From df9477dfa60ebb5d31bc142e58ce46535c17abce Mon Sep 17 00:00:00 2001 From: trav90 Date: Wed, 17 Oct 2018 05:59:08 -0500 Subject: Update aom to slightly newer commit ID --- third_party/aom/aom_dsp/aom_convolve.c | 267 +- third_party/aom/aom_dsp/aom_convolve.h | 5 + third_party/aom/aom_dsp/aom_dsp.cmake | 84 +- third_party/aom/aom_dsp/aom_dsp.mk | 42 +- third_party/aom/aom_dsp/aom_dsp_common.h | 2 + third_party/aom/aom_dsp/aom_dsp_rtcd_defs.pl | 177 +- third_party/aom/aom_dsp/arm/avg_neon.c | 38 - .../aom/aom_dsp/arm/bilinear_filter_media.asm | 240 -- third_party/aom/aom_dsp/arm/sad_media.asm | 98 - .../aom/aom_dsp/arm/subpel_variance_media.c | 81 - .../arm/variance_halfpixvar16x16_h_media.asm | 185 -- .../arm/variance_halfpixvar16x16_hv_media.asm | 225 -- .../arm/variance_halfpixvar16x16_v_media.asm | 187 -- third_party/aom/aom_dsp/arm/variance_media.asm | 361 --- third_party/aom/aom_dsp/avg.c | 42 - third_party/aom/aom_dsp/binary_codes_reader.c | 53 +- third_party/aom/aom_dsp/binary_codes_reader.h | 34 +- third_party/aom/aom_dsp/bitreader.h | 62 +- third_party/aom/aom_dsp/bitreader_buffer.c | 2 +- third_party/aom/aom_dsp/bitwriter.h | 63 +- third_party/aom/aom_dsp/dkboolreader.c | 110 - third_party/aom/aom_dsp/dkboolreader.h | 181 -- third_party/aom/aom_dsp/dkboolwriter.c | 44 - third_party/aom/aom_dsp/dkboolwriter.h | 104 - third_party/aom/aom_dsp/intrapred.c | 162 +- third_party/aom/aom_dsp/inv_txfm.c | 864 ++++++ third_party/aom/aom_dsp/loopfilter.c | 110 +- third_party/aom/aom_dsp/mips/avg_msa.c | 57 - third_party/aom/aom_dsp/prob.c | 17 - third_party/aom/aom_dsp/prob.h | 12 +- third_party/aom/aom_dsp/sad.c | 54 +- third_party/aom/aom_dsp/simd/v64_intrinsics.h | 4 +- third_party/aom/aom_dsp/variance.c | 365 +-- third_party/aom/aom_dsp/variance.h | 13 +- .../aom/aom_dsp/x86/aom_convolve_hip_sse2.c | 195 ++ .../aom_dsp/x86/aom_highbd_convolve_hip_ssse3.c | 203 ++ third_party/aom/aom_dsp/x86/avg_intrin_sse2.c | 46 - third_party/aom/aom_dsp/x86/highbd_convolve_avx2.c | 656 +++-- third_party/aom/aom_dsp/x86/inv_txfm_avx2.c | 1238 +++++++++ third_party/aom/aom_dsp/x86/inv_txfm_common_avx2.h | 80 + third_party/aom/aom_dsp/x86/inv_txfm_sse2.c | 103 + .../aom/aom_dsp/x86/masked_sad_intrin_ssse3.c | 498 ++-- .../aom/aom_dsp/x86/masked_variance_intrin_ssse3.c | 2799 +++++++------------- third_party/aom/aom_dsp/x86/obmc_intrinsic_ssse3.h | 45 + third_party/aom/aom_dsp/x86/obmc_sad_sse4.c | 1 + third_party/aom/aom_dsp/x86/obmc_variance_sse4.c | 3 +- third_party/aom/aom_dsp/x86/synonyms.h | 28 - third_party/aom/aom_dsp/x86/txfm_common_avx2.h | 44 +- 48 files changed, 5193 insertions(+), 5091 deletions(-) delete mode 100644 third_party/aom/aom_dsp/arm/bilinear_filter_media.asm delete mode 100644 third_party/aom/aom_dsp/arm/sad_media.asm delete mode 100644 third_party/aom/aom_dsp/arm/subpel_variance_media.c delete mode 100644 third_party/aom/aom_dsp/arm/variance_halfpixvar16x16_h_media.asm delete mode 100644 third_party/aom/aom_dsp/arm/variance_halfpixvar16x16_hv_media.asm delete mode 100644 third_party/aom/aom_dsp/arm/variance_halfpixvar16x16_v_media.asm delete mode 100644 third_party/aom/aom_dsp/arm/variance_media.asm delete mode 100644 third_party/aom/aom_dsp/dkboolreader.c delete mode 100644 third_party/aom/aom_dsp/dkboolreader.h delete mode 100644 third_party/aom/aom_dsp/dkboolwriter.c delete mode 100644 third_party/aom/aom_dsp/dkboolwriter.h delete mode 100644 third_party/aom/aom_dsp/mips/avg_msa.c create mode 100644 third_party/aom/aom_dsp/x86/aom_convolve_hip_sse2.c create mode 100644 third_party/aom/aom_dsp/x86/aom_highbd_convolve_hip_ssse3.c create mode 100644 third_party/aom/aom_dsp/x86/inv_txfm_avx2.c create mode 100644 third_party/aom/aom_dsp/x86/inv_txfm_common_avx2.h create mode 100644 third_party/aom/aom_dsp/x86/obmc_intrinsic_ssse3.h (limited to 'third_party/aom/aom_dsp') diff --git a/third_party/aom/aom_dsp/aom_convolve.c b/third_party/aom/aom_dsp/aom_convolve.c index 74f4c00fb..4dac6aacc 100644 --- a/third_party/aom/aom_dsp/aom_convolve.c +++ b/third_party/aom/aom_dsp/aom_convolve.c @@ -337,14 +337,14 @@ static void convolve_add_src_horiz(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) { - int x, y; + int x, y, k; src -= SUBPEL_TAPS / 2 - 1; for (y = 0; y < h; ++y) { int x_q4 = x0_q4; for (x = 0; x < w; ++x) { const uint8_t *const src_x = &src[x_q4 >> SUBPEL_BITS]; const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK]; - int k, sum = 0; + int sum = 0; for (k = 0; k < SUBPEL_TAPS; ++k) sum += src_x[k] * x_filter[k]; dst[x] = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS) + src_x[SUBPEL_TAPS / 2 - 1]); @@ -359,7 +359,7 @@ static void convolve_add_src_vert(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 x, y; + int x, y, k; src -= src_stride * (SUBPEL_TAPS / 2 - 1); for (x = 0; x < w; ++x) { @@ -367,7 +367,7 @@ static void convolve_add_src_vert(const uint8_t *src, ptrdiff_t src_stride, 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]; - int k, sum = 0; + int sum = 0; for (k = 0; k < SUBPEL_TAPS; ++k) sum += src_y[k * src_stride] * y_filter[k]; dst[y * dst_stride] = @@ -446,6 +446,127 @@ void aom_convolve8_add_src_c(const uint8_t *src, ptrdiff_t src_stride, convolve_add_src(src, src_stride, dst, dst_stride, filters_x, x0_q4, x_step_q4, filters_y, y0_q4, y_step_q4, w, h); } + +static void convolve_add_src_horiz_hip(const uint8_t *src, ptrdiff_t src_stride, + uint16_t *dst, ptrdiff_t dst_stride, + const InterpKernel *x_filters, int x0_q4, + int x_step_q4, int w, int h) { + const int bd = 8; + int x, y, k; + src -= SUBPEL_TAPS / 2 - 1; + for (y = 0; y < h; ++y) { + int x_q4 = x0_q4; + for (x = 0; x < w; ++x) { + const uint8_t *const src_x = &src[x_q4 >> SUBPEL_BITS]; + const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK]; + int sum = ((int)src_x[SUBPEL_TAPS / 2 - 1] << FILTER_BITS) + + (1 << (bd + FILTER_BITS - 1)); + for (k = 0; k < SUBPEL_TAPS; ++k) sum += src_x[k] * x_filter[k]; + dst[x] = + (uint16_t)clamp(ROUND_POWER_OF_TWO(sum, FILTER_BITS - EXTRAPREC_BITS), + 0, EXTRAPREC_CLAMP_LIMIT(bd) - 1); + x_q4 += x_step_q4; + } + src += src_stride; + dst += dst_stride; + } +} + +static void convolve_add_src_vert_hip(const uint16_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) { + const int bd = 8; + int x, y, k; + src -= src_stride * (SUBPEL_TAPS / 2 - 1); + + for (x = 0; x < w; ++x) { + int y_q4 = y0_q4; + for (y = 0; y < h; ++y) { + const uint16_t *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride]; + const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK]; + int sum = + ((int)src_y[(SUBPEL_TAPS / 2 - 1) * src_stride] << FILTER_BITS) - + (1 << (bd + FILTER_BITS + EXTRAPREC_BITS - 1)); + for (k = 0; k < SUBPEL_TAPS; ++k) + sum += src_y[k * src_stride] * y_filter[k]; + dst[y * dst_stride] = + clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS + EXTRAPREC_BITS)); + y_q4 += y_step_q4; + } + ++src; + ++dst; + } +} + +static void convolve_add_src_hip(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) { + uint16_t temp[MAX_EXT_SIZE * MAX_SB_SIZE]; + 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); + + convolve_add_src_horiz_hip(src - src_stride * (SUBPEL_TAPS / 2 - 1), + src_stride, temp, MAX_SB_SIZE, x_filters, x0_q4, + x_step_q4, w, intermediate_height); + convolve_add_src_vert_hip(temp + MAX_SB_SIZE * (SUBPEL_TAPS / 2 - 1), + MAX_SB_SIZE, dst, dst_stride, y_filters, y0_q4, + y_step_q4, w, h); +} + +void aom_convolve8_add_src_horiz_hip_c(const uint8_t *src, ptrdiff_t src_stride, + uint16_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); + + (void)filter_y; + (void)y_step_q4; + + convolve_add_src_horiz_hip(src, src_stride, dst, dst_stride, filters_x, x0_q4, + x_step_q4, w, h); +} + +void aom_convolve8_add_src_vert_hip_c(const uint16_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_y = get_filter_base(filter_y); + const int y0_q4 = get_filter_offset(filter_y, filters_y); + + (void)filter_x; + (void)x_step_q4; + + convolve_add_src_vert_hip(src, src_stride, dst, dst_stride, filters_y, y0_q4, + y_step_q4, w, h); +} + +void aom_convolve8_add_src_hip_c(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); + + convolve_add_src_hip(src, src_stride, dst, dst_stride, filters_x, x0_q4, + x_step_q4, filters_y, y0_q4, y_step_q4, w, h); +} #endif // CONFIG_LOOP_RESTORATION #if CONFIG_HIGHBITDEPTH @@ -721,7 +842,7 @@ static void highbd_convolve_add_src_horiz(const uint8_t *src8, const InterpKernel *x_filters, int x0_q4, int x_step_q4, int w, int h, int bd) { - int x, y; + int x, y, k; uint16_t *src = CONVERT_TO_SHORTPTR(src8); uint16_t *dst = CONVERT_TO_SHORTPTR(dst8); src -= SUBPEL_TAPS / 2 - 1; @@ -730,7 +851,7 @@ static void highbd_convolve_add_src_horiz(const uint8_t *src8, for (x = 0; x < w; ++x) { const uint16_t *const src_x = &src[x_q4 >> SUBPEL_BITS]; const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK]; - int k, sum = 0; + int sum = 0; for (k = 0; k < SUBPEL_TAPS; ++k) sum += src_x[k] * x_filter[k]; dst[x] = clip_pixel_highbd( ROUND_POWER_OF_TWO(sum, FILTER_BITS) + src_x[SUBPEL_TAPS / 2 - 1], @@ -748,7 +869,7 @@ static void highbd_convolve_add_src_vert(const uint8_t *src8, const InterpKernel *y_filters, int y0_q4, int y_step_q4, int w, int h, int bd) { - int x, y; + int x, y, k; uint16_t *src = CONVERT_TO_SHORTPTR(src8); uint16_t *dst = CONVERT_TO_SHORTPTR(dst8); src -= src_stride * (SUBPEL_TAPS / 2 - 1); @@ -757,7 +878,7 @@ static void highbd_convolve_add_src_vert(const uint8_t *src8, for (y = 0; y < h; ++y) { const uint16_t *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride]; const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK]; - int k, sum = 0; + int sum = 0; for (k = 0; k < SUBPEL_TAPS; ++k) sum += src_y[k * src_stride] * y_filter[k]; dst[y * dst_stride] = @@ -850,5 +971,135 @@ void aom_highbd_convolve8_add_src_c(const uint8_t *src, ptrdiff_t src_stride, highbd_convolve_add_src(src, src_stride, dst, dst_stride, filters_x, x0_q4, x_step_q4, filters_y, y0_q4, y_step_q4, w, h, bd); } + +static void highbd_convolve_add_src_horiz_hip( + const uint8_t *src8, ptrdiff_t src_stride, uint16_t *dst, + ptrdiff_t dst_stride, const InterpKernel *x_filters, int x0_q4, + int x_step_q4, int w, int h, int bd) { + const int extraprec_clamp_limit = EXTRAPREC_CLAMP_LIMIT(bd); + int x, y, k; + uint16_t *src = CONVERT_TO_SHORTPTR(src8); + src -= SUBPEL_TAPS / 2 - 1; + for (y = 0; y < h; ++y) { + int x_q4 = x0_q4; + for (x = 0; x < w; ++x) { + const uint16_t *const src_x = &src[x_q4 >> SUBPEL_BITS]; + const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK]; + int sum = ((int)src_x[SUBPEL_TAPS / 2 - 1] << FILTER_BITS) + + (1 << (bd + FILTER_BITS - 1)); + for (k = 0; k < SUBPEL_TAPS; ++k) sum += src_x[k] * x_filter[k]; + dst[x] = + (uint16_t)clamp(ROUND_POWER_OF_TWO(sum, FILTER_BITS - EXTRAPREC_BITS), + 0, extraprec_clamp_limit - 1); + x_q4 += x_step_q4; + } + src += src_stride; + dst += dst_stride; + } +} + +static void highbd_convolve_add_src_vert_hip( + const uint16_t *src, ptrdiff_t src_stride, uint8_t *dst8, + ptrdiff_t dst_stride, const InterpKernel *y_filters, int y0_q4, + int y_step_q4, int w, int h, int bd) { + int x, y, k; + uint16_t *dst = CONVERT_TO_SHORTPTR(dst8); + src -= src_stride * (SUBPEL_TAPS / 2 - 1); + for (x = 0; x < w; ++x) { + int y_q4 = y0_q4; + for (y = 0; y < h; ++y) { + const uint16_t *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride]; + const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK]; + int sum = + ((int)src_y[(SUBPEL_TAPS / 2 - 1) * src_stride] << FILTER_BITS) - + (1 << (bd + FILTER_BITS + EXTRAPREC_BITS - 1)); + for (k = 0; k < SUBPEL_TAPS; ++k) + sum += src_y[k * src_stride] * y_filter[k]; + dst[y * dst_stride] = clip_pixel_highbd( + ROUND_POWER_OF_TWO(sum, FILTER_BITS + EXTRAPREC_BITS), bd); + y_q4 += y_step_q4; + } + ++src; + ++dst; + } +} + +static void highbd_convolve_add_src_hip( + 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, int bd) { + // 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. + uint16_t temp[MAX_EXT_SIZE * MAX_SB_SIZE]; + 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); + + highbd_convolve_add_src_horiz_hip( + src - src_stride * (SUBPEL_TAPS / 2 - 1), src_stride, temp, MAX_SB_SIZE, + x_filters, x0_q4, x_step_q4, w, intermediate_height, bd); + highbd_convolve_add_src_vert_hip(temp + MAX_SB_SIZE * (SUBPEL_TAPS / 2 - 1), + MAX_SB_SIZE, dst, dst_stride, y_filters, + y0_q4, y_step_q4, w, h, bd); +} + +void aom_highbd_convolve8_add_src_horiz_hip_c( + const uint8_t *src, ptrdiff_t src_stride, uint16_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) { + const InterpKernel *const filters_x = get_filter_base(filter_x); + const int x0_q4 = get_filter_offset(filter_x, filters_x); + (void)filter_y; + (void)y_step_q4; + + highbd_convolve_add_src_horiz_hip(src, src_stride, dst, dst_stride, filters_x, + x0_q4, x_step_q4, w, h, bd); +} + +void aom_highbd_convolve8_add_src_vert_hip_c( + const uint16_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) { + const InterpKernel *const filters_y = get_filter_base(filter_y); + const int y0_q4 = get_filter_offset(filter_y, filters_y); + (void)filter_x; + (void)x_step_q4; + + highbd_convolve_add_src_vert_hip(src, src_stride, dst, dst_stride, filters_y, + y0_q4, y_step_q4, w, h, bd); +} + +void aom_highbd_convolve8_add_src_hip_c(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) { + 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); + + highbd_convolve_add_src_hip(src, src_stride, dst, dst_stride, filters_x, + x0_q4, x_step_q4, filters_y, y0_q4, y_step_q4, w, + h, bd); +} + #endif // CONFIG_LOOP_RESTORATION #endif // CONFIG_HIGHBITDEPTH diff --git a/third_party/aom/aom_dsp/aom_convolve.h b/third_party/aom/aom_dsp/aom_convolve.h index d0de6c5d2..c7943dced 100644 --- a/third_party/aom/aom_dsp/aom_convolve.h +++ b/third_party/aom/aom_dsp/aom_convolve.h @@ -36,6 +36,11 @@ extern "C" { #define MAX_EXT_SIZE 135 #endif // CONFIG_AV1 && CONFIG_EXT_PARTITION +#if CONFIG_AV1 && CONFIG_LOOP_RESTORATION +#define EXTRAPREC_BITS 2 +#define EXTRAPREC_CLAMP_LIMIT(bd) (1 << ((bd) + 1 + EXTRAPREC_BITS)) +#endif + typedef void (*convolve_fn_t)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, diff --git a/third_party/aom/aom_dsp/aom_dsp.cmake b/third_party/aom/aom_dsp/aom_dsp.cmake index f00348cbc..5a49ae817 100644 --- a/third_party/aom/aom_dsp/aom_dsp.cmake +++ b/third_party/aom/aom_dsp/aom_dsp.cmake @@ -8,6 +8,9 @@ ## 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. ## +if (NOT AOM_AOM_DSP_AOM_DSP_CMAKE_) +set(AOM_AOM_DSP_AOM_DSP_CMAKE_ 1) + set(AOM_DSP_COMMON_SOURCES "${AOM_ROOT}/aom_dsp/aom_convolve.c" "${AOM_ROOT}/aom_dsp/aom_convolve.h" @@ -23,7 +26,6 @@ set(AOM_DSP_COMMON_SOURCES "${AOM_ROOT}/aom_dsp/loopfilter.c" "${AOM_ROOT}/aom_dsp/prob.c" "${AOM_ROOT}/aom_dsp/prob.h" - "${AOM_ROOT}/aom_dsp/sad.c" "${AOM_ROOT}/aom_dsp/simd/v128_intrinsics.h" "${AOM_ROOT}/aom_dsp/simd/v128_intrinsics_c.h" "${AOM_ROOT}/aom_dsp/simd/v256_intrinsics.h" @@ -62,8 +64,10 @@ set(AOM_DSP_COMMON_INTRIN_SSE4_1 set(AOM_DSP_COMMON_INTRIN_AVX2 "${AOM_ROOT}/aom_dsp/x86/aom_subpixel_8t_intrin_avx2.c" - "${AOM_ROOT}/aom_dsp/x86/fwd_txfm_avx2.c" - "${AOM_ROOT}/aom_dsp/x86/loopfilter_avx2.c") + "${AOM_ROOT}/aom_dsp/x86/loopfilter_avx2.c" + "${AOM_ROOT}/aom_dsp/x86/inv_txfm_avx2.c" + "${AOM_ROOT}/aom_dsp/x86/inv_txfm_common_avx2.h" + "${AOM_ROOT}/aom_dsp/x86/txfm_common_avx2.h") set(AOM_DSP_COMMON_ASM_NEON "${AOM_ROOT}/aom_dsp/arm/aom_convolve8_avg_neon_asm.asm" @@ -175,6 +179,8 @@ set(AOM_DSP_COMMON_INTRIN_MSA if (CONFIG_HIGHBITDEPTH) set(AOM_DSP_COMMON_ASM_SSE2 ${AOM_DSP_COMMON_ASM_SSE2} + "${AOM_ROOT}/aom_dsp/x86/aom_high_subpixel_8t_sse2.asm" + "${AOM_ROOT}/aom_dsp/x86/aom_high_subpixel_bilinear_sse2.asm" "${AOM_ROOT}/aom_dsp/x86/highbd_intrapred_sse2.asm") set(AOM_DSP_COMMON_INTRIN_SSE2 @@ -198,7 +204,7 @@ if (CONFIG_ANS) set(AOM_DSP_COMMON_SOURCES ${AOM_DSP_COMMON_SOURCES} "${AOM_ROOT}/aom_dsp/ans.h") -elseif (CONFIG_DAALA_EC) +else () set(AOM_DSP_COMMON_SOURCES ${AOM_DSP_COMMON_SOURCES} "${AOM_ROOT}/aom_dsp/entcode.c" @@ -221,7 +227,7 @@ if (CONFIG_AV1) "${AOM_ROOT}/aom_dsp/x86/inv_txfm_sse2.h") endif () -if (CONFIG_DECODERS) +if (CONFIG_AV1_DECODER) set(AOM_DSP_DECODER_SOURCES "${AOM_ROOT}/aom_dsp/binary_codes_reader.c" "${AOM_ROOT}/aom_dsp/binary_codes_reader.h" @@ -233,22 +239,17 @@ if (CONFIG_DECODERS) set(AOM_DSP_DECODER_SOURCES ${AOM_DSP_DECODER_SOURCES} "${AOM_ROOT}/aom_dsp/ansreader.h") - elseif (CONFIG_DAALA_EC) + else () set(AOM_DSP_DECODER_SOURCES ${AOM_DSP_DECODER_SOURCES} "${AOM_ROOT}/aom_dsp/daalaboolreader.c" "${AOM_ROOT}/aom_dsp/daalaboolreader.h" "${AOM_ROOT}/aom_dsp/entdec.c" "${AOM_ROOT}/aom_dsp/entdec.h") - else () - set(AOM_DSP_DECODER_SOURCES - ${AOM_DSP_DECODER_SOURCES} - "${AOM_ROOT}/aom_dsp/dkboolreader.c" - "${AOM_ROOT}/aom_dsp/dkboolreader.h") endif () endif () -if (CONFIG_ENCODERS) +if (CONFIG_AV1_ENCODER) set(AOM_DSP_ENCODER_SOURCES "${AOM_ROOT}/aom_dsp/binary_codes_writer.c" "${AOM_ROOT}/aom_dsp/binary_codes_writer.h" @@ -257,6 +258,7 @@ if (CONFIG_ENCODERS) "${AOM_ROOT}/aom_dsp/bitwriter_buffer.h" "${AOM_ROOT}/aom_dsp/psnr.c" "${AOM_ROOT}/aom_dsp/psnr.h" + "${AOM_ROOT}/aom_dsp/sad.c" "${AOM_ROOT}/aom_dsp/variance.c" "${AOM_ROOT}/aom_dsp/variance.h") @@ -282,6 +284,9 @@ if (CONFIG_ENCODERS) set(AOM_DSP_ENCODER_ASM_SSE4_1 "${AOM_ROOT}/aom_dsp/x86/sad_sse4.asm") set(AOM_DSP_ENCODER_INTRIN_AVX2 + "${AOM_ROOT}/aom_dsp/x86/fwd_dct32x32_impl_avx2.h" + "${AOM_ROOT}/aom_dsp/x86/fwd_txfm_avx2.c" + "${AOM_ROOT}/aom_dsp/x86/fwd_txfm_avx2.h" "${AOM_ROOT}/aom_dsp/x86/sad4d_avx2.c" "${AOM_ROOT}/aom_dsp/x86/sad_avx2.c" "${AOM_ROOT}/aom_dsp/x86/sad_impl_avx2.c" @@ -310,11 +315,6 @@ if (CONFIG_ENCODERS) "${AOM_ROOT}/aom_dsp/x86/variance_sse2.c" "${AOM_ROOT}/aom_dsp/x86/sum_squares_sse2.c") - set(AOM_DSP_ENCODER_INTRIN_SSSE3 - ${AOM_DSP_ENCODER_INTRIN_SSSE3} - "${AOM_ROOT}/aom_dsp/x86/masked_sad_intrin_ssse3.c" - "${AOM_ROOT}/aom_dsp/x86/masked_variance_intrin_ssse3.c") - set(AOM_DSP_ENCODER_ASM_SSSE3_X86_64 ${AOM_DSP_ENCODER_ASM_SSSE3_X86_64} "${AOM_ROOT}/aom_dsp/x86/avg_ssse3_x86_64.asm" @@ -325,7 +325,6 @@ if (CONFIG_ENCODERS) "${AOM_ROOT}/aom_dsp/x86/quantize_avx_x86_64.asm") set(AOM_DSP_ENCODER_INTRIN_MSA - "${AOM_ROOT}/aom_dsp/mips/avg_msa.c" "${AOM_ROOT}/aom_dsp/mips/sad_msa.c" "${AOM_ROOT}/aom_dsp/mips/subtract_msa.c" "${AOM_ROOT}/aom_dsp/mips/variance_msa.c" @@ -345,9 +344,7 @@ if (CONFIG_ENCODERS) "${AOM_ROOT}/aom_dsp/x86/highbd_sad4d_sse2.asm" "${AOM_ROOT}/aom_dsp/x86/highbd_sad_sse2.asm" "${AOM_ROOT}/aom_dsp/x86/highbd_subpel_variance_impl_sse2.asm" - "${AOM_ROOT}/aom_dsp/x86/highbd_variance_impl_sse2.asm" - "${AOM_ROOT}/aom_dsp/x86/aom_high_subpixel_8t_sse2.asm" - "${AOM_ROOT}/aom_dsp/x86/aom_high_subpixel_bilinear_sse2.asm") + "${AOM_ROOT}/aom_dsp/x86/highbd_variance_impl_sse2.asm") set(AOM_DSP_ENCODER_INTRIN_SSE2 ${AOM_DSP_ENCODER_INTRIN_SSE2} @@ -368,18 +365,13 @@ if (CONFIG_ENCODERS) "${AOM_ROOT}/aom_dsp/answriter.h" "${AOM_ROOT}/aom_dsp/buf_ans.c" "${AOM_ROOT}/aom_dsp/buf_ans.h") - elseif (CONFIG_DAALA_EC) + else () set(AOM_DSP_ENCODER_SOURCES ${AOM_DSP_ENCODER_SOURCES} "${AOM_ROOT}/aom_dsp/daalaboolwriter.c" "${AOM_ROOT}/aom_dsp/daalaboolwriter.h" "${AOM_ROOT}/aom_dsp/entenc.c" "${AOM_ROOT}/aom_dsp/entenc.h") - else () - set(AOM_DSP_ENCODER_SOURCES - ${AOM_DSP_ENCODER_SOURCES} - "${AOM_ROOT}/aom_dsp/dkboolwriter.c" - "${AOM_ROOT}/aom_dsp/dkboolwriter.h") endif () if (CONFIG_INTERNAL_STATS) @@ -392,6 +384,18 @@ if (CONFIG_ENCODERS) endif () endif () +if (CONFIG_LOOP_RESTORATION) + set(AOM_DSP_COMMON_INTRIN_SSE2 + ${AOM_DSP_COMMON_INTRIN_SSE2} + "${AOM_ROOT}/aom_dsp/x86/aom_convolve_hip_sse2.c") + + if (CONFIG_HIGHBITDEPTH) + set(AOM_DSP_COMMON_INTRIN_SSSE3 + ${AOM_DSP_COMMON_INTRIN_SSSE3} + "${AOM_ROOT}/aom_dsp/x86/aom_highbd_convolve_hip_ssse3.c") + endif () +endif () + if (CONFIG_MOTION_VAR) set(AOM_DSP_ENCODER_INTRIN_SSE4_1 ${AOM_DSP_ENCODER_INTRIN_SSE4_1} @@ -406,13 +410,13 @@ function (setup_aom_dsp_targets) set(AOM_LIB_TARGETS ${AOM_LIB_TARGETS} aom_dsp_common) target_sources(aom PUBLIC $) - if (CONFIG_DECODERS) + if (CONFIG_AV1_DECODER) add_library(aom_dsp_decoder OBJECT ${AOM_DSP_DECODER_SOURCES}) set(AOM_LIB_TARGETS ${AOM_LIB_TARGETS} aom_dsp_decoder) target_sources(aom PUBLIC $) endif () - if (CONFIG_ENCODERS) + if (CONFIG_AV1_ENCODER) add_library(aom_dsp_encoder OBJECT ${AOM_DSP_ENCODER_SOURCES}) set(AOM_LIB_TARGETS ${AOM_LIB_TARGETS} aom_dsp_encoder) target_sources(aom PUBLIC $) @@ -422,14 +426,14 @@ function (setup_aom_dsp_targets) add_asm_library("aom_dsp_common_sse2" "AOM_DSP_COMMON_ASM_SSE2" "aom") add_intrinsics_object_library("-msse2" "sse2" "aom_dsp_common" "AOM_DSP_COMMON_INTRIN_SSE2") - if (CONFIG_ENCODERS) + if (CONFIG_AV1_ENCODER) add_asm_library("aom_dsp_encoder_sse2" "AOM_DSP_ENCODER_ASM_SSE2" "aom") add_intrinsics_object_library("-msse2" "sse2" "aom_dsp_encoder" "AOM_DSP_ENCODER_INTRIN_SSE2") endif() endif () - if (HAVE_SSE3 AND CONFIG_ENCODERS) + if (HAVE_SSE3 AND CONFIG_AV1_ENCODER) add_asm_library("aom_dsp_encoder_sse3" "AOM_DSP_ENCODER_INTRIN_SSE3" "aom") endif () @@ -438,21 +442,19 @@ function (setup_aom_dsp_targets) add_intrinsics_object_library("-mssse3" "ssse3" "aom_dsp_common" "AOM_DSP_COMMON_INTRIN_SSSE3") - if (CONFIG_ENCODERS) + if (CONFIG_AV1_ENCODER) if ("${AOM_TARGET_CPU}" STREQUAL "x86_64") list(APPEND AOM_DSP_ENCODER_ASM_SSSE3 ${AOM_DSP_ENCODER_ASM_SSSE3_X86_64}) endif () add_asm_library("aom_dsp_encoder_ssse3" "AOM_DSP_ENCODER_ASM_SSSE3" "aom") - add_intrinsics_object_library("-mssse3" "ssse3" "aom_dsp_encoder" - "AOM_DSP_ENCODER_INTRIN_SSSE3") endif () endif () if (HAVE_SSE4_1) add_intrinsics_object_library("-msse4.1" "sse4_1" "aom_dsp_common" "AOM_DSP_COMMON_INTRIN_SSE4_1") - if (CONFIG_ENCODERS) + if (CONFIG_AV1_ENCODER) if (AOM_DSP_ENCODER_INTRIN_SSE4_1) add_intrinsics_object_library("-msse4.1" "sse4_1" "aom_dsp_encoder" "AOM_DSP_ENCODER_INTRIN_SSE4_1") @@ -463,14 +465,16 @@ function (setup_aom_dsp_targets) endif () if (HAVE_AVX AND "${AOM_TARGET_CPU}" STREQUAL "x86_64") - add_asm_library("aom_dsp_encoder_avx" "AOM_DSP_ENCODER_AVX_ASM_X86_64" - "aom") + if (CONFIG_AV1_ENCODER) + add_asm_library("aom_dsp_encoder_avx" "AOM_DSP_ENCODER_AVX_ASM_X86_64" + "aom") + endif () endif () if (HAVE_AVX2) add_intrinsics_object_library("-mavx2" "avx2" "aom_dsp_common" "AOM_DSP_COMMON_INTRIN_AVX2") - if (CONFIG_ENCODERS) + if (CONFIG_AV1_ENCODER) add_intrinsics_object_library("-mavx2" "avx2" "aom_dsp_encoder" "AOM_DSP_ENCODER_INTRIN_AVX2") endif () @@ -497,7 +501,7 @@ function (setup_aom_dsp_targets) if (HAVE_MSA) add_intrinsics_object_library("" "msa" "aom_dsp_common" "AOM_DSP_COMMON_INTRIN_MSA") - if (CONFIG_ENCODERS) + if (CONFIG_AV1_ENCODER) add_intrinsics_object_library("" "msa" "aom_dsp_encoder" "AOM_DSP_ENCODER_INTRIN_MSA") endif () @@ -507,3 +511,5 @@ function (setup_aom_dsp_targets) # $AOM_LIB_TARGETS. set(AOM_LIB_TARGETS ${AOM_LIB_TARGETS} PARENT_SCOPE) endfunction () + +endif () # AOM_AOM_DSP_AOM_DSP_CMAKE_ diff --git a/third_party/aom/aom_dsp/aom_dsp.mk b/third_party/aom/aom_dsp/aom_dsp.mk index 8c7241b83..6e2d5630e 100644 --- a/third_party/aom/aom_dsp/aom_dsp.mk +++ b/third_party/aom/aom_dsp/aom_dsp.mk @@ -22,19 +22,16 @@ DSP_SRCS-yes += prob.h DSP_SRCS-yes += prob.c DSP_SRCS-$(CONFIG_ANS) += ans.h -ifeq ($(CONFIG_ENCODERS),yes) +ifeq ($(CONFIG_AV1_ENCODER),yes) ifeq ($(CONFIG_ANS),yes) DSP_SRCS-yes += answriter.h DSP_SRCS-yes += buf_ans.h DSP_SRCS-yes += buf_ans.c -else ifeq ($(CONFIG_DAALA_EC),yes) +else DSP_SRCS-yes += entenc.c DSP_SRCS-yes += entenc.h DSP_SRCS-yes += daalaboolwriter.c DSP_SRCS-yes += daalaboolwriter.h -else -DSP_SRCS-yes += dkboolwriter.h -DSP_SRCS-yes += dkboolwriter.c endif DSP_SRCS-yes += bitwriter.h DSP_SRCS-yes += bitwriter_buffer.c @@ -49,17 +46,14 @@ DSP_SRCS-$(CONFIG_INTERNAL_STATS) += psnrhvs.c DSP_SRCS-$(CONFIG_INTERNAL_STATS) += fastssim.c endif -ifeq ($(CONFIG_DECODERS),yes) +ifeq ($(CONFIG_AV1_DECODER),yes) ifeq ($(CONFIG_ANS),yes) DSP_SRCS-yes += ansreader.h -else ifeq ($(CONFIG_DAALA_EC),yes) +else DSP_SRCS-yes += entdec.c DSP_SRCS-yes += entdec.h DSP_SRCS-yes += daalaboolreader.c DSP_SRCS-yes += daalaboolreader.h -else -DSP_SRCS-yes += dkboolreader.h -DSP_SRCS-yes += dkboolreader.c endif DSP_SRCS-yes += bitreader.h DSP_SRCS-yes += bitreader_buffer.c @@ -71,7 +65,7 @@ endif # intra predictions DSP_SRCS-yes += intrapred.c -ifeq ($(CONFIG_DAALA_EC),yes) +ifneq ($(CONFIG_ANS),yes) DSP_SRCS-yes += entcode.c DSP_SRCS-yes += entcode.h endif @@ -205,6 +199,7 @@ endif # CONFIG_HIGHBITDEPTH DSP_SRCS-yes += txfm_common.h DSP_SRCS-yes += x86/txfm_common_intrin.h DSP_SRCS-$(HAVE_SSE2) += x86/txfm_common_sse2.h +DSP_SRCS-$(HAVE_SSSE3) += x86/obmc_intrinsic_ssse3.h DSP_SRCS-$(HAVE_MSA) += mips/txfm_macros_msa.h # forward transform @@ -239,6 +234,8 @@ DSP_SRCS-$(HAVE_SSE2) += x86/inv_txfm_sse2.h DSP_SRCS-$(HAVE_SSE2) += x86/inv_txfm_sse2.c DSP_SRCS-$(HAVE_SSE2) += x86/inv_wht_sse2.asm DSP_SRCS-$(HAVE_SSSE3) += x86/inv_txfm_ssse3.c +DSP_SRCS-$(HAVE_AVX2) += x86/inv_txfm_common_avx2.h +DSP_SRCS-$(HAVE_AVX2) += x86/inv_txfm_avx2.c ifeq ($(HAVE_NEON_ASM),yes) DSP_SRCS-yes += arm/save_reg_neon$(ASM) @@ -278,6 +275,13 @@ DSP_SRCS-$(HAVE_DSPR2) += mips/itrans16_dspr2.c DSP_SRCS-$(HAVE_DSPR2) += mips/itrans32_dspr2.c DSP_SRCS-$(HAVE_DSPR2) += mips/itrans32_cols_dspr2.c endif # CONFIG_HIGHBITDEPTH + +ifeq ($(CONFIG_LOOP_RESTORATION),yes) +DSP_SRCS-$(HAVE_SSE2) += x86/aom_convolve_hip_sse2.c +ifeq ($(CONFIG_HIGHBITDEPTH),yes) +DSP_SRCS-$(HAVE_SSSE3) += x86/aom_highbd_convolve_hip_ssse3.c +endif +endif # CONFIG_LOOP_RESTORATION endif # CONFIG_AV1 # quantization @@ -298,7 +302,6 @@ endif DSP_SRCS-yes += avg.c DSP_SRCS-$(HAVE_SSE2) += x86/avg_intrin_sse2.c DSP_SRCS-$(HAVE_NEON) += arm/avg_neon.c -DSP_SRCS-$(HAVE_MSA) += mips/avg_msa.c DSP_SRCS-$(HAVE_NEON) += arm/hadamard_neon.c ifeq ($(ARCH_X86_64),yes) DSP_SRCS-$(HAVE_SSSE3) += x86/avg_ssse3_x86_64.asm @@ -317,11 +320,10 @@ DSP_SRCS-yes += sum_squares.c DSP_SRCS-$(HAVE_SSE2) += x86/sum_squares_sse2.c endif # CONFIG_AV1_ENCODER -ifeq ($(CONFIG_ENCODERS),yes) +ifeq ($(CONFIG_AV1_ENCODER),yes) DSP_SRCS-yes += sad.c DSP_SRCS-yes += subtract.c -DSP_SRCS-$(HAVE_MEDIA) += arm/sad_media$(ASM) DSP_SRCS-$(HAVE_NEON) += arm/sad4d_neon.c DSP_SRCS-$(HAVE_NEON) += arm/sad_neon.c DSP_SRCS-$(HAVE_NEON) += arm/subtract_neon.c @@ -364,18 +366,12 @@ DSP_SRCS-$(HAVE_SSE2) += x86/highbd_sad4d_sse2.asm DSP_SRCS-$(HAVE_SSE2) += x86/highbd_sad_sse2.asm endif # CONFIG_HIGHBITDEPTH -endif # CONFIG_ENCODERS +endif # CONFIG_AV1_ENCODER -ifneq ($(filter yes,$(CONFIG_ENCODERS)),) +ifneq ($(filter yes,$(CONFIG_AV1_ENCODER)),) DSP_SRCS-yes += variance.c DSP_SRCS-yes += variance.h -DSP_SRCS-$(HAVE_MEDIA) += arm/bilinear_filter_media$(ASM) -DSP_SRCS-$(HAVE_MEDIA) += arm/subpel_variance_media.c -DSP_SRCS-$(HAVE_MEDIA) += arm/variance_halfpixvar16x16_h_media$(ASM) -DSP_SRCS-$(HAVE_MEDIA) += arm/variance_halfpixvar16x16_hv_media$(ASM) -DSP_SRCS-$(HAVE_MEDIA) += arm/variance_halfpixvar16x16_v_media$(ASM) -DSP_SRCS-$(HAVE_MEDIA) += arm/variance_media$(ASM) DSP_SRCS-$(HAVE_NEON) += arm/subpel_variance_neon.c DSP_SRCS-$(HAVE_NEON) += arm/variance_neon.c @@ -402,7 +398,7 @@ DSP_SRCS-$(HAVE_SSE4_1) += x86/highbd_variance_sse4.c DSP_SRCS-$(HAVE_SSE2) += x86/highbd_variance_impl_sse2.asm DSP_SRCS-$(HAVE_SSE2) += x86/highbd_subpel_variance_impl_sse2.asm endif # CONFIG_HIGHBITDEPTH -endif # CONFIG_ENCODERS +endif # CONFIG_AV1_ENCODER DSP_SRCS-no += $(DSP_SRCS_REMOVE-yes) diff --git a/third_party/aom/aom_dsp/aom_dsp_common.h b/third_party/aom/aom_dsp/aom_dsp_common.h index 47ffbeb6c..82f9a95e9 100644 --- a/third_party/aom/aom_dsp/aom_dsp_common.h +++ b/third_party/aom/aom_dsp/aom_dsp_common.h @@ -31,6 +31,8 @@ extern "C" { #define AOMMIN(x, y) (((x) < (y)) ? (x) : (y)) #define AOMMAX(x, y) (((x) > (y)) ? (x) : (y)) +#define NELEMENTS(x) (sizeof((x)) / sizeof((x)[0])) + #define IMPLIES(a, b) (!(a) || (b)) // Logical 'a implies b' (or 'a -> b') #define IS_POWER_OF_TWO(x) (((x) & ((x)-1)) == 0) diff --git a/third_party/aom/aom_dsp/aom_dsp_rtcd_defs.pl b/third_party/aom/aom_dsp/aom_dsp_rtcd_defs.pl index b4ef0d92f..8047cbc09 100755 --- a/third_party/aom/aom_dsp/aom_dsp_rtcd_defs.pl +++ b/third_party/aom/aom_dsp/aom_dsp_rtcd_defs.pl @@ -49,6 +49,9 @@ if (aom_config("CONFIG_TX64X64") eq "yes") { @pred_names = qw/dc dc_top dc_left dc_128 v h d207e d63e d45e d117 d135 d153/; if (aom_config("CONFIG_ALT_INTRA") eq "yes") { push @pred_names, qw/paeth smooth/; + if (aom_config("CONFIG_SMOOTH_HV") eq "yes") { + push @pred_names, qw/smooth_v smooth_h/; + } } else { push @pred_names, 'tm'; } @@ -168,10 +171,14 @@ if (aom_config("CONFIG_LOOP_RESTORATION") eq "yes") { add_proto qw/void aom_convolve8_add_src/, "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"; add_proto qw/void aom_convolve8_add_src_horiz/, "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"; add_proto qw/void aom_convolve8_add_src_vert/, "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"; + add_proto qw/void aom_convolve8_add_src_hip/, "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"; + add_proto qw/void aom_convolve8_add_src_horiz_hip/, "const uint8_t *src, ptrdiff_t src_stride, uint16_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"; + add_proto qw/void aom_convolve8_add_src_vert_hip/, "const uint16_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"; specialize qw/aom_convolve8_add_src ssse3/; specialize qw/aom_convolve8_add_src_horiz ssse3/; specialize qw/aom_convolve8_add_src_vert ssse3/; + specialize qw/aom_convolve8_add_src_hip sse2/; } # CONFIG_LOOP_RESTORATION # TODO(any): These need to be extended to up to 128x128 block sizes @@ -215,8 +222,12 @@ if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") { add_proto qw/void aom_highbd_convolve8_add_src/, "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 bps"; add_proto qw/void aom_highbd_convolve8_add_src_horiz/, "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 bps"; add_proto qw/void aom_highbd_convolve8_add_src_vert/, "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 bps"; + add_proto qw/void aom_highbd_convolve8_add_src_hip/, "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 bps"; + add_proto qw/void aom_highbd_convolve8_add_src_horiz_hip/, "const uint8_t *src, ptrdiff_t src_stride, uint16_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 bps"; + add_proto qw/void aom_highbd_convolve8_add_src_vert_hip/, "const uint16_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 bps"; specialize qw/aom_highbd_convolve8_add_src/, "$sse2_x86_64"; + specialize qw/aom_highbd_convolve8_add_src_hip ssse3/; # The _horiz/_vert functions are currently unused, so we don't bother # specialising them. } # CONFIG_LOOP_RESTORATION @@ -434,29 +445,30 @@ if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") { specialize qw/aom_idct8x8_1_add sse2/; add_proto qw/void aom_idct16x16_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride"; - specialize qw/aom_idct16x16_256_add sse2/; + specialize qw/aom_idct16x16_256_add sse2 avx2/; add_proto qw/void aom_idct16x16_38_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride"; + specialize qw/aom_idct16x16_38_add avx2/; add_proto qw/void aom_idct16x16_10_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride"; - specialize qw/aom_idct16x16_10_add sse2/; + specialize qw/aom_idct16x16_10_add sse2 avx2/; add_proto qw/void aom_idct16x16_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride"; - specialize qw/aom_idct16x16_1_add sse2/; + specialize qw/aom_idct16x16_1_add sse2 avx2/; add_proto qw/void aom_idct32x32_1024_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride"; - specialize qw/aom_idct32x32_1024_add sse2 ssse3/; + specialize qw/aom_idct32x32_1024_add sse2 ssse3 avx2/; add_proto qw/void aom_idct32x32_135_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride"; - specialize qw/aom_idct32x32_135_add sse2 ssse3/; + specialize qw/aom_idct32x32_135_add sse2 ssse3 avx2/; # Need to add 135 eob idct32x32 implementations. $aom_idct32x32_135_add_sse2=aom_idct32x32_1024_add_sse2; add_proto qw/void aom_idct32x32_34_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride"; - specialize qw/aom_idct32x32_34_add sse2 ssse3/; + specialize qw/aom_idct32x32_34_add sse2 ssse3 avx2/; add_proto qw/void aom_idct32x32_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride"; - specialize qw/aom_idct32x32_1_add sse2/; + specialize qw/aom_idct32x32_1_add sse2 avx2/; add_proto qw/void aom_highbd_idct4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd"; specialize qw/aom_highbd_idct4x4_16_add sse2/; @@ -479,21 +491,22 @@ if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") { specialize qw/aom_idct8x8_12_add sse2 ssse3 neon dspr2 msa/; add_proto qw/void aom_idct16x16_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride"; - specialize qw/aom_idct16x16_1_add sse2 neon dspr2 msa/; + specialize qw/aom_idct16x16_1_add sse2 avx2 neon dspr2 msa/; add_proto qw/void aom_idct16x16_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride"; - specialize qw/aom_idct16x16_256_add sse2 neon dspr2 msa/; + specialize qw/aom_idct16x16_256_add sse2 avx2 neon dspr2 msa/; add_proto qw/void aom_idct16x16_38_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride"; + specialize qw/aom_idct16x16_38_add avx2/; add_proto qw/void aom_idct16x16_10_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride"; - specialize qw/aom_idct16x16_10_add sse2 neon dspr2 msa/; + specialize qw/aom_idct16x16_10_add sse2 avx2 neon dspr2 msa/; add_proto qw/void aom_idct32x32_1024_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride"; - specialize qw/aom_idct32x32_1024_add sse2 ssse3 neon dspr2 msa/; + specialize qw/aom_idct32x32_1024_add sse2 ssse3 avx2 neon dspr2 msa/; add_proto qw/void aom_idct32x32_135_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride"; - specialize qw/aom_idct32x32_135_add sse2 ssse3 neon dspr2 msa/; + specialize qw/aom_idct32x32_135_add sse2 ssse3 avx2 neon dspr2 msa/; # Need to add 135 eob idct32x32 implementations. $aom_idct32x32_135_add_sse2=aom_idct32x32_1024_add_sse2; $aom_idct32x32_135_add_neon=aom_idct32x32_1024_add_neon; @@ -501,12 +514,12 @@ if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") { $aom_idct32x32_135_add_msa=aom_idct32x32_1024_add_msa; add_proto qw/void aom_idct32x32_34_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride"; - specialize qw/aom_idct32x32_34_add sse2 ssse3 neon dspr2 msa/; + specialize qw/aom_idct32x32_34_add sse2 ssse3 avx2 neon dspr2 msa/; # Need to add 34 eob idct32x32 neon implementation. $aom_idct32x32_34_add_neon=aom_idct32x32_1024_add_neon; add_proto qw/void aom_idct32x32_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride"; - specialize qw/aom_idct32x32_1_add sse2 neon dspr2 msa/; + specialize qw/aom_idct32x32_1_add sse2 avx2 neon dspr2 msa/; add_proto qw/void aom_iwht4x4_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride"; specialize qw/aom_iwht4x4_1_add msa/; @@ -578,7 +591,7 @@ if (aom_config("CONFIG_AV1") eq "yes") { } } # CONFIG_AV1 -if (aom_config("CONFIG_ENCODERS") eq "yes") { +if (aom_config("CONFIG_AV1_ENCODER") eq "yes") { # # Block subtraction # @@ -604,13 +617,8 @@ if (aom_config("CONFIG_AV1_ENCODER") eq "yes") { # # Avg # - add_proto qw/unsigned int aom_avg_8x8/, "const uint8_t *, int p"; specialize qw/aom_avg_8x8 sse2 neon msa/; - add_proto qw/unsigned int aom_avg_4x4/, "const uint8_t *, int p"; - specialize qw/aom_avg_4x4 sse2 neon msa/; if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") { - add_proto qw/unsigned int aom_highbd_avg_8x8/, "const uint8_t *, int p"; - add_proto qw/unsigned int aom_highbd_avg_4x4/, "const uint8_t *, int p"; add_proto qw/void aom_highbd_subtract_block/, "int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride, int bd"; specialize qw/aom_highbd_subtract_block sse2/; } @@ -652,22 +660,22 @@ foreach (@block_sizes) { add_proto qw/unsigned int/, "aom_sad${w}x${h}_avg", "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred"; } -specialize qw/aom_sad128x128 avx2 sse2/; -specialize qw/aom_sad128x64 avx2 sse2/; -specialize qw/aom_sad64x128 avx2 sse2/; -specialize qw/aom_sad64x64 avx2 neon msa sse2/; -specialize qw/aom_sad64x32 avx2 msa sse2/; -specialize qw/aom_sad32x64 avx2 msa sse2/; -specialize qw/aom_sad32x32 avx2 neon msa sse2/; -specialize qw/aom_sad32x16 avx2 msa sse2/; -specialize qw/aom_sad16x32 msa sse2/; -specialize qw/aom_sad16x16 media neon msa sse2/; -specialize qw/aom_sad16x8 neon msa sse2/; -specialize qw/aom_sad8x16 neon msa sse2/; -specialize qw/aom_sad8x8 neon msa sse2/; -specialize qw/aom_sad8x4 msa sse2/; -specialize qw/aom_sad4x8 msa sse2/; -specialize qw/aom_sad4x4 neon msa sse2/; +specialize qw/aom_sad128x128 avx2 sse2/; +specialize qw/aom_sad128x64 avx2 sse2/; +specialize qw/aom_sad64x128 avx2 sse2/; +specialize qw/aom_sad64x64 avx2 neon msa sse2/; +specialize qw/aom_sad64x32 avx2 msa sse2/; +specialize qw/aom_sad32x64 avx2 msa sse2/; +specialize qw/aom_sad32x32 avx2 neon msa sse2/; +specialize qw/aom_sad32x16 avx2 msa sse2/; +specialize qw/aom_sad16x32 msa sse2/; +specialize qw/aom_sad16x16 neon msa sse2/; +specialize qw/aom_sad16x8 neon msa sse2/; +specialize qw/aom_sad8x16 neon msa sse2/; +specialize qw/aom_sad8x8 neon msa sse2/; +specialize qw/aom_sad8x4 msa sse2/; +specialize qw/aom_sad4x8 msa sse2/; +specialize qw/aom_sad4x4 neon msa sse2/; specialize qw/aom_sad128x128_avg avx2 sse2/; specialize qw/aom_sad128x64_avg avx2 sse2/; @@ -727,14 +735,14 @@ if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") { if (aom_config("CONFIG_EXT_INTER") eq "yes") { foreach (@block_sizes) { ($w, $h) = @$_; - add_proto qw/unsigned int/, "aom_masked_sad${w}x${h}", "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *mask, int mask_stride"; + add_proto qw/unsigned int/, "aom_masked_sad${w}x${h}", "const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask"; specialize "aom_masked_sad${w}x${h}", qw/ssse3/; } if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") { foreach (@block_sizes) { ($w, $h) = @$_; - add_proto qw/unsigned int/, "aom_highbd_masked_sad${w}x${h}", "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *mask, int mask_stride"; + add_proto qw/unsigned int/, "aom_highbd_masked_sad${w}x${h}", "const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask"; specialize "aom_highbd_masked_sad${w}x${h}", qw/ssse3/; } } @@ -876,9 +884,9 @@ if (aom_config("CONFIG_INTERNAL_STATS") eq "yes") { add_proto qw/void aom_highbd_ssim_parms_8x8/, "const uint16_t *s, int sp, const uint16_t *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"; } } -} # CONFIG_ENCODERS +} # CONFIG_AV1_ENCODER -if (aom_config("CONFIG_ENCODERS") eq "yes") { +if (aom_config("CONFIG_AV1_ENCODER") eq "yes") { # # Specialty Variance @@ -896,10 +904,10 @@ add_proto qw/unsigned int aom_mse16x8/, "const uint8_t *src_ptr, int source_str add_proto qw/unsigned int aom_mse8x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int recon_stride, unsigned int *sse"; add_proto qw/unsigned int aom_mse8x8/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int recon_stride, unsigned int *sse"; -specialize qw/aom_mse16x16 sse2 avx2 media neon msa/; -specialize qw/aom_mse16x8 sse2 msa/; -specialize qw/aom_mse8x16 sse2 msa/; -specialize qw/aom_mse8x8 sse2 msa/; +specialize qw/aom_mse16x16 sse2 avx2 neon msa/; +specialize qw/aom_mse16x8 sse2 msa/; +specialize qw/aom_mse8x16 sse2 msa/; +specialize qw/aom_mse8x8 sse2 msa/; if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") { foreach $bd (8, 10, 12) { @@ -956,33 +964,33 @@ foreach (@block_sizes) { add_proto qw/uint32_t/, "aom_sub_pixel_avg_variance${w}x${h}", "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred"; } -specialize qw/aom_variance64x64 sse2 avx2 neon msa/; -specialize qw/aom_variance64x32 sse2 avx2 neon msa/; -specialize qw/aom_variance32x64 sse2 neon msa/; -specialize qw/aom_variance32x32 sse2 avx2 neon msa/; -specialize qw/aom_variance32x16 sse2 avx2 msa/; -specialize qw/aom_variance16x32 sse2 msa/; -specialize qw/aom_variance16x16 sse2 avx2 media neon msa/; -specialize qw/aom_variance16x8 sse2 neon msa/; -specialize qw/aom_variance8x16 sse2 neon msa/; -specialize qw/aom_variance8x8 sse2 media neon msa/; -specialize qw/aom_variance8x4 sse2 msa/; -specialize qw/aom_variance4x8 sse2 msa/; -specialize qw/aom_variance4x4 sse2 msa/; - -specialize qw/aom_sub_pixel_variance64x64 avx2 neon msa sse2 ssse3/; -specialize qw/aom_sub_pixel_variance64x32 msa sse2 ssse3/; -specialize qw/aom_sub_pixel_variance32x64 msa sse2 ssse3/; -specialize qw/aom_sub_pixel_variance32x32 avx2 neon msa sse2 ssse3/; -specialize qw/aom_sub_pixel_variance32x16 msa sse2 ssse3/; -specialize qw/aom_sub_pixel_variance16x32 msa sse2 ssse3/; -specialize qw/aom_sub_pixel_variance16x16 media neon msa sse2 ssse3/; -specialize qw/aom_sub_pixel_variance16x8 msa sse2 ssse3/; -specialize qw/aom_sub_pixel_variance8x16 msa sse2 ssse3/; -specialize qw/aom_sub_pixel_variance8x8 media neon msa sse2 ssse3/; -specialize qw/aom_sub_pixel_variance8x4 msa sse2 ssse3/; -specialize qw/aom_sub_pixel_variance4x8 msa sse2 ssse3/; -specialize qw/aom_sub_pixel_variance4x4 msa sse2 ssse3/; +specialize qw/aom_variance64x64 sse2 avx2 neon msa/; +specialize qw/aom_variance64x32 sse2 avx2 neon