diff options
Diffstat (limited to 'media/libaom/src/third_party/libyuv/source')
45 files changed, 46385 insertions, 0 deletions
diff --git a/media/libaom/src/third_party/libyuv/source/compare.cc b/media/libaom/src/third_party/libyuv/source/compare.cc new file mode 100644 index 000000000..46aa8473d --- /dev/null +++ b/media/libaom/src/third_party/libyuv/source/compare.cc @@ -0,0 +1,373 @@ +/* + * Copyright 2011 The LibYuv Project Authors. All rights reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#include "libyuv/compare.h" + +#include <float.h> +#include <math.h> +#ifdef _OPENMP +#include <omp.h> +#endif + +#include "libyuv/basic_types.h" +#include "libyuv/cpu_id.h" +#include "libyuv/row.h" +#include "libyuv/video_common.h" + +#ifdef __cplusplus +namespace libyuv { +extern "C" { +#endif + +// hash seed of 5381 recommended. +// Internal C version of HashDjb2 with int sized count for efficiency. +uint32 HashDjb2_C(const uint8* src, int count, uint32 seed); + +// This module is for Visual C x86 +#if !defined(LIBYUV_DISABLE_X86) && \ + (defined(_M_IX86) || \ + (defined(__x86_64__) || (defined(__i386__) && !defined(__pic__)))) +#define HAS_HASHDJB2_SSE41 +uint32 HashDjb2_SSE41(const uint8* src, int count, uint32 seed); + +#ifdef VISUALC_HAS_AVX2 +#define HAS_HASHDJB2_AVX2 +uint32 HashDjb2_AVX2(const uint8* src, int count, uint32 seed); +#endif + +#endif // HAS_HASHDJB2_SSE41 + +// hash seed of 5381 recommended. +LIBYUV_API +uint32 HashDjb2(const uint8* src, uint64 count, uint32 seed) { + const int kBlockSize = 1 << 15; // 32768; + int remainder; + uint32 (*HashDjb2_SSE)(const uint8* src, int count, uint32 seed) = HashDjb2_C; +#if defined(HAS_HASHDJB2_SSE41) + if (TestCpuFlag(kCpuHasSSE41)) { + HashDjb2_SSE = HashDjb2_SSE41; + } +#endif +#if defined(HAS_HASHDJB2_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + HashDjb2_SSE = HashDjb2_AVX2; + } +#endif + + while (count >= (uint64)(kBlockSize)) { + seed = HashDjb2_SSE(src, kBlockSize, seed); + src += kBlockSize; + count -= kBlockSize; + } + remainder = (int)(count) & ~15; + if (remainder) { + seed = HashDjb2_SSE(src, remainder, seed); + src += remainder; + count -= remainder; + } + remainder = (int)(count) & 15; + if (remainder) { + seed = HashDjb2_C(src, remainder, seed); + } + return seed; +} + +static uint32 ARGBDetectRow_C(const uint8* argb, int width) { + int x; + for (x = 0; x < width - 1; x += 2) { + if (argb[0] != 255) { // First byte is not Alpha of 255, so not ARGB. + return FOURCC_BGRA; + } + if (argb[3] != 255) { // 4th byte is not Alpha of 255, so not BGRA. + return FOURCC_ARGB; + } + if (argb[4] != 255) { // Second pixel first byte is not Alpha of 255. + return FOURCC_BGRA; + } + if (argb[7] != 255) { // Second pixel 4th byte is not Alpha of 255. + return FOURCC_ARGB; + } + argb += 8; + } + if (width & 1) { + if (argb[0] != 255) { // First byte is not Alpha of 255, so not ARGB. + return FOURCC_BGRA; + } + if (argb[3] != 255) { // 4th byte is not Alpha of 255, so not BGRA. + return FOURCC_ARGB; + } + } + return 0; +} + +// Scan an opaque argb image and return fourcc based on alpha offset. +// Returns FOURCC_ARGB, FOURCC_BGRA, or 0 if unknown. +LIBYUV_API +uint32 ARGBDetect(const uint8* argb, int stride_argb, int width, int height) { + uint32 fourcc = 0; + int h; + + // Coalesce rows. + if (stride_argb == width * 4) { + width *= height; + height = 1; + stride_argb = 0; + } + for (h = 0; h < height && fourcc == 0; ++h) { + fourcc = ARGBDetectRow_C(argb, width); + argb += stride_argb; + } + return fourcc; +} + +uint32 SumSquareError_C(const uint8* src_a, const uint8* src_b, int count); +#if !defined(LIBYUV_DISABLE_NEON) && \ + (defined(__ARM_NEON__) || defined(LIBYUV_NEON) || defined(__aarch64__)) +#define HAS_SUMSQUAREERROR_NEON +uint32 SumSquareError_NEON(const uint8* src_a, const uint8* src_b, int count); +#endif +#if !defined(LIBYUV_DISABLE_X86) && \ + (defined(_M_IX86) || defined(__x86_64__) || defined(__i386__)) +#define HAS_SUMSQUAREERROR_SSE2 +uint32 SumSquareError_SSE2(const uint8* src_a, const uint8* src_b, int count); +#endif + +#ifdef VISUALC_HAS_AVX2 +#define HAS_SUMSQUAREERROR_AVX2 +uint32 SumSquareError_AVX2(const uint8* src_a, const uint8* src_b, int count); +#endif + +// TODO(fbarchard): Refactor into row function. +LIBYUV_API +uint64 ComputeSumSquareError(const uint8* src_a, const uint8* src_b, + int count) { + // SumSquareError returns values 0 to 65535 for each squared difference. + // Up to 65536 of those can be summed and remain within a uint32. + // After each block of 65536 pixels, accumulate into a uint64. + const int kBlockSize = 65536; + int remainder = count & (kBlockSize - 1) & ~31; + uint64 sse = 0; + int i; + uint32 (*SumSquareError)(const uint8* src_a, const uint8* src_b, int count) = + SumSquareError_C; +#if defined(HAS_SUMSQUAREERROR_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + SumSquareError = SumSquareError_NEON; + } +#endif +#if defined(HAS_SUMSQUAREERROR_SSE2) + if (TestCpuFlag(kCpuHasSSE2)) { + // Note only used for multiples of 16 so count is not checked. + SumSquareError = SumSquareError_SSE2; + } +#endif +#if defined(HAS_SUMSQUAREERROR_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + // Note only used for multiples of 32 so count is not checked. + SumSquareError = SumSquareError_AVX2; + } +#endif +#ifdef _OPENMP +#pragma omp parallel for reduction(+: sse) +#endif + for (i = 0; i < (count - (kBlockSize - 1)); i += kBlockSize) { + sse += SumSquareError(src_a + i, src_b + i, kBlockSize); + } + src_a += count & ~(kBlockSize - 1); + src_b += count & ~(kBlockSize - 1); + if (remainder) { + sse += SumSquareError(src_a, src_b, remainder); + src_a += remainder; + src_b += remainder; + } + remainder = count & 31; + if (remainder) { + sse += SumSquareError_C(src_a, src_b, remainder); + } + return sse; +} + +LIBYUV_API +uint64 ComputeSumSquareErrorPlane(const uint8* src_a, int stride_a, + const uint8* src_b, int stride_b, + int width, int height) { + uint64 sse = 0; + int h; + // Coalesce rows. + if (stride_a == width && + stride_b == width) { + width *= height; + height = 1; + stride_a = stride_b = 0; + } + for (h = 0; h < height; ++h) { + sse += ComputeSumSquareError(src_a, src_b, width); + src_a += stride_a; + src_b += stride_b; + } + return sse; +} + +LIBYUV_API +double SumSquareErrorToPsnr(uint64 sse, uint64 count) { + double psnr; + if (sse > 0) { + double mse = (double)(count) / (double)(sse); + psnr = 10.0 * log10(255.0 * 255.0 * mse); + } else { + psnr = kMaxPsnr; // Limit to prevent divide by 0 + } + + if (psnr > kMaxPsnr) + psnr = kMaxPsnr; + + return psnr; +} + +LIBYUV_API +double CalcFramePsnr(const uint8* src_a, int stride_a, + const uint8* src_b, int stride_b, + int width, int height) { + const uint64 samples = width * height; + const uint64 sse = ComputeSumSquareErrorPlane(src_a, stride_a, + src_b, stride_b, + width, height); + return SumSquareErrorToPsnr(sse, samples); +} + +LIBYUV_API +double I420Psnr(const uint8* src_y_a, int stride_y_a, + const uint8* src_u_a, int stride_u_a, + const uint8* src_v_a, int stride_v_a, + const uint8* src_y_b, int stride_y_b, + const uint8* src_u_b, int stride_u_b, + const uint8* src_v_b, int stride_v_b, + int width, int height) { + const uint64 sse_y = ComputeSumSquareErrorPlane(src_y_a, stride_y_a, + src_y_b, stride_y_b, + width, height); + const int width_uv = (width + 1) >> 1; + const int height_uv = (height + 1) >> 1; + const uint64 sse_u = ComputeSumSquareErrorPlane(src_u_a, stride_u_a, + src_u_b, stride_u_b, + width_uv, height_uv); + const uint64 sse_v = ComputeSumSquareErrorPlane(src_v_a, stride_v_a, + src_v_b, stride_v_b, + width_uv, height_uv); + const uint64 samples = width * height + 2 * (width_uv * height_uv); + const uint64 sse = sse_y + sse_u + sse_v; + return SumSquareErrorToPsnr(sse, samples); +} + +static const int64 cc1 = 26634; // (64^2*(.01*255)^2 +static const int64 cc2 = 239708; // (64^2*(.03*255)^2 + +static double Ssim8x8_C(const uint8* src_a, int stride_a, + const uint8* src_b, int stride_b) { + int64 sum_a = 0; + int64 sum_b = 0; + int64 sum_sq_a = 0; + int64 sum_sq_b = 0; + int64 sum_axb = 0; + + int i; + for (i = 0; i < 8; ++i) { + int j; + for (j = 0; j < 8; ++j) { + sum_a += src_a[j]; + sum_b += src_b[j]; + sum_sq_a += src_a[j] * src_a[j]; + sum_sq_b += src_b[j] * src_b[j]; + sum_axb += src_a[j] * src_b[j]; + } + + src_a += stride_a; + src_b += stride_b; + } + + { + const int64 count = 64; + // scale the constants by number of pixels + const int64 c1 = (cc1 * count * count) >> 12; + const int64 c2 = (cc2 * count * count) >> 12; + + const int64 sum_a_x_sum_b = sum_a * sum_b; + + const int64 ssim_n = (2 * sum_a_x_sum_b + c1) * + (2 * count * sum_axb - 2 * sum_a_x_sum_b + c2); + + const int64 sum_a_sq = sum_a*sum_a; + const int64 sum_b_sq = sum_b*sum_b; + + const int64 ssim_d = (sum_a_sq + sum_b_sq + c1) * + (count * sum_sq_a - sum_a_sq + + count * sum_sq_b - sum_b_sq + c2); + + if (ssim_d == 0.0) { + return DBL_MAX; + } + return ssim_n * 1.0 / ssim_d; + } +} + +// We are using a 8x8 moving window with starting location of each 8x8 window +// on the 4x4 pixel grid. Such arrangement allows the windows to overlap +// block boundaries to penalize blocking artifacts. +LIBYUV_API +double CalcFrameSsim(const uint8* src_a, int stride_a, + const uint8* src_b, int stride_b, + int width, int height) { + int samples = 0; + double ssim_total = 0; + double (*Ssim8x8)(const uint8* src_a, int stride_a, + const uint8* src_b, int stride_b) = Ssim8x8_C; + + // sample point start with each 4x4 location + int i; + for (i = 0; i < height - 8; i += 4) { + int j; + for (j = 0; j < width - 8; j += 4) { + ssim_total += Ssim8x8(src_a + j, stride_a, src_b + j, stride_b); + samples++; + } + + src_a += stride_a * 4; + src_b += stride_b * 4; + } + + ssim_total /= samples; + return ssim_total; +} + +LIBYUV_API +double I420Ssim(const uint8* src_y_a, int stride_y_a, + const uint8* src_u_a, int stride_u_a, + const uint8* src_v_a, int stride_v_a, + const uint8* src_y_b, int stride_y_b, + const uint8* src_u_b, int stride_u_b, + const uint8* src_v_b, int stride_v_b, + int width, int height) { + const double ssim_y = CalcFrameSsim(src_y_a, stride_y_a, + src_y_b, stride_y_b, width, height); + const int width_uv = (width + 1) >> 1; + const int height_uv = (height + 1) >> 1; + const double ssim_u = CalcFrameSsim(src_u_a, stride_u_a, + src_u_b, stride_u_b, + width_uv, height_uv); + const double ssim_v = CalcFrameSsim(src_v_a, stride_v_a, + src_v_b, stride_v_b, + width_uv, height_uv); + return ssim_y * 0.8 + 0.1 * (ssim_u + ssim_v); +} + +#ifdef __cplusplus +} // extern "C" +} // namespace libyuv +#endif diff --git a/media/libaom/src/third_party/libyuv/source/compare_common.cc b/media/libaom/src/third_party/libyuv/source/compare_common.cc new file mode 100644 index 000000000..c546b5182 --- /dev/null +++ b/media/libaom/src/third_party/libyuv/source/compare_common.cc @@ -0,0 +1,42 @@ +/* + * Copyright 2012 The LibYuv Project Authors. All rights reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#include "libyuv/basic_types.h" + +#ifdef __cplusplus +namespace libyuv { +extern "C" { +#endif + +uint32 SumSquareError_C(const uint8* src_a, const uint8* src_b, int count) { + uint32 sse = 0u; + int i; + for (i = 0; i < count; ++i) { + int diff = src_a[i] - src_b[i]; + sse += (uint32)(diff * diff); + } + return sse; +} + +// hash seed of 5381 recommended. +// Internal C version of HashDjb2 with int sized count for efficiency. +uint32 HashDjb2_C(const uint8* src, int count, uint32 seed) { + uint32 hash = seed; + int i; + for (i = 0; i < count; ++i) { + hash += (hash << 5) + src[i]; + } + return hash; +} + +#ifdef __cplusplus +} // extern "C" +} // namespace libyuv +#endif diff --git a/media/libaom/src/third_party/libyuv/source/compare_gcc.cc b/media/libaom/src/third_party/libyuv/source/compare_gcc.cc new file mode 100644 index 000000000..247cb33bb --- /dev/null +++ b/media/libaom/src/third_party/libyuv/source/compare_gcc.cc @@ -0,0 +1,152 @@ +/* + * Copyright 2012 The LibYuv Project Authors. All rights reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#include "libyuv/basic_types.h" +#include "libyuv/row.h" + +#ifdef __cplusplus +namespace libyuv { +extern "C" { +#endif + +#if !defined(LIBYUV_DISABLE_X86) && (defined(__x86_64__) || defined(__i386__)) + +uint32 SumSquareError_SSE2(const uint8* src_a, const uint8* src_b, int count) { + uint32 sse; + asm volatile ( // NOLINT + "pxor %%xmm0,%%xmm0 \n" + "pxor %%xmm5,%%xmm5 \n" + LABELALIGN + "1: \n" + "movdqu " MEMACCESS(0) ",%%xmm1 \n" + "lea " MEMLEA(0x10, 0) ",%0 \n" + "movdqu " MEMACCESS(1) ",%%xmm2 \n" + "lea " MEMLEA(0x10, 1) ",%1 \n" + "movdqa %%xmm1,%%xmm3 \n" + "psubusb %%xmm2,%%xmm1 \n" + "psubusb %%xmm3,%%xmm2 \n" + "por %%xmm2,%%xmm1 \n" + "movdqa %%xmm1,%%xmm2 \n" + "punpcklbw %%xmm5,%%xmm1 \n" + "punpckhbw %%xmm5,%%xmm2 \n" + "pmaddwd %%xmm1,%%xmm1 \n" + "pmaddwd %%xmm2,%%xmm2 \n" + "paddd %%xmm1,%%xmm0 \n" + "paddd %%xmm2,%%xmm0 \n" + "sub $0x10,%2 \n" + "jg 1b \n" + + "pshufd $0xee,%%xmm0,%%xmm1 \n" + "paddd %%xmm1,%%xmm0 \n" + "pshufd $0x1,%%xmm0,%%xmm1 \n" + "paddd %%xmm1,%%xmm0 \n" + "movd %%xmm0,%3 \n" + + : "+r"(src_a), // %0 + "+r"(src_b), // %1 + "+r"(count), // %2 + "=g"(sse) // %3 + :: "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm5" + ); // NOLINT + return sse; +} + +#endif // defined(__x86_64__) || defined(__i386__) + +#if !defined(LIBYUV_DISABLE_X86) && \ + (defined(__x86_64__) || (defined(__i386__) && !defined(__pic__))) +#define HAS_HASHDJB2_SSE41 +static uvec32 kHash16x33 = { 0x92d9e201, 0, 0, 0 }; // 33 ^ 16 +static uvec32 kHashMul0 = { + 0x0c3525e1, // 33 ^ 15 + 0xa3476dc1, // 33 ^ 14 + 0x3b4039a1, // 33 ^ 13 + 0x4f5f0981, // 33 ^ 12 +}; +static uvec32 kHashMul1 = { + 0x30f35d61, // 33 ^ 11 + 0x855cb541, // 33 ^ 10 + 0x040a9121, // 33 ^ 9 + 0x747c7101, // 33 ^ 8 +}; +static uvec32 kHashMul2 = { + 0xec41d4e1, // 33 ^ 7 + 0x4cfa3cc1, // 33 ^ 6 + 0x025528a1, // 33 ^ 5 + 0x00121881, // 33 ^ 4 +}; +static uvec32 kHashMul3 = { + 0x00008c61, // 33 ^ 3 + 0x00000441, // 33 ^ 2 + 0x00000021, // 33 ^ 1 + 0x00000001, // 33 ^ 0 +}; + +uint32 HashDjb2_SSE41(const uint8* src, int count, uint32 seed) { + uint32 hash; + asm volatile ( // NOLINT + "movd %2,%%xmm0 \n" + "pxor %%xmm7,%%xmm7 \n" + "movdqa %4,%%xmm6 \n" + LABELALIGN + "1: \n" + "movdqu " MEMACCESS(0) ",%%xmm1 \n" + "lea " MEMLEA(0x10, 0) ",%0 \n" + "pmulld %%xmm6,%%xmm0 \n" + "movdqa %5,%%xmm5 \n" + "movdqa %%xmm1,%%xmm2 \n" + "punpcklbw %%xmm7,%%xmm2 \n" + "movdqa %%xmm2,%%xmm3 \n" + "punpcklwd %%xmm7,%%xmm3 \n" + "pmulld %%xmm5,%%xmm3 \n" + "movdqa %6,%%xmm5 \n" + "movdqa %%xmm2,%%xmm4 \n" + "punpckhwd %%xmm7,%%xmm4 \n" + "pmulld %%xmm5,%%xmm4 \n" + "movdqa %7,%%xmm5 \n" + "punpckhbw %%xmm7,%%xmm1 \n" + "movdqa %%xmm1,%%xmm2 \n" + "punpcklwd %%xmm7,%%xmm2 \n" + "pmulld %%xmm5,%%xmm2 \n" + "movdqa %8,%%xmm5 \n" + "punpckhwd %%xmm7,%%xmm1 \n" + "pmulld %%xmm5,%%xmm1 \n" + "paddd %%xmm4,%%xmm3 \n" + "paddd %%xmm2,%%xmm1 \n" + "paddd %%xmm3,%%xmm1 \n" + "pshufd $0xe,%%xmm1,%%xmm2 \n" + "paddd %%xmm2,%%xmm1 \n" + "pshufd $0x1,%%xmm1,%%xmm2 \n" + "paddd %%xmm2,%%xmm1 \n" + "paddd %%xmm1,%%xmm0 \n" + "sub $0x10,%1 \n" + "jg 1b \n" + "movd %%xmm0,%3 \n" + : "+r"(src), // %0 + "+r"(count), // %1 + "+rm"(seed), // %2 + "=g"(hash) // %3 + : "m"(kHash16x33), // %4 + "m"(kHashMul0), // %5 + "m"(kHashMul1), // %6 + "m"(kHashMul2), // %7 + "m"(kHashMul3) // %8 + : "memory", "cc" + , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7" + ); // NOLINT + return hash; +} +#endif // defined(__x86_64__) || (defined(__i386__) && !defined(__pic__))) + +#ifdef __cplusplus +} // extern "C" +} // namespace libyuv +#endif + diff --git a/media/libaom/src/third_party/libyuv/source/compare_neon.cc b/media/libaom/src/third_party/libyuv/source/compare_neon.cc new file mode 100644 index 000000000..ef006ec41 --- /dev/null +++ b/media/libaom/src/third_party/libyuv/source/compare_neon.cc @@ -0,0 +1,65 @@ +/* + * Copyright 2012 The LibYuv Project Authors. All rights reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#include "libyuv/basic_types.h" +#include "libyuv/row.h" + +#ifdef __cplusplus +namespace libyuv { +extern "C" { +#endif + +#if !defined(LIBYUV_DISABLE_NEON) && defined(__ARM_NEON__) && \ + !defined(__aarch64__) + +uint32 SumSquareError_NEON(const uint8* src_a, const uint8* src_b, int count) { + volatile uint32 sse; + asm