32(v_sq_4_d, v_sq_5_d); + const __m128i v_sum_67_d = _mm_add_epi32(v_sq_6_d, v_sq_7_d); + + const __m128i v_sum_0123_d = _mm_add_epi32(v_sum_01_d, v_sum_23_d); + const __m128i v_sum_4567_d = _mm_add_epi32(v_sum_45_d, v_sum_67_d); + + const __m128i v_sum_d = _mm_add_epi32(v_sum_0123_d, v_sum_4567_d); + + v_acc0_q = _mm_add_epi64(v_acc0_q, _mm_and_si128(v_sum_d, v_zext_mask_q)); + v_acc1_q = _mm_add_epi64(v_acc1_q, _mm_srli_epi64(v_sum_d, 32)); + + src += 64; + } + + v_acc0_q = _mm_add_epi64(v_acc0_q, v_acc1_q); + v_acc0_q = _mm_add_epi64(v_acc0_q, _mm_srli_si128(v_acc0_q, 8)); + +#if ARCH_X86_64 + return (uint64_t)_mm_cvtsi128_si64(v_acc0_q); +#else + { + uint64_t tmp; + _mm_storel_epi64((__m128i *)&tmp, v_acc0_q); + return tmp; + } +#endif +} + +uint64_t aom_sum_squares_i16_sse2(const int16_t *src, uint32_t n) { + if (n % 64 == 0) { + return aom_sum_squares_i16_64n_sse2(src, n); + } else if (n > 64) { + int k = n & ~(64 - 1); + return aom_sum_squares_i16_64n_sse2(src, k) + + aom_sum_squares_i16_c(src + k, n - k); + } else { + return aom_sum_squares_i16_c(src, n); + } +} diff --git a/third_party/aom/aom_dsp/x86/synonyms.h b/third_party/aom/aom_dsp/x86/synonyms.h new file mode 100644 index 000000000..bef606dae --- /dev/null +++ b/third_party/aom/aom_dsp/x86/synonyms.h @@ -0,0 +1,120 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#ifndef AOM_DSP_X86_SYNONYMS_H_ +#define AOM_DSP_X86_SYNONYMS_H_ + +#include <immintrin.h> + +#include "./aom_config.h" +#include "aom/aom_integer.h" + +/** + * Various reusable shorthands for x86 SIMD intrinsics. + * + * Intrinsics prefixed with xx_ operate on or return 128bit XMM registers. + * Intrinsics prefixed with yy_ operate on or return 256bit YMM registers. + */ + +// Loads and stores to do away with the tedium of casting the address +// to the right type. +static INLINE __m128i xx_loadl_32(const void *a) { + return _mm_cvtsi32_si128(*(const uint32_t *)a); +} + +static INLINE __m128i xx_loadl_64(const void *a) { + return _mm_loadl_epi64((const __m128i *)a); +} + +static INLINE __m128i xx_load_128(const void *a) { + return _mm_load_si128((const __m128i *)a); +} + +static INLINE __m128i xx_loadu_128(const void *a) { + return _mm_loadu_si128((const __m128i *)a); +} + +static INLINE void xx_storel_32(void *const a, const __m128i v) { + *(uint32_t *)a = _mm_cvtsi128_si32(v); +} + +static INLINE void xx_storel_64(void *const a, const __m128i v) { + _mm_storel_epi64((__m128i *)a, v); +} + +static INLINE void xx_store_128(void *const a, const __m128i v) { + _mm_store_si128((__m128i *)a, v); +} + +static INLINE void xx_storeu_128(void *const a, const __m128i v) { + _mm_storeu_si128((__m128i *)a, v); +} + +static INLINE __m128i xx_round_epu16(__m128i v_val_w) { + return _mm_avg_epu16(v_val_w, _mm_setzero_si128()); +} + +static INLINE __m128i xx_roundn_epu16(__m128i v_val_w, int bits) { + const __m128i v_s_w = _mm_srli_epi16(v_val_w, bits - 1); + return _mm_avg_epu16(v_s_w, _mm_setzero_si128()); +} + +static INLINE __m128i xx_roundn_epu32(__m128i v_val_d, int bits) { + const __m128i v_bias_d = _mm_set1_epi32((1 << bits) >> 1); + const __m128i v_tmp_d = _mm_add_epi32(v_val_d, v_bias_d); + return _mm_srli_epi32(v_tmp_d, bits); +} + +// This is equivalent to ROUND_POWER_OF_TWO(v_val_d, bits) +static INLINE __m128i xx_roundn_epi32_unsigned(__m128i v_val_d, int bits) { + const __m128i v_bias_d = _mm_set1_epi32((1 << bits) >> 1); + const __m128i v_tmp_d = _mm_add_epi32(v_val_d, v_bias_d); + return _mm_srai_epi32(v_tmp_d, bits); +} + +// This is equivalent to ROUND_POWER_OF_TWO_SIGNED(v_val_d, bits) +static INLINE __m128i xx_roundn_epi32(__m128i v_val_d, int bits) { + const __m128i v_bias_d = _mm_set1_epi32((1 << bits) >> 1); + const __m128i v_sign_d = _mm_srai_epi32(v_val_d, 31); + const __m128i v_tmp_d = + _mm_add_epi32(_mm_add_epi32(v_val_d, v_bias_d), v_sign_d); + return _mm_srai_epi32(v_tmp_d, bits); +} + +#ifdef __SSSE3__ +static INLINE int32_t xx_hsum_epi32_si32(__m128i v_d) { + v_d = _mm_hadd_epi32(v_d, v_d); + v_d = _mm_hadd_epi32(v_d, v_d); + return _mm_cvtsi128_si32(v_d); +} + +static INLINE int64_t xx_hsum_epi64_si64(__m128i v_q) { + v_q = _mm_add_epi64(v_q, _mm_srli_si128(v_q, 8)); +#if ARCH_X86_64 + return _mm_cvtsi128_si64(v_q); +#else + { + int64_t tmp; + _mm_storel_epi64((__m128i *)&tmp, v_q); + return tmp; + } +#endif +} + +static INLINE int64_t xx_hsum_epi32_si64(__m128i v_d) { + const __m128i v_sign_d = _mm_cmplt_epi32(v_d, _mm_setzero_si128()); + const __m128i v_0_q = _mm_unpacklo_epi32(v_d, v_sign_d); + const __m128i v_1_q = _mm_unpackhi_epi32(v_d, v_sign_d); + return xx_hsum_epi64_si64(_mm_add_epi64(v_0_q, v_1_q)); +} +#endif // __SSSE3__ + +#endif // AOM_DSP_X86_SYNONYMS_H_ diff --git a/third_party/aom/aom_dsp/x86/txfm_common_avx2.h b/third_party/aom/aom_dsp/x86/txfm_common_avx2.h new file mode 100644 index 000000000..39e9b8e2a --- /dev/null +++ b/third_party/aom/aom_dsp/x86/txfm_common_avx2.h @@ -0,0 +1,204 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#ifndef AOM_DSP_X86_TXFM_COMMON_AVX2_H +#define AOM_DSP_X86_TXFM_COMMON_AVX2_H + +#include <immintrin.h> + +#include "aom_dsp/txfm_common.h" + +#define pair256_set_epi16(a, b) \ + _mm256_set_epi16((int16_t)(b), (int16_t)(a), (int16_t)(b), (int16_t)(a), \ + (int16_t)(b), (int16_t)(a), (int16_t)(b), (int16_t)(a), \ + (int16_t)(b), (int16_t)(a), (int16_t)(b), (int16_t)(a), \ + (int16_t)(b), (int16_t)(a), (int16_t)(b), (int16_t)(a)) + +#define pair256_set_epi32(a, b) \ + _mm256_set_epi32((int)(b), (int)(a), (int)(b), (int)(a), (int)(b), (int)(a), \ + (int)(b), (int)(a)) + +static INLINE void mm256_reverse_epi16(__m256i *u) { + const __m256i control = _mm256_set_epi16( + 0x0100, 0x0302, 0x0504, 0x0706, 0x0908, 0x0B0A, 0x0D0C, 0x0F0E, 0x0100, + 0x0302, 0x0504, 0x0706, 0x0908, 0x0B0A, 0x0D0C, 0x0F0E); + __m256i v = _mm256_shuffle_epi8(*u, control); + *u = _mm256_permute2x128_si256(v, v, 1); +} + +static INLINE void mm256_transpose_16x16(__m256i *in) { + __m256i tr0_0 = _mm256_unpacklo_epi16(in[0], in[1]); + __m256i tr0_1 = _mm256_unpackhi_epi16(in[0], in[1]); + __m256i tr0_2 = _mm256_unpacklo_epi16(in[2], in[3]); + __m256i tr0_3 = _mm256_unpackhi_epi16(in[2], in[3]); + __m256i tr0_4 = _mm256_unpacklo_epi16(in[4], in[5]); + __m256i tr0_5 = _mm256_unpackhi_epi16(in[4], in[5]); + __m256i tr0_6 = _mm256_unpacklo_epi16(in[6], in[7]); + __m256i tr0_7 = _mm256_unpackhi_epi16(in[6], in[7]); + + __m256i tr0_8 = _mm256_unpacklo_epi16(in[8], in[9]); + __m256i tr0_9 = _mm256_unpackhi_epi16(in[8], in[9]); + __m256i tr0_a = _mm256_unpacklo_epi16(in[10], in[11]); + __m256i tr0_b = _mm256_unpackhi_epi16(in[10], in[11]); + __m256i tr0_c = _mm256_unpacklo_epi16(in[12], in[13]); + __m256i tr0_d = _mm256_unpackhi_epi16(in[12], in[13]); + __m256i tr0_e = _mm256_unpacklo_epi16(in[14], in[15]); + __m256i tr0_f = _mm256_unpackhi_epi16(in[14], in[15]); + + // 00 10 01 11 02 12 03 13 08 18 09 19 0a 1a 0b 1b + // 04 14 05 15 06 16 07 17 0c 1c 0d 1d 0e 1e 0f 1f + // 20 30 21 31 22 32 23 33 28 38 29 39 2a 3a 2b 3b + // 24 34 25 35 26 36 27 37 2c 3c 2d 3d 2e 3e 2f 3f + // 40 50 41 51 42 52 43 53 48 58 49 59 4a 5a 4b 5b + // 44 54 45 55 46 56 47 57 4c 5c 4d 5d 4e 5e 4f 5f + // 60 70 61 71 62 72 63 73 68 78 69 79 6a 7a 6b 7b + // 64 74 65 75 66 76 67 77 6c 7c 6d 7d 6e 7e 6f 7f + + // 80 90 81 91 82 92 83 93 88 98 89 99 8a 9a 8b 9b + // 84 94 85 95 86 96 87 97 8c 9c 8d 9d 8e 9e 8f 9f + // a0 b0 a1 b1 a2 b2 a3 b3 a8 b8 a9 b9 aa ba ab bb + // a4 b4 a5 b5 a6 b6 a7 b7 ac bc ad bd ae be af bf + // c0 d0 c1 d1 c2 d2 c3 d3 c8 d8 c9 d9 ca da cb db + // c4 d4 c5 d5 c6 d6 c7 d7 cc dc cd dd ce de cf df + // e0 f0 e1 f1 e2 f2 e3 f3 e8 f8 e9 f9 ea fa eb fb + // e4 f4 e5 f5 e6 f6 e7 f7 ec fc ed fd ee fe ef ff + + __m256i tr1_0 = _mm256_unpacklo_epi32(tr0_0, tr0_2); + __m256i tr1_1 = _mm256_unpackhi_epi32(tr0_0, tr0_2); + __m256i tr1_2 = _mm256_unpacklo_epi32(tr0_1, tr0_3); + __m256i tr1_3 = _mm256_unpackhi_epi32(tr0_1, tr0_3); + __m256i tr1_4 = _mm256_unpacklo_epi32(tr0_4, tr0_6); + __m256i tr1_5 = _mm256_unpackhi_epi32(tr0_4, tr0_6); + __m256i tr1_6 = _mm256_unpacklo_epi32(tr0_5, tr0_7); + __m256i tr1_7 = _mm256_unpackhi_epi32(tr0_5, tr0_7); + + __m256i tr1_8 = _mm256_unpacklo_epi32(tr0_8, tr0_a); + __m256i tr1_9 = _mm256_unpackhi_epi32(tr0_8, tr0_a); + __m256i tr1_a = _mm256_unpacklo_epi32(tr0_9, tr0_b); + __m256i tr1_b = _mm256_unpackhi_epi32(tr0_9, tr0_b); + __m256i tr1_c = _mm256_unpacklo_epi32(tr0_c, tr0_e); + __m256i tr1_d = _mm256_unpackhi_epi32(tr0_c, tr0_e); + __m256i tr1_e = _mm256_unpacklo_epi32(tr0_d, tr0_f); + __m256i tr1_f = _mm256_unpackhi_epi32(tr0_d, tr0_f); + + // 00 10 20 30 01 11 21 31 08 18 28 38 09 19 29 39 + // 02 12 22 32 03 13 23 33 0a 1a 2a 3a 0b 1b 2b 3b + // 04 14 24 34 05 15 25 35 0c 1c 2c 3c 0d 1d 2d 3d + // 06 16 26 36 07 17 27 37 0e 1e 2e 3e 0f 1f 2f 3f + // 40 50 60 70 41 51 61 71 48 58 68 78 49 59 69 79 + // 42 52 62 72 43 53 63 73 4a 5a 6a 7a 4b 5b 6b 7b + // 44 54 64 74 45 55 65 75 4c 5c 6c 7c 4d 5d 6d 7d + // 46 56 66 76 47 57 67 77 4e 5e 6e 7e 4f 5f 6f 7f + + // 80 90 a0 b0 81 91 a1 b1 88 98 a8 b8 89 99 a9 b9 + // 82 92 a2 b2 83 93 a3 b3 8a 9a aa ba 8b 9b ab bb + // 84 94 a4 b4 85 95 a5 b5 8c 9c ac bc 8d 9d ad bd + // 86 96 a6 b6 87 97 a7 b7 8e ae 9e be 8f 9f af bf + // c0 d0 e0 f0 c1 d1 e1 f1 c8 d8 e8 f8 c9 d9 e9 f9 + // c2 d2 e2 f2 c3 d3 e3 f3 ca da ea fa cb db eb fb + // c4 d4 e4 f4 c5 d5 e5 f5 cc dc ef fc cd dd ed fd + // c6 d6 e6 f6 c7 d7 e7 f7 ce de ee fe cf df ef ff + + tr0_0 = _mm256_unpacklo_epi64(tr1_0, tr1_4); + tr0_1 = _mm256_unpackhi_epi64(tr1_0, tr1_4); + tr0_2 = _mm256_unpacklo_epi64(tr1_1, tr1_5); + tr0_3 = _mm256_unpackhi_epi64(tr1_1, tr1_5); + tr0_4 = _mm256_unpacklo_epi64(tr1_2, tr1_6); + tr0_5 = _mm256_unpackhi_epi64(tr1_2, tr1_6); + tr0_6 = _mm256_unpacklo_epi64(tr1_3, tr1_7); + tr0_7 = _mm256_unpackhi_epi64(tr1_3, tr1_7); + + tr0_8 = _mm256_unpacklo_epi64(tr1_8, tr1_c); + tr0_9 = _mm256_unpackhi_epi64(tr1_8, tr1_c); + tr0_a = _mm256_unpacklo_epi64(tr1_9, tr1_d); + tr0_b = _mm256_unpackhi_epi64(tr1_9, tr1_d); + tr0_c = _mm256_unpacklo_epi64(tr1_a, tr1_e); + tr0_d = _mm256_unpackhi_epi64(tr1_a, tr1_e); + tr0_e = _mm256_unpacklo_epi64(tr1_b, tr1_f); + tr0_f = _mm256_unpackhi_epi64(tr1_b, tr1_f); + + // 00 10 20 30 40 50 60 70 08 18 28 38 48 58 68 78 + // 01 11 21 31 41 51 61 71 09 19 29 39 49 59 69 79 + // 02 12 22 32 42 52 62 72 0a 1a 2a 3a 4a 5a 6a 7a + // 03 13 23 33 43 53 63 73 0b 1b 2b 3b 4b 5b 6b 7b + // 04 14 24 34 44 54 64 74 0c 1c 2c 3c 4c 5c 6c 7c + // 05 15 25 35 45 55 65 75 0d 1d 2d 3d 4d 5d 6d 7d + // 06 16 26 36 46 56 66 76 0e 1e 2e 3e 4e 5e 6e 7e + // 07 17 27 37 47 57 67 77 0f 1f 2f 3f 4f 5f 6f 7f + + // 80 90 a0 b0 c0 d0 e0 f0 88 98 a8 b8 c8 d8 e8 f8 + // 81 91 a1 b1 c1 d1 e1 f1 89 99 a9 b9 c9 d9 e9 f9 + // 82 92 a2 b2 c2 d2 e2 f2 8a 9a aa ba ca da ea fa + // 83 93 a3 b3 c3 d3 e3 f3 8b 9b ab bb cb db eb fb + // 84 94 a4 b4 c4 d4 e4 f4 8c 9c ac bc cc dc ef fc + // 85 95 a5 b5 c5 d5 e5 f5 8d 9d ad bd cd dd ed fd + // 86 96 a6 b6 c6 d6 e6 f6 8e ae 9e be ce de ee fe + // 87 97 a7 b7 c7 d7 e7 f7 8f 9f af bf cf df ef ff + + in[0] = _mm256_permute2x128_si256(tr0_0, tr0_8, 0x20); // 0010 0000 + in[8] = _mm256_permute2x128_si256(tr0_0, tr0_8, 0x31); // 0011 0001 + in[1] = _mm256_permute2x128_si256(tr0_1, tr0_9, 0x20); + in[9] = _mm256_permute2x128_si256(tr0_1, tr0_9, 0x31); + in[2] = _mm256_permute2x128_si256(tr0_2, tr0_a, 0x20); + in[10] = _mm256_permute2x128_si256(tr0_2, tr0_a, 0x31); + in[3] = _mm256_permute2x128_si256(tr0_3, tr0_b, 0x20); + in[11] = _mm256_permute2x128_si256(tr0_3, tr0_b, 0x31); + + in[4] = _mm256_permute2x128_si256(tr0_4, tr0_c, 0x20); + in[12] = _mm256_permute2x128_si256(tr0_4, tr0_c, 0x31); + in[5] = _mm256_permute2x128_si256(tr0_5, tr0_d, 0x20); + in[13] = _mm256_permute2x128_si256(tr0_5, tr0_d, 0x31); + in[6] = _mm256_permute2x128_si256(tr0_6, tr0_e, 0x20); + in[14] = _mm256_permute2x128_si256(tr0_6, tr0_e, 0x31); + in[7] = _mm256_permute2x128_si256(tr0_7, tr0_f, 0x20); + in[15] = _mm256_permute2x128_si256(tr0_7, tr0_f, 0x31); +} + +static INLINE __m256i butter_fly(__m256i a0, __m256i a1, const __m256i cospi) { + const __m256i dct_rounding = _mm256_set1_epi32(DCT_CONST_ROUNDING); + __m256i y0 = _mm256_madd_epi16(a0, cospi); + __m256i y1 = _mm256_madd_epi16(a1, cospi); + + y0 = _mm256_add_epi32(y0, dct_rounding); + y1 = _mm256_add_epi32(y1, dct_rounding); + y0 = _mm256_srai_epi32(y0, DCT_CONST_BITS); + y1 = _mm256_srai_epi32(y1, DCT_CONST_BITS); + + return _mm256_packs_epi32(y0, y1); +} + +static INLINE void txfm_scaling16_avx2(const int16_t c, __m256i *in) { + const __m256i zero = _mm256_setzero_si256(); + const __m256i sqrt2_epi16 = _mm256_set1_epi16(c); + const __m256i dct_const_rounding = _mm256_set1_epi32(DCT_CONST_ROUNDING); + __m256i u0, u1; + int i = 0; + + while (i < 16) { + in[i] = _mm256_slli_epi16(in[i], 1); + + u0 = _mm256_unpacklo_epi16(zero, in[i]); + u1 = _mm256_unpackhi_epi16(zero, in[i]); + + u0 = _mm256_madd_epi16(u0, sqrt2_epi16); + u1 = _mm256_madd_epi16(u1, sqrt2_epi16); + + u0 = _mm256_add_epi32(u0, dct_const_rounding); + u1 = _mm256_add_epi32(u1, dct_const_rounding); + + u0 = _mm256_srai_epi32(u0, DCT_CONST_BITS); + u1 = _mm256_srai_epi32(u1, DCT_CONST_BITS); + in[i] = _mm256_packs_epi32(u0, u1); + i++; + } +} + +#endif // AOM_DSP_X86_TXFM_COMMON_AVX2_H diff --git a/third_party/aom/aom_dsp/x86/txfm_common_intrin.h b/third_party/aom/aom_dsp/x86/txfm_common_intrin.h new file mode 100644 index 000000000..e4ac56339 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/txfm_common_intrin.h @@ -0,0 +1,31 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#ifndef _AOM_DSP_X86_TXFM_COMMON_INTRIN_H_ +#define _AOM_DSP_X86_TXFM_COMMON_INTRIN_H_ + +// Note: +// This header file should be put below any x86 intrinsics head file + +static INLINE void storeu_output(const __m128i *poutput, tran_low_t *dst_ptr) { +#if CONFIG_HIGHBITDEPTH + const __m128i zero = _mm_setzero_si128(); + const __m128i sign_bits = _mm_cmplt_epi16(*poutput, zero); + __m128i out0 = _mm_unpacklo_epi16(*poutput, sign_bits); + __m128i out1 = _mm_unpackhi_epi16(*poutput, sign_bits); + _mm_storeu_si128((__m128i *)(dst_ptr), out0); + _mm_storeu_si128((__m128i *)(dst_ptr + 4), out1); +#else + _mm_storeu_si128((__m128i *)(dst_ptr), *poutput); +#endif // CONFIG_HIGHBITDEPTH +} + +#endif // _AOM_DSP_X86_TXFM_COMMON_INTRIN_H_ diff --git a/third_party/aom/aom_dsp/x86/txfm_common_sse2.h b/third_party/aom/aom_dsp/x86/txfm_common_sse2.h new file mode 100644 index 000000000..4257d8b9c --- /dev/null +++ b/third_party/aom/aom_dsp/x86/txfm_common_sse2.h @@ -0,0 +1,326 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#ifndef AOM_DSP_X86_TXFM_COMMON_SSE2_H_ +#define AOM_DSP_X86_TXFM_COMMON_SSE2_H_ + +#include <emmintrin.h> +#include "aom/aom_integer.h" +#include "aom_dsp/x86/synonyms.h" + +#define pair_set_epi16(a, b) \ + _mm_set_epi16((int16_t)(b), (int16_t)(a), (int16_t)(b), (int16_t)(a), \ + (int16_t)(b), (int16_t)(a), (int16_t)(b), (int16_t)(a)) + +#define dual_set_epi16(a, b) \ + _mm_set_epi16((int16_t)(b), (int16_t)(b), (int16_t)(b), (int16_t)(b), \ + (int16_t)(a), (int16_t)(a), (int16_t)(a), (int16_t)(a)) + +#define octa_set_epi16(a, b, c, d, e, f, g, h) \ + _mm_setr_epi16((int16_t)(a), (int16_t)(b), (int16_t)(c), (int16_t)(d), \ + (int16_t)(e), (int16_t)(f), (int16_t)(g), (int16_t)(h)) + +// Reverse the 8 16 bit words in __m128i +static INLINE __m128i mm_reverse_epi16(const __m128i x) { + const __m128i a = _mm_shufflelo_epi16(x, 0x1b); + const __m128i b = _mm_shufflehi_epi16(a, 0x1b); + return _mm_shuffle_epi32(b, 0x4e); +} + +#if CONFIG_EXT_TX +// Identity transform (both forward and inverse). +static INLINE void idtx16_8col(__m128i *in) { + const __m128i k__zero_epi16 = _mm_set1_epi16((int16_t)0); + const __m128i k__sqrt2_epi16 = _mm_set1_epi16((int16_t)Sqrt2); + const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING); + + __m128i v0, v1, v2, v3, v4, v5, v6, v7; + __m128i u0, u1, u2, u3, u4, u5, u6, u7; + __m128i x0, x1, x2, x3, x4, x5, x6, x7; + __m128i y0, y1, y2, y3, y4, y5, y6, y7; + + in[0] = _mm_slli_epi16(in[0], 1); + in[1] = _mm_slli_epi16(in[1], 1); + in[2] = _mm_slli_epi16(in[2], 1); + in[3] = _mm_slli_epi16(in[3], 1); + in[4] = _mm_slli_epi16(in[4], 1); + in[5] = _mm_slli_epi16(in[5], 1); + in[6] = _mm_slli_epi16(in[6], 1); + in[7] = _mm_slli_epi16(in[7], 1); + in[8] = _mm_slli_epi16(in[8], 1); + in[9] = _mm_slli_epi16(in[9], 1); + in[10] = _mm_slli_epi16(in[10], 1); + in[11] = _mm_slli_epi16(in[11], 1); + in[12] = _mm_slli_epi16(in[12], 1); + in[13] = _mm_slli_epi16(in[13], 1); + in[14] = _mm_slli_epi16(in[14], 1); + in[15] = _mm_slli_epi16(in[15], 1); + + v0 = _mm_unpacklo_epi16(in[0], k__zero_epi16); + v1 = _mm_unpacklo_epi16(in[1], k__zero_epi16); + v2 = _mm_unpacklo_epi16(in[2], k__zero_epi16); + v3 = _mm_unpacklo_epi16(in[3], k__zero_epi16); + v4 = _mm_unpacklo_epi16(in[4], k__zero_epi16); + v5 = _mm_unpacklo_epi16(in[5], k__zero_epi16); + v6 = _mm_unpacklo_epi16(in[6], k__zero_epi16); + v7 = _mm_unpacklo_epi16(in[7], k__zero_epi16); + + u0 = _mm_unpacklo_epi16(in[8], k__zero_epi16); + u1 = _mm_unpacklo_epi16(in[9], k__zero_epi16); + u2 = _mm_unpacklo_epi16(in[10], k__zero_epi16); + u3 = _mm_unpacklo_epi16(in[11], k__zero_epi16); + u4 = _mm_unpacklo_epi16(in[12], k__zero_epi16); + u5 = _mm_unpacklo_epi16(in[13], k__zero_epi16); + u6 = _mm_unpacklo_epi16(in[14], k__zero_epi16); + u7 = _mm_unpacklo_epi16(in[15], k__zero_epi16); + + x0 = _mm_unpackhi_epi16(in[0], k__zero_epi16); + x1 = _mm_unpackhi_epi16(in[1], k__zero_epi16); + x2 = _mm_unpackhi_epi16(in[2], k__zero_epi16); + x3 = _mm_unpackhi_epi16(in[3], k__zero_epi16); + x4 = _mm_unpackhi_epi16(in[4], k__zero_epi16); + x5 = _mm_unpackhi_epi16(in[5], k__zero_epi16); + x6 = _mm_unpackhi_epi16(in[6], k__zero_epi16); + x7 = _mm_unpackhi_epi16(in[7], k__zero_epi16); + + y0 = _mm_unpackhi_epi16(in[8], k__zero_epi16); + y1 = _mm_unpackhi_epi16(in[9], k__zero_epi16); + y2 = _mm_unpackhi_epi16(in[10], k__zero_epi16); + y3 = _mm_unpackhi_epi16(in[11], k__zero_epi16); + y4 = _mm_unpackhi_epi16(in[12], k__zero_epi16); + y5 = _mm_unpackhi_epi16(in[13], k__zero_epi16); + y6 = _mm_unpackhi_epi16(in[14], k__zero_epi16); + y7 = _mm_unpackhi_epi16(in[15], k__zero_epi16); + + v0 = _mm_madd_epi16(v0, k__sqrt2_epi16); + v1 = _mm_madd_epi16(v1, k__sqrt2_epi16); + v2 = _mm_madd_epi16(v2, k__sqrt2_epi16); + v3 = _mm_madd_epi16(v3, k__sqrt2_epi16); + v4 = _mm_madd_epi16(v4, k__sqrt2_epi16); + v5 = _mm_madd_epi16(v5, k__sqrt2_epi16); + v6 = _mm_madd_epi16(v6, k__sqrt2_epi16); + v7 = _mm_madd_epi16(v7, k__sqrt2_epi16); + + x0 = _mm_madd_epi16(x0, k__sqrt2_epi16); + x1 = _mm_madd_epi16(x1, k__sqrt2_epi16); + x2 = _mm_madd_epi16(x2, k__sqrt2_epi16); + x3 = _mm_madd_epi16(x3, k__sqrt2_epi16); + x4 = _mm_madd_epi16(x4, k__sqrt2_epi16); + x5 = _mm_madd_epi16(x5, k__sqrt2_epi16); + x6 = _mm_madd_epi16(x6, k__sqrt2_epi16); + x7 = _mm_madd_epi16(x7, k__sqrt2_epi16); + + u0 = _mm_madd_epi16(u0, k__sqrt2_epi16); + u1 = _mm_madd_epi16(u1, k__sqrt2_epi16); + u2 = _mm_madd_epi16(u2, k__sqrt2_epi16); + u3 = _mm_madd_epi16(u3, k__sqrt2_epi16); + u4 = _mm_madd_epi16(u4, k__sqrt2_epi16); + u5 = _mm_madd_epi16(u5, k__sqrt2_epi16); + u6 = _mm_madd_epi16(u6, k__sqrt2_epi16); + u7 = _mm_madd_epi16(u7, k__sqrt2_epi16); + + y0 = _mm_madd_epi16(y0, k__sqrt2_epi16); + y1 = _mm_madd_epi16(y1, k__sqrt2_epi16); + y2 = _mm_madd_epi16(y2, k__sqrt2_epi16); + y3 = _mm_madd_epi16(y3, k__sqrt2_epi16); + y4 = _mm_madd_epi16(y4, k__sqrt2_epi16); + y5 = _mm_madd_epi16(y5, k__sqrt2_epi16); + y6 = _mm_madd_epi16(y6, k__sqrt2_epi16); + y7 = _mm_madd_epi16(y7, k__sqrt2_epi16); + + v0 = _mm_add_epi32(v0, k__DCT_CONST_ROUNDING); + v1 = _mm_add_epi32(v1, k__DCT_CONST_ROUNDING); + v2 = _mm_add_epi32(v2, k__DCT_CONST_ROUNDING); + v3 = _mm_add_epi32(v3, k__DCT_CONST_ROUNDING); + v4 = _mm_add_epi32(v4, k__DCT_CONST_ROUNDING); + v5 = _mm_add_epi32(v5, k__DCT_CONST_ROUNDING); + v6 = _mm_add_epi32(v6, k__DCT_CONST_ROUNDING); + v7 = _mm_add_epi32(v7, k__DCT_CONST_ROUNDING); + + x0 = _mm_add_epi32(x0, k__DCT_CONST_ROUNDING); + x1 = _mm_add_epi32(x1, k__DCT_CONST_ROUNDING); + x2 = _mm_add_epi32(x2, k__DCT_CONST_ROUNDING); + x3 = _mm_add_epi32(x3, k__DCT_CONST_ROUNDING); + x4 = _mm_add_epi32(x4, k__DCT_CONST_ROUNDING); + x5 = _mm_add_epi32(x5, k__DCT_CONST_ROUNDING); + x6 = _mm_add_epi32(x6, k__DCT_CONST_ROUNDING); + x7 = _mm_add_epi32(x7, k__DCT_CONST_ROUNDING); + + u0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING); + u1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING); + u2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING); + u3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING); + u4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING); + u5 = _mm_add_epi32(u5, k__DCT_CONST_ROUNDING); + u6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING); + u7 = _mm_add_epi32(u7, k__DCT_CONST_ROUNDING); + + y0 = _mm_add_epi32(y0, k__DCT_CONST_ROUNDING); + y1 = _mm_add_epi32(y1, k__DCT_CONST_ROUNDING); + y2 = _mm_add_epi32(y2, k__DCT_CONST_ROUNDING); + y3 = _mm_add_epi32(y3, k__DCT_CONST_ROUNDING); + y4 = _mm_add_epi32(y4, k__DCT_CONST_ROUNDING); + y5 = _mm_add_epi32(y5, k__DCT_CONST_ROUNDING); + y6 = _mm_add_epi32(y6, k__DCT_CONST_ROUNDING); + y7 = _mm_add_epi32(y7, k__DCT_CONST_ROUNDING); + + v0 = _mm_srai_epi32(v0, DCT_CONST_BITS); + v1 = _mm_srai_epi32(v1, DCT_CONST_BITS); + v2 = _mm_srai_epi32(v2, DCT_CONST_BITS); + v3 = _mm_srai_epi32(v3, DCT_CONST_BITS); + v4 = _mm_srai_epi32(v4, DCT_CONST_BITS); + v5 = _mm_srai_epi32(v5, DCT_CONST_BITS); + v6 = _mm_srai_epi32(v6, DCT_CONST_BITS); + v7 = _mm_srai_epi32(v7, DCT_CONST_BITS); + + x0 = _mm_srai_epi32(x0, DCT_CONST_BITS); + x1 = _mm_srai_epi32(x1, DCT_CONST_BITS); + x2 = _mm_srai_epi32(x2, DCT_CONST_BITS); + x3 = _mm_srai_epi32(x3, DCT_CONST_BITS); + x4 = _mm_srai_epi32(x4, DCT_CONST_BITS); + x5 = _mm_srai_epi32(x5, DCT_CONST_BITS); + x6 = _mm_srai_epi32(x6, DCT_CONST_BITS); + x7 = _mm_srai_epi32(x7, DCT_CONST_BITS); + + u0 = _mm_srai_epi32(u0, DCT_CONST_BITS); + u1 = _mm_srai_epi32(u1, DCT_CONST_BITS); + u2 = _mm_srai_epi32(u2, DCT_CONST_BITS); + u3 = _mm_srai_epi32(u3, DCT_CONST_BITS); + u4 = _mm_srai_epi32(u4, DCT_CONST_BITS); + u5 = _mm_srai_epi32(u5, DCT_CONST_BITS); + u6 = _mm_srai_epi32(u6, DCT_CONST_BITS); + u7 = _mm_srai_epi32(u7, DCT_CONST_BITS); + + y0 = _mm_srai_epi32(y0, DCT_CONST_BITS); + y1 = _mm_srai_epi32(y1, DCT_CONST_BITS); + y2 = _mm_srai_epi32(y2, DCT_CONST_BITS); + y3 = _mm_srai_epi32(y3, DCT_CONST_BITS); + y4 = _mm_srai_epi32(y4, DCT_CONST_BITS); + y5 = _mm_srai_epi32(y5, DCT_CONST_BITS); + y6 = _mm_srai_epi32(y6, DCT_CONST_BITS); + y7 = _mm_srai_epi32(y7, DCT_CONST_BITS); + + in[0] = _mm_packs_epi32(v0, x0); + in[1] = _mm_packs_epi32(v1, x1); + in[2] = _mm_packs_epi32(v2, x2); + in[3] = _mm_packs_epi32(v3, x3); + in[4] = _mm_packs_epi32(v4, x4); + in[5] = _mm_packs_epi32(v5, x5); + in[6] = _mm_packs_epi32(v6, x6); + in[7] = _mm_packs_epi32(v7, x7); + + in[8] = _mm_packs_epi32(u0, y0); + in[9] = _mm_packs_epi32(u1, y1); + in[10] = _mm_packs_epi32(u2, y2); + in[11] = _mm_packs_epi32(u3, y3); + in[12] = _mm_packs_epi32(u4, y4); + in[13] = _mm_packs_epi32(u5, y5); + in[14] = _mm_packs_epi32(u6, y6); + in[15] = _mm_packs_epi32(u7, y7); +} +#endif // CONFIG_EXT_TX + +static INLINE void scale_sqrt2_8x4(__m128i *in) { + // Implements ROUND_POWER_OF_TWO(input * Sqrt2, DCT_CONST_BITS), for 32 + // consecutive elements. + const __m128i v_scale_w = _mm_set1_epi16((int16_t)Sqrt2); + + const __m128i v_p0l_w = _mm_mullo_epi16(in[0], v_scale_w); + const __m128i v_p0h_w = _mm_mulhi_epi16(in[0], v_scale_w); + const __m128i v_p1l_w = _mm_mullo_epi16(in[1], v_scale_w); + const __m128i v_p1h_w = _mm_mulhi_epi16(in[1], v_scale_w); + const __m128i v_p2l_w = _mm_mullo_epi16(in[2], v_scale_w); + const __m128i v_p2h_w = _mm_mulhi_epi16(in[2], v_scale_w); + const __m128i v_p3l_w = _mm_mullo_epi16(in[3], v_scale_w); + const __m128i v_p3h_w = _mm_mulhi_epi16(in[3], v_scale_w); + + const __m128i v_p0a_d = _mm_unpacklo_epi16(v_p0l_w, v_p0h_w); + const __m128i v_p0b_d = _mm_unpackhi_epi16(v_p0l_w, v_p0h_w); + const __m128i v_p1a_d = _mm_unpacklo_epi16(v_p1l_w, v_p1h_w); + const __m128i v_p1b_d = _mm_unpackhi_epi16(v_p1l_w, v_p1h_w); + const __m128i v_p2a_d = _mm_unpacklo_epi16(v_p2l_w, v_p2h_w); + const __m128i v_p2b_d = _mm_unpackhi_epi16(v_p2l_w, v_p2h_w); + const __m128i v_p3a_d = _mm_unpacklo_epi16(v_p3l_w, v_p3h_w); + const __m128i v_p3b_d = _mm_unpackhi_epi16(v_p3l_w, v_p3h_w); + + in[0] = _mm_packs_epi32(xx_roundn_epi32_unsigned(v_p0a_d, DCT_CONST_BITS), + xx_roundn_epi32_unsigned(v_p0b_d, DCT_CONST_BITS)); + in[1] = _mm_packs_epi32(xx_roundn_epi32_unsigned(v_p1a_d, DCT_CONST_BITS), + xx_roundn_epi32_unsigned(v_p1b_d, DCT_CONST_BITS)); + in[2] = _mm_packs_epi32(xx_roundn_epi32_unsigned(v_p2a_d, DCT_CONST_BITS), + xx_roundn_epi32_unsigned(v_p2b_d, DCT_CONST_BITS)); + in[3] = _mm_packs_epi32(xx_roundn_epi32_unsigned(v_p3a_d, DCT_CONST_BITS), + xx_roundn_epi32_unsigned(v_p3b_d, DCT_CONST_BITS)); +} + +static INLINE void scale_sqrt2_8x8(__m128i *in) { + // Implements 'ROUND_POWER_OF_TWO_SIGNED(input * Sqrt2, DCT_CONST_BITS)' + // for each element. + const __m128i v_scale_w = _mm_set1_epi16((int16_t)Sqrt2); + + const __m128i v_p0l_w = _mm_mullo_epi16(in[0], v_scale_w); + const __m128i v_p0h_w = _mm_mulhi_epi16(in[0], v_scale_w); + const __m128i v_p1l_w = _mm_mullo_epi16(in[1], v_scale_w); + const __m128i v_p1h_w = _mm_mulhi_epi16(in[1], v_scale_w); + const __m128i v_p2l_w = _mm_mullo_epi16(in[2], v_scale_w); + const __m128i v_p2h_w = _mm_mulhi_epi16(in[2], v_scale_w); + const __m128i v_p3l_w = _mm_mullo_epi16(in[3], v_scale_w); + const __m128i v_p3h_w = _mm_mulhi_epi16(in[3], v_scale_w); + const __m128i v_p4l_w = _mm_mullo_epi16(in[4], v_scale_w); + const __m128i v_p4h_w = _mm_mulhi_epi16(in[4], v_scale_w); + const __m128i v_p5l_w = _mm_mullo_epi16(in[5], v_scale_w); + const __m128i v_p5h_w = _mm_mulhi_epi16(in[5], v_scale_w); + const __m128i v_p6l_w = _mm_mullo_epi16(in[6], v_scale_w); + const __m128i v_p6h_w = _mm_mulhi_epi16(in[6], v_scale_w); + const __m128i v_p7l_w = _mm_mullo_epi16(in[7], v_scale_w); + const __m128i v_p7h_w = _mm_mulhi_epi16(in[7], v_scale_w); + + const __m128i v_p0a_d = _mm_unpacklo_epi16(v_p0l_w, v_p0h_w); + const __m128i v_p0b_d = _mm_unpackhi_epi16(v_p0l_w, v_p0h_w); + const __m128i v_p1a_d = _mm_unpacklo_epi16(v_p1l_w, v_p1h_w); + const __m128i v_p1b_d = _mm_unpackhi_epi16(v_p1l_w, v_p1h_w); + const __m128i v_p2a_d = _mm_unpacklo_epi16(v_p2l_w, v_p2h_w); + const __m128i v_p2b_d = _mm_unpackhi_epi16(v_p2l_w, v_p2h_w); + const __m128i v_p3a_d = _mm_unpacklo_epi16(v_p3l_w, v_p3h_w); + const __m128i v_p3b_d = _mm_unpackhi_epi16(v_p3l_w, v_p3h_w); + const __m128i v_p4a_d = _mm_unpacklo_epi16(v_p4l_w, v_p4h_w); + const __m128i v_p4b_d = _mm_unpackhi_epi16(v_p4l_w, v_p4h_w); + const __m128i v_p5a_d = _mm_unpacklo_epi16(v_p5l_w, v_p5h_w); + const __m128i v_p5b_d = _mm_unpackhi_epi16(v_p5l_w, v_p5h_w); + const __m128i v_p6a_d = _mm_unpacklo_epi16(v_p6l_w, v_p6h_w); + const __m128i v_p6b_d = _mm_unpackhi_epi16(v_p6l_w, v_p6h_w); + const __m128i v_p7a_d = _mm_unpacklo_epi16(v_p7l_w, v_p7h_w); + const __m128i v_p7b_d = _mm_unpackhi_epi16(v_p7l_w, v_p7h_w); + + in[0] = _mm_packs_epi32(xx_roundn_epi32_unsigned(v_p0a_d, DCT_CONST_BITS), + xx_roundn_epi32_unsigned(v_p0b_d, DCT_CONST_BITS)); + in[1] = _mm_packs_epi32(xx_roundn_epi32_unsigned(v_p1a_d, DCT_CONST_BITS), + xx_roundn_epi32_unsigned(v_p1b_d, DCT_CONST_BITS)); + in[2] = _mm_packs_epi32(xx_roundn_epi32_unsigned(v_p2a_d, DCT_CONST_BITS), + xx_roundn_epi32_unsigned(v_p2b_d, DCT_CONST_BITS)); + in[3] = _mm_packs_epi32(xx_roundn_epi32_unsigned(v_p3a_d, DCT_CONST_BITS), + xx_roundn_epi32_unsigned(v_p3b_d, DCT_CONST_BITS)); + in[4] = _mm_packs_epi32(xx_roundn_epi32_unsigned(v_p4a_d, DCT_CONST_BITS), + xx_roundn_epi32_unsigned(v_p4b_d, DCT_CONST_BITS)); + in[5] = _mm_packs_epi32(xx_roundn_epi32_unsigned(v_p5a_d, DCT_CONST_BITS), + xx_roundn_epi32_unsigned(v_p5b_d, DCT_CONST_BITS)); + in[6] = _mm_packs_epi32(xx_roundn_epi32_unsigned(v_p6a_d, DCT_CONST_BITS), + xx_roundn_epi32_unsigned(v_p6b_d, DCT_CONST_BITS)); + in[7] = _mm_packs_epi32(xx_roundn_epi32_unsigned(v_p7a_d, DCT_CONST_BITS), + xx_roundn_epi32_unsigned(v_p7b_d, DCT_CONST_BITS)); +} + +static INLINE void scale_sqrt2_8x16(__m128i *in) { + scale_sqrt2_8x8(in); + scale_sqrt2_8x8(in + 8); +} + +#endif // AOM_DSP_X86_TXFM_COMMON_SSE2_H_ diff --git a/third_party/aom/aom_dsp/x86/variance_avx2.c b/third_party/aom/aom_dsp/x86/variance_avx2.c new file mode 100644 index 000000000..18a70dffe --- /dev/null +++ b/third_party/aom/aom_dsp/x86/variance_avx2.c @@ -0,0 +1,192 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include <immintrin.h> +#include "./aom_dsp_rtcd.h" + +typedef void (*get_var_avx2)(const uint8_t *src, int src_stride, + const uint8_t *ref, int ref_stride, + unsigned int *sse, int *sum); + +void aom_get32x32var_avx2(const uint8_t *src, int src_stride, + const uint8_t *ref, int ref_stride, unsigned int *sse, + int *sum); + +static void variance_avx2(const uint8_t *src, int src_stride, + const uint8_t *ref, int ref_stride, int w, int h, + unsigned int *sse, int *sum, get_var_avx2 var_fn, + int block_size) { + int i, j; + + *sse = 0; + *sum = 0; + + for (i = 0; i < h; i += 16) { + for (j = 0; j < w; j += block_size) { + unsigned int sse0; + int sum0; + var_fn(&src[src_stride * i + j], src_stride, &ref[ref_stride * i + j], + ref_stride, &sse0, &sum0); + *sse += sse0; + *sum += sum0; + } + } +} + +unsigned int aom_variance16x16_avx2(const uint8_t *src, int src_stride, + const uint8_t *ref, int ref_stride, + unsigned int *sse) { + int sum; + unsigned int variance; + variance_avx2(src, src_stride, ref, ref_stride, 16, 16, sse, &sum, + aom_get16x16var_avx2, 16); + + variance = *sse - (((uint32_t)((int64_t)sum * sum)) >> 8); + _mm256_zeroupper(); + return variance; +} + +unsigned int aom_mse16x16_avx2(const uint8_t *src, int src_stride, + const uint8_t *ref, int ref_stride, + unsigned int *sse) { + int sum; + aom_get16x16var_avx2(src, src_stride, ref, ref_stride, sse, &sum); + _mm256_zeroupper(); + return *sse; +} + +unsigned int aom_variance32x16_avx2(const uint8_t *src, int src_stride, + const uint8_t *ref, int ref_stride, + unsigned int *sse) { + int