msa/; +specialize qw/aom_variance32x64 sse2 neon msa/; +specialize qw/aom_variance32x32 sse2 avx2 neon msa/; +specialize qw/aom_variance32x16 sse2 avx2 msa/; +specialize qw/aom_variance16x32 sse2 msa/; +specialize qw/aom_variance16x16 sse2 avx2 neon msa/; +specialize qw/aom_variance16x8 sse2 neon msa/; +specialize qw/aom_variance8x16 sse2 neon msa/; +specialize qw/aom_variance8x8 sse2 neon msa/; +specialize qw/aom_variance8x4 sse2 msa/; +specialize qw/aom_variance4x8 sse2 msa/; +specialize qw/aom_variance4x4 sse2 msa/; + +specialize qw/aom_sub_pixel_variance64x64 avx2 neon msa sse2 ssse3/; +specialize qw/aom_sub_pixel_variance64x32 msa sse2 ssse3/; +specialize qw/aom_sub_pixel_variance32x64 msa sse2 ssse3/; +specialize qw/aom_sub_pixel_variance32x32 avx2 neon msa sse2 ssse3/; +specialize qw/aom_sub_pixel_variance32x16 msa sse2 ssse3/; +specialize qw/aom_sub_pixel_variance16x32 msa sse2 ssse3/; +specialize qw/aom_sub_pixel_variance16x16 neon msa sse2 ssse3/; +specialize qw/aom_sub_pixel_variance16x8 msa sse2 ssse3/; +specialize qw/aom_sub_pixel_variance8x16 msa sse2 ssse3/; +specialize qw/aom_sub_pixel_variance8x8 neon msa sse2 ssse3/; +specialize qw/aom_sub_pixel_variance8x4 msa sse2 ssse3/; +specialize qw/aom_sub_pixel_variance4x8 msa sse2 ssse3/; +specialize qw/aom_sub_pixel_variance4x4 msa sse2 ssse3/; specialize qw/aom_sub_pixel_avg_variance64x64 avx2 msa sse2 ssse3/; specialize qw/aom_sub_pixel_avg_variance64x32 msa sse2 ssse3/; @@ -1034,19 +1042,15 @@ if (aom_config("CONFIG_EXT_INTER") eq "yes") { # foreach (@block_sizes) { ($w, $h) = @$_; - add_proto qw/unsigned int/, "aom_masked_variance${w}x${h}", "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *mask, int mask_stride, unsigned int *sse"; - add_proto qw/unsigned int/, "aom_masked_sub_pixel_variance${w}x${h}", "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, const uint8_t *mask, int mask_stride, unsigned int *sse"; - specialize "aom_masked_variance${w}x${h}", qw/ssse3/; + add_proto qw/unsigned int/, "aom_masked_sub_pixel_variance${w}x${h}", "const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse"; specialize "aom_masked_sub_pixel_variance${w}x${h}", qw/ssse3/; } if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") { - foreach $bd ("_", "_10_", "_12_") { + foreach $bd ("_8_", "_10_", "_12_") { foreach (@block_sizes) { ($w, $h) = @$_; - add_proto qw/unsigned int/, "aom_highbd${bd}masked_variance${w}x${h}", "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *mask, int mask_stride, unsigned int *sse"; - add_proto qw/unsigned int/, "aom_highbd${bd}masked_sub_pixel_variance${w}x${h}", "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, const uint8_t *m, int m_stride, unsigned int *sse"; - specialize "aom_highbd${bd}masked_variance${w}x${h}", qw/ssse3/; + add_proto qw/unsigned int/, "aom_highbd${bd}masked_sub_pixel_variance${w}x${h}", "const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse"; specialize "aom_highbd${bd}masked_sub_pixel_variance${w}x${h}", qw/ssse3/; } } @@ -1119,13 +1123,13 @@ add_proto qw/uint32_t aom_sub_pixel_avg_variance4x4/, "const uint8_t *src_ptr, i # Specialty Subpixel # add_proto qw/uint32_t aom_variance_halfpixvar16x16_h/, "const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int ref_stride, uint32_t *sse"; - specialize qw/aom_variance_halfpixvar16x16_h sse2 media/; + specialize qw/aom_variance_halfpixvar16x16_h sse2/; add_proto qw/uint32_t aom_variance_halfpixvar16x16_v/, "const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int ref_stride, uint32_t *sse"; - specialize qw/aom_variance_halfpixvar16x16_v sse2 media/; + specialize qw/aom_variance_halfpixvar16x16_v sse2/; add_proto qw/uint32_t aom_variance_halfpixvar16x16_hv/, "const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int ref_stride, uint32_t *sse"; - specialize qw/aom_variance_halfpixvar16x16_hv sse2 media/; + specialize qw/aom_variance_halfpixvar16x16_hv sse2/; # # Comp Avg @@ -1490,6 +1494,15 @@ if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") { } # CONFIG_HIGHBITDEPTH -} # CONFIG_ENCODERS +if (aom_config("CONFIG_EXT_INTER") eq "yes") { + add_proto qw/void aom_comp_mask_pred/, "uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask"; + add_proto qw/void aom_comp_mask_upsampled_pred/, "uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask"; + if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") { + add_proto qw/void aom_highbd_comp_mask_pred/, "uint16_t *comp_pred, const uint8_t *pred8, int width, int height, const uint8_t *ref8, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask"; + add_proto qw/void aom_highbd_comp_mask_upsampled_pred/, "uint16_t *comp_pred, const uint8_t *pred8, int width, int height, const uint8_t *ref8, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask"; + } +} + +} # CONFIG_AV1_ENCODER 1; diff --git a/third_party/aom/aom_dsp/arm/avg_neon.c b/third_party/aom/aom_dsp/arm/avg_neon.c index e730ccbcc..6ff760017 100644 --- a/third_party/aom/aom_dsp/arm/avg_neon.c +++ b/third_party/aom/aom_dsp/arm/avg_neon.c @@ -25,44 +25,6 @@ static INLINE unsigned int horizontal_add_u16x8(const uint16x8_t v_16x8) { return vget_lane_u32(c, 0); } -unsigned int aom_avg_4x4_neon(const uint8_t *s, int p) { - uint16x8_t v_sum; - uint32x2_t v_s0 = vdup_n_u32(0); - uint32x2_t v_s1 = vdup_n_u32(0); - v_s0 = vld1_lane_u32((const uint32_t *)s, v_s0, 0); - v_s0 = vld1_lane_u32((const uint32_t *)(s + p), v_s0, 1); - v_s1 = vld1_lane_u32((const uint32_t *)(s + 2 * p), v_s1, 0); - v_s1 = vld1_lane_u32((const uint32_t *)(s + 3 * p), v_s1, 1); - v_sum = vaddl_u8(vreinterpret_u8_u32(v_s0), vreinterpret_u8_u32(v_s1)); - return (horizontal_add_u16x8(v_sum) + 8) >> 4; -} - -unsigned int aom_avg_8x8_neon(const uint8_t *s, int p) { - uint8x8_t v_s0 = vld1_u8(s); - const uint8x8_t v_s1 = vld1_u8(s + p); - uint16x8_t v_sum = vaddl_u8(v_s0, v_s1); - - v_s0 = vld1_u8(s + 2 * p); - v_sum = vaddw_u8(v_sum, v_s0); - - v_s0 = vld1_u8(s + 3 * p); - v_sum = vaddw_u8(v_sum, v_s0); - - v_s0 = vld1_u8(s + 4 * p); - v_sum = vaddw_u8(v_sum, v_s0); - - v_s0 = vld1_u8(s + 5 * p); - v_sum = vaddw_u8(v_sum, v_s0); - - v_s0 = vld1_u8(s + 6 * p); - v_sum = vaddw_u8(v_sum, v_s0); - - v_s0 = vld1_u8(s + 7 * p); - v_sum = vaddw_u8(v_sum, v_s0); - - return (horizontal_add_u16x8(v_sum) + 32) >> 6; -} - // coeff: 16 bits, dynamic range [-32640, 32640]. // length: value range {16, 64, 256, 1024}. int aom_satd_neon(const int16_t *coeff, int length) { diff --git a/third_party/aom/aom_dsp/arm/bilinear_filter_media.asm b/third_party/aom/aom_dsp/arm/bilinear_filter_media.asm deleted file mode 100644 index 17b7d25f9..000000000 --- a/third_party/aom/aom_dsp/arm/bilinear_filter_media.asm +++ /dev/null @@ -1,240 +0,0 @@ -; -; 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. -; - -; - - - EXPORT |aom_filter_block2d_bil_first_pass_media| - EXPORT |aom_filter_block2d_bil_second_pass_media| - - AREA |.text|, CODE, READONLY ; name this block of code - -;------------------------------------- -; r0 unsigned char *src_ptr, -; r1 unsigned short *dst_ptr, -; r2 unsigned int src_pitch, -; r3 unsigned int height, -; stack unsigned int width, -; stack const short *aom_filter -;------------------------------------- -; The output is transposed stroed in output array to make it easy for second pass filtering. -|aom_filter_block2d_bil_first_pass_media| PROC - stmdb sp!, {r4 - r11, lr} - - ldr r11, [sp, #40] ; aom_filter address - ldr r4, [sp, #36] ; width - - mov r12, r3 ; outer-loop counter - - add r7, r2, r4 ; preload next row - pld [r0, r7] - - sub r2, r2, r4 ; src increment for height loop - - ldr r5, [r11] ; load up filter coefficients - - mov r3, r3, lsl #1 ; height*2 - add r3, r3, #2 ; plus 2 to make output buffer 4-bit aligned since height is actually (height+1) - - mov r11, r1 ; save dst_ptr for each row - - cmp r5, #128 ; if filter coef = 128, then skip the filter - beq bil_null_1st_filter - -|bil_height_loop_1st_v6| - ldrb r6, [r0] ; load source data - ldrb r7, [r0, #1] - ldrb r8, [r0, #2] - mov lr, r4, lsr #2 ; 4-in-parellel loop counter - -|bil_width_loop_1st_v6| - ldrb r9, [r0, #3] - ldrb r10, [r0, #4] - - pkhbt r6, r6, r7, lsl #16 ; src[1] | src[0] - pkhbt r7, r7, r8, lsl #16 ; src[2] | src[1] - - smuad r6, r6, r5 ; apply the filter - pkhbt r8, r8, r9, lsl #16 ; src[3] | src[2] - smuad r7, r7, r5 - pkhbt r9, r9, r10, lsl #16 ; src[4] | src[3] - - smuad r8, r8, r5 - smuad r9, r9, r5 - - add r0, r0, #4 - subs lr, lr, #1 - - add r6, r6, #0x40 ; round_shift_and_clamp - add r7, r7, #0x40 - usat r6, #16, r6, asr #7 - usat r7, #16, r7, asr #7 - - strh r6, [r1], r3 ; result is transposed and stored - - add r8, r8, #0x40 ; round_shift_and_clamp - strh r7, [r1], r3 - add r9, r9, #0x40 - usat r8, #16, r8, asr #7 - usat r9, #16, r9, asr #7 - - strh r8, [r1], r3 ; result is transposed and stored - - ldrneb r6, [r0] ; load source data - strh r9, [r1], r3 - - ldrneb r7, [r0, #1] - ldrneb r8, [r0, #2] - - bne bil_width_loop_1st_v6 - - add r0, r0, r2 ; move to next input row - subs r12, r12, #1 - - add r9, r2, r4, lsl #1 ; adding back block width - pld [r0, r9] ; preload next row - - add r11, r11, #2 ; move over to next column - mov r1, r11 - - bne bil_height_loop_1st_v6 - - ldmia sp!, {r4 - r11, pc} - -|bil_null_1st_filter| -|bil_height_loop_null_1st| - mov lr, r4, lsr #2 ; loop counter - -|bil_width_loop_null_1st| - ldrb r6, [r0] ; load data - ldrb r7, [r0, #1] - ldrb r8, [r0, #2] - ldrb r9, [r0, #3] - - strh r6, [r1], r3 ; store it to immediate buffer - add r0, r0, #4 - strh r7, [r1], r3 - subs lr, lr, #1 - strh r8, [r1], r3 - strh r9, [r1], r3 - - bne bil_width_loop_null_1st - - subs r12, r12, #1 - add r0, r0, r2 ; move to next input line - add r11, r11, #2 ; move over to next column - mov r1, r11 - - bne bil_height_loop_null_1st - - ldmia sp!, {r4 - r11, pc} - - ENDP ; |aom_filter_block2d_bil_first_pass_media| - - -;--------------------------------- -; r0 unsigned short *src_ptr, -; r1 unsigned char *dst_ptr, -; r2 int dst_pitch, -; r3 unsigned int height, -; stack unsigned int width, -; stack const short *aom_filter -;--------------------------------- -|aom_filter_block2d_bil_second_pass_media| PROC - stmdb sp!, {r4 - r11, lr} - - ldr r11, [sp, #40] ; aom_filter address - ldr r4, [sp, #36] ; width - - ldr r5, [r11] ; load up filter coefficients - mov r12, r4 ; outer-loop counter = width, since we work on transposed data matrix - mov r11, r1 - - cmp r5, #128 ; if filter coef = 128, then skip the filter - beq bil_null_2nd_filter - -|bil_height_loop_2nd| - ldr r6, [r0] ; load the data - ldr r8, [r0, #4] - ldrh r10, [r0, #8] - mov lr, r3, lsr #2 ; loop counter - -|bil_width_loop_2nd| - pkhtb r7, r6, r8 ; src[1] | src[2] - pkhtb r9, r8, r10 ; src[3] | src[4] - - smuad r6, r6, r5 ; apply filter - smuad r8, r8, r5 ; apply filter - - subs lr, lr, #1 - - smuadx r7, r7, r5 ; apply filter - smuadx r9, r9, r5 ; apply filter - - add r0, r0, #8 - - add r6, r6, #0x40 ; round_shift_and_clamp - add r7, r7, #0x40 - usat r6, #8, r6, asr #7 - usat r7, #8, r7, asr #7 - strb r6, [r1], r2 ; the result is transposed back and stored - - add r8, r8, #0x40 ; round_shift_and_clamp - strb r7, [r1], r2 - add r9, r9, #0x40 - usat r8, #8, r8, asr #7 - usat r9, #8, r9, asr #7 - strb r8, [r1], r2 ; the result is transposed back and stored - - ldrne r6, [r0] ; load data - strb r9, [r1], r2 - ldrne r8, [r0, #4] - ldrneh r10, [r0, #8] - - bne bil_width_loop_2nd - - subs r12, r12, #1 - add r0, r0, #4 ; update src for next row - add r11, r11, #1 - mov r1, r11 - - bne bil_height_loop_2nd - ldmia sp!, {r4 - r11, pc} - -|bil_null_2nd_filter| -|bil_height_loop_null_2nd| - mov lr, r3, lsr #2 - -|bil_width_loop_null_2nd| - ldr r6, [r0], #4 ; load data - subs lr, lr, #1 - ldr r8, [r0], #4 - - strb r6, [r1], r2 ; store data - mov r7, r6, lsr #16 - strb r7, [r1], r2 - mov r9, r8, lsr #16 - strb r8, [r1], r2 - strb r9, [r1], r2 - - bne bil_width_loop_null_2nd - - subs r12, r12, #1 - add r0, r0, #4 - add r11, r11, #1 - mov r1, r11 - - bne bil_height_loop_null_2nd - - ldmia sp!, {r4 - r11, pc} - ENDP ; |aom_filter_block2d_second_pass_media| - - END diff --git a/third_party/aom/aom_dsp/arm/sad_media.asm b/third_party/aom/aom_dsp/arm/sad_media.asm deleted file mode 100644 index 49ddb6764..000000000 --- a/third_party/aom/aom_dsp/arm/sad_media.asm +++ /dev/null @@ -1,98 +0,0 @@ -; -; 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. -; - -; - - - EXPORT |aom_sad16x16_media| - - ARM - REQUIRE8 - PRESERVE8 - - AREA ||.text||, CODE, READONLY, ALIGN=2 - -; r0 const unsigned char *src_ptr -; r1 int src_stride -; r2 const unsigned char *ref_ptr -; r3 int ref_stride -|aom_sad16x16_media| PROC - stmfd sp!, {r4-r12, lr} - - pld [r0, r1, lsl #0] - pld [r2, r3, lsl #0] - pld [r0, r1, lsl #1] - pld [r2, r3, lsl #1] - - mov r4, #0 ; sad = 0; - mov r5, #8 ; loop count - -loop - ; 1st row - ldr r6, [r0, #0x0] ; load 4 src pixels (1A) - ldr r8, [r2, #0x0] ; load 4 ref pixels (1A) - ldr r7, [r0, #0x4] ; load 4 src pixels (1A) - ldr r9, [r2, #0x4] ; load 4 ref pixels (1A) - ldr r10, [r0, #0x8] ; load 4 src pixels (1B) - ldr r11, [r0, #0xC] ; load 4 src pixels (1B) - - usada8 r4, r8, r6, r4 ; calculate sad for 4 pixels - usad8 r8, r7, r9 ; calculate sad for 4 pixels - - ldr r12, [r2, #0x8] ; load 4 ref pixels (1B) - ldr lr, [r2, #0xC] ; load 4 ref pixels (1B) - - add r0, r0, r1 ; set src pointer to next row - add r2, r2, r3 ; set dst pointer to next row - - pld [r0, r1, lsl #1] - pld [r2, r3, lsl #1] - - usada8 r4, r10, r12, r4 ; calculate sad for 4 pixels - usada8 r8, r11, lr, r8 ; calculate sad for 4 pixels - - ldr r6, [r0, #0x0] ; load 4 src pixels (2A) - ldr r7, [r0, #0x4] ; load 4 src pixels (2A) - add r4, r4, r8 ; add partial sad values - - ; 2nd row - ldr r8, [r2, #0x0] ; load 4 ref pixels (2A) - ldr r9, [r2, #0x4] ; load 4 ref pixels (2A) - ldr r10, [r0, #0x8] ; load 4 src pixels (2B) - ldr r11, [r0, #0xC] ; load 4 src pixels (2B) - - usada8 r4, r6, r8, r4 ; calculate sad for 4 pixels - usad8 r8, r7, r9 ; calculate sad for 4 pixels - - ldr r12, [r2, #0x8] ; load 4 ref pixels (2B) - ldr lr, [r2, #0xC] ; load 4 ref pixels (2B) - - add r0, r0, r1 ; set src pointer to next row - add r2, r2, r3 ; set dst pointer to next row - - usada8 r4, r10, r12, r4 ; calculate sad for 4 pixels - usada8 r8, r11, lr, r8 ; calculate sad for 4 pixels - - pld [r0, r1, lsl #1] - pld [r2, r3, lsl #1] - - subs r5, r5, #1 ; decrement loop counter - add r4, r4, r8 ; add partial sad values - - bne loop - - mov r0, r4 ; return sad - ldmfd sp!, {r4-r12, pc} - - ENDP - - END - diff --git a/third_party/aom/aom_dsp/arm/subpel_variance_media.c b/third_party/aom/aom_dsp/arm/subpel_variance_media.c deleted file mode 100644 index 46ec028d3..000000000 --- a/third_party/aom/aom_dsp/arm/subpel_variance_media.c +++ /dev/null @@ -1,81 +0,0 @@ -/* - * 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/aom_integer.h" -#include "aom_ports/mem.h" - -#if HAVE_MEDIA -static const int16_t bilinear_filters_media[8][2] = { { 128, 0 }, { 112, 16 }, - { 96, 32 }, { 80, 48 }, - { 64, 64 }, { 48, 80 }, - { 32, 96 }, { 16, 112 } }; - -extern void aom_filter_block2d_bil_first_pass_media( - const uint8_t *src_ptr, uint16_t *dst_ptr, uint32_t src_pitch, - uint32_t height, uint32_t width, const int16_t *filter); - -extern void aom_filter_block2d_bil_second_pass_media( - const uint16_t *src_ptr, uint8_t *dst_ptr, int32_t src_pitch, - uint32_t height, uint32_t width, const int16_t *filter); - -unsigned int aom_sub_pixel_variance8x8_media( - const uint8_t *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, - const uint8_t *dst_ptr, int dst_pixels_per_line, unsigned int *sse) { - uint16_t first_pass[10 * 8]; - uint8_t second_pass[8 * 8]; - const int16_t *HFilter, *VFilter; - - HFilter = bilinear_filters_media[xoffset]; - VFilter = bilinear_filters_media[yoffset]; - - aom_filter_block2d_bil_first_pass_media(src_ptr, first_pass, - src_pixels_per_line, 9, 8, HFilter); - aom_filter_block2d_bil_second_pass_media(first_pass, second_pass, 8, 8, 8, - VFilter); - - return aom_variance8x8_media(second_pass, 8, dst_ptr, dst_pixels_per_line, - sse); -} - -unsigned int aom_sub_pixel_variance16x16_media( - const uint8_t *src_ptr, int src_pixels_per_line, int xoffset, int yoffset, - const uint8_t *dst_ptr, int dst_pixels_per_line, unsigned int *sse) { - uint16_t first_pass[36 * 16]; - uint8_t second_pass[20 * 16]; - const int16_t *HFilter, *VFilter; - unsigned int var; - - if (xoffset == 4 && yoffset == 0) { - var = aom_variance_halfpixvar16x16_h_media( - src_ptr, src_pixels_per_line, dst_ptr, dst_pixels_per_line, sse); - } else if (xoffset == 0 && yoffset == 4) { - var = aom_variance_halfpixvar16x16_v_media( - src_ptr, src_pixels_per_line, dst_ptr, dst_pixels_per_line, sse); - } else if (xoffset == 4 && yoffset == 4) { - var = aom_variance_halfpixvar16x16_hv_media( - src_ptr, src_pixels_per_line, dst_ptr, dst_pixels_per_line, sse); - } else { - HFilter = bilinear_filters_media[xoffset]; - VFilter = bilinear_filters_media[yoffset]; - - aom_filter_block2d_bil_first_pass_media( - src_ptr, first_pass, src_pixels_per_line, 17, 16, HFilter); - aom_filter_block2d_bil_second_pass_media(first_pass, second_pass, 16, 16, - 16, VFilter); - - var = aom_variance16x16_media(second_pass, 16, dst_ptr, dst_pixels_per_line, - sse); - } - return var; -} -#endif // HAVE_MEDIA diff --git a/third_party/aom/aom_dsp/arm/variance_halfpixvar16x16_h_media.asm b/third_party/aom/aom_dsp/arm/variance_halfpixvar16x16_h_media.asm deleted file mode 100644 index 1e5c9178e..000000000 --- a/third_party/aom/aom_dsp/arm/variance_halfpixvar16x16_h_media.asm +++ /dev/null @@ -1,185 +0,0 @@ -; -; 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. -; - -; - - - EXPORT |aom_variance_halfpixvar16x16_h_media| - - ARM - REQUIRE8 - PRESERVE8 - - AREA ||.text||, CODE, READONLY, ALIGN=2 - -; r0 unsigned char *src_ptr -; r1 int source_stride -; r2 unsigned char *ref_ptr -; r3 int recon_stride -; stack unsigned int *sse -|aom_variance_halfpixvar16x16_h_media| PROC - - stmfd sp!, {r4-r12, lr} - - pld [r0, r1, lsl #0] - pld [r2, r3, lsl #0] - - mov r8, #0 ; initialize sum = 0 - ldr r10, c80808080 - mov r11, #0 ; initialize sse = 0 - mov r12, #16 ; set loop counter to 16 (=block height) - mov lr, #0 ; constant zero -loop - ; 1st 4 pixels - ldr r4, [r0, #0] ; load 4 src pixels - ldr r6, [r0, #1] ; load 4 src pixels with 1 byte offset - ldr r5, [r2, #0] ; load 4 ref pixels - - ; bilinear interpolation - mvn r6, r6 - uhsub8 r4, r4, r6 - eor r4, r4, r10 - - usub8 r6, r4, r5 ; calculate difference - pld [r0, r1, lsl #1] - sel r7, r6, lr ; select bytes with positive difference - usub8 r6, r5, r4 ; calculate difference with reversed operands - pld [r2, r3, lsl #1] - sel r6, r6, lr ; select bytes with negative difference - - ; calculate partial sums - usad8 r4, r7, lr ; calculate sum of positive differences - usad8 r5, r6, lr ; calculate sum of negative differences - orr r6, r6, r7 ; differences of all 4 pixels - ; calculate total sum - adds r8, r8, r4 ; add positive differences to sum - subs r8, r8, r5 ; subtract negative differences from sum - - ; calculate sse - uxtb16 r5, r6 ; byte (two pixels) to halfwords - uxtb16 r7, r6, ror #8 ; another two pixels to halfwords - smlad r11, r5, r5, r11 ; dual signed multiply, add and accumulate (1) - - ; 2nd 4 pixels - ldr r4, [r0, #4] ; load 4 src pixels - ldr r6, [r0, #5] ; load 4 src pixels with 1 byte offset - ldr r5, [r2, #4] ; load 4 ref pixels - - ; bilinear interpolation - mvn r6, r6 - uhsub8 r4, r4, r6 - eor r4, r4, r10 - - smlad r11, r7, r7, r11 ; dual signed multiply, add and accumulate (2) - - usub8 r6, r4, r5 ; calculate difference - sel r7, r6, lr ; select bytes with positive difference - usub8 r6, r5, r4 ; calculate difference with reversed operands - sel r6, r6, lr ; select bytes with negative difference - - ; calculate partial sums - usad8 r4, r7, lr ; calculate sum of positive differences - usad8 r5, r6, lr ; calculate sum of negative differences - orr r6, r6, r7 ; differences of all 4 pixels - - ; calculate total sum - add r8, r8, r4 ; add positive differences to sum - sub r8, r8, r5 ; subtract negative differences from sum - - ; calculate sse - uxtb16 r5, r6 ; byte (two pixels) to halfwords - uxtb16 r7, r6, ror #8 ; another two pixels to halfwords - smlad r11, r5, r5, r11 ; dual signed multiply, add and accumulate (1) - - ; 3rd 4 pixels - ldr r4, [r0, #8] ; load 4 src pixels - ldr r6, [r0, #9] ; load 4 src pixels with 1 byte offset - ldr r5, [r2, #8] ; load 4 ref pixels - - ; bilinear interpolation - mvn r6, r6 - uhsub8 r4, r4, r6 - eor r4, r4, r10 - - smlad r11, r7, r7, r11 ; dual signed multiply, add and accumulate (2) - - usub8 r6, r4, r5 ; calculate difference - sel r7, r6, lr ; select bytes with positive difference - usub8 r6, r5, r4 ; calculate difference with reversed operands - sel r6, r6, lr ; select bytes with negative difference - - ; calculate partial sums - usad8 r4, r7, lr ; calculate sum of positive differences - usad8 r5, r6, lr ; calculate sum of negative differences - orr r6, r6, r7 ; differences of all 4 pixels - - ; calculate total sum - add r8, r8, r4 ; add positive differences to sum - sub r8, r8, r5 ; subtract negative differences from sum - - ; calculate sse - uxtb16 r5, r6 ; byte (two pixels) to halfwords - uxtb16 r7, r6, ror #8 ; another two pixels to halfwords - smlad r11, r5, r5, r11 ; dual signed multiply, add and accumulate (1) - - ; 4th 4 pixels - ldr r4, [r0, #12] ; load 4 src pixels - ldr r6, [r0, #13] ; load 4 src pixels with 1 byte offset - ldr r5, [r2, #12] ; load 4 ref pixels - - ; bilinear interpolation - mvn r6, r6 - uhsub8 r4, r4, r6 - eor r4, r4, r10 - - smlad r11, r7, r7, r11 ; dual signed multiply, add and accumulate (2) - - usub8 r6, r4, r5 ; calculate difference - add r0, r0, r1 ; set src_ptr to next row - sel r7, r6, lr ; select bytes with positive difference - usub8 r6, r5, r4 ; calculate difference with reversed operands - add r2, r2, r3 ; set dst_ptr to next row - sel r6, r6, lr ; select bytes with negative difference - - ; calculate partial sums - usad8 r4, r7, lr ; calculate sum of positive differences - usad8 r5, r6, lr ; calculate sum of negative differences - orr r6, r6, r7 ; differences of all 4 pixels - - ; calculate total sum - add r8, r8, r4 ; add positive differences to sum - sub r8, r8, r5 ; subtract negative differences from sum - - ; calculate sse - uxtb16 r5, r6 ; byte (two pixels) to halfwords - uxtb16 r7, r6, ror #8 ; another two pixels to halfwords - smlad r11, r5, r5, r11 ; dual signed multiply, add and accumulate (1) - smlad r11, r7, r7, r11 ; dual signed multiply, add and accumulate (2) - - subs r12, r12, #1 - - bne loop - - ; return stuff - ldr r6, [sp, #40] ; get address of sse - mul r0, r8, r8 ; sum * sum - str r11, [r6] ; store sse - sub r0, r11, r0, lsr #8 ; return (sse - ((sum * sum) >> 8)) - - ldmfd sp!, {r4-r12, pc} - - ENDP - -c80808080 - DCD 0x80808080 - - END - diff --git a/third_party/aom/aom_dsp/arm/variance_halfpixvar16x16_hv_media.asm b/third_party/aom/aom_dsp/arm/variance_halfpixvar16x16_hv_media.asm deleted file mode 100644 index 9e0af830e..000000000 --- a/third_party/aom/aom_dsp/arm/variance_halfpixvar16x16_hv_media.asm +++ /dev/null @@ -1,225 +0,0 @@ -; -; 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. -; - -; - - - EXPORT |aom_variance_halfpixvar16x16_hv_media| - - ARM - REQUIRE8 - PRESERVE8 - - AREA ||.text||, CODE, READONLY, ALIGN=2 - -; r0 unsigned char *src_ptr -; r1 int source_stride -; r2 unsigned char *ref_ptr -; r3 int recon_stride -; stack unsigned int *sse -|aom_variance_halfpixvar16x16_hv_media| PROC - - stmfd sp!, {r4-r12, lr} - - pld [r0, r1, lsl #0] - pld [r2, r3, lsl #0] - - mov r8, #0 ; initialize sum = 0 - ldr r10, c80808080 - mov r11, #0 ; initialize sse = 0 - mov r12, #16 ; set loop counter to 16 (=block height) - mov lr, #0 ; constant zero -loop - add r9, r0, r1 ; pointer to pixels on the next row - ; 1st 4 pixels - ldr r4, [r0, #0] ; load source pixels a, row N - ldr r6, [r0, #1] ; load source pixels b, row N - ldr r5, [r9, #0] ; load source pixels c, row N+1 - ldr r7, [r9, #1] ; load source pixels d, row N+1 - - ; x = (a + b + 1) >> 1, interpolate pixels horizontally on row N - mvn r6, r6 - uhsub8 r4, r4, r6 - eor r4, r4, r10 - ; y = (c + d + 1) >> 1, interpolate pixels horizontally on row N+1 - mvn r7, r7 - uhsub8 r5, r5, r7 - eor r5, r5, r10 - ; z = (x + y + 1) >> 1, interpolate half pixel values vertically - mvn r5, r5 - uhsub8 r4, r4, r5 - ldr r5, [r2, #0] ; load 4 ref pixels - eor r4, r4, r10 - - usub8 r6, r4, r5 ; calculate difference - pld [r0, r1, lsl #1] - sel r7, r6, lr ; select bytes with positive difference - usub8 r6, r5, r4 ; calculate difference with reversed operands - pld [r2, r3, lsl #1] - sel r6, r6, lr ; select bytes with negative difference - - ; calculate partial sums - usad8 r4, r7, lr ; calculate sum of positive differences - usad8 r5, r6, lr ; calculate sum of negative differences - orr r6, r6, r7 ; differences of all 4 pixels - ; calculate total sum - adds r8, r8, r4 ; add positive differences to sum - subs r8, r8, r5 ; subtract negative differences from sum - - ; calculate sse - uxtb16 r5, r6 ; byte (two pixels) to halfwords - uxtb16 r7, r6, ror #8 ; another two pixels to halfwords - smlad r11, r5, r5, r11 ; dual signed multiply, add and accumulate (1) - - ; 2nd 4 pixels - ldr r4, [r0, #4] ; load source pixels a, row N - ldr r6, [r0, #5] ; load source pixels b, row N - ldr r5, [r9, #4] ; load source pixels c, row N+1 - - smlad r11, r7, r7, r11 ; dual signed multiply, add and accumulate (2) - - ldr r7, [r9, #5] ; load source pixels d, row N+1 - - ; x = (a + b + 1) >> 1, interpolate pixels horizontally on row N - mvn r6, r6 - uhsub8 r4, r4, r6 - eor r4, r4, r10 - ; y = (c + d + 1) >> 1, interpolate pixels horizontally on row N+1 - mvn r7, r7 - uhsub8 r5, r5, r7 - eor r5, r5, r10 - ; z = (x + y + 1) >> 1, interpolate half pixel values vertically - mvn r5, r5 - uhsub8 r4, r4, r5 - ldr r5, [r2, #4] ; load 4 ref pixels - eor r4, r4, r10 - - usub8 r6, r4, r5 ; calculate difference - sel r7, r6, lr ; select bytes with positive difference - usub8 r6, r5, r4 ; calculate difference with reversed operands - sel r6, r6, lr ; select bytes with negative difference - - ; calculate partial sums - usad8 r4, r7, lr ; calculate sum of positive differences - usad8 r5, r6, lr ; calculate sum of negative differences - orr r6, r6, r7 ; differences of all 4 pixels - - ; calculate total sum - add r8, r8, r4 ; add positive differences to sum - sub r8, r8, r5 ; subtract negative differences from sum - - ; calculate sse - uxtb16 r5, r6 ; byte (two pixels) to halfwords - uxtb16 r7, r6, ror #8 ; another two pixels to halfwords - smlad r11, r5, r5, r11 ; dual signed multiply, add and accumulate (1) - - ; 3rd 4 pixels - ldr r4, [r0, #8] ; load source pixels a, row N - ldr r6, [r0, #9] ; load source pixels b, row N - ldr r5, [r9, #8] ; load source pixels c, row N+1 - - smlad r11, r7, r7, r11 ; dual signed multiply, add and accumulate (2) - - ldr r7, [r9, #9] ; load source pixels d, row N+1 - - ; x = (a + b + 1) >> 1, interpolate pixels horizontally on row N - mvn r6, r6 - uhsub8 r4, r4, r6 - eor r4, r4, r10 - ; y = (c + d + 1) >> 1, interpolate pixels horizontally on row N+1 - mvn r7, r7 - uhsub8 r5, r5, r7 - eor r5, r5, r10 - ; z = (x + y + 1) >> 1, interpolate half pixel values vertically - mvn r5, r5 - uhsub8 r4, r4, r5 - ldr r5, [r2, #8] ; load 4 ref pixels - eor r4, r4, r10 - - usub8 r6, r4, r5 ; calculate difference - sel r7, r6, lr ; select bytes with positive difference - usub8 r6, r5, r4 ; calculate difference with reversed operands - sel r6, r6, lr ; select bytes with negative difference - - ; calculate partial sums - usad8 r4, r7, lr ; calculate sum of positive differences - usad8 r5, r6, lr ; calculate sum of negative differences - orr r6, r6, r7 ; differences of all 4 pixels - - ; calculate total sum - add r8, r8, r4 ; add positive differences to sum - sub r8, r8, r5 ; subtract negative differences from sum - - ; calculate sse - uxtb16 r5, r6 ; byte (two pixels) to halfwords - uxtb16 r7, r6, ror #8 ; another two pixels to halfwords - smlad r11, r5, r5, r11 ; dual signed multiply, add and accumulate (1) - - ; 4th 4 pixels - ldr r4, [r0, #12] ; load source pixels a, row N - ldr r6, [r0, #13] ; load source pixels b, row N - ldr r5, [r9, #12] ; load source pixels c, row N+1 - smlad r11, r7, r7, r11 ; dual signed multiply, add and accumulate (2) - ldr r7, [r9, #13] ; load source pixels d, row N+1 - - ; x = (a + b + 1) >> 1, interpolate pixels horizontally on row N - mvn r6, r6 - uhsub8 r4, r4, r6 - eor r4, r4, r10 - ; y = (c + d + 1) >> 1, interpolate pixels horizontally on row N+1 - mvn r7, r7 - uhsub8 r5, r5, r7 - eor r5, r5, r10 - ; z = (x + y + 1) >> 1, interpolate half pixel values vertically - mvn r5, r5 - uhsub8 r4, r4, r5 - ldr r5, [r2, #12] ; load 4 ref pixels - eor r4, r4, r10 - - usub8 r6, r4, r5 ; calculate difference - add r0, r0, r1 ; set src_ptr to next row - sel r7, r6, lr ; select bytes with positive difference - usub8 r6, r5, r4 ; calculate difference with reversed operands - add r2, r2, r3 ; set dst_ptr to next row - sel r6, r6, lr ; select bytes with negative difference - - ; calculate partial sums - usad8 r4, r7, lr ; calculate sum of positive differences - usad8 r5, r6, lr ; calculate sum of negative differences - orr r6, r6, r7 ; differences of all 4 pixels - - ; calculate total sum - add r8, r8, r4 ; add positive differences to sum - sub r8, r8, r5 ; subtract negative differences from sum - - ; calculate sse - uxtb16 r5, r6 ; byte (two pixels) to halfwords - uxtb16 r7, r6, ror #8 ; another two pixels to halfwords - smlad r11, r5, r5, r11 ; dual signed multiply, add and accumulate (1) - subs r12, r12, #1 - smlad r11, r7, r7, r11 ; dual signed multiply, add and accumulate (2) - - bne loop - - ; return stuff - ldr r6, [sp, #40] ; get address of sse - mul r0, r8, r8 ; sum * sum - str r11, [r6] ; store sse - sub r0, r11, r0, lsr #8 ; return (sse - ((sum * sum) >> 8)) - - ldmfd sp!, {r4-r12, pc} - - ENDP - -c80808080 - DCD 0x80808080 - - END diff --git a/third_party/aom/aom_dsp/arm/variance_halfpixvar16x16_v_media.asm b/third_party/aom/aom_dsp/arm/variance_halfpixvar16x16_v_media.asm deleted file mode 100644 index 545b68179..000000000 --- a/third_party/aom/aom_dsp/arm/variance_halfpixvar16x16_v_media.asm +++ /dev/null @@ -1,187 +0,0 @@ -; -; 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. -; - -; - - - EXPORT |aom_variance_halfpixvar16x16_v_media| - - ARM - REQUIRE8 - PRESERVE8 - - AREA ||.text||, CODE, READONLY, ALIGN=2 - -; r0 unsigned char *src_ptr -; r1 int source_stride -; r2 unsigned char *ref_ptr -; r3 int recon_stride -; stack unsigned int *sse -|aom_variance_halfpixvar16x16_v_media| PROC - - stmfd sp!, {r4-r12, lr} - - pld [r0, r1, lsl #0] - pld [r2, r3, lsl #0] - - mov r8, #0 ; initialize sum = 0 - ldr r10, c80808080 - mov r11, #0 ; initialize sse = 0 - mov r12, #16 ; set loop counter to 16 (=block height) - mov lr, #0 ; constant zero -loop - add r9, r0, r1 ; set src pointer to next row - ; 1st 4 pixels - ldr r4, [r0, #0] ; load 4 src pixels - ldr r6, [r9, #0] ; load 4 src pixels from next row - ldr r5, [r2, #0] ; load 4 ref pixels - - ; bilinear interpolation - mvn r6, r6 - uhsub8 r4, r4, r6 - eor r4, r4, r10 - - usub8 r6, r4, r5 ; calculate difference - pld [r0, r1, lsl #1] - sel r7, r6, lr ; select bytes with positive difference - usub8 r6, r5, r4 ; calculate difference with reversed operands - pld [r2, r3, lsl #1] - sel r6, r6, lr ; select bytes with negative difference - - ; calculate partial sums - usad8 r4, r7, lr ; calculate sum of positive differences - usad8 r5, r6, lr ; calculate sum of negative differences - orr r6, r6, r7 ; differences of all 4 pixels - ; calculate total sum - adds r8, r8, r4 ; add positive differences to sum - subs r8, r8, r5 ; subtract negative differences from sum - - ; calculate sse - uxtb16 r5, r6 ; byte (two pixels) to halfwords - uxtb16 r7, r6, ror #8 ; another two pixels to halfwords - smlad r11, r5, r5, r11 ; dual signed multiply, add and accumulate (1) - - ; 2nd 4 pixels - ldr r4, [r0, #4] ; load 4 src pixels - ldr r6, [r9, #4] ; load 4 src pixels from next row - ldr r5, [r2, #4] ; load 4 ref pixels - - ; bilinear interpolation - mvn r6, r6 - uhsub8 r4, r4, r6 - eor r4, r4, r10 - - smlad r11, r7, r7, r11 ; dual signed multiply, add and accumulate (2) - - usub8 r6, r4, r5 ; calculate difference - sel r7, r6, lr ; select bytes with positive difference - usub8 r6, r5, r4 ; calculate difference with reversed operands - sel r6, r6, lr ; select bytes with negative difference - - ; calculate partial sums - usad8 r4, r7, lr ; calculate sum of positive differences - usad8 r5, r6, lr ; calculate sum of negative differences - orr r6, r6, r7 ; differences of all 4 pixels - - ; calculate total sum - add r8, r8, r4 ; add positive differences to sum - sub r8, r8, r5 ; subtract negative differences from sum - - ; calculate sse - uxtb16 r5, r6 ; byte (two pixels) to halfwords - uxtb16 r7, r6, ror #8 ; another two pixels to halfwords - smlad r11, r5, r5, r11 ; dual signed multiply, add and accumulate (1) - - ; 3rd 4 pixels - ldr r4, [r0, #8] ; load 4 src pixels - ldr r6, [r9, #8] ; load 4 src pixels from next row - ldr r5, [r2, #8] ; load 4 ref pixels - - ; bilinear interpolation - mvn r6, r6 - uhsub8 r4, r4, r6 - eor r4, r4, r10 - - smlad r11, r7, r7, r11 ; dual signed multiply, add and accumulate (2) - - usub8 r6, r4, r5 ; calculate difference - sel r7, r6, lr ; select bytes with positive difference - usub8 r6, r5, r4 ; calculate difference with reversed operands - sel r6, r6, lr ; select bytes with negative difference - - ; calculate partial sums - usad8 r4, r7, lr ; calculate sum of positive differences - usad8 r5, r6, lr ; calculate sum of negative differences - orr r6, r6, r7 ; differences of all 4 pixels - - ; calculate total sum - add r8, r8, r4 ; add positive differences to sum - sub r8, r8, r5 ; subtract negative differences from sum - - ; calculate sse - uxtb16 r5, r6 ; byte (two pixels) to halfwords - uxtb16 r7, r6, ror #8 ; another two pixels to halfwords - smlad r11, r5, r5, r11 ; dual signed multiply, add and accumulate (1) - - ; 4th 4 pixels - ldr r4, [r0, #12] ; load 4 src pixels - ldr r6, [r9, #12] ; load 4 src pixels from next row - ldr r5, [r2, #12] ; load 4 ref pixels - - ; bilinear interpolation - mvn r6, r6 - uhsub8 r4, r4, r6 - eor r4, r4, r10 - - smlad r11, r7, r7, r11 ; dual signed multiply, add and accumulate (2) - - usub8 r6, r4, r5 ; calculate difference - add r0, r0, r1 ; set src_ptr to next row - sel r7, r6, lr ; select bytes with positive difference - usub8 r6, r5, r4 ; calculate difference with reversed operands - add r2, r2, r3 ; set dst_ptr to next row - sel r6, r6, lr ; select bytes with negative difference - - ; calculate partial sums - usad8 r4, r7, lr ; calculate sum of positive differences - usad8 r5, r6, lr ; calculate sum of negative differences - orr r6, r6, r7 ; differences of all 4 pixels - - ; calculate total sum - add r8, r8, r4 ; add positive differences to sum - sub r8, r8, r5 ; subtract negative differences from sum - - ; calculate sse - uxtb16 r5, r6 ; byte (two pixels) to halfwords - uxtb16 r7, r6, ror #8 ; another two pixels to halfwords - smlad r11, r5, r5, r11 ; dual signed multiply, add and accumulate (1) - smlad r11, r7, r7, r11 ; dual signed multiply, add and accumulate (2) - - - subs r12, r12, #1 - - bne loop - - ; return stuff - ldr r6, [sp, #40] ; get address of sse - mul r0, r8, r8 ; sum * sum - str r11, [r6] ; store sse - sub r0, r11, r0, lsr #8 ; return (sse - ((sum * sum) >> 8)) - - ldmfd sp!, {r4-r12, pc} - - ENDP - -c80808080 - DCD 0x80808080 - - END - diff --git a/third_party/aom/aom_dsp/arm/variance_media.asm b/third_party/aom/aom_dsp/arm/variance_media.asm deleted file mode 100644 index fdc311a81..000000000 --- a/third_party/aom/aom_dsp/arm/variance_media.asm +++ /dev/null @@ -1,361 +0,0 @@ -; -; 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. -; - -; - - - EXPORT |aom_variance16x16_media| - EXPORT |aom_variance8x8_media| - EXPORT |aom_mse16x16_media| - - ARM - REQUIRE8 - PRESERVE8 - - AREA ||.text||, CODE, READONLY, ALIGN=2 - -; r0 unsigned char *src_ptr -; r1 int source_stride -; r2 unsigned char *ref_ptr -; r3 int recon_stride -; stack unsigned int *sse -|aom_variance16x16_media| PROC - - stmfd sp!, {r4-r12, lr} - - pld [r0, r1, lsl #0] - pld [r2, r3, lsl #0] - - mov r8, #0 ; initialize sum = 0 - mov r11, #0 ; initialize sse = 0 - mov r12, #16 ; set loop counter to 16 (=block height) - -loop16x16 - ; 1st 4 pixels - ldr r4, [r0, #0] ; load 4 src pixels - ldr r5, [r2, #0] ; load 4 ref pixels - - mov lr, #0 ; constant zero - - usub8 r6, r4, r5 ; calculate difference - pld [r0, r1, lsl #1] - sel r7, r6, lr ; select bytes with positive difference - usub8 r9, r5, r4 ; calculate difference with reversed operands - pld [r2, r3, lsl #1] - sel r6, r9, lr ; select bytes with negative difference - - ; calculate partial sums - usad8 r4, r7, lr ; calculate sum of positive differences - usad8 r5, r6, lr ; calculate sum of negative differences - orr r6, r6, r7 ; differences of all 4 pixels - ; calculate total sum - adds r8, r8, r4 ; add positive differences to sum - subs r8, r8, r5 ; subtract negative differences from sum - - ; calculate sse - uxtb16 r5, r6 ; byte (two pixels) to halfwords - uxtb16 r10, r6, ror #8 ; another two pixels to halfwords - smlad r11, r5, r5, r11 ; dual signed multiply, add and accumulate (1) - - ; 2nd 4 pixels - ldr r4, [r0, #4] ; load 4 src pixels - ldr r5, [r2, #4] ; load 4 ref pixels - smlad r11, r10, r10, r11 ; dual signed multiply, add and accumulate (2) - - usub8 r6, r4, r5 ; calculate difference - sel r7, r6, lr ; select bytes with positive difference - usub8 r9, r5, r4 ; calculate difference with reversed operands - sel r6, r9, lr ; select bytes with negative difference - - ; calculate partial sums - usad8 r4, r7, lr ; calculate sum of positive differences - usad8 r5, r6, lr ; calculate sum of negative differences - orr r6, r6, r7 ; differences of all 4 pixels - - ; calculate total sum - add r8, r8, r4 ; add positive differences to sum - sub r8, r8, r5 ; subtract negative differences from sum - - ; calculate sse - uxtb16 r5, r6 ; byte (two pixels) to halfwords - uxtb16 r10, r6, ror #8 ; another two pixels to halfwords - smlad r11, r5, r5, r11 ; dual signed multiply, add and accumulate (1) - - ; 3rd 4 pixels - ldr r4, [r0, #8] ; load 4 src pixels - ldr r5, [r2, #8] ; load 4 ref pixels - smlad r11, r10, r10, r11 ; dual signed multiply, add and accumulate (2) - - usub8 r6, r4, r5 ; calculate difference - sel r7, r6, lr ; select bytes with positive difference - usub8 r9, r5, r4 ; calculate difference with reversed operands - sel r6, r9, lr ; select bytes with negative difference - - ; calculate partial sums - usad8 r4, r7, lr ; calculate sum of positive differences - usad8 r5, r6, lr ; calculate sum of negative differences - orr r6, r6, r7 ; differences of all 4 pixels - - ; calculate total sum - add r8, r8, r4 ; add positive differences to sum - sub r8, r8, r5 ; subtract negative differences from sum - - ; calculate sse - uxtb16 r5, r6 ; byte (two pixels) to halfwords - uxtb16 r10, r6, ror #8 ; another two pixels to halfwords - smlad r11, r5, r5, r11 ; dual signed multiply, add and accumulate (1) - - ; 4th 4 pixels - ldr r4, [r0, #12] ; load 4 src pixels - ldr r5, [r2, #12] ; load 4 ref pixels - smlad r11, r10, r10, r11 ; dual signed multiply, add and accumulate (2) - - usub8 r6, r4, r5 ; calculate difference - add r0, r0, r1 ; set src_ptr to next row - sel r7, r6, lr ; select bytes with positive difference - usub8 r9, r5, r4 ; calculate difference with reversed operands - add r2, r2, r3 ; set dst_ptr to next row - sel r6, r9, lr ; select bytes with negative difference - - ; calculate partial sums - usad8 r4, r7, lr ; calculate sum of positive differences - usad8 r5, r6, lr ; calculate sum of negative differences - orr r6, r6, r7 ; differences of all 4 pixels - - ; calculate total sum - add r8, r8, r4 ; add positive differences to sum - sub r8, r8, r5 ; subtract negative differences from sum - - ; calculate sse - uxtb16 r5, r6 ; byte (two pixels) to halfwords - uxtb16 r10, r6, ror #8 ; another two pixels to halfwords - smlad r11, r5, r5, r11 ; dual signed multiply, add and accumulate (1) - smlad r11, r10, r10, r11 ; dual signed multiply, add and accumulate (2) - - - subs r12, r12, #1 - - bne loop16x16 - - ; return stuff - ldr r6, [sp, #40] ; get address of sse - mul r0, r8, r8 ; sum * sum - str r11, [r6] ; store sse - sub r0, r11, r0, lsr #8 ; return (sse - ((sum * sum) >> 8)) - - ldmfd sp!, {r4-r12, pc} - - ENDP - -; r0 unsigned char *src_ptr -; r1 int source_stride -; r2 unsigned char *ref_ptr -; r3 int recon_stride -; stack unsigned int *sse -|aom_variance8x8_media| PROC - - push {r4-r10, lr} - - pld [r0, r1, lsl #0] - pld [r2, r3, lsl #0] - - mov r12, #8 ; set loop counter to 8 (=block height) - mov r4, #0 ; initialize sum = 0 - mov r5, #0 ; initialize sse = 0 - -loop8x8 - ; 1st 4 pixels - ldr r6, [r0, #0x0] ; load 4 src pixels - ldr r7, [r2, #0x0] ; load 4 ref pixels - - mov lr, #0 ; constant zero - - usub8 r8, r6, r7 ; calculate difference - pld [r0, r1, lsl #1] - sel r10, r8, lr ; select bytes with positive difference - usub8 r9, r7, r6 ; calculate difference with reversed operands - pld [r2, r3, lsl #1] - sel r8, r9, lr ; select bytes with negative difference - - ; calculate partial sums - usad8 r6, r10, lr ; calculate sum of positive differences - usad8 r7, r8, lr ; calculate sum of negative differences - orr r8, r8, r10 ; differences of all 4 pixels - ; calculate total sum - add r4, r4, r6 ; add positive differences to sum - sub r4, r4, r7 ; subtract negative differences from sum - - ; calculate sse - uxtb16 r7, r8 ; byte (two pixels) to halfwords - uxtb16 r10, r8, ror #8 ; another two pixels to halfwords - smlad r5, r7, r7, r5 ; dual signed multiply, add and accumulate (1) - - ; 2nd 4 pixels - ldr r6, [r0, #0x4] ; load 4 src pixels - ldr r7, [r2, #0x4] ; load 4 ref pixels - smlad r5, r10, r10, r5 ; dual signed multiply, add and accumulate (2) - - usub8 r8, r6, r7 ; calculate difference - add r0, r0, r1 ; set src_ptr to next row - sel r10, r8, lr ; select bytes with positive difference - usub8 r9, r7, r6 ; calculate difference with reversed operands - add r2, r2, r3 ; set dst_ptr to next row - sel r8, r9, lr ; select bytes with negative difference - - ; calculate partial sums - usad8 r6, r10, lr ; calculate sum of positive differences - usad8 r7, r8, lr ; calculate sum of negative differences - orr r8, r8, r10 ; differences of all 4 pixels - - ; calculate total sum - add r4, r4, r6 ; add positive differences to sum - sub r4, r4, r7 ; subtract negative differences from sum - - ; calculate sse - uxtb16 r7, r8 ; byte (two pixels) to halfwords - uxtb16 r10, r8, ror #8 ; another two pixels to halfwords - smlad r5, r7, r7, r5 ; dual signed multiply, add and accumulate (1) - subs r12, r12, #1 ; next row - smlad r5, r10, r10, r5 ; dual signed multiply, add and accumulate (2) - - bne loop8x8 - - ; return stuff - ldr r8, [sp, #32] ; get address of sse - mul r1, r4, r4 ; sum * sum - str r5, [r8] ; store sse - sub r0, r5, r1, ASR #6 ; return (sse - ((sum * sum) >> 6)) - - pop {r4-r10, pc} - - ENDP - -; r0 unsigned char *src_ptr -; r1 int source_stride -; r2 unsigned char *ref_ptr -; r3 int recon_stride -; stack unsigned int *sse -; -;note: Based on aom_variance16x16_media. In this function, sum is never used. -; So, we can remove this part of calculation. - -|aom_mse16x16_media| PROC - - push {r4-r9, lr} - - pld [r0, r1, lsl #0] - pld [r2, r3, lsl #0] - - mov r12, #16 ; set loop counter to 16 (=block height) - mov r4, #0 ; initialize sse = 0 - -loopmse - ; 1st 4 pixels - ldr r5, [r0, #0x0] ; load 4 src pixels - ldr r6, [r2, #0x0] ; load 4 ref pixels - - mov lr, #0 ; constant zero - - usub8 r8, r5, r6 ; calculate difference - pld [r0, r1, lsl #1] - sel r7, r8, lr ; select bytes with positive difference - usub8 r9, r6, r5 ; calculate difference with reversed operands - pld [r2, r3, lsl #1] - sel r8, r9, lr ; select bytes with negative difference - - ; calculate partial sums - usad8 r5, r7, lr ; calculate sum of positive differences - usad8 r6, r8, lr ; calculate sum of negative differences - orr r8, r8, r7 ; differences of all 4 pixels - - ldr r5, [r0, #0x4] ; load 4 src pixels - - ; calculate sse - uxtb16 r6, r8 ; byte (two pixels) to halfwords - uxtb16 r7, r8, ror #8 ; another two pixels to halfwords - smlad r4, r6, r6, r4 ; dual signed multiply, add and accumulate (1) - - ; 2nd 4 pixels - ldr r6, [r2, #0x4] ; load 4 ref pixels - smlad r4, r7, r7, r4 ; dual signed multiply, add and accumulate (2) - - usub8 r8, r5, r6 ; calculate difference - sel r7, r8, lr ; select bytes with positive difference - usub8 r9, r6, r5 ; calculate difference with reversed operands - sel r8, r9, lr ; select bytes with negative difference - - ; calculate partial sums - usad8 r5, r7, lr ; calculate sum of positive differences - usad8 r6, r8, lr ; calculate sum of negative differences - orr r8, r8, r7 ; differences of all 4 pixels - ldr r5, [r0, #0x8] ; load 4 src pixels - ; calculate sse - uxtb16 r6, r8 ; byte (two pixels) to halfwords - uxtb16 r7, r8, ror #8 ; another two pixels to halfwords - smlad r4, r6, r6, r4 ; dual signed multiply, add and accumulate (1) - - ; 3rd 4 pixels - ldr r6, [r2, #0x8] ; load 4 ref pixels - smlad r4, r7, r7, r4 ; dual signed multiply, add and accumulate (2) - - usub8 r8, r5, r6 ; calculate difference - sel r7, r8, lr ; select bytes with positive difference - usub8 r9, r6, r5 ; calculate difference with reversed operands - sel r8, r9, lr ; select bytes with negative difference - - ; calculate partial sums - usad8 r5, r7, lr ; calculate sum of positive differences - usad8 r6, r8, lr ; calculate sum of negative differences - orr r8, r8, r7 ; differences of all 4 pixels - - ldr r5, [r0, #0xc] ; load 4 src pixels - - ; calculate sse - uxtb16 r6, r8 ; byte (two pixels) to halfwords - uxtb16 r7, r8, ror #8 ; another two pixels to halfwords - smlad r4, r6, r6, r4 ; dual signed multiply, add and accumulate (1) - - ; 4th 4 pixels - ldr r6, [r2, #0xc] ; load 4 ref pixels - smlad r4, r7, r7, r4 ; dual signed multiply, add and accumulate (2) - - usub8 r8, r5, r6 ; calculate difference - add r0, r0, r1 ; set src_ptr to next row - sel r7, r8, lr ; select bytes with positive difference - usub8 r9, r6, r5 ; calculate difference with reversed operands - add r2, r2, r3 ; set dst_ptr to next row - sel r8, r9, lr ; select bytes with negative difference - - ; calculate partial sums - usad8 r5, r7, lr ; calculate sum of positive differences - usad8 r6, r8, lr ; calculate sum of negative differences - orr r8, r8, r7 ; differences of all 4 pixels - - subs r12, r12, #1 ; next row - - ; calculate sse - uxtb16 r6, r8 ; byte (two pixels) to halfwords - uxtb16 r7, r8, ror #8 ; another two pixels to halfwords - smlad r4, r6, r6, r4 ; dual signed multiply, add and accumulate (1) - smlad r4, r7, r7, r4 ; dual signed multiply, add and accumulate (2) - - bne loopmse - - ; return stuff - ldr r1, [sp, #28] ; get address of sse - mov r0, r4 ; return sse - str r4, [r1] ; store sse - - pop {r4-r9, pc} - - ENDP - - END diff --git a/third_party/aom/aom_dsp/avg.c b/third_party/aom/aom_dsp/avg.c index eb6059705..f732224fd 100644 --- a/third_party/aom/aom_dsp/avg.c +++ b/third_party/aom/aom_dsp/avg.c @@ -13,26 +13,6 @@ #include "./aom_dsp_rtcd.h" #include "aom_ports/mem.h" -unsigned int aom_avg_8x8_c(const uint8_t *src, int stride) { - int i, j; - int sum = 0; - for (i = 0; i < 8; ++i, src += stride) - for (j = 0; j < 8; sum += src[j], ++j) { - } - - return ROUND_POWER_OF_TWO(sum, 6); -} - -unsigned int aom_avg_4x4_c(const uint8_t *src, int stride) { - int i, j; - int sum = 0; - for (i = 0; i < 4; ++i, src += stride) - for (j = 0; j < 4; sum += src[j], ++j) { - } - - return ROUND_POWER_OF_TWO(sum, 4); -} - // src_diff: first pass, 9 bit, dynamic range [-255, 255] // second pass, 12 bit, dynamic range [-2040, 2040] static void hadamard_col8(const int16_t *src_diff, int src_stride, @@ -192,28 +172,6 @@ void aom_minmax_8x8_c(const uint8_t *src, int src_stride, const uint8_t *ref, } #if CONFIG_HIGHBITDEPTH -unsigned int aom_highbd_avg_8x8_c(const uint8_t *src, int stride) { - int i, j; - int sum = 0; - const uint16_t *s = CONVERT_TO_SHORTPTR(src); - for (i = 0; i < 8; ++i, s += stride) - for (j = 0; j < 8; sum += s[j], ++j) { - } - - return ROUND_POWER_OF_TWO(sum, 6); -} - -unsigned int aom_highbd_avg_4x4_c(const uint8_t *src, int stride) { - int i, j; - int sum = 0; - const uint16_t *s = CONVERT_TO_SHORTPTR(src); - for (i = 0; i < 4; ++i, s += stride) - for (j = 0; j < 4; sum += s[j], ++j) { - } - - return ROUND_POWER_OF_TWO(sum, 4); -} - void aom_highbd_minmax_8x8_c(const uint8_t *s8, int p, const uint8_t *d8, int dp, int *min, int *max) { int i, j; diff --git a/third_party/aom/aom_dsp/binary_codes_reader.c b/third_party/aom/aom_dsp/binary_codes_reader.c index 96c4cb436..bf304dada 100644 --- a/third_party/aom/aom_dsp/binary_codes_reader.c +++ b/third_party/aom/aom_dsp/binary_codes_reader.c @@ -9,7 +9,7 @@ * PATENTS file, you can obtain it at www.aomedia.org/license/patent. */ -#include "aom_dsp/bitreader.h" +#include "aom_dsp/binary_codes_reader.h" #include "av1/common/common.h" @@ -33,26 +33,28 @@ static uint16_t inv_recenter_finite_nonneg(uint16_t n, uint16_t r, uint16_t v) { } } -int16_t aom_read_primitive_symmetric(aom_reader *r, unsigned int mag_bits) { - if (aom_read_bit(r, NULL)) { - int s = aom_read_bit(r, NULL); - int16_t x = aom_read_literal(r, mag_bits, NULL) + 1; +int16_t aom_read_primitive_symmetric_(aom_reader *r, + unsigned int mag_bits ACCT_STR_PARAM) { + if (aom_read_bit(r, ACCT_STR_NAME)) { + int s = aom_read_bit(r, ACCT_STR_NAME); + int16_t x = aom_read_literal(r, mag_bits, ACCT_STR_NAME) + 1; return (s > 0 ? -x : x); } else { return 0; } } -uint16_t aom_read_primitive_quniform(aom_reader *r, uint16_t n) { +uint16_t aom_read_primitive_quniform_(aom_reader *r, + uint16_t n ACCT_STR_PARAM) { if (n <= 1) return 0; const int l = get_msb(n - 1) + 1; const int m = (1 << l) - n; - const int v = aom_read_literal(r, l - 1, NULL); - return v < m ? v : (v << 1) - m + aom_read_bit(r, NULL); + const int v = aom_read_literal(r, l - 1, ACCT_STR_NAME); + return v < m ? v : (v << 1) - m + aom_read_bit(r, ACCT_STR_NAME); } -uint16_t aom_read_primitive_refbilevel(aom_reader *r, uint16_t n, uint16_t p, - uint16_t ref) { +uint16_t aom_read_primitive_refbilevel_(aom_reader *r, uint16_t n, uint16_t p, + uint16_t ref ACCT_STR_PARAM) { if (n <= 1) return 0; assert(p > 0 && p <= n); assert(ref < n); @@ -64,10 +66,10 @@ uint16_t aom_read_primitive_refbilevel(aom_reader *r, uint16_t n, uint16_t p, lolimit = n - p; } int v; - if (aom_read_bit(r, NULL)) { - v = aom_read_primitive_quniform(r, p) + lolimit; + if (aom_read_bit(r, ACCT_STR_NAME)) { + v = aom_read_primitive_quniform(r, p, ACCT_STR_NAME) + lolimit; } else { - v = aom_read_primitive_quniform(r, n - p); + v = aom_read_primitive_quniform(r, n - p, ACCT_STR_NAME); if (v >= lolimit) v += p; } return v; @@ -75,7 +77,8 @@ uint16_t aom_read_primitive_refbilevel(aom_reader *r, uint16_t n, uint16_t p, // Decode finite subexponential code that for a symbol v in [0, n-1] with // parameter k -uint16_t aom_read_primitive_subexpfin(aom_reader *r, uint16_t n, uint16_t k) { +uint16_t aom_read_primitive_subexpfin_(aom_reader *r, uint16_t n, + uint16_t k ACCT_STR_PARAM) { int i = 0; int mk = 0; uint16_t v; @@ -83,14 +86,14 @@ uint16_t aom_read_primitive_subexpfin(aom_reader *r, uint16_t n, uint16_t k) { int b = (i ? k + i - 1 : k); int a = (1 << b); if (n <= mk + 3 * a) { - v = aom_read_primitive_quniform(r, n - mk) + mk; + v = aom_read_primitive_quniform(r, n - mk, ACCT_STR_NAME) + mk; break; } else { - if (aom_read_bit(r, NULL)) { + if (aom_read_bit(r, ACCT_STR_NAME)) { i = i + 1; mk += a; } else { - v = aom_read_literal(r, b, NULL) + mk; + v = aom_read_literal(r, b, ACCT_STR_NAME) + mk; break; } } @@ -101,17 +104,19 @@ uint16_t aom_read_primitive_subexpfin(aom_reader *r, uint16_t n, uint16_t k) { // Decode finite subexponential code that for a symbol v in [0, n-1] with // parameter k // based on a reference ref also in [0, n-1]. -uint16_t aom_read_primitive_refsubexpfin(aom_reader *r, uint16_t n, uint16_t k, - uint16_t ref) { - return inv_recenter_finite_nonneg(n, ref, - aom_read_primitive_subexpfin(r, n, k)); +uint16_t aom_read_primitive_refsubexpfin_(aom_reader *r, uint16_t n, uint16_t k, + uint16_t ref ACCT_STR_PARAM) { + return inv_recenter_finite_nonneg( + n, ref, aom_read_primitive_subexpfin(r, n, k, ACCT_STR_NAME)); } // Decode finite subexponential code that for a symbol v in [-(n-1), n-1] with // parameter k based on a reference ref also in [-(n-1), n-1]. -int16_t aom_read_signed_primitive_refsubexpfin(aom_reader *r, uint16_t n, - uint16_t k, int16_t ref) { +int16_t aom_read_signed_primitive_refsubexpfin_(aom_reader *r, uint16_t n, + uint16_t k, + int16_t ref ACCT_STR_PARAM) { ref += n - 1; const uint16_t scaled_n = (n << 1) - 1; - return aom_read_primitive_refsubexpfin(r, scaled_n, k, ref) - n + 1; + return aom_read_primitive_refsubexpfin(r, scaled_n, k, ref, ACCT_STR_NAME) - + n + 1; } diff --git a/third_party/aom/aom_dsp/binary_codes_reader.h b/third_party/aom/aom_dsp/binary_codes_reader.h index 738d91da8..1540cf46b 100644 --- a/third_party/aom/aom_dsp/binary_codes_reader.h +++ b/third_party/aom/aom_dsp/binary_codes_reader.h @@ -21,16 +21,32 @@ extern "C" { #include "aom/aom_integer.h" #include "aom_dsp/bitreader.h" -int16_t aom_read_primitive_symmetric(aom_reader *r, unsigned int mag_bits); +#define aom_read_primitive_symmetric(r, n, ACCT_STR_NAME) \ + aom_read_primitive_symmetric_(r, n ACCT_STR_ARG(ACCT_STR_NAME)) +#define aom_read_primitive_quniform(r, n, ACCT_STR_NAME) \ + aom_read_primitive_quniform_(r, n ACCT_STR_ARG(ACCT_STR_NAME)) +#define aom_read_primitive_refbilevel(r, n, p, ref, ACCT_STR_NAME) \ + aom_read_primitive_refbilevel_(r, n, p, ref ACCT_STR_ARG(ACCT_STR_NAME)) +#define aom_read_primitive_subexpfin(r, n, k, ACCT_STR_NAME) \ + aom_read_primitive_subexpfin_(r, n, k ACCT_STR_ARG(ACCT_STR_NAME)) +#define aom_read_primitive_refsubexpfin(r, n, k, ref, ACCT_STR_NAME) \ + aom_read_primitive_refsubexpfin_(r, n, k, ref ACCT_STR_ARG(ACCT_STR_NAME)) +#define aom_read_signed_primitive_refsubexpfin(r, n, k, ref, ACCT_STR_NAME) \ + aom_read_signed_primitive_refsubexpfin_(r, n, k, \ + ref ACCT_STR_ARG(ACCT_STR_NAME)) -uint16_t aom_read_primitive_quniform(aom_reader *r, uint16_t n); -uint16_t aom_read_primitive_refbilevel(aom_reader *r, uint16_t n, uint16_t p, - uint16_t ref); -uint16_t aom_read_primitive_subexpfin(aom_reader *r, uint16_t n, uint16_t k); -uint16_t aom_read_primitive_refsubexpfin(aom_reader *r, uint16_t n, uint16_t k, - uint16_t ref); -int16_t aom_read_signed_primitive_refsubexpfin(aom_reader *r, uint16_t n, - uint16_t k, int16_t ref); +int16_t aom_read_primitive_symmetric_(aom_reader *r, + unsigned int mag_bits ACCT_STR_PARAM); +uint16_t aom_read_primitive_quniform_(aom_reader *r, uint16_t n ACCT_STR_PARAM); +uint16_t aom_read_primitive_refbilevel_(aom_reader *r, uint16_t n, uint16_t p, + uint16_t ref ACCT_STR_PARAM); +uint16_t aom_read_primitive_subexpfin_(aom_reader *r, uint16_t n, + uint16_t k ACCT_STR_PARAM); +uint16_t aom_read_primitive_refsubexpfin_(aom_reader *r, uint16_t n, uint16_t k, + uint16_t ref ACCT_STR_PARAM); +int16_t aom_read_signed_primitive_refsubexpfin_(aom_reader *r, uint16_t n, + uint16_t k, + int16_t ref ACCT_STR_PARAM); #ifdef __cplusplus } // extern "C" #endif diff --git a/third_party/aom/aom_dsp/bitreader.h b/third_party/aom/aom_dsp/bitreader.h index 9cd34dd48..5bad70cb3 100644 --- a/third_party/aom/aom_dsp/bitreader.h +++ b/third_party/aom/aom_dsp/bitreader.h @@ -16,18 +16,13 @@ #include #include "./aom_config.h" -#if CONFIG_EC_ADAPT && !CONFIG_EC_MULTISYMBOL -#error "CONFIG_EC_ADAPT is enabled without enabling CONFIG_EC_MULTISYMBOL." -#endif #include "aom/aomdx.h" #include "aom/aom_integer.h" #if CONFIG_ANS #include "aom_dsp/ansreader.h" -#elif CONFIG_DAALA_EC -#include "aom_dsp/daalaboolreader.h" #else -#include "aom_dsp/dkboolreader.h" +#include "aom_dsp/daalaboolreader.h" #endif #include "aom_dsp/prob.h" #include "av1/common/odintrin.h" @@ -61,26 +56,20 @@ extern "C" { #if CONFIG_ANS typedef struct AnsDecoder aom_reader; -#elif CONFIG_DAALA_EC -typedef struct daala_reader aom_reader; #else -typedef struct aom_dk_reader aom_reader; +typedef struct daala_reader aom_reader; #endif static INLINE int aom_reader_init(aom_reader *r, const uint8_t *buffer, size_t size, aom_decrypt_cb decrypt_cb, void *decrypt_state) { -#if CONFIG_ANS (void)decrypt_cb; (void)decrypt_state; +#if CONFIG_ANS if (size > INT_MAX) return 1; return ans_read_init(r, buffer, (int)size); -#elif CONFIG_DAALA_EC - (void)decrypt_cb; - (void)decrypt_state; - return aom_daala_reader_init(r, buffer, (int)size); #else - return aom_dk_reader_init(r, buffer, size, decrypt_cb, decrypt_state); + return aom_daala_reader_init(r, buffer, (int)size); #endif } @@ -89,20 +78,16 @@ static INLINE const uint8_t *aom_reader_find_end(aom_reader *r) { (void)r; assert(0 && "Use the raw buffer size with ANS"); return NULL; -#elif CONFIG_DAALA_EC - return aom_daala_reader_find_end(r); #else - return aom_dk_reader_find_end(r); + return aom_daala_reader_find_end(r); #endif } static INLINE int aom_reader_has_error(aom_reader *r) { #if CONFIG_ANS return ans_reader_has_error(r); -#elif CONFIG_DAALA_EC - return aom_daala_reader_has_error(r); #else - return aom_dk_reader_has_error(r); + return aom_daala_reader_has_error(r); #endif } @@ -112,10 +97,8 @@ static INLINE uint32_t aom_reader_tell(const aom_reader *r) { (void)r; assert(0 && "aom_reader_tell() is unimplemented for ANS"); return 0; -#elif CONFIG_DAALA_EC - return aom_daala_reader_tell(r); #else - return aom_dk_reader_tell(r); + return aom_daala_reader_tell(r); #endif } @@ -125,10 +108,8 @@ static INLINE uint32_t aom_reader_tell_frac(const aom_reader *r) { (void)r; assert(0 && "aom_reader_tell_frac() is unimplemented for ANS"); return 0; -#elif CONFIG_DAALA_EC - return aom_daala_reader_tell_frac(r); #else - return aom_dk_reader_tell_frac(r); + return aom_daala_reader_tell_frac(r); #endif } @@ -155,10 +136,8 @@ static INLINE int aom_read_(aom_reader *r, int prob ACCT_STR_PARAM) { int ret; #if CONFIG_ANS ret = rabs_read(r, prob); -#elif CONFIG_DAALA_EC - ret = aom_daala_read(r, prob); #else - ret = aom_dk_read(r, prob); + ret = aom_daala_read(r, prob); #endif #if CONFIG_ACCOUNTING if (ACCT_STR_NAME) aom_process_accounting(r, ACCT_STR_NAME); @@ -171,7 +150,7 @@ static INLINE int aom_read_bit_(aom_reader *r ACCT_STR_PARAM) { int ret; #if CONFIG_ANS ret = rabs_read_bit(r); // Non trivial optimization at half probability -#elif CONFIG_DAALA_EC && CONFIG_RAWBITS +#elif CONFIG_RAWBITS // Note this uses raw bits and is not the same as aom_daala_read(r, 128); // Calls to this function are omitted from raw symbol accounting. ret = aom_daala_read_bit(r); @@ -194,28 +173,14 @@ static INLINE int aom_read_literal_(aom_reader *r, int bits ACCT_STR_PARAM) { return literal; } -static INLINE int aom_read_tree_as_bits(aom_reader *r, - const aom_tree_index *tree, - const aom_prob *probs) { - aom_tree_index i = 0; - - while ((i = tree[i + aom_read(r, probs[i >> 1], NULL)]) > 0) continue; - return -i; -} - -#if CONFIG_EC_MULTISYMBOL static INLINE int aom_read_cdf_(aom_reader *r, const aom_cdf_prob *cdf, int nsymbs ACCT_STR_PARAM) { int ret; #if CONFIG_ANS (void)nsymbs; ret = rans_read(r, cdf); -#elif CONFIG_DAALA_EC - ret = daala_read_symbol(r, cdf, nsymbs); #else -#error \ - "CONFIG_EC_MULTISYMBOL is selected without a valid backing entropy " \ - "coder. Enable daala_ec or ans for a valid configuration." + ret = daala_read_symbol(r, cdf, nsymbs); #endif #if CONFIG_ACCOUNTING @@ -253,16 +218,11 @@ static INLINE int aom_read_tree_as_cdf(aom_reader *r, } while (i > 0); return -i; } -#endif // CONFIG_EC_MULTISYMBOL static INLINE int aom_read_tree_(aom_reader *r, const aom_tree_index *tree, const aom_prob *probs ACCT_STR_PARAM) { int ret; -#if CONFIG_EC_MULTISYMBOL ret = aom_read_tree_as_cdf(r, tree, probs); -#else - ret = aom_read_tree_as_bits(r, tree, probs); -#endif #if CONFIG_ACCOUNTING if (ACCT_STR_NAME) aom_process_accounting(r, ACCT_STR_NAME); #endif diff --git a/third_party/aom/aom_dsp/bitreader_buffer.c b/third_party/aom/aom_dsp/bitreader_buffer.c index 009682b4c..e51b1cc3a 100644 --- a/third_party/aom/aom_dsp/bitreader_buffer.c +++ b/third_party/aom/aom_dsp/bitreader_buffer.c @@ -24,7 +24,7 @@ int aom_rb_read_bit(struct aom_read_bit_buffer *rb) { rb->bit_offset = off + 1; return bit; } else { - rb->error_handler(rb->error_handler_data); + if (rb->error_handler) rb->error_handler(rb->error_handler_data); return 0; } } diff --git a/third_party/aom/aom_dsp/bitwriter.h b/third_party/aom/aom_dsp/bitwriter.h index 6e3fac260..588e47bf3 100644 --- a/third_party/aom/aom_dsp/bitwriter.h +++ b/third_party/aom/aom_dsp/bitwriter.h @@ -14,16 +14,11 @@ #include #include "./aom_config.h" -#if CONFIG_EC_ADAPT && !CONFIG_EC_MULTISYMBOL -#error "CONFIG_EC_ADAPT is enabled without enabling CONFIG_EC_MULTISYMBOL" -#endif #if CONFIG_ANS #include "aom_dsp/buf_ans.h" -#elif CONFIG_DAALA_EC -#include "aom_dsp/daalaboolwriter.h" #else -#include "aom_dsp/dkboolwriter.h" +#include "aom_dsp/daalaboolwriter.h" #endif #include "aom_dsp/prob.h" @@ -38,10 +33,8 @@ extern "C" { #if CONFIG_ANS typedef struct BufAnsCoder aom_writer; -#elif CONFIG_DAALA_EC -typedef struct daala_writer aom_writer; #else -typedef struct aom_dk_writer aom_writer; +typedef struct daala_writer aom_writer; #endif typedef struct TOKEN_STATS { @@ -72,10 +65,8 @@ static INLINE void aom_start_encode(aom_writer *bc, uint8_t *buffer) { (void)bc; (void)buffer; assert(0 && "buf_ans requires a more complicated startup procedure"); -#elif CONFIG_DAALA_EC - aom_daala_start_encode(bc, buffer); #else - aom_dk_start_encode(bc, buffer); + aom_daala_start_encode(bc, buffer); #endif } @@ -83,20 +74,16 @@ static INLINE void aom_stop_encode(aom_writer *bc) { #if CONFIG_ANS (void)bc; assert(0 && "buf_ans requires a more complicated shutdown procedure"); -#elif CONFIG_DAALA_EC - aom_daala_stop_encode(bc); #else - aom_dk_stop_encode(bc); + aom_daala_stop_encode(bc); #endif } static INLINE void aom_write(aom_writer *br, int bit, int probability) { #if CONFIG_ANS buf_rabs_write(br, bit, probability); -#elif CONFIG_DAALA_EC - aom_daala_write(br, bit, probability); #else - aom_dk_write(br, bit, probability); + aom_daala_write(br, bit, probability); #endif } @@ -113,7 +100,7 @@ static INLINE void aom_write_record(aom_writer *br, int bit, int probability, static INLINE void aom_write_bit(aom_writer *w, int bit) { #if CONFIG_ANS buf_rabs_write_bit(w, bit); -#elif CONFIG_DAALA_EC && CONFIG_RAWBITS +#elif CONFIG_RAWBITS // Note this uses raw bits and is not the same as aom_daala_write(r, 128); aom_daala_write_bit(w, bit); #else @@ -137,28 +124,6 @@ static INLINE void aom_write_literal(aom_writer *w, int data, int bits) { for (bit = bits - 1; bit >= 0; bit--) aom_write_bit(w, 1 & (data >> bit)); } -static INLINE void aom_write_tree_as_bits(aom_writer *w, - const aom_tree_index *tr, - const aom_prob *probs, int bits, - int len, aom_tree_index i) { - do { - const int bit = (bits >> --len) & 1; - aom_write(w, bit, probs[i >> 1]); - i = tr[i + bit]; - } while (len); -} - -static INLINE void aom_write_tree_as_bits_record( - aom_writer *w, const aom_tree_index *tr, const aom_prob *probs, int bits, - int len, aom_tree_index i, TOKEN_STATS *token_stats) { - do { - const int bit = (bits >> --len) & 1; - aom_write_record(w, bit, probs[i >> 1], token_stats); - i = tr[i + bit]; - } while (len); -} - -#if CONFIG_EC_MULTISYMBOL static INLINE void aom_write_cdf(aom_writer *w, int symb, const aom_cdf_prob *cdf, int nsymbs) { #if CONFIG_ANS @@ -167,12 +132,8 @@ static INLINE void aom_write_cdf(aom_writer *w, int symb, const aom_cdf_prob cum_prob = symb > 0 ? cdf[symb - 1] : 0; const aom_cdf_prob prob = cdf[symb] - cum_prob; buf_rans_write(w, cum_prob, prob); -#elif CONFIG_DAALA_EC - daala_write_symbol(w, symb, cdf, nsymbs); #else -#error \ - "CONFIG_EC_MULTISYMBOL is selected without a valid backing entropy " \ - "coder. Enable daala_ec or ans for a valid configuration." + daala_write_symbol(w, symb, cdf, nsymbs); #endif } @@ -223,16 +184,10 @@ static INLINE void aom_write_tree_as_cdf(aom_writer *w, } while (len); } -#endif // CONFIG_EC_MULTISYMBOL - static INLINE void aom_write_tree(aom_writer *w, const aom_tree_index *tree, const aom_prob *probs, int bits, int len, aom_tree_index i) { -#if CONFIG_EC_MULTISYMBOL aom_write_tree_as_cdf(w, tree, probs, bits, len, i); -#else - aom_write_tree_as_bits(w, tree, probs, bits, len, i); -#endif } static INLINE void aom_write_tree_record(aom_writer *w, @@ -240,12 +195,8 @@ static INLINE void aom_write_tree_record(aom_writer *w, const aom_prob *probs, int bits, int len, aom_tree_index i, TOKEN_STATS *token_stats) { -#if CONFIG_EC_MULTISYMBOL (void)token_stats; aom_write_tree_as_cdf(w, tree, probs, bits, len, i); -#else - aom_write_tree_as_bits_record(w, tree, probs, bits, len, i, token_stats); -#endif } #ifdef __cplusplus diff --git a/third_party/aom/aom_dsp/dkboolreader.c b/third_party/aom/aom_dsp/dkboolreader.c deleted file mode 100644 index 288d5f1ce..000000000 --- a/third_party/aom/aom_dsp/dkboolreader.c +++ /dev/null @@ -1,110 +0,0 @@ -/* - * 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 - -#include "./aom_config.h" - -#include "aom_dsp/dkboolreader.h" -#include "aom_dsp/prob.h" -#include "aom_dsp/aom_dsp_common.h" -#include "aom_ports/mem.h" -#include "aom_mem/aom_mem.h" -#include "aom_util/endian_inl.h" - -static INLINE int aom_dk_read_bit(struct aom_dk_reader *r) { - return aom_dk_read(r, 128); // aom_prob_half -} - -int aom_dk_reader_init(struct aom_dk_reader *r, const uint8_t *buffer, - size_t size, aom_decrypt_cb decrypt_cb, - void *decrypt_state) { - if (size && !buffer) { - return 1; - } else { - r->buffer_end = buffer + size; - r->buffer_start = r->buffer = buffer; - r->value = 0; - r->count = -8; - r->range = 255; - r->decrypt_cb = decrypt_cb; - r->decrypt_state = decrypt_state; - aom_dk_reader_fill(r); -#if CONFIG_ACCOUNTING - r->accounting = NULL; -#endif - return aom_dk_read_bit(r) != 0; // marker bit - } -} - -void aom_dk_reader_fill(struct aom_dk_reader *r) { - const uint8_t *const buffer_end = r->buffer_end; - const uint8_t *buffer = r->buffer; - const uint8_t *buffer_start = buffer; - BD_VALUE value = r->value; - int count = r->count; - const size_t bytes_left = buffer_end - buffer; - const size_t bits_left = bytes_left * CHAR_BIT; - int shift = BD_VALUE_SIZE - CHAR_BIT - (count + CHAR_BIT); - - if (r->decrypt_cb) { - size_t n = AOMMIN(sizeof(r->clear_buffer), bytes_left); - r->decrypt_cb(r->decrypt_state, buffer, r->clear_buffer, (int)n); - buffer = r->clear_buffer; - buffer_start = r->clear_buffer; - } - if (bits_left > BD_VALUE_SIZE) { - const int bits = (shift & 0xfffffff8) + CHAR_BIT; - BD_VALUE nv; - BD_VALUE big_endian_values; - memcpy(&big_endian_values, buffer, sizeof(BD_VALUE)); -#if SIZE_MAX == 0xffffffffffffffffULL - big_endian_values = HToBE64(big_endian_values); -#else - big_endian_values = HToBE32(big_endian_values); -#endif - nv = big_endian_values >> (BD_VALUE_SIZE - bits); - count += bits; - buffer += (bits >> 3); - value = r->value | (nv << (shift & 0x7)); - } else { - const int bits_over = (int)(shift + CHAR_BIT - (int)bits_left); - int loop_end = 0; - if (bits_over >= 0) { - count += LOTS_OF_BITS; - loop_end = bits_over; - } - - if (bits_over < 0 || bits_left) { - while (shift >= loop_end) { - count += CHAR_BIT; - value |= (BD_VALUE)*buffer++ << shift; - shift -= CHAR_BIT; - } - } - } - - // NOTE: Variable 'buffer' may not relate to 'r->buffer' after decryption, - // so we increase 'r->buffer' by the amount that 'buffer' moved, rather than - // assign 'buffer' to 'r->buffer'. - r->buffer += buffer - buffer_start; - r->value = value; - r->count = count; -} - -const uint8_t *aom_dk_reader_find_end(struct aom_dk_reader *r) { - // Find the end of the coded buffer - while (r->count > CHAR_BIT && r->count < BD_VALUE_SIZE) { - r->count -= CHAR_BIT; - r->buffer--; - } - return r->buffer; -} diff --git a/third_party/aom/aom_dsp/dkboolreader.h b/third_party/aom/aom_dsp/dkboolreader.h deleted file mode 100644 index f0bc84381..000000000 --- a/third_party/aom/aom_dsp/dkboolreader.h +++ /dev/null @@ -1,181 +0,0 @@ -/* - * 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_DKBOOLREADER_H_ -#define AOM_DSP_DKBOOLREADER_H_ - -#include -#include -#include - -#include "./aom_config.h" -#if CONFIG_BITSTREAM_DEBUG -#include -#include -#include "aom_util/debug_util.h" -#endif // CONFIG_BITSTREAM_DEBUG - -#include "aom_ports/mem.h" -#include "aom/aomdx.h" -#include "aom/aom_integer.h" -#include "aom_dsp/prob.h" -#if CONFIG_ACCOUNTING -#include "av1/decoder/accounting.h" -#endif - -#ifdef __cplusplus -extern "C" { -#endif - -typedef size_t BD_VALUE; - -#define BD_VALUE_SIZE ((int)sizeof(BD_VALUE) * CHAR_BIT) - -// This is meant to be a large, positive constant that can still be efficiently -// loaded as an immediate (on platforms like ARM, for example). -// Even relatively modest values like 100 would work fine. -#define LOTS_OF_BITS 0x40000000 - -struct aom_dk_reader { - // Be careful when reordering this struct, it may impact the cache negatively. - BD_VALUE value; - unsigned int range; - int count; - const uint8_t *buffer_start; - const uint8_t *buffer_end; - const uint8_t *buffer; - aom_decrypt_cb decrypt_cb; - void *decrypt_state; - uint8_t clear_buffer[sizeof(BD_VALUE) + 1]; -#if CONFIG_ACCOUNTING - Accounting *accounting; -#endif -}; - -int aom_dk_reader_init(struct aom_dk_reader *r, const uint8_t *buffer, - size_t size, aom_decrypt_cb decrypt_cb, - void *decrypt_state); - -void aom_dk_reader_fill(struct aom_dk_reader *r); - -const uint8_t *aom_dk_reader_find_end(struct aom_dk_reader *r); - -static INLINE uint32_t aom_dk_reader_tell(const struct aom_dk_reader *r) { - const uint32_t bits_read = - (uint32_t)((r->buffer - r->buffer_start) * CHAR_BIT); - const int count = - (r->count < LOTS_OF_BITS) ? r->count : r->count - LOTS_OF_BITS; - assert(r->buffer >= r->buffer_start); - return bits_read - (count + CHAR_BIT); -} - -/*The resolution of fractional-precision bit usage measurements, i.e., - 3 => 1/8th bits.*/ -#define DK_BITRES (3) - -static INLINE uint32_t aom_dk_reader_tell_frac(const struct aom_dk_reader *r) { - uint32_t num_bits; - uint32_t range; - int l; - int i; - num_bits = aom_dk_reader_tell(r) << DK_BITRES; - range = r->range; - l = 0; - for (i = DK_BITRES; i-- > 0;) { - int b; - range = range * range >> 7; - b = (int)(range >> 8); - l = l << 1 | b; - range >>= b; - } - return num_bits - l; -} - -static INLINE int aom_dk_reader_has_error(struct aom_dk_reader *r) { - // Check if we have reached the end of the buffer. - // - // Variable 'count' stores the number of bits in the 'value' buffer, minus - // 8. The top byte is part of the algorithm, and the remainder is buffered - // to be shifted into it. So if count == 8, the top 16 bits of 'value' are - // occupied, 8 for the algorithm and 8 in the buffer. - // - // When reading a byte from the user's buffer, count is filled with 8 and - // one byte is filled into the value buffer. When we reach the end of the - // data, count is additionally filled with LOTS_OF_BITS. So when - // count == LOTS_OF_BITS - 1, the user's data has been exhausted. - // - // 1 if we have tried to decode bits after the end of stream was encountered. - // 0 No error. - return r->count > BD_VALUE_SIZE && r->count < LOTS_OF_BITS; -} - -static INLINE int aom_dk_read(struct aom_dk_reader *r, int prob) { - unsigned int bit = 0; - BD_VALUE value; - BD_VALUE bigsplit; - int count; - unsigned int range; - unsigned int split = (r->range * prob + (256 - prob)) >> CHAR_BIT; - - if (r->count < 0) aom_dk_reader_fill(r); - - value = r->value; - count = r->count; - - bigsplit = (BD_VALUE)split << (BD_VALUE_SIZE - CHAR_BIT); - - range = split; - - if (value >= bigsplit) { - range = r->range - split; - value = value - bigsplit; - bit = 1; - } - - { - register int shift = aom_norm[range]; - range <<= shift; - value <<= shift; - count -= shift; - } - r->value = value; - r->count = count; - r->range = range; - -#if CONFIG_BITSTREAM_DEBUG - { - int ref_bit, ref_prob; - const int queue_r = bitstream_queue_get_read(); - const int frame_idx = bitstream_queue_get_frame_read(); - bitstream_queue_pop(&ref_bit, &ref_prob); - if (prob != ref_prob) { - fprintf( - stderr, - "\n *** prob error, frame_idx_r %d prob %d ref_prob %d queue_r %d\n", - frame_idx, prob, ref_prob, queue_r); - assert(0); - } - if ((int)bit != ref_bit) { - fprintf(stderr, "\n *** bit error, frame_idx_r %d bit %d ref_bit %d\n", - frame_idx, bit, ref_bit); - assert(0); - } - } -#endif // CONFIG_BITSTREAM_DEBUG - - return bit; -} - -#ifdef __cplusplus -} // extern "C" -#endif - -#endif // AOM_DSP_DKBOOLREADER_H_ diff --git a/third_party/aom/aom_dsp/dkboolwriter.c b/third_party/aom/aom_dsp/dkboolwriter.c deleted file mode 100644 index fc98e7c9b..000000000 --- a/third_party/aom/aom_dsp/dkboolwriter.c +++ /dev/null @@ -1,44 +0,0 @@ -/* - * 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 - -#include "./dkboolwriter.h" - -static INLINE void aom_dk_write_bit(aom_dk_writer *w, int bit) { - aom_dk_write(w, bit, 128); // aom_prob_half -} - -void aom_dk_start_encode(aom_dk_writer *br, uint8_t *source) { - br->lowvalue = 0; - br->range = 255; - br->count = -24; - br->buffer = source; - br->pos = 0; - aom_dk_write_bit(br, 0); -} - -void aom_dk_stop_encode(aom_dk_writer *br) { - int i; - -#if CONFIG_BITSTREAM_DEBUG - bitstream_queue_set_skip_write(1); -#endif // CONFIG_BITSTREAM_DEBUG - - for (i = 0; i < 32; i++) aom_dk_write_bit(br, 0); - -#if CONFIG_BITSTREAM_DEBUG - bitstream_queue_set_skip_write(0); -#endif // CONFIG_BITSTREAM_DEBUG - - // Ensure there's no ambigous collision with any index marker bytes - if ((br->buffer[br->pos - 1] & 0xe0) == 0xc0) br->buffer[br->pos++] = 0; -} diff --git a/third_party/aom/aom_dsp/dkboolwriter.h b/third_party/aom/aom_dsp/dkboolwriter.h deleted file mode 100644 index 835436885..000000000 --- a/third_party/aom/aom_dsp/dkboolwriter.h +++ /dev/null @@ -1,104 +0,0 @@ -/* - * 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_DKBOOLWRITER_H_ -#define AOM_DSP_DKBOOLWRITER_H_ - -#include "./aom_config.h" - -#if CONFIG_BITSTREAM_DEBUG -#include -#include "aom_util/debug_util.h" -#endif // CONFIG_BITSTREAM_DEBUG - -#include "aom_dsp/prob.h" -#include "aom_ports/mem.h" - -#ifdef __cplusplus -extern "C" { -#endif - -typedef struct aom_dk_writer { - unsigned int lowvalue; - unsigned int range; - int count; - unsigned int pos; - uint8_t *buffer; -} aom_dk_writer; - -void aom_dk_start_encode(aom_dk_writer *bc, uint8_t *buffer); -void aom_dk_stop_encode(aom_dk_writer *bc); - -static INLINE void aom_dk_write(aom_dk_writer *br, int bit, int probability) { - unsigned int split; - int count = br->count; - unsigned int range = br->range; - unsigned int lowvalue = br->lowvalue; - register int shift; - -#if CONFIG_BITSTREAM_DEBUG - // int queue_r = 0; - // int frame_idx_r = 0; - // int queue_w = bitstream_queue_get_write(); - // int frame_idx_w = bitstream_queue_get_frame_write(); - // if (frame_idx_w == frame_idx_r && queue_w == queue_r) { - // fprintf(stderr, "\n *** bitstream queue at frame_idx_w %d queue_w %d\n", - // frame_idx_w, queue_w); - // } - bitstream_queue_push(bit, probability); -#endif // CONFIG_BITSTREAM_DEBUG - - split = 1 + (((range - 1) * probability) >> 8); - - range = split; - - if (bit) { - lowvalue += split; - range = br->range - split; - } - - shift = aom_norm[range]; - - range <<= shift; - count += shift; - - if (count >= 0) { - int offset = shift - count; - - if ((lowvalue << (offset - 1)) & 0x80000000) { - int x = br->pos - 1; - - while (x >= 0 && br->buffer[x] == 0xff) { - br->buffer[x] = 0; - x--; - } - - br->buffer[x] += 1; - } - - br->buffer[br->pos++] = (lowvalue >> (24 - offset)); - lowvalue <<= offset; - shift = count; - lowvalue &= 0xffffff; - count -= 8; - } - - lowvalue <<= shift; - br->count = count; - br->lowvalue = lowvalue; - br->range = range; -} - -#ifdef __cplusplus -} // extern "C" -#endif - -#endif // AOM_DSP_DKBOOLWRITER_H_ diff --git a/third_party/aom/aom_dsp/intrapred.c b/third_party/aom/aom_dsp/intrapred.c index 1f0870b64..370d0374b 100644 --- a/third_party/aom/aom_dsp/intrapred.c +++ b/third_party/aom/aom_dsp/intrapred.c @@ -208,33 +208,30 @@ static const int sm_weight_log2_scale = 8; #if CONFIG_TX64X64 // max(block_size_wide[BLOCK_LARGEST], block_size_high[BLOCK_LARGEST]) #define MAX_BLOCK_DIM 64 -#define NUM_BLOCK_DIMS 6 // log2(MAX_BLOCK_DIM) #else #define MAX_BLOCK_DIM 32 -#define NUM_BLOCK_DIMS 5 #endif // CONFIG_TX64X64 -static const uint8_t sm_weight_arrays[NUM_BLOCK_DIMS][MAX_BLOCK_DIM] = { +static const uint8_t sm_weight_arrays[2 * MAX_BLOCK_DIM] = { + // Unused, because we always offset by bs, which is at least 2. + 0, 0, // bs = 2 - { 255, 128 }, + 255, 128, // bs = 4 - { 255, 149, 85, 64 }, + 255, 149, 85, 64, // bs = 8 - { 255, 197, 146, 105, 73, 50, 37, 32 }, + 255, 197, 146, 105, 73, 50, 37, 32, // bs = 16 - { 255, 225, 196, 170, 145, 123, 102, 84, 68, 54, 43, 33, 26, 20, 17, 16 }, + 255, 225, 196, 170, 145, 123, 102, 84, 68, 54, 43, 33, 26, 20, 17, 16, // bs = 32 - { - 255, 240, 225, 210, 196, 182, 169, 157, 145, 133, 122, - 111, 101, 92, 83, 74, 66, 59, 52, 45, 39, 34, - 29, 25, 21, 17, 14, 12, 10, 9, 8, 8 }, + 255, 240, 225, 210, 196, 182, 169, 157, 145, 133, 122, 111, 101, 92, 83, 74, + 66, 59, 52, 45, 39, 34, 29, 25, 21, 17, 14, 12, 10, 9, 8, 8, #if CONFIG_TX64X64 // bs = 64 - { 255, 248, 240, 233, 225, 218, 210, 203, 196, 189, 182, 176, 169, - 163, 156, 150, 144, 138, 133, 127, 121, 116, 111, 106, 101, 96, - 91, 86, 82, 77, 73, 69, 65, 61, 57, 54, 50, 47, 44, - 41, 38, 35, 32, 29, 27, 25, 22, 20, 18, 16, 15, 13, - 12, 10, 9, 8, 7, 6, 6, 5, 5, 4, 4, 4 }, + 255, 248, 240, 233, 225, 218, 210, 203, 196, 189, 182, 176, 169, 163, 156, + 150, 144, 138, 133, 127, 121, 116, 111, 106, 101, 96, 91, 86, 82, 77, 73, 69, + 65, 61, 57, 54, 50, 47, 44, 41, 38, 35, 32, 29, 27, 25, 22, 20, 18, 16, 15, + 13, 12, 10, 9, 8, 7, 6, 6, 5, 5, 4, 4, 4, #endif // CONFIG_TX64X64 }; @@ -250,10 +247,7 @@ static INLINE void smooth_predictor(uint8_t *dst, ptrdiff_t stride, int bs, const uint8_t *above, const uint8_t *left) { const uint8_t below_pred = left[bs - 1]; // estimated by bottom-left pixel const uint8_t right_pred = above[bs - 1]; // estimated by top-right pixel - const int arr_index = get_msb(bs) - 1; - assert(arr_index >= 0); - assert(arr_index < NUM_BLOCK_DIMS); - const uint8_t *const sm_weights = sm_weight_arrays[arr_index]; + const uint8_t *const sm_weights = sm_weight_arrays + bs; // scale = 2 * 2^sm_weight_log2_scale const int log2_scale = 1 + sm_weight_log2_scale; const uint16_t scale = (1 << sm_weight_log2_scale); @@ -277,6 +271,64 @@ static INLINE void smooth_predictor(uint8_t *dst, ptrdiff_t stride, int bs, } } +#if CONFIG_SMOOTH_HV +static INLINE void smooth_v_predictor(uint8_t *dst, ptrdiff_t stride, int bs, + const uint8_t *above, + const uint8_t *left) { + const uint8_t below_pred = left[bs - 1]; // estimated by bottom-left pixel + const uint8_t *const sm_weights = sm_weight_arrays + bs; + // scale = 2^sm_weight_log2_scale + const int log2_scale = sm_weight_log2_scale; + const uint16_t scale = (1 << sm_weight_log2_scale); + sm_weights_sanity_checks(sm_weights, scale, log2_scale + sizeof(*dst)); + + int r; + for (r = 0; r < bs; r++) { + int c; + for (c = 0; c < bs; ++c) { + const uint8_t pixels[] = { above[c], below_pred }; + const uint8_t weights[] = { sm_weights[r], scale - sm_weights[r] }; + uint32_t this_pred = 0; + assert(scale >= sm_weights[r]); + int i; + for (i = 0; i < 2; ++i) { + this_pred += weights[i] * pixels[i]; + } + dst[c] = clip_pixel(divide_round(this_pred, log2_scale)); + } + dst += stride; + } +} + +static INLINE void smooth_h_predictor(uint8_t *dst, ptrdiff_t stride, int bs, + const uint8_t *above, + const uint8_t *left) { + const uint8_t right_pred = above[bs - 1]; // estimated by top-right pixel + const uint8_t *const sm_weights = sm_weight_arrays + bs; + // scale = 2^sm_weight_log2_scale + const int log2_scale = sm_weight_log2_scale; + const uint16_t scale = (1 << sm_weight_log2_scale); + sm_weights_sanity_checks(sm_weights, scale, log2_scale + sizeof(*dst)); + + int r; + for (r = 0; r < bs; r++) { + int c; + for (c = 0; c < bs; ++c) { + const uint8_t pixels[] = { left[r], right_pred }; + const uint8_t weights[] = { sm_weights[c], scale - sm_weights[c] }; + uint32_t this_pred = 0; + assert(scale >= sm_weights[c]); + int i; + for (i = 0; i < 2; ++i) { + this_pred += weights[i] * pixels[i]; + } + dst[c] = clip_pixel(divide_round(this_pred, log2_scale)); + } + dst += stride; + } +} +#endif // CONFIG_SMOOTH_HV + #else static INLINE void tm_predictor(uint8_t *dst, ptrdiff_t stride, int bs, @@ -743,10 +795,7 @@ static INLINE void highbd_smooth_predictor(uint16_t *dst, ptrdiff_t stride, const uint16_t *left, int bd) { const uint16_t below_pred = left[bs - 1]; // estimated by bottom-left pixel const uint16_t right_pred = above[bs - 1]; // estimated by top-right pixel - const int arr_index = get_msb(bs) - 1; - assert(arr_index >= 0); - assert(arr_index < NUM_BLOCK_DIMS); - const uint8_t *const sm_weights = sm_weight_arrays[arr_index]; + const uint8_t *const sm_weights = sm_weight_arrays + bs; // scale = 2 * 2^sm_weight_log2_scale const int log2_scale = 1 + sm_weight_log2_scale; const uint16_t scale = (1 << sm_weight_log2_scale); @@ -770,6 +819,64 @@ static INLINE void highbd_smooth_predictor(uint16_t *dst, ptrdiff_t stride, } } +#if CONFIG_SMOOTH_HV +static INLINE void highbd_smooth_v_predictor(uint16_t *dst, ptrdiff_t stride, + int bs, const uint16_t *above, + const uint16_t *left, int bd) { + const uint16_t below_pred = left[bs - 1]; // estimated by bottom-left pixel + const uint8_t *const sm_weights = sm_weight_arrays + bs; + // scale = 2^sm_weight_log2_scale + const int log2_scale = sm_weight_log2_scale; + const uint16_t scale = (1 << sm_weight_log2_scale); + sm_weights_sanity_checks(sm_weights, scale, log2_scale + sizeof(*dst)); + + int r; + for (r = 0; r < bs; r++) { + int c; + for (c = 0; c < bs; ++c) { + const uint16_t pixels[] = { above[c], below_pred }; + const uint8_t weights[] = { sm_weights[r], scale - sm_weights[r] }; + uint32_t this_pred = 0; + assert(scale >= sm_weights[r]); + int i; + for (i = 0; i < 2; ++i) { + this_pred += weights[i] * pixels[i]; + } + dst[c] = clip_pixel_highbd(divide_round(this_pred, log2_scale), bd); + } + dst += stride; + } +} + +static INLINE void highbd_smooth_h_predictor(uint16_t *dst, ptrdiff_t stride, + int bs, const uint16_t *above, + const uint16_t *left, int bd) { + const uint16_t right_pred = above[bs - 1]; // estimated by top-right pixel + const uint8_t *const sm_weights = sm_weight_arrays + bs; + // scale = 2^sm_weight_log2_scale + const int log2_scale = sm_weight_log2_scale; + const uint16_t scale = (1 << sm_weight_log2_scale); + sm_weights_sanity_checks(sm_weights, scale, log2_scale + sizeof(*dst)); + + int r; + for (r = 0; r < bs; r++) { + int c; + for (c = 0; c < bs; ++c) { + const uint16_t pixels[] = { left[r], right_pred }; + const uint8_t weights[] = { sm_weights[c], scale - sm_weights[c] }; + uint32_t this_pred = 0; + assert(scale >= sm_weights[c]); + int i; + for (i = 0; i < 2; ++i) { + this_pred += weights[i] * pixels[i]; + } + dst[c] = clip_pixel_highbd(divide_round(this_pred, log2_scale), bd); + } + dst += stride; + } +} +#endif + #else static INLINE void highbd_tm_predictor(uint16_t *dst, ptrdiff_t stride, int bs, const uint16_t *above, @@ -879,6 +986,7 @@ static INLINE void highbd_dc_predictor(uint16_t *dst, ptrdiff_t stride, int bs, intra_pred_sized(type, 16) \ intra_pred_sized(type, 32) \ intra_pred_sized(type, 64) \ + intra_pred_highbd_sized(type, 2) \ intra_pred_highbd_sized(type, 4) \ intra_pred_highbd_sized(type, 8) \ intra_pred_highbd_sized(type, 16) \ @@ -958,8 +1066,12 @@ intra_pred_above_4x4(d153) intra_pred_allsizes(v) intra_pred_allsizes(h) #if CONFIG_ALT_INTRA -intra_pred_allsizes(paeth) intra_pred_allsizes(smooth) +#if CONFIG_SMOOTH_HV +intra_pred_allsizes(smooth_v) +intra_pred_allsizes(smooth_h) +#endif // CONFIG_SMOOTH_HV +intra_pred_allsizes(paeth) #else intra_pred_allsizes(tm) #endif // CONFIG_ALT_INTRA diff --git a/third_party/aom/aom_dsp/inv_txfm.c b/third_party/aom/aom_dsp/inv_txfm.c index bb995856a..6e7d8c928 100644 --- a/third_party/aom/aom_dsp/inv_txfm.c +++ b/third_party/aom/aom_dsp/inv_txfm.c @@ -1442,4 +1442,868 @@ void aom_highbd_idct4x4_1_add_c(const tran_low_t *input, uint8_t *dest8, } } +void aom_highbd_idct8_c(const tran_low_t *input, tran_low_t *output, int bd) { + tran_low_t step1[8], step2[8]; + tran_high_t temp1, temp2; + // stage 1 + step1[0] = input[0]; + step1[2] = input[4]; + step1[1] = input[2]; + step1[3] = input[6]; + temp1 = input[1] * cospi_28_64 - input[7] * cospi_4_64; + temp2 = input[1] * cospi_4_64 + input[7] * cospi_28_64; + step1[4] = HIGHBD_WRAPLOW(dct_const_round_shift(temp1), bd); + step1[7] = HIGHBD_WRAPLOW(dct_const_round_shift(temp2), bd); + temp1 = input[5] * cospi_12_64 - input[3] * cospi_20_64; + temp2 = input[5] * cospi_20_64 + input[3] * cospi_12_64; + step1[5] = HIGHBD_WRAPLOW(dct_const_round_shift(temp1), bd); + step1[6] = HIGHBD_WRAPLOW(dct_const_round_shift(temp2), bd); + + // stage 2 & stage 3 - even half + aom_highbd_idct4_c(step1, step1, bd); + + // stage 2 - odd half + step2[4] = HIGHBD_WRAPLOW(step1[4] + step1[5], bd); + step2[5] = HIGHBD_WRAPLOW(step1[4] - step1[5], bd); + step2[6] = HIGHBD_WRAPLOW(-step1[6] + step1[7], bd); + step2[7] = HIGHBD_WRAPLOW(step1[6] + step1[7], bd); + + // stage 3 - odd half + step1[4] = step2[4]; + temp1 = (step2[6] - step2[5]) * cospi_16_64; + temp2 = (step2[5] + step2[6]) * cospi_16_64; + step1[5] = HIGHBD_WRAPLOW(dct_const_round_shift(temp1), bd); + step1[6] = HIGHBD_WRAPLOW(dct_const_round_shift(temp2), bd); + step1[7] = step2[7]; + + // stage 4 + output[0] = HIGHBD_WRAPLOW(step1[0] + step1[7], bd); + output[1] = HIGHBD_WRAPLOW(step1[1] + step1[6], bd); + output[2] = HIGHBD_WRAPLOW(step1[2] + step1[5], bd); + output[3] = HIGHBD_WRAPLOW(step1[3] + step1[4], bd); + output[4] = HIGHBD_WRAPLOW(step1[3] - step1[4], bd); + output[5] = HIGHBD_WRAPLOW(step1[2] - step1[5], bd); + output[6] = HIGHBD_WRAPLOW(step1[1] - step1[6], bd); + output[7] = HIGHBD_WRAPLOW(step1[0] - step1[7], bd); +} + +void aom_highbd_iadst4_c(const tran_low_t *input, tran_low_t *output, int bd) { + tran_high_t s0, s1, s2, s3, s4, s5, s6, s7; + + tran_low_t x0 = input[0]; + tran_low_t x1 = input[1]; + tran_low_t x2 = input[2]; + tran_low_t x3 = input[3]; + (void)bd; + + if (!