volatile ( + "vmov.u8 q8, #0 \n" + "vmov.u8 q10, #0 \n" + "vmov.u8 q9, #0 \n" + "vmov.u8 q11, #0 \n" + + ".p2align 2 \n" + "1: \n" + MEMACCESS(0) + "vld1.8 {q0}, [%0]! \n" + MEMACCESS(1) + "vld1.8 {q1}, [%1]! \n" + "subs %2, %2, #16 \n" + "vsubl.u8 q2, d0, d2 \n" + "vsubl.u8 q3, d1, d3 \n" + "vmlal.s16 q8, d4, d4 \n" + "vmlal.s16 q9, d6, d6 \n" + "vmlal.s16 q10, d5, d5 \n" + "vmlal.s16 q11, d7, d7 \n" + "bgt 1b \n" + + "vadd.u32 q8, q8, q9 \n" + "vadd.u32 q10, q10, q11 \n" + "vadd.u32 q11, q8, q10 \n" + "vpaddl.u32 q1, q11 \n" + "vadd.u64 d0, d2, d3 \n" + "vmov.32 %3, d0[0] \n" + : "+r"(src_a), + "+r"(src_b), + "+r"(count), + "=r"(sse) + : + : "memory", "cc", "q0", "q1", "q2", "q3", "q8", "q9", "q10", "q11"); + return sse; +} + +#endif // defined(__ARM_NEON__) && !defined(__aarch64__) + +#ifdef __cplusplus +} // extern "C" +} // namespace libyuv +#endif diff --git a/media/libaom/src/third_party/libyuv/source/compare_neon64.cc b/media/libaom/src/third_party/libyuv/source/compare_neon64.cc new file mode 100644 index 000000000..6d1e5e1bc --- /dev/null +++ b/media/libaom/src/third_party/libyuv/source/compare_neon64.cc @@ -0,0 +1,63 @@ +/* + * Copyright 2012 The LibYuv Project Authors. All rights reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#include "libyuv/basic_types.h" +#include "libyuv/row.h" + +#ifdef __cplusplus +namespace libyuv { +extern "C" { +#endif + +#if !defined(LIBYUV_DISABLE_NEON) && defined(__aarch64__) + +uint32 SumSquareError_NEON(const uint8* src_a, const uint8* src_b, int count) { + volatile uint32 sse; + asm volatile ( + "eor v16.16b, v16.16b, v16.16b \n" + "eor v18.16b, v18.16b, v18.16b \n" + "eor v17.16b, v17.16b, v17.16b \n" + "eor v19.16b, v19.16b, v19.16b \n" + + ".p2align 2 \n" + "1: \n" + MEMACCESS(0) + "ld1 {v0.16b}, [%0], #16 \n" + MEMACCESS(1) + "ld1 {v1.16b}, [%1], #16 \n" + "subs %w2, %w2, #16 \n" + "usubl v2.8h, v0.8b, v1.8b \n" + "usubl2 v3.8h, v0.16b, v1.16b \n" + "smlal v16.4s, v2.4h, v2.4h \n" + "smlal v17.4s, v3.4h, v3.4h \n" + "smlal2 v18.4s, v2.8h, v2.8h \n" + "smlal2 v19.4s, v3.8h, v3.8h \n" + "b.gt 1b \n" + + "add v16.4s, v16.4s, v17.4s \n" + "add v18.4s, v18.4s, v19.4s \n" + "add v19.4s, v16.4s, v18.4s \n" + "addv s0, v19.4s \n" + "fmov %w3, s0 \n" + : "+r"(src_a), + "+r"(src_b), + "+r"(count), + "=r"(sse) + : + : "cc", "v0", "v1", "v2", "v3", "v16", "v17", "v18", "v19"); + return sse; +} + +#endif // !defined(LIBYUV_DISABLE_NEON) && defined(__aarch64__) + +#ifdef __cplusplus +} // extern "C" +} // namespace libyuv +#endif diff --git a/media/libaom/src/third_party/libyuv/source/compare_win.cc b/media/libaom/src/third_party/libyuv/source/compare_win.cc new file mode 100644 index 000000000..19806f275 --- /dev/null +++ b/media/libaom/src/third_party/libyuv/source/compare_win.cc @@ -0,0 +1,229 @@ +/* + * Copyright 2012 The LibYuv Project Authors. All rights reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#include "libyuv/basic_types.h" +#include "libyuv/row.h" + +#ifdef __cplusplus +namespace libyuv { +extern "C" { +#endif + +// This module is for Visual C x86. +#if !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86) && \ + defined(_MSC_VER) && !defined(__clang__) + +__declspec(naked) +uint32 SumSquareError_SSE2(const uint8* src_a, const uint8* src_b, int count) { + __asm { + mov eax, [esp + 4] // src_a + mov edx, [esp + 8] // src_b + mov ecx, [esp + 12] // count + pxor xmm0, xmm0 + pxor xmm5, xmm5 + + wloop: + movdqu xmm1, [eax] + lea eax, [eax + 16] + movdqu xmm2, [edx] + lea edx, [edx + 16] + movdqa xmm3, xmm1 // abs trick + psubusb xmm1, xmm2 + psubusb xmm2, xmm3 + por xmm1, xmm2 + movdqa xmm2, xmm1 + punpcklbw xmm1, xmm5 + punpckhbw xmm2, xmm5 + pmaddwd xmm1, xmm1 + pmaddwd xmm2, xmm2 + paddd xmm0, xmm1 + paddd xmm0, xmm2 + sub ecx, 16 + jg wloop + + pshufd xmm1, xmm0, 0xee + paddd xmm0, xmm1 + pshufd xmm1, xmm0, 0x01 + paddd xmm0, xmm1 + movd eax, xmm0 + ret + } +} + +// Visual C 2012 required for AVX2. +#if _MSC_VER >= 1700 +// C4752: found Intel(R) Advanced Vector Extensions; consider using /arch:AVX. +#pragma warning(disable: 4752) +__declspec(naked) +uint32 SumSquareError_AVX2(const uint8* src_a, const uint8* src_b, int count) { + __asm { + mov eax, [esp + 4] // src_a + mov edx, [esp + 8] // src_b + mov ecx, [esp + 12] // count + vpxor ymm0, ymm0, ymm0 // sum + vpxor ymm5, ymm5, ymm5 // constant 0 for unpck + sub edx, eax + + wloop: + vmovdqu ymm1, [eax] + vmovdqu ymm2, [eax + edx] + lea eax, [eax + 32] + vpsubusb ymm3, ymm1, ymm2 // abs difference trick + vpsubusb ymm2, ymm2, ymm1 + vpor ymm1, ymm2, ymm3 + vpunpcklbw ymm2, ymm1, ymm5 // u16. mutates order. + vpunpckhbw ymm1, ymm1, ymm5 + vpmaddwd ymm2, ymm2, ymm2 // square + hadd to u32. + vpmaddwd ymm1, ymm1, ymm1 + vpaddd ymm0, ymm0, ymm1 + vpaddd ymm0, ymm0, ymm2 + sub ecx, 32 + jg wloop + + vpshufd ymm1, ymm0, 0xee // 3, 2 + 1, 0 both lanes. + vpaddd ymm0, ymm0, ymm1 + vpshufd ymm1, ymm0, 0x01 // 1 + 0 both lanes. + vpaddd ymm0, ymm0, ymm1 + vpermq ymm1, ymm0, 0x02 // high + low lane. + vpaddd ymm0, ymm0, ymm1 + vmovd eax, xmm0 + vzeroupper + ret + } +} +#endif // _MSC_VER >= 1700 + +#define HAS_HASHDJB2_SSE41 +static uvec32 kHash16x33 = { 0x92d9e201, 0, 0, 0 }; // 33 ^ 16 +static uvec32 kHashMul0 = { + 0x0c3525e1, // 33 ^ 15 + 0xa3476dc1, // 33 ^ 14 + 0x3b4039a1, // 33 ^ 13 + 0x4f5f0981, // 33 ^ 12 +}; +static uvec32 kHashMul1 = { + 0x30f35d61, // 33 ^ 11 + 0x855cb541, // 33 ^ 10 + 0x040a9121, // 33 ^ 9 + 0x747c7101, // 33 ^ 8 +}; +static uvec32 kHashMul2 = { + 0xec41d4e1, // 33 ^ 7 + 0x4cfa3cc1, // 33 ^ 6 + 0x025528a1, // 33 ^ 5 + 0x00121881, // 33 ^ 4 +}; +static uvec32 kHashMul3 = { + 0x00008c61, // 33 ^ 3 + 0x00000441, // 33 ^ 2 + 0x00000021, // 33 ^ 1 + 0x00000001, // 33 ^ 0 +}; + +// 27: 66 0F 38 40 C6 pmulld xmm0,xmm6 +// 44: 66 0F 38 40 DD pmulld xmm3,xmm5 +// 59: 66 0F 38 40 E5 pmulld xmm4,xmm5 +// 72: 66 0F 38 40 D5 pmulld xmm2,xmm5 +// 83: 66 0F 38 40 CD pmulld xmm1,xmm5 +#define pmulld(reg) _asm _emit 0x66 _asm _emit 0x0F _asm _emit 0x38 \ + _asm _emit 0x40 _asm _emit reg + +__declspec(naked) +uint32 HashDjb2_SSE41(const uint8* src, int count, uint32 seed) { + __asm { + mov eax, [esp + 4] // src + mov ecx, [esp + 8] // count + movd xmm0, [esp + 12] // seed + + pxor xmm7, xmm7 // constant 0 for unpck + movdqa xmm6, kHash16x33 + + wloop: + movdqu xmm1, [eax] // src[0-15] + lea eax, [eax + 16] + pmulld(0xc6) // pmulld xmm0,xmm6 hash *= 33 ^ 16 + movdqa xmm5, kHashMul0 + movdqa xmm2, xmm1 + punpcklbw xmm2, xmm7 // src[0-7] + movdqa xmm3, xmm2 + punpcklwd xmm3, xmm7 // src[0-3] + pmulld(0xdd) // pmulld xmm3, xmm5 + movdqa xmm5, kHashMul1 + movdqa xmm4, xmm2 + punpckhwd xmm4, xmm7 // src[4-7] + pmulld(0xe5) // pmulld xmm4, xmm5 + movdqa xmm5, kHashMul2 + punpckhbw xmm1, xmm7 // src[8-15] + movdqa xmm2, xmm1 + punpcklwd xmm2, xmm7 // src[8-11] + pmulld(0xd5) // pmulld xmm2, xmm5 + movdqa xmm5, kHashMul3 + punpckhwd xmm1, xmm7 // src[12-15] + pmulld(0xcd) // pmulld xmm1, xmm5 + paddd xmm3, xmm4 // add 16 results + paddd xmm1, xmm2 + paddd xmm1, xmm3 + + pshufd xmm2, xmm1, 0x0e // upper 2 dwords + paddd xmm1, xmm2 + pshufd xmm2, xmm1, 0x01 + paddd xmm1, xmm2 + paddd xmm0, xmm1 + sub ecx, 16 + jg wloop + + movd eax, xmm0 // return hash + ret + } +} + +// Visual C 2012 required for AVX2. +#if _MSC_VER >= 1700 +__declspec(naked) +uint32 HashDjb2_AVX2(const uint8* src, int count, uint32 seed) { + __asm { + mov eax, [esp + 4] // src + mov ecx, [esp + 8] // count + movd xmm0, [esp + 12] // seed + movdqa xmm6, kHash16x33 + + wloop: + vpmovzxbd xmm3, dword ptr [eax] // src[0-3] + pmulld xmm0, xmm6 // hash *= 33 ^ 16 + vpmovzxbd xmm4, dword ptr [eax + 4] // src[4-7] + pmulld xmm3, kHashMul0 + vpmovzxbd xmm2, dword ptr [eax + 8] // src[8-11] + pmulld xmm4, kHashMul1 + vpmovzxbd xmm1, dword ptr [eax + 12] // src[12-15] + pmulld xmm2, kHashMul2 + lea eax, [eax + 16] + pmulld xmm1, kHashMul3 + paddd xmm3, xmm4 // add 16 results + paddd xmm1, xmm2 + paddd xmm1, xmm3 + pshufd xmm2, xmm1, 0x0e // upper 2 dwords + paddd xmm1, xmm2 + pshufd xmm2, xmm1, 0x01 + paddd xmm1, xmm2 + paddd xmm0, xmm1 + sub ecx, 16 + jg wloop + + movd eax, xmm0 // return hash + ret + } +} +#endif // _MSC_VER >= 1700 +#endif // !