sum; + unsigned int variance; + variance_avx2(src, src_stride, ref, ref_stride, 32, 16, sse, &sum, + aom_get32x32var_avx2, 32); + + variance = *sse - (uint32_t)(((int64_t)sum * sum) >> 9); + _mm256_zeroupper(); + return variance; +} + +unsigned int aom_variance32x32_avx2(const uint8_t *src, int src_stride, + const uint8_t *ref, int ref_stride, + unsigned int *sse) { + int sum; + unsigned int variance; + variance_avx2(src, src_stride, ref, ref_stride, 32, 32, sse, &sum, + aom_get32x32var_avx2, 32); + + variance = *sse - (uint32_t)(((int64_t)sum * sum) >> 10); + _mm256_zeroupper(); + return variance; +} + +unsigned int aom_variance64x64_avx2(const uint8_t *src, int src_stride, + const uint8_t *ref, int ref_stride, + unsigned int *sse) { + int sum; + unsigned int variance; + variance_avx2(src, src_stride, ref, ref_stride, 64, 64, sse, &sum, + aom_get32x32var_avx2, 32); + + variance = *sse - (uint32_t)(((int64_t)sum * sum) >> 12); + _mm256_zeroupper(); + return variance; +} + +unsigned int aom_variance64x32_avx2(const uint8_t *src, int src_stride, + const uint8_t *ref, int ref_stride, + unsigned int *sse) { + int sum; + unsigned int variance; + variance_avx2(src, src_stride, ref, ref_stride, 64, 32, sse, &sum, + aom_get32x32var_avx2, 32); + + variance = *sse - (uint32_t)(((int64_t)sum * sum) >> 11); + _mm256_zeroupper(); + return variance; +} + +unsigned int aom_sub_pixel_variance32xh_avx2(const uint8_t *src, int src_stride, + int x_offset, int y_offset, + const uint8_t *dst, int dst_stride, + int height, unsigned int *sse); + +unsigned int aom_sub_pixel_avg_variance32xh_avx2( + const uint8_t *src, int src_stride, int x_offset, int y_offset, + const uint8_t *dst, int dst_stride, const uint8_t *sec, int sec_stride, + int height, unsigned int *sseptr); + +unsigned int aom_sub_pixel_variance64x64_avx2(const uint8_t *src, + int src_stride, int x_offset, + int y_offset, const uint8_t *dst, + int dst_stride, + unsigned int *sse) { + unsigned int sse1; + const int se1 = aom_sub_pixel_variance32xh_avx2( + src, src_stride, x_offset, y_offset, dst, dst_stride, 64, &sse1); + unsigned int sse2; + const int se2 = + aom_sub_pixel_variance32xh_avx2(src + 32, src_stride, x_offset, y_offset, + dst + 32, dst_stride, 64, &sse2); + const int se = se1 + se2; + unsigned int variance; + *sse = sse1 + sse2; + + variance = *sse - (uint32_t)(((int64_t)se * se) >> 12); + _mm256_zeroupper(); + return variance; +} + +unsigned int aom_sub_pixel_variance32x32_avx2(const uint8_t *src, + int src_stride, int x_offset, + int y_offset, const uint8_t *dst, + int dst_stride, + unsigned int *sse) { + const int se = aom_sub_pixel_variance32xh_avx2( + src, src_stride, x_offset, y_offset, dst, dst_stride, 32, sse); + + const unsigned int variance = *sse - (uint32_t)(((int64_t)se * se) >> 10); + _mm256_zeroupper(); + return variance; +} + +unsigned int aom_sub_pixel_avg_variance64x64_avx2( + const uint8_t *src, int src_stride, int x_offset, int y_offset, + const uint8_t *dst, int dst_stride, unsigned int *sse, const uint8_t *sec) { + unsigned int sse1; + const int se1 = aom_sub_pixel_avg_variance32xh_avx2( + src, src_stride, x_offset, y_offset, dst, dst_stride, sec, 64, 64, &sse1); + unsigned int sse2; + const int se2 = aom_sub_pixel_avg_variance32xh_avx2( + src + 32, src_stride, x_offset, y_offset, dst + 32, dst_stride, sec + 32, + 64, 64, &sse2); + const int se = se1 + se2; + unsigned int variance; + + *sse = sse1 + sse2; + + variance = *sse - (uint32_t)(((int64_t)se * se) >> 12); + _mm256_zeroupper(); + return variance; +} + +unsigned int aom_sub_pixel_avg_variance32x32_avx2( + const uint8_t *src, int src_stride, int x_offset, int y_offset, + const uint8_t *dst, int dst_stride, unsigned int *sse, const uint8_t *sec) { + // Process 32 elements in parallel. + const int se = aom_sub_pixel_avg_variance32xh_avx2( + src, src_stride, x_offset, y_offset, dst, dst_stride, sec, 32, 32, sse); + + const unsigned int variance = *sse - (uint32_t)(((int64_t)se * se) >> 10); + _mm256_zeroupper(); + return variance; +} diff --git a/third_party/aom/aom_dsp/x86/variance_impl_avx2.c b/third_party/aom/aom_dsp/x86/variance_impl_avx2.c new file mode 100644 index 000000000..999b541e3 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/variance_impl_avx2.c @@ -0,0 +1,713 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include <immintrin.h> // AVX2 + +#include "./aom_dsp_rtcd.h" +#include "aom_ports/mem.h" + +/* clang-format off */ +DECLARE_ALIGNED(32, static const uint8_t, bilinear_filters_avx2[512]) = { + 16, 0, 16, 0, 16, 0, 16, 0, 16, 0, 16, 0, 16, 0, 16, 0, + 16, 0, 16, 0, 16, 0, 16, 0, 16, 0, 16, 0, 16, 0, 16, 0, + 14, 2, 14, 2, 14, 2, 14, 2, 14, 2, 14, 2, 14, 2, 14, 2, + 14, 2, 14, 2, 14, 2, 14, 2, 14, 2, 14, 2, 14, 2, 14, 2, + 12, 4, 12, 4, 12, 4, 12, 4, 12, 4, 12, 4, 12, 4, 12, 4, + 12, 4, 12, 4, 12, 4, 12, 4, 12, 4, 12, 4, 12, 4, 12, 4, + 10, 6, 10, 6, 10, 6, 10, 6, 10, 6, 10, 6, 10, 6, 10, 6, + 10, 6, 10, 6, 10, 6, 10, 6, 10, 6, 10, 6, 10, 6, 10, 6, + 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, + 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, + 6, 10, 6, 10, 6, 10, 6, 10, 6, 10, 6, 10, 6, 10, 6, 10, + 6, 10, 6, 10, 6, 10, 6, 10, 6, 10, 6, 10, 6, 10, 6, 10, + 4, 12, 4, 12, 4, 12, 4, 12, 4, 12, 4, 12, 4, 12, 4, 12, + 4, 12, 4, 12, 4, 12, 4, 12, 4, 12, 4, 12, 4, 12, 4, 12, + 2, 14, 2, 14, 2, 14, 2, 14, 2, 14, 2, 14, 2, 14, 2, 14, + 2, 14, 2, 14, 2, 14, 2, 14, 2, 14, 2, 14, 2, 14, 2, 14, +}; +/* clang-format on */ + +void aom_get16x16var_avx2(const unsigned char *src_ptr, int source_stride, + const unsigned char *ref_ptr, int recon_stride, + unsigned int *SSE, int *Sum) { + __m256i src, src_expand_low, src_expand_high, ref, ref_expand_low; + __m256i ref_expand_high, madd_low, madd_high; + unsigned int i, src_2strides, ref_2strides; + __m256i zero_reg = _mm256_set1_epi16(0); + __m256i sum_ref_src = _mm256_set1_epi16(0); + __m256i madd_ref_src = _mm256_set1_epi16(0); + + // processing two strides in a 256 bit register reducing the number + // of loop stride by half (comparing to the sse2 code) + src_2strides = source_stride << 1; + ref_2strides = recon_stride << 1; + for (i = 0; i < 8; i++) { + src = _mm256_castsi128_si256(_mm_loadu_si128((__m128i const *)(src_ptr))); + src = _mm256_inserti128_si256( + src, _mm_loadu_si128((__m128i const *)(src_ptr + source_stride)), 1); + + ref = _mm256_castsi128_si256(_mm_loadu_si128((__m128i const *)(ref_ptr))); + ref = _mm256_inserti128_si256( + ref, _mm_loadu_si128((__m128i const *)(ref_ptr + recon_stride)), 1); + + // expanding to 16 bit each lane + src_expand_low = _mm256_unpacklo_epi8(src, zero_reg); + src_expand_high = _mm256_unpackhi_epi8(src, zero_reg); + + ref_expand_low = _mm256_unpacklo_epi8(ref, zero_reg); + ref_expand_high = _mm256_unpackhi_epi8(ref, zero_reg); + + // src-ref + src_expand_low = _mm256_sub_epi16(src_expand_low, ref_expand_low); + src_expand_high = _mm256_sub_epi16(src_expand_high, ref_expand_high); + + // madd low (src - ref) + madd_low = _mm256_madd_epi16(src_expand_low, src_expand_low); + + // add high to low + src_expand_low = _mm256_add_epi16(src_expand_low, src_expand_high); + + // madd high (src - ref) + madd_high = _mm256_madd_epi16(src_expand_high, src_expand_high); + + sum_ref_src = _mm256_add_epi16(sum_ref_src, src_expand_low); + + // add high to low + madd_ref_src = + _mm256_add_epi32(madd_ref_src, _mm256_add_epi32(madd_low, madd_high)); + + src_ptr += src_2strides; + ref_ptr += ref_2strides; + } + + { + __m128i sum_res, madd_res; + __m128i expand_sum_low, expand_sum_high, expand_sum; + __m128i expand_madd_low, expand_madd_high, expand_madd; + __m128i ex_expand_sum_low, ex_expand_sum_high, ex_expand_sum; + + // extract the low lane and add it to the high lane + sum_res = _mm_add_epi16(_mm256_castsi256_si128(sum_ref_src), + _mm256_extractf128_si256(sum_ref_src, 1)); + + madd_res = _mm_add_epi32(_mm256_castsi256_si128(madd_ref_src), + _mm256_extractf128_si256(madd_ref_src, 1)); + + // padding each 2 bytes with another 2 zeroed bytes + expand_sum_low = + _mm_unpacklo_epi16(_mm256_castsi256_si128(zero_reg), sum_res); + expand_sum_high = + _mm_unpackhi_epi16(_mm256_castsi256_si128(zero_reg), sum_res); + + // shifting the sign 16 bits right + expand_sum_low = _mm_srai_epi32(expand_sum_low, 16); + expand_sum_high = _mm_srai_epi32(expand_sum_high, 16); + + expand_sum = _mm_add_epi32(expand_sum_low, expand_sum_high); + + // expand each 32 bits of the madd result to 64 bits + expand_madd_low = + _mm_unpacklo_epi32(madd_res, _mm256_castsi256_si128(zero_reg)); + expand_madd_high = + _mm_unpackhi_epi32(madd_res, _mm256_castsi256_si128(zero_reg)); + + expand_madd = _mm_add_epi32(expand_madd_low, expand_madd_high); + + ex_expand_sum_low = + _mm_unpacklo_epi32(expand_sum, _mm256_castsi256_si128(zero_reg)); + ex_expand_sum_high = + _mm_unpackhi_epi32(expand_sum, _mm256_castsi256_si128(zero_reg)); + + ex_expand_sum = _mm_add_epi32(ex_expand_sum_low, ex_expand_sum_high); + + // shift 8 bytes eight + madd_res = _mm_srli_si128(expand_madd, 8); + sum_res = _mm_srli_si128(ex_expand_sum, 8); + + madd_res = _mm_add_epi32(madd_res, expand_madd); + sum_res = _mm_add_epi32(sum_res, ex_expand_sum); + + *((int *)SSE) = _mm_cvtsi128_si32(madd_res); + + *((int *)Sum) = _mm_cvtsi128_si32(sum_res); + } + _mm256_zeroupper(); +} + +void aom_get32x32var_avx2(const unsigned char *src_ptr, int source_stride, + const unsigned char *ref_ptr, int recon_stride, + unsigned int *SSE, int *Sum) { + __m256i src, src_expand_low, src_expand_high, ref, ref_expand_low; + __m256i ref_expand_high, madd_low, madd_high; + unsigned int i; + __m256i zero_reg = _mm256_set1_epi16(0); + __m256i sum_ref_src = _mm256_set1_epi16(0); + __m256i madd_ref_src = _mm256_set1_epi16(0); + + // processing 32 elements in parallel + for (i = 0; i < 16; i++) { + src = _mm256_loadu_si256((__m256i const *)(src_ptr)); + + ref = _mm256_loadu_si256((__m256i const *)(ref_ptr)); + + // expanding to 16 bit each lane + src_expand_low = _mm256_unpacklo_epi8(src, zero_reg); + src_expand_high = _mm256_unpackhi_epi8(src, zero_reg); + + ref_expand_low = _mm256_unpacklo_epi8(ref, zero_reg); + ref_expand_high = _mm256_unpackhi_epi8(ref, zero_reg); + + // src-ref + src_expand_low = _mm256_sub_epi16(src_expand_low, ref_expand_low); + src_expand_high = _mm256_sub_epi16(src_expand_high, ref_expand_high); + + // madd low (src - ref) + madd_low = _mm256_madd_epi16(src_expand_low, src_expand_low); + + // add high to low + src_expand_low = _mm256_add_epi16(src_expand_low, src_expand_high); + + // madd high (src - ref) + madd_high = _mm256_madd_epi16(src_expand_high, src_expand_high); + + sum_ref_src = _mm256_add_epi16(sum_ref_src, src_expand_low); + + // add high to low + madd_ref_src = + _mm256_add_epi32(madd_ref_src, _mm256_add_epi32(madd_low, madd_high)); + + src_ptr += source_stride; + ref_ptr += recon_stride; + } + + { + __m256i expand_sum_low, expand_sum_high, expand_sum; + __m256i expand_madd_low, expand_madd_high, expand_madd; + __m256i ex_expand_sum_low, ex_expand_sum_high, ex_expand_sum; + + // padding each 2 bytes with another 2 zeroed bytes + expand_sum_low = _mm256_unpacklo_epi16(zero_reg, sum_ref_src); + expand_sum_high = _mm256_unpackhi_epi16(zero_reg, sum_ref_src); + + // shifting the sign 16 bits right + expand_sum_low = _mm256_srai_epi32(expand_sum_low, 16); + expand_sum_high = _mm256_srai_epi32(expand_sum_high, 16); + + expand_sum = _mm256_add_epi32(expand_sum_low, expand_sum_high); + + // expand each 32 bits of the madd result to 64 bits + expand_madd_low = _mm256_unpacklo_epi32(madd_ref_src, zero_reg); + expand_madd_high = _mm256_unpackhi_epi32(madd_ref_src, zero_reg); + + expand_madd = _mm256_add_epi32(expand_madd_low, expand_madd_high); + + ex_expand_sum_low = _mm256_unpacklo_epi32(expand_sum, zero_reg); + ex_expand_sum_high = _mm256_unpackhi_epi32(expand_sum, zero_reg); + + ex_expand_sum = _mm256_add_epi32(ex_expand_sum_low, ex_expand_sum_high); + + // shift 8 bytes eight + madd_ref_src = _mm256_srli_si256(expand_madd, 8); + sum_ref_src = _mm256_srli_si256(ex_expand_sum, 8); + + madd_ref_src = _mm256_add_epi32(madd_ref_src, expand_madd); + sum_ref_src = _mm256_add_epi32(sum_ref_src, ex_expand_sum); + + // extract the low lane and the high lane and add the results + *((int *)SSE) = + _mm_cvtsi128_si32(_mm256_castsi256_si128(madd_ref_src)) + + _mm_cvtsi128_si32(_mm256_extractf128_si256(madd_ref_src, 1)); + + *((int *)Sum) = _mm_cvtsi128_si32(_mm256_castsi256_si128(sum_ref_src)) + + _mm_cvtsi128_si32(_mm256_extractf128_si256(sum_ref_src, 1)); + } + _mm256_zeroupper(); +} + +#define FILTER_SRC(filter) \ + /* filter the source */ \ + exp_src_lo = _mm256_maddubs_epi16(exp_src_lo, filter); \ + exp_src_hi = _mm256_maddubs_epi16(exp_src_hi, filter); \ + \ + /* add 8 to source */ \ + exp_src_lo = _mm256_add_epi16(exp_src_lo, pw8); \ + exp_src_hi = _mm256_add_epi16(exp_src_hi, pw8); \ + \ + /* divide source by 16 */ \ + exp_src_lo = _mm256_srai_epi16(exp_src_lo, 4); \ + exp_src_hi = _mm256_srai_epi16(exp_src_hi, 4); + +#define MERGE_WITH_SRC(src_reg, reg) \ + exp_src_lo = _mm256_unpacklo_epi8(src_reg, reg); \ + exp_src_hi = _mm256_unpackhi_epi8(src_reg, reg); + +#define LOAD_SRC_DST \ + /* load source and destination */ \ + src_reg = _mm256_loadu_si256((__m256i const *)(src)); \ + dst_reg = _mm256_loadu_si256((__m256i const *)(dst)); + +#define AVG_NEXT_SRC(src_reg, size_stride) \ + src_next_reg = _mm256_loadu_si256((__m256i const *)(src + size_stride)); \ + /* average between current and next stride source */ \ + src_reg = _mm256_avg_epu8(src_reg, src_next_reg); + +#define MERGE_NEXT_SRC(src_reg, size_stride) \ + src_next_reg = _mm256_loadu_si256((__m256i const *)(src + size_stride)); \ + MERGE_WITH_SRC(src_reg, src_next_reg) + +#define CALC_SUM_SSE_INSIDE_LOOP \ + /* expand each byte to 2 bytes */ \ + exp_dst_lo = _mm256_unpacklo_epi8(dst_reg, zero_reg); \ + exp_dst_hi = _mm256_unpackhi_epi8(dst_reg, zero_reg); \ + /* source - dest */ \ + exp_src_lo = _mm256_sub_epi16(exp_src_lo, exp_dst_lo); \ + exp_src_hi = _mm256_sub_epi16(exp_src_hi, exp_dst_hi); \ + /* caculate sum */ \ + sum_reg = _mm256_add_epi16(sum_reg, exp_src_lo); \ + exp_src_lo = _mm256_madd_epi16(exp_src_lo, exp_src_lo); \ + sum_reg = _mm256_add_epi16(sum_reg, exp_src_hi); \ + exp_src_hi = _mm256_madd_epi16(exp_src_hi, exp_src_hi); \ + /* calculate sse */ \ + sse_reg = _mm256_add_epi32(sse_reg, exp_src_lo); \ + sse_reg = _mm256_add_epi32(sse_reg, exp_src_hi); + +// final calculation to sum and sse +#define CALC_SUM_AND_SSE \ + res_cmp = _mm256_cmpgt_epi16(zero_reg, sum_reg); \ + sse_reg_hi = _mm256_srli_si256(sse_reg, 8); \ + sum_reg_lo = _mm256_unpacklo_epi16(sum_reg, res_cmp); \ + sum_reg_hi = _mm256_unpackhi_epi16(sum_reg, res_cmp); \ + sse_reg = _mm256_add_epi32(sse_reg, sse_reg_hi); \ + sum_reg = _mm256_add_epi32(sum_reg_lo, sum_reg_hi); \ + \ + sse_reg_hi = _mm256_srli_si256(sse_reg, 4); \ + sum_reg_hi = _mm256_srli_si256(sum_reg, 8); \ + \ + sse_reg = _mm256_add_epi32(sse_reg, sse_reg_hi); \ + sum_reg = _mm256_add_epi32(sum_reg, sum_reg_hi); \ + *((int *)sse) = _mm_cvtsi128_si32(_mm256_castsi256_si128(sse_reg)) + \ + _mm_cvtsi128_si32(_mm256_extractf128_si256(sse_reg, 1)); \ + sum_reg_hi = _mm256_srli_si256(sum_reg, 4); \ + sum_reg = _mm256_add_epi32(sum_reg, sum_reg_hi); \ + sum = _mm_cvtsi128_si32(_mm256_castsi256_si128(sum_reg)) + \ + _mm_cvtsi128_si32(_mm256_extractf128_si256(sum_reg, 1)); + +unsigned int aom_sub_pixel_variance32xh_avx2(const uint8_t *src, int src_stride, + int x_offset, int y_offset, + const uint8_t *dst, int dst_stride, + int height, unsigned int *sse) { + __m256i src_reg, dst_reg, exp_src_lo, exp_src_hi, exp_dst_lo, exp_dst_hi; + __m256i sse_reg, sum_reg, sse_reg_hi, res_cmp, sum_reg_lo, sum_reg_hi; + __m256i zero_reg; + int i, sum; + sum_reg = _mm256_set1_epi16(0); + sse_reg = _mm256_set1_epi16(0); + zero_reg = _mm256_set1_epi16(0); + + // x_offset = 0 and y_offset = 0 + if (x_offset == 0) { + if (y_offset == 0) { + for (i = 0; i < height; i++) { + LOAD_SRC_DST + // expend each byte to 2 bytes + MERGE_WITH_SRC(src_reg, zero_reg) + CALC_SUM_SSE_INSIDE_LOOP + src += src_stride; + dst += dst_stride; + } + // x_offset = 0 and y_offset = 8 + } else if (y_offset == 8) { + __m256i src_next_reg; + for (i = 0; i < height; i++) { + LOAD_SRC_DST + AVG_NEXT_SRC(src_reg, src_stride) + // expend each byte to 2 bytes + MERGE_WITH_SRC(src_reg, zero_reg) + CALC_SUM_SSE_INSIDE_LOOP + src += src_stride; + dst += dst_stride; + } + // x_offset = 0 and y_offset = bilin interpolation + } else { + __m256i filter, pw8, src_next_reg; + + y_offset <<= 5; + filter = _mm256_load_si256( + (__m256i const *)(bilinear_filters_avx2 + y_offset)); + pw8 = _mm256_set1_epi16(8); + for (i = 0; i < height; i++) { + LOAD_SRC_DST + MERGE_NEXT_SRC(src_reg, src_stride) + FILTER_SRC(filter) + CALC_SUM_SSE_INSIDE_LOOP + src += src_stride; + dst += dst_stride; + } + } + // x_offset = 8 and y_offset = 0 + } else if (x_offset == 8) { + if (y_offset == 0) { + __m256i src_next_reg; + for (i = 0; i < height; i++) { + LOAD_SRC_DST + AVG_NEXT_SRC(src_reg, 1) + // expand each byte to 2 bytes + MERGE_WITH_SRC(src_reg, zero_reg) + CALC_SUM_SSE_INSIDE_LOOP + src += src_stride; + dst += dst_stride; + } + // x_offset = 8 and y_offset = 8 + } else if (y_offset == 8) { + __m256i src_next_reg, src_avg; + // load source and another source starting from the next + // following byte + src_reg = _mm256_loadu_si256((__m256i const *)(src)); + AVG_NEXT_SRC(src_reg, 1) + for (i = 0; i < height; i++) { + src_avg = src_reg; + src += src_stride; + LOAD_SRC_DST + AVG_NEXT_SRC(src_reg, 1) + // average between previous average to current average + src_avg = _mm256_avg_epu8(src_avg, src_reg); + // expand each byte to 2 bytes + MERGE_WITH_SRC(src_avg, zero_reg) + // save current source average + CALC_SUM_SSE_INSIDE_LOOP + dst += dst_stride; + } + // x_offset = 8 and y_offset = bilin interpolation + } else { + __m256i filter, pw8, src_next_reg, src_avg; + y_offset <<= 5; + filter = _mm256_load_si256( + (__m256i const *)(bilinear_filters_avx2 + y_offset)); + pw8 = _mm256_set1_epi16(8); + // load source and another source starting from the next + // following byte + src_reg = _mm256_loadu_si256((__m256i const *)(src)); + AVG_NEXT_SRC(src_reg, 1) + for (i = 0; i < height; i++) { + // save current source average + src_avg = src_reg; + src += src_stride; + LOAD_SRC_DST + AVG_NEXT_SRC(src_reg, 1) + MERGE_WITH_SRC(src_avg, src_reg) + FILTER_SRC(filter) + CALC_SUM_SSE_INSIDE_LOOP + dst += dst_stride; + } + } + // x_offset = bilin interpolation and y_offset = 0 + } else { + if (y_offset == 0) { + __m256i filter, pw8, src_next_reg; + x_offset <<= 5; + filter = _mm256_load_si256( + (__m256i const *)(bilinear_filters_avx2 + x_offset)); + pw8 = _mm256_set1_epi16(8); + for (i = 0; i < height; i++) { + LOAD_SRC_DST + MERGE_NEXT_SRC(src_reg, 1) + FILTER_SRC(filter) + CALC_SUM_SSE_INSIDE_LOOP + src += src_stride; + dst += dst_stride; + } + // x_offset = bilin interpolation and y_offset = 8 + } else if (y_offset == 8) { + __m256i filter, pw8, src_next_reg, src_pack; + x_offset <<= 5; + filter = _mm256_load_si256( + (__m256i const *)(bilinear_filters_avx2 + x_offset)); + pw8 = _mm256_set1_epi16(8); + src_reg = _mm256_loadu_si256((__m256i const *)(src)); + MERGE_NEXT_SRC(src_reg, 1) + FILTER_SRC(filter) + // convert each 16 bit to 8 bit to each low and high lane source + src_pack = _mm256_packus_epi16(exp_src_lo, exp_src_hi); + for (i = 0; i < height; i++) { + src += src_stride; + LOAD_SRC_DST + MERGE_NEXT_SRC(src_reg, 1) + FILTER_SRC(filter) + src_reg = _mm256_packus_epi16(exp_src_lo, exp_src_hi); + // average between previous pack to the current + src_pack = _mm256_avg_epu8(src_pack, src_reg); + MERGE_WITH_SRC(src_pack, zero_reg) + CALC_SUM_SSE_INSIDE_LOOP + src_pack = src_reg; + dst += dst_stride; + } + // x_offset = bilin interpolation and y_offset = bilin interpolation + } else { + __m256i xfilter, yfilter, pw8, src_next_reg, src_pack; + x_offset <<= 5; + xfilter = _mm256_load_si256( + (__m256i const *)(bilinear_filters_avx2 + x_offset)); + y_offset <<= 5; + yfilter = _mm256_load_si256( + (__m256i const *)(bilinear_filters_avx2 + y_offset)); + pw8 = _mm256_set1_epi16(8); + // load source and another source starting from the next + // following byte + src_reg = _mm256_loadu_si256((__m256i const *)(src)); + MERGE_NEXT_SRC(src_reg, 1) + + FILTER_SRC(xfilter) + // convert each 16 bit to 8 bit to each low and high lane source + src_pack = _mm256_packus_epi16(exp_src_lo, exp_src_hi); + for (i = 0; i < height; i++) { + src += src_stride; + LOAD_SRC_DST + MERGE_NEXT_SRC(src_reg, 1) + FILTER_SRC(xfilter) + src_reg = _mm256_packus_epi16(exp_src_lo, exp_src_hi); + // merge previous pack to current pack source + MERGE_WITH_SRC(src_pack, src_reg) + // filter the source + FILTER_SRC(yfilter) + src_pack = src_reg; + CALC_SUM_SSE_INSIDE_LOOP + dst += dst_stride; + } + } + } + CALC_SUM_AND_SSE + _mm256_zeroupper(); + return sum; +} + +unsigned int aom_sub_pixel_avg_variance32xh_avx2( + const uint8_t *src, int src_stride, int x_offset, int y_offset, + const uint8_t *dst, int dst_stride, const uint8_t *sec, int sec_stride, + int height, unsigned int *sse) { + __m256i sec_reg; + __m256i src_reg, dst_reg, exp_src_lo, exp_src_hi, exp_dst_lo, exp_dst_hi; + __m256i sse_reg, sum_reg, sse_reg_hi, res_cmp, sum_reg_lo, sum_reg_hi; + __m256i zero_reg; + int i, sum; + sum_reg = _mm256_set1_epi16(0); + sse_reg = _mm256_set1_epi16(0); + zero_reg = _mm256_set1_epi16(0); + + // x_offset = 0 and y_offset = 0 + if (x_offset == 0) { + if (y_offset == 0) { + for (i = 0; i < height; i++) { + LOAD_SRC_DST + sec_reg = _mm256_loadu_si256((__m256i const *)(sec)); + src_reg = _mm256_avg_epu8(src_reg, sec_reg); + sec += sec_stride; + // expend each byte to 2 bytes + MERGE_WITH_SRC(src_reg, zero_reg) + CALC_SUM_SSE_INSIDE_LOOP + src += src_stride; + dst += dst_stride; + } + } else if (y_offset == 8) { + __m256i src_next_reg; + for (i = 0; i < height; i++) { + LOAD_SRC_DST + AVG_NEXT_SRC(src_reg, src_stride) + sec_reg = _mm256_loadu_si256((__m256i const *)(sec)); + src_reg = _mm256_avg_epu8(src_reg, sec_reg); + sec += sec_stride; + // expend each byte to 2 bytes + MERGE_WITH_SRC(src_reg, zero_reg) + CALC_SUM_SSE_INSIDE_LOOP + src += src_stride; + dst += dst_stride; + } + // x_offset = 0 and y_offset = bilin interpolation + } else { + __m256i filter, pw8, src_next_reg; + + y_offset <<= 5; + filter = _mm256_load_si256( + (__m256i const *)(bilinear_filters_avx2 + y_offset)); + pw8 = _mm256_set1_epi16(8); + for (i = 0; i < height; i++) { + LOAD_SRC_DST + MERGE_NEXT_SRC(src_reg, src_stride) + FILTER_SRC(filter) + src_reg = _mm256_packus_epi16(exp_src_lo, exp_src_hi); + sec_reg = _mm256_loadu_si256((__m256i const *)(sec)); + src_reg = _mm256_avg_epu8(src_reg, sec_reg); + sec += sec_stride; + MERGE_WITH_SRC(src_reg, zero_reg) + CALC_SUM_SSE_INSIDE_LOOP + src += src_stride; + dst += dst_stride; + } + } + // x_offset = 8 and y_offset = 0 + } else if (x_offset == 8) { + if (y_offset == 0) { + __m256i src_next_reg; + for (i = 0; i < height; i++) { + LOAD_SRC_DST + AVG_NEXT_SRC(src_reg, 1) + sec_reg = _mm256_loadu_si256((__m256i const *)(sec)); + src_reg = _mm256_avg_epu8(src_reg, sec_reg); + sec += sec_stride; + // expand each byte to 2 bytes + MERGE_WITH_SRC(src_reg, zero_reg) + CALC_SUM_SSE_INSIDE_LOOP + src += src_stride; + dst += dst_stride; + } + // x_offset = 8 and y_offset = 8 + } else if (y_offset == 8) { + __m256i src_next_reg, src_avg; + // load source and another source starting from the next + // following byte + src_reg = _mm256_loadu_si256((__m256i const *)(src)); + AVG_NEXT_SRC(src_reg, 1) + for (i = 0; i < height; i++) { + // save current source average + src_avg = src_reg; + src += src_stride; + LOAD_SRC_DST + AVG_NEXT_SRC(src_reg, 1) + // average between previous average to current average + src_avg = _mm256_avg_epu8(src_avg, src_reg); + sec_reg = _mm256_loadu_si256((__m256i const *)(sec)); + src_avg = _mm256_avg_epu8(src_avg, sec_reg); + sec += sec_stride; + // expand each byte to 2 bytes + MERGE_WITH_SRC(src_avg, zero_reg) + CALC_SUM_SSE_INSIDE_LOOP + dst += dst_stride; + } + // x_offset = 8 and y_offset = bilin interpolation + } else { + __m256i filter, pw8, src_next_reg, src_avg; + y_offset <<= 5; + filter = _mm256_load_si256( + (__m256i const *)(bilinear_filters_avx2 + y_offset)); + pw8 = _mm256_set1_epi16(8); + // load source and another source starting from the next + // following byte + src_reg = _mm256_loadu_si256((__m256i const *)(src)); + AVG_NEXT_SRC(src_reg, 1) + for (i = 0; i < height; i++) { + // save current source average + src_avg = src_reg; + src += src_stride; + LOAD_SRC_DST + AVG_NEXT_SRC(src_reg, 1) + MERGE_WITH_SRC(src_avg, src_reg) + FILTER_SRC(filter) + src_avg = _mm256_packus_epi16(exp_src_lo, exp_src_hi); + sec_reg = _mm256_loadu_si256((__m256i const *)(sec)); + src_avg = _mm256_avg_epu8(src_avg, sec_reg); + // expand each byte to 2 bytes + MERGE_WITH_SRC(src_avg, zero_reg) + sec += sec_stride; + CALC_SUM_SSE_INSIDE_LOOP + dst += dst_stride; + } + } + // x_offset = bilin interpolation and y_offset = 0 + } else { + if (y_offset == 0) { + __m256i filter, pw8, src_next_reg; + x_offset <<= 5; + filter = _mm256_load_si256( + (__m256i const *)(bilinear_filters_avx2 + x_offset)); + pw8 = _mm256_set1_epi16(8); + for (i = 0; i < height; i++) { + LOAD_SRC_DST + MERGE_NEXT_SRC(src_reg, 1) + FILTER_SRC(filter) + src_reg = _mm256_packus_epi16(exp_src_lo, exp_src_hi); + sec_reg = _mm256_loadu_si256((__m256i const *)(sec)); + src_reg = _mm256_avg_epu8(src_reg, sec_reg); + MERGE_WITH_SRC(src_reg, zero_reg) + sec += sec_stride; + CALC_SUM_SSE_INSIDE_LOOP + src += src_stride; + dst += dst_stride; + } + // x_offset = bilin interpolation and y_offset = 8 + } else if (y_offset == 8) { + __m256i filter, pw8, src_next_reg, src_pack; + x_offset <<= 5; + filter = _mm256_load_si256( + (__m256i const *)(bilinear_filters_avx2 + x_offset)); + pw8 = _mm256_set1_epi16(8); + src_reg = _mm256_loadu_si256((__m256i const *)(src)); + MERGE_NEXT_SRC(src_reg, 1) + FILTER_SRC(filter) + // convert each 16 bit to 8 bit to each low and high lane source + src_pack = _mm256_packus_epi16(exp_src_lo, exp_src_hi); + for (i = 0; i < height; i++) { + src += src_stride; + LOAD_SRC_DST + MERGE_NEXT_SRC(src_reg, 1) + FILTER_SRC(filter) + src_reg = _mm256_packus_epi16(exp_src_lo, exp_src_hi); + // average between previous pack to the current + src_pack = _mm256_avg_epu8(src_pack, src_reg); + sec_reg = _mm256_loadu_si256((__m256i const *)(sec)); + src_pack = _mm256_avg_epu8(src_pack, sec_reg); + sec += sec_stride; + MERGE_WITH_SRC(src_pack, zero_reg) + src_pack = src_reg; + CALC_SUM_SSE_INSIDE_LOOP + dst += dst_stride; + } + // x_offset = bilin interpolation and y_offset = bilin interpolation + } else { + __m256i xfilter, yfilter, pw8, src_next_reg, src_pack; + x_offset <<= 5; + xfilter = _mm256_load_si256( + (__m256i const *)(bilinear_filters_avx2 + x_offset)); + y_offset <<= 5; + yfilter = _mm256_load_si256( + (__m256i const *)(bilinear_filters_avx2 + y_offset)); + pw8 = _mm256_set1_epi16(8); + // load source and another source starting from the next + // following byte + src_reg = _mm256_loadu_si256((__m256i const *)(src)); + MERGE_NEXT_SRC(src_reg, 1) + + FILTER_SRC(xfilter) + // convert each 16 bit to 8 bit to each low and high lane source + src_pack = _mm256_packus_epi16(exp_src_lo, exp_src_hi); + for (i = 0; i < height; i++) { + src += src_stride; + LOAD_SRC_DST + MERGE_NEXT_SRC(src_reg, 1) + FILTER_SRC(xfilter) + src_reg = _mm256_packus_epi16(exp_src_lo, exp_src_hi); + // merge previous pack to current pack source + MERGE_WITH_SRC(src_pack, src_reg) + // filter the source + FILTER_SRC(yfilter) + src_pack = _mm256_packus_epi16(exp_src_lo, exp_src_hi); + sec_reg = _mm256_loadu_si256((__m256i const *)(sec)); + src_pack = _mm256_avg_epu8(src_pack, sec_reg); + MERGE_WITH_SRC(src_pack, zero_reg) + src_pack = src_reg; + sec += sec_stride; + CALC_SUM_SSE_INSIDE_LOOP + dst += dst_stride; + } + } + } + CALC_SUM_AND_SSE + _mm256_zeroupper(); + return sum; +} diff --git a/third_party/aom/aom_dsp/x86/variance_sse2.c b/third_party/aom/aom_dsp/x86/variance_sse2.c new file mode 100644 index 000000000..d9563aa7f --- /dev/null +++ b/third_party/aom/aom_dsp/x86/variance_sse2.c @@ -0,0 +1,690 @@ +/* + * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include <assert.h> +#include <emmintrin.h> // SSE2 + +#include "./aom_config.h" +#include "./aom_dsp_rtcd.h" + +#include "aom_ports/mem.h" + +typedef void (*getNxMvar_fn_t)(const unsigned char *src, int src_stride, + const unsigned char *ref, int ref_stride, + unsigned int *sse, int *sum); + +unsigned int aom_get_mb_ss_sse2(const int16_t *src) { + __m128i vsum = _mm_setzero_si128(); + int i; + + for (i = 0; i < 32; ++i) { + const __m128i v = _mm_loadu_si128((const __m128i *)src); + vsum = _mm_add_epi32(vsum, _mm_madd_epi16(v, v)); + src += 8; + } + + vsum = _mm_add_epi32(vsum, _mm_srli_si128(vsum, 8)); + vsum = _mm_add_epi32(vsum, _mm_srli_si128(vsum, 4)); + return _mm_cvtsi128_si32(vsum); +} + +#define READ64(p, stride, i) \ + _mm_unpacklo_epi8( \ + _mm_cvtsi32_si128(*(const uint32_t *)(p + i * stride)), \ + _mm_cvtsi32_si128(*(const uint32_t *)(p + (i + 1) * stride))) + +static void get4x4var_sse2(const uint8_t *src, int src_stride, + const uint8_t *ref, int ref_stride, + unsigned int *sse, int *sum) { + const __m128i zero = _mm_setzero_si128(); + const __m128i src0 = _mm_unpacklo_epi8(READ64(src, src_stride, 0), zero); + const __m128i src1 = _mm_unpacklo_epi8(READ64(src, src_stride, 2), zero); + const __m128i ref0 = _mm_unpacklo_epi8(READ64(ref, ref_stride, 0), zero); + const __m128i ref1 = _mm_unpacklo_epi8(READ64(ref, ref_stride, 2), zero); + const __m128i diff0 = _mm_sub_epi16(src0, ref0); + const __m128i diff1 = _mm_sub_epi16(src1, ref1); + + // sum + __m128i vsum = _mm_add_epi16(diff0, diff1); + vsum = _mm_add_epi16(vsum, _mm_srli_si128(vsum, 8)); + vsum = _mm_add_epi16(vsum, _mm_srli_si128(vsum, 4)); + vsum = _mm_add_epi16(vsum, _mm_srli_si128(vsum, 2)); + *sum = (int16_t)_mm_extract_epi16(vsum, 0); + + // sse + vsum = + _mm_add_epi32(_mm_madd_epi16(diff0, diff0), _mm_madd_epi16(diff1, diff1)); + vsum = _mm_add_epi32(vsum, _mm_srli_si128(vsum, 8)); + vsum = _mm_add_epi32(vsum, _mm_srli_si128(vsum, 4)); + *sse = _mm_cvtsi128_si32(vsum); +} + +void aom_get8x8var_sse2(const uint8_t *src, int src_stride, const uint8_t *ref, + int ref_stride, unsigned int *sse, int *sum) { + const __m128i zero = _mm_setzero_si128(); + __m128i vsum = _mm_setzero_si128(); + __m128i vsse = _mm_setzero_si128(); + int i; + + for (i = 0; i < 8; i += 2) { + const __m128i src0 = _mm_unpacklo_epi8( + _mm_loadl_epi64((const __m128i *)(src + i * src_stride)), zero); + const __m128i ref0 = _mm_unpacklo_epi8( + _mm_loadl_epi64((const __m128i *)(ref + i * ref_stride)), zero); + const __m128i diff0 = _mm_sub_epi16(src0, ref0); + + const __m128i src1 = _mm_unpacklo_epi8( + _mm_loadl_epi64((const __m128i *)(src + (i + 1) * src_stride)), zero); + const __m128i ref1 = _mm_unpacklo_epi8( + _mm_loadl_epi64((const __m128i *)(ref + (i + 1) * ref_stride)), zero); + const __m128i diff1 = _mm_sub_epi16(src1, ref1); + + vsum = _mm_add_epi16(vsum, diff0); + vsum = _mm_add_epi16(vsum, diff1); + vsse = _mm_add_epi32(vsse, _mm_madd_epi16(diff0, diff0)); + vsse = _mm_add_epi32(vsse, _mm_madd_epi16(diff1, diff1)); + } + + // sum + vsum = _mm_add_epi16(vsum, _mm_srli_si128(vsum, 8)); + vsum = _mm_add_epi16(vsum, _mm_srli_si128(vsum, 4)); + vsum = _mm_add_epi16(vsum, _mm_srli_si128(vsum, 2)); + *sum = (int16_t)_mm_extract_epi16(vsum, 0); + + // sse + vsse = _mm_add_epi32(vsse, _mm_srli_si128(vsse, 8)); + vsse = _mm_add_epi32(vsse, _mm_srli_si128(vsse, 4)); + *sse = _mm_cvtsi128_si32(vsse); +} + +void aom_get16x16var_sse2(const uint8_t *src, int src_stride, + const uint8_t *ref, int ref_stride, unsigned int *sse, + int *sum) { + const __m128i zero = _mm_setzero_si128(); + __m128i vsum = _mm_setzero_si128(); + __m128i vsse = _mm_setzero_si128(); + int i; + + for (i = 0; i < 16; ++i) { + const __m128i s = _mm_loadu_si128((const __m128i *)src); + const __m128i r = _mm_loadu_si128((const __m128i *)ref); + + const __m128i src0 = _mm_unpacklo_epi8(s, zero); + const __m128i ref0 = _mm_unpacklo_epi8(r, zero); + const __m128i diff0 = _mm_sub_epi16(src0, ref0); + + const __m128i src1 = _mm_unpackhi_epi8(s, zero); + const __m128i ref1 = _mm_unpackhi_epi8(r, zero); + const __m128i diff1 = _mm_sub_epi16(src1, ref1); + + vsum = _mm_add_epi16(vsum, diff0); + vsum = _mm_add_epi16(vsum, diff1); + vsse = _mm_add_epi32(vsse, _mm_madd_epi16(diff0, diff0)); + vsse = _mm_add_epi32(vsse, _mm_madd_epi16(diff1, diff1)); + + src += src_stride; + ref += ref_stride; + } + + // sum + vsum = _mm_add_epi16(vsum, _mm_srli_si128(vsum, 8)); + vsum = _mm_add_epi16(vsum, _mm_srli_si128(vsum, 4)); + *sum = + (int16_t)_mm_extract_epi16(vsum, 0) + (int16_t)_mm_extract_epi16(vsum, 1); + + // sse + vsse = _mm_add_epi32(vsse, _mm_srli_si128(vsse, 8)); + vsse = _mm_add_epi32(vsse, _mm_srli_si128(vsse, 4)); + *sse = _mm_cvtsi128_si32(vsse); +} + +static void variance_sse2(const unsigned char *src, int src_stride, + const unsigned char *ref, int ref_stride, int w, + int h, unsigned int *sse, int *sum, + getNxMvar_fn_t var_fn, int block_size) { + int i, j; + + *sse = 0; + *sum = 0; + + for (i = 0; i < h; i += block_size) { + for (j = 0; j < w; j += block_size) { + unsigned int sse0; + int sum0; + var_fn(src + src_stride * i + j, src_stride, ref + ref_stride * i + j, + ref_stride, &sse0, &sum0); + *sse += sse0; + *sum += sum0; + } + } +} + +unsigned int aom_variance4x4_sse2(const uint8_t *src, int src_stride, + const uint8_t *ref, int ref_stride, + unsigned int *sse) { + int sum; + get4x4var_sse2(src, src_stride, ref, ref_stride, sse, &sum); + assert(sum <= 255 * 4 * 4); + assert(sum >= -255 * 4 * 4); + return *sse - ((sum * sum) >> 4); +} + +unsigned int aom_variance8x4_sse2(const uint8_t *src, int src_stride, + const uint8_t *ref, int ref_stride, + unsigned int *sse) { + int sum; + variance_sse2(src, src_stride, ref, ref_stride, 8, 4, sse, &sum, + get4x4var_sse2, 4); + assert(sum <= 255 * 8 * 4); + assert(sum >= -255 * 8 * 4); + return *sse - ((sum * sum) >> 5); +} + +unsigned int aom_variance4x8_sse2(const uint8_t *src, int src_stride, + const uint8_t *ref, int ref_stride, + unsigned int *sse) { + int sum; + variance_sse2(src, src_stride, ref, ref_stride, 4, 8, sse, &sum, + get4x4var_sse2, 4); + assert(sum <= 255 * 8 * 4); + assert(sum >= -255 * 8 * 4); + return *sse - ((sum * sum) >> 5); +} + +unsigned int aom_variance8x8_sse2(const unsigned char *src, int src_stride, + const unsigned char *ref, int ref_stride, + unsigned int *sse) { + int sum; + aom_get8x8var_sse2(src, src_stride, ref, ref_stride, sse, &sum); + assert(sum <= 255 * 8 * 8); + assert(sum >= -255 * 8 * 8); + return *sse - ((sum * sum) >> 6); +} + +unsigned int aom_variance16x8_sse2(const unsigned char *src, int src_stride, + const unsigned char *ref, int ref_stride, + unsigned int *sse) { + int sum; + variance_sse2(src, src_stride, ref, ref_stride, 16, 8, sse, &sum, + aom_get8x8var_sse2, 8); + assert(sum <= 255 * 16 * 8); + assert(sum >= -255 * 16 * 8); + return *sse - ((sum * sum) >> 7); +} + +unsigned int aom_variance8x16_sse2(const unsigned char *src, int src_stride, + const unsigned char *ref, int ref_stride, + unsigned int *sse) { + int sum; + variance_sse2(src, src_stride, ref, ref_stride, 8, 16, sse, &sum, + aom_get8x8var_sse2, 8); + assert(sum <= 255 * 16 * 8); + assert(sum >= -255 * 16 * 8); + return *sse - ((sum * sum) >> 7); +} + +unsigned int aom_variance16x16_sse2(const unsigned char *src, int src_stride, + const unsigned char *ref, int ref_stride, + unsigned int *sse) { + int sum; + aom_get16x16var_sse2(src, src_stride, ref, ref_stride, sse, &sum); + assert(sum <= 255 * 16 * 16); + assert(sum >= -255 * 16 * 16); + return *sse - ((uint32_t)((int64_t)sum * sum) >> 8); +} + +unsigned int aom_variance32x32_sse2(const uint8_t *src, int src_stride, + const uint8_t *ref, int ref_stride, + unsigned int *sse) { + int sum; + variance_sse2(src, src_stride, ref, ref_stride, 32, 32, sse, &sum, + aom_get16x16var_sse2, 16); + assert(sum <= 255 * 32 * 32); + assert(sum >= -255 * 32 * 32); + return *sse - (unsigned int)(((int64_t)sum * sum) >> 10); +} + +unsigned int aom_variance32x16_sse2(const uint8_t *src, int src_stride, + const uint8_t *ref, int ref_stride, + unsigned int *sse) { + int sum; + variance_sse2(src, src_stride, ref, ref_stride, 32, 16, sse, &sum, + aom_get16x16var_sse2, 16); + assert(sum <= 255 * 32 * 16); + assert(sum >= -255 * 32 * 16); + return *sse - (unsigned int)(((int64_t)sum * sum) >> 9); +} + +unsigned int aom_variance16x32_sse2(const uint8_t *src, int src_stride, + const uint8_t *ref, int ref_stride, + unsigned int *sse) { + int sum; + variance_sse2(src, src_stride, ref, ref_stride, 16, 32, sse, &sum, + aom_get16x16var_sse2, 16); + assert(sum <= 255 * 32 * 16); + assert(sum >= -255 * 32 * 16); + return *sse - (unsigned int)(((int64_t)sum * sum) >> 9); +} + +unsigned int aom_variance64x64_sse2(const uint8_t *src, int src_stride, + const uint8_t *ref, int ref_stride, + unsigned int *sse) { + int sum; + variance_sse2(src, src_stride, ref, ref_stride, 64, 64, sse, &sum, + aom_get16x16var_sse2, 16); + assert(sum <= 255 * 64 * 64); + assert(sum >= -255 * 64 * 64); + return *sse - (unsigned int)(((int64_t)sum * sum) >> 12); +} + +unsigned int aom_variance64x32_sse2(const uint8_t *src, int src_stride, + const uint8_t *ref, int ref_stride, + unsigned int *sse) { + int sum; + variance_sse2(src, src_stride, ref, ref_stride, 64, 32, sse, &sum, + aom_get16x16var_sse2, 16); + assert(sum <= 255 * 64 * 32); + assert(sum >= -255 * 64 * 32); + return *sse - (unsigned int)(((int64_t)sum * sum) >> 11); +} + +unsigned int aom_variance32x64_sse2(const uint8_t *src, int src_stride, + const uint8_t *ref, int ref_stride, + unsigned int *sse) { + int sum; + variance_sse2(src, src_stride, ref, ref_stride, 32, 64, sse, &sum, + aom_get16x16var_sse2, 16); + assert(sum <= 255 * 64 * 32); + assert(sum >= -255 * 64 * 32); + return *sse - (unsigned int)(((int64_t)sum * sum) >> 11); +} + +unsigned int aom_mse8x8_sse2(const uint8_t *src, int src_stride, + const uint8_t *ref, int ref_stride, + unsigned int *sse) { + aom_variance8x8_sse2(src, src_stride, ref, ref_stride, sse); + return *sse; +} + +unsigned int aom_mse8x16_sse2(const uint8_t *src, int src_stride, + const uint8_t *ref, int ref_stride, + unsigned int *sse) { + aom_variance8x16_sse2(src, src_stride, ref, ref_stride, sse); + return *sse; +} + +unsigned int aom_mse16x8_sse2(const uint8_t *src, int src_stride, + const uint8_t *ref, int ref_stride, + unsigned int *sse) { + aom_variance16x8_sse2(src, src_stride, ref, ref_stride, sse); + return *sse; +} + +unsigned int aom_mse16x16_sse2(const uint8_t *src, int src_stride, + const uint8_t *ref, int ref_stride, + unsigned int *sse) { + aom_variance16x16_sse2(src, src_stride, ref, ref_stride, sse); + return *sse; +} + +// The 2 unused parameters are place holders for PIC enabled build. +// These definitions are for functions defined in subpel_variance.asm +#define DECL(w, opt) \ + int aom_sub_pixel_variance##w##xh_##opt( \ + const uint8_t *src, ptrdiff_t src_stride, int x_offset, int y_offset, \ + const uint8_t *dst, ptrdiff_t dst_stride, int height, unsigned int *sse, \ + void *unused0, void *unused) +#define DECLS(opt1, opt2) \ + DECL(4, opt1); \ + DECL(8, opt1); \ + DECL(16, opt1) + +DECLS(sse2, sse2); +DECLS(ssse3, ssse3); +#undef DECLS +#undef DECL + +#define FN(w, h, wf, wlog2, hlog2, opt, cast_prod, cast) \ + unsigned int aom_sub_pixel_variance##w##x##h##_##opt( \ + const uint8_t *src, int src_stride, int x_offset, int y_offset, \ + const uint8_t *dst, int dst_stride, unsigned int *sse_ptr) { \ + unsigned int sse; \ + int se = aom_sub_pixel_variance##wf##xh_##opt(src, src_stride, x_offset, \ + y_offset, dst, dst_stride, \ + h, &sse, NULL, NULL); \ + if (w > wf) { \ + unsigned int sse2; \ + int se2 = aom_sub_pixel_variance##wf##xh_##opt( \ + src + 16, src_stride, x_offset, y_offset, dst + 16, dst_stride, h, \ + &sse2, NULL, NULL); \ + se += se2; \ + sse += sse2; \ + if (w > wf * 2) { \ + se2 = aom_sub_pixel_variance##wf##xh_##opt( \ + src + 32, src_stride, x_offset, y_offset, dst + 32, dst_stride, h, \ + &sse2, NULL, NULL); \ + se += se2; \ + sse += sse2; \ + se2 = aom_sub_pixel_variance##wf##xh_##opt( \ + src + 48, src_stride, x_offset, y_offset, dst + 48, dst_stride, h, \ + &sse2, NULL, NULL); \ + se += se2; \ + sse += sse2; \ + } \ + } \ + *sse_ptr = sse; \ + return sse - (unsigned int)(cast_prod(cast se * se) >> (wlog2 + hlog2)); \ + } + +#define FNS(opt1, opt2) \ + FN(64, 64, 16, 6, 6, opt1, (int64_t), (int64_t)); \ + FN(64, 32, 16, 6, 5, opt1, (int64_t), (int64_t)); \ + FN(32, 64, 16, 5, 6, opt1, (int64_t), (int64_t)); \ + FN(32, 32, 16, 5, 5, opt1, (int64_t), (int64_t)); \ + FN(32, 16, 16, 5, 4, opt1, (int64_t), (int64_t)); \ + FN(16, 32, 16, 4, 5, opt1, (int64_t), (int64_t)); \ + FN(16, 16, 16, 4, 4, opt1, (uint32_t), (int64_t)); \ + FN(16, 8, 16, 4, 3, opt1, (int32_t), (int32_t)); \ + FN(8, 16, 8, 3, 4, opt1, (int32_t), (int32_t)); \ + FN(8, 8, 8, 3, 3, opt1, (int32_t), (int32_t)); \ + FN(8, 4, 8, 3, 2, opt1, (int32_t), (int32_t)); \ + FN(4, 8, 4, 2, 3, opt1, (int32_t), (int32_t)); \ + FN(4, 4, 4, 2, 2, opt1, (int32_t), (int32_t)) + +FNS(sse2, sse2); +FNS(ssse3, ssse3); + +#undef FNS +#undef FN + +// The 2 unused parameters are place holders for PIC enabled build. +#define DECL(w, opt) \ + int aom_sub_pixel_avg_variance##w##xh_##opt( \ + const uint8_t *src, ptrdiff_t src_stride, int x_offset, int y_offset, \ + const uint8_t *dst, ptrdiff_t dst_stride, const uint8_t *sec, \ + ptrdiff_t sec_stride, int height, unsigned int *sse, void *unused0, \ + void *unused) +#define DECLS(opt1, opt2) \ + DECL(4, opt1); \ + DECL(8, opt1); \ + DECL(16, opt1) + +DECLS(sse2, sse2); +DECLS(ssse3, ssse3); +#undef DECL +#undef DECLS + +#define FN(w, h, wf, wlog2, hlog2, opt, cast_prod, cast) \ + unsigned int aom_sub_pixel_avg_variance##w##x##h##_##opt( \ + const uint8_t *src, int src_stride, int x_offset, int y_offset, \ + const uint8_t *dst, int dst_stride, unsigned int *sseptr, \ + const uint8_t *sec) { \ + unsigned int sse; \ + int se = aom_sub_pixel_avg_variance##wf##xh_##opt( \ + src, src_stride, x_offset, y_offset, dst, dst_stride, sec, w, h, &sse, \ + NULL, NULL); \ + if (w > wf) { \ + unsigned int sse2; \ + int se2 = aom_sub_pixel_avg_variance##wf##xh_##opt( \ + src + 16, src_stride, x_offset, y_offset, dst + 16, dst_stride, \ + sec + 16, w, h, &sse2, NULL, NULL); \ + se += se2; \ + sse += sse2; \ + if (w > wf * 2) { \ + se2 = aom_sub_pixel_avg_variance##wf##xh_##opt( \ + src + 32, src_stride, x_offset, y_offset, dst + 32, dst_stride, \ + sec + 32, w, h, &sse2, NULL, NULL); \ + se += se2; \ + sse += sse2; \ + se2 = aom_sub_pixel_avg_variance##wf##xh_##opt( \ + src + 48, src_stride, x_offset, y_offset, dst + 48, dst_stride, \ + sec + 48, w, h, &sse2, NULL, NULL); \ + se += se2; \ + sse += sse2; \ + } \ + } \ + *sseptr = sse; \ + return sse - (unsigned int)(cast_prod(cast se * se) >> (wlog2 + hlog2)); \ + } + +#define FNS(opt1, opt2) \ + FN(64, 64, 16, 6, 6, opt1, (int64_t), (int64_t)); \ + FN(64, 32, 16, 6, 5, opt1, (int64_t), (int64_t)); \ + FN(32, 64, 16, 5, 6, opt1, (int64_t), (int64_t)); \ + FN(32, 32, 16, 5, 5, opt1, (int64_t), (int64_t)); \ + FN(32, 16, 16, 5, 4, opt1, (int64_t), (int64_t)); \ + FN(16, 32, 16, 4, 5, opt1, (int64_t), (int64_t)); \ + FN(16, 16, 16, 4, 4, opt1, (uint32_t), (int64_t)); \ + FN(16, 8, 16, 4, 3, opt1, (uint32_t), (int32_t)); \ + FN(8, 16, 8, 3, 4, opt1, (uint32_t), (int32_t)); \ + FN(8, 8, 8, 3, 3, opt1, (uint32_t), (int32_t)); \ + FN(8, 4, 8, 3, 2, opt1, (uint32_t), (int32_t)); \ + FN(4, 8, 4, 2, 3, opt1, (uint32_t), (int32_t)); \ + FN(4, 4, 4, 2, 2, opt1, (uint32_t), (int32_t)) + +FNS(sse2, sse); +FNS(ssse3, ssse3); + +#undef FNS +#undef FN + +void aom_upsampled_pred_sse2(uint8_t *comp_pred, int width, int height, + const uint8_t *ref, int ref_stride) { + int i, j; + int stride = ref_stride << 3; + + if (width >= 16) { + // read 16 points at one time + for (i = 0; i < height; i++) { + for (j = 0; j < width; j += 16) { + __m128i s0 = _mm_loadu_si128((const __m128i *)ref); + __m128i s1 = _mm_loadu_si128((const __m128i *)(ref + 16)); + __m128i s2 = _mm_loadu_si128((const __m128i *)(ref + 32)); + __m128i s3 = _mm_loadu_si128((const __m128i *)(ref + 48)); + __m128i s4 = _mm_loadu_si128((const __m128i *)(ref + 64)); + __m128i s5 = _mm_loadu_si128((const __m128i *)(ref + 80)); + __m128i s6 = _mm_loadu_si128((const __m128i *)(ref + 96)); + __m128i s7 = _mm_loadu_si128((const __m128i *)(ref + 112)); + __m128i t0, t1, t2, t3; + + t0 = _mm_unpacklo_epi8(s0, s1); + s1 = _mm_unpackhi_epi8(s0, s1); + t1 = _mm_unpacklo_epi8(s2, s3); + s3 = _mm_unpackhi_epi8(s2, s3); + t2 = _mm_unpacklo_epi8(s4, s5); + s5 = _mm_unpackhi_epi8(s4, s5); + t3 = _mm_unpacklo_epi8(s6, s7); + s7 = _mm_unpackhi_epi8(s6, s7); + + s0 = _mm_unpacklo_epi8(t0, s1); + s2 = _mm_unpacklo_epi8(t1, s3); + s4 = _mm_unpacklo_epi8(t2, s5); + s6 = _mm_unpacklo_epi8(t3, s7); + s0 = _mm_unpacklo_epi32(s0, s2); + s4 = _mm_unpacklo_epi32(s4, s6); + s0 = _mm_unpacklo_epi64(s0, s4); + + _mm_storeu_si128((__m128i *)(comp_pred), s0); + comp_pred += 16; + ref += 16 * 8; + } + ref += stride - (width << 3); + } + } else if (width >= 8) { + // read 8 points at one time + for (i = 0; i < height; i++) { + for (j = 0; j < width; j += 8) { + __m128i s0 = _mm_loadu_si128((const __m128i *)ref); + __m128i s1 = _mm_loadu_si128((const __m128i *)(ref + 16)); + __m128i s2 = _mm_loadu_si128((const __m128i *)(ref + 32)); + __m128i s3 = _mm_loadu_si128((const __m128i *)(ref + 48)); + __m128i t0, t1; + + t0 = _mm_unpacklo_epi8(s0, s1); + s1 = _mm_unpackhi_epi8(s0, s1); + t1 = _mm_unpacklo_epi8(s2, s3); + s3 = _mm_unpackhi_epi8(s2, s3); + + s0 = _mm_unpacklo_epi8(t0, s1); + s2 = _mm_unpacklo_epi8(t1, s3); + s0 = _mm_unpacklo_epi32(s0, s2); + + _mm_storel_epi64((__m128i *)(comp_pred), s0); + comp_pred += 8; + ref += 8 * 8; + } + ref += stride - (width << 3); + } + } else { + // read 4 points at one time + for (i = 0; i < height; i++) { + for (j = 0; j < width; j += 4) { + __m128i s0 = _mm_loadu_si128((const __m128i *)ref); + __m128i s1 = _mm_loadu_si128((const __m128i *)(ref + 16)); + __m128i t0; + + t0 = _mm_unpacklo_epi8(s0, s1); + s1 = _mm_unpackhi_epi8(s0, s1); + s0 = _mm_unpacklo_epi8(t0, s1); + + *(int *)comp_pred = _mm_cvtsi128_si32(s0); + comp_pred += 4; + ref += 4 * 8; + } + ref += stride - (width << 3); + } + } +} + +void aom_comp_avg_upsampled_pred_sse2(uint8_t *comp_pred, const uint8_t *pred, + int width, int height, const uint8_t *ref, + int ref_stride) { + const __m128i zero = _mm_set1_epi16(0); + const __m128i one = _mm_set1_epi16(1); + int i, j; + int stride = ref_stride << 3; + + if (width >= 16) { + // read 16 points at one time + for (i = 0; i < height; i++) { + for (j = 0; j < width; j += 16) { + __m128i s0 = _mm_loadu_si128((const __m128i *)ref); + __m128i s1 = _mm_loadu_si128((const __m128i *)(ref + 16)); + __m128i s2 = _mm_loadu_si128((const __m128i *)(ref + 32)); + __m128i s3 = _mm_loadu_si128((const __m128i *)(ref + 48)); + __m128i s4 = _mm_loadu_si128((const __m128i *)(ref + 64)); + __m128i s5 = _mm_loadu_si128((const __m128i *)(ref + 80)); + __m128i s6 = _mm_loadu_si128((const __m128i *)(ref + 96)); + __m128i s7 = _mm_loadu_si128((const __m128i *)(ref + 112)); + __m128i p0 = _mm_loadu_si128((const __m128i *)pred); + __m128i p1; + __m128i t0, t1, t2, t3; + + t0 = _mm_unpacklo_epi8(s0, s1); + s1 = _mm_unpackhi_epi8(s0, s1); + t1 = _mm_unpacklo_epi8(s2, s3); + s3 = _mm_unpackhi_epi8(s2, s3); + t2 = _mm_unpacklo_epi8(s4, s5); + s5 = _mm_unpackhi_epi8(s4, s5); + t3 = _mm_unpacklo_epi8(s6, s7); + s7 = _mm_unpackhi_epi8(s6, s7); + + s0 = _mm_unpacklo_epi8(t0, s1); + s2 = _mm_unpacklo_epi8(t1, s3); + s4 = _mm_unpacklo_epi8(t2, s5); + s6 = _mm_unpacklo_epi8(t3, s7); + + s0 = _mm_unpacklo_epi32(s0, s2); + s4 = _mm_unpacklo_epi32(s4, s6); + s0 = _mm_unpacklo_epi8(s0, zero); + s4 = _mm_unpacklo_epi8(s4, zero); + + p1 = _mm_unpackhi_epi8(p0, zero); + p0 = _mm_unpacklo_epi8(p0, zero); + p0 = _mm_adds_epu16(s0, p0); + p1 = _mm_adds_epu16(s4, p1); + p0 = _mm_adds_epu16(p0, one); + p1 = _mm_adds_epu16(p1, one); + + p0 = _mm_srli_epi16(p0, 1); + p1 = _mm_srli_epi16(p1, 1); + p0 = _mm_packus_epi16(p0, p1); + + _mm_storeu_si128((__m128i *)(comp_pred), p0); + comp_pred += 16; + pred += 16; + ref += 16 * 8; + } + ref += stride - (width << 3); + } + } else if (width >= 8) { + // read 8 points at one time + for (i = 0; i < height; i++) { + for (j = 0; j < width; j += 8) { + __m128i s0 = _mm_loadu_si128((const __m128i *)ref); + __m128i s1 = _mm_loadu_si128((const __m128i *)(ref + 16)); + __m128i s2 = _mm_loadu_si128((const __m128i *)(ref + 32)); + __m128i s3 = _mm_loadu_si128((const __m128i *)(ref + 48)); + __m128i p0 = _mm_loadl_epi64((const __m128i *)pred); + __m128i t0, t1; + + t0 = _mm_unpacklo_epi8(s0, s1); + s1 = _mm_unpackhi_epi8(s0, s1); + t1 = _mm_unpacklo_epi8(s2, s3); + s3 = _mm_unpackhi_epi8(s2, s3); + + s0 = _mm_unpacklo_epi8(t0, s1); + s2 = _mm_unpacklo_epi8(t1, s3); + s0 = _mm_unpacklo_epi32(s0, s2); + s0 = _mm_unpacklo_epi8(s0, zero); + + p0 = _mm_unpacklo_epi8(p0, zero); + p0 = _mm_adds_epu16(s0, p0); + p0 = _mm_adds_epu16(p0, one); + p0 = _mm_srli_epi16(p0, 1); + p0 = _mm_packus_epi16(p0, zero); + + _mm_storel_epi64((__m128i *)(comp_pred), p0); + comp_pred += 8; + pred += 8; + ref += 8 * 8; + } + ref += stride - (width << 3); + } + } else { + // read 4 points at one time + for (i = 0; i < height; i++) { + for (j = 0; j < width; j += 4) { + __m128i s0 = _mm_loadu_si128((const __m128i *)ref); + __m128i s1 = _mm_loadu_si128((const __m128i *)(ref + 16)); + __m128i p0 = _mm_cvtsi32_si128(*(const uint32_t *)pred); + __m128i t0; + + t0 = _mm_unpacklo_epi8(s0, s1); + s1 = _mm_unpackhi_epi8(s0, s1); + s0 = _mm_unpacklo_epi8(t0, s1); + s0 = _mm_unpacklo_epi8(s0, zero); + + p0 = _mm_unpacklo_epi8(p0, zero); + p0 = _mm_adds_epu16(s0, p0); + p0 = _mm_adds_epu16(p0, one); + p0 = _mm_srli_epi16(p0, 1); + p0 = _mm_packus_epi16(p0, zero); + + *(int *)comp_pred = _mm_cvtsi128_si32(p0); + comp_pred += 4; + pred += 4; + ref += 4 * 8; + } + ref += stride - (width << 3); + } + } +} |