(x0 | x1 | x2 | x3)) { + memset(output, 0, 4 * sizeof(*output)); + return; + } + + s0 = sinpi_1_9 * x0; + s1 = sinpi_2_9 * x0; + s2 = sinpi_3_9 * x1; + s3 = sinpi_4_9 * x2; + s4 = sinpi_1_9 * x2; + s5 = sinpi_2_9 * x3; + s6 = sinpi_4_9 * x3; + s7 = (tran_high_t)HIGHBD_WRAPLOW(x0 - x2 + x3, bd); + + s0 = s0 + s3 + s5; + s1 = s1 - s4 - s6; + s3 = s2; + s2 = sinpi_3_9 * s7; + + // 1-D transform scaling factor is sqrt(2). + // The overall dynamic range is 14b (input) + 14b (multiplication scaling) + // + 1b (addition) = 29b. + // Hence the output bit depth is 15b. + output[0] = HIGHBD_WRAPLOW(dct_const_round_shift(s0 + s3), bd); + output[1] = HIGHBD_WRAPLOW(dct_const_round_shift(s1 + s3), bd); + output[2] = HIGHBD_WRAPLOW(dct_const_round_shift(s2), bd); + output[3] = HIGHBD_WRAPLOW(dct_const_round_shift(s0 + s1 - s3), bd); +} + +void aom_highbd_iadst8_c(const tran_low_t *input, tran_low_t *output, int bd) { + tran_high_t s0, s1, s2, s3, s4, s5, s6, s7; + + tran_low_t x0 = input[7]; + tran_low_t x1 = input[0]; + tran_low_t x2 = input[5]; + tran_low_t x3 = input[2]; + tran_low_t x4 = input[3]; + tran_low_t x5 = input[4]; + tran_low_t x6 = input[1]; + tran_low_t x7 = input[6]; + (void)bd; + + if (!(x0 | x1 | x2 | x3 | x4 | x5 | x6 | x7)) { + memset(output, 0, 8 * sizeof(*output)); + return; + } + + // stage 1 + s0 = cospi_2_64 * x0 + cospi_30_64 * x1; + s1 = cospi_30_64 * x0 - cospi_2_64 * x1; + s2 = cospi_10_64 * x2 + cospi_22_64 * x3; + s3 = cospi_22_64 * x2 - cospi_10_64 * x3; + s4 = cospi_18_64 * x4 + cospi_14_64 * x5; + s5 = cospi_14_64 * x4 - cospi_18_64 * x5; + s6 = cospi_26_64 * x6 + cospi_6_64 * x7; + s7 = cospi_6_64 * x6 - cospi_26_64 * x7; + + x0 = HIGHBD_WRAPLOW(dct_const_round_shift(s0 + s4), bd); + x1 = HIGHBD_WRAPLOW(dct_const_round_shift(s1 + s5), bd); + x2 = HIGHBD_WRAPLOW(dct_const_round_shift(s2 + s6), bd); + x3 = HIGHBD_WRAPLOW(dct_const_round_shift(s3 + s7), bd); + x4 = HIGHBD_WRAPLOW(dct_const_round_shift(s0 - s4), bd); + x5 = HIGHBD_WRAPLOW(dct_const_round_shift(s1 - s5), bd); + x6 = HIGHBD_WRAPLOW(dct_const_round_shift(s2 - s6), bd); + x7 = HIGHBD_WRAPLOW(dct_const_round_shift(s3 - s7), bd); + + // stage 2 + s0 = x0; + s1 = x1; + s2 = x2; + s3 = x3; + s4 = cospi_8_64 * x4 + cospi_24_64 * x5; + s5 = cospi_24_64 * x4 - cospi_8_64 * x5; + s6 = -cospi_24_64 * x6 + cospi_8_64 * x7; + s7 = cospi_8_64 * x6 + cospi_24_64 * x7; + + x0 = HIGHBD_WRAPLOW(s0 + s2, bd); + x1 = HIGHBD_WRAPLOW(s1 + s3, bd); + x2 = HIGHBD_WRAPLOW(s0 - s2, bd); + x3 = HIGHBD_WRAPLOW(s1 - s3, bd); + x4 = HIGHBD_WRAPLOW(dct_const_round_shift(s4 + s6), bd); + x5 = HIGHBD_WRAPLOW(dct_const_round_shift(s5 + s7), bd); + x6 = HIGHBD_WRAPLOW(dct_const_round_shift(s4 - s6), bd); + x7 = HIGHBD_WRAPLOW(dct_const_round_shift(s5 - s7), bd); + + // stage 3 + s2 = cospi_16_64 * (x2 + x3); + s3 = cospi_16_64 * (x2 - x3); + s6 = cospi_16_64 * (x6 + x7); + s7 = cospi_16_64 * (x6 - x7); + + x2 = HIGHBD_WRAPLOW(dct_const_round_shift(s2), bd); + x3 = HIGHBD_WRAPLOW(dct_const_round_shift(s3), bd); + x6 = HIGHBD_WRAPLOW(dct_const_round_shift(s6), bd); + x7 = HIGHBD_WRAPLOW(dct_const_round_shift(s7), bd); + + output[0] = HIGHBD_WRAPLOW(x0, bd); + output[1] = HIGHBD_WRAPLOW(-x4, bd); + output[2] = HIGHBD_WRAPLOW(x6, bd); + output[3] = HIGHBD_WRAPLOW(-x2, bd); + output[4] = HIGHBD_WRAPLOW(x3, bd); + output[5] = HIGHBD_WRAPLOW(-x7, bd); + output[6] = HIGHBD_WRAPLOW(x5, bd); + output[7] = HIGHBD_WRAPLOW(-x1, bd); +} + +void aom_highbd_idct16_c(const tran_low_t *input, tran_low_t *output, int bd) { + tran_low_t step1[16], step2[16]; + tran_high_t temp1, temp2; + (void)bd; + + // stage 1 + step1[0] = input[0 / 2]; + step1[1] = input[16 / 2]; + step1[2] = input[8 / 2]; + step1[3] = input[24 / 2]; + step1[4] = input[4 / 2]; + step1[5] = input[20 / 2]; + step1[6] = input[12 / 2]; + step1[7] = input[28 / 2]; + step1[8] = input[2 / 2]; + step1[9] = input[18 / 2]; + step1[10] = input[10 / 2]; + step1[11] = input[26 / 2]; + step1[12] = input[6 / 2]; + step1[13] = input[22 / 2]; + step1[14] = input[14 / 2]; + step1[15] = input[30 / 2]; + + // stage 2 + step2[0] = step1[0]; + step2[1] = step1[1]; + step2[2] = step1[2]; + step2[3] = step1[3]; + step2[4] = step1[4]; + step2[5] = step1[5]; + step2[6] = step1[6]; + step2[7] = step1[7]; + + temp1 = step1[8] * cospi_30_64 - step1[15] * cospi_2_64; + temp2 = step1[8] * cospi_2_64 + step1[15] * cospi_30_64; + step2[8] = HIGHBD_WRAPLOW(dct_const_round_shift(temp1), bd); + step2[15] = HIGHBD_WRAPLOW(dct_const_round_shift(temp2), bd); + + temp1 = step1[9] * cospi_14_64 - step1[14] * cospi_18_64; + temp2 = step1[9] * cospi_18_64 + step1[14] * cospi_14_64; + step2[9] = HIGHBD_WRAPLOW(dct_const_round_shift(temp1), bd); + step2[14] = HIGHBD_WRAPLOW(dct_const_round_shift(temp2), bd); + + temp1 = step1[10] * cospi_22_64 - step1[13] * cospi_10_64; + temp2 = step1[10] * cospi_10_64 + step1[13] * cospi_22_64; + step2[10] = HIGHBD_WRAPLOW(dct_const_round_shift(temp1), bd); + step2[13] = HIGHBD_WRAPLOW(dct_const_round_shift(temp2), bd); + + temp1 = step1[11] * cospi_6_64 - step1[12] * cospi_26_64; + temp2 = step1[11] * cospi_26_64 + step1[12] * cospi_6_64; + step2[11] = HIGHBD_WRAPLOW(dct_const_round_shift(temp1), bd); + step2[12] = HIGHBD_WRAPLOW(dct_const_round_shift(temp2), bd); + + // stage 3 + step1[0] = step2[0]; + step1[1] = step2[1]; + step1[2] = step2[2]; + step1[3] = step2[3]; + + temp1 = step2[4] * cospi_28_64 - step2[7] * cospi_4_64; + temp2 = step2[4] * cospi_4_64 + step2[7] * cospi_28_64; + step1[4] = HIGHBD_WRAPLOW(dct_const_round_shift(temp1), bd); + step1[7] = HIGHBD_WRAPLOW(dct_const_round_shift(temp2), bd); + temp1 = step2[5] * cospi_12_64 - step2[6] * cospi_20_64; + temp2 = step2[5] * cospi_20_64 + step2[6] * cospi_12_64; + step1[5] = HIGHBD_WRAPLOW(dct_const_round_shift(temp1), bd); + step1[6] = HIGHBD_WRAPLOW(dct_const_round_shift(temp2), bd); + + step1[8] = HIGHBD_WRAPLOW(step2[8] + step2[9], bd); + step1[9] = HIGHBD_WRAPLOW(step2[8] - step2[9], bd); + step1[10] = HIGHBD_WRAPLOW(-step2[10] + step2[11], bd); + step1[11] = HIGHBD_WRAPLOW(step2[10] + step2[11], bd); + step1[12] = HIGHBD_WRAPLOW(step2[12] + step2[13], bd); + step1[13] = HIGHBD_WRAPLOW(step2[12] - step2[13], bd); + step1[14] = HIGHBD_WRAPLOW(-step2[14] + step2[15], bd); + step1[15] = HIGHBD_WRAPLOW(step2[14] + step2[15], bd); + + // stage 4 + temp1 = (step1[0] + step1[1]) * cospi_16_64; + temp2 = (step1[0] - step1[1]) * cospi_16_64; + step2[0] = HIGHBD_WRAPLOW(dct_const_round_shift(temp1), bd); + step2[1] = HIGHBD_WRAPLOW(dct_const_round_shift(temp2), bd); + temp1 = step1[2] * cospi_24_64 - step1[3] * cospi_8_64; + temp2 = step1[2] * cospi_8_64 + step1[3] * cospi_24_64; + step2[2] = HIGHBD_WRAPLOW(dct_const_round_shift(temp1), bd); + step2[3] = HIGHBD_WRAPLOW(dct_const_round_shift(temp2), bd); + step2[4] = HIGHBD_WRAPLOW(step1[4] + step1[5], bd); + step2[5] = HIGHBD_WRAPLOW(step1[4] - step1[5], bd); + step2[6] = HIGHBD_WRAPLOW(-step1[6] + step1[7], bd); + step2[7] = HIGHBD_WRAPLOW(step1[6] + step1[7], bd); + + step2[8] = step1[8]; + step2[15] = step1[15]; + temp1 = -step1[9] * cospi_8_64 + step1[14] * cospi_24_64; + temp2 = step1[9] * cospi_24_64 + step1[14] * cospi_8_64; + step2[9] = HIGHBD_WRAPLOW(dct_const_round_shift(temp1), bd); + step2[14] = HIGHBD_WRAPLOW(dct_const_round_shift(temp2), bd); + temp1 = -step1[10] * cospi_24_64 - step1[13] * cospi_8_64; + temp2 = -step1[10] * cospi_8_64 + step1[13] * cospi_24_64; + step2[10] = HIGHBD_WRAPLOW(dct_const_round_shift(temp1), bd); + step2[13] = HIGHBD_WRAPLOW(dct_const_round_shift(temp2), bd); + step2[11] = step1[11]; + step2[12] = step1[12]; + + // stage 5 + step1[0] = HIGHBD_WRAPLOW(step2[0] + step2[3], bd); + step1[1] = HIGHBD_WRAPLOW(step2[1] + step2[2], bd); + step1[2] = HIGHBD_WRAPLOW(step2[1] - step2[2], bd); + step1[3] = HIGHBD_WRAPLOW(step2[0] - step2[3], bd); + step1[4] = step2[4]; + temp1 = (step2[6] - step2[5]) * cospi_16_64; + temp2 = (step2[5] + step2[6]) * cospi_16_64; + step1[5] = HIGHBD_WRAPLOW(dct_const_round_shift(temp1), bd); + step1[6] = HIGHBD_WRAPLOW(dct_const_round_shift(temp2), bd); + step1[7] = step2[7]; + + step1[8] = HIGHBD_WRAPLOW(step2[8] + step2[11], bd); + step1[9] = HIGHBD_WRAPLOW(step2[9] + step2[10], bd); + step1[10] = HIGHBD_WRAPLOW(step2[9] - step2[10], bd); + step1[11] = HIGHBD_WRAPLOW(step2[8] - step2[11], bd); + step1[12] = HIGHBD_WRAPLOW(-step2[12] + step2[15], bd); + step1[13] = HIGHBD_WRAPLOW(-step2[13] + step2[14], bd); + step1[14] = HIGHBD_WRAPLOW(step2[13] + step2[14], bd); + step1[15] = HIGHBD_WRAPLOW(step2[12] + step2[15], bd); + + // stage 6 + step2[0] = HIGHBD_WRAPLOW(step1[0] + step1[7], bd); + step2[1] = HIGHBD_WRAPLOW(step1[1] + step1[6], bd); + step2[2] = HIGHBD_WRAPLOW(step1[2] + step1[5], bd); + step2[3] = HIGHBD_WRAPLOW(step1[3] + step1[4], bd); + step2[4] = HIGHBD_WRAPLOW(step1[3] - step1[4], bd); + step2[5] = HIGHBD_WRAPLOW(step1[2] - step1[5], bd); + step2[6] = HIGHBD_WRAPLOW(step1[1] - step1[6], bd); + step2[7] = HIGHBD_WRAPLOW(step1[0] - step1[7], bd); + step2[8] = step1[8]; + step2[9] = step1[9]; + temp1 = (-step1[10] + step1[13]) * cospi_16_64; + temp2 = (step1[10] + step1[13]) * cospi_16_64; + step2[10] = HIGHBD_WRAPLOW(dct_const_round_shift(temp1), bd); + step2[13] = HIGHBD_WRAPLOW(dct_const_round_shift(temp2), bd); + temp1 = (-step1[11] + step1[12]) * cospi_16_64; + temp2 = (step1[11] + step1[12]) * cospi_16_64; + step2[11] = HIGHBD_WRAPLOW(dct_const_round_shift(temp1), bd); + step2[12] = HIGHBD_WRAPLOW(dct_const_round_shift(temp2), bd); + step2[14] = step1[14]; + step2[15] = step1[15]; + + // stage 7 + output[0] = HIGHBD_WRAPLOW(step2[0] + step2[15], bd); + output[1] = HIGHBD_WRAPLOW(step2[1] + step2[14], bd); + output[2] = HIGHBD_WRAPLOW(step2[2] + step2[13], bd); + output[3] = HIGHBD_WRAPLOW(step2[3] + step2[12], bd); + output[4] = HIGHBD_WRAPLOW(step2[4] + step2[11], bd); + output[5] = HIGHBD_WRAPLOW(step2[5] + step2[10], bd); + output[6] = HIGHBD_WRAPLOW(step2[6] + step2[9], bd); + output[7] = HIGHBD_WRAPLOW(step2[7] + step2[8], bd); + output[8] = HIGHBD_WRAPLOW(step2[7] - step2[8], bd); + output[9] = HIGHBD_WRAPLOW(step2[6] - step2[9], bd); + output[10] = HIGHBD_WRAPLOW(step2[5] - step2[10], bd); + output[11] = HIGHBD_WRAPLOW(step2[4] - step2[11], bd); + output[12] = HIGHBD_WRAPLOW(step2[3] - step2[12], bd); + output[13] = HIGHBD_WRAPLOW(step2[2] - step2[13], bd); + output[14] = HIGHBD_WRAPLOW(step2[1] - step2[14], bd); + output[15] = HIGHBD_WRAPLOW(step2[0] - step2[15], bd); +} + +void aom_highbd_iadst16_c(const tran_low_t *input, tran_low_t *output, int bd) { + tran_high_t s0, s1, s2, s3, s4, s5, s6, s7, s8; + tran_high_t s9, s10, s11, s12, s13, s14, s15; + + tran_low_t x0 = input[15]; + tran_low_t x1 = input[0]; + tran_low_t x2 = input[13]; + tran_low_t x3 = input[2]; + tran_low_t x4 = input[11]; + tran_low_t x5 = input[4]; + tran_low_t x6 = input[9]; + tran_low_t x7 = input[6]; + tran_low_t x8 = input[7]; + tran_low_t x9 = input[8]; + tran_low_t x10 = input[5]; + tran_low_t x11 = input[10]; + tran_low_t x12 = input[3]; + tran_low_t x13 = input[12]; + tran_low_t x14 = input[1]; + tran_low_t x15 = input[14]; + (void)bd; + + if (!(x0 | x1 | x2 | x3 | x4 | x5 | x6 | x7 | x8 | x9 | x10 | x11 | x12 | + x13 | x14 | x15)) { + memset(output, 0, 16 * sizeof(*output)); + return; + } + + // stage 1 + s0 = x0 * cospi_1_64 + x1 * cospi_31_64; + s1 = x0 * cospi_31_64 - x1 * cospi_1_64; + s2 = x2 * cospi_5_64 + x3 * cospi_27_64; + s3 = x2 * cospi_27_64 - x3 * cospi_5_64; + s4 = x4 * cospi_9_64 + x5 * cospi_23_64; + s5 = x4 * cospi_23_64 - x5 * cospi_9_64; + s6 = x6 * cospi_13_64 + x7 * cospi_19_64; + s7 = x6 * cospi_19_64 - x7 * cospi_13_64; + s8 = x8 * cospi_17_64 + x9 * cospi_15_64; + s9 = x8 * cospi_15_64 - x9 * cospi_17_64; + s10 = x10 * cospi_21_64 + x11 * cospi_11_64; + s11 = x10 * cospi_11_64 - x11 * cospi_21_64; + s12 = x12 * cospi_25_64 + x13 * cospi_7_64; + s13 = x12 * cospi_7_64 - x13 * cospi_25_64; + s14 = x14 * cospi_29_64 + x15 * cospi_3_64; + s15 = x14 * cospi_3_64 - x15 * cospi_29_64; + + x0 = HIGHBD_WRAPLOW(dct_const_round_shift(s0 + s8), bd); + x1 = HIGHBD_WRAPLOW(dct_const_round_shift(s1 + s9), bd); + x2 = HIGHBD_WRAPLOW(dct_const_round_shift(s2 + s10), bd); + x3 = HIGHBD_WRAPLOW(dct_const_round_shift(s3 + s11), bd); + x4 = HIGHBD_WRAPLOW(dct_const_round_shift(s4 + s12), bd); + x5 = HIGHBD_WRAPLOW(dct_const_round_shift(s5 + s13), bd); + x6 = HIGHBD_WRAPLOW(dct_const_round_shift(s6 + s14), bd); + x7 = HIGHBD_WRAPLOW(dct_const_round_shift(s7 + s15), bd); + x8 = HIGHBD_WRAPLOW(dct_const_round_shift(s0 - s8), bd); + x9 = HIGHBD_WRAPLOW(dct_const_round_shift(s1 - s9), bd); + x10 = HIGHBD_WRAPLOW(dct_const_round_shift(s2 - s10), bd); + x11 = HIGHBD_WRAPLOW(dct_const_round_shift(s3 - s11), bd); + x12 = HIGHBD_WRAPLOW(dct_const_round_shift(s4 - s12), bd); + x13 = HIGHBD_WRAPLOW(dct_const_round_shift(s5 - s13), bd); + x14 = HIGHBD_WRAPLOW(dct_const_round_shift(s6 - s14), bd); + x15 = HIGHBD_WRAPLOW(dct_const_round_shift(s7 - s15), bd); + + // stage 2 + s0 = x0; + s1 = x1; + s2 = x2; + s3 = x3; + s4 = x4; + s5 = x5; + s6 = x6; + s7 = x7; + s8 = x8 * cospi_4_64 + x9 * cospi_28_64; + s9 = x8 * cospi_28_64 - x9 * cospi_4_64; + s10 = x10 * cospi_20_64 + x11 * cospi_12_64; + s11 = x10 * cospi_12_64 - x11 * cospi_20_64; + s12 = -x12 * cospi_28_64 + x13 * cospi_4_64; + s13 = x12 * cospi_4_64 + x13 * cospi_28_64; + s14 = -x14 * cospi_12_64 + x15 * cospi_20_64; + s15 = x14 * cospi_20_64 + x15 * cospi_12_64; + + x0 = HIGHBD_WRAPLOW(s0 + s4, bd); + x1 = HIGHBD_WRAPLOW(s1 + s5, bd); + x2 = HIGHBD_WRAPLOW(s2 + s6, bd); + x3 = HIGHBD_WRAPLOW(s3 + s7, bd); + x4 = HIGHBD_WRAPLOW(s0 - s4, bd); + x5 = HIGHBD_WRAPLOW(s1 - s5, bd); + x6 = HIGHBD_WRAPLOW(s2 - s6, bd); + x7 = HIGHBD_WRAPLOW(s3 - s7, bd); + x8 = HIGHBD_WRAPLOW(dct_const_round_shift(s8 + s12), bd); + x9 = HIGHBD_WRAPLOW(dct_const_round_shift(s9 + s13), bd); + x10 = HIGHBD_WRAPLOW(dct_const_round_shift(s10 + s14), bd); + x11 = HIGHBD_WRAPLOW(dct_const_round_shift(s11 + s15), bd); + x12 = HIGHBD_WRAPLOW(dct_const_round_shift(s8 - s12), bd); + x13 = HIGHBD_WRAPLOW(dct_const_round_shift(s9 - s13), bd); + x14 = HIGHBD_WRAPLOW(dct_const_round_shift(s10 - s14), bd); + x15 = HIGHBD_WRAPLOW(dct_const_round_shift(s11 - s15), bd); + + // stage 3 + s0 = x0; + s1 = x1; + s2 = x2; + s3 = x3; + s4 = x4 * cospi_8_64 + x5 * cospi_24_64; + s5 = x4 * cospi_24_64 - x5 * cospi_8_64; + s6 = -x6 * cospi_24_64 + x7 * cospi_8_64; + s7 = x6 * cospi_8_64 + x7 * cospi_24_64; + s8 = x8; + s9 = x9; + s10 = x10; + s11 = x11; + s12 = x12 * cospi_8_64 + x13 * cospi_24_64; + s13 = x12 * cospi_24_64 - x13 * cospi_8_64; + s14 = -x14 * cospi_24_64 + x15 * cospi_8_64; + s15 = x14 * cospi_8_64 + x15 * cospi_24_64; + + x0 = HIGHBD_WRAPLOW(s0 + s2, bd); + x1 = HIGHBD_WRAPLOW(s1 + s3, bd); + x2 = HIGHBD_WRAPLOW(s0 - s2, bd); + x3 = HIGHBD_WRAPLOW(s1 - s3, bd); + x4 = HIGHBD_WRAPLOW(dct_const_round_shift(s4 + s6), bd); + x5 = HIGHBD_WRAPLOW(dct_const_round_shift(s5 + s7), bd); + x6 = HIGHBD_WRAPLOW(dct_const_round_shift(s4 - s6), bd); + x7 = HIGHBD_WRAPLOW(dct_const_round_shift(s5 - s7), bd); + x8 = HIGHBD_WRAPLOW(s8 + s10, bd); + x9 = HIGHBD_WRAPLOW(s9 + s11, bd); + x10 = HIGHBD_WRAPLOW(s8 - s10, bd); + x11 = HIGHBD_WRAPLOW(s9 - s11, bd); + x12 = HIGHBD_WRAPLOW(dct_const_round_shift(s12 + s14), bd); + x13 = HIGHBD_WRAPLOW(dct_const_round_shift(s13 + s15), bd); + x14 = HIGHBD_WRAPLOW(dct_const_round_shift(s12 - s14), bd); + x15 = HIGHBD_WRAPLOW(dct_const_round_shift(s13 - s15), bd); + + // stage 4 + s2 = (-cospi_16_64) * (x2 + x3); + s3 = cospi_16_64 * (x2 - x3); + s6 = cospi_16_64 * (x6 + x7); + s7 = cospi_16_64 * (-x6 + x7); + s10 = cospi_16_64 * (x10 + x11); + s11 = cospi_16_64 * (-x10 + x11); + s14 = (-cospi_16_64) * (x14 + x15); + s15 = cospi_16_64 * (x14 - x15); + + x2 = HIGHBD_WRAPLOW(dct_const_round_shift(s2), bd); + x3 = HIGHBD_WRAPLOW(dct_const_round_shift(s3), bd); + x6 = HIGHBD_WRAPLOW(dct_const_round_shift(s6), bd); + x7 = HIGHBD_WRAPLOW(dct_const_round_shift(s7), bd); + x10 = HIGHBD_WRAPLOW(dct_const_round_shift(s10), bd); + x11 = HIGHBD_WRAPLOW(dct_const_round_shift(s11), bd); + x14 = HIGHBD_WRAPLOW(dct_const_round_shift(s14), bd); + x15 = HIGHBD_WRAPLOW(dct_const_round_shift(s15), bd); + + output[0] = HIGHBD_WRAPLOW(x0, bd); + output[1] = HIGHBD_WRAPLOW(-x8, bd); + output[2] = HIGHBD_WRAPLOW(x12, bd); + output[3] = HIGHBD_WRAPLOW(-x4, bd); + output[4] = HIGHBD_WRAPLOW(x6, bd); + output[5] = HIGHBD_WRAPLOW(x14, bd); + output[6] = HIGHBD_WRAPLOW(x10, bd); + output[7] = HIGHBD_WRAPLOW(x2, bd); + output[8] = HIGHBD_WRAPLOW(x3, bd); + output[9] = HIGHBD_WRAPLOW(x11, bd); + output[10] = HIGHBD_WRAPLOW(x15, bd); + output[11] = HIGHBD_WRAPLOW(x7, bd); + output[12] = HIGHBD_WRAPLOW(x5, bd); + output[13] = HIGHBD_WRAPLOW(-x13, bd); + output[14] = HIGHBD_WRAPLOW(x9, bd); + output[15] = HIGHBD_WRAPLOW(-x1, bd); +} + +void aom_highbd_idct32_c(const tran_low_t *input, tran_low_t *output, int bd) { + tran_low_t step1[32], step2[32]; + tran_high_t temp1, temp2; + (void)bd; + + // stage 1 + step1[0] = input[0]; + step1[1] = input[16]; + step1[2] = input[8]; + step1[3] = input[24]; + step1[4] = input[4]; + step1[5] = input[20]; + step1[6] = input[12]; + step1[7] = input[28]; + step1[8] = input[2]; + step1[9] = input[18]; + step1[10] = input[10]; + step1[11] = input[26]; + step1[12] = input[6]; + step1[13] = input[22]; + step1[14] = input[14]; + step1[15] = input[30]; + + temp1 = input[1] * cospi_31_64 - input[31] * cospi_1_64; + temp2 = input[1] * cospi_1_64 + input[31] * cospi_31_64; + step1[16] = HIGHBD_WRAPLOW(dct_const_round_shift(temp1), bd); + step1[31] = HIGHBD_WRAPLOW(dct_const_round_shift(temp2), bd); + + temp1 = input[17] * cospi_15_64 - input[15] * cospi_17_64; + temp2 = input[17] * cospi_17_64 + input[15] * cospi_15_64; + step1[17] = HIGHBD_WRAPLOW(dct_const_round_shift(temp1), bd); + step1[30] = HIGHBD_WRAPLOW(dct_const_round_shift(temp2), bd); + + temp1 = input[9] * cospi_23_64 - input[23] * cospi_9_64; + temp2 = input[9] * cospi_9_64 + input[23] * cospi_23_64; + step1[18] = HIGHBD_WRAPLOW(dct_const_round_shift(temp1), bd); + step1[29] = HIGHBD_WRAPLOW(dct_const_round_shift(temp2), bd); + + temp1 = input[25] * cospi_7_64 - input[7] * cospi_25_64; + temp2 = input[25] * cospi_25_64 + input[7] * cospi_7_64; + step1[19] = HIGHBD_WRAPLOW(dct_const_round_shift(temp1), bd); + step1[28] = HIGHBD_WRAPLOW(dct_const_round_shift(temp2), bd); + + temp1 = input[5] * cospi_27_64 - input[27] * cospi_5_64; + temp2 = input[5] * cospi_5_64 + input[27] * cospi_27_64; + step1[20] = HIGHBD_WRAPLOW(dct_const_round_shift(temp1), bd); + step1[27] = HIGHBD_WRAPLOW(dct_const_round_shift(temp2), bd); + + temp1 = input[21] * cospi_11_64 - input[11] * cospi_21_64; + temp2 = input[21] * cospi_21_64 + input[11] * cospi_11_64; + step1[21] = HIGHBD_WRAPLOW(dct_const_round_shift(temp1), bd); + step1[26] = HIGHBD_WRAPLOW(dct_const_round_shift(temp2), bd); + + temp1 = input[13] * cospi_19_64 - input[19] * cospi_13_64; + temp2 = input[13] * cospi_13_64 + input[19] * cospi_19_64; + step1[22] = HIGHBD_WRAPLOW(dct_const_round_shift(temp1), bd); + step1[25] = HIGHBD_WRAPLOW(dct_const_round_shift(temp2), bd); + + temp1 = input[29] * cospi_3_64 - input[3] * cospi_29_64; + temp2 = input[29] * cospi_29_64 + input[3] * cospi_3_64; + step1[23] = HIGHBD_WRAPLOW(dct_const_round_shift(temp1), bd); + step1[24] = HIGHBD_WRAPLOW(dct_const_round_shift(temp2), bd); + + // stage 2 + step2[0] = step1[0]; + step2[1] = step1[1]; + step2[2] = step1[2]; + step2[3] = step1[3]; + step2[4] = step1[4]; + step2[5] = step1[5]; + step2[6] = step1[6]; + step2[7] = step1[7]; + + temp1 = step1[8] * cospi_30_64 - step1[15] * cospi_2_64; + temp2 = step1[8] * cospi_2_64 + step1[15] * cospi_30_64; + step2[8] = HIGHBD_WRAPLOW(dct_const_round_shift(temp1), bd); + step2[15] = HIGHBD_WRAPLOW(dct_const_round_shift(temp2), bd); + + temp1 = step1[9] * cospi_14_64 - step1[14] * cospi_18_64; + temp2 = step1[9] * cospi_18_64 + step1[14] * cospi_14_64; + step2[9] = HIGHBD_WRAPLOW(dct_const_round_shift(temp1), bd); + step2[14] = HIGHBD_WRAPLOW(dct_const_round_shift(temp2), bd); + + temp1 = step1[10] * cospi_22_64 - step1[13] * cospi_10_64; + temp2 = step1[10] * cospi_10_64 + step1[13] * cospi_22_64; + step2[10] = HIGHBD_WRAPLOW(dct_const_round_shift(temp1), bd); + step2[13] = HIGHBD_WRAPLOW(dct_const_round_shift(temp2), bd); + + temp1 = step1[11] * cospi_6_64 - step1[12] * cospi_26_64; + temp2 = step1[11] * cospi_26_64 + step1[12] * cospi_6_64; + step2[11] = HIGHBD_WRAPLOW(dct_const_round_shift(temp1), bd); + step2[12] = HIGHBD_WRAPLOW(dct_const_round_shift(temp2), bd); + + step2[16] = HIGHBD_WRAPLOW(step1[16] + step1[17], bd); + step2[17] = HIGHBD_WRAPLOW(step1[16] - step1[17], bd); + step2[18] = HIGHBD_WRAPLOW(-step1[18] + step1[19], bd); + step2[19] = HIGHBD_WRAPLOW(step1[18] + step1[19], bd); + step2[20] = HIGHBD_WRAPLOW(step1[20] + step1[21], bd); + step2[21] = HIGHBD_WRAPLOW(step1[20] - step1[21], bd); + step2[22] = HIGHBD_WRAPLOW(-step1[22] + step1[23], bd); + step2[23] = HIGHBD_WRAPLOW(step1[22] + step1[23], bd); + step2[24] = HIGHBD_WRAPLOW(step1[24] + step1[25], bd); + step2[25] = HIGHBD_WRAPLOW(step1[24] - step1[25], bd); + step2[26] = HIGHBD_WRAPLOW(-step1[26] + step1[27], bd); + step2[27] = HIGHBD_WRAPLOW(step1[26] + step1[27], bd); + step2[28] = HIGHBD_WRAPLOW(step1[28] + step1[29], bd); + step2[29] = HIGHBD_WRAPLOW(step1[28] - step1[29], bd); + step2[30] = HIGHBD_WRAPLOW(-step1[30] + step1[31], bd); + step2[31] = HIGHBD_WRAPLOW(step1[30] + step1[31], bd); + + // stage 3 + step1[0] = step2[0]; + step1[1] = step2[1]; + step1[2] = step2[2]; + step1[3] = step2[3]; + + temp1 = step2[4] * cospi_28_64 - step2[7] * cospi_4_64; + temp2 = step2[4] * cospi_4_64 + step2[7] * cospi_28_64; + step1[4] = HIGHBD_WRAPLOW(dct_const_round_shift(temp1), bd); + step1[7] = HIGHBD_WRAPLOW(dct_const_round_shift(temp2), bd); + temp1 = step2[5] * cospi_12_64 - step2[6] * cospi_20_64; + temp2 = step2[5] * cospi_20_64 + step2[6] * cospi_12_64; + step1[5] = HIGHBD_WRAPLOW(dct_const_round_shift(temp1), bd); + step1[6] = HIGHBD_WRAPLOW(dct_const_round_shift(temp2), bd); + + step1[8] = HIGHBD_WRAPLOW(step2[8] + step2[9], bd); + step1[9] = HIGHBD_WRAPLOW(step2[8] - step2[9], bd); + step1[10] = HIGHBD_WRAPLOW(-step2[10] + step2[11], bd); + step1[11] = HIGHBD_WRAPLOW(step2[10] + step2[11], bd); + step1[12] = HIGHBD_WRAPLOW(step2[12] + step2[13], bd); + step1[13] = HIGHBD_WRAPLOW(step2[12] - step2[13], bd); + step1[14] = HIGHBD_WRAPLOW(-step2[14] + step2[15], bd); + step1[15] = HIGHBD_WRAPLOW(step2[14] + step2[15], bd); + + step1[16] = step2[16]; + step1[31] = step2[31]; + temp1 = -step2[17] * cospi_4_64 + step2[30] * cospi_28_64; + temp2 = step2[17] * cospi_28_64 + step2[30] * cospi_4_64; + step1[17] = HIGHBD_WRAPLOW(dct_const_round_shift(temp1), bd); + step1[30] = HIGHBD_WRAPLOW(dct_const_round_shift(temp2), bd); + temp1 = -step2[18] * cospi_28_64 - step2[29] * cospi_4_64; + temp2 = -step2[18] * cospi_4_64 + step2[29] * cospi_28_64; + step1[18] = HIGHBD_WRAPLOW(dct_const_round_shift(temp1), bd); + step1[29] = HIGHBD_WRAPLOW(dct_const_round_shift(temp2), bd); + step1[19] = step2[19]; + step1[20] = step2[20]; + temp1 = -step2[21] * cospi_20_64 + step2[26] * cospi_12_64; + temp2 = step2[21] * cospi_12_64 + step2[26] * cospi_20_64; + step1[21] = HIGHBD_WRAPLOW(dct_const_round_shift(temp1), bd); + step1[26] = HIGHBD_WRAPLOW(dct_const_round_shift(temp2), bd); + temp1 = -step2[22] * cospi_12_64 - step2[25] * cospi_20_64; + temp2 = -step2[22] * cospi_20_64 + step2[25] * cospi_12_64; + step1[22] = HIGHBD_WRAPLOW(dct_const_round_shift(temp1), bd); + step1[25] = HIGHBD_WRAPLOW(dct_const_round_shift(temp2), bd); + step1[23] = step2[23]; + step1[24] = step2[24]; + step1[27] = step2[27]; + step1[28] = step2[28]; + + // stage 4 + temp1 = (step1[0] + step1[1]) * cospi_16_64; + temp2 = (step1[0] - step1[1]) * cospi_16_64; + step2[0] = HIGHBD_WRAPLOW(dct_const_round_shift(temp1), bd); + step2[1] = HIGHBD_WRAPLOW(dct_const_round_shift(temp2), bd); + temp1 = step1[2] * cospi_24_64 - step1[3] * cospi_8_64; + temp2 = step1[2] * cospi_8_64 + step1[3] * cospi_24_64; + step2[2] = HIGHBD_WRAPLOW(dct_const_round_shift(temp1), bd); + step2[3] = HIGHBD_WRAPLOW(dct_const_round_shift(temp2), bd); + step2[4] = HIGHBD_WRAPLOW(step1[4] + step1[5], bd); + step2[5] = HIGHBD_WRAPLOW(step1[4] - step1[5], bd); + step2[6] = HIGHBD_WRAPLOW(-step1[6] + step1[7], bd); + step2[7] = HIGHBD_WRAPLOW(step1[6] + step1[7], bd); + + step2[8] = step1[8]; + step2[15] = step1[15]; + temp1 = -step1[9] * cospi_8_64 + step1[14] * cospi_24_64; + temp2 = step1[9] * cospi_24_64 + step1[14] * cospi_8_64; + step2[9] = HIGHBD_WRAPLOW(dct_const_round_shift(temp1), bd); + step2[14] = HIGHBD_WRAPLOW(dct_const_round_shift(temp2), bd); + temp1 = -step1[10] * cospi_24_64 - step1[13] * cospi_8_64; + temp2 = -step1[10] * cospi_8_64 + step1[13] * cospi_24_64; + step2[10] = HIGHBD_WRAPLOW(dct_const_round_shift(temp1), bd); + step2[13] = HIGHBD_WRAPLOW(dct_const_round_shift(temp2), bd); + step2[11] = step1[11]; + step2[12] = step1[12]; + + step2[16] = HIGHBD_WRAPLOW(step1[16] + step1[19], bd); + step2[17] = HIGHBD_WRAPLOW(step1[17] + step1[18], bd); + step2[18] = HIGHBD_WRAPLOW(step1[17] - step1[18], bd); + step2[19] = HIGHBD_WRAPLOW(step1[16] - step1[19], bd); + step2[20] = HIGHBD_WRAPLOW(-step1[20] + step1[23], bd); + step2[21] = HIGHBD_WRAPLOW(-step1[21] + step1[22], bd); + step2[22] = HIGHBD_WRAPLOW(step1[21] + step1[22], bd); + step2[23] = HIGHBD_WRAPLOW(step1[20] + step1[23], bd); + + step2[24] = HIGHBD_WRAPLOW(step1[24] + step1[27], bd); + step2[25] = HIGHBD_WRAPLOW(step1[25] + step1[26], bd); + step2[26] = HIGHBD_WRAPLOW(step1[25] - step1[26], bd); + step2[27] = HIGHBD_WRAPLOW(step1[24] - step1[27], bd); + step2[28] = HIGHBD_WRAPLOW(-step1[28] + step1[31], bd); + step2[29] = HIGHBD_WRAPLOW(-step1[29] + step1[30], bd); + step2[30] = HIGHBD_WRAPLOW(step1[29] + step1[30], bd); + step2[31] = HIGHBD_WRAPLOW(step1[28] + step1[31], bd); + + // stage 5 + step1[0] = HIGHBD_WRAPLOW(step2[0] + step2[3], bd); + step1[1] = HIGHBD_WRAPLOW(step2[1] + step2[2], bd); + step1[2] = HIGHBD_WRAPLOW(step2[1] - step2[2], bd); + step1[3] = HIGHBD_WRAPLOW(step2[0] - step2[3], bd); + step1[4] = step2[4]; + temp1 = (step2[6] - step2[5]) * cospi_16_64; + temp2 = (step2[5] + step2[6]) * cospi_16_64; + step1[5] = HIGHBD_WRAPLOW(dct_const_round_shift(temp1), bd); + step1[6] = HIGHBD_WRAPLOW(dct_const_round_shift(temp2), bd); + step1[7] = step2[7]; + + step1[8] = HIGHBD_WRAPLOW(step2[8] + step2[11], bd); + step1[9] = HIGHBD_WRAPLOW(step2[9] + step2[10], bd); + step1[10] = HIGHBD_WRAPLOW(step2[9] - step2[10], bd); + step1[11] = HIGHBD_WRAPLOW(step2[8] - step2[11], bd); + step1[12] = HIGHBD_WRAPLOW(-step2[12] + step2[15], bd); + step1[13] = HIGHBD_WRAPLOW(-step2[13] + step2[14], bd); + step1[14] = HIGHBD_WRAPLOW(step2[13] + step2[14], bd); + step1[15] = HIGHBD_WRAPLOW(step2[12] + step2[15], bd); + + step1[16] = step2[16]; + step1[17] = step2[17]; + temp1 = -step2[18] * cospi_8_64 + step2[29] * cospi_24_64; + temp2 = step2[18] * cospi_24_64 + step2[29] * cospi_8_64; + step1[18] = HIGHBD_WRAPLOW(dct_const_round_shift(temp1), bd); + step1[29] = HIGHBD_WRAPLOW(dct_const_round_shift(temp2), bd); + temp1 = -step2[19] * cospi_8_64 + step2[28] * cospi_24_64; + temp2 = step2[19] * cospi_24_64 + step2[28] * cospi_8_64; + step1[19] = HIGHBD_WRAPLOW(dct_const_round_shift(temp1), bd); + step1[28] = HIGHBD_WRAPLOW(dct_const_round_shift(temp2), bd); + temp1 = -step2[20] * cospi_24_64 - step2[27] * cospi_8_64; + temp2 = -step2[20] * cospi_8_64 + step2[27] * cospi_24_64; + step1[20] = HIGHBD_WRAPLOW(dct_const_round_shift(temp1), bd); + step1[27] = HIGHBD_WRAPLOW(dct_const_round_shift(temp2), bd); + temp1 = -step2[21] * cospi_24_64 - step2[26] * cospi_8_64; + temp2 = -step2[21] * cospi_8_64 + step2[26] * cospi_24_64; + step1[21] = HIGHBD_WRAPLOW(dct_const_round_shift(temp1), bd); + step1[26] = HIGHBD_WRAPLOW(dct_const_round_shift(temp2), bd); + step1[22] = step2[22]; + step1[23] = step2[23]; + step1[24] = step2[24]; + step1[25] = step2[25]; + step1[30] = step2[30]; + step1[31] = step2[31]; + + // stage 6 + step2[0] = HIGHBD_WRAPLOW(step1[0] + step1[7], bd); + step2[1] = HIGHBD_WRAPLOW(step1[1] + step1[6], bd); + step2[2] = HIGHBD_WRAPLOW(step1[2] + step1[5], bd); + step2[3] = HIGHBD_WRAPLOW(step1[3] + step1[4], bd); + step2[4] = HIGHBD_WRAPLOW(step1[3] - step1[4], bd); + step2[5] = HIGHBD_WRAPLOW(step1[2] - step1[5], bd); + step2[6] = HIGHBD_WRAPLOW(step1[1] - step1[6], bd); + step2[7] = HIGHBD_WRAPLOW(step1[0] - step1[7], bd); + step2[8] = step1[8]; + step2[9] = step1[9]; + temp1 = (-step1[10] + step1[13]) * cospi_16_64; + temp2 = (step1[10] + step1[13]) * cospi_16_64; + step2[10] = HIGHBD_WRAPLOW(dct_const_round_shift(temp1), bd); + step2[13] = HIGHBD_WRAPLOW(dct_const_round_shift(temp2), bd); + temp1 = (-step1[11] + step1[12]) * cospi_16_64; + temp2 = (step1[11] + step1[12]) * cospi_16_64; + step2[11] = HIGHBD_WRAPLOW(dct_const_round_shift(temp1), bd); + step2[12] = HIGHBD_WRAPLOW(dct_const_round_shift(temp2), bd); + step2[14] = step1[14]; + step2[15] = step1[15]; + + step2[16] = HIGHBD_WRAPLOW(step1[16] + step1[23], bd); + step2[17] = HIGHBD_WRAPLOW(step1[17] + step1[22], bd); + step2[18] = HIGHBD_WRAPLOW(step1[18] + step1[21], bd); + step2[19] = HIGHBD_WRAPLOW(step1[19] + step1[20], bd); + step2[20] = HIGHBD_WRAPLOW(step1[19] - step1[20], bd); + step2[21] = HIGHBD_WRAPLOW(step1[18] - step1[21], bd); + step2[22] = HIGHBD_WRAPLOW(step1[17] - step1[22], bd); + step2[23] = HIGHBD_WRAPLOW(step1[16] - step1[23], bd); + + step2[24] = HIGHBD_WRAPLOW(-step1[24] + step1[31], bd); + step2[25] = HIGHBD_WRAPLOW(-step1[25] + step1[30], bd); + step2[26] = HIGHBD_WRAPLOW(-step1[26] + step1[29], bd); + step2[27] = HIGHBD_WRAPLOW(-step1[27] + step1[28], bd); + step2[28] = HIGHBD_WRAPLOW(step1[27] + step1[28], bd); + step2[29] = HIGHBD_WRAPLOW(step1[26] + step1[29], bd); + step2[30] = HIGHBD_WRAPLOW(step1[25] + step1[30], bd); + step2[31] = HIGHBD_WRAPLOW(step1[24] + step1[31], bd); + + // stage 7 + step1[0] = HIGHBD_WRAPLOW(step2[0] + step2[15], bd); + step1[1] = HIGHBD_WRAPLOW(step2[1] + step2[14], bd); + step1[2] = HIGHBD_WRAPLOW(step2[2] + step2[13], bd); + step1[3] = HIGHBD_WRAPLOW(step2[3] + step2[12], bd); + step1[4] = HIGHBD_WRAPLOW(step2[4] + step2[11], bd); + step1[5] = HIGHBD_WRAPLOW(step2[5] + step2[10], bd); + step1[6] = HIGHBD_WRAPLOW(step2[6] + step2[9], bd); + step1[7] = HIGHBD_WRAPLOW(step2[7] + step2[8], bd); + step1[8] = HIGHBD_WRAPLOW(step2[7] - step2[8], bd); + step1[9] = HIGHBD_WRAPLOW(step2[6] - step2[9], bd); + step1[10] = HIGHBD_WRAPLOW(step2[5] - step2[10], bd); + step1[11] = HIGHBD_WRAPLOW(step2[4] - step2[11], bd); + step1[12] = HIGHBD_WRAPLOW(step2[3] - step2[12], bd); + step1[13] = HIGHBD_WRAPLOW(step2[2] - step2[13], bd); + step1[14] = HIGHBD_WRAPLOW(step2[1] - step2[14], bd); + step1[15] = HIGHBD_WRAPLOW(step2[0] - step2[15], bd); + + step1[16] = step2[16]; + step1[17] = step2[17]; + step1[18] = step2[18]; + step1[19] = step2[19]; + temp1 = (-step2[20] + step2[27]) * cospi_16_64; + temp2 = (step2[20] + step2[27]) * cospi_16_64; + step1[20] = HIGHBD_WRAPLOW(dct_const_round_shift(temp1), bd); + step1[27] = HIGHBD_WRAPLOW(dct_const_round_shift(temp2), bd); + temp1 = (-step2[21] + step2[26]) * cospi_16_64; + temp2 = (step2[21] + step2[26]) * cospi_16_64; + step1[21] = HIGHBD_WRAPLOW(dct_const_round_shift(temp1), bd); + step1[26] = HIGHBD_WRAPLOW(dct_const_round_shift(temp2), bd); + temp1 = (-step2[22] + step2[25]) * cospi_16_64; + temp2 = (step2[22] + step2[25]) * cospi_16_64; + step1[22] = HIGHBD_WRAPLOW(dct_const_round_shift(temp1), bd); + step1[25] = HIGHBD_WRAPLOW(dct_const_round_shift(temp2), bd); + temp1 = (-step2[23] + step2[24]) * cospi_16_64; + temp2 = (step2[23] + step2[24]) * cospi_16_64; + step1[23] = HIGHBD_WRAPLOW(dct_const_round_shift(temp1), bd); + step1[24] = HIGHBD_WRAPLOW(dct_const_round_shift(temp2), bd); + step1[28] = step2[28]; + step1[29] = step2[29]; + step1[30] = step2[30]; + step1[31] = step2[31]; + + // final stage + output[0] = HIGHBD_WRAPLOW(step1[0] + step1[31], bd); + output[1] = HIGHBD_WRAPLOW(step1[1] + step1[30], bd); + output[2] = HIGHBD_WRAPLOW(step1[2] + step1[29], bd); + output[3] = HIGHBD_WRAPLOW(step1[3] + step1[28], bd); + output[4] = HIGHBD_WRAPLOW(step1[4] + step1[27], bd); + output[5] = HIGHBD_WRAPLOW(step1[5] + step1[26], bd); + output[6] = HIGHBD_WRAPLOW(step1[6] + step1[25], bd); + output[7] = HIGHBD_WRAPLOW(step1[7] + step1[24], bd); + output[8] = HIGHBD_WRAPLOW(step1[8] + step1[23], bd); + output[9] = HIGHBD_WRAPLOW(step1[9] + step1[22], bd); + output[10] = HIGHBD_WRAPLOW(step1[10] + step1[21], bd); + output[11] = HIGHBD_WRAPLOW(step1[11] + step1[20], bd); + output[12] = HIGHBD_WRAPLOW(step1[12] + step1[19], bd); + output[13] = HIGHBD_WRAPLOW(step1[13] + step1[18], bd); + output[14] = HIGHBD_WRAPLOW(step1[14] + step1[17], bd); + output[15] = HIGHBD_WRAPLOW(step1[15] + step1[16], bd); + output[16] = HIGHBD_WRAPLOW(step1[15] - step1[16], bd); + output[17] = HIGHBD_WRAPLOW(step1[14] - step1[17], bd); + output[18] = HIGHBD_WRAPLOW(step1[13] - step1[18], bd); + output[19] = HIGHBD_WRAPLOW(step1[12] - step1[19], bd); + output[20] = HIGHBD_WRAPLOW(step1[11] - step1[20], bd); + output[21] = HIGHBD_WRAPLOW(step1[10] - step1[21], bd); + output[22] = HIGHBD_WRAPLOW(step1[9] - step1[22], bd); + output[23] = HIGHBD_WRAPLOW(step1[8] - step1[23], bd); + output[24] = HIGHBD_WRAPLOW(step1[7] - step1[24], bd); + output[25] = HIGHBD_WRAPLOW(step1[6] - step1[25], bd); + output[26] = HIGHBD_WRAPLOW(step1[5] - step1[26], bd); + output[27] = HIGHBD_WRAPLOW(step1[4] - step1[27], bd); + output[28] = HIGHBD_WRAPLOW(step1[3] - step1[28], bd); + output[29] = HIGHBD_WRAPLOW(step1[2] - step1[29], bd); + output[30] = HIGHBD_WRAPLOW(step1[1] - step1[30], bd); + output[31] = HIGHBD_WRAPLOW(step1[0] - step1[31], bd); +} + #endif // CONFIG_HIGHBITDEPTH diff --git a/third_party/aom/aom_dsp/loopfilter.c b/third_party/aom/aom_dsp/loopfilter.c index e2e839219..7ea1e6b89 100644 --- a/third_party/aom/aom_dsp/loopfilter.c +++ b/third_party/aom/aom_dsp/loopfilter.c @@ -149,10 +149,15 @@ void aom_lpf_horizontal_4_c(uint8_t *s, int p /* pitch */, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh) { int i; +#if CONFIG_PARALLEL_DEBLOCKING && CONFIG_CB4X4 + int count = 4; +#else + int count = 8; +#endif // loop filter designed to work using chars so that we can make maximum use // of 8 bit simd instructions. - for (i = 0; i < 8; ++i) { + for (i = 0; i < count; ++i) { #if !CONFIG_PARALLEL_DEBLOCKING const uint8_t p3 = s[-4 * p], p2 = s[-3 * p], p1 = s[-2 * p], p0 = s[-p]; const uint8_t q0 = s[0 * p], q1 = s[1 * p], q2 = s[2 * p], q3 = s[3 * p]; @@ -179,10 +184,15 @@ void aom_lpf_horizontal_4_dual_c(uint8_t *s, int p, const uint8_t *blimit0, void aom_lpf_vertical_4_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh) { int i; +#if CONFIG_PARALLEL_DEBLOCKING && CONFIG_CB4X4 + int count = 4; +#else + int count = 8; +#endif // loop filter designed to work using chars so that we can make maximum use // of 8 bit simd instructions. - for (i = 0; i < 8; ++i) { + for (i = 0; i < count; ++i) { #if !CONFIG_PARALLEL_DEBLOCKING const uint8_t p3 = s[-4], p2 = s[-3], p1 = s[-2], p0 = s[-1]; const uint8_t q0 = s[0], q1 = s[1], q2 = s[2], q3 = s[3]; @@ -206,7 +216,7 @@ void aom_lpf_vertical_4_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, aom_lpf_vertical_4_c(s + 8 * pitch, pitch, blimit1, limit1, thresh1); } -static INLINE void filter8(int8_t mask, uint8_t thresh, uint8_t flat, +static INLINE void filter8(int8_t mask, uint8_t thresh, int8_t flat, uint8_t *op3, uint8_t *op2, uint8_t *op1, uint8_t *op0, uint8_t *oq0, uint8_t *oq1, uint8_t *oq2, uint8_t *oq3) { @@ -229,10 +239,15 @@ static INLINE void filter8(int8_t mask, uint8_t thresh, uint8_t flat, void aom_lpf_horizontal_8_c(uint8_t *s, int p, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh) { int i; +#if CONFIG_PARALLEL_DEBLOCKING && CONFIG_CB4X4 + int count = 4; +#else + int count = 8; +#endif // loop filter designed to work using chars so that we can make maximum use // of 8 bit simd instructions. - for (i = 0; i < 8; ++i) { + for (i = 0; i < count; ++i) { const uint8_t p3 = s[-4 * p], p2 = s[-3 * p], p1 = s[-2 * p], p0 = s[-p]; const uint8_t q0 = s[0 * p], q1 = s[1 * p], q2 = s[2 * p], q3 = s[3 * p]; @@ -256,8 +271,13 @@ void aom_lpf_horizontal_8_dual_c(uint8_t *s, int p, const uint8_t *blimit0, void aom_lpf_vertical_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh) { int i; +#if CONFIG_PARALLEL_DEBLOCKING && CONFIG_CB4X4 + int count = 4; +#else + int count = 8; +#endif - for (i = 0; i < 8; ++i) { + for (i = 0; i < count; ++i) { const uint8_t p3 = s[-4], p2 = s[-3], p1 = s[-2], p0 = s[-1]; const uint8_t q0 = s[0], q1 = s[1], q2 = s[2], q3 = s[3]; const int8_t mask = @@ -278,8 +298,8 @@ void aom_lpf_vertical_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, } #if PARALLEL_DEBLOCKING_11_TAP -static INLINE void filter12(int8_t mask, uint8_t thresh, uint8_t flat, - uint8_t flat2, uint8_t *op5, uint8_t *op4, +static INLINE void filter12(int8_t mask, uint8_t thresh, int8_t flat, + int8_t flat2, uint8_t *op5, uint8_t *op4, uint8_t *op3, uint8_t *op2, uint8_t *op1, uint8_t *op0, uint8_t *oq0, uint8_t *oq1, uint8_t *oq2, uint8_t *oq3, uint8_t *oq4, @@ -308,8 +328,8 @@ static INLINE void filter12(int8_t mask, uint8_t thresh, uint8_t flat, #endif #if PARALLEL_DEBLOCKING_9_TAP -static INLINE void filter10(int8_t mask, uint8_t thresh, uint8_t flat, - uint8_t flat2, uint8_t *op4, uint8_t *op3, +static INLINE void filter10(int8_t mask, uint8_t thresh, int8_t flat, + int8_t flat2, uint8_t *op4, uint8_t *op3, uint8_t *op2, uint8_t *op1, uint8_t *op0, uint8_t *oq0, uint8_t *oq1, uint8_t *oq2, uint8_t *oq3, uint8_t *oq4) { @@ -332,8 +352,8 @@ static INLINE void filter10(int8_t mask, uint8_t thresh, uint8_t flat, } #endif -static INLINE void filter16(int8_t mask, uint8_t thresh, uint8_t flat, - uint8_t flat2, uint8_t *op7, uint8_t *op6, +static INLINE void filter16(int8_t mask, uint8_t thresh, int8_t flat, + int8_t flat2, uint8_t *op7, uint8_t *op6, uint8_t *op5, uint8_t *op4, uint8_t *op3, uint8_t *op2, uint8_t *op1, uint8_t *op0, uint8_t *oq0, uint8_t *oq1, uint8_t *oq2, @@ -390,10 +410,15 @@ static void mb_lpf_horizontal_edge_w(uint8_t *s, int p, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count) { int i; +#if CONFIG_PARALLEL_DEBLOCKING && CONFIG_CB4X4 + int step = 4; +#else + int step = 8; +#endif // loop filter designed to work using chars so that we can make maximum use // of 8 bit simd instructions. - for (i = 0; i < 8 * count; ++i) { + for (i = 0; i < step * count; ++i) { const uint8_t p7 = s[-8 * p], p6 = s[-7 * p], p5 = s[-6 * p], p4 = s[-5 * p], p3 = s[-4 * p], p2 = s[-3 * p], p1 = s[-2 * p], p0 = s[-p]; @@ -436,7 +461,11 @@ void aom_lpf_horizontal_edge_8_c(uint8_t *s, int p, const uint8_t *blimit, void aom_lpf_horizontal_edge_16_c(uint8_t *s, int p, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh) { +#if CONFIG_PARALLEL_DEBLOCKING && CONFIG_CB4X4 + mb_lpf_horizontal_edge_w(s, p, blimit, limit, thresh, 1); +#else mb_lpf_horizontal_edge_w(s, p, blimit, limit, thresh, 2); +#endif } static void mb_lpf_vertical_edge_w(uint8_t *s, int p, const uint8_t *blimit, @@ -478,7 +507,11 @@ static void mb_lpf_vertical_edge_w(uint8_t *s, int p, const uint8_t *blimit, void aom_lpf_vertical_16_c(uint8_t *s, int p, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh) { +#if CONFIG_PARALLEL_DEBLOCKING && CONFIG_CB4X4 + mb_lpf_vertical_edge_w(s, p, blimit, limit, thresh, 4); +#else mb_lpf_vertical_edge_w(s, p, blimit, limit, thresh, 8); +#endif } void aom_lpf_vertical_16_dual_c(uint8_t *s, int p, const uint8_t *blimit, @@ -596,10 +629,15 @@ void aom_highbd_lpf_horizontal_4_c(uint16_t *s, int p /* pitch */, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd) { int i; +#if CONFIG_PARALLEL_DEBLOCKING && CONFIG_CB4X4 + int count = 4; +#else + int count = 8; +#endif // loop filter designed to work using chars so that we can make maximum use // of 8 bit simd instructions. - for (i = 0; i < 8; ++i) { + for (i = 0; i < count; ++i) { #if !CONFIG_PARALLEL_DEBLOCKING const uint16_t p3 = s[-4 * p]; const uint16_t p2 = s[-3 * p]; @@ -636,10 +674,15 @@ void aom_highbd_lpf_vertical_4_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd) { int i; +#if CONFIG_PARALLEL_DEBLOCKING && CONFIG_CB4X4 + int count = 4; +#else + int count = 8; +#endif // loop filter designed to work using chars so that we can make maximum use // of 8 bit simd instructions. - for (i = 0; i < 8; ++i) { + for (i = 0; i < count; ++i) { #if !CONFIG_PARALLEL_DEBLOCKING const uint16_t p3 = s[-4], p2 = s[-3], p1 = s[-2], p0 = s[-1]; const uint16_t q0 = s[0], q1 = s[1], q2 = s[2], q3 = s[3]; @@ -665,7 +708,7 @@ void aom_highbd_lpf_vertical_4_dual_c( bd); } -static INLINE void highbd_filter8(int8_t mask, uint8_t thresh, uint8_t flat, +static INLINE void highbd_filter8(int8_t mask, uint8_t thresh, int8_t flat, uint16_t *op3, uint16_t *op2, uint16_t *op1, uint16_t *op0, uint16_t *oq0, uint16_t *oq1, uint16_t *oq2, uint16_t *oq3, int bd) { @@ -689,10 +732,15 @@ void aom_highbd_lpf_horizontal_8_c(uint16_t *s, int p, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd) { int i; +#if CONFIG_PARALLEL_DEBLOCKING && CONFIG_CB4X4 + int count = 4; +#else + int count = 8; +#endif // loop filter designed to work using chars so that we can make maximum use // of 8 bit simd instructions. - for (i = 0; i < 8; ++i) { + for (i = 0; i < count; ++i) { const uint16_t p3 = s[-4 * p], p2 = s[-3 * p], p1 = s[-2 * p], p0 = s[-p]; const uint16_t q0 = s[0 * p], q1 = s[1 * p], q2 = s[2 * p], q3 = s[3 * p]; @@ -718,8 +766,13 @@ void aom_highbd_lpf_vertical_8_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd) { int i; +#if CONFIG_PARALLEL_DEBLOCKING && CONFIG_CB4X4 + int count = 4; +#else + int count = 8; +#endif - for (i = 0; i < 8; ++i) { + for (i = 0; i < count; ++i) { const uint16_t p3 = s[-4], p2 = s[-3], p1 = s[-2], p0 = s[-1]; const uint16_t q0 = s[0], q1 = s[1], q2 = s[2], q3 = s[3]; const int8_t mask = @@ -741,8 +794,8 @@ void aom_highbd_lpf_vertical_8_dual_c( bd); } -static INLINE void highbd_filter16(int8_t mask, uint8_t thresh, uint8_t flat, - uint8_t flat2, uint16_t *op7, uint16_t *op6, +static INLINE void highbd_filter16(int8_t mask, uint8_t thresh, int8_t flat, + int8_t flat2, uint16_t *op7, uint16_t *op6, uint16_t *op5, uint16_t *op4, uint16_t *op3, uint16_t *op2, uint16_t *op1, uint16_t *op0, uint16_t *oq0, uint16_t *oq1, uint16_t *oq2, @@ -813,10 +866,15 @@ static void highbd_mb_lpf_horizontal_edge_w(uint16_t *s, int p, const uint8_t *thresh, int count, int bd) { int i; +#if CONFIG_PARALLEL_DEBLOCKING && CONFIG_CB4X4 + int step = 4; +#else + int step = 8; +#endif // loop filter designed to work using chars so that we can make maximum use // of 8 bit simd instructions. - for (i = 0; i < 8 * count; ++i) { + for (i = 0; i < step * count; ++i) { const uint16_t p3 = s[-4 * p]; const uint16_t p2 = s[-3 * p]; const uint16_t p1 = s[-2 * p]; @@ -852,7 +910,11 @@ void aom_highbd_lpf_horizontal_edge_16_c(uint16_t *s, int p, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd) { +#if CONFIG_PARALLEL_DEBLOCKING && CONFIG_CB4X4 + highbd_mb_lpf_horizontal_edge_w(s, p, blimit, limit, thresh, 1, bd); +#else highbd_mb_lpf_horizontal_edge_w(s, p, blimit, limit, thresh, 2, bd); +#endif } static void highbd_mb_lpf_vertical_edge_w(uint16_t *s, int p, @@ -888,13 +950,21 @@ static void highbd_mb_lpf_vertical_edge_w(uint16_t *s, int p, void aom_highbd_lpf_vertical_16_c(uint16_t *s, int p, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd) { +#if CONFIG_PARALLEL_DEBLOCKING && CONFIG_CB4X4 + highbd_mb_lpf_vertical_edge_w(s, p, blimit, limit, thresh, 4, bd); +#else highbd_mb_lpf_vertical_edge_w(s, p, blimit, limit, thresh, 8, bd); +#endif } void aom_highbd_lpf_vertical_16_dual_c(uint16_t *s, int p, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd) { +#if CONFIG_PARALLEL_DEBLOCKING && CONFIG_CB4X4 + highbd_mb_lpf_vertical_edge_w(s, p, blimit, limit, thresh, 8, bd); +#else highbd_mb_lpf_vertical_edge_w(s, p, blimit, limit, thresh, 16, bd); +#endif } #endif // CONFIG_HIGHBITDEPTH diff --git a/third_party/aom/aom_dsp/mips/avg_msa.c b/third_party/aom/aom_dsp/mips/avg_msa.c deleted file mode 100644 index 0e1728155..000000000 --- a/third_party/aom/aom_dsp/mips/avg_msa.c +++ /dev/null @@ -1,57 +0,0 @@ -/* - * 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/mips/macros_msa.h" - -uint32_t aom_avg_8x8_msa(const uint8_t *src, int32_t src_stride) { - uint32_t sum_out; - v16u8 src0, src1, src2, src3, src4, src5, src6, src7; - v8u16 sum0, sum1, sum2, sum3, sum4, sum5, sum6, sum7; - v4u32 sum = { 0 }; - - LD_UB8(src, src_stride, src0, src1, src2, src3, src4, src5, src6, src7); - HADD_UB4_UH(src0, src1, src2, src3, sum0, sum1, sum2, sum3); - HADD_UB4_UH(src4, src5, src6, src7, sum4, sum5, sum6, sum7); - ADD4(sum0, sum1, sum2, sum3, sum4, sum5, sum6, sum7, sum0, sum2, sum4, sum6); - ADD2(sum0, sum2, sum4, sum6, sum0, sum4); - sum0 += sum4; - - sum = __msa_hadd_u_w(sum0, sum0); - sum0 = (v8u16)__msa_pckev_h((v8i16)sum, (v8i16)sum); - sum = __msa_hadd_u_w(sum0, sum0); - sum = (v4u32)__msa_srari_w((v4i32)sum, 6); - sum_out = __msa_copy_u_w((v4i32)sum, 0); - - return sum_out; -} - -uint32_t aom_avg_4x4_msa(const uint8_t *src, int32_t src_stride) { - uint32_t sum_out; - uint32_t src0, src1, src2, src3; - v16u8 vec = { 0 }; - v8u16 sum0; - v4u32 sum1; - v2u64 sum2; - - LW4(src, src_stride, src0, src1, src2, src3); - INSERT_W4_UB(src0, src1, src2, src3, vec); - - sum0 = __msa_hadd_u_h(vec, vec); - sum1 = __msa_hadd_u_w(sum0, sum0); - sum0 = (v8u16)__msa_pckev_h((v8i16)sum1, (v8i16)sum1); - sum1 = __msa_hadd_u_w(sum0, sum0); - sum2 = __msa_hadd_u_d(sum1, sum1); - sum1 = (v4u32)__msa_srari_w((v4i32)sum2, 4); - sum_out = __msa_copy_u_w((v4i32)sum1, 0); - - return sum_out; -} diff --git a/third_party/aom/aom_dsp/prob.c b/third_party/aom/aom_dsp/prob.c index c60bfdac5..eefe7521f 100644 --- a/third_party/aom/aom_dsp/prob.c +++ b/third_party/aom/aom_dsp/prob.c @@ -11,25 +11,10 @@ #include "./aom_config.h" -#if CONFIG_EC_MULTISYMBOL #include -#endif #include "aom_dsp/prob.h" -const uint8_t aom_norm[256] = { - 0, 7, 6, 6, 5, 5, 5, 5, 4, 4, 4, 4, 4, 4, 4, 4, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, - 3, 3, 3, 3, 3, 3, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, - 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, - 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, - 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, - 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, - 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, - 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, - 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, - 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 -}; - static unsigned int tree_merge_probs_impl(unsigned int i, const aom_tree_index *tree, const aom_prob *pre_probs, @@ -53,7 +38,6 @@ void aom_tree_merge_probs(const aom_tree_index *tree, const aom_prob *pre_probs, tree_merge_probs_impl(0, tree, pre_probs, counts, probs); } -#if CONFIG_EC_MULTISYMBOL typedef struct tree_node tree_node; struct tree_node { @@ -233,4 +217,3 @@ void av1_indices_from_tree(int *ind, int *inv, const aom_tree_index *tree) { int stack_index = 0; tree_to_index(&stack_index, ind, inv, tree, 0, 0); } -#endif diff --git a/third_party/aom/aom_dsp/prob.h b/third_party/aom/aom_dsp/prob.h index 808592923..ec6654ab7 100644 --- a/third_party/aom/aom_dsp/prob.h +++ b/third_party/aom/aom_dsp/prob.h @@ -20,7 +20,7 @@ #include "aom_ports/bitops.h" #include "aom_ports/mem.h" -#if CONFIG_DAALA_EC +#if !CONFIG_ANS #include "aom_dsp/entcode.h" #endif @@ -33,14 +33,12 @@ typedef uint8_t aom_prob; // TODO(negge): Rename this aom_prob once we remove vpxbool. typedef uint16_t aom_cdf_prob; -#if CONFIG_EC_MULTISYMBOL #define CDF_SIZE(x) ((x) + 1) -#endif #define CDF_PROB_BITS 15 #define CDF_PROB_TOP (1 << CDF_PROB_BITS) -#if CONFIG_DAALA_EC +#if !CONFIG_ANS #define AOM_ICDF OD_ICDF #else #define AOM_ICDF(x) (x) @@ -117,7 +115,6 @@ static INLINE aom_prob mode_mv_merge_probs(aom_prob pre_prob, void aom_tree_merge_probs(const aom_tree_index *tree, const aom_prob *pre_probs, const unsigned int *counts, aom_prob *probs); -#if CONFIG_EC_MULTISYMBOL int tree_to_cdf(const aom_tree_index *tree, const aom_prob *probs, aom_tree_index root, aom_cdf_prob *cdf, aom_tree_index *ind, int *pth, int *len); @@ -150,9 +147,6 @@ static INLINE void av1_tree_to_cdf(const aom_tree_index *tree, } while (0) void av1_indices_from_tree(int *ind, int *inv, const aom_tree_index *tree); -#endif - -DECLARE_ALIGNED(16, extern const uint8_t, aom_norm[256]); #if CONFIG_EC_ADAPT static INLINE void update_cdf(aom_cdf_prob *cdf, int val, int nsymbs) { @@ -165,7 +159,7 @@ static INLINE void update_cdf(aom_cdf_prob *cdf, int val, int nsymbs) { tmp = AOM_ICDF(tmp0); diff = ((CDF_PROB_TOP - (nsymbs << rate2)) >> rate) << rate; // Single loop (faster) -#if CONFIG_DAALA_EC && CONFIG_EC_SMALLMUL +#if !CONFIG_ANS && CONFIG_EC_SMALLMUL for (i = 0; i < nsymbs - 1; ++i, tmp -= tmp0) { tmp -= (i == val ? diff : 0); cdf[i] += ((tmp - cdf[i]) >> rate); diff --git a/third_party/aom/aom_dsp/sad.c b/third_party/aom/aom_dsp/sad.c index 3e1070519..2cc172ba5 100644 --- a/third_party/aom/aom_dsp/sad.c +++ b/third_party/aom/aom_dsp/sad.c @@ -16,6 +16,7 @@ #include "aom/aom_integer.h" #include "aom_ports/mem.h" +#include "aom_dsp/blend.h" /* Sum the difference between every corresponding element of the buffers. */ static INLINE unsigned int sad(const uint8_t *a, int a_stride, const uint8_t *b, @@ -311,15 +312,20 @@ highbd_sadMxNx4D(4, 4) #if CONFIG_AV1 && CONFIG_EXT_INTER static INLINE - unsigned int masked_sad(const uint8_t *a, int a_stride, const uint8_t *b, + unsigned int masked_sad(const uint8_t *src, int src_stride, + const uint8_t *a, int a_stride, const uint8_t *b, int b_stride, const uint8_t *m, int m_stride, int width, int height) { int y, x; unsigned int sad = 0; for (y = 0; y < height; y++) { - for (x = 0; x < width; x++) sad += m[x] * abs(a[x] - b[x]); + for (x = 0; x < width; x++) { + const uint8_t pred = AOM_BLEND_A64(m[x], a[x], b[x]); + sad += abs(pred - src[x]); + } + src += src_stride; a += a_stride; b += b_stride; m += m_stride; @@ -329,12 +335,17 @@ highbd_sadMxNx4D(4, 4) return sad; } -#define MASKSADMxN(m, n) \ - unsigned int aom_masked_sad##m##x##n##_c( \ - const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, \ - const uint8_t *msk, int msk_stride) { \ - return masked_sad(src, src_stride, ref, ref_stride, msk, msk_stride, m, \ - n); \ +#define MASKSADMxN(m, n) \ + unsigned int aom_masked_sad##m##x##n##_c( \ + const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, \ + const uint8_t *second_pred, const uint8_t *msk, int msk_stride, \ + int invert_mask) { \ + if (!invert_mask) \ + return masked_sad(src, src_stride, ref, ref_stride, second_pred, m, msk, \ + msk_stride, m, n); \ + else \ + return masked_sad(src, src_stride, second_pred, m, ref, ref_stride, msk, \ + msk_stride, m, n); \ } /* clang-format off */ @@ -360,18 +371,24 @@ MASKSADMxN(4, 4) #if CONFIG_HIGHBITDEPTH static INLINE - unsigned int highbd_masked_sad(const uint8_t *a8, int a_stride, + unsigned int highbd_masked_sad(const uint8_t *src8, int src_stride, + const uint8_t *a8, int a_stride, const uint8_t *b8, int b_stride, const uint8_t *m, int m_stride, int width, int height) { int y, x; unsigned int sad = 0; + const uint16_t *src = CONVERT_TO_SHORTPTR(src8); const uint16_t *a = CONVERT_TO_SHORTPTR(a8); const uint16_t *b = CONVERT_TO_SHORTPTR(b8); for (y = 0; y < height; y++) { - for (x = 0; x < width; x++) sad += m[x] * abs(a[x] - b[x]); + for (x = 0; x < width; x++) { + const uint16_t pred = AOM_BLEND_A64(m[x], a[x], b[x]); + sad += abs(pred - src[x]); + } + src += src_stride; a += a_stride; b += b_stride; m += m_stride; @@ -381,12 +398,17 @@ MASKSADMxN(4, 4) return sad; } -#define HIGHBD_MASKSADMXN(m, n) \ - unsigned int aom_highbd_masked_sad##m##x##n##_c( \ - 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(src, src_stride, ref, ref_stride, msk, \ - msk_stride, m, n); \ +#define HIGHBD_MASKSADMXN(m, n) \ + unsigned int aom_highbd_masked_sad##m##x##n##_c( \ + const uint8_t *src8, int src_stride, const uint8_t *ref8, \ + int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, \ + int msk_stride, int invert_mask) { \ + if (!invert_mask) \ + return highbd_masked_sad(src8, src_stride, ref8, ref_stride, \ + second_pred8, m, msk, msk_stride, m, n); \ + else \ + return highbd_masked_sad(src8, src_stride, second_pred8, m, ref8, \ + ref_stride, msk, msk_stride, m, n); \ } #if CONFIG_EXT_PARTITION diff --git a/third_party/aom/aom_dsp/simd/v64_intrinsics.h b/third_party/aom/aom_dsp/simd/v64_intrinsics.h index ee2b683a4..5c0042d8c 100644 --- a/third_party/aom/aom_dsp/simd/v64_intrinsics.h +++ b/third_party/aom/aom_dsp/simd/v64_intrinsics.h @@ -60,7 +60,9 @@ SIMD_INLINE void v64_store_aligned(void *p, v64 a) { c_v64_store_aligned(p, a); } -SIMD_INLINE v64 v64_align(v64 a, v64 b, c) { return c_v64_align(a, b, c); } +SIMD_INLINE v64 v64_align(v64 a, v64 b, unsigned int c) { + return c_v64_align(a, b, c); +} SIMD_INLINE v64 v64_zero() { return c_v64_zero(); } SIMD_INLINE v64 v64_dup_8(uint8_t x) { return c_v64_dup_8(x); } diff --git a/third_party/aom/aom_dsp/variance.c b/third_party/aom/aom_dsp/variance.c index 9fc0db783..79677c92f 100644 --- a/third_party/aom/aom_dsp/variance.c +++ b/third_party/aom/aom_dsp/variance.c @@ -18,6 +18,7 @@ #include "aom_dsp/variance.h" #include "aom_dsp/aom_filter.h" +#include "aom_dsp/blend.h" uint32_t aom_get4x4sse_cs_c(const uint8_t *a, int a_stride, const uint8_t *b, int b_stride) { @@ -672,297 +673,215 @@ void aom_highbd_comp_avg_upsampled_pred_c(uint16_t *comp_pred, #endif // CONFIG_HIGHBITDEPTH #if CONFIG_AV1 && CONFIG_EXT_INTER -void masked_variance(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 *sum) { +void aom_comp_mask_pred_c(uint8_t *comp_pred, const uint8_t *pred, int width, + int height, const uint8_t *ref, int ref_stride, + const uint8_t *mask, int mask_stride, + int invert_mask) { int i, j; - int64_t sum64 = 0; - uint64_t sse64 = 0; - - for (i = 0; i < h; i++) { - for (j = 0; j < w; j++) { - const int diff = (a[j] - b[j]) * (m[j]); - sum64 += diff; - sse64 += diff * diff; + for (i = 0; i < height; ++i) { + for (j = 0; j < width; ++j) { + if (!invert_mask) + comp_pred[j] = AOM_BLEND_A64(mask[j], ref[j], pred[j]); + else + comp_pred[j] = AOM_BLEND_A64(mask[j], pred[j], ref[j]); } - - a += a_stride; - b += b_stride; - m += m_stride; + comp_pred += width; + pred += width; + ref += ref_stride; + mask += mask_stride; } - sum64 = (sum64 >= 0) ? sum64 : -sum64; - *sum = (int)ROUND_POWER_OF_TWO(sum64, 6); - *sse = (uint32_t)ROUND_POWER_OF_TWO(sse64, 12); } -#define MASK_VAR(W, H) \ - unsigned int aom_masked_variance##W##x##H##_c( \ - const uint8_t *a, int a_stride, const uint8_t *b, int b_stride, \ - const uint8_t *m, int m_stride, unsigned int *sse) { \ - int sum; \ - masked_variance(a, a_stride, b, b_stride, m, m_stride, W, H, sse, &sum); \ - return *sse - (unsigned int)(((int64_t)sum * sum) / (W * H)); \ +void aom_comp_mask_upsampled_pred_c(uint8_t *comp_pred, const uint8_t *pred, + int width, int height, const uint8_t *ref, + int ref_stride, const uint8_t *mask, + int mask_stride, int invert_mask) { + int i, j; + int stride = ref_stride << 3; + + for (i = 0; i < height; i++) { + for (j = 0; j < width; j++) { + if (!invert_mask) + comp_pred[j] = AOM_BLEND_A64(mask[j], ref[(j << 3)], pred[j]); + else + comp_pred[j] = AOM_BLEND_A64(mask[j], pred[j], ref[(j << 3)]); + } + comp_pred += width; + pred += width; + ref += stride; + mask += mask_stride; } +} #define MASK_SUBPIX_VAR(W, H) \ unsigned int aom_masked_sub_pixel_variance##W##x##H##_c( \ 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, \ + const uint8_t *ref, int ref_stride, const uint8_t *second_pred, \ + const uint8_t *msk, int msk_stride, int invert_mask, \ unsigned int *sse) { \ uint16_t fdata3[(H + 1) * W]; \ uint8_t temp2[H * W]; \ + DECLARE_ALIGNED(16, uint8_t, temp3[H * W]); \ \ var_filter_block2d_bil_first_pass(src, fdata3, src_stride, 1, H + 1, W, \ bilinear_filters_2t[xoffset]); \ var_filter_block2d_bil_second_pass(fdata3, temp2, W, W, H, W, \ bilinear_filters_2t[yoffset]); \ \ - return aom_masked_variance##W##x##H##_c(temp2, W, dst, dst_stride, msk, \ - msk_stride, sse); \ + aom_comp_mask_pred_c(temp3, second_pred, W, H, temp2, W, msk, msk_stride, \ + invert_mask); \ + return aom_variance##W##x##H##_c(temp3, W, ref, ref_stride, sse); \ } -MASK_VAR(4, 4) MASK_SUBPIX_VAR(4, 4) - -MASK_VAR(4, 8) MASK_SUBPIX_VAR(4, 8) - -MASK_VAR(8, 4) MASK_SUBPIX_VAR(8, 4) - -MASK_VAR(8, 8) MASK_SUBPIX_VAR(8, 8) - -MASK_VAR(8, 16) MASK_SUBPIX_VAR(8, 16) - -MASK_VAR(16, 8) MASK_SUBPIX_VAR(16, 8) - -MASK_VAR(16, 16) MASK_SUBPIX_VAR(16, 16) - -MASK_VAR(16, 32) MASK_SUBPIX_VAR(16, 32) - -MASK_VAR(32, 16) MASK_SUBPIX_VAR(32, 16) - -MASK_VAR(32, 32) MASK_SUBPIX_VAR(32, 32) - -MASK_VAR(32, 64) MASK_SUBPIX_VAR(32, 64) - -MASK_VAR(64, 32) MASK_SUBPIX_VAR(64, 32) - -MASK_VAR(64, 64) MASK_SUBPIX_VAR(64, 64) - #if CONFIG_EXT_PARTITION -MASK_VAR(64, 128) MASK_SUBPIX_VAR(64, 128) - -MASK_VAR(128, 64) MASK_SUBPIX_VAR(128, 64) - -MASK_VAR(128, 128) MASK_SUBPIX_VAR(128, 128) #endif // CONFIG_EXT_PARTITION #if CONFIG_HIGHBITDEPTH -void highbd_masked_variance64(const uint8_t *a8, int a_stride, - const uint8_t *b8, int b_stride, const uint8_t *m, - int m_stride, int w, int h, uint64_t *sse, - int64_t *sum) { +void aom_highbd_comp_mask_pred_c(uint16_t *comp_pred, const uint8_t *pred8, + int width, int height, const uint8_t *ref8, + int ref_stride, const uint8_t *mask, + int mask_stride, int invert_mask) { int i, j; - uint16_t *a = CONVERT_TO_SHORTPTR(a8); - uint16_t *b = CONVERT_TO_SHORTPTR(b8); - - *sum = 0; - *sse = 0; - - for (i = 0; i < h; i++) { - for (j = 0; j < w; j++) { - const int diff = (a[j] - b[j]) * (m[j]); - *sum += (int64_t)diff; - *sse += (int64_t)diff * diff; + uint16_t *pred = CONVERT_TO_SHORTPTR(pred8); + uint16_t *ref = CONVERT_TO_SHORTPTR(ref8); + for (i = 0; i < height; ++i) { + for (j = 0; j < width; ++j) { + if (!invert_mask) + comp_pred[j] = AOM_BLEND_A64(mask[j], ref[j], pred[j]); + else + comp_pred[j] = AOM_BLEND_A64(mask[j], pred[j], ref[j]); } - - a += a_stride; - b += b_stride; - m += m_stride; + comp_pred += width; + pred += width; + ref += ref_stride; + mask += mask_stride; } - *sum = (*sum >= 0) ? *sum : -*sum; - *sum = ROUND_POWER_OF_TWO(*sum, 6); - *sse = ROUND_POWER_OF_TWO(*sse, 12); } -void highbd_masked_variance(const uint8_t *a8, int a_stride, const uint8_t *b8, - int b_stride, const uint8_t *m, int m_stride, int w, - int h, unsigned int *sse, int *sum) { - int64_t sum64; - uint64_t sse64; - highbd_masked_variance64(a8, a_stride, b8, b_stride, m, m_stride, w, h, - &sse64, &sum64); - *sum = (int)sum64; - *sse = (unsigned int)sse64; -} - -void highbd_10_masked_variance(const uint8_t *a8, int a_stride, - const uint8_t *b8, int b_stride, - const uint8_t *m, int m_stride, int w, int h, - unsigned int *sse, int *sum) { - int64_t sum64; - uint64_t sse64; - highbd_masked_variance64(a8, a_stride, b8, b_stride, m, m_stride, w, h, - &sse64, &sum64); - *sum = (int)ROUND_POWER_OF_TWO(sum64, 2); - *sse = (unsigned int)ROUND_POWER_OF_TWO(sse64, 4); -} - -void highbd_12_masked_variance(const uint8_t *a8, int a_stride, - const uint8_t *b8, int b_stride, - const uint8_t *m, int m_stride, int w, int h, - unsigned int *sse, int *sum) { - int64_t sum64; - uint64_t sse64; - highbd_masked_variance64(a8, a_stride, b8, b_stride, m, m_stride, w, h, - &sse64, &sum64); - *sum = (int)ROUND_POWER_OF_TWO(sum64, 4); - *sse = (unsigned int)ROUND_POWER_OF_TWO(sse64, 8); -} +void aom_highbd_comp_mask_upsampled_pred_c(uint16_t *comp_pred, + const uint8_t *pred8, int width, + int height, const uint8_t *ref8, + int ref_stride, const uint8_t *mask, + int mask_stride, int invert_mask) { + int i, j; + int stride = ref_stride << 3; -#define HIGHBD_MASK_VAR(W, H) \ - unsigned int aom_highbd_masked_variance##W##x##H##_c( \ - const uint8_t *a, int a_stride, const uint8_t *b, int b_stride, \ - const uint8_t *m, int m_stride, unsigned int *sse) { \ - int sum; \ - highbd_masked_variance(a, a_stride, b, b_stride, m, m_stride, W, H, sse, \ - &sum); \ - return *sse - (unsigned int)(((int64_t)sum * sum) / (W * H)); \ - } \ - \ - unsigned int aom_highbd_10_masked_variance##W##x##H##_c( \ - const uint8_t *a, int a_stride, const uint8_t *b, int b_stride, \ - const uint8_t *m, int m_stride, unsigned int *sse) { \ - int sum; \ - int64_t var; \ - highbd_10_masked_variance(a, a_stride, b, b_stride, m, m_stride, 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_masked_variance##W##x##H##_c( \ - const uint8_t *a, int a_stride, const uint8_t *b, int b_stride, \ - const uint8_t *m, int m_stride, unsigned int *sse) { \ - int sum; \ - int64_t var; \ - highbd_12_masked_variance(a, a_stride, b, b_stride, m, m_stride, W, H, \ - sse, &sum); \ - var = (int64_t)(*sse) - (((int64_t)sum * sum) / (W * H)); \ - return (var >= 0) ? (uint32_t)var : 0; \ + uint16_t *pred = CONVERT_TO_SHORTPTR(pred8); + uint16_t *ref = CONVERT_TO_SHORTPTR(ref8); + for (i = 0; i < height; ++i) { + for (j = 0; j < width; ++j) { + if (!invert_mask) + comp_pred[j] = AOM_BLEND_A64(mask[j], ref[j << 3], pred[j]); + else + comp_pred[j] = AOM_BLEND_A64(mask[j], pred[j], ref[j << 3]); + } + comp_pred += width; + pred += width; + ref += stride; + mask += mask_stride; } +} -#define HIGHBD_MASK_SUBPIX_VAR(W, H) \ - unsigned int aom_highbd_masked_sub_pixel_variance##W##x##H##_c( \ - 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) { \ - uint16_t fdata3[(H + 1) * W]; \ - uint16_t temp2[H * W]; \ - \ - aom_highbd_var_filter_block2d_bil_first_pass( \ - src, fdata3, src_stride, 1, H + 1, W, bilinear_filters_2t[xoffset]); \ - aom_highbd_var_filter_block2d_bil_second_pass( \ - fdata3, temp2, W, W, H, W, bilinear_filters_2t[yoffset]); \ - \ - return aom_highbd_masked_variance##W##x##H##_c( \ - CONVERT_TO_BYTEPTR(temp2), W, dst, dst_stride, msk, msk_stride, sse); \ - } \ - \ - unsigned int aom_highbd_10_masked_sub_pixel_variance##W##x##H##_c( \ - 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) { \ - uint16_t fdata3[(H + 1) * W]; \ - uint16_t temp2[H * W]; \ - \ - aom_highbd_var_filter_block2d_bil_first_pass( \ - src, fdata3, src_stride, 1, H + 1, W, bilinear_filters_2t[xoffset]); \ - aom_highbd_var_filter_block2d_bil_second_pass( \ - fdata3, temp2, W, W, H, W, bilinear_filters_2t[yoffset]); \ - \ - return aom_highbd_10_masked_variance##W##x##H##_c( \ - CONVERT_TO_BYTEPTR(temp2), W, dst, dst_stride, msk, msk_stride, sse); \ - } \ - \ - unsigned int aom_highbd_12_masked_sub_pixel_variance##W##x##H##_c( \ - 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) { \ - uint16_t fdata3[(H + 1) * W]; \ - uint16_t temp2[H * W]; \ - \ - aom_highbd_var_filter_block2d_bil_first_pass( \ - src, fdata3, src_stride, 1, H + 1, W, bilinear_filters_2t[xoffset]); \ - aom_highbd_var_filter_block2d_bil_second_pass( \ - fdata3, temp2, W, W, H, W, bilinear_filters_2t[yoffset]); \ - \ - return aom_highbd_12_masked_variance##W##x##H##_c( \ - CONVERT_TO_BYTEPTR(temp2), W, dst, dst_stride, msk, msk_stride, sse); \ +#define HIGHBD_MASK_SUBPIX_VAR(W, H) \ + unsigned int aom_highbd_8_masked_sub_pixel_variance##W##x##H##_c( \ + const uint8_t *src, int src_stride, int xoffset, int yoffset, \ + const uint8_t *ref, int ref_stride, const uint8_t *second_pred, \ + const uint8_t *msk, int msk_stride, int invert_mask, \ + unsigned int *sse) { \ + uint16_t fdata3[(H + 1) * W]; \ + uint16_t temp2[H * W]; \ + DECLARE_ALIGNED(16, uint16_t, temp3[H * W]); \ + \ + aom_highbd_var_filter_block2d_bil_first_pass( \ + src, fdata3, src_stride, 1, H + 1, W, bilinear_filters_2t[xoffset]); \ + aom_highbd_var_filter_block2d_bil_second_pass( \ + fdata3, temp2, W, W, H, W, bilinear_filters_2t[yoffset]); \ + \ + aom_highbd_comp_mask_pred_c(temp3, second_pred, W, H, \ + CONVERT_TO_BYTEPTR(temp2), W, msk, msk_stride, \ + invert_mask); \ + \ + return aom_highbd_8_variance##W##x##H##_c(CONVERT_TO_BYTEPTR(temp3), W, \ + ref, ref_stride, sse); \ + } \ + \ + unsigned int aom_highbd_10_masked_sub_pixel_variance##W##x##H##_c( \ + const uint8_t *src, int src_stride, int xoffset, int yoffset, \ + const uint8_t *ref, int ref_stride, const uint8_t *second_pred, \ + const uint8_t *msk, int msk_stride, int invert_mask, \ + unsigned int *sse) { \ + uint16_t fdata3[(H + 1) * W]; \ + uint16_t temp2[H * W]; \ + DECLARE_ALIGNED(16, uint16_t, temp3[H * W]); \ + \ + aom_highbd_var_filter_block2d_bil_first_pass( \ + src, fdata3, src_stride, 1, H + 1, W, bilinear_filters_2t[xoffset]); \ + aom_highbd_var_filter_block2d_bil_second_pass( \ + fdata3, temp2, W, W, H, W, bilinear_filters_2t[yoffset]); \ + \ + aom_highbd_comp_mask_pred_c(temp3, second_pred, W, H, \ + CONVERT_TO_BYTEPTR(temp2), W, msk, msk_stride, \ + invert_mask); \ + \ + return aom_highbd_10_variance##W##x##H##_c(CONVERT_TO_BYTEPTR(temp3), W, \ + ref, ref_stride, sse); \ + } \ + \ + unsigned int aom_highbd_12_masked_sub_pixel_variance##W##x##H##_c( \ + const uint8_t *src, int src_stride, int xoffset, int yoffset, \ + const uint8_t *ref, int ref_stride, const uint8_t *second_pred, \ + const uint8_t *msk, int msk_stride, int invert_mask, \ + unsigned int *sse) { \ + uint16_t fdata3[(H + 1) * W]; \ + uint16_t temp2[H * W]; \ + DECLARE_ALIGNED(16, uint16_t, temp3[H * W]); \ + \ + aom_highbd_var_filter_block2d_bil_first_pass( \ + src, fdata3, src_stride, 1, H + 1, W, bilinear_filters_2t[xoffset]); \ + aom_highbd_var_filter_block2d_bil_second_pass( \ + fdata3, temp2, W, W, H, W, bilinear_filters_2t[yoffset]); \ + \ + aom_highbd_comp_mask_pred_c(temp3, second_pred, W, H, \ + CONVERT_TO_BYTEPTR(temp2), W, msk, msk_stride, \ + invert_mask); \ + \ + return aom_highbd_12_variance##W##x##H##_c(CONVERT_TO_BYTEPTR(temp3), W, \ + ref, ref_stride, sse); \ } -HIGHBD_MASK_VAR(4, 4) HIGHBD_MASK_SUBPIX_VAR(4, 4) - -HIGHBD_MASK_VAR(4, 8) HIGHBD_MASK_SUBPIX_VAR(4, 8) - -HIGHBD_MASK_VAR(8, 4) HIGHBD_MASK_SUBPIX_VAR(8, 4) - -HIGHBD_MASK_VAR(8, 8) HIGHBD_MASK_SUBPIX_VAR(8, 8) - -HIGHBD_MASK_VAR(8, 16) HIGHBD_MASK_SUBPIX_VAR(8, 16) - -HIGHBD_MASK_VAR(16, 8) HIGHBD_MASK_SUBPIX_VAR(16, 8) - -HIGHBD_MASK_VAR(16, 16) HIGHBD_MASK_SUBPIX_VAR(16, 16) - -HIGHBD_MASK_VAR(16, 32) HIGHBD_MASK_SUBPIX_VAR(16, 32) - -HIGHBD_MASK_VAR(32, 16) HIGHBD_MASK_SUBPIX_VAR(32, 16) - -HIGHBD_MASK_VAR(32, 32) HIGHBD_MASK_SUBPIX_VAR(32, 32) - -HIGHBD_MASK_VAR(32, 64) HIGHBD_MASK_SUBPIX_VAR(32, 64) - -HIGHBD_MASK_VAR(64, 32) HIGHBD_MASK_SUBPIX_VAR(64, 32) - -HIGHBD_MASK_VAR(64, 64) HIGHBD_MASK_SUBPIX_VAR(64, 64) - #if CONFIG_EXT_PARTITION -HIGHBD_MASK_VAR(64, 128) HIGHBD_MASK_SUBPIX_VAR(64, 128) - -HIGHBD_MASK_VAR(128, 64) HIGHBD_MASK_SUBPIX_VAR(128, 64) - -HIGHBD_MASK_VAR(128, 128) HIGHBD_MASK_SUBPIX_VAR(128, 128) #endif // CONFIG_EXT_PARTITION #endif // CONFIG_HIGHBITDEPTH diff --git a/third_party/aom/aom_dsp/variance.h b/third_party/aom/aom_dsp/variance.h index 7c925cfac..20f0895cb 100644 --- a/third_party/aom/aom_dsp/variance.h +++ b/third_party/aom/aom_dsp/variance.h @@ -57,15 +57,13 @@ typedef unsigned int (*aom_subp_avg_variance_fn_t)( #if CONFIG_AV1 && CONFIG_EXT_INTER typedef unsigned int (*aom_masked_sad_fn_t)(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, - const uint8_t *msk_ptr, - int msk_stride); -typedef unsigned int (*aom_masked_variance_fn_t)( - const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, - const uint8_t *msk, int msk_stride, unsigned int *sse); + const uint8_t *second_pred, + const uint8_t *msk, int msk_stride, + int invert_mask); typedef unsigned int (*aom_masked_subpixvariance_fn_t)( const uint8_t *src, int src_stride, int xoffset, int yoffset, - const uint8_t *ref, int ref_stride, const uint8_t *msk, int msk_stride, - unsigned int *sse); + const uint8_t *ref, int ref_stride, const uint8_t *second_pred, + const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse); #endif // CONFIG_AV1 && CONFIG_EXT_INTER #if CONFIG_AV1 && CONFIG_MOTION_VAR @@ -94,7 +92,6 @@ typedef struct aom_variance_vtable { aom_sad_multi_d_fn_t sdx4df; #if CONFIG_EXT_INTER aom_masked_sad_fn_t msdf; - aom_masked_variance_fn_t mvf; aom_masked_subpixvariance_fn_t msvf; #endif // CONFIG_EXT_INTER #if CONFIG_MOTION_VAR diff --git a/third_party/aom/aom_dsp/x86/aom_convolve_hip_sse2.c b/third_party/aom/aom_dsp/x86/aom_convolve_hip_sse2.c new file mode 100644 index 000000000..14352895d --- /dev/null +++ b/third_party/aom/aom_dsp/x86/aom_convolve_hip_sse2.c @@ -0,0 +1,195 @@ +/* + * 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 +#include + +#include "./aom_dsp_rtcd.h" +#include "aom_dsp/aom_convolve.h" +#include "aom_dsp/aom_dsp_common.h" +#include "aom_dsp/aom_filter.h" + +void aom_convolve8_add_src_hip_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) { + const int bd = 8; + assert(x_step_q4 == 16 && y_step_q4 == 16); + assert(!(w & 7)); + (void)x_step_q4; + (void)y_step_q4; + + uint16_t temp[(MAX_SB_SIZE + SUBPEL_TAPS - 1) * MAX_SB_SIZE]; + int intermediate_height = h + SUBPEL_TAPS - 1; + int i, j; + const int center_tap = ((SUBPEL_TAPS - 1) / 2); + const uint8_t *const src_ptr = src - center_tap * src_stride - center_tap; + + const __m128i zero = _mm_setzero_si128(); + // Add an offset to account for the "add_src" part of the convolve function. + const __m128i offset = _mm_insert_epi16(zero, 1 << FILTER_BITS, 3); + + /* Horizontal filter */ + { + const __m128i coeffs_x = + _mm_add_epi16(_mm_loadu_si128((__m128i *)filter_x), offset); + + // coeffs 0 1 0 1 2 3 2 3 + const __m128i tmp_0 = _mm_unpacklo_epi32(coeffs_x, coeffs_x); + // coeffs 4 5 4 5 6 7 6 7 + const __m128i tmp_1 = _mm_unpackhi_epi32(coeffs_x, coeffs_x); + + // coeffs 0 1 0 1 0 1 0 1 + const __m128i coeff_01 = _mm_unpacklo_epi64(tmp_0, tmp_0); + // coeffs 2 3 2 3 2 3 2 3 + const __m128i coeff_23 = _mm_unpackhi_epi64(tmp_0, tmp_0); + // coeffs 4 5 4 5 4 5 4 5 + const __m128i coeff_45 = _mm_unpacklo_epi64(tmp_1, tmp_1); + // coeffs 6 7 6 7 6 7 6 7 + const __m128i coeff_67 = _mm_unpackhi_epi64(tmp_1, tmp_1); + + const __m128i round_const = + _mm_set1_epi32((1 << (FILTER_BITS - EXTRAPREC_BITS - 1)) + + (1 << (bd + FILTER_BITS - 1))); + + for (i = 0; i < intermediate_height; ++i) { + for (j = 0; j < w; j += 8) { + const __m128i data = + _mm_loadu_si128((__m128i *)&src_ptr[i * src_stride + j]); + + // Filter even-index pixels + const __m128i src_0 = _mm_unpacklo_epi8(data, zero); + const __m128i res_0 = _mm_madd_epi16(src_0, coeff_01); + const __m128i src_2 = _mm_unpacklo_epi8(_mm_srli_si128(data, 2), zero); + const __m128i res_2 = _mm_madd_epi16(src_2, coeff_23); + const __m128i src_4 = _mm_unpacklo_epi8(_mm_srli_si128(data, 4), zero); + const __m128i res_4 = _mm_madd_epi16(src_4, coeff_45); + const __m128i src_6 = _mm_unpacklo_epi8(_mm_srli_si128(data, 6), zero); + const __m128i res_6 = _mm_madd_epi16(src_6, coeff_67); + + __m128i res_even = _mm_add_epi32(_mm_add_epi32(res_0, res_4), + _mm_add_epi32(res_2, res_6)); + res_even = _mm_srai_epi32(_mm_add_epi32(res_even, round_const), + FILTER_BITS - EXTRAPREC_BITS); + + // Filter odd-index pixels + const __m128i src_1 = _mm_unpacklo_epi8(_mm_srli_si128(data, 1), zero); + const __m128i res_1 = _mm_madd_epi16(src_1, coeff_01); + const __m128i src_3 = _mm_unpacklo_epi8(_mm_srli_si128(data, 3), zero); + const __m128i res_3 = _mm_madd_epi16(src_3, coeff_23); + const __m128i src_5 = _mm_unpacklo_epi8(_mm_srli_si128(data, 5), zero); + const __m128i res_5 = _mm_madd_epi16(src_5, coeff_45); + const __m128i src_7 = _mm_unpacklo_epi8(_mm_srli_si128(data, 7), zero); + const __m128i res_7 = _mm_madd_epi16(src_7, coeff_67); + + __m128i res_odd = _mm_add_epi32(_mm_add_epi32(res_1, res_5), + _mm_add_epi32(res_3, res_7)); + res_odd = _mm_srai_epi32(_mm_add_epi32(res_odd, round_const), + FILTER_BITS - EXTRAPREC_BITS); + + // Pack in the column order 0, 2, 4, 6, 1, 3, 5, 7 + __m128i res = _mm_packs_epi32(res_even, res_odd); + res = _mm_min_epi16(_mm_max_epi16(res, zero), + _mm_set1_epi16(EXTRAPREC_CLAMP_LIMIT(bd) - 1)); + _mm_storeu_si128((__m128i *)&temp[i * MAX_SB_SIZE + j], res); + } + } + } + + /* Vertical filter */ + { + const __m128i coeffs_y = + _mm_add_epi16(_mm_loadu_si128((__m128i *)filter_y), offset); + + // coeffs 0 1 0 1 2 3 2 3 + const __m128i tmp_0 = _mm_unpacklo_epi32(coeffs_y, coeffs_y); + // coeffs 4 5 4 5 6 7 6 7 + const __m128i tmp_1 = _mm_unpackhi_epi32(coeffs_y, coeffs_y); + + // coeffs 0 1 0 1 0 1 0 1 + const __m128i coeff_01 = _mm_unpacklo_epi64(tmp_0, tmp_0); + // coeffs 2 3 2 3 2 3 2 3 + const __m128i coeff_23 = _mm_unpackhi_epi64(tmp_0, tmp_0); + // coeffs 4 5 4 5 4 5 4 5 + const __m128i coeff_45 = _mm_unpacklo_epi64(tmp_1, tmp_1); + // coeffs 6 7 6 7 6 7 6 7 + const __m128i coeff_67 = _mm_unpackhi_epi64(tmp_1, tmp_1); + + const __m128i round_const = + _mm_set1_epi32((1 << (FILTER_BITS + EXTRAPREC_BITS - 1)) - + (1 << (bd + FILTER_BITS + EXTRAPREC_BITS - 1))); + + for (i = 0; i < h; ++i) { + for (j = 0; j < w; j += 8) { + // Filter even-index pixels + const uint16_t *data = &temp[i * MAX_SB_SIZE + j]; + const __m128i src_0 = + _mm_unpacklo_epi16(*(__m128i *)(data + 0 * MAX_SB_SIZE), + *(__m128i *)(data + 1 * MAX_SB_SIZE)); + const __m128i src_2 = + _mm_unpacklo_epi16(*(__m128i *)(data + 2 * MAX_SB_SIZE), + *(__m128i *)(data + 3 * MAX_SB_SIZE)); + const __m128i src_4 = + _mm_unpacklo_epi16(*(__m128i *)(data + 4 * MAX_SB_SIZE), + *(__m128i *)(data + 5 * MAX_SB_SIZE)); + const __m128i src_6 = + _mm_unpacklo_epi16(*(__m128i *)(data + 6 * MAX_SB_SIZE), + *(__m128i *)(data + 7 * MAX_SB_SIZE)); + + const __m128i res_0 = _mm_madd_epi16(src_0, coeff_01); + const __m128i res_2 = _mm_madd_epi16(src_2, coeff_23); + const __m128i res_4 = _mm_madd_epi16(src_4, coeff_45); + const __m128i res_6 = _mm_madd_epi16(src_6, coeff_67); + + const __m128i res_even = _mm_add_epi32(_mm_add_epi32(res_0, res_2), + _mm_add_epi32(res_4, res_6)); + + // Filter odd-index pixels + const __m128i src_1 = + _mm_unpackhi_epi16(*(__m128i *)(data + 0 * MAX_SB_SIZE), + *(__m128i *)(data + 1 * MAX_SB_SIZE)); + const __m128i src_3 = + _mm_unpackhi_epi16(*(__m128i *)(data + 2 * MAX_SB_SIZE), + *(__m128i *)(data + 3 * MAX_SB_SIZE)); + const __m128i src_5 = + _mm_unpackhi_epi16(*(__m128i *)(data + 4 * MAX_SB_SIZE), + *(__m128i *)(data + 5 * MAX_SB_SIZE)); + const __m128i src_7 = + _mm_unpackhi_epi16(*(__m128i *)(data + 6 * MAX_SB_SIZE), + *(__m128i *)(data + 7 * MAX_SB_SIZE)); + + const __m128i res_1 = _mm_madd_epi16(src_1, coeff_01); + const __m128i res_3 = _mm_madd_epi16(src_3, coeff_23); + const __m128i res_5 = _mm_madd_epi16(src_5, coeff_45); + const __m128i res_7 = _mm_madd_epi16(src_7, coeff_67); + + const __m128i res_odd = _mm_add_epi32(_mm_add_epi32(res_1, res_3), + _mm_add_epi32(res_5, res_7)); + + // Rearrange pixels back into the order 0 ... 7 + const __m128i res_lo = _mm_unpacklo_epi32(res_even, res_odd); + const __m128i res_hi = _mm_unpackhi_epi32(res_even, res_odd); + + const __m128i res_lo_round = _mm_srai_epi32( + _mm_add_epi32(res_lo, round_const), FILTER_BITS + EXTRAPREC_BITS); + const __m128i res_hi_round = _mm_srai_epi32( + _mm_add_epi32(res_hi, round_const), FILTER_BITS + EXTRAPREC_BITS); + + const __m128i res_16bit = _mm_packs_epi32(res_lo_round, res_hi_round); + __m128i res_8bit = _mm_packus_epi16(res_16bit, res_16bit); + + __m128i *const p = (__m128i *)&dst[i * dst_stride + j]; + _mm_storel_epi64(p, res_8bit); + } + } + } +} diff --git a/third_party/aom/aom_dsp/x86/aom_highbd_convolve_hip_ssse3.c b/third_party/aom/aom_dsp/x86/aom_highbd_convolve_hip_ssse3.c new file mode 100644 index 000000000..74ce80e50 --- /dev/null +++ b/third_party/aom/aom_dsp/x86/aom_highbd_convolve_hip_ssse3.c @@ -0,0 +1,203 @@ +/* + * 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 +#include + +#include "./aom_dsp_rtcd.h" +#include "aom_dsp/aom_convolve.h" +#include "aom_dsp/aom_dsp_common.h" +#include "aom_dsp/aom_filter.h" + +#if EXTRAPREC_BITS > 2 +#error "Highbd high-prec convolve filter only supports EXTRAPREC_BITS <= 2" +#error "(need to use 32-bit intermediates for EXTRAPREC_BITS > 2)" +#endif + +void aom_highbd_convolve8_add_src_hip_ssse3( + 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) { + assert(x_step_q4 == 16 && y_step_q4 == 16); + assert(!(w & 7)); + (void)x_step_q4; + (void)y_step_q4; + + const uint16_t *const src = CONVERT_TO_SHORTPTR(src8); + uint16_t *const dst = CONVERT_TO_SHORTPTR(dst8); + + uint16_t temp[(MAX_SB_SIZE + SUBPEL_TAPS - 1) * MAX_SB_SIZE]; + int intermediate_height = h + SUBPEL_TAPS - 1; + int i, j; + const int center_tap = ((SUBPEL_TAPS - 1) / 2); + const uint16_t *const src_ptr = src - center_tap * src_stride - center_tap; + + const __m128i zero = _mm_setzero_si128(); + // Add an offset to account for the "add_src" part of the convolve function. + const __m128i offset = _mm_insert_epi16(zero, 1 << FILTER_BITS, 3); + + /* Horizontal filter */ + { + const __m128i coeffs_x = + _mm_add_epi16(_mm_loadu_si128((__m128i *)filter_x), offset); + + // coeffs 0 1 0 1 2 3 2 3 + const __m128i tmp_0 = _mm_unpacklo_epi32(coeffs_x, coeffs_x); + // coeffs 4 5 4 5 6 7 6 7 + const __m128i tmp_1 = _mm_unpackhi_epi32(coeffs_x, coeffs_x); + + // coeffs 0 1 0 1 0 1 0 1 + const __m128i coeff_01 = _mm_unpacklo_epi64(tmp_0, tmp_0); + // coeffs 2 3 2 3 2 3 2 3 + const __m128i coeff_23 = _mm_unpackhi_epi64(tmp_0, tmp_0); + // coeffs 4 5 4 5 4 5 4 5 + const __m128i coeff_45 = _mm_unpacklo_epi64(tmp_1, tmp_1); + // coeffs 6 7 6 7 6 7 6 7 + const __m128i coeff_67 = _mm_unpackhi_epi64(tmp_1, tmp_1); + + const __m128i round_const = + _mm_set1_epi32((1 << (FILTER_BITS - EXTRAPREC_BITS - 1)) + + (1 << (bd + FILTER_BITS - 1))); + + for (i = 0; i < intermediate_height; ++i) { + for (j = 0; j < w; j += 8) { + const __m128i data = + _mm_loadu_si128((__m128i *)&src_ptr[i * src_stride + j]); + const __m128i data2 = + _mm_loadu_si128((__m128i *)&src_ptr[i * src_stride + j + 8]); + + // Filter even-index pixels + const __m128i res_0 = _mm_madd_epi16(data, coeff_01); + const __m128i res_2 = + _mm_madd_epi16(_mm_alignr_epi8(data2, data, 4), coeff_23); + const __m128i res_4 = + _mm_madd_epi16(_mm_alignr_epi8(data2, data, 8), coeff_45); + const __m128i res_6 = + _mm_madd_epi16(_mm_alignr_epi8(data2, data, 12), coeff_67); + + __m128i res_even = _mm_add_epi32(_mm_add_epi32(res_0, res_4), + _mm_add_epi32(res_2, res_6)); + res_even = _mm_srai_epi32(_mm_add_epi32(res_even, round_const), + FILTER_BITS - EXTRAPREC_BITS); + + // Filter odd-index pixels + const __m128i res_1 = + _mm_madd_epi16(_mm_alignr_epi8(data2, data, 2), coeff_01); + const __m128i res_3 = + _mm_madd_epi16(_mm_alignr_epi8(data2, data, 6), coeff_23); + const __m128i res_5 = + _mm_madd_epi16(_mm_alignr_epi8(data2, data, 10), coeff_45); + const __m128i res_7 = + _mm_madd_epi16(_mm_alignr_epi8(data2, data, 14), coeff_67); + + __m128i res_odd = _mm_add_epi32(_mm_add_epi32(res_1, res_5), + _mm_add_epi32(res_3, res_7)); + res_odd = _mm_srai_epi32(_mm_add_epi32(res_odd, round_const), + FILTER_BITS - EXTRAPREC_BITS); + + // Pack in the column order 0, 2, 4, 6, 1, 3, 5, 7 + const __m128i maxval = _mm_set1_epi16((EXTRAPREC_CLAMP_LIMIT(bd)) - 1); + __m128i res = _mm_packs_epi32(res_even, res_odd); + res = _mm_min_epi16(_mm_max_epi16(res, zero), maxval); + _mm_storeu_si128((__m128i *)&temp[i * MAX_SB_SIZE + j], res); + } + } + } + + /* Vertical filter */ + { + const __m128i coeffs_y = + _mm_add_epi16(_mm_loadu_si128((__m128i *)filter_y), offset); + + // coeffs 0 1 0 1 2 3 2 3 + const __m128i tmp_0 = _mm_unpacklo_epi32(coeffs_y, coeffs_y); + // coeffs 4 5 4 5 6 7 6 7 + const __m128i tmp_1 = _mm_unpackhi_epi32(coeffs_y, coeffs_y); + + // coeffs 0 1 0 1 0 1 0 1 + const __m128i coeff_01 = _mm_unpacklo_epi64(tmp_0, tmp_0); + // coeffs 2 3 2 3 2 3 2 3 + const __m128i coeff_23 = _mm_unpackhi_epi64(tmp_0, tmp_0); + // coeffs 4 5 4 5 4 5 4 5 + const __m128i coeff_45 = _mm_unpacklo_epi64(tmp_1, tmp_1); + // coeffs 6 7 6 7 6 7 6 7 + const __m128i coeff_67 = _mm_unpackhi_epi64(tmp_1, tmp_1); + + const __m128i round_const = + _mm_set1_epi32((1 << (FILTER_BITS + EXTRAPREC_BITS - 1)) - + (1 << (bd + FILTER_BITS + EXTRAPREC_BITS - 1))); + + for (i = 0; i < h; ++i) { + for (j = 0; j < w; j += 8) { + // Filter even-index pixels + const uint16_t *data = &temp[i * MAX_SB_SIZE + j]; + const __m128i src_0 = + _mm_unpacklo_epi16(*(__m128i *)(data + 0 * MAX_SB_SIZE), + *(__m128i *)(data + 1 * MAX_SB_SIZE)); + const __m128i src_2 = + _mm_unpacklo_epi16(*(__m128i *)(data + 2 * MAX_SB_SIZE), + *(__m128i *)(data + 3 * MAX_SB_SIZE)); + const __m128i src_4 = + _mm_unpacklo_epi16(*(__m128i *)(data + 4 * MAX_SB_SIZE), + *(__m128i *)(data + 5 * MAX_SB_SIZE)); + const __m128i src_6 = + _mm_unpacklo_epi16(*(__m128i *)(data + 6 * MAX_SB_SIZE), + *(__m128i *)(data + 7 * MAX_SB_SIZE)); + + const __m128i res_0 = _mm_madd_epi16(src_0, coeff_01); + const __m128i res_2 = _mm_madd_epi16(src_2, coeff_23); + const __m128i res_4 = _mm_madd_epi16(src_4, coeff_45); + const __m128i res_6 = _mm_madd_epi16(src_6, coeff_67); + + const __m128i res_even = _mm_add_epi32(_mm_add_epi32(res_0, res_2), + _mm_add_epi32(res_4, res_6)); + + // Filter odd-index pixels + const __m128i src_1 = + _mm_unpackhi_epi16(*(__m128i *)(data + 0 * MAX_SB_SIZE), + *(__m128i *)(data + 1 * MAX_SB_SIZE)); + const __m128i src_3 = + _mm_unpackhi_epi16(*(__m128i *)(data + 2 * MAX_SB_SIZE), + *(__m128i *)(data + 3 * MAX_SB_SIZE)); + const __m128i src_5 = + _mm_unpackhi_epi16(*(__m128i *)(data + 4 * MAX_SB_SIZE), + *(__m128i *)(data + 5 * MAX_SB_SIZE)); + const __m128i src_7 = + _mm_unpackhi_epi16(*(__m128i *)(data + 6 * MAX_SB_SIZE), + *(__m128i *)(data + 7 * MAX_SB_SIZE)); + + const __m128i res_1 = _mm_madd_epi16(src_1, coeff_01); + const __m128i res_3 = _mm_madd_epi16(src_3, coeff_23); + const __m128i res_5 = _mm_madd_epi16(src_5, coeff_45); + const __m128i res_7 = _mm_madd_epi16(src_7, coeff_67); + + const __m128i res_odd = _mm_add_epi32(_mm_add_epi32(res_1, res_3), + _mm_add_epi32(res_5, res_7)); + + // Rearrange pixels back into the order 0 ... 7 + const __m128i res_lo = _mm_unpacklo_epi32(res_even, res_odd); + const __m128i res_hi = _mm_unpackhi_epi32(res_even, res_odd); + + const __m128i res_lo_round = _mm_srai_epi32( + _mm_add_epi32(res_lo, round_const), FILTER_BITS + EXTRAPREC_BITS); + const __m128i res_hi_round = _mm_srai_epi32( + _mm_add_epi32(res_hi, round_const), FILTER_BITS + EXTRAPREC_BITS); + + const __m128i maxval = _mm_set1_epi16((1 << bd) - 1); + __m128i res_16bit = _mm_packs_epi32(res_lo_round, res_hi_round); + res_16bit = _mm_min_epi16(_mm_max_epi16(res_16bit, zero), maxval); + + __m128i *const p = (__m128i *)&dst[i * dst_stride + j]; + _mm_storeu_si128(p, res_16bit); + } + } + } +} diff --git a/third_party/aom/aom_dsp/x86/avg_intrin_sse2.c b/third_party/aom/aom_dsp/x86/avg_intrin_sse2.c index bcdc20f63..1a6457402 100644 --- a/third_party/aom/aom_dsp/x86/avg_intrin_sse2.c +++ b/third_party/aom/aom_dsp/x86/avg_intrin_sse2.c @@ -94,52 +94,6 @@ void aom_minmax_8x8_sse2(const uint8_t *s, int p, const uint8_t *d, int dp, *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]; diff --git a/third_party/aom/aom_dsp/x86/highbd_convolve_avx2.c b/third_party/aom/aom_dsp/x86/highbd_convolve_avx2.c index 7d96e26ae..133640eb7 100644 --- a/third_party/aom/aom_dsp/x86/highbd_convolve_avx2.c +++ b/third_party/aom/aom_dsp/x86/highbd_convolve_avx2.c @@ -14,30 +14,6 @@ #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 @@ -216,6 +192,27 @@ void aom_highbd_convolve_avg_avx2(const uint8_t *src8, ptrdiff_t src_stride, } } +// ----------------------------------------------------------------------------- +// Horizontal and vertical filtering + +#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 }; + // ----------------------------------------------------------------------------- // Horizontal Filtering @@ -248,52 +245,30 @@ static INLINE void pack_16_pixels(const __m256i *s0, const __m256i *s1, 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_8x1_pixels(const uint16_t *src, __m256i *x) { + __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); } -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_8x2_pixels(const uint16_t *src, ptrdiff_t stride, + __m256i *x) { + __m256i s0, s1; + s0 = _mm256_loadu_si256((const __m256i *)src); + s1 = _mm256_loadu_si256((const __m256i *)(src + stride)); + pack_16_pixels(&s0, &s1, 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); +static INLINE void pack_16x1_pixels(const uint16_t *src, __m256i *x) { + __m256i s0, s1; + s0 = _mm256_loadu_si256((const __m256i *)src); + s1 = _mm256_loadu_si256((const __m256i *)(src + 8)); + pack_16_pixels(&s0, &s1, x); } // Note: @@ -323,51 +298,49 @@ static INLINE void filter_8x1_pixels(const __m256i *sig /*sig[4]*/, 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); + { + 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) { +static INLINE void store_8x1_pixels(const __m256i *y, const __m256i *mask, + uint16_t *dst) { 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) { +static INLINE void store_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; +static INLINE void store_16x1_pixels(const __m256i *y0, const __m256i *y1, + const __m256i *mask, uint16_t *dst) { __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) { +static void aom_highbd_filter_block1d8_h8_avx2( + const uint16_t *src_ptr, ptrdiff_t src_pitch, 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); @@ -379,32 +352,22 @@ static void filter_block_width8_horiz( 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); + store_8x2_pixels(&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); + pack_8x1_pixels(src_ptr, signal); filter_8x1_pixels(signal, ff, &res0); - write_8x1(&res0, &res1, &max, dst_ptr, dst_pitch); + store_8x1_pixels(&res0, &max, dst_ptr); } } -static void aom_highbd_filter_block1d8_h8_avx2( - const uint16_t *src, ptrdiff_t src_pitch, uint16_t *dst, +static void aom_highbd_filter_block1d16_h8_avx2( + const uint16_t *src_ptr, ptrdiff_t src_pitch, uint16_t *dst_ptr, 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); @@ -413,23 +376,17 @@ static void filter_block_width16_horiz(const uint16_t *src_ptr, src_ptr -= 3; do { - pack_16x1_pixels(src_ptr, src_pitch, signal); + pack_16x1_pixels(src_ptr, signal); filter_8x1_pixels(signal, ff, &res0); filter_8x1_pixels(&signal[4], ff, &res1); - write_16x1(&res0, &res1, &max, dst_ptr, dst_pitch); + store_16x1_pixels(&res0, &res1, &max, dst_ptr); 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) { @@ -493,16 +450,6 @@ static INLINE void filter_16_2t_pixels(const __m256i *sig, const __m256i *f, *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)); @@ -511,10 +458,9 @@ static INLINE void filter_8x1_2t_pixels(const __m256i *sig, const __m256i *f, *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) { +static void aom_highbd_filter_block1d8_h2_avx2( + const uint16_t *src_ptr, ptrdiff_t src_pitch, 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); @@ -524,8 +470,8 @@ static void filter_block_width8_2t_horiz( 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); + filter_16_2t_pixels(signal, &ff, &res0, &res1); + store_8x2_pixels(&res0, &res1, &max, dst_ptr, dst_pitch); height -= 2; src_ptr += src_pitch << 1; dst_ptr += dst_pitch << 1; @@ -534,24 +480,13 @@ static void filter_block_width8_2t_horiz( 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); + store_8x1_pixels(&res0, &max, dst_ptr); } } -static void aom_highbd_filter_block1d8_h2_avx2( - const uint16_t *src, ptrdiff_t src_pitch, uint16_t *dst, +static void aom_highbd_filter_block1d16_h2_avx2( + const uint16_t *src_ptr, ptrdiff_t src_pitch, uint16_t *dst_ptr, 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); @@ -561,21 +496,15 @@ static void filter_block_width16_2t_horiz(const uint16_t *src_ptr, 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); + filter_16_2t_pixels(signal, &ff, &res0, &res1); + store_16x1_pixels(&res0, &res1, &max, dst_ptr); 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) { @@ -638,22 +567,9 @@ static INLINE void update_pixels(__m256i *sig) { } } -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) { +static void aom_highbd_filter_block1d8_v8_avx2( + const uint16_t *src_ptr, ptrdiff_t src_pitch, 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); @@ -666,27 +582,13 @@ static void filter_block_width8_vert(const uint16_t *src_ptr, pack_8x9_pixels(src_ptr, src_pitch, signal); filter_8x9_pixels(signal, ff, &res0, &res1); - write_8x2(&res0, &res1, &max, dst_ptr, dst_pitch); + store_8x2_pixels(&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); + } while (height > 0); } static void pack_16x9_init(const uint16_t *src, ptrdiff_t pitch, __m256i *sig) { @@ -770,13 +672,15 @@ static INLINE void filter_16x9_pixels(const __m256i *sig, const __m256i *f, 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); + { + 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, +static INLINE void store_16x2_pixels(const __m256i *y0, const __m256i *y1, const __m256i *mask, uint16_t *dst, ptrdiff_t pitch) { __m256i p = _mm256_min_epi16(*y0, *mask); @@ -785,26 +689,14 @@ static INLINE void write_16x2_pixels(const __m256i *y0, const __m256i *y1, _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) { +static void aom_highbd_filter_block1d16_v8_avx2( + const uint16_t *src_ptr, ptrdiff_t src_pitch, 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); @@ -816,29 +708,16 @@ static void filter_block_width16_vert(const uint16_t *src_ptr, 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); + store_16x2_pixels(&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); + } while (height > 0); } +// ----------------------------------------------------------------------------- // 2-tap vertical filtering static void pack_16x2_init(const uint16_t *src, __m256i *sig) { @@ -859,12 +738,9 @@ static INLINE void filter_16x2_2t_pixels(const __m256i *sig, const __m256i *f, 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) { +static void aom_highbd_filter_block1d16_v2_avx2( + const uint16_t *src_ptr, ptrdiff_t src_pitch, 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; @@ -875,7 +751,7 @@ static void filter_block_width16_2t_vert(const uint16_t *src_ptr, 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); + store_16x1_pixels(&res0, &res1, &max, dst_ptr); src_ptr += src_pitch; dst_ptr += dst_pitch; @@ -883,13 +759,6 @@ static void filter_block_width16_2t_vert(const uint16_t *src_ptr, } 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); @@ -920,22 +789,16 @@ static INLINE void filter_8_2t_pixels(const __m128i *sig, const __m128i *f, *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) { +static INLINE void store_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) { +static void aom_highbd_filter_block1d8_v2_avx2( + const uint16_t *src_ptr, ptrdiff_t src_pitch, 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; @@ -946,7 +809,7 @@ static void filter_block_width8_2t_vert(const uint16_t *src_ptr, 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); + store_8x1_2t_pixels_ver(&res0, &res1, &max, dst_ptr); src_ptr += src_pitch; dst_ptr += dst_pitch; @@ -954,20 +817,10 @@ static void filter_block_width8_2t_vert(const uint16_t *src_ptr, } 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; +static INLINE void store_8x1_avg_pixels(const __m256i *y0, const __m256i *mask, + uint16_t *dst) { const __m128i a0 = _mm256_castsi256_si128(*y0); const __m128i a1 = _mm256_extractf128_si256(*y0, 1); __m128i res = _mm_packus_epi32(a0, a1); @@ -977,9 +830,9 @@ static void write_8x1_avg_pixels(const __m256i *y0, const __m256i *y1, _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) { +static INLINE void store_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)); @@ -991,10 +844,8 @@ static void write_8x2_avg_pixels(const __m256i *y0, const __m256i *y1, _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; +static INLINE void store_16x1_avg_pixels(const __m256i *y0, const __m256i *y1, + const __m256i *mask, uint16_t *dst) { __m256i a = _mm256_packus_epi32(*y0, *y1); const __m256i pix = _mm256_loadu_si256((const __m256i *)dst); a = _mm256_min_epi16(a, *mask); @@ -1002,21 +853,7 @@ static void write_16x1_avg_pixels(const __m256i *y0, const __m256i *y1, _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, +static INLINE void store_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); @@ -1030,20 +867,10 @@ static INLINE void write_16x2_avg_pixels(const __m256i *y0, const __m256i *y1, _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) { +static INLINE void store_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); @@ -1052,96 +879,229 @@ static void write_8x1_2t_avg_pixels_ver(const __m128i *y0, const __m128i *y1, } static void aom_highbd_filter_block1d8_h8_avg_avx2( - const uint16_t *src, ptrdiff_t src_pitch, uint16_t *dst, + const uint16_t *src_ptr, ptrdiff_t src_pitch, uint16_t *dst_ptr, 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); + __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); + store_8x2_avg_pixels(&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, signal); + filter_8x1_pixels(signal, ff, &res0); + store_8x1_avg_pixels(&res0, &max, dst_ptr); + } } static void aom_highbd_filter_block1d16_h8_avg_avx2( - const uint16_t *src, ptrdiff_t src_pitch, uint16_t *dst, + const uint16_t *src_ptr, ptrdiff_t src_pitch, uint16_t *dst_ptr, 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); + __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, signal); + filter_8x1_pixels(signal, ff, &res0); + filter_8x1_pixels(&signal[4], ff, &res1); + store_16x1_avg_pixels(&res0, &res1, &max, dst_ptr); + height -= 1; + src_ptr += src_pitch; + dst_ptr += dst_pitch; + } while (height > 0); } static void aom_highbd_filter_block1d8_v8_avg_avx2( - const uint16_t *src, ptrdiff_t src_pitch, uint16_t *dst, + const uint16_t *src_ptr, ptrdiff_t src_pitch, uint16_t *dst_ptr, 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); + __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); + store_8x2_avg_pixels(&res0, &res1, &max, dst_ptr, dst_pitch); + update_pixels(signal); + + src_ptr += src_pitch << 1; + dst_ptr += dst_pitch << 1; + height -= 2; + } while (height > 0); } static void aom_highbd_filter_block1d16_v8_avg_avx2( - const uint16_t *src, ptrdiff_t src_pitch, uint16_t *dst, + const uint16_t *src_ptr, ptrdiff_t src_pitch, uint16_t *dst_ptr, 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); -} + __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); + store_16x2_avg_pixels(&res0, &res1, &max, dst_ptr, dst_pitch); + update_16x9_pixels(signal); -// 2-tap averaging + src_ptr += src_pitch << 1; + dst_ptr += dst_pitch << 1; + height -= 2; + } while (height > 0); +} static void aom_highbd_filter_block1d8_h2_avg_avx2( - const uint16_t *src, ptrdiff_t src_pitch, uint16_t *dst, + const uint16_t *src_ptr, ptrdiff_t src_pitch, uint16_t *dst_ptr, 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); + __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_16_2t_pixels(signal, &ff, &res0, &res1); + store_8x2_avg_pixels(&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); + store_8x1_avg_pixels(&res0, &max, dst_ptr); + } } static void aom_highbd_filter_block1d16_h2_avg_avx2( - const uint16_t *src, ptrdiff_t src_pitch, uint16_t *dst, + const uint16_t *src_ptr, ptrdiff_t src_pitch, uint16_t *dst_ptr, 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); + __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_16_2t_pixels(signal, &ff, &res0, &res1); + store_16x1_avg_pixels(&res0, &res1, &max, dst_ptr); + height -= 1; + src_ptr += src_pitch; + dst_ptr += dst_pitch; + } while (height > 0); } static void aom_highbd_filter_block1d16_v2_avg_avx2( - const uint16_t *src, ptrdiff_t src_pitch, uint16_t *dst, + const uint16_t *src_ptr, ptrdiff_t src_pitch, uint16_t *dst_ptr, 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); + __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); + store_16x1_avg_pixels(&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_avg_avx2( - const uint16_t *src, ptrdiff_t src_pitch, uint16_t *dst, + const uint16_t *src_ptr, ptrdiff_t src_pitch, uint16_t *dst_ptr, 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); -} + __m128i signal[3], res0, res1; + const __m128i max = _mm_set1_epi16((1 << bd) - 1); + __m128i ff; -typedef void HbdFilter1dFunc(const uint16_t *, ptrdiff_t, uint16_t *, ptrdiff_t, - uint32_t, const int16_t *, int); + pack_8x1_2t_filter(filter, &ff); + pack_8x2_init(src_ptr, signal); -#define HIGHBD_FUNC(width, dir, avg, opt) \ - aom_highbd_filter_block1d##width##_##dir##_##avg##opt + do { + pack_8x2_2t_pixels_ver(src_ptr, src_pitch, signal); + filter_8_2t_pixels(signal, &ff, &res0, &res1); + store_8x1_2t_avg_pixels_ver(&res0, &res1, &max, dst_ptr); -HbdFilter1dFunc HIGHBD_FUNC(4, h8, , sse2); -HbdFilter1dFunc HIGHBD_FUNC(4, h2, , sse2); -HbdFilter1dFunc HIGHBD_FUNC(4, v8, , sse2); -HbdFilter1dFunc HIGHBD_FUNC(4, v2, , sse2); + src_ptr += src_pitch; + dst_ptr += dst_pitch; + height -= 1; + } while (height > 0); +} -#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) +void aom_highbd_filter_block1d4_h8_sse2(const uint16_t *, ptrdiff_t, uint16_t *, + ptrdiff_t, uint32_t, const int16_t *, + int); +void aom_highbd_filter_block1d4_h2_sse2(const uint16_t *, ptrdiff_t, uint16_t *, + ptrdiff_t, uint32_t, const int16_t *, + int); +void aom_highbd_filter_block1d4_v8_sse2(const uint16_t *, ptrdiff_t, uint16_t *, + ptrdiff_t, uint32_t, const int16_t *, + int); +void aom_highbd_filter_block1d4_v2_sse2(const uint16_t *, ptrdiff_t, uint16_t *, + ptrdiff_t, uint32_t, const int16_t *, + int); +#define aom_highbd_filter_block1d4_h8_avx2 aom_highbd_filter_block1d4_h8_sse2 +#define aom_highbd_filter_block1d4_h2_avx2 aom_highbd_filter_block1d4_h2_sse2 +#define aom_highbd_filter_block1d4_v8_avx2 aom_highbd_filter_block1d4_v8_sse2 +#define aom_highbd_filter_block1d4_v2_avx2 aom_highbd_filter_block1d4_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) +void aom_highbd_filter_block1d4_h8_avg_sse2(const uint16_t *, ptrdiff_t, + uint16_t *, ptrdiff_t, uint32_t, + const int16_t *, int); +void aom_highbd_filter_block1d4_h2_avg_sse2(const uint16_t *, ptrdiff_t, + uint16_t *, ptrdiff_t, uint32_t, + const int16_t *, int); +void aom_highbd_filter_block1d4_v8_avg_sse2(const uint16_t *, ptrdiff_t, + uint16_t *, ptrdiff_t, uint32_t, + const int16_t *, int); +void aom_highbd_filter_block1d4_v2_avg_sse2(const uint16_t *, ptrdiff_t, + uint16_t *, ptrdiff_t, uint32_t, + const int16_t *, int); +#define aom_highbd_filter_block1d4_h8_avg_avx2 \ + aom_highbd_filter_block1d4_h8_avg_sse2 +#define aom_highbd_filter_block1d4_h2_avg_avx2 \ + aom_highbd_filter_block1d4_h2_avg_sse2 +#define aom_highbd_filter_block1d4_v8_avg_avx2 \ + aom_highbd_filter_block1d4_v8_avg_sse2 +#define aom_highbd_filter_block1d4_v2_avg_avx2 \ + aom_highbd_filter_block1d4_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_, diff --git a/third_party/aom/aom_dsp/x86/inv_txfm_avx2.c b/third_party/aom/aom_dsp/x86/inv_txfm_avx2.c new file mode 100644 index 000000000..a9d6a127c --- /dev/null +++ b/third_party/aom/aom_dsp/x86/inv_txfm_avx2.c @@ -0,0 +1,1238 @@ +/* + * 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 + +#include "./aom_dsp_rtcd.h" +#include "aom_dsp/inv_txfm.h" +#include "aom_dsp/x86/inv_txfm_common_avx2.h" +#include "aom_dsp/x86/txfm_common_avx2.h" + +void aom_idct16x16_256_add_avx2(const tran_low_t *input, uint8_t *dest, + int stride) { + __m256i in[16]; + load_buffer_16x16(input, in); + mm256_transpose_16x16(in, in); + av1_idct16_avx2(in); + mm256_transpose_16x16(in, in); + av1_idct16_avx2(in); + store_buffer_16xN(in, stride, dest, 16); +} + +static INLINE void transpose_col_to_row_nz4x4(__m256i *in /*in[4]*/) { + const __m256i u0 = _mm256_unpacklo_epi16(in[0], in[1]); + const __m256i u1 = _mm256_unpacklo_epi16(in[2], in[3]); + const __m256i v0 = _mm256_unpacklo_epi32(u0, u1); + const __m256i v1 = _mm256_unpackhi_epi32(u0, u1); + in[0] = _mm256_permute4x64_epi64(v0, 0xA8); + in[1] = _mm256_permute4x64_epi64(v0, 0xA9); + in[2] = _mm256_permute4x64_epi64(v1, 0xA8); + in[3] = _mm256_permute4x64_epi64(v1, 0xA9); +} + +#define MM256_SHUFFLE_EPI64(x0, x1, imm8) \ + _mm256_castpd_si256(_mm256_shuffle_pd(_mm256_castsi256_pd(x0), \ + _mm256_castsi256_pd(x1), imm8)) + +static INLINE void transpose_col_to_row_nz4x16(__m256i *in /*in[16]*/) { + int i; + for (i = 0; i < 16; i += 4) { + transpose_col_to_row_nz4x4(&in[i]); + } + + for (i = 0; i < 4; ++i) { + in[i] = MM256_SHUFFLE_EPI64(in[i], in[i + 4], 0); + in[i + 8] = MM256_SHUFFLE_EPI64(in[i + 8], in[i + 12], 0); + } + + for (i = 0; i < 4; ++i) { + in[i] = _mm256_permute2x128_si256(in[i], in[i + 8], 0x20); + } +} + +// Coefficients 0-7 before the final butterfly +static INLINE void idct16_10_first_half(const __m256i *in, __m256i *out) { + const __m256i c2p28 = pair256_set_epi16(2 * cospi_28_64, 2 * cospi_28_64); + const __m256i c2p04 = pair256_set_epi16(2 * cospi_4_64, 2 * cospi_4_64); + const __m256i v4 = _mm256_mulhrs_epi16(in[2], c2p28); + const __m256i v7 = _mm256_mulhrs_epi16(in[2], c2p04); + + const __m256i c2p16 = pair256_set_epi16(2 * cospi_16_64, 2 * cospi_16_64); + const __m256i v0 = _mm256_mulhrs_epi16(in[0], c2p16); + const __m256i v1 = v0; + + const __m256i cospi_p16_p16 = _mm256_set1_epi16((int16_t)cospi_16_64); + const __m256i cospi_p16_m16 = pair256_set_epi16(cospi_16_64, -cospi_16_64); + __m256i v5, v6; + unpack_butter_fly(&v7, &v4, &cospi_p16_m16, &cospi_p16_p16, &v5, &v6); + + out[0] = _mm256_add_epi16(v0, v7); + out[1] = _mm256_add_epi16(v1, v6); + out[2] = _mm256_add_epi16(v1, v5); + out[3] = _mm256_add_epi16(v0, v4); + out[4] = _mm256_sub_epi16(v0, v4); + out[5] = _mm256_sub_epi16(v1, v5); + out[6] = _mm256_sub_epi16(v1, v6); + out[7] = _mm256_sub_epi16(v0, v7); +} + +// Coefficients 8-15 before the final butterfly +static INLINE void idct16_10_second_half(const __m256i *in, __m256i *out) { + const __m256i c2p30 = pair256_set_epi16(2 * cospi_30_64, 2 * cospi_30_64); + const __m256i c2p02 = pair256_set_epi16(2 * cospi_2_64, 2 * cospi_2_64); + const __m256i t0 = _mm256_mulhrs_epi16(in[1], c2p30); + const __m256i t7 = _mm256_mulhrs_epi16(in[1], c2p02); + + const __m256i c2m26 = pair256_set_epi16(-2 * cospi_26_64, -2 * cospi_26_64); + const __m256i c2p06 = pair256_set_epi16(2 * cospi_6_64, 2 * cospi_6_64); + const __m256i t3 = _mm256_mulhrs_epi16(in[3], c2m26); + const __m256i t4 = _mm256_mulhrs_epi16(in[3], c2p06); + + const __m256i cospi_m08_p24 = pair256_set_epi16(-cospi_8_64, cospi_24_64); + const __m256i cospi_p24_p08 = pair256_set_epi16(cospi_24_64, cospi_8_64); + const __m256i cospi_m24_m08 = pair256_set_epi16(-cospi_24_64, -cospi_8_64); + + __m256i t1, t2, t5, t6; + unpack_butter_fly(&t0, &t7, &cospi_m08_p24, &cospi_p24_p08, &t1, &t6); + unpack_butter_fly(&t3, &t4, &cospi_m24_m08, &cospi_m08_p24, &t2, &t5); + + out[0] = _mm256_add_epi16(t0, t3); + out[1] = _mm256_add_epi16(t1, t2); + out[6] = _mm256_add_epi16(t6, t5); + out[7] = _mm256_add_epi16(t7, t4); + + const __m256i v2 = _mm256_sub_epi16(t1, t2); + const __m256i v3 = _mm256_sub_epi16(t0, t3); + const __m256i v4 = _mm256_sub_epi16(t7, t4); + const __m256i v5 = _mm256_sub_epi16(t6, t5); + const __m256i cospi_p16_p16 = _mm256_set1_epi16((int16_t)cospi_16_64); + const __m256i cospi_p16_m16 = pair256_set_epi16(cospi_16_64, -cospi_16_64); + unpack_butter_fly(&v5, &v2, &cospi_p16_m16, &cospi_p16_p16, &out[2], &out[5]); + unpack_butter_fly(&v4, &v3, &cospi_p16_m16, &cospi_p16_p16, &out[3], &out[4]); +} + +static INLINE void add_sub_butterfly(const __m256i *in, __m256i *out, + int size) { + int i = 0; + const int num = size >> 1; + const int bound = size - 1; + while (i < num) { + out[i] = _mm256_add_epi16(in[i], in[bound - i]); + out[bound - i] = _mm256_sub_epi16(in[i], in[bound - i]); + i++; + } +} + +static INLINE void idct16_10(__m256i *in /*in[16]*/) { + __m256i out[16]; + idct16_10_first_half(in, out); + idct16_10_second_half(in, &out[8]); + add_sub_butterfly(out, in, 16); +} + +void aom_idct16x16_10_add_avx2(const tran_low_t *input, uint8_t *dest, + int stride) { + __m256i in[16]; + + load_coeff(input, &in[0]); + load_coeff(input + 16, &in[1]); + load_coeff(input + 32, &in[2]); + load_coeff(input + 48, &in[3]); + + transpose_col_to_row_nz4x4(in); + idct16_10(in); + + transpose_col_to_row_nz4x16(in); + idct16_10(in); + + store_buffer_16xN(in, stride, dest, 16); +} + +// Note: +// For 16x16 int16_t matrix +// transpose first 8 columns into first 8 rows. +// Since only upper-left 8x8 are non-zero, the input are first 8 rows (in[8]). +// After transposing, the 8 row vectors are in in[8]. +void transpose_col_to_row_nz8x8(__m256i *in /*in[8]*/) { + __m256i u0 = _mm256_unpacklo_epi16(in[0], in[1]); + __m256i u1 = _mm256_unpackhi_epi16(in[0], in[1]); + __m256i u2 = _mm256_unpacklo_epi16(in[2], in[3]); + __m256i u3 = _mm256_unpackhi_epi16(in[2], in[3]); + + const __m256i v0 = _mm256_unpacklo_epi32(u0, u2); + const __m256i v1 = _mm256_unpackhi_epi32(u0, u2); + const __m256i v2 = _mm256_unpacklo_epi32(u1, u3); + const __m256i v3 = _mm256_unpackhi_epi32(u1, u3); + + u0 = _mm256_unpacklo_epi16(in[4], in[5]); + u1 = _mm256_unpackhi_epi16(in[4], in[5]); + u2 = _mm256_unpacklo_epi16(in[6], in[7]); + u3 = _mm256_unpackhi_epi16(in[6], in[7]); + + const __m256i v4 = _mm256_unpacklo_epi32(u0, u2); + const __m256i v5 = _mm256_unpackhi_epi32(u0, u2); + const __m256i v6 = _mm256_unpacklo_epi32(u1, u3); + const __m256i v7 = _mm256_unpackhi_epi32(u1, u3); + + in[0] = MM256_SHUFFLE_EPI64(v0, v4, 0); + in[1] = MM256_SHUFFLE_EPI64(v0, v4, 3); + in[2] = MM256_SHUFFLE_EPI64(v1, v5, 0); + in[3] = MM256_SHUFFLE_EPI64(v1, v5, 3); + in[4] = MM256_SHUFFLE_EPI64(v2, v6, 0); + in[5] = MM256_SHUFFLE_EPI64(v2, v6, 3); + in[6] = MM256_SHUFFLE_EPI64(v3, v7, 0); + in[7] = MM256_SHUFFLE_EPI64(v3, v7, 3); +} + +// Note: +// For 16x16 int16_t matrix +// transpose first 8 columns into first 8 rows. +// Since only matrix left 8x16 are non-zero, the input are total 16 rows +// (in[16]). +// After transposing, the 8 row vectors are in in[8]. All else are zero. +static INLINE void transpose_col_to_row_nz8x16(__m256i *in /*in[16]*/) { + transpose_col_to_row_nz8x8(in); + transpose_col_to_row_nz8x8(&in[8]); + + int i; + for (i = 0; i < 8; ++i) { + in[i] = _mm256_permute2x128_si256(in[i], in[i + 8], 0x20); + } +} + +static INLINE void idct16_38_first_half(const __m256i *in, __m256i *out) { + const __m256i c2p28 = pair256_set_epi16(2 * cospi_28_64, 2 * cospi_28_64); + const __m256i c2p04 = pair256_set_epi16(2 * cospi_4_64, 2 * cospi_4_64); + __m256i t4 = _mm256_mulhrs_epi16(in[2], c2p28); + __m256i t7 = _mm256_mulhrs_epi16(in[2], c2p04); + + const __m256i c2m20 = pair256_set_epi16(-2 * cospi_20_64, -2 * cospi_20_64); + const __m256i c2p12 = pair256_set_epi16(2 * cospi_12_64, 2 * cospi_12_64); + __m256i t5 = _mm256_mulhrs_epi16(in[6], c2m20); + __m256i t6 = _mm256_mulhrs_epi16(in[6], c2p12); + + const __m256i c2p16 = pair256_set_epi16(2 * cospi_16_64, 2 * cospi_16_64); + const __m256i c2p24 = pair256_set_epi16(2 * cospi_24_64, 2 * cospi_24_64); + const __m256i c2p08 = pair256_set_epi16(2 * cospi_8_64, 2 * cospi_8_64); + const __m256i u0 = _mm256_mulhrs_epi16(in[0], c2p16); + const __m256i u1 = _mm256_mulhrs_epi16(in[0], c2p16); + const __m256i u2 = _mm256_mulhrs_epi16(in[4], c2p24); + const __m256i u3 = _mm256_mulhrs_epi16(in[4], c2p08); + + const __m256i u4 = _mm256_add_epi16(t4, t5); + const __m256i u5 = _mm256_sub_epi16(t4, t5); + const __m256i u6 = _mm256_sub_epi16(t7, t6); + const __m256i u7 = _mm256_add_epi16(t7, t6); + + const __m256i t0 = _mm256_add_epi16(u0, u3); + const __m256i t1 = _mm256_add_epi16(u1, u2); + const __m256i t2 = _mm256_sub_epi16(u1, u2); + const __m256i t3 = _mm256_sub_epi16(u0, u3); + + t4 = u4; + t7 = u7; + + const __m256i cospi_p16_p16 = _mm256_set1_epi16((int16_t)cospi_16_64); + const __m256i cospi_p16_m16 = pair256_set_epi16(cospi_16_64, -cospi_16_64); + unpack_butter_fly(&u6, &u5, &cospi_p16_m16, &cospi_p16_p16, &t5, &t6); + + out[0] = _mm256_add_epi16(t0, t7); + out[1] = _mm256_add_epi16(t1, t6); + out[2] = _mm256_add_epi16(t2, t5); + out[3] = _mm256_add_epi16(t3, t4); + out[4] = _mm256_sub_epi16(t3, t4); + out[5] = _mm256_sub_epi16(t2, t5); + out[6] = _mm256_sub_epi16(t1, t6); + out[7] = _mm256_sub_epi16(t0, t7); +} + +static INLINE void idct16_38_second_half(const __m256i *in, __m256i *out) { + const __m256i c2p30 = pair256_set_epi16(2 * cospi_30_64, 2 * cospi_30_64); + const __m256i c2p02 = pair256_set_epi16(2 * cospi_2_64, 2 * cospi_2_64); + __m256i t0 = _mm256_mulhrs_epi16(in[1], c2p30); + __m256i t7 = _mm256_mulhrs_epi16(in[1], c2p02); + + const __m256i c2m18 = pair256_set_epi16(-2 * cospi_18_64, -2 * cospi_18_64); + const __m256i c2p14 = pair256_set_epi16(2 * cospi_14_64, 2 * cospi_14_64); + __m256i t1 = _mm256_mulhrs_epi16(in[7], c2m18); + __m256i t6 = _mm256_mulhrs_epi16(in[7], c2p14); + + const __m256i c2p22 = pair256_set_epi16(2 * cospi_22_64, 2 * cospi_22_64); + const __m256i c2p10 = pair256_set_epi16(2 * cospi_10_64, 2 * cospi_10_64); + __m256i t2 = _mm256_mulhrs_epi16(in[5], c2p22); + __m256i t5 = _mm256_mulhrs_epi16(in[5], c2p10); + + const __m256i c2m26 = pair256_set_epi16(-2 * cospi_26_64, -2 * cospi_26_64); + const __m256i c2p06 = pair256_set_epi16(2 * cospi_6_64, 2 * cospi_6_64); + __m256i t3 = _mm256_mulhrs_epi16(in[3], c2m26); + __m256i t4 = _mm256_mulhrs_epi16(in[3], c2p06); + + __m256i v0, v1, v2, v3, v4, v5, v6, v7; + v0 = _mm256_add_epi16(t0, t1); + v1 = _mm256_sub_epi16(t0, t1); + v2 = _mm256_sub_epi16(t3, t2); + v3 = _mm256_add_epi16(t2, t3); + v4 = _mm256_add_epi16(t4, t5); + v5 = _mm256_sub_epi16(t4, t5); + v6 = _mm256_sub_epi16(t7, t6); + v7 = _mm256_add_epi16(t6, t7); + + t0 = v0; + t7 = v7; + t3 = v3; + t4 = v4; + const __m256i cospi_m08_p24 = pair256_set_epi16(-cospi_8_64, cospi_24_64); + const __m256i cospi_p24_p08 = pair256_set_epi16(cospi_24_64, cospi_8_64); + const __m256i cospi_m24_m08 = pair256_set_epi16(-cospi_24_64, -cospi_8_64); + unpack_butter_fly(&v1, &v6, &cospi_m08_p24, &cospi_p24_p08, &t1, &t6); + unpack_butter_fly(&v2, &v5, &cospi_m24_m08, &cospi_m08_p24, &t2, &t5); + + v0 = _mm256_add_epi16(t0, t3); + v1 = _mm256_add_epi16(t1, t2); + v2 = _mm256_sub_epi16(t1, t2); + v3 = _mm256_sub_epi16(t0, t3); + v4 = _mm256_sub_epi16(t7, t4); + v5 = _mm256_sub_epi16(t6, t5); + v6 = _mm256_add_epi16(t6, t5); + v7 = _mm256_add_epi16(t7, t4); + + // stage 6, (8-15) + out[0] = v0; + out[1] = v1; + out[6] = v6; + out[7] = v7; + const __m256i cospi_p16_p16 = _mm256_set1_epi16((int16_t)cospi_16_64); + const __m256i cospi_p16_m16 = pair256_set_epi16(cospi_16_64, -cospi_16_64); + unpack_butter_fly(&v5, &v2, &cospi_p16_m16, &cospi_p16_p16, &out[2], &out[5]); + unpack_butter_fly(&v4, &v3, &cospi_p16_m16, &cospi_p16_p16, &out[3], &out[4]); +} + +static INLINE void idct16_38(__m256i *in /*in[16]*/) { + __m256i out[16]; + idct16_38_first_half(in, out); + idct16_38_second_half(in, &out[8]); + add_sub_butterfly(out, in, 16); +} + +void aom_idct16x16_38_add_avx2(const tran_low_t *input, uint8_t *dest, + int stride) { + __m256i in[16]; + + int i; + for (i = 0; i < 8; ++i) { + load_coeff(input + (i << 4), &in[i]); + } + + transpose_col_to_row_nz8x8(in); + idct16_38(in); + + transpose_col_to_row_nz8x16(in); + idct16_38(in); + + store_buffer_16xN(in, stride, dest, 16); +} + +static INLINE int calculate_dc(const tran_low_t *input) { + int dc = (int)dct_const_round_shift(input[0] * cospi_16_64); + dc = (int)dct_const_round_shift(dc * cospi_16_64); + dc = ROUND_POWER_OF_TWO(dc, IDCT_ROUNDING_POS); + return dc; +} + +void aom_idct16x16_1_add_avx2(const tran_low_t *input, uint8_t *dest, + int stride) { + const int dc = calculate_dc(input); + if (dc == 0) return; + + const __m256i dc_value = _mm256_set1_epi16(dc); + + int i; + for (i = 0; i < 16; ++i) { + recon_and_store(&dc_value, dest); + dest += stride; + } +} + +// ----------------------------------------------------------------------------- +// 32x32 partial IDCT + +void aom_idct32x32_1_add_avx2(const tran_low_t *input, uint8_t *dest, + int stride) { + const int dc = calculate_dc(input); + if (dc == 0) return; + + const __m256i dc_value = _mm256_set1_epi16(dc); + + int i; + for (i = 0; i < 32; ++i) { + recon_and_store(&dc_value, dest); + recon_and_store(&dc_value, dest + 16); + dest += stride; + } +} + +static void load_buffer_32x16(const tran_low_t *input, __m256i *in /*in[32]*/) { + int i; + for (i = 0; i < 16; ++i) { + load_coeff(input, &in[i]); + load_coeff(input + 16, &in[i + 16]); + input += 32; + } +} + +// Note: +// We extend SSSE3 operations to AVX2. Instead of operating on __m128i, we +// operate coefficients on __m256i. Our operation capacity doubles for each +// instruction. +#define BUTTERFLY_PAIR(x0, x1, co0, co1) \ + do { \ + tmp0 = _mm256_madd_epi16(x0, co0); \ + tmp1 = _mm256_madd_epi16(x1, co0); \ + tmp2 = _mm256_madd_epi16(x0, co1); \ + tmp3 = _mm256_madd_epi16(x1, co1); \ + tmp0 = _mm256_add_epi32(tmp0, rounding); \ + tmp1 = _mm256_add_epi32(tmp1, rounding); \ + tmp2 = _mm256_add_epi32(tmp2, rounding); \ + tmp3 = _mm256_add_epi32(tmp3, rounding); \ + tmp0 = _mm256_srai_epi32(tmp0, DCT_CONST_BITS); \ + tmp1 = _mm256_srai_epi32(tmp1, DCT_CONST_BITS); \ + tmp2 = _mm256_srai_epi32(tmp2, DCT_CONST_BITS); \ + tmp3 = _mm256_srai_epi32(tmp3, DCT_CONST_BITS); \ + } while (0) + +static INLINE void butterfly(const __m256i *x0, const __m256i *x1, + const __m256i *c0, const __m256i *c1, __m256i *y0, + __m256i *y1) { + __m256i tmp0, tmp1, tmp2, tmp3, u0, u1; + const __m256i rounding = _mm256_set1_epi32(DCT_CONST_ROUNDING); + + u0 = _mm256_unpacklo_epi16(*x0, *x1); + u1 = _mm256_unpackhi_epi16(*x0, *x1); + BUTTERFLY_PAIR(u0, u1, *c0, *c1); + *y0 = _mm256_packs_epi32(tmp0, tmp1); + *y1 = _mm256_packs_epi32(tmp2, tmp3); +} + +static INLINE void butterfly_self(__m256i *x0, __m256i *x1, const __m256i *c0, + const __m256i *c1) { + __m256i tmp0, tmp1, tmp2, tmp3, u0, u1; + const __m256i rounding = _mm256_set1_epi32(DCT_CONST_ROUNDING); + + u0 = _mm256_unpacklo_epi16(*x0, *x1); + u1 = _mm256_unpackhi_epi16(*x0, *x1); + BUTTERFLY_PAIR(u0, u1, *c0, *c1); + *x0 = _mm256_packs_epi32(tmp0, tmp1); + *x1 = _mm256_packs_epi32(tmp2, tmp3); +} + +// For each 16x32 block __m256i in[32], +// Input with index, 2, 6, 10, 14, 18, 22, 26, 30 +// output pixels: 8-15 in __m256i in[32] +static void idct32_full_16x32_quarter_2(const __m256i *in /*in[32]*/, + __m256i *out /*out[16]*/) { + __m256i u8, u9, u10, u11, u12, u13, u14, u15; // stp2_ + __m256i v8, v9, v10, v11, v12, v13, v14, v15; // stp1_ + + { + const __m256i stg2_0 = pair256_set_epi16(cospi_30_64, -cospi_2_64); + const __m256i stg2_1 = pair256_set_epi16(cospi_2_64, cospi_30_64); + const __m256i stg2_2 = pair256_set_epi16(cospi_14_64, -cospi_18_64); + const __m256i stg2_3 = pair256_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 = _mm256_add_epi16(u8, u9); + v9 = _mm256_sub_epi16(u8, u9); + v14 = _mm256_sub_epi16(u15, u14); + v15 = _mm256_add_epi16(u15, u14); + + { + const __m256i stg2_4 = pair256_set_epi16(cospi_22_64, -cospi_10_64); + const __m256i stg2_5 = pair256_set_epi16(cospi_10_64, cospi_22_64); + const __m256i stg2_6 = pair256_set_epi16(cospi_6_64, -cospi_26_64); + const __m256i stg2_7 = pair256_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 = _mm256_sub_epi16(u11, u10); + v11 = _mm256_add_epi16(u11, u10); + v12 = _mm256_add_epi16(u12, u13); + v13 = _mm256_sub_epi16(u12, u13); + + { + const __m256i stg4_4 = pair256_set_epi16(-cospi_8_64, cospi_24_64); + const __m256i stg4_5 = pair256_set_epi16(cospi_24_64, cospi_8_64); + const __m256i stg4_6 = pair256_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] = _mm256_add_epi16(v8, v11); + out[1] = _mm256_add_epi16(v9, v10); + out[6] = _mm256_add_epi16(v14, v13); + out[7] = _mm256_add_epi16(v15, v12); + + out[2] = _mm256_sub_epi16(v9, v10); + out[3] = _mm256_sub_epi16(v8, v11); + out[4] = _mm256_sub_epi16(v15, v12); + out[5] = _mm256_sub_epi16(v14, v13); + + { + const __m256i stg4_0 = pair256_set_epi16(cospi_16_64, cospi_16_64); + const __m256i stg6_0 = pair256_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 __m256i in[32], +// Input with index, 0, 4, 8, 12, 16, 20, 24, 28 +// output pixels: 0-7 in __m256i in[32] +static void idct32_full_16x32_quarter_1(const __m256i *in /*in[32]*/, + __m256i *out /*out[8]*/) { + __m256i u0, u1, u2, u3, u4, u5, u6, u7; // stp1_ + __m256i v0, v1, v2, v3, v4, v5, v6, v7; // stp2_ + + { + const __m256i stg3_0 = pair256_set_epi16(cospi_28_64, -cospi_4_64); + const __m256i stg3_1 = pair256_set_epi16(cospi_4_64, cospi_28_64); + const __m256i stg3_2 = pair256_set_epi16(cospi_12_64, -cospi_20_64); + const __m256i stg3_3 = pair256_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 = _mm256_add_epi16(u4, u5); + v5 = _mm256_sub_epi16(u4, u5); + v6 = _mm256_sub_epi16(u7, u6); + v7 = _mm256_add_epi16(u7, u6); + + { + const __m256i stg4_0 = pair256_set_epi16(cospi_16_64, cospi_16_64); + const __m256i stg4_1 = pair256_set_epi16(cospi_16_64, -cospi_16_64); + const __m256i stg4_2 = pair256_set_epi16(cospi_24_64, -cospi_8_64); + const __m256i stg4_3 = pair256_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 = _mm256_add_epi16(u0, u3); + v1 = _mm256_add_epi16(u1, u2); + v2 = _mm256_sub_epi16(u1, u2); + v3 = _mm256_sub_epi16(u0, u3); + + out[0] = _mm256_add_epi16(v0, v7); + out[1] = _mm256_add_epi16(v1, v6); + out[2] = _mm256_add_epi16(v2, v5); + out[3] = _mm256_add_epi16(v3, v4); + out[4] = _mm256_sub_epi16(v3, v4); + out[5] = _mm256_sub_epi16(v2, v5); + out[6] = _mm256_sub_epi16(v1, v6); + out[7] = _mm256_sub_epi16(v0, v7); +} + +// For each 8x32 block __m256i 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 __m256i in[32] +// We avoid hide an offset, 16, inside this function. So we output 0-15 into +// array out[16] +static void idct32_full_16x32_quarter_3_4(const __m256i *in /*in[32]*/, + __m256i *out /*out[16]*/) { + __m256i v16, v17, v18, v19, v20, v21, v22, v23; + __m256i v24, v25, v26, v27, v28, v29, v30, v31; + __m256i u16, u17, u18, u19, u20, u21, u22, u23; + __m256i u24, u25, u26, u27, u28, u29, u30, u31; + + { + const __m256i stg1_0 = pair256_set_epi16(cospi_31_64, -cospi_1_64); + const __m256i stg1_1 = pair256_set_epi16(cospi_1_64, cospi_31_64); + const __m256i stg1_2 = pair256_set_epi16(cospi_15_64, -cospi_17_64); + const __m256i stg1_3 = pair256_set_epi16(cospi_17_64, cospi_15_64); + const __m256i stg1_4 = pair256_set_epi16(cospi_23_64, -cospi_9_64); + const __m256i stg1_5 = pair256_set_epi16(cospi_9_64, cospi_23_64); + const __m256i stg1_6 = pair256_set_epi16(cospi_7_64, -cospi_25_64); + const __m256i stg1_7 = pair256_set_epi16(cospi_25_64, cospi_7_64); + const __m256i stg1_8 = pair256_set_epi16(cospi_27_64, -cospi_5_64); + const __m256i stg1_9 = pair256_set_epi16(cospi_5_64, cospi_27_64); + const __m256i stg1_10 = pair256_set_epi16(cospi_11_64, -cospi_21_64); + const __m256i stg1_11 = pair256_set_epi16(cospi_21_64, cospi_11_64); + const __m256i stg1_12 = pair256_set_epi16(cospi_19_64, -cospi_13_64); + const __m256i stg1_13 = pair256_set_epi16(cospi_13_64, cospi_19_64); + const __m256i stg1_14 = pair256_set_epi16(cospi_3_64, -cospi_29_64); + const __m256i stg1_15 = pair256_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 = _mm256_add_epi16(u16, u17); + v17 = _mm256_sub_epi16(u16, u17); + v18 = _mm256_sub_epi16(u19, u18); + v19 = _mm256_add_epi16(u19, u18); + + v20 = _mm256_add_epi16(u20, u21); + v21 = _mm256_sub_epi16(u20, u21); + v22 = _mm256_sub_epi16(u23, u22); + v23 = _mm256_add_epi16(u23, u22); + + v24 = _mm256_add_epi16(u24, u25); + v25 = _mm256_sub_epi16(u24, u25); + v26 = _mm256_sub_epi16(u27, u26); + v27 = _mm256_add_epi16(u27, u26); + + v28 = _mm256_add_epi16(u28, u29); + v29 = _mm256_sub_epi16(u28, u29); + v30 = _mm256_sub_epi16(u31, u30); + v31 = _mm256_add_epi16(u31, u30); + + { + const __m256i stg3_4 = pair256_set_epi16(-cospi_4_64, cospi_28_64); + const __m256i stg3_5 = pair256_set_epi16(cospi_28_64, cospi_4_64); + const __m256i stg3_6 = pair256_set_epi16(-cospi_28_64, -cospi_4_64); + const __m256i stg3_8 = pair256_set_epi16(-cospi_20_64, cospi_12_64); + const __m256i stg3_9 = pair256_set_epi16(cospi_12_64, cospi_20_64); + const __m256i stg3_10 = pair256_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 = _mm256_add_epi16(v16, v19); + u17 = _mm256_add_epi16(v17, v18); + u18 = _mm256_sub_epi16(v17, v18); + u19 = _mm256_sub_epi16(v16, v19); + u20 = _mm256_sub_epi16(v23, v20); + u21 = _mm256_sub_epi16(v22, v21); + u22 = _mm256_add_epi16(v22, v21); + u23 = _mm256_add_epi16(v23, v20); + + u24 = _mm256_add_epi16(v24, v27); + u25 = _mm256_add_epi16(v25, v26); + u26 = _mm256_sub_epi16(v25, v26); + u27 = _mm256_sub_epi16(v24, v27); + + u28 = _mm256_sub_epi16(v31, v28); + u29 = _mm256_sub_epi16(v30, v29); + u30 = _mm256_add_epi16(v29, v30); + u31 = _mm256_add_epi16(v28, v31); + + { + const __m256i stg4_4 = pair256_set_epi16(-cospi_8_64, cospi_24_64); + const __m256i stg4_5 = pair256_set_epi16(cospi_24_64, cospi_8_64); + const __m256i stg4_6 = pair256_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] = _mm256_add_epi16(u16, u23); + out[1] = _mm256_add_epi16(u17, u22); + out[2] = _mm256_add_epi16(u18, u21); + out[3] = _mm256_add_epi16(u19, u20); + out[4] = _mm256_sub_epi16(u19, u20); + out[5] = _mm256_sub_epi16(u18, u21); + out[6] = _mm256_sub_epi16(u17, u22); + out[7] = _mm256_sub_epi16(u16, u23); + + out[8] = _mm256_sub_epi16(u31, u24); + out[9] = _mm256_sub_epi16(u30, u25); + out[10] = _mm256_sub_epi16(u29, u26); + out[11] = _mm256_sub_epi16(u28, u27); + out[12] = _mm256_add_epi16(u27, u28); + out[13] = _mm256_add_epi16(u26, u29); + out[14] = _mm256_add_epi16(u25, u30); + out[15] = _mm256_add_epi16(u24, u31); + + { + const __m256i stg4_0 = pair256_set_epi16(cospi_16_64, cospi_16_64); + const __m256i stg6_0 = pair256_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_16x32_quarter_1_2(const __m256i *in /*in[32]*/, + __m256i *out /*out[32]*/) { + __m256i temp[16]; + idct32_full_16x32_quarter_1(in, temp); + idct32_full_16x32_quarter_2(in, &temp[8]); + add_sub_butterfly(temp, out, 16); +} + +static void idct32_16x32(const __m256i *in /*in[32]*/, + __m256i *out /*out[32]*/) { + __m256i temp[32]; + idct32_full_16x32_quarter_1_2(in, temp); + idct32_full_16x32_quarter_3_4(in, &temp[16]); + add_sub_butterfly(temp, out, 32); +} + +void aom_idct32x32_1024_add_avx2(const tran_low_t *input, uint8_t *dest, + int stride) { + __m256i col[64], in[32]; + int i; + + for (i = 0; i < 2; ++i) { + load_buffer_32x16(input, in); + input += 32 << 4; + + mm256_transpose_16x16(in, in); + mm256_transpose_16x16(&in[16], &in[16]); + idct32_16x32(in, col + (i << 5)); + } + + for (i = 0; i < 2; ++i) { + int j = i << 4; + mm256_transpose_16x16(col + j, in); + mm256_transpose_16x16(col + j + 32, &in[16]); + idct32_16x32(in, in); + store_buffer_16xN(in, stride, dest, 32); + dest += 16; + } +} + +// 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_16x32_135_quarter_1(const __m256i *in /*in[16]*/, + __m256i *out /*out[8]*/) { + __m256i u0, u1, u2, u3, u4, u5, u6, u7; + __m256i v0, v1, v2, v3, v4, v5, v6, v7; + + { + const __m256i stk4_0 = pair256_set_epi16(2 * cospi_16_64, 2 * cospi_16_64); + const __m256i stk4_2 = pair256_set_epi16(2 * cospi_24_64, 2 * cospi_24_64); + const __m256i stk4_3 = pair256_set_epi16(2 * cospi_8_64, 2 * cospi_8_64); + u0 = _mm256_mulhrs_epi16(in[0], stk4_0); + u2 = _mm256_mulhrs_epi16(in[8], stk4_2); + u3 = _mm256_mulhrs_epi16(in[8], stk4_3); + u1 = u0; + } + + v0 = _mm256_add_epi16(u0, u3); + v1 = _mm256_add_epi16(u1, u2); + v2 = _mm256_sub_epi16(u1, u2); + v3 = _mm256_sub_epi16(u0, u3); + + { + const __m256i stk3_0 = pair256_set_epi16(2 * cospi_28_64, 2 * cospi_28_64); + const __m256i stk3_1 = pair256_set_epi16(2 * cospi_4_64, 2 * cospi_4_64); + const __m256i stk3_2 = + pair256_set_epi16(-2 * cospi_20_64, -2 * cospi_20_64); + const __m256i stk3_3 = pair256_set_epi16(2 * cospi_12_64, 2 * cospi_12_64); + u4 = _mm256_mulhrs_epi16(in[4], stk3_0); + u7 = _mm256_mulhrs_epi16(in[4], stk3_1); + u5 = _mm256_mulhrs_epi16(in[12], stk3_2); + u6 = _mm256_mulhrs_epi16(in[12], stk3_3); + } + + v4 = _mm256_add_epi16(u4, u5); + v5 = _mm256_sub_epi16(u4, u5); + v6 = _mm256_sub_epi16(u7, u6); + v7 = _mm256_add_epi16(u7, u6); + + { + const __m256i stg4_0 = pair256_set_epi16(cospi_16_64, cospi_16_64); + const __m256i stg4_1 = pair256_set_epi16(cospi_16_64, -cospi_16_64); + butterfly(&v6, &v5, &stg4_1, &stg4_0, &v5, &v6); + } + + out[0] = _mm256_add_epi16(v0, v7); + out[1] = _mm256_add_epi16(v1, v6); + out[2] = _mm256_add_epi16(v2, v5); + out[3] = _mm256_add_epi16(v3, v4); + out[4] = _mm256_sub_epi16(v3, v4); + out[5] = _mm256_sub_epi16(v2, v5); + out[6] = _mm256_sub_epi16(v1, v6); + out[7] = _mm256_sub_epi16(v0, v7); +} + +static void idct32_16x32_135_quarter_2(const __m256i *in /*in[16]*/, + __m256i *out /*out[8]*/) { + __m256i u8, u9, u10, u11, u12, u13, u14, u15; + __m256i v8, v9, v10, v11, v12, v13, v14, v15; + + { + const __m256i stk2_0 = pair256_set_epi16(2 * cospi_30_64, 2 * cospi_30_64); + const __m256i stk2_1 = pair256_set_epi16(2 * cospi_2_64, 2 * cospi_2_64); + const __m256i stk2_2 = + pair256_set_epi16(-2 * cospi_18_64, -2 * cospi_18_64); + const __m256i stk2_3 = pair256_set_epi16(2 * cospi_14_64, 2 * cospi_14_64); + const __m256i stk2_4 = pair256_set_epi16(2 * cospi_22_64, 2 * cospi_22_64); + const __m256i stk2_5 = pair256_set_epi16(2 * cospi_10_64, 2 * cospi_10_64); + const __m256i stk2_6 = + pair256_set_epi16(-2 * cospi_26_64, -2 * cospi_26_64); + const __m256i stk2_7 = pair256_set_epi16(2 * cospi_6_64, 2 * cospi_6_64); + u8 = _mm256_mulhrs_epi16(in[2], stk2_0); + u15 = _mm256_mulhrs_epi16(in[2], stk2_1); + u9 = _mm256_mulhrs_epi16(in[14], stk2_2); + u14 = _mm256_mulhrs_epi16(in[14], stk2_3); + u10 = _mm256_mulhrs_epi16(in[10], stk2_4); + u13 = _mm256_mulhrs_epi16(in[10], stk2_5); + u11 = _mm256_mulhrs_epi16(in[6], stk2_6); + u12 = _mm256_mulhrs_epi16(in[6], stk2_7); + } + + v8 = _mm256_add_epi16(u8, u9); + v9 = _mm256_sub_epi16(u8, u9); + v10 = _mm256_sub_epi16(u11, u10); + v11 = _mm256_add_epi16(u11, u10); + v12 = _mm256_add_epi16(u12, u13); + v13 = _mm256_sub_epi16(u12, u13); + v14 = _mm256_sub_epi16(u15, u14); + v15 = _mm256_add_epi16(u15, u14); + + { + const __m256i stg4_4 = pair256_set_epi16(-cospi_8_64, cospi_24_64); + const __m256i stg4_5 = pair256_set_epi16(cospi_24_64, cospi_8_64); + const __m256i stg4_6 = pair256_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] = _mm256_add_epi16(v8, v11); + out[1] = _mm256_add_epi16(v9, v10); + out[2] = _mm256_sub_epi16(v9, v10); + out[3] = _mm256_sub_epi16(v8, v11); + out[4] = _mm256_sub_epi16(v15, v12); + out[5] = _mm256_sub_epi16(v14, v13); + out[6] = _mm256_add_epi16(v14, v13); + out[7] = _mm256_add_epi16(v15, v12); + + { + const __m256i stg4_0 = pair256_set_epi16(cospi_16_64, cospi_16_64); + const __m256i stg6_0 = pair256_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_16x32_quarter_1_2(const __m256i *in /*in[16]*/, + __m256i *out /*out[32]*/) { + __m256i temp[16]; + idct32_16x32_135_quarter_1(in, temp); + idct32_16x32_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_16x32_quarter_3_4(const __m256i *in /*in[16]*/, + __m256i *out /*out[32]*/) { + __m256i v16, v17, v18, v19, v20, v21, v22, v23; + __m256i v24, v25, v26, v27, v28, v29, v30, v31; + __m256i u16, u17, u18, u19, u20, u21, u22, u23; + __m256i u24, u25, u26, u27, u28, u29, u30, u31; + + { + const __m256i stk1_0 = pair256_set_epi16(2 * cospi_31_64, 2 * cospi_31_64); + const __m256i stk1_1 = pair256_set_epi16(2 * cospi_1_64, 2 * cospi_1_64); + const __m256i stk1_2 = + pair256_set_epi16(-2 * cospi_17_64, -2 * cospi_17_64); + const __m256i stk1_3 = pair256_set_epi16(2 * cospi_15_64, 2 * cospi_15_64); + + const __m256i stk1_4 = pair256_set_epi16(2 * cospi_23_64, 2 * cospi_23_64); + const __m256i stk1_5 = pair256_set_epi16(2 * cospi_9_64, 2 * cospi_9_64); + const __m256i stk1_6 = + pair256_set_epi16(-2 * cospi_25_64, -2 * cospi_25_64); + const __m256i stk1_7 = pair256_set_epi16(2 * cospi_7_64, 2 * cospi_7_64); + const __m256i stk1_8 = pair256_set_epi16(2 * cospi_27_64, 2 * cospi_27_64); + const __m256i stk1_9 = pair256_set_epi16(2 * cospi_5_64, 2 * cospi_5_64); + const __m256i stk1_10 = + pair256_set_epi16(-2 * cospi_21_64, -2 * cospi_21_64); + const __m256i stk1_11 = pair256_set_epi16(2 * cospi_11_64, 2 * cospi_11_64); + + const __m256i stk1_12 = pair256_set_epi16(2 * cospi_19_64, 2 * cospi_19_64); + const __m256i stk1_13 = pair256_set_epi16(2 * cospi_13_64, 2 * cospi_13_64); + const __m256i stk1_14 = + pair256_set_epi16(-2 * cospi_29_64, -2 * cospi_29_64); + const __m256i stk1_15 = pair256_set_epi16(2 * cospi_3_64, 2 * cospi_3_64); + u16 = _mm256_mulhrs_epi16(in[1], stk1_0); + u31 = _mm256_mulhrs_epi16(in[1], stk1_1); + u17 = _mm256_mulhrs_epi16(in[15], stk1_2); + u30 = _mm256_mulhrs_epi16(in[15], stk1_3); + + u18 = _mm256_mulhrs_epi16(in[9], stk1_4); + u29 = _mm256_mulhrs_epi16(in[9], stk1_5); + u19 = _mm256_mulhrs_epi16(in[7], stk1_6); + u28 = _mm256_mulhrs_epi16(in[7], stk1_7); + + u20 = _mm256_mulhrs_epi16(in[5], stk1_8); + u27 = _mm256_mulhrs_epi16(in[5], stk1_9); + u21 = _mm256_mulhrs_epi16(in[11], stk1_10); + u26 = _mm256_mulhrs_epi16(in[11], stk1_11); + + u22 = _mm256_mulhrs_epi16(in[13], stk1_12); + u25 = _mm256_mulhrs_epi16(in[13], stk1_13); + u23 = _mm256_mulhrs_epi16(in[3], stk1_14); + u24 = _mm256_mulhrs_epi16(in[3], stk1_15); + } + + v16 = _mm256_add_epi16(u16, u17); + v17 = _mm256_sub_epi16(u16, u17); + v18 = _mm256_sub_epi16(u19, u18); + v19 = _mm256_add_epi16(u19, u18); + + v20 = _mm256_add_epi16(u20, u21); + v21 = _mm256_sub_epi16(u20, u21); + v22 = _mm256_sub_epi16(u23, u22); + v23 = _mm256_add_epi16(u23, u22); + + v24 = _mm256_add_epi16(u24, u25); + v25 = _mm256_sub_epi16(u24, u25); + v26 = _mm256_sub_epi16(u27, u26); + v27 = _mm256_add_epi16(u27, u26); + + v28 = _mm256_add_epi16(u28, u29); + v29 = _mm256_sub_epi16(u28, u29); + v30 = _mm256_sub_epi16(u31, u30); + v31 = _mm256_add_epi16(u31, u30); + + { + const __m256i stg3_4 = pair256_set_epi16(-cospi_4_64, cospi_28_64); + const __m256i stg3_5 = pair256_set_epi16(cospi_28_64, cospi_4_64); + const __m256i stg3_6 = pair256_set_epi16(-cospi_28_64, -cospi_4_64); + const __m256i stg3_8 = pair256_set_epi16(-cospi_20_64, cospi_12_64); + const __m256i stg3_9 = pair256_set_epi16(cospi_12_64, cospi_20_64); + const __m256i stg3_10 = pair256_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 = _mm256_add_epi16(v16, v19); + u17 = _mm256_add_epi16(v17, v18); + u18 = _mm256_sub_epi16(v17, v18); + u19 = _mm256_sub_epi16(v16, v19); + u20 = _mm256_sub_epi16(v23, v20); + u21 = _mm256_sub_epi16(v22, v21); + u22 = _mm256_add_epi16(v22, v21); + u23 = _mm256_add_epi16(v23, v20); + + u24 = _mm256_add_epi16(v24, v27); + u25 = _mm256_add_epi16(v25, v26); + u26 = _mm256_sub_epi16(v25, v26); + u27 = _mm256_sub_epi16(v24, v27); + u28 = _mm256_sub_epi16(v31, v28); + u29 = _mm256_sub_epi16(v30, v29); + u30 = _mm256_add_epi16(v29, v30); + u31 = _mm256_add_epi16(v28, v31); + + { + const __m256i stg4_4 = pair256_set_epi16(-cospi_8_64, cospi_24_64); + const __m256i stg4_5 = pair256_set_epi16(cospi_24_64, cospi_8_64); + const __m256i stg4_6 = pair256_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] = _mm256_add_epi16(u16, u23); + out[1] = _mm256_add_epi16(u17, u22); + out[2] = _mm256_add_epi16(u18, u21); + out[3] = _mm256_add_epi16(u19, u20); + v20 = _mm256_sub_epi16(u19, u20); + v21 = _mm256_sub_epi16(u18, u21); + v22 = _mm256_sub_epi16(u17, u22); + v23 = _mm256_sub_epi16(u16, u23); + + v24 = _mm256_sub_epi16(u31, u24); + v25 = _mm256_sub_epi16(u30, u25); + v26 = _mm256_sub_epi16(u29, u26); + v27 = _mm256_sub_epi16(u28, u27); + out[12] = _mm256_add_epi16(u27, u28); + out[13] = _mm256_add_epi16(u26, u29); + out[14] = _mm256_add_epi16(u25, u30); + out[15] = _mm256_add_epi16(u24, u31); + + { + const __m256i stg4_0 = pair256_set_epi16(cospi_16_64, cospi_16_64); + const __m256i stg6_0 = pair256_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]); + } +} + +// 16x16 block input __m256i in[32], output 16x32 __m256i in[32] +static void idct32_16x32_135(__m256i *in /*in[32]*/) { + __m256i out[32]; + idct32_16x32_quarter_1_2(in, out); + idct32_16x32_quarter_3_4(in, &out[16]); + add_sub_butterfly(out, in, 32); +} + +static INLINE void load_buffer_from_32x32(const tran_low_t *coeff, __m256i *in, + int size) { + int i = 0; + while (i < size) { + load_coeff(coeff + (i << 5), &in[i]); + i += 1; + } +} + +static INLINE void zero_buffer(__m256i *in, int num) { + int i; + for (i = 0; i < num; ++i) { + in[i] = _mm256_setzero_si256(); + } +} + +// Only upper-left 16x16 has non-zero coeff +void aom_idct32x32_135_add_avx2(const tran_low_t *input, uint8_t *dest, + int stride) { + __m256i in[32]; + zero_buffer(in, 32); + load_buffer_from_32x32(input, in, 16); + mm256_transpose_16x16(in, in); + idct32_16x32_135(in); + + __m256i out[32]; + mm256_transpose_16x16(in, out); + idct32_16x32_135(out); + store_buffer_16xN(out, stride, dest, 32); + mm256_transpose_16x16(&in[16], in); + idct32_16x32_135(in); + store_buffer_16xN(in, stride, dest + 16, 32); +} + +static void idct32_34_first_half(const __m256i *in, __m256i *stp1) { + const __m256i stk2_0 = pair256_set_epi16(2 * cospi_30_64, 2 * cospi_30_64); + const __m256i stk2_1 = pair256_set_epi16(2 * cospi_2_64, 2 * cospi_2_64); + const __m256i stk2_6 = pair256_set_epi16(-2 * cospi_26_64, -2 * cospi_26_64); + const __m256i stk2_7 = pair256_set_epi16(2 * cospi_6_64, 2 * cospi_6_64); + + const __m256i stk3_0 = pair256_set_epi16(2 * cospi_28_64, 2 * cospi_28_64); + const __m256i stk3_1 = pair256_set_epi16(2 * cospi_4_64, 2 * cospi_4_64); + + const __m256i stg4_0 = pair256_set_epi16(cospi_16_64, cospi_16_64); + const __m256i stk4_0 = pair256_set_epi16(2 * cospi_16_64, 2 * cospi_16_64); + const __m256i stg4_1 = pair256_set_epi16(cospi_16_64, -cospi_16_64); + const __m256i stg4_4 = pair256_set_epi16(-cospi_8_64, cospi_24_64); + const __m256i stg4_5 = pair256_set_epi16(cospi_24_64, cospi_8_64); + const __m256i stg4_6 = pair256_set_epi16(-cospi_24_64, -cospi_8_64); + + const __m256i stg6_0 = pair256_set_epi16(-cospi_16_64, cospi_16_64); + __m256i u0, u1, u2, u3, u4, u5, u6, u7; + __m256i x0, x1, x4, x5, x6, x7; + __m256i v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15; + + // phase 1 + + // 0, 15 + u2 = _mm256_mulhrs_epi16(in[2], stk2_1); // stp2_15 + u3 = _mm256_mulhrs_epi16(in[6], stk2_7); // stp2_12 + v15 = _mm256_add_epi16(u2, u3); + // in[0], in[4] + x0 = _mm256_mulhrs_epi16(in[0], stk4_0); // stp1[0] + x7 = _mm256_mulhrs_epi16(in[4], stk3_1); // stp1[7] + v0 = _mm256_add_epi16(x0, x7); // stp2_0 + stp1[0] = _mm256_add_epi16(v0, v15); + stp1[15] = _mm256_sub_epi16(v0, v15); + + // in[2], in[6] + u0 = _mm256_mulhrs_epi16(in[2], stk2_0); // stp2_8 + u1 = _mm256_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 = _mm256_add_epi16(u0, u1); + v9 = _mm256_add_epi16(u4, u6); + v10 = _mm256_sub_epi16(u4, u6); + v11 = _mm256_sub_epi16(u0, u1); + v12 = _mm256_sub_epi16(u2, u3); + v13 = _mm256_sub_epi16(u5, u7); + v14 = _mm256_add_epi16(u5, u7); + + butterfly_self(&v10, &v13, &stg6_0, &stg4_0); + butterfly_self(&v11, &v12, &stg6_0, &stg4_0); + + // 1, 14 + x1 = _mm256_mulhrs_epi16(in[0], stk4_0); // stp1[1], stk4_1 = stk4_0 + // stp1[2] = stp1[0], stp1[3] = stp1[1] + x4 = _mm256_mulhrs_epi16(in[4], stk3_0); // stp1[4] + butterfly(&x7, &x4, &stg4_1, &stg4_0, &x5, &x6); + v1 = _mm256_add_epi16(x1, x6); // stp2_1 + v2 = _mm256_add_epi16(x0, x5); // stp2_2 + stp1[1] = _mm256_add_epi16(v1, v14); + stp1[14] = _mm256_sub_epi16(v1, v14); + + stp1[2] = _mm256_add_epi16(v2, v13); + stp1[13] = _mm256_sub_epi16(v2, v13); + + v3 = _mm256_add_epi16(x1, x4); // stp2_3 + v4 = _mm256_sub_epi16(x1, x4); // stp2_4 + + v5 = _mm256_sub_epi16(x0, x5); // stp2_5 + + v6 = _mm256_sub_epi16(x1, x6); // stp2_6 + v7 = _mm256_sub_epi16(x0, x7); // stp2_7 + stp1[3] = _mm256_add_epi16(v3, v12); + stp1[12] = _mm256_sub_epi16(v3, v12); + + stp1[6] = _mm256_add_epi16(v6, v9); + stp1[9] = _mm256_sub_epi16(v6, v9); + + stp1[7] = _mm256_add_epi16(v7, v8); + stp1[8] = _mm256_sub_epi16(v7, v8); + + stp1[4] = _mm256_add_epi16(v4, v11); + stp1[11] = _mm256_sub_epi16(v4, v11); + + stp1[5] = _mm256_add_epi16(v5, v10); + stp1[10] = _mm256_sub_epi16(v5, v10); +} + +static void idct32_34_second_half(const __m256i *in, __m256i *stp1) { + const __m256i stk1_0 = pair256_set_epi16(2 * cospi_31_64, 2 * cospi_31_64); + const __m256i stk1_1 = pair256_set_epi16(2 * cospi_1_64, 2 * cospi_1_64); + const __m256i stk1_6 = pair256_set_epi16(-2 * cospi_25_64, -2 * cospi_25_64); + const __m256i stk1_7 = pair256_set_epi16(2 * cospi_7_64, 2 * cospi_7_64); + const __m256i stk1_8 = pair256_set_epi16(2 * cospi_27_64, 2 * cospi_27_64); + const __m256i stk1_9 = pair256_set_epi16(2 * cospi_5_64, 2 * cospi_5_64); + const __m256i stk1_14 = pair256_set_epi16(-2 * cospi_29_64, -2 * cospi_29_64); + const __m256i stk1_15 = pair256_set_epi16(2 * cospi_3_64, 2 * cospi_3_64); + const __m256i stg3_4 = pair256_set_epi16(-cospi_4_64, cospi_28_64); + const __m256i stg3_5 = pair256_set_epi16(cospi_28_64, cospi_4_64); + const __m256i stg3_6 = pair256_set_epi16(-cospi_28_64, -cospi_4_64); + const __m256i stg3_8 = pair256_set_epi16(-cospi_20_64, cospi_12_64); + const __m256i stg3_9 = pair256_set_epi16(cospi_12_64, cospi_20_64); + const __m256i stg3_10 = pair256_set_epi16(-cospi_12_64, -cospi_20_64); + + const __m256i stg4_0 = pair256_set_epi16(cospi_16_64, cospi_16_64); + const __m256i stg4_4 = pair256_set_epi16(-cospi_8_64, cospi_24_64); + const __m256i stg4_5 = pair256_set_epi16(cospi_24_64, cospi_8_64); + const __m256i stg4_6 = pair256_set_epi16(-cospi_24_64, -cospi_8_64); + + const __m256i stg6_0 = pair256_set_epi16(-cospi_16_64, cospi_16_64); + __m256i v16, v17, v18, v19, v20, v21, v22, v23; + __m256i v24, v25, v26, v27, v28, v29, v30, v31; + __m256i u16, u17, u18, u19, u20, u21, u22, u23; + __m256i u24, u25, u26, u27, u28, u29, u30, u31; + + v16 = _mm256_mulhrs_epi16(in[1], stk1_0); + v31 = _mm256_mulhrs_epi16(in[1], stk1_1); + + v19 = _mm256_mulhrs_epi16(in[7], stk1_6); + v28 = _mm256_mulhrs_epi16(in[7], stk1_7); + + v20 = _mm256_mulhrs_epi16(in[5], stk1_8); + v27 = _mm256_mulhrs_epi16(in[5], stk1_9); + + v23 = _mm256_mulhrs_epi16(in[3], stk1_14); + v24 = _mm256_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 = _mm256_add_epi16(v16, v19); + u17 = _mm256_add_epi16(v17, v18); + u18 = _mm256_sub_epi16(v17, v18); + u19 = _mm256_sub_epi16(v16, v19); + u20 = _mm256_sub_epi16(v23, v20); + u21 = _mm256_sub_epi16(v22, v21); + u22 = _mm256_add_epi16(v22, v21); + u23 = _mm256_add_epi16(v23, v20); + u24 = _mm256_add_epi16(v24, v27); + u27 = _mm256_sub_epi16(v24, v27); + u25 = _mm256_add_epi16(v25, v26); + u26 = _mm256_sub_epi16(v25, v26); + u28 = _mm256_sub_epi16(v31, v28); + u31 = _mm256_add_epi16(v28, v31); + u29 = _mm256_sub_epi16(v30, v29); + u30 = _mm256_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[0] = _mm256_add_epi16(u16, u23); + stp1[7] = _mm256_sub_epi16(u16, u23); + + stp1[1] = _mm256_add_epi16(u17, u22); + stp1[6] = _mm256_sub_epi16(u17, u22); + + stp1[2] = _mm256_add_epi16(u18, u21); + stp1[5] = _mm256_sub_epi16(u18, u21); + + stp1[3] = _mm256_add_epi16(u19, u20); + stp1[4] = _mm256_sub_epi16(u19, u20); + + stp1[8] = _mm256_sub_epi16(u31, u24); + stp1[15] = _mm256_add_epi16(u24, u31); + + stp1[9] = _mm256_sub_epi16(u30, u25); + stp1[14] = _mm256_add_epi16(u25, u30); + + stp1[10] = _mm256_sub_epi16(u29, u26); + stp1[13] = _mm256_add_epi16(u26, u29); + + stp1[11] = _mm256_sub_epi16(u28, u27); + stp1[12] = _mm256_add_epi16(u27, u28); + + butterfly_self(&stp1[4], &stp1[11], &stg6_0, &stg4_0); + butterfly_self(&stp1[5], &stp1[10], &stg6_0, &stg4_0); + butterfly_self(&stp1[6], &stp1[9], &stg6_0, &stg4_0); + butterfly_self(&stp1[7], &stp1[8], &stg6_0, &stg4_0); +} + +// 16x16 block input __m256i in[32], output 16x32 __m256i in[32] +static void idct32_16x32_34(__m256i *in /*in[32]*/) { + __m256i out[32]; + idct32_34_first_half(in, out); + idct32_34_second_half(in, &out[16]); + add_sub_butterfly(out, in, 32); +} + +// Only upper-left 8x8 has non-zero coeff +void aom_idct32x32_34_add_avx2(const tran_low_t *input, uint8_t *dest, + int stride) { + __m256i in[32]; + zero_buffer(in, 32); + load_buffer_from_32x32(input, in, 8); + mm256_transpose_16x16(in, in); + idct32_16x32_34(in); + + __m256i out[32]; + mm256_transpose_16x16(in, out); + idct32_16x32_34(out); + store_buffer_16xN(out, stride, dest, 32); + mm256_transpose_16x16(&in[16], in); + idct32_16x32_34(in); + store_buffer_16xN(in, stride, dest + 16, 32); +} diff --git a/third_party/aom/aom_dsp/x86/inv_txfm_common_avx2.h b/third_party/aom/aom_dsp/x86/inv_txfm_common_avx2.h new file mode 100644 index 000000000..4238e651b --- /dev/null +++ b/third_party/aom/aom_dsp/x86/inv_txfm_common_avx2.h @@ -0,0 +1,80 @@ +/* + * 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. + */ + +#ifndef AOM_DSP_X86_INV_TXFM_COMMON_AVX2_H +#define AOM_DSP_X86_INV_TXFM_COMMON_AVX2_H + +#include + +#include "aom_dsp/txfm_common.h" +#include "aom_dsp/x86/txfm_common_avx2.h" + +static INLINE void load_coeff(const tran_low_t *coeff, __m256i *in) { +#if CONFIG_HIGHBITDEPTH + *in = _mm256_setr_epi16( + (int16_t)coeff[0], (int16_t)coeff[1], (int16_t)coeff[2], + (int16_t)coeff[3], (int16_t)coeff[4], (int16_t)coeff[5], + (int16_t)coeff[6], (int16_t)coeff[7], (int16_t)coeff[8], + (int16_t)coeff[9], (int16_t)coeff[10], (int16_t)coeff[11], + (int16_t)coeff[12], (int16_t)coeff[13], (int16_t)coeff[14], + (int16_t)coeff[15]); +#else + *in = _mm256_loadu_si256((const __m256i *)coeff); +#endif +} + +static INLINE void load_buffer_16x16(const tran_low_t *coeff, __m256i *in) { + int i = 0; + while (i < 16) { + load_coeff(coeff + (i << 4), &in[i]); + i += 1; + } +} + +static INLINE void recon_and_store(const __m256i *res, uint8_t *output) { + const __m128i zero = _mm_setzero_si128(); + __m128i x = _mm_loadu_si128((__m128i const *)output); + __m128i p0 = _mm_unpacklo_epi8(x, zero); + __m128i p1 = _mm_unpackhi_epi8(x, zero); + + p0 = _mm_add_epi16(p0, _mm256_castsi256_si128(*res)); + p1 = _mm_add_epi16(p1, _mm256_extractf128_si256(*res, 1)); + x = _mm_packus_epi16(p0, p1); + _mm_storeu_si128((__m128i *)output, x); +} + +#define IDCT_ROUNDING_POS (6) +static INLINE void store_buffer_16xN(__m256i *in, const int stride, + uint8_t *output, int num) { + const __m256i rounding = _mm256_set1_epi16(1 << (IDCT_ROUNDING_POS - 1)); + int i = 0; + + while (i < num) { + in[i] = _mm256_adds_epi16(in[i], rounding); + in[i] = _mm256_srai_epi16(in[i], IDCT_ROUNDING_POS); + recon_and_store(&in[i], output + i * stride); + i += 1; + } +} + +static INLINE void unpack_butter_fly(const __m256i *a0, const __m256i *a1, + const __m256i *c0, const __m256i *c1, + __m256i *b0, __m256i *b1) { + __m256i x0, x1; + x0 = _mm256_unpacklo_epi16(*a0, *a1); + x1 = _mm256_unpackhi_epi16(*a0, *a1); + *b0 = butter_fly(&x0, &x1, c0); + *b1 = butter_fly(&x0, &x1, c1); +} + +void av1_idct16_avx2(__m256i *in); + +#endif // AOM_DSP_X86_INV_TXFM_COMMON_AVX2_H diff --git a/third_party/aom/aom_dsp/x86/inv_txfm_sse2.c b/third_party/aom/aom_dsp/x86/inv_txfm_sse2.c index 5795a1845..be200df4c 100644 --- a/third_party/aom/aom_dsp/x86/inv_txfm_sse2.c +++ b/third_party/aom/aom_dsp/x86/inv_txfm_sse2.c @@ -3628,4 +3628,107 @@ void aom_highbd_idct4x4_16_add_sse2(const tran_low_t *input, uint8_t *dest8, } } +void aom_highbd_idct8x8_10_add_sse2(const tran_low_t *input, uint8_t *dest8, + int stride, int bd) { + tran_low_t out[8 * 8] = { 0 }; + tran_low_t *outptr = out; + int i, j, test; + __m128i inptr[8]; + __m128i min_input, max_input, temp1, temp2, sign_bits; + uint16_t *dest = CONVERT_TO_SHORTPTR(dest8); + const __m128i zero = _mm_set1_epi16(0); + const __m128i sixteen = _mm_set1_epi16(16); + const __m128i max = _mm_set1_epi16(6201); + const __m128i min = _mm_set1_epi16(-6201); + int optimised_cols = 0; + + // Load input into __m128i & pack to 16 bits + for (i = 0; i < 8; i++) { + temp1 = _mm_loadu_si128((const __m128i *)(input + 8 * i)); + temp2 = _mm_loadu_si128((const __m128i *)(input + 8 * i + 4)); + inptr[i] = _mm_packs_epi32(temp1, temp2); + } + + // Find the min & max for the row transform + // only first 4 row has non-zero coefs + max_input = _mm_max_epi16(inptr[0], inptr[1]); + min_input = _mm_min_epi16(inptr[0], inptr[1]); + for (i = 2; i < 4; i++) { + max_input = _mm_max_epi16(max_input, inptr[i]); + min_input = _mm_min_epi16(min_input, inptr[i]); + } + max_input = _mm_cmpgt_epi16(max_input, max); + min_input = _mm_cmplt_epi16(min_input, min); + temp1 = _mm_or_si128(max_input, min_input); + test = _mm_movemask_epi8(temp1); + + if (!test) { + // Do the row transform + aom_idct8_sse2(inptr); + + // Find the min & max for the column transform + // N.B. Only first 4 cols contain non-zero coeffs + max_input = _mm_max_epi16(inptr[0], inptr[1]); + min_input = _mm_min_epi16(inptr[0], inptr[1]); + for (i = 2; i < 8; i++) { + max_input = _mm_max_epi16(max_input, inptr[i]); + min_input = _mm_min_epi16(min_input, inptr[i]); + } + max_input = _mm_cmpgt_epi16(max_input, max); + min_input = _mm_cmplt_epi16(min_input, min); + temp1 = _mm_or_si128(max_input, min_input); + test = _mm_movemask_epi8(temp1); + + if (test) { + // Use fact only first 4 rows contain non-zero coeffs + array_transpose_4X8(inptr, inptr); + for (i = 0; i < 4; i++) { + sign_bits = _mm_cmplt_epi16(inptr[i], zero); + temp1 = _mm_unpackhi_epi16(inptr[i], sign_bits); + temp2 = _mm_unpacklo_epi16(inptr[i], sign_bits); + _mm_storeu_si128((__m128i *)(outptr + 4 * (2 * i + 1)), temp1); + _mm_storeu_si128((__m128i *)(outptr + 4 * (2 * i)), temp2); + } + } 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_idct8_c(input, outptr, bd); + input += 8; + outptr += 8; + } + } + + if (optimised_cols) { + aom_idct8_sse2(inptr); + + // Final round & shift and Reconstruction and Store + { + __m128i d[8]; + for (i = 0; i < 8; i++) { + inptr[i] = _mm_add_epi16(inptr[i], sixteen); + d[i] = _mm_loadu_si128((const __m128i *)(dest + stride * i)); + inptr[i] = _mm_srai_epi16(inptr[i], 5); + d[i] = clamp_high_sse2(_mm_adds_epi16(d[i], inptr[i]), bd); + // Store + _mm_storeu_si128((__m128i *)(dest + stride * i), d[i]); + } + } + } else { + // Run the un-optimised column transform + tran_low_t temp_in[8], temp_out[8]; + for (i = 0; i < 8; ++i) { + for (j = 0; j < 8; ++j) temp_in[j] = out[j * 8 + i]; + aom_highbd_idct8_c(temp_in, temp_out, bd); + for (j = 0; j < 8; ++j) { + dest[j * stride + i] = highbd_clip_pixel_add( + dest[j * stride + i], ROUND_POWER_OF_TWO(temp_out[j], 5), bd); + } + } + } +} + #endif // CONFIG_HIGHBITDEPTH 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 index 5166e9e0a..9d16a3e84 100644 --- a/third_party/aom/aom_dsp/x86/masked_sad_intrin_ssse3.c +++ b/third_party/aom/aom_dsp/x86/masked_sad_intrin_ssse3.c @@ -1,5 +1,5 @@ /* - * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * 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 @@ -9,49 +9,70 @@ * PATENTS file, you can obtain it at www.aomedia.org/license/patent. */ -#include -#include +#include #include -#include "aom_ports/mem.h" #include "./aom_config.h" +#include "./aom_dsp_rtcd.h" +#include "aom_dsp/blend.h" #include "aom/aom_integer.h" +#include "aom_dsp/x86/synonyms.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, +// For width a multiple of 16 +static INLINE unsigned int masked_sad_ssse3(const uint8_t *src_ptr, + int src_stride, + 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); + const uint8_t *src_ptr, int src_stride, 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); \ + const uint8_t *src_ptr, int src_stride, 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 *second_pred, const uint8_t *msk, int msk_stride, \ + int invert_mask) { \ + if (!invert_mask) \ + return masked_sad_ssse3(src, src_stride, ref, ref_stride, second_pred, \ + m, msk, msk_stride, m, n); \ + else \ + return masked_sad_ssse3(src, src_stride, second_pred, m, ref, \ + ref_stride, msk, msk_stride, m, n); \ + } + +#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 *second_pred, const uint8_t *msk, int msk_stride, \ + int invert_mask) { \ + if (!invert_mask) \ + return masked_sad8xh_ssse3(src, src_stride, ref, ref_stride, \ + second_pred, 8, msk, msk_stride, n); \ + else \ + return masked_sad8xh_ssse3(src, src_stride, second_pred, 8, ref, \ + ref_stride, msk, msk_stride, n); \ + } + +#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 *second_pred, const uint8_t *msk, int msk_stride, \ + int invert_mask) { \ + if (!invert_mask) \ + return masked_sad4xh_ssse3(src, src_stride, ref, ref_stride, \ + second_pred, 4, msk, msk_stride, n); \ + else \ + return masked_sad4xh_ssse3(src, src_stride, second_pred, 4, ref, \ + ref_stride, msk, msk_stride, n); \ } #if CONFIG_EXT_PARTITION @@ -67,165 +88,181 @@ 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, +static INLINE unsigned int masked_sad_ssse3(const uint8_t *src_ptr, + int src_stride, + 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; + int x, y; __m128i res = _mm_setzero_si128(); - __m128i one = _mm_set1_epi16(1); - // For each row + const __m128i mask_max = _mm_set1_epi8((1 << AOM_BLEND_A64_ROUND_BITS)); + 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)); + const __m128i src = _mm_loadu_si128((const __m128i *)&src_ptr[x]); + const __m128i a = _mm_loadu_si128((const __m128i *)&a_ptr[x]); + const __m128i b = _mm_loadu_si128((const __m128i *)&b_ptr[x]); + const __m128i m = _mm_loadu_si128((const __m128i *)&m_ptr[x]); + const __m128i m_inv = _mm_sub_epi8(mask_max, m); + + // Calculate 16 predicted pixels. + // Note that the maximum value of any entry of 'pred_l' or 'pred_r' + // is 64 * 255, so we have plenty of space to add rounding constants. + const __m128i data_l = _mm_unpacklo_epi8(a, b); + const __m128i mask_l = _mm_unpacklo_epi8(m, m_inv); + __m128i pred_l = _mm_maddubs_epi16(data_l, mask_l); + pred_l = xx_roundn_epu16(pred_l, AOM_BLEND_A64_ROUND_BITS); + + const __m128i data_r = _mm_unpackhi_epi8(a, b); + const __m128i mask_r = _mm_unpackhi_epi8(m, m_inv); + __m128i pred_r = _mm_maddubs_epi16(data_r, mask_r); + pred_r = xx_roundn_epu16(pred_r, AOM_BLEND_A64_ROUND_BITS); + + const __m128i pred = _mm_packus_epi16(pred_l, pred_r); + res = _mm_add_epi32(res, _mm_sad_epu8(pred, src)); } - // Move onto the next row + + src_ptr += src_stride; 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; + // At this point, we have two 32-bit partial SADs in lanes 0 and 2 of 'res'. + int32_t sad = + _mm_cvtsi128_si32(res) + _mm_cvtsi128_si32(_mm_srli_si128(res, 8)); + return (sad + 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) { + const uint8_t *src_ptr, int src_stride, 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)); + const __m128i mask_max = _mm_set1_epi8((1 << AOM_BLEND_A64_ROUND_BITS)); - // Move onto the next rows + for (y = 0; y < height; y += 2) { + const __m128i src = _mm_unpacklo_epi64( + _mm_loadl_epi64((const __m128i *)src_ptr), + _mm_loadl_epi64((const __m128i *)&src_ptr[src_stride])); + const __m128i a0 = _mm_loadl_epi64((const __m128i *)a_ptr); + const __m128i a1 = _mm_loadl_epi64((const __m128i *)&a_ptr[a_stride]); + const __m128i b0 = _mm_loadl_epi64((const __m128i *)b_ptr); + const __m128i b1 = _mm_loadl_epi64((const __m128i *)&b_ptr[b_stride]); + const __m128i m = + _mm_unpacklo_epi64(_mm_loadl_epi64((const __m128i *)m_ptr), + _mm_loadl_epi64((const __m128i *)&m_ptr[m_stride])); + const __m128i m_inv = _mm_sub_epi8(mask_max, m); + + const __m128i data_l = _mm_unpacklo_epi8(a0, b0); + const __m128i mask_l = _mm_unpacklo_epi8(m, m_inv); + __m128i pred_l = _mm_maddubs_epi16(data_l, mask_l); + pred_l = xx_roundn_epu16(pred_l, AOM_BLEND_A64_ROUND_BITS); + + const __m128i data_r = _mm_unpacklo_epi8(a1, b1); + const __m128i mask_r = _mm_unpackhi_epi8(m, m_inv); + __m128i pred_r = _mm_maddubs_epi16(data_r, mask_r); + pred_r = xx_roundn_epu16(pred_r, AOM_BLEND_A64_ROUND_BITS); + + const __m128i pred = _mm_packus_epi16(pred_l, pred_r); + res = _mm_add_epi32(res, _mm_sad_epu8(pred, src)); + + src_ptr += src_stride * 2; 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; + int32_t sad = + _mm_cvtsi128_si32(res) + _mm_cvtsi128_si32(_mm_srli_si128(res, 8)); + return (sad + 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) { + const uint8_t *src_ptr, int src_stride, 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; + const __m128i mask_max = _mm_set1_epi8((1 << AOM_BLEND_A64_ROUND_BITS)); + + for (y = 0; y < height; y += 2) { + // Load two rows at a time, this seems to be a bit faster + // than four rows at a time in this case. + const __m128i src = _mm_unpacklo_epi32( + _mm_cvtsi32_si128(*(uint32_t *)src_ptr), + _mm_cvtsi32_si128(*(uint32_t *)&src_ptr[src_stride])); + const __m128i a = + _mm_unpacklo_epi32(_mm_cvtsi32_si128(*(uint32_t *)a_ptr), + _mm_cvtsi32_si128(*(uint32_t *)&a_ptr[a_stride])); + const __m128i b = + _mm_unpacklo_epi32(_mm_cvtsi32_si128(*(uint32_t *)b_ptr), + _mm_cvtsi32_si128(*(uint32_t *)&b_ptr[b_stride])); + const __m128i m = + _mm_unpacklo_epi32(_mm_cvtsi32_si128(*(uint32_t *)m_ptr), + _mm_cvtsi32_si128(*(uint32_t *)&m_ptr[m_stride])); + const __m128i m_inv = _mm_sub_epi8(mask_max, m); + + const __m128i data = _mm_unpacklo_epi8(a, b); + const __m128i mask = _mm_unpacklo_epi8(m, m_inv); + __m128i pred_16bit = _mm_maddubs_epi16(data, mask); + pred_16bit = xx_roundn_epu16(pred_16bit, AOM_BLEND_A64_ROUND_BITS); + + const __m128i pred = _mm_packus_epi16(pred_16bit, _mm_setzero_si128()); + res = _mm_add_epi32(res, _mm_sad_epu8(pred, src)); + + src_ptr += src_stride * 2; + a_ptr += a_stride * 2; + b_ptr += b_stride * 2; + m_ptr += m_stride * 2; } - // 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; + // At this point, the SAD is stored in lane 0 of 'res' + int32_t sad = _mm_cvtsi128_si32(res); + return (sad + 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); -} - +// For width a multiple of 8 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); + const uint8_t *src8, int src_stride, const uint8_t *a8, int a_stride, + const uint8_t *b8, 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); + const uint8_t *src8, int src_stride, const uint8_t *a8, int a_stride, + const uint8_t *b8, 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); \ + const uint8_t *src8, int src_stride, const uint8_t *ref8, \ + int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, \ + int msk_stride, int invert_mask) { \ + if (!invert_mask) \ + return highbd_masked_sad_ssse3(src8, src_stride, ref8, ref_stride, \ + second_pred8, m, msk, msk_stride, m, n); \ + else \ + return highbd_masked_sad_ssse3(src8, src_stride, second_pred8, m, ref8, \ + ref_stride, msk, msk_stride, m, n); \ + } + +#define HIGHBD_MASKSAD4XN_SSSE3(n) \ + unsigned int aom_highbd_masked_sad4x##n##_ssse3( \ + const uint8_t *src8, int src_stride, const uint8_t *ref8, \ + int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, \ + int msk_stride, int invert_mask) { \ + if (!invert_mask) \ + return highbd_masked_sad4xh_ssse3(src8, src_stride, ref8, ref_stride, \ + second_pred8, 4, msk, msk_stride, n); \ + else \ + return highbd_masked_sad4xh_ssse3(src8, src_stride, second_pred8, 4, \ + ref8, ref_stride, msk, msk_stride, n); \ } #if CONFIG_EXT_PARTITION @@ -244,91 +281,124 @@ 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); + const uint8_t *src8, int src_stride, const uint8_t *a8, int a_stride, + const uint8_t *b8, int b_stride, const uint8_t *m_ptr, int m_stride, + int width, int height) { + const uint16_t *src_ptr = CONVERT_TO_SHORTPTR(src8); + const uint16_t *a_ptr = CONVERT_TO_SHORTPTR(a8); + const uint16_t *b_ptr = CONVERT_TO_SHORTPTR(b8); + int x, y; __m128i res = _mm_setzero_si128(); - // For each row + const __m128i mask_max = _mm_set1_epi16((1 << AOM_BLEND_A64_ROUND_BITS)); + const __m128i round_const = + _mm_set1_epi32((1 << AOM_BLEND_A64_ROUND_BITS) >> 1); + const __m128i one = _mm_set1_epi16(1); + 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)); + const __m128i src = _mm_loadu_si128((const __m128i *)&src_ptr[x]); + const __m128i a = _mm_loadu_si128((const __m128i *)&a_ptr[x]); + const __m128i b = _mm_loadu_si128((const __m128i *)&b_ptr[x]); + // Zero-extend mask to 16 bits + const __m128i m = _mm_unpacklo_epi8( + _mm_loadl_epi64((const __m128i *)&m_ptr[x]), _mm_setzero_si128()); + const __m128i m_inv = _mm_sub_epi16(mask_max, m); + + const __m128i data_l = _mm_unpacklo_epi16(a, b); + const __m128i mask_l = _mm_unpacklo_epi16(m, m_inv); + __m128i pred_l = _mm_madd_epi16(data_l, mask_l); + pred_l = _mm_srai_epi32(_mm_add_epi32(pred_l, round_const), + AOM_BLEND_A64_ROUND_BITS); + + const __m128i data_r = _mm_unpackhi_epi16(a, b); + const __m128i mask_r = _mm_unpackhi_epi16(m, m_inv); + __m128i pred_r = _mm_madd_epi16(data_r, mask_r); + pred_r = _mm_srai_epi32(_mm_add_epi32(pred_r, round_const), + AOM_BLEND_A64_ROUND_BITS); + + // Note: the maximum value in pred_l/r is (2^bd)-1 < 2^15, + // so it is safe to do signed saturation here. + const __m128i pred = _mm_packs_epi32(pred_l, pred_r); + // There is no 16-bit SAD instruction, so we have to synthesize + // an 8-element SAD. We do this by storing 4 32-bit partial SADs, + // and accumulating them at the end + const __m128i diff = _mm_abs_epi16(_mm_sub_epi16(pred, src)); + res = _mm_add_epi32(res, _mm_madd_epi16(diff, one)); } - // Move onto the next row + + src_ptr += src_stride; 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; + // At this point, we have four 32-bit partial SADs stored in 'res'. + res = _mm_hadd_epi32(res, res); + res = _mm_hadd_epi32(res, res); + int sad = _mm_cvtsi128_si32(res); + return (sad + 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) { + const uint8_t *src8, int src_stride, const uint8_t *a8, int a_stride, + const uint8_t *b8, int b_stride, const uint8_t *m_ptr, int m_stride, + int height) { + const uint16_t *src_ptr = CONVERT_TO_SHORTPTR(src8); + const uint16_t *a_ptr = CONVERT_TO_SHORTPTR(a8); + const uint16_t *b_ptr = CONVERT_TO_SHORTPTR(b8); 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 + const __m128i mask_max = _mm_set1_epi16((1 << AOM_BLEND_A64_ROUND_BITS)); + const __m128i round_const = + _mm_set1_epi32((1 << AOM_BLEND_A64_ROUND_BITS) >> 1); + const __m128i one = _mm_set1_epi16(1); + 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 + const __m128i src = _mm_unpacklo_epi64( + _mm_loadl_epi64((const __m128i *)src_ptr), + _mm_loadl_epi64((const __m128i *)&src_ptr[src_stride])); + const __m128i a = + _mm_unpacklo_epi64(_mm_loadl_epi64((const __m128i *)a_ptr), + _mm_loadl_epi64((const __m128i *)&a_ptr[a_stride])); + const __m128i b = + _mm_unpacklo_epi64(_mm_loadl_epi64((const __m128i *)b_ptr), + _mm_loadl_epi64((const __m128i *)&b_ptr[b_stride])); + // Zero-extend mask to 16 bits + const __m128i m = _mm_unpacklo_epi8( + _mm_unpacklo_epi32( + _mm_cvtsi32_si128(*(const uint32_t *)m_ptr), + _mm_cvtsi32_si128(*(const uint32_t *)&m_ptr[m_stride])), + _mm_setzero_si128()); + const __m128i m_inv = _mm_sub_epi16(mask_max, m); + + const __m128i data_l = _mm_unpacklo_epi16(a, b); + const __m128i mask_l = _mm_unpacklo_epi16(m, m_inv); + __m128i pred_l = _mm_madd_epi16(data_l, mask_l); + pred_l = _mm_srai_epi32(_mm_add_epi32(pred_l, round_const), + AOM_BLEND_A64_ROUND_BITS); + + const __m128i data_r = _mm_unpackhi_epi16(a, b); + const __m128i mask_r = _mm_unpackhi_epi16(m, m_inv); + __m128i pred_r = _mm_madd_epi16(data_r, mask_r); + pred_r = _mm_srai_epi32(_mm_add_epi32(pred_r, round_const), + AOM_BLEND_A64_ROUND_BITS); + + const __m128i pred = _mm_packs_epi32(pred_l, pred_r); + const __m128i diff = _mm_abs_epi16(_mm_sub_epi16(pred, src)); + res = _mm_add_epi32(res, _mm_madd_epi16(diff, one)); + + src_ptr += src_stride * 2; 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; + res = _mm_hadd_epi32(res, res); + res = _mm_hadd_epi32(res, res); + int sad = _mm_cvtsi128_si32(res); + return (sad + 31) >> 6; } -#endif // CONFIG_HIGHBITDEPTH + +#endif 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 index fe14597f6..be9d437d2 100644 --- a/third_party/aom/aom_dsp/x86/masked_variance_intrin_ssse3.c +++ b/third_party/aom/aom_dsp/x86/masked_variance_intrin_ssse3.c @@ -1,5 +1,5 @@ /* - * Copyright (c) 2016, Alliance for Open Media. All rights reserved + * 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 @@ -9,1940 +9,1003 @@ * PATENTS file, you can obtain it at www.aomedia.org/license/patent. */ -#include #include -#include +#include #include #include "./aom_config.h" +#include "./aom_dsp_rtcd.h" +#include "aom_dsp/blend.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); \ +#include "aom_dsp/x86/synonyms.h" + +// For width a multiple of 16 +static void bilinear_filter(const uint8_t *src, int src_stride, int xoffset, + int yoffset, uint8_t *dst, int w, int h); + +static void bilinear_filter8xh(const uint8_t *src, int src_stride, int xoffset, + int yoffset, uint8_t *dst, int h); + +static void bilinear_filter4xh(const uint8_t *src, int src_stride, int xoffset, + int yoffset, uint8_t *dst, int h); + +// For width a multiple of 16 +static void masked_variance(const uint8_t *src_ptr, int src_stride, + 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, unsigned int *sse, int *sum_); + +static void masked_variance8xh(const uint8_t *src_ptr, int src_stride, + const uint8_t *a_ptr, const uint8_t *b_ptr, + const uint8_t *m_ptr, int m_stride, int height, + unsigned int *sse, int *sum_); + +static void masked_variance4xh(const uint8_t *src_ptr, int src_stride, + const uint8_t *a_ptr, const uint8_t *b_ptr, + const uint8_t *m_ptr, int m_stride, int height, + unsigned int *sse, int *sum_); + +#define MASK_SUBPIX_VAR_SSSE3(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 *ref, int ref_stride, const uint8_t *second_pred, \ + const uint8_t *msk, int msk_stride, int invert_mask, \ + unsigned int *sse) { \ + int sum; \ + uint8_t temp[(H + 1) * W]; \ + \ + bilinear_filter(src, src_stride, xoffset, yoffset, temp, W, H); \ + \ + if (!invert_mask) \ + masked_variance(ref, ref_stride, temp, W, second_pred, W, msk, \ + msk_stride, W, H, sse, &sum); \ + else \ + masked_variance(ref, ref_stride, second_pred, W, temp, W, msk, \ + msk_stride, W, H, sse, &sum); \ + return *sse - (uint32_t)(((int64_t)sum * sum) / (W * H)); \ + } + +#define MASK_SUBPIX_VAR8XH_SSSE3(H) \ + unsigned int aom_masked_sub_pixel_variance8x##H##_ssse3( \ + const uint8_t *src, int src_stride, int xoffset, int yoffset, \ + const uint8_t *ref, int ref_stride, const uint8_t *second_pred, \ + const uint8_t *msk, int msk_stride, int invert_mask, \ + unsigned int *sse) { \ + int sum; \ + uint8_t temp[(H + 1) * 8]; \ + \ + bilinear_filter8xh(src, src_stride, xoffset, yoffset, temp, H); \ + \ + if (!invert_mask) \ + masked_variance8xh(ref, ref_stride, temp, second_pred, msk, msk_stride, \ + H, sse, &sum); \ + else \ + masked_variance8xh(ref, ref_stride, second_pred, temp, msk, msk_stride, \ + H, sse, &sum); \ + return *sse - (uint32_t)(((int64_t)sum * sum) / (8 * H)); \ + } + +#define MASK_SUBPIX_VAR4XH_SSSE3(H) \ + unsigned int aom_masked_sub_pixel_variance4x##H##_ssse3( \ + const uint8_t *src, int src_stride, int xoffset, int yoffset, \ + const uint8_t *ref, int ref_stride, const uint8_t *second_pred, \ + const uint8_t *msk, int msk_stride, int invert_mask, \ + unsigned int *sse) { \ + int sum; \ + uint8_t temp[(H + 1) * 4]; \ + \ + bilinear_filter4xh(src, src_stride, xoffset, yoffset, temp, H); \ + \ + if (!invert_mask) \ + masked_variance4xh(ref, ref_stride, temp, second_pred, msk, msk_stride, \ + H, sse, &sum); \ + else \ + masked_variance4xh(ref, ref_stride, second_pred, temp, msk, msk_stride, \ + H, sse, &sum); \ + return *sse - (uint32_t)(((int64_t)sum * sum) / (4 * H)); \ } -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 - +MASK_SUBPIX_VAR_SSSE3(128, 128) +MASK_SUBPIX_VAR_SSSE3(128, 64) +MASK_SUBPIX_VAR_SSSE3(64, 128) #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) { +MASK_SUBPIX_VAR_SSSE3(64, 64) +MASK_SUBPIX_VAR_SSSE3(64, 32) +MASK_SUBPIX_VAR_SSSE3(32, 64) +MASK_SUBPIX_VAR_SSSE3(32, 32) +MASK_SUBPIX_VAR_SSSE3(32, 16) +MASK_SUBPIX_VAR_SSSE3(16, 32) +MASK_SUBPIX_VAR_SSSE3(16, 16) +MASK_SUBPIX_VAR_SSSE3(16, 8) +MASK_SUBPIX_VAR8XH_SSSE3(16) +MASK_SUBPIX_VAR8XH_SSSE3(8) +MASK_SUBPIX_VAR8XH_SSSE3(4) +MASK_SUBPIX_VAR4XH_SSSE3(8) +MASK_SUBPIX_VAR4XH_SSSE3(4) + +static INLINE __m128i filter_block(const __m128i a, const __m128i b, + const __m128i filter) { + __m128i v0 = _mm_unpacklo_epi8(a, b); + v0 = _mm_maddubs_epi16(v0, filter); + v0 = xx_roundn_epu16(v0, FILTER_BITS); + + __m128i v1 = _mm_unpackhi_epi8(a, b); + v1 = _mm_maddubs_epi16(v1, filter); + v1 = xx_roundn_epu16(v1, FILTER_BITS); + + return _mm_packus_epi16(v0, v1); +} + +static void bilinear_filter(const uint8_t *src, int src_stride, int xoffset, + int yoffset, uint8_t *dst, int w, int h) { 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; + // Horizontal filter + if (xoffset == 0) { + uint8_t *b = dst; + for (i = 0; i < h + 1; ++i) { + for (j = 0; j < w; j += 16) { + __m128i x = _mm_loadu_si128((__m128i *)&src[j]); + _mm_storeu_si128((__m128i *)&b[j], x); + } + src += src_stride; + b += w; } - // 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); + } else if (xoffset == 4) { + uint8_t *b = dst; + for (i = 0; i < h + 1; ++i) { + for (j = 0; j < w; j += 16) { + __m128i x = _mm_loadu_si128((__m128i *)&src[j]); + __m128i y = _mm_loadu_si128((__m128i *)&src[j + 16]); + __m128i z = _mm_alignr_epi8(y, x, 1); + _mm_storeu_si128((__m128i *)&b[j], _mm_avg_epu8(x, z)); + } + src += src_stride; + b += w; } - 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; + } else { + uint8_t *b = dst; + const uint8_t *hfilter = bilinear_filters_2t[xoffset]; + const __m128i hfilter_vec = _mm_set1_epi16(hfilter[0] | (hfilter[1] << 8)); + for (i = 0; i < h + 1; ++i) { + for (j = 0; j < w; j += 16) { + const __m128i x = _mm_loadu_si128((__m128i *)&src[j]); + const __m128i y = _mm_loadu_si128((__m128i *)&src[j + 16]); + const __m128i z = _mm_alignr_epi8(y, x, 1); + const __m128i res = filter_block(x, z, hfilter_vec); + _mm_storeu_si128((__m128i *)&b[j], res); + } + + src += src_stride; + b += w; } - // 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); + // Vertical filter + if (yoffset == 0) { + // The data is already in 'dst', so no need to filter + } else if (yoffset == 4) { + for (i = 0; i < h; ++i) { + for (j = 0; j < w; j += 16) { + __m128i x = _mm_loadu_si128((__m128i *)&dst[j]); + __m128i y = _mm_loadu_si128((__m128i *)&dst[j + w]); + _mm_storeu_si128((__m128i *)&dst[j], _mm_avg_epu8(x, y)); + } + dst += 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); + } else { + const uint8_t *vfilter = bilinear_filters_2t[yoffset]; + const __m128i vfilter_vec = _mm_set1_epi16(vfilter[0] | (vfilter[1] << 8)); + for (i = 0; i < h; ++i) { + for (j = 0; j < w; j += 16) { + const __m128i x = _mm_loadu_si128((__m128i *)&dst[j]); + const __m128i y = _mm_loadu_si128((__m128i *)&dst[j + w]); + const __m128i res = filter_block(x, y, vfilter_vec); + _mm_storeu_si128((__m128i *)&dst[j], res); + } + + dst += 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); -} +static INLINE __m128i filter_block_2rows(const __m128i a0, const __m128i b0, + const __m128i a1, const __m128i b1, + const __m128i filter) { + __m128i v0 = _mm_unpacklo_epi8(a0, b0); + v0 = _mm_maddubs_epi16(v0, filter); + v0 = xx_roundn_epu16(v0, FILTER_BITS); -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); + __m128i v1 = _mm_unpacklo_epi8(a1, b1); + v1 = _mm_maddubs_epi16(v1, filter); + v1 = xx_roundn_epu16(v1, FILTER_BITS); + + return _mm_packus_epi16(v0, v1); } -// 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) { +static void bilinear_filter8xh(const uint8_t *src, int src_stride, int xoffset, + int yoffset, uint8_t *dst, 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); + // Horizontal filter + if (xoffset == 0) { + uint8_t *b = dst; + for (i = 0; i < h + 1; ++i) { + __m128i x = _mm_loadl_epi64((__m128i *)src); + _mm_storel_epi64((__m128i *)b, x); + src += src_stride; + b += 8; + } + } else if (xoffset == 4) { + uint8_t *b = dst; + for (i = 0; i < h + 1; ++i) { + __m128i x = _mm_loadu_si128((__m128i *)src); + __m128i z = _mm_srli_si128(x, 1); + _mm_storel_epi64((__m128i *)b, _mm_avg_epu8(x, z)); + src += src_stride; + b += 8; } - // 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)); - } + uint8_t *b = dst; + const uint8_t *hfilter = bilinear_filters_2t[xoffset]; + const __m128i hfilter_vec = _mm_set1_epi16(hfilter[0] | (hfilter[1] << 8)); + for (i = 0; i < h; i += 2) { + const __m128i x0 = _mm_loadu_si128((__m128i *)src); + const __m128i z0 = _mm_srli_si128(x0, 1); + const __m128i x1 = _mm_loadu_si128((__m128i *)&src[src_stride]); + const __m128i z1 = _mm_srli_si128(x1, 1); + const __m128i res = filter_block_2rows(x0, z0, x1, z1, hfilter_vec); + _mm_storeu_si128((__m128i *)b, res); - 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); + src += src_stride * 2; + b += 16; } - // 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); + // Handle i = h separately + const __m128i x0 = _mm_loadu_si128((__m128i *)src); + const __m128i z0 = _mm_srli_si128(x0, 1); + + __m128i v0 = _mm_unpacklo_epi8(x0, z0); + v0 = _mm_maddubs_epi16(v0, hfilter_vec); + v0 = xx_roundn_epu16(v0, FILTER_BITS); + + _mm_storel_epi64((__m128i *)b, _mm_packus_epi16(v0, v0)); + } + + // Vertical filter + if (yoffset == 0) { + // The data is already in 'dst', so no need to filter + } else if (yoffset == 4) { + for (i = 0; i < h; ++i) { + __m128i x = _mm_loadl_epi64((__m128i *)dst); + __m128i y = _mm_loadl_epi64((__m128i *)&dst[8]); + _mm_storel_epi64((__m128i *)dst, _mm_avg_epu8(x, y)); + dst += 8; } + } else { + const uint8_t *vfilter = bilinear_filters_2t[yoffset]; + const __m128i vfilter_vec = _mm_set1_epi16(vfilter[0] | (vfilter[1] << 8)); + for (i = 0; i < h; i += 2) { + const __m128i x = _mm_loadl_epi64((__m128i *)dst); + const __m128i y = _mm_loadl_epi64((__m128i *)&dst[8]); + const __m128i z = _mm_loadl_epi64((__m128i *)&dst[16]); + const __m128i res = filter_block_2rows(x, y, y, z, vfilter_vec); + _mm_storeu_si128((__m128i *)dst, res); - // 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; + dst += 16; + } } - 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) { +static void bilinear_filter4xh(const uint8_t *src, int src_stride, int xoffset, + int yoffset, uint8_t *dst, 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); + // Horizontal filter + if (xoffset == 0) { + uint8_t *b = dst; + for (i = 0; i < h + 1; ++i) { + __m128i x = xx_loadl_32((__m128i *)src); + xx_storel_32((__m128i *)b, x); + src += src_stride; + b += 4; } - // 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); + } else if (xoffset == 4) { + uint8_t *b = dst; + for (i = 0; i < h + 1; ++i) { + __m128i x = _mm_loadl_epi64((__m128i *)src); + __m128i z = _mm_srli_si128(x, 1); + xx_storel_32((__m128i *)b, _mm_avg_epu8(x, z)); + src += src_stride; + b += 4; } - // 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); + } else { + uint8_t *b = dst; + const uint8_t *hfilter = bilinear_filters_2t[xoffset]; + const __m128i hfilter_vec = _mm_set1_epi16(hfilter[0] | (hfilter[1] << 8)); + for (i = 0; i < h; i += 4) { + const __m128i x0 = _mm_loadl_epi64((__m128i *)src); + const __m128i z0 = _mm_srli_si128(x0, 1); + const __m128i x1 = _mm_loadl_epi64((__m128i *)&src[src_stride]); + const __m128i z1 = _mm_srli_si128(x1, 1); + const __m128i x2 = _mm_loadl_epi64((__m128i *)&src[src_stride * 2]); + const __m128i z2 = _mm_srli_si128(x2, 1); + const __m128i x3 = _mm_loadl_epi64((__m128i *)&src[src_stride * 3]); + const __m128i z3 = _mm_srli_si128(x3, 1); + + const __m128i a0 = _mm_unpacklo_epi32(x0, x1); + const __m128i b0 = _mm_unpacklo_epi32(z0, z1); + const __m128i a1 = _mm_unpacklo_epi32(x2, x3); + const __m128i b1 = _mm_unpacklo_epi32(z2, z3); + const __m128i res = filter_block_2rows(a0, b0, a1, b1, hfilter_vec); + _mm_storeu_si128((__m128i *)b, res); + + src += src_stride * 4; + b += 16; } - // 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); + // Handle i = h separately + const __m128i x = _mm_loadl_epi64((__m128i *)src); + const __m128i z = _mm_srli_si128(x, 1); + + __m128i v0 = _mm_unpacklo_epi8(x, z); + v0 = _mm_maddubs_epi16(v0, hfilter_vec); + v0 = xx_roundn_epu16(v0, FILTER_BITS); + + xx_storel_32((__m128i *)b, _mm_packus_epi16(v0, v0)); + } + + // Vertical filter + if (yoffset == 0) { + // The data is already in 'dst', so no need to filter + } else if (yoffset == 4) { + for (i = 0; i < h; ++i) { + __m128i x = xx_loadl_32((__m128i *)dst); + __m128i y = xx_loadl_32((__m128i *)&dst[4]); + xx_storel_32((__m128i *)dst, _mm_avg_epu8(x, y)); + dst += 4; + } + } else { + const uint8_t *vfilter = bilinear_filters_2t[yoffset]; + const __m128i vfilter_vec = _mm_set1_epi16(vfilter[0] | (vfilter[1] << 8)); + for (i = 0; i < h; i += 4) { + const __m128i a = xx_loadl_32((__m128i *)dst); + const __m128i b = xx_loadl_32((__m128i *)&dst[4]); + const __m128i c = xx_loadl_32((__m128i *)&dst[8]); + const __m128i d = xx_loadl_32((__m128i *)&dst[12]); + const __m128i e = xx_loadl_32((__m128i *)&dst[16]); + + const __m128i a0 = _mm_unpacklo_epi32(a, b); + const __m128i b0 = _mm_unpacklo_epi32(b, c); + const __m128i a1 = _mm_unpacklo_epi32(c, d); + const __m128i b1 = _mm_unpacklo_epi32(d, e); + const __m128i res = filter_block_2rows(a0, b0, a1, b1, vfilter_vec); + _mm_storeu_si128((__m128i *)dst, res); + + dst += 16; } - // 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 INLINE void accumulate_block(const __m128i src, const __m128i a, + const __m128i b, const __m128i m, + __m128i *sum, __m128i *sum_sq) { + const __m128i zero = _mm_setzero_si128(); + const __m128i one = _mm_set1_epi16(1); + const __m128i mask_max = _mm_set1_epi8((1 << AOM_BLEND_A64_ROUND_BITS)); + const __m128i m_inv = _mm_sub_epi8(mask_max, m); + + // Calculate 16 predicted pixels. + // Note that the maximum value of any entry of 'pred_l' or 'pred_r' + // is 64 * 255, so we have plenty of space to add rounding constants. + const __m128i data_l = _mm_unpacklo_epi8(a, b); + const __m128i mask_l = _mm_unpacklo_epi8(m, m_inv); + __m128i pred_l = _mm_maddubs_epi16(data_l, mask_l); + pred_l = xx_roundn_epu16(pred_l, AOM_BLEND_A64_ROUND_BITS); + + const __m128i data_r = _mm_unpackhi_epi8(a, b); + const __m128i mask_r = _mm_unpackhi_epi8(m, m_inv); + __m128i pred_r = _mm_maddubs_epi16(data_r, mask_r); + pred_r = xx_roundn_epu16(pred_r, AOM_BLEND_A64_ROUND_BITS); + + const __m128i src_l = _mm_unpacklo_epi8(src, zero); + const __m128i src_r = _mm_unpackhi_epi8(src, zero); + const __m128i diff_l = _mm_sub_epi16(pred_l, src_l); + const __m128i diff_r = _mm_sub_epi16(pred_r, src_r); + + // Update partial sums and partial sums of squares + *sum = + _mm_add_epi32(*sum, _mm_madd_epi16(_mm_add_epi16(diff_l, diff_r), one)); + *sum_sq = + _mm_add_epi32(*sum_sq, _mm_add_epi32(_mm_madd_epi16(diff_l, diff_l), + _mm_madd_epi16(diff_r, diff_r))); +} + +static void masked_variance(const uint8_t *src_ptr, int src_stride, + 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, unsigned int *sse, int *sum_) { + int x, y; + __m128i sum = _mm_setzero_si128(), sum_sq = _mm_setzero_si128(); + + for (y = 0; y < height; y++) { + for (x = 0; x < width; x += 16) { + const __m128i src = _mm_loadu_si128((const __m128i *)&src_ptr[x]); + const __m128i a = _mm_loadu_si128((const __m128i *)&a_ptr[x]); + const __m128i b = _mm_loadu_si128((const __m128i *)&b_ptr[x]); + const __m128i m = _mm_loadu_si128((const __m128i *)&m_ptr[x]); + accumulate_block(src, a, b, m, &sum, &sum_sq); + } -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); + src_ptr += src_stride; + a_ptr += a_stride; + b_ptr += b_stride; + m_ptr += m_stride; + } + // Reduce down to a single sum and sum of squares + sum = _mm_hadd_epi32(sum, sum_sq); + sum = _mm_hadd_epi32(sum, sum); + *sum_ = _mm_cvtsi128_si32(sum); + *sse = _mm_cvtsi128_si32(_mm_srli_si128(sum, 4)); +} + +static void masked_variance8xh(const uint8_t *src_ptr, int src_stride, + const uint8_t *a_ptr, const uint8_t *b_ptr, + const uint8_t *m_ptr, int m_stride, int height, + unsigned int *sse, int *sum_) { + int y; + __m128i sum = _mm_setzero_si128(), sum_sq = _mm_setzero_si128(); + + for (y = 0; y < height; y += 2) { + __m128i src = _mm_unpacklo_epi64( + _mm_loadl_epi64((const __m128i *)src_ptr), + _mm_loadl_epi64((const __m128i *)&src_ptr[src_stride])); + const __m128i a = _mm_loadu_si128((const __m128i *)a_ptr); + const __m128i b = _mm_loadu_si128((const __m128i *)b_ptr); + const __m128i m = + _mm_unpacklo_epi64(_mm_loadl_epi64((const __m128i *)m_ptr), + _mm_loadl_epi64((const __m128i *)&m_ptr[m_stride])); + accumulate_block(src, a, b, m, &sum, &sum_sq); + + src_ptr += src_stride * 2; + a_ptr += 16; + b_ptr += 16; + m_ptr += m_stride * 2; + } + // Reduce down to a single sum and sum of squares + sum = _mm_hadd_epi32(sum, sum_sq); + sum = _mm_hadd_epi32(sum, sum); + *sum_ = _mm_cvtsi128_si32(sum); + *sse = _mm_cvtsi128_si32(_mm_srli_si128(sum, 4)); +} + +static void masked_variance4xh(const uint8_t *src_ptr, int src_stride, + const uint8_t *a_ptr, const uint8_t *b_ptr, + const uint8_t *m_ptr, int m_stride, int height, + unsigned int *sse, int *sum_) { + int y; + __m128i sum = _mm_setzero_si128(), sum_sq = _mm_setzero_si128(); + + for (y = 0; y < height; y += 4) { + // Load four rows at a time + __m128i src = + _mm_setr_epi32(*(uint32_t *)src_ptr, *(uint32_t *)&src_ptr[src_stride], + *(uint32_t *)&src_ptr[src_stride * 2], + *(uint32_t *)&src_ptr[src_stride * 3]); + const __m128i a = _mm_loadu_si128((const __m128i *)a_ptr); + const __m128i b = _mm_loadu_si128((const __m128i *)b_ptr); + const __m128i m = _mm_setr_epi32( + *(uint32_t *)m_ptr, *(uint32_t *)&m_ptr[m_stride], + *(uint32_t *)&m_ptr[m_stride * 2], *(uint32_t *)&m_ptr[m_stride * 3]); + accumulate_block(src, a, b, m, &sum, &sum_sq); + + src_ptr += src_stride * 4; + a_ptr += 16; + b_ptr += 16; + m_ptr += m_stride * 4; + } + // Reduce down to a single sum and sum of squares + sum = _mm_hadd_epi32(sum, sum_sq); + sum = _mm_hadd_epi32(sum, sum); + *sum_ = _mm_cvtsi128_si32(sum); + *sse = _mm_cvtsi128_si32(_mm_srli_si128(sum, 4)); } -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)); -} +#if CONFIG_HIGHBITDEPTH +// For width a multiple of 8 +static void highbd_bilinear_filter(const uint16_t *src, int src_stride, + int xoffset, int yoffset, uint16_t *dst, + int w, int h); + +static void highbd_bilinear_filter4xh(const uint16_t *src, int src_stride, + int xoffset, int yoffset, uint16_t *dst, + int h); + +// For width a multiple of 8 +static void highbd_masked_variance(const uint16_t *src_ptr, int src_stride, + const uint16_t *a_ptr, int a_stride, + const uint16_t *b_ptr, int b_stride, + const uint8_t *m_ptr, int m_stride, + int width, int height, uint64_t *sse, + int *sum_); + +static void highbd_masked_variance4xh(const uint16_t *src_ptr, int src_stride, + const uint16_t *a_ptr, + const uint16_t *b_ptr, + const uint8_t *m_ptr, int m_stride, + int height, int *sse, int *sum_); + +#define HIGHBD_MASK_SUBPIX_VAR_SSSE3(W, H) \ + unsigned int aom_highbd_8_masked_sub_pixel_variance##W##x##H##_ssse3( \ + const uint8_t *src8, int src_stride, int xoffset, int yoffset, \ + const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, \ + const uint8_t *msk, int msk_stride, int invert_mask, uint32_t *sse) { \ + uint64_t sse64; \ + int sum; \ + uint16_t temp[(H + 1) * W]; \ + const uint16_t *src = CONVERT_TO_SHORTPTR(src8); \ + const uint16_t *ref = CONVERT_TO_SHORTPTR(ref8); \ + const uint16_t *second_pred = CONVERT_TO_SHORTPTR(second_pred8); \ + \ + highbd_bilinear_filter(src, src_stride, xoffset, yoffset, temp, W, H); \ + \ + if (!invert_mask) \ + highbd_masked_variance(ref, ref_stride, temp, W, second_pred, W, msk, \ + msk_stride, W, H, &sse64, &sum); \ + else \ + highbd_masked_variance(ref, ref_stride, second_pred, W, temp, W, msk, \ + msk_stride, W, H, &sse64, &sum); \ + *sse = (uint32_t)sse64; \ + return *sse - (uint32_t)(((int64_t)sum * sum) / (W * H)); \ + } \ + 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 *ref8, int ref_stride, const uint8_t *second_pred8, \ + const uint8_t *msk, int msk_stride, int invert_mask, uint32_t *sse) { \ + uint64_t sse64; \ + int sum; \ + uint16_t temp[(H + 1) * W]; \ + const uint16_t *src = CONVERT_TO_SHORTPTR(src8); \ + const uint16_t *ref = CONVERT_TO_SHORTPTR(ref8); \ + const uint16_t *second_pred = CONVERT_TO_SHORTPTR(second_pred8); \ + \ + highbd_bilinear_filter(src, src_stride, xoffset, yoffset, temp, W, H); \ + \ + if (!invert_mask) \ + highbd_masked_variance(ref, ref_stride, temp, W, second_pred, W, msk, \ + msk_stride, W, H, &sse64, &sum); \ + else \ + highbd_masked_variance(ref, ref_stride, second_pred, W, temp, W, msk, \ + msk_stride, W, H, &sse64, &sum); \ + *sse = (uint32_t)ROUND_POWER_OF_TWO(sse64, 4); \ + sum = ROUND_POWER_OF_TWO(sum, 2); \ + return *sse - (uint32_t)(((int64_t)sum * sum) / (W * H)); \ + } \ + 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 *ref8, int ref_stride, const uint8_t *second_pred8, \ + const uint8_t *msk, int msk_stride, int invert_mask, uint32_t *sse) { \ + uint64_t sse64; \ + int sum; \ + uint16_t temp[(H + 1) * W]; \ + const uint16_t *src = CONVERT_TO_SHORTPTR(src8); \ + const uint16_t *ref = CONVERT_TO_SHORTPTR(ref8); \ + const uint16_t *second_pred = CONVERT_TO_SHORTPTR(second_pred8); \ + \ + highbd_bilinear_filter(src, src_stride, xoffset, yoffset, temp, W, H); \ + \ + if (!invert_mask) \ + highbd_masked_variance(ref, ref_stride, temp, W, second_pred, W, msk, \ + msk_stride, W, H, &sse64, &sum); \ + else \ + highbd_masked_variance(ref, ref_stride, second_pred, W, temp, W, msk, \ + msk_stride, W, H, &sse64, &sum); \ + *sse = (uint32_t)ROUND_POWER_OF_TWO(sse64, 8); \ + sum = ROUND_POWER_OF_TWO(sum, 4); \ + return *sse - (uint32_t)(((int64_t)sum * sum) / (W * H)); \ + } + +#define HIGHBD_MASK_SUBPIX_VAR4XH_SSSE3(H) \ + unsigned int aom_highbd_8_masked_sub_pixel_variance4x##H##_ssse3( \ + const uint8_t *src8, int src_stride, int xoffset, int yoffset, \ + const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, \ + const uint8_t *msk, int msk_stride, int invert_mask, uint32_t *sse) { \ + int sse_; \ + int sum; \ + uint16_t temp[(H + 1) * 4]; \ + const uint16_t *src = CONVERT_TO_SHORTPTR(src8); \ + const uint16_t *ref = CONVERT_TO_SHORTPTR(ref8); \ + const uint16_t *second_pred = CONVERT_TO_SHORTPTR(second_pred8); \ + \ + highbd_bilinear_filter4xh(src, src_stride, xoffset, yoffset, temp, H); \ + \ + if (!invert_mask) \ + highbd_masked_variance4xh(ref, ref_stride, temp, second_pred, msk, \ + msk_stride, H, &sse_, &sum); \ + else \ + highbd_masked_variance4xh(ref, ref_stride, second_pred, temp, msk, \ + msk_stride, H, &sse_, &sum); \ + *sse = (uint32_t)sse_; \ + return *sse - (uint32_t)(((int64_t)sum * sum) / (4 * H)); \ + } \ + unsigned int aom_highbd_10_masked_sub_pixel_variance4x##H##_ssse3( \ + const uint8_t *src8, int src_stride, int xoffset, int yoffset, \ + const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, \ + const uint8_t *msk, int msk_stride, int invert_mask, uint32_t *sse) { \ + int sse_; \ + int sum; \ + uint16_t temp[(H + 1) * 4]; \ + const uint16_t *src = CONVERT_TO_SHORTPTR(src8); \ + const uint16_t *ref = CONVERT_TO_SHORTPTR(ref8); \ + const uint16_t *second_pred = CONVERT_TO_SHORTPTR(second_pred8); \ + \ + highbd_bilinear_filter4xh(src, src_stride, xoffset, yoffset, temp, H); \ + \ + if (!invert_mask) \ + highbd_masked_variance4xh(ref, ref_stride, temp, second_pred, msk, \ + msk_stride, H, &sse_, &sum); \ + else \ + highbd_masked_variance4xh(ref, ref_stride, second_pred, temp, msk, \ + msk_stride, H, &sse_, &sum); \ + *sse = (uint32_t)ROUND_POWER_OF_TWO(sse_, 4); \ + sum = ROUND_POWER_OF_TWO(sum, 2); \ + return *sse - (uint32_t)(((int64_t)sum * sum) / (4 * H)); \ + } \ + unsigned int aom_highbd_12_masked_sub_pixel_variance4x##H##_ssse3( \ + const uint8_t *src8, int src_stride, int xoffset, int yoffset, \ + const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, \ + const uint8_t *msk, int msk_stride, int invert_mask, uint32_t *sse) { \ + int sse_; \ + int sum; \ + uint16_t temp[(H + 1) * 4]; \ + const uint16_t *src = CONVERT_TO_SHORTPTR(src8); \ + const uint16_t *ref = CONVERT_TO_SHORTPTR(ref8); \ + const uint16_t *second_pred = CONVERT_TO_SHORTPTR(second_pred8); \ + \ + highbd_bilinear_filter4xh(src, src_stride, xoffset, yoffset, temp, H); \ + \ + if (!invert_mask) \ + highbd_masked_variance4xh(ref, ref_stride, temp, second_pred, msk, \ + msk_stride, H, &sse_, &sum); \ + else \ + highbd_masked_variance4xh(ref, ref_stride, second_pred, temp, msk, \ + msk_stride, H, &sse_, &sum); \ + *sse = (uint32_t)ROUND_POWER_OF_TWO(sse_, 8); \ + sum = ROUND_POWER_OF_TWO(sum, 4); \ + return *sse - (uint32_t)(((int64_t)sum * sum) / (4 * 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) { +#if CONFIG_EXT_PARTITION +HIGHBD_MASK_SUBPIX_VAR_SSSE3(128, 128) +HIGHBD_MASK_SUBPIX_VAR_SSSE3(128, 64) +HIGHBD_MASK_SUBPIX_VAR_SSSE3(64, 128) +#endif +HIGHBD_MASK_SUBPIX_VAR_SSSE3(64, 64) +HIGHBD_MASK_SUBPIX_VAR_SSSE3(64, 32) +HIGHBD_MASK_SUBPIX_VAR_SSSE3(32, 64) +HIGHBD_MASK_SUBPIX_VAR_SSSE3(32, 32) +HIGHBD_MASK_SUBPIX_VAR_SSSE3(32, 16) +HIGHBD_MASK_SUBPIX_VAR_SSSE3(16, 32) +HIGHBD_MASK_SUBPIX_VAR_SSSE3(16, 16) +HIGHBD_MASK_SUBPIX_VAR_SSSE3(16, 8) +HIGHBD_MASK_SUBPIX_VAR_SSSE3(8, 16) +HIGHBD_MASK_SUBPIX_VAR_SSSE3(8, 8) +HIGHBD_MASK_SUBPIX_VAR_SSSE3(8, 4) +HIGHBD_MASK_SUBPIX_VAR4XH_SSSE3(8) +HIGHBD_MASK_SUBPIX_VAR4XH_SSSE3(4) + +static INLINE __m128i highbd_filter_block(const __m128i a, const __m128i b, + const __m128i filter) { + __m128i v0 = _mm_unpacklo_epi16(a, b); + v0 = _mm_madd_epi16(v0, filter); + v0 = xx_roundn_epu32(v0, FILTER_BITS); + + __m128i v1 = _mm_unpackhi_epi16(a, b); + v1 = _mm_madd_epi16(v1, filter); + v1 = xx_roundn_epu32(v1, FILTER_BITS); + + return _mm_packs_epi32(v0, v1); +} + +static void highbd_bilinear_filter(const uint16_t *src, int src_stride, + int xoffset, int yoffset, uint16_t *dst, + int w, int h) { 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; + // Horizontal filter + if (xoffset == 0) { + uint16_t *b = dst; + for (i = 0; i < h + 1; ++i) { + for (j = 0; j < w; j += 8) { + __m128i x = _mm_loadu_si128((__m128i *)&src[j]); + _mm_storeu_si128((__m128i *)&b[j], x); + } + src += src_stride; + b += w; } - // 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); + } else if (xoffset == 4) { + uint16_t *b = dst; + for (i = 0; i < h + 1; ++i) { + for (j = 0; j < w; j += 8) { + __m128i x = _mm_loadu_si128((__m128i *)&src[j]); + __m128i y = _mm_loadu_si128((__m128i *)&src[j + 8]); + __m128i z = _mm_alignr_epi8(y, x, 2); + _mm_storeu_si128((__m128i *)&b[j], _mm_avg_epu16(x, z)); + } + src += src_stride; + b += w; } - 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; + } else { + uint16_t *b = dst; + const uint8_t *hfilter = bilinear_filters_2t[xoffset]; + const __m128i hfilter_vec = _mm_set1_epi32(hfilter[0] | (hfilter[1] << 16)); + for (i = 0; i < h + 1; ++i) { + for (j = 0; j < w; j += 8) { + const __m128i x = _mm_loadu_si128((__m128i *)&src[j]); + const __m128i y = _mm_loadu_si128((__m128i *)&src[j + 8]); + const __m128i z = _mm_alignr_epi8(y, x, 2); + const __m128i res = highbd_filter_block(x, z, hfilter_vec); + _mm_storeu_si128((__m128i *)&b[j], res); + } + + src += src_stride; + b += w; } - // 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); + // Vertical filter + if (yoffset == 0) { + // The data is already in 'dst', so no need to filter + } else if (yoffset == 4) { + for (i = 0; i < h; ++i) { + for (j = 0; j < w; j += 8) { + __m128i x = _mm_loadu_si128((__m128i *)&dst[j]); + __m128i y = _mm_loadu_si128((__m128i *)&dst[j + w]); + _mm_storeu_si128((__m128i *)&dst[j], _mm_avg_epu16(x, y)); + } + dst += w; + } + } else { + const uint8_t *vfilter = bilinear_filters_2t[yoffset]; + const __m128i vfilter_vec = _mm_set1_epi32(vfilter[0] | (vfilter[1] << 16)); + for (i = 0; i < h; ++i) { + for (j = 0; j < w; j += 8) { + const __m128i x = _mm_loadu_si128((__m128i *)&dst[j]); + const __m128i y = _mm_loadu_si128((__m128i *)&dst[j + w]); + const __m128i res = highbd_filter_block(x, y, vfilter_vec); + _mm_storeu_si128((__m128i *)&dst[j], res); + } + + dst += w; } - // 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); +static INLINE __m128i highbd_filter_block_2rows(const __m128i a0, + const __m128i b0, + const __m128i a1, + const __m128i b1, + const __m128i filter) { + __m128i v0 = _mm_unpacklo_epi16(a0, b0); + v0 = _mm_madd_epi16(v0, filter); + v0 = xx_roundn_epu32(v0, FILTER_BITS); + + __m128i v1 = _mm_unpacklo_epi16(a1, b1); + v1 = _mm_madd_epi16(v1, filter); + v1 = xx_roundn_epu32(v1, FILTER_BITS); + + return _mm_packs_epi32(v0, v1); } -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) { +static void highbd_bilinear_filter4xh(const uint16_t *src, int src_stride, + int xoffset, int yoffset, uint16_t *dst, + int h) { 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); + // Horizontal filter + if (xoffset == 0) { + uint16_t *b = dst; + for (i = 0; i < h + 1; ++i) { + __m128i x = _mm_loadl_epi64((__m128i *)src); + _mm_storel_epi64((__m128i *)b, x); + src += src_stride; + b += 4; } - // 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); + } else if (xoffset == 4) { + uint16_t *b = dst; + for (i = 0; i < h + 1; ++i) { + __m128i x = _mm_loadu_si128((__m128i *)src); + __m128i z = _mm_srli_si128(x, 2); + _mm_storel_epi64((__m128i *)b, _mm_avg_epu16(x, z)); + src += src_stride; + b += 4; } - // 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); + } else { + uint16_t *b = dst; + const uint8_t *hfilter = bilinear_filters_2t[xoffset]; + const __m128i hfilter_vec = _mm_set1_epi32(hfilter[0] | (hfilter[1] << 16)); + for (i = 0; i < h; i += 2) { + const __m128i x0 = _mm_loadu_si128((__m128i *)src); + const __m128i z0 = _mm_srli_si128(x0, 2); + const __m128i x1 = _mm_loadu_si128((__m128i *)&src[src_stride]); + const __m128i z1 = _mm_srli_si128(x1, 2); + const __m128i res = + highbd_filter_block_2rows(x0, z0, x1, z1, hfilter_vec); + _mm_storeu_si128((__m128i *)b, res); + + src += src_stride * 2; + b += 8; } - // 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); + // Process i = h separately + __m128i x = _mm_loadu_si128((__m128i *)src); + __m128i z = _mm_srli_si128(x, 2); + + __m128i v0 = _mm_unpacklo_epi16(x, z); + v0 = _mm_madd_epi16(v0, hfilter_vec); + v0 = xx_roundn_epu32(v0, FILTER_BITS); + + _mm_storel_epi64((__m128i *)b, _mm_packs_epi32(v0, v0)); + } + + // Vertical filter + if (yoffset == 0) { + // The data is already in 'dst', so no need to filter + } else if (yoffset == 4) { + for (i = 0; i < h; ++i) { + __m128i x = _mm_loadl_epi64((__m128i *)dst); + __m128i y = _mm_loadl_epi64((__m128i *)&dst[4]); + _mm_storel_epi64((__m128i *)dst, _mm_avg_epu16(x, y)); + dst += 4; } - // 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); -} + } else { + const uint8_t *vfilter = bilinear_filters_2t[yoffset]; + const __m128i vfilter_vec = _mm_set1_epi32(vfilter[0] | (vfilter[1] << 16)); + for (i = 0; i < h; i += 2) { + const __m128i x = _mm_loadl_epi64((__m128i *)dst); + const __m128i y = _mm_loadl_epi64((__m128i *)&dst[4]); + const __m128i z = _mm_loadl_epi64((__m128i *)&dst[8]); + const __m128i res = highbd_filter_block_2rows(x, y, y, z, vfilter_vec); + _mm_storeu_si128((__m128i *)dst, res); -// 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); \ - } \ + dst += 8; + } } +} -// 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); \ - } +static void highbd_masked_variance(const uint16_t *src_ptr, int src_stride, + const uint16_t *a_ptr, int a_stride, + const uint16_t *b_ptr, int b_stride, + const uint8_t *m_ptr, int m_stride, + int width, int height, uint64_t *sse, + int *sum_) { + int x, y; + // Note on bit widths: + // The maximum value of 'sum' is (2^12 - 1) * 128 * 128 =~ 2^26, + // so this can be kept as four 32-bit values. + // But the maximum value of 'sum_sq' is (2^12 - 1)^2 * 128 * 128 =~ 2^38, + // so this must be stored as two 64-bit values. + __m128i sum = _mm_setzero_si128(), sum_sq = _mm_setzero_si128(); + const __m128i mask_max = _mm_set1_epi16((1 << AOM_BLEND_A64_ROUND_BITS)); + const __m128i round_const = + _mm_set1_epi32((1 << AOM_BLEND_A64_ROUND_BITS) >> 1); + const __m128i zero = _mm_setzero_si128(); + + for (y = 0; y < height; y++) { + for (x = 0; x < width; x += 8) { + const __m128i src = _mm_loadu_si128((const __m128i *)&src_ptr[x]); + const __m128i a = _mm_loadu_si128((const __m128i *)&a_ptr[x]); + const __m128i b = _mm_loadu_si128((const __m128i *)&b_ptr[x]); + const __m128i m = + _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)&m_ptr[x]), zero); + const __m128i m_inv = _mm_sub_epi16(mask_max, m); + + // Calculate 8 predicted pixels. + const __m128i data_l = _mm_unpacklo_epi16(a, b); + const __m128i mask_l = _mm_unpacklo_epi16(m, m_inv); + __m128i pred_l = _mm_madd_epi16(data_l, mask_l); + pred_l = _mm_srai_epi32(_mm_add_epi32(pred_l, round_const), + AOM_BLEND_A64_ROUND_BITS); + + const __m128i data_r = _mm_unpackhi_epi16(a, b); + const __m128i mask_r = _mm_unpackhi_epi16(m, m_inv); + __m128i pred_r = _mm_madd_epi16(data_r, mask_r); + pred_r = _mm_srai_epi32(_mm_add_epi32(pred_r, round_const), + AOM_BLEND_A64_ROUND_BITS); + + const __m128i src_l = _mm_unpacklo_epi16(src, zero); + const __m128i src_r = _mm_unpackhi_epi16(src, zero); + __m128i diff_l = _mm_sub_epi32(pred_l, src_l); + __m128i diff_r = _mm_sub_epi32(pred_r, src_r); + + // Update partial sums and partial sums of squares + sum = _mm_add_epi32(sum, _mm_add_epi32(diff_l, diff_r)); + // A trick: Now each entry of diff_l and diff_r is stored in a 32-bit + // field, but the range of values is only [-(2^12 - 1), 2^12 - 1]. + // So we can re-pack into 16-bit fields and use _mm_madd_epi16 + // to calculate the squares and partially sum them. + const __m128i tmp = _mm_packs_epi32(diff_l, diff_r); + const __m128i prod = _mm_madd_epi16(tmp, tmp); + // Then we want to sign-extend to 64 bits and accumulate + const __m128i sign = _mm_srai_epi32(prod, 31); + const __m128i tmp_0 = _mm_unpacklo_epi32(prod, sign); + const __m128i tmp_1 = _mm_unpackhi_epi32(prod, sign); + sum_sq = _mm_add_epi64(sum_sq, _mm_add_epi64(tmp_0, tmp_1)); + } -#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); \ - } + src_ptr += src_stride; + a_ptr += a_stride; + b_ptr += b_stride; + m_ptr += m_stride; + } + // Reduce down to a single sum and sum of squares + sum = _mm_hadd_epi32(sum, zero); + sum = _mm_hadd_epi32(sum, zero); + *sum_ = _mm_cvtsi128_si32(sum); + sum_sq = _mm_add_epi64(sum_sq, _mm_srli_si128(sum_sq, 8)); + _mm_storel_epi64((__m128i *)sse, sum_sq); +} + +static void highbd_masked_variance4xh(const uint16_t *src_ptr, int src_stride, + const uint16_t *a_ptr, + const uint16_t *b_ptr, + const uint8_t *m_ptr, int m_stride, + int height, int *sse, int *sum_) { + int y; + // Note: For this function, h <= 8 (or maybe 16 if we add 4:1 partitions). + // So the maximum value of sum is (2^12 - 1) * 4 * 16 =~ 2^18 + // and the maximum value of sum_sq is (2^12 - 1)^2 * 4 * 16 =~ 2^30. + // So we can safely pack sum_sq into 32-bit fields, which is slightly more + // convenient. + __m128i sum = _mm_setzero_si128(), sum_sq = _mm_setzero_si128(); + const __m128i mask_max = _mm_set1_epi16((1 << AOM_BLEND_A64_ROUND_BITS)); + const __m128i round_const = + _mm_set1_epi32((1 << AOM_BLEND_A64_ROUND_BITS) >> 1); + const __m128i zero = _mm_setzero_si128(); + + for (y = 0; y < height; y += 2) { + __m128i src = _mm_unpacklo_epi64( + _mm_loadl_epi64((const __m128i *)src_ptr), + _mm_loadl_epi64((const __m128i *)&src_ptr[src_stride])); + const __m128i a = _mm_loadu_si128((const __m128i *)a_ptr); + const __m128i b = _mm_loadu_si128((const __m128i *)b_ptr); + const __m128i m = _mm_unpacklo_epi8( + _mm_unpacklo_epi32( + _mm_cvtsi32_si128(*(const uint32_t *)m_ptr), + _mm_cvtsi32_si128(*(const uint32_t *)&m_ptr[m_stride])), + zero); + const __m128i m_inv = _mm_sub_epi16(mask_max, m); + + const __m128i data_l = _mm_unpacklo_epi16(a, b); + const __m128i mask_l = _mm_unpacklo_epi16(m, m_inv); + __m128i pred_l = _mm_madd_epi16(data_l, mask_l); + pred_l = _mm_srai_epi32(_mm_add_epi32(pred_l, round_const), + AOM_BLEND_A64_ROUND_BITS); + + const __m128i data_r = _mm_unpackhi_epi16(a, b); + const __m128i mask_r = _mm_unpackhi_epi16(m, m_inv); + __m128i pred_r = _mm_madd_epi16(data_r, mask_r); + pred_r = _mm_srai_epi32(_mm_add_epi32(pred_r, round_const), + AOM_BLEND_A64_ROUND_BITS); + + const __m128i src_l = _mm_unpacklo_epi16(src, zero); + const __m128i src_r = _mm_unpackhi_epi16(src, zero); + __m128i diff_l = _mm_sub_epi32(pred_l, src_l); + __m128i diff_r = _mm_sub_epi32(pred_r, src_r); + + // Update partial sums and partial sums of squares + sum = _mm_add_epi32(sum, _mm_add_epi32(diff_l, diff_r)); + const __m128i tmp = _mm_packs_epi32(diff_l, diff_r); + const __m128i prod = _mm_madd_epi16(tmp, tmp); + sum_sq = _mm_add_epi32(sum_sq, prod); + + src_ptr += src_stride * 2; + a_ptr += 8; + b_ptr += 8; + m_ptr += m_stride * 2; + } + // Reduce down to a single sum and sum of squares + sum = _mm_hadd_epi32(sum, sum_sq); + sum = _mm_hadd_epi32(sum, zero); + *sum_ = _mm_cvtsi128_si32(sum); + *sse = _mm_cvtsi128_si32(_mm_srli_si128(sum, 4)); +} -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_intrinsic_ssse3.h b/third_party/aom/aom_dsp/x86/obmc_intrinsic_ssse3.h new file mode 100644 index 000000000..73589a32a --- /dev/null +++ b/third_party/aom/aom_dsp/x86/obmc_intrinsic_ssse3.h @@ -0,0 +1,45 @@ +/* + * 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. + */ + +#ifndef AOM_DSP_X86_OBMC_INTRINSIC_SSSE3_H_ +#define AOM_DSP_X86_OBMC_INTRINSIC_SSSE3_H_ + +#include + +#include "./aom_config.h" + +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 // AOM_DSP_X86_OBMC_INTRINSIC_SSSE3_H_ diff --git a/third_party/aom/aom_dsp/x86/obmc_sad_sse4.c b/third_party/aom/aom_dsp/x86/obmc_sad_sse4.c index ad77f974c..21632644f 100644 --- a/third_party/aom/aom_dsp/x86/obmc_sad_sse4.c +++ b/third_party/aom/aom_dsp/x86/obmc_sad_sse4.c @@ -17,6 +17,7 @@ #include "aom/aom_integer.h" #include "aom_dsp/aom_dsp_common.h" +#include "aom_dsp/x86/obmc_intrinsic_ssse3.h" #include "aom_dsp/x86/synonyms.h" //////////////////////////////////////////////////////////////////////////////// diff --git a/third_party/aom/aom_dsp/x86/obmc_variance_sse4.c b/third_party/aom/aom_dsp/x86/obmc_variance_sse4.c index efb3659cf..1797ded80 100644 --- a/third_party/aom/aom_dsp/x86/obmc_variance_sse4.c +++ b/third_party/aom/aom_dsp/x86/obmc_variance_sse4.c @@ -17,8 +17,9 @@ #include "aom/aom_integer.h" #include "aom_dsp/aom_dsp_common.h" -#include "aom_dsp/x86/synonyms.h" #include "aom_dsp/aom_filter.h" +#include "aom_dsp/x86/obmc_intrinsic_ssse3.h" +#include "aom_dsp/x86/synonyms.h" //////////////////////////////////////////////////////////////////////////////// // 8 bit diff --git a/third_party/aom/aom_dsp/x86/synonyms.h b/third_party/aom/aom_dsp/x86/synonyms.h index bef606dae..cd049a454 100644 --- a/third_party/aom/aom_dsp/x86/synonyms.h +++ b/third_party/aom/aom_dsp/x86/synonyms.h @@ -89,32 +89,4 @@ static INLINE __m128i xx_roundn_epi32(__m128i v_val_d, int bits) { 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 index 39e9b8e2a..4f7a60c22 100644 --- a/third_party/aom/aom_dsp/x86/txfm_common_avx2.h +++ b/third_party/aom/aom_dsp/x86/txfm_common_avx2.h @@ -34,7 +34,8 @@ static INLINE void mm256_reverse_epi16(__m256i *u) { *u = _mm256_permute2x128_si256(v, v, 1); } -static INLINE void mm256_transpose_16x16(__m256i *in) { +// Note: in and out could have the same value +static INLINE void mm256_transpose_16x16(const __m256i *in, __m256i *out) { __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]); @@ -143,29 +144,30 @@ static INLINE void mm256_transpose_16x16(__m256i *in) { // 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); + out[0] = _mm256_permute2x128_si256(tr0_0, tr0_8, 0x20); // 0010 0000 + out[8] = _mm256_permute2x128_si256(tr0_0, tr0_8, 0x31); // 0011 0001 + out[1] = _mm256_permute2x128_si256(tr0_1, tr0_9, 0x20); + out[9] = _mm256_permute2x128_si256(tr0_1, tr0_9, 0x31); + out[2] = _mm256_permute2x128_si256(tr0_2, tr0_a, 0x20); + out[10] = _mm256_permute2x128_si256(tr0_2, tr0_a, 0x31); + out[3] = _mm256_permute2x128_si256(tr0_3, tr0_b, 0x20); + out[11] = _mm256_permute2x128_si256(tr0_3, tr0_b, 0x31); + + out[4] = _mm256_permute2x128_si256(tr0_4, tr0_c, 0x20); + out[12] = _mm256_permute2x128_si256(tr0_4, tr0_c, 0x31); + out[5] = _mm256_permute2x128_si256(tr0_5, tr0_d, 0x20); + out[13] = _mm256_permute2x128_si256(tr0_5, tr0_d, 0x31); + out[6] = _mm256_permute2x128_si256(tr0_6, tr0_e, 0x20); + out[14] = _mm256_permute2x128_si256(tr0_6, tr0_e, 0x31); + out[7] = _mm256_permute2x128_si256(tr0_7, tr0_f, 0x20); + out[15] = _mm256_permute2x128_si256(tr0_7, tr0_f, 0x31); } -static INLINE __m256i butter_fly(__m256i a0, __m256i a1, const __m256i cospi) { +static INLINE __m256i butter_fly(const __m256i *a0, const __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); + __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); -- cgit v1.2.3