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86) + +#ifdef __cplusplus +} // extern "C" +} // namespace libyuv +#endif diff --git a/media/libaom/src/third_party/libyuv/source/convert.cc b/media/libaom/src/third_party/libyuv/source/convert.cc new file mode 100644 index 000000000..3ad6bd7a4 --- /dev/null +++ b/media/libaom/src/third_party/libyuv/source/convert.cc @@ -0,0 +1,1389 @@ +/* + * Copyright 2011 The LibYuv Project Authors. All rights reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#include "libyuv/convert.h" + +#include "libyuv/basic_types.h" +#include "libyuv/cpu_id.h" +#include "libyuv/planar_functions.h" +#include "libyuv/rotate.h" +#include "libyuv/scale.h" // For ScalePlane() +#include "libyuv/row.h" + +#ifdef __cplusplus +namespace libyuv { +extern "C" { +#endif + +#define SUBSAMPLE(v, a, s) (v < 0) ? (-((-v + a) >> s)) : ((v + a) >> s) +static __inline int Abs(int v) { + return v >= 0 ? v : -v; +} + +// Any I4xx To I420 format with mirroring. +static int I4xxToI420(const uint8* src_y, int src_stride_y, + const uint8* src_u, int src_stride_u, + const uint8* src_v, int src_stride_v, + uint8* dst_y, int dst_stride_y, + uint8* dst_u, int dst_stride_u, + uint8* dst_v, int dst_stride_v, + int src_y_width, int src_y_height, + int src_uv_width, int src_uv_height) { + const int dst_y_width = Abs(src_y_width); + const int dst_y_height = Abs(src_y_height); + const int dst_uv_width = SUBSAMPLE(dst_y_width, 1, 1); + const int dst_uv_height = SUBSAMPLE(dst_y_height, 1, 1); + if (src_y_width == 0 || src_y_height == 0 || + src_uv_width == 0 || src_uv_height == 0) { + return -1; + } + ScalePlane(src_y, src_stride_y, src_y_width, src_y_height, + dst_y, dst_stride_y, dst_y_width, dst_y_height, + kFilterBilinear); + ScalePlane(src_u, src_stride_u, src_uv_width, src_uv_height, + dst_u, dst_stride_u, dst_uv_width, dst_uv_height, + kFilterBilinear); + ScalePlane(src_v, src_stride_v, src_uv_width, src_uv_height, + dst_v, dst_stride_v, dst_uv_width, dst_uv_height, + kFilterBilinear); + return 0; +} + +// Copy I420 with optional flipping +// TODO(fbarchard): Use Scale plane which supports mirroring, but ensure +// is does row coalescing. +LIBYUV_API +int I420Copy(const uint8* src_y, int src_stride_y, + const uint8* src_u, int src_stride_u, + const uint8* src_v, int src_stride_v, + uint8* dst_y, int dst_stride_y, + uint8* dst_u, int dst_stride_u, + uint8* dst_v, int dst_stride_v, + int width, int height) { + int halfwidth = (width + 1) >> 1; + int halfheight = (height + 1) >> 1; + if (!src_y || !src_u || !src_v || + !dst_y || !dst_u || !dst_v || + width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + halfheight = (height + 1) >> 1; + src_y = src_y + (height - 1) * src_stride_y; + src_u = src_u + (halfheight - 1) * src_stride_u; + src_v = src_v + (halfheight - 1) * src_stride_v; + src_stride_y = -src_stride_y; + src_stride_u = -src_stride_u; + src_stride_v = -src_stride_v; + } + + if (dst_y) { + CopyPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height); + } + // Copy UV planes. + CopyPlane(src_u, src_stride_u, dst_u, dst_stride_u, halfwidth, halfheight); + CopyPlane(src_v, src_stride_v, dst_v, dst_stride_v, halfwidth, halfheight); + return 0; +} + +// 422 chroma is 1/2 width, 1x height +// 420 chroma is 1/2 width, 1/2 height +LIBYUV_API +int I422ToI420(const uint8* src_y, int src_stride_y, + const uint8* src_u, int src_stride_u, + const uint8* src_v, int src_stride_v, + uint8* dst_y, int dst_stride_y, + uint8* dst_u, int dst_stride_u, + uint8* dst_v, int dst_stride_v, + int width, int height) { + const int src_uv_width = SUBSAMPLE(width, 1, 1); + return I4xxToI420(src_y, src_stride_y, + src_u, src_stride_u, + src_v, src_stride_v, + dst_y, dst_stride_y, + dst_u, dst_stride_u, + dst_v, dst_stride_v, + width, height, + src_uv_width, height); +} + +// 444 chroma is 1x width, 1x height +// 420 chroma is 1/2 width, 1/2 height +LIBYUV_API +int I444ToI420(const uint8* src_y, int src_stride_y, + const uint8* src_u, int src_stride_u, + const uint8* src_v, int src_stride_v, + uint8* dst_y, int dst_stride_y, + uint8* dst_u, int dst_stride_u, + uint8* dst_v, int dst_stride_v, + int width, int height) { + return I4xxToI420(src_y, src_stride_y, + src_u, src_stride_u, + src_v, src_stride_v, + dst_y, dst_stride_y, + dst_u, dst_stride_u, + dst_v, dst_stride_v, + width, height, + width, height); +} + +// 411 chroma is 1/4 width, 1x height +// 420 chroma is 1/2 width, 1/2 height +LIBYUV_API +int I411ToI420(const uint8* src_y, int src_stride_y, + const uint8* src_u, int src_stride_u, + const uint8* src_v, int src_stride_v, + uint8* dst_y, int dst_stride_y, + uint8* dst_u, int dst_stride_u, + uint8* dst_v, int dst_stride_v, + int width, int height) { + const int src_uv_width = SUBSAMPLE(width, 3, 2); + return I4xxToI420(src_y, src_stride_y, + src_u, src_stride_u, + src_v, src_stride_v, + dst_y, dst_stride_y, + dst_u, dst_stride_u, + dst_v, dst_stride_v, + width, height, + src_uv_width, height); +} + +// I400 is greyscale typically used in MJPG +LIBYUV_API +int I400ToI420(const uint8* src_y, int src_stride_y, + uint8* dst_y, int dst_stride_y, + uint8* dst_u, int dst_stride_u, + uint8* dst_v, int dst_stride_v, + int width, int height) { + int halfwidth = (width + 1) >> 1; + int halfheight = (height + 1) >> 1; + if (!src_y || !dst_y || !dst_u || !dst_v || + width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + halfheight = (height + 1) >> 1; + src_y = src_y + (height - 1) * src_stride_y; + src_stride_y = -src_stride_y; + } + CopyPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height); + SetPlane(dst_u, dst_stride_u, halfwidth, halfheight, 128); + SetPlane(dst_v, dst_stride_v, halfwidth, halfheight, 128); + return 0; +} + +static void CopyPlane2(const uint8* src, int src_stride_0, int src_stride_1, + uint8* dst, int dst_stride, + int width, int height) { + int y; + void (*CopyRow)(const uint8* src, uint8* dst, int width) = CopyRow_C; +#if defined(HAS_COPYROW_SSE2) + if (TestCpuFlag(kCpuHasSSE2)) { + CopyRow = IS_ALIGNED(width, 32) ? CopyRow_SSE2 : CopyRow_Any_SSE2; + } +#endif +#if defined(HAS_COPYROW_AVX) + if (TestCpuFlag(kCpuHasAVX)) { + CopyRow = IS_ALIGNED(width, 64) ? CopyRow_AVX : CopyRow_Any_AVX; + } +#endif +#if defined(HAS_COPYROW_ERMS) + if (TestCpuFlag(kCpuHasERMS)) { + CopyRow = CopyRow_ERMS; + } +#endif +#if defined(HAS_COPYROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + CopyRow = IS_ALIGNED(width, 32) ? CopyRow_NEON : CopyRow_Any_NEON; + } +#endif +#if defined(HAS_COPYROW_MIPS) + if (TestCpuFlag(kCpuHasMIPS)) { + CopyRow = CopyRow_MIPS; + } +#endif + + // Copy plane + for (y = 0; y < height - 1; y += 2) { + CopyRow(src, dst, width); + CopyRow(src + src_stride_0, dst + dst_stride, width); + src += src_stride_0 + src_stride_1; + dst += dst_stride * 2; + } + if (height & 1) { + CopyRow(src, dst, width); + } +} + +// Support converting from FOURCC_M420 +// Useful for bandwidth constrained transports like USB 1.0 and 2.0 and for +// easy conversion to I420. +// M420 format description: +// M420 is row biplanar 420: 2 rows of Y and 1 row of UV. +// Chroma is half width / half height. (420) +// src_stride_m420 is row planar. Normally this will be the width in pixels. +// The UV plane is half width, but 2 values, so src_stride_m420 applies to +// this as well as the two Y planes. +static int X420ToI420(const uint8* src_y, + int src_stride_y0, int src_stride_y1, + const uint8* src_uv, int src_stride_uv, + uint8* dst_y, int dst_stride_y, + uint8* dst_u, int dst_stride_u, + uint8* dst_v, int dst_stride_v, + int width, int height) { + int y; + int halfwidth = (width + 1) >> 1; + int halfheight = (height + 1) >> 1; + void (*SplitUVRow)(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int pix) = + SplitUVRow_C; + if (!src_y || !src_uv || + !dst_y || !dst_u || !dst_v || + width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + halfheight = (height + 1) >> 1; + dst_y = dst_y + (height - 1) * dst_stride_y; + dst_u = dst_u + (halfheight - 1) * dst_stride_u; + dst_v = dst_v + (halfheight - 1) * dst_stride_v; + dst_stride_y = -dst_stride_y; + dst_stride_u = -dst_stride_u; + dst_stride_v = -dst_stride_v; + } + // Coalesce rows. + if (src_stride_y0 == width && + src_stride_y1 == width && + dst_stride_y == width) { + width *= height; + height = 1; + src_stride_y0 = src_stride_y1 = dst_stride_y = 0; + } + // Coalesce rows. + if (src_stride_uv == halfwidth * 2 && + dst_stride_u == halfwidth && + dst_stride_v == halfwidth) { + halfwidth *= halfheight; + halfheight = 1; + src_stride_uv = dst_stride_u = dst_stride_v = 0; + } +#if defined(HAS_SPLITUVROW_SSE2) + if (TestCpuFlag(kCpuHasSSE2)) { + SplitUVRow = SplitUVRow_Any_SSE2; + if (IS_ALIGNED(halfwidth, 16)) { + SplitUVRow = SplitUVRow_SSE2; + } + } +#endif +#if defined(HAS_SPLITUVROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + SplitUVRow = SplitUVRow_Any_AVX2; + if (IS_ALIGNED(halfwidth, 32)) { + SplitUVRow = SplitUVRow_AVX2; + } + } +#endif +#if defined(HAS_SPLITUVROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + SplitUVRow = SplitUVRow_Any_NEON; + if (IS_ALIGNED(halfwidth, 16)) { + SplitUVRow = SplitUVRow_NEON; + } + } +#endif +#if defined(HAS_SPLITUVROW_MIPS_DSPR2) + if (TestCpuFlag(kCpuHasMIPS_DSPR2) && + IS_ALIGNED(src_uv, 4) && IS_ALIGNED(src_stride_uv, 4) && + IS_ALIGNED(dst_u, 4) && IS_ALIGNED(dst_stride_u, 4) && + IS_ALIGNED(dst_v, 4) && IS_ALIGNED(dst_stride_v, 4)) { + SplitUVRow = SplitUVRow_Any_MIPS_DSPR2; + if (IS_ALIGNED(halfwidth, 16)) { + SplitUVRow = SplitUVRow_MIPS_DSPR2; + } + } +#endif + + if (dst_y) { + if (src_stride_y0 == src_stride_y1) { + CopyPlane(src_y, src_stride_y0, dst_y, dst_stride_y, width, height); + } else { + CopyPlane2(src_y, src_stride_y0, src_stride_y1, dst_y, dst_stride_y, + width, height); + } + } + + for (y = 0; y < halfheight; ++y) { + // Copy a row of UV. + SplitUVRow(src_uv, dst_u, dst_v, halfwidth); + dst_u += dst_stride_u; + dst_v += dst_stride_v; + src_uv += src_stride_uv; + } + return 0; +} + +// Convert NV12 to I420. +LIBYUV_API +int NV12ToI420(const uint8* src_y, int src_stride_y, + const uint8* src_uv, int src_stride_uv, + uint8* dst_y, int dst_stride_y, + uint8* dst_u, int dst_stride_u, + uint8* dst_v, int dst_stride_v, + int width, int height) { + return X420ToI420(src_y, src_stride_y, src_stride_y, + src_uv, src_stride_uv, + dst_y, dst_stride_y, + dst_u, dst_stride_u, + dst_v, dst_stride_v, + width, height); +} + +// Convert NV21 to I420. Same as NV12 but u and v pointers swapped. +LIBYUV_API +int NV21ToI420(const uint8* src_y, int src_stride_y, + const uint8* src_vu, int src_stride_vu, + uint8* dst_y, int dst_stride_y, + uint8* dst_u, int dst_stride_u, + uint8* dst_v, int dst_stride_v, + int width, int height) { + return X420ToI420(src_y, src_stride_y, src_stride_y, + src_vu, src_stride_vu, + dst_y, dst_stride_y, + dst_v, dst_stride_v, + dst_u, dst_stride_u, + width, height); +} + +// Convert M420 to I420. +LIBYUV_API +int M420ToI420(const uint8* src_m420, int src_stride_m420, + uint8* dst_y, int dst_stride_y, + uint8* dst_u, int dst_stride_u, + uint8* dst_v, int dst_stride_v, + int width, int height) { + return X420ToI420(src_m420, src_stride_m420, src_stride_m420 * 2, + src_m420 + src_stride_m420 * 2, src_stride_m420 * 3, + dst_y, dst_stride_y, + dst_u, dst_stride_u, + dst_v, dst_stride_v, + width, height); +} + +// Convert YUY2 to I420. +LIBYUV_API +int YUY2ToI420(const uint8* src_yuy2, int src_stride_yuy2, + uint8* dst_y, int dst_stride_y, + uint8* dst_u, int dst_stride_u, + uint8* dst_v, int dst_stride_v, + int width, int height) { + int y; + void (*YUY2ToUVRow)(const uint8* src_yuy2, int src_stride_yuy2, + uint8* dst_u, uint8* dst_v, int pix) = YUY2ToUVRow_C; + void (*YUY2ToYRow)(const uint8* src_yuy2, + uint8* dst_y, int pix) = YUY2ToYRow_C; + // Negative height means invert the image. + if (height < 0) { + height = -height; + src_yuy2 = src_yuy2 + (height - 1) * src_stride_yuy2; + src_stride_yuy2 = -src_stride_yuy2; + } +#if defined(HAS_YUY2TOYROW_SSE2) + if (TestCpuFlag(kCpuHasSSE2)) { + YUY2ToUVRow = YUY2ToUVRow_Any_SSE2; + YUY2ToYRow = YUY2ToYRow_Any_SSE2; + if (IS_ALIGNED(width, 16)) { + YUY2ToUVRow = YUY2ToUVRow_SSE2; + YUY2ToYRow = YUY2ToYRow_SSE2; + } + } +#endif +#if defined(HAS_YUY2TOYROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + YUY2ToUVRow = YUY2ToUVRow_Any_AVX2; + YUY2ToYRow = YUY2ToYRow_Any_AVX2; + if (IS_ALIGNED(width, 32)) { + YUY2ToUVRow = YUY2ToUVRow_AVX2; + YUY2ToYRow = YUY2ToYRow_AVX2; + } + } +#endif +#if defined(HAS_YUY2TOYROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + YUY2ToYRow = YUY2ToYRow_Any_NEON; + YUY2ToUVRow = YUY2ToUVRow_Any_NEON; + if (IS_ALIGNED(width, 16)) { + YUY2ToYRow = YUY2ToYRow_NEON; + YUY2ToUVRow = YUY2ToUVRow_NEON; + } + } +#endif + + for (y = 0; y < height - 1; y += 2) { + YUY2ToUVRow(src_yuy2, src_stride_yuy2, dst_u, dst_v, width); + YUY2ToYRow(src_yuy2, dst_y, width); + YUY2ToYRow(src_yuy2 + src_stride_yuy2, dst_y + dst_stride_y, width); + src_yuy2 += src_stride_yuy2 * 2; + dst_y += dst_stride_y * 2; + dst_u += dst_stride_u; + dst_v += dst_stride_v; + } + if (height & 1) { + YUY2ToUVRow(src_yuy2, 0, dst_u, dst_v, width); + YUY2ToYRow(src_yuy2, dst_y, width); + } + return 0; +} + +// Convert UYVY to I420. +LIBYUV_API +int UYVYToI420(const uint8* src_uyvy, int src_stride_uyvy, + uint8* dst_y, int dst_stride_y, + uint8* dst_u, int dst_stride_u, + uint8* dst_v, int dst_stride_v, + int width, int height) { + int y; + void (*UYVYToUVRow)(const uint8* src_uyvy, int src_stride_uyvy, + uint8* dst_u, uint8* dst_v, int pix) = UYVYToUVRow_C; + void (*UYVYToYRow)(const uint8* src_uyvy, + uint8* dst_y, int pix) = UYVYToYRow_C; + // Negative height means invert the image. + if (height < 0) { + height = -height; + src_uyvy = src_uyvy + (height - 1) * src_stride_uyvy; + src_stride_uyvy = -src_stride_uyvy; + } +#if defined(HAS_UYVYTOYROW_SSE2) + if (TestCpuFlag(kCpuHasSSE2)) { + UYVYToUVRow = UYVYToUVRow_Any_SSE2; + UYVYToYRow = UYVYToYRow_Any_SSE2; + if (IS_ALIGNED(width, 16)) { + UYVYToUVRow = UYVYToUVRow_SSE2; + UYVYToYRow = UYVYToYRow_SSE2; + } + } +#endif +#if defined(HAS_UYVYTOYROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + UYVYToUVRow = UYVYToUVRow_Any_AVX2; + UYVYToYRow = UYVYToYRow_Any_AVX2; + if (IS_ALIGNED(width, 32)) { + UYVYToUVRow = UYVYToUVRow_AVX2; + UYVYToYRow = UYVYToYRow_AVX2; + } + } +#endif +#if defined(HAS_UYVYTOYROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + UYVYToYRow = UYVYToYRow_Any_NEON; + UYVYToUVRow = UYVYToUVRow_Any_NEON; + if (IS_ALIGNED(width, 16)) { + UYVYToYRow = UYVYToYRow_NEON; + UYVYToUVRow = UYVYToUVRow_NEON; + } + } +#endif + + for (y = 0; y < height - 1; y += 2) { + UYVYToUVRow(src_uyvy, src_stride_uyvy, dst_u, dst_v, width); + UYVYToYRow(src_uyvy, dst_y, width); + UYVYToYRow(src_uyvy + src_stride_uyvy, dst_y + dst_stride_y, width); + src_uyvy += src_stride_uyvy * 2; + dst_y += dst_stride_y * 2; + dst_u += dst_stride_u; + dst_v += dst_stride_v; + } + if (height & 1) { + UYVYToUVRow(src_uyvy, 0, dst_u, dst_v, width); + UYVYToYRow(src_uyvy, dst_y, width); + } + return 0; +} + +// Convert ARGB to I420. +LIBYUV_API +int ARGBToI420(const uint8* src_argb, int src_stride_argb, + uint8* dst_y, int dst_stride_y, + uint8* dst_u, int dst_stride_u, + uint8* dst_v, int dst_stride_v, + int width, int height) { + int y; + void (*ARGBToUVRow)(const uint8* src_argb0, int src_stride_argb, + uint8* dst_u, uint8* dst_v, int width) = ARGBToUVRow_C; + void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int pix) = + ARGBToYRow_C; + if (!src_argb || + !dst_y || !dst_u || !dst_v || + width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + src_argb = src_argb + (height - 1) * src_stride_argb; + src_stride_argb = -src_stride_argb; + } +#if defined(HAS_ARGBTOYROW_SSSE3) && defined(HAS_ARGBTOUVROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + ARGBToUVRow = ARGBToUVRow_Any_SSSE3; + ARGBToYRow = ARGBToYRow_Any_SSSE3; + if (IS_ALIGNED(width, 16)) { + ARGBToUVRow = ARGBToUVRow_SSSE3; + ARGBToYRow = ARGBToYRow_SSSE3; + } + } +#endif +#if defined(HAS_ARGBTOYROW_AVX2) && defined(HAS_ARGBTOUVROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + ARGBToUVRow = ARGBToUVRow_Any_AVX2; + ARGBToYRow = ARGBToYRow_Any_AVX2; + if (IS_ALIGNED(width, 32)) { + ARGBToUVRow = ARGBToUVRow_AVX2; + ARGBToYRow = ARGBToYRow_AVX2; + } + } +#endif +#if defined(HAS_ARGBTOYROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + ARGBToYRow = ARGBToYRow_Any_NEON; + if (IS_ALIGNED(width, 8)) { + ARGBToYRow = ARGBToYRow_NEON; + } + } +#endif +#if defined(HAS_ARGBTOUVROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + ARGBToUVRow = ARGBToUVRow_Any_NEON; + if (IS_ALIGNED(width, 16)) { + ARGBToUVRow = ARGBToUVRow_NEON; + } + } +#endif + + for (y = 0; y < height - 1; y += 2) { + ARGBToUVRow(src_argb, src_stride_argb, dst_u, dst_v, width); + ARGBToYRow(src_argb, dst_y, width); + ARGBToYRow(src_argb + src_stride_argb, dst_y + dst_stride_y, width); + src_argb += src_stride_argb * 2; + dst_y += dst_stride_y * 2; + dst_u += dst_stride_u; + dst_v += dst_stride_v; + } + if (height & 1) { + ARGBToUVRow(src_argb, 0, dst_u, dst_v, width); + ARGBToYRow(src_argb, dst_y, width); + } + return 0; +} + +// Convert BGRA to I420. +LIBYUV_API +int BGRAToI420(const uint8* src_bgra, int src_stride_bgra, + uint8* dst_y, int dst_stride_y, + uint8* dst_u, int dst_stride_u, + uint8* dst_v, int dst_stride_v, + int width, int height) { + int y; + void (*BGRAToUVRow)(const uint8* src_bgra0, int src_stride_bgra, + uint8* dst_u, uint8* dst_v, int width) = BGRAToUVRow_C; + void (*BGRAToYRow)(const uint8* src_bgra, uint8* dst_y, int pix) = + BGRAToYRow_C; + if (!src_bgra || + !dst_y || !dst_u || !dst_v || + width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + src_bgra = src_bgra + (height - 1) * src_stride_bgra; + src_stride_bgra = -src_stride_bgra; + } +#if defined(HAS_BGRATOYROW_SSSE3) && defined(HAS_BGRATOUVROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + BGRAToUVRow = BGRAToUVRow_Any_SSSE3; + BGRAToYRow = BGRAToYRow_Any_SSSE3; + if (IS_ALIGNED(width, 16)) { + BGRAToUVRow = BGRAToUVRow_SSSE3; + BGRAToYRow = BGRAToYRow_SSSE3; + } + } +#endif +#if defined(HAS_BGRATOYROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + BGRAToYRow = BGRAToYRow_Any_NEON; + if (IS_ALIGNED(width, 8)) { + BGRAToYRow = BGRAToYRow_NEON; + } + } +#endif +#if defined(HAS_BGRATOUVROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + BGRAToUVRow = BGRAToUVRow_Any_NEON; + if (IS_ALIGNED(width, 16)) { + BGRAToUVRow = BGRAToUVRow_NEON; + } + } +#endif + + for (y = 0; y < height - 1; y += 2) { + BGRAToUVRow(src_bgra, src_stride_bgra, dst_u, dst_v, width); + BGRAToYRow(src_bgra, dst_y, width); + BGRAToYRow(src_bgra + src_stride_bgra, dst_y + dst_stride_y, width); + src_bgra += src_stride_bgra * 2; + dst_y += dst_stride_y * 2; + dst_u += dst_stride_u; + dst_v += dst_stride_v; + } + if (height & 1) { + BGRAToUVRow(src_bgra, 0, dst_u, dst_v, width); + BGRAToYRow(src_bgra, dst_y, width); + } + return 0; +} + +// Convert ABGR to I420. +LIBYUV_API +int ABGRToI420(const uint8* src_abgr, int src_stride_abgr, + uint8* dst_y, int dst_stride_y, + uint8* dst_u, int dst_stride_u, + uint8* dst_v, int dst_stride_v, + int width, int height) { + int y; + void (*ABGRToUVRow)(const uint8* src_abgr0, int src_stride_abgr, + uint8* dst_u, uint8* dst_v, int width) = ABGRToUVRow_C; + void (*ABGRToYRow)(const uint8* src_abgr, uint8* dst_y, int pix) = + ABGRToYRow_C; + if (!src_abgr || + !dst_y || !dst_u || !dst_v || + width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + src_abgr = src_abgr + (height - 1) * src_stride_abgr; + src_stride_abgr = -src_stride_abgr; + } +#if defined(HAS_ABGRTOYROW_SSSE3) && defined(HAS_ABGRTOUVROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + ABGRToUVRow = ABGRToUVRow_Any_SSSE3; + ABGRToYRow = ABGRToYRow_Any_SSSE3; + if (IS_ALIGNED(width, 16)) { + ABGRToUVRow = ABGRToUVRow_SSSE3; + ABGRToYRow = ABGRToYRow_SSSE3; + } + } +#endif +#if defined(HAS_ABGRTOYROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + ABGRToYRow = ABGRToYRow_Any_NEON; + if (IS_ALIGNED(width, 8)) { + ABGRToYRow = ABGRToYRow_NEON; + } + } +#endif +#if defined(HAS_ABGRTOUVROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + ABGRToUVRow = ABGRToUVRow_Any_NEON; + if (IS_ALIGNED(width, 16)) { + ABGRToUVRow = ABGRToUVRow_NEON; + } + } +#endif + + for (y = 0; y < height - 1; y += 2) { + ABGRToUVRow(src_abgr, src_stride_abgr, dst_u, dst_v, width); + ABGRToYRow(src_abgr, dst_y, width); + ABGRToYRow(src_abgr + src_stride_abgr, dst_y + dst_stride_y, width); + src_abgr += src_stride_abgr * 2; + dst_y += dst_stride_y * 2; + dst_u += dst_stride_u; + dst_v += dst_stride_v; + } + if (height & 1) { + ABGRToUVRow(src_abgr, 0, dst_u, dst_v, width); + ABGRToYRow(src_abgr, dst_y, width); + } + return 0; +} + +// Convert RGBA to I420. +LIBYUV_API +int RGBAToI420(const uint8* src_rgba, int src_stride_rgba, + uint8* dst_y, int dst_stride_y, + uint8* dst_u, int dst_stride_u, + uint8* dst_v, int dst_stride_v, + int width, int height) { + int y; + void (*RGBAToUVRow)(const uint8* src_rgba0, int src_stride_rgba, + uint8* dst_u, uint8* dst_v, int width) = RGBAToUVRow_C; + void (*RGBAToYRow)(const uint8* src_rgba, uint8* dst_y, int pix) = + RGBAToYRow_C; + if (!src_rgba || + !dst_y || !dst_u || !dst_v || + width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + src_rgba = src_rgba + (height - 1) * src_stride_rgba; + src_stride_rgba = -src_stride_rgba; + } +#if defined(HAS_RGBATOYROW_SSSE3) && defined(HAS_RGBATOUVROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + RGBAToUVRow = RGBAToUVRow_Any_SSSE3; + RGBAToYRow = RGBAToYRow_Any_SSSE3; + if (IS_ALIGNED(width, 16)) { + RGBAToUVRow = RGBAToUVRow_SSSE3; + RGBAToYRow = RGBAToYRow_SSSE3; + } + } +#endif +#if defined(HAS_RGBATOYROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + RGBAToYRow = RGBAToYRow_Any_NEON; + if (IS_ALIGNED(width, 8)) { + RGBAToYRow = RGBAToYRow_NEON; + } + } +#endif +#if defined(HAS_RGBATOUVROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + RGBAToUVRow = RGBAToUVRow_Any_NEON; + if (IS_ALIGNED(width, 16)) { + RGBAToUVRow = RGBAToUVRow_NEON; + } + } +#endif + + for (y = 0; y < height - 1; y += 2) { + RGBAToUVRow(src_rgba, src_stride_rgba, dst_u, dst_v, width); + RGBAToYRow(src_rgba, dst_y, width); + RGBAToYRow(src_rgba + src_stride_rgba, dst_y + dst_stride_y, width); + src_rgba += src_stride_rgba * 2; + dst_y += dst_stride_y * 2; + dst_u += dst_stride_u; + dst_v += dst_stride_v; + } + if (height & 1) { + RGBAToUVRow(src_rgba, 0, dst_u, dst_v, width); + RGBAToYRow(src_rgba, dst_y, width); + } + return 0; +} + +// Convert RGB24 to I420. +LIBYUV_API +int RGB24ToI420(const uint8* src_rgb24, int src_stride_rgb24, + uint8* dst_y, int dst_stride_y, + uint8* dst_u, int dst_stride_u, + uint8* dst_v, int dst_stride_v, + int width, int height) { + int y; +#if defined(HAS_RGB24TOYROW_NEON) + void (*RGB24ToUVRow)(const uint8* src_rgb24, int src_stride_rgb24, + uint8* dst_u, uint8* dst_v, int width) = RGB24ToUVRow_C; + void (*RGB24ToYRow)(const uint8* src_rgb24, uint8* dst_y, int pix) = + RGB24ToYRow_C; +#else + void (*RGB24ToARGBRow)(const uint8* src_rgb, uint8* dst_argb, int pix) = + RGB24ToARGBRow_C; + void (*ARGBToUVRow)(const uint8* src_argb0, int src_stride_argb, + uint8* dst_u, uint8* dst_v, int width) = ARGBToUVRow_C; + void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int pix) = + ARGBToYRow_C; +#endif + if (!src_rgb24 || !dst_y || !dst_u || !dst_v || + width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + src_rgb24 = src_rgb24 + (height - 1) * src_stride_rgb24; + src_stride_rgb24 = -src_stride_rgb24; + } + +// Neon version does direct RGB24 to YUV. +#if defined(HAS_RGB24TOYROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + RGB24ToUVRow = RGB24ToUVRow_Any_NEON; + RGB24ToYRow = RGB24ToYRow_Any_NEON; + if (IS_ALIGNED(width, 8)) { + RGB24ToYRow = RGB24ToYRow_NEON; + if (IS_ALIGNED(width, 16)) { + RGB24ToUVRow = RGB24ToUVRow_NEON; + } + } + } +// Other platforms do intermediate conversion from RGB24 to ARGB. +#else +#if defined(HAS_RGB24TOARGBROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + RGB24ToARGBRow = RGB24ToARGBRow_Any_SSSE3; + if (IS_ALIGNED(width, 16)) { + RGB24ToARGBRow = RGB24ToARGBRow_SSSE3; + } + } +#endif +#if defined(HAS_ARGBTOYROW_SSSE3) && defined(HAS_ARGBTOUVROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + ARGBToUVRow = ARGBToUVRow_Any_SSSE3; + ARGBToYRow = ARGBToYRow_Any_SSSE3; + if (IS_ALIGNED(width, 16)) { + ARGBToUVRow = ARGBToUVRow_SSSE3; + ARGBToYRow = ARGBToYRow_SSSE3; + } + } +#endif +#if defined(HAS_ARGBTOYROW_AVX2) && defined(HAS_ARGBTOUVROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + ARGBToUVRow = ARGBToUVRow_Any_AVX2; + ARGBToYRow = ARGBToYRow_Any_AVX2; + if (IS_ALIGNED(width, 32)) { + ARGBToUVRow = ARGBToUVRow_AVX2; + ARGBToYRow = ARGBToYRow_AVX2; + } + } +#endif + { + // Allocate 2 rows of ARGB. + const int kRowSize = (width * 4 + 31) & ~31; + align_buffer_64(row, kRowSize * 2); +#endif + + for (y = 0; y < height - 1; y += 2) { +#if defined(HAS_RGB24TOYROW_NEON) + RGB24ToUVRow(src_rgb24, src_stride_rgb24, dst_u, dst_v, width); + RGB24ToYRow(src_rgb24, dst_y, width); + RGB24ToYRow(src_rgb24 + src_stride_rgb24, dst_y + dst_stride_y, width); +#else + RGB24ToARGBRow(src_rgb24, row, width); + RGB24ToARGBRow(src_rgb24 + src_stride_rgb24, row + kRowSize, width); + ARGBToUVRow(row, kRowSize, dst_u, dst_v, width); + ARGBToYRow(row, dst_y, width); + ARGBToYRow(row + kRowSize, dst_y + dst_stride_y, width); +#endif + src_rgb24 += src_stride_rgb24 * 2; + dst_y += dst_stride_y * 2; + dst_u += dst_stride_u; + dst_v += dst_stride_v; + } + if (height & 1) { +#if defined(HAS_RGB24TOYROW_NEON) + RGB24ToUVRow(src_rgb24, 0, dst_u, dst_v, width); + RGB24ToYRow(src_rgb24, dst_y, width); +#else + RGB24ToARGBRow(src_rgb24, row, width); + ARGBToUVRow(row, 0, dst_u, dst_v, width); + ARGBToYRow(row, dst_y, width); +#endif + } +#if !defined(HAS_RGB24TOYROW_NEON) + free_aligned_buffer_64(row); + } +#endif + return 0; +} + +// Convert RAW to I420. +LIBYUV_API +int RAWToI420(const uint8* src_raw, int src_stride_raw, + uint8* dst_y, int dst_stride_y, + uint8* dst_u, int dst_stride_u, + uint8* dst_v, int dst_stride_v, + int width, int height) { + int y; +#if defined(HAS_RAWTOYROW_NEON) + void (*RAWToUVRow)(const uint8* src_raw, int src_stride_raw, + uint8* dst_u, uint8* dst_v, int width) = RAWToUVRow_C; + void (*RAWToYRow)(const uint8* src_raw, uint8* dst_y, int pix) = + RAWToYRow_C; +#else + void (*RAWToARGBRow)(const uint8* src_rgb, uint8* dst_argb, int pix) = + RAWToARGBRow_C; + void (*ARGBToUVRow)(const uint8* src_argb0, int src_stride_argb, + uint8* dst_u, uint8* dst_v, int width) = ARGBToUVRow_C; + void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int pix) = + ARGBToYRow_C; +#endif + if (!src_raw || !dst_y || !dst_u || !dst_v || + width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + src_raw = src_raw + (height - 1) * src_stride_raw; + src_stride_raw = -src_stride_raw; + } + +// Neon version does direct RAW to YUV. +#if defined(HAS_RAWTOYROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + RAWToUVRow = RAWToUVRow_Any_NEON; + RAWToYRow = RAWToYRow_Any_NEON; + if (IS_ALIGNED(width, 8)) { + RAWToYRow = RAWToYRow_NEON; + if (IS_ALIGNED(width, 16)) { + RAWToUVRow = RAWToUVRow_NEON; + } + } + } +// Other platforms do intermediate conversion from RAW to ARGB. +#else +#if defined(HAS_RAWTOARGBROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + RAWToARGBRow = RAWToARGBRow_Any_SSSE3; + if (IS_ALIGNED(width, 16)) { + RAWToARGBRow = RAWToARGBRow_SSSE3; + } + } +#endif +#if defined(HAS_ARGBTOYROW_SSSE3) && defined(HAS_ARGBTOUVROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + ARGBToUVRow = ARGBToUVRow_Any_SSSE3; + ARGBToYRow = ARGBToYRow_Any_SSSE3; + if (IS_ALIGNED(width, 16)) { + ARGBToUVRow = ARGBToUVRow_SSSE3; + ARGBToYRow = ARGBToYRow_SSSE3; + } + } +#endif +#if defined(HAS_ARGBTOYROW_AVX2) && defined(HAS_ARGBTOUVROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + ARGBToUVRow = ARGBToUVRow_Any_AVX2; + ARGBToYRow = ARGBToYRow_Any_AVX2; + if (IS_ALIGNED(width, 32)) { + ARGBToUVRow = ARGBToUVRow_AVX2; + ARGBToYRow = ARGBToYRow_AVX2; + } + } +#endif + { + // Allocate 2 rows of ARGB. + const int kRowSize = (width * 4 + 31) & ~31; + align_buffer_64(row, kRowSize * 2); +#endif + + for (y = 0; y < height - 1; y += 2) { +#if defined(HAS_RAWTOYROW_NEON) + RAWToUVRow(src_raw, src_stride_raw, dst_u, dst_v, width); + RAWToYRow(src_raw, dst_y, width); + RAWToYRow(src_raw + src_stride_raw, dst_y + dst_stride_y, width); +#else + RAWToARGBRow(src_raw, row, width); + RAWToARGBRow(src_raw + src_stride_raw, row + kRowSize, width); + ARGBToUVRow(row, kRowSize, dst_u, dst_v, width); + ARGBToYRow(row, dst_y, width); + ARGBToYRow(row + kRowSize, dst_y + dst_stride_y, width); +#endif + src_raw += src_stride_raw * 2; + dst_y += dst_stride_y * 2; + dst_u += dst_stride_u; + dst_v += dst_stride_v; + } + if (height & 1) { +#if defined(HAS_RAWTOYROW_NEON) + RAWToUVRow(src_raw, 0, dst_u, dst_v, width); + RAWToYRow(src_raw, dst_y, width); +#else + RAWToARGBRow(src_raw, row, width); + ARGBToUVRow(row, 0, dst_u, dst_v, width); + ARGBToYRow(row, dst_y, width); +#endif + } +#if !defined(HAS_RAWTOYROW_NEON) + free_aligned_buffer_64(row); + } +#endif + return 0; +} + +// Convert RGB565 to I420. +LIBYUV_API +int RGB565ToI420(const uint8* src_rgb565, int src_stride_rgb565, + uint8* dst_y, int dst_stride_y, + uint8* dst_u, int dst_stride_u, + uint8* dst_v, int dst_stride_v, + int width, int height) { + int y; +#if defined(HAS_RGB565TOYROW_NEON) + void (*RGB565ToUVRow)(const uint8* src_rgb565, int src_stride_rgb565, + uint8* dst_u, uint8* dst_v, int width) = RGB565ToUVRow_C; + void (*RGB565ToYRow)(const uint8* src_rgb565, uint8* dst_y, int pix) = + RGB565ToYRow_C; +#else + void (*RGB565ToARGBRow)(const uint8* src_rgb, uint8* dst_argb, int pix) = + RGB565ToARGBRow_C; + void (*ARGBToUVRow)(const uint8* src_argb0, int src_stride_argb, + uint8* dst_u, uint8* dst_v, int width) = ARGBToUVRow_C; + void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int pix) = + ARGBToYRow_C; +#endif + if (!src_rgb565 || !dst_y || !dst_u || !dst_v || + width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + src_rgb565 = src_rgb565 + (height - 1) * src_stride_rgb565; + src_stride_rgb565 = -src_stride_rgb565; + } + +// Neon version does direct RGB565 to YUV. +#if defined(HAS_RGB565TOYROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + RGB565ToUVRow = RGB565ToUVRow_Any_NEON; + RGB565ToYRow = RGB565ToYRow_Any_NEON; + if (IS_ALIGNED(width, 8)) { + RGB565ToYRow = RGB565ToYRow_NEON; + if (IS_ALIGNED(width, 16)) { + RGB565ToUVRow = RGB565ToUVRow_NEON; + } + } + } +// Other platforms do intermediate conversion from RGB565 to ARGB. +#else +#if defined(HAS_RGB565TOARGBROW_SSE2) + if (TestCpuFlag(kCpuHasSSE2)) { + RGB565ToARGBRow = RGB565ToARGBRow_Any_SSE2; + if (IS_ALIGNED(width, 8)) { + RGB565ToARGBRow = RGB565ToARGBRow_SSE2; + } + } +#endif +#if defined(HAS_RGB565TOARGBROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + RGB565ToARGBRow = RGB565ToARGBRow_Any_AVX2; + if (IS_ALIGNED(width, 16)) { + RGB565ToARGBRow = RGB565ToARGBRow_AVX2; + } + } +#endif +#if defined(HAS_ARGBTOYROW_SSSE3) && defined(HAS_ARGBTOUVROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + ARGBToUVRow = ARGBToUVRow_Any_SSSE3; + ARGBToYRow = ARGBToYRow_Any_SSSE3; + if (IS_ALIGNED(width, 16)) { + ARGBToUVRow = ARGBToUVRow_SSSE3; + ARGBToYRow = ARGBToYRow_SSSE3; + } + } +#endif +#if defined(HAS_ARGBTOYROW_AVX2) && defined(HAS_ARGBTOUVROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + ARGBToUVRow = ARGBToUVRow_Any_AVX2; + ARGBToYRow = ARGBToYRow_Any_AVX2; + if (IS_ALIGNED(width, 32)) { + ARGBToUVRow = ARGBToUVRow_AVX2; + ARGBToYRow = ARGBToYRow_AVX2; + } + } +#endif + { + // Allocate 2 rows of ARGB. + const int kRowSize = (width * 4 + 31) & ~31; + align_buffer_64(row, kRowSize * 2); +#endif + + for (y = 0; y < height - 1; y += 2) { +#if defined(HAS_RGB565TOYROW_NEON) + RGB565ToUVRow(src_rgb565, src_stride_rgb565, dst_u, dst_v, width); + RGB565ToYRow(src_rgb565, dst_y, width); + RGB565ToYRow(src_rgb565 + src_stride_rgb565, dst_y + dst_stride_y, width); +#else + RGB565ToARGBRow(src_rgb565, row, width); + RGB565ToARGBRow(src_rgb565 + src_stride_rgb565, row + kRowSize, width); + ARGBToUVRow(row, kRowSize, dst_u, dst_v, width); + ARGBToYRow(row, dst_y, width); + ARGBToYRow(row + kRowSize, dst_y + dst_stride_y, width); +#endif + src_rgb565 += src_stride_rgb565 * 2; + dst_y += dst_stride_y * 2; + dst_u += dst_stride_u; + dst_v += dst_stride_v; + } + if (height & 1) { +#if defined(HAS_RGB565TOYROW_NEON) + RGB565ToUVRow(src_rgb565, 0, dst_u, dst_v, width); + RGB565ToYRow(src_rgb565, dst_y, width); +#else + RGB565ToARGBRow(src_rgb565, row, width); + ARGBToUVRow(row, 0, dst_u, dst_v, width); + ARGBToYRow(row, dst_y, width); +#endif + } +#if !defined(HAS_RGB565TOYROW_NEON) + free_aligned_buffer_64(row); + } +#endif + return 0; +} + +// Convert ARGB1555 to I420. +LIBYUV_API +int ARGB1555ToI420(const uint8* src_argb1555, int src_stride_argb1555, + uint8* dst_y, int dst_stride_y, + uint8* dst_u, int dst_stride_u, + uint8* dst_v, int dst_stride_v, + int width, int height) { + int y; +#if defined(HAS_ARGB1555TOYROW_NEON) + void (*ARGB1555ToUVRow)(const uint8* src_argb1555, int src_stride_argb1555, + uint8* dst_u, uint8* dst_v, int width) = ARGB1555ToUVRow_C; + void (*ARGB1555ToYRow)(const uint8* src_argb1555, uint8* dst_y, int pix) = + ARGB1555ToYRow_C; +#else + void (*ARGB1555ToARGBRow)(const uint8* src_rgb, uint8* dst_argb, int pix) = + ARGB1555ToARGBRow_C; + void (*ARGBToUVRow)(const uint8* src_argb0, int src_stride_argb, + uint8* dst_u, uint8* dst_v, int width) = ARGBToUVRow_C; + void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int pix) = + ARGBToYRow_C; +#endif + if (!src_argb1555 || !dst_y || !dst_u || !dst_v || + width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + src_argb1555 = src_argb1555 + (height - 1) * src_stride_argb1555; + src_stride_argb1555 = -src_stride_argb1555; + } + +// Neon version does direct ARGB1555 to YUV. +#if defined(HAS_ARGB1555TOYROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + ARGB1555ToUVRow = ARGB1555ToUVRow_Any_NEON; + ARGB1555ToYRow = ARGB1555ToYRow_Any_NEON; + if (IS_ALIGNED(width, 8)) { + ARGB1555ToYRow = ARGB1555ToYRow_NEON; + if (IS_ALIGNED(width, 16)) { + ARGB1555ToUVRow = ARGB1555ToUVRow_NEON; + } + } + } +// Other platforms do intermediate conversion from ARGB1555 to ARGB. +#else +#if defined(HAS_ARGB1555TOARGBROW_SSE2) + if (TestCpuFlag(kCpuHasSSE2)) { + ARGB1555ToARGBRow = ARGB1555ToARGBRow_Any_SSE2; + if (IS_ALIGNED(width, 8)) { + ARGB1555ToARGBRow = ARGB1555ToARGBRow_SSE2; + } + } +#endif +#if defined(HAS_ARGB1555TOARGBROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + ARGB1555ToARGBRow = ARGB1555ToARGBRow_Any_AVX2; + if (IS_ALIGNED(width, 16)) { + ARGB1555ToARGBRow = ARGB1555ToARGBRow_AVX2; + } + } +#endif +#if defined(HAS_ARGBTOYROW_SSSE3) && defined(HAS_ARGBTOUVROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + ARGBToUVRow = ARGBToUVRow_Any_SSSE3; + ARGBToYRow = ARGBToYRow_Any_SSSE3; + if (IS_ALIGNED(width, 16)) { + ARGBToUVRow = ARGBToUVRow_SSSE3; + ARGBToYRow = ARGBToYRow_SSSE3; + } + } +#endif +#if defined(HAS_ARGBTOYROW_AVX2) && defined(HAS_ARGBTOUVROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + ARGBToUVRow = ARGBToUVRow_Any_AVX2; + ARGBToYRow = ARGBToYRow_Any_AVX2; + if (IS_ALIGNED(width, 32)) { + ARGBToUVRow = ARGBToUVRow_AVX2; + ARGBToYRow = ARGBToYRow_AVX2; + } + } +#endif + { + // Allocate 2 rows of ARGB. + const int kRowSize = (width * 4 + 31) & ~31; + align_buffer_64(row, kRowSize * 2); +#endif + + for (y = 0; y < height - 1; y += 2) { +#if defined(HAS_ARGB1555TOYROW_NEON) + ARGB1555ToUVRow(src_argb1555, src_stride_argb1555, dst_u, dst_v, width); + ARGB1555ToYRow(src_argb1555, dst_y, width); + ARGB1555ToYRow(src_argb1555 + src_stride_argb1555, dst_y + dst_stride_y, + width); +#else + ARGB1555ToARGBRow(src_argb1555, row, width); + ARGB1555ToARGBRow(src_argb1555 + src_stride_argb1555, row + kRowSize, + width); + ARGBToUVRow(row, kRowSize, dst_u, dst_v, width); + ARGBToYRow(row, dst_y, width); + ARGBToYRow(row + kRowSize, dst_y + dst_stride_y, width); +#endif + src_argb1555 += src_stride_argb1555 * 2; + dst_y += dst_stride_y * 2; + dst_u += dst_stride_u; + dst_v += dst_stride_v; + } + if (height & 1) { +#if defined(HAS_ARGB1555TOYROW_NEON) + ARGB1555ToUVRow(src_argb1555, 0, dst_u, dst_v, width); + ARGB1555ToYRow(src_argb1555, dst_y, width); +#else + ARGB1555ToARGBRow(src_argb1555, row, width); + ARGBToUVRow(row, 0, dst_u, dst_v, width); + ARGBToYRow(row, dst_y, width); +#endif + } +#if !defined(HAS_ARGB1555TOYROW_NEON) + free_aligned_buffer_64(row); + } +#endif + return 0; +} + +// Convert ARGB4444 to I420. +LIBYUV_API +int ARGB4444ToI420(const uint8* src_argb4444, int src_stride_argb4444, + uint8* dst_y, int dst_stride_y, + uint8* dst_u, int dst_stride_u, + uint8* dst_v, int dst_stride_v, + int width, int height) { + int y; +#if defined(HAS_ARGB4444TOYROW_NEON) + void (*ARGB4444ToUVRow)(const uint8* src_argb4444, int src_stride_argb4444, + uint8* dst_u, uint8* dst_v, int width) = ARGB4444ToUVRow_C; + void (*ARGB4444ToYRow)(const uint8* src_argb4444, uint8* dst_y, int pix) = + ARGB4444ToYRow_C; +#else + void (*ARGB4444ToARGBRow)(const uint8* src_rgb, uint8* dst_argb, int pix) = + ARGB4444ToARGBRow_C; + void (*ARGBToUVRow)(const uint8* src_argb0, int src_stride_argb, + uint8* dst_u, uint8* dst_v, int width) = ARGBToUVRow_C; + void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int pix) = + ARGBToYRow_C; +#endif + if (!src_argb4444 || !dst_y || !dst_u || !dst_v || + width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + src_argb4444 = src_argb4444 + (height - 1) * src_stride_argb4444; + src_stride_argb4444 = -src_stride_argb4444; + } + +// Neon version does direct ARGB4444 to YUV. +#if defined(HAS_ARGB4444TOYROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + ARGB4444ToUVRow = ARGB4444ToUVRow_Any_NEON; + ARGB4444ToYRow = ARGB4444ToYRow_Any_NEON; + if (IS_ALIGNED(width, 8)) { + ARGB4444ToYRow = ARGB4444ToYRow_NEON; + if (IS_ALIGNED(width, 16)) { + ARGB4444ToUVRow = ARGB4444ToUVRow_NEON; + } + } + } +// Other platforms do intermediate conversion from ARGB4444 to ARGB. +#else +#if defined(HAS_ARGB4444TOARGBROW_SSE2) + if (TestCpuFlag(kCpuHasSSE2)) { + ARGB4444ToARGBRow = ARGB4444ToARGBRow_Any_SSE2; + if (IS_ALIGNED(width, 8)) { + ARGB4444ToARGBRow = ARGB4444ToARGBRow_SSE2; + } + } +#endif +#if defined(HAS_ARGB4444TOARGBROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + ARGB4444ToARGBRow = ARGB4444ToARGBRow_Any_AVX2; + if (IS_ALIGNED(width, 16)) { + ARGB4444ToARGBRow = ARGB4444ToARGBRow_AVX2; + } + } +#endif +#if defined(HAS_ARGBTOYROW_SSSE3) && defined(HAS_ARGBTOUVROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + ARGBToUVRow = ARGBToUVRow_Any_SSSE3; + ARGBToYRow = ARGBToYRow_Any_SSSE3; + if (IS_ALIGNED(width, 16)) { + ARGBToUVRow = ARGBToUVRow_SSSE3; + ARGBToYRow = ARGBToYRow_SSSE3; + } + } +#endif +#if defined(HAS_ARGBTOYROW_AVX2) && defined(HAS_ARGBTOUVROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + ARGBToUVRow = ARGBToUVRow_Any_AVX2; + ARGBToYRow = ARGBToYRow_Any_AVX2; + if (IS_ALIGNED(width, 32)) { + ARGBToUVRow = ARGBToUVRow_AVX2; + ARGBToYRow = ARGBToYRow_AVX2; + } + } +#endif + { + // Allocate 2 rows of ARGB. + const int kRowSize = (width * 4 + 31) & ~31; + align_buffer_64(row, kRowSize * 2); +#endif + + for (y = 0; y < height - 1; y += 2) { +#if defined(HAS_ARGB4444TOYROW_NEON) + ARGB4444ToUVRow(src_argb4444, src_stride_argb4444, dst_u, dst_v, width); + ARGB4444ToYRow(src_argb4444, dst_y, width); + ARGB4444ToYRow(src_argb4444 + src_stride_argb4444, dst_y + dst_stride_y, + width); +#else + ARGB4444ToARGBRow(src_argb4444, row, width); + ARGB4444ToARGBRow(src_argb4444 + src_stride_argb4444, row + kRowSize, + width); + ARGBToUVRow(row, kRowSize, dst_u, dst_v, width); + ARGBToYRow(row, dst_y, width); + ARGBToYRow(row + kRowSize, dst_y + dst_stride_y, width); +#endif + src_argb4444 += src_stride_argb4444 * 2; + dst_y += dst_stride_y * 2; + dst_u += dst_stride_u; + dst_v += dst_stride_v; + } + if (height & 1) { +#if defined(HAS_ARGB4444TOYROW_NEON) + ARGB4444ToUVRow(src_argb4444, 0, dst_u, dst_v, width); + ARGB4444ToYRow(src_argb4444, dst_y, width); +#else + ARGB4444ToARGBRow(src_argb4444, row, width); + ARGBToUVRow(row, 0, dst_u, dst_v, width); + ARGBToYRow(row, dst_y, width); +#endif + } +#if !defined(HAS_ARGB4444TOYROW_NEON) + free_aligned_buffer_64(row); + } +#endif + return 0; +} + +#ifdef __cplusplus +} // extern "C" +} // namespace libyuv +#endif diff --git a/media/libaom/src/third_party/libyuv/source/convert_argb.cc b/media/libaom/src/third_party/libyuv/source/convert_argb.cc new file mode 100644 index 000000000..44756bc41 --- /dev/null +++ b/media/libaom/src/third_party/libyuv/source/convert_argb.cc @@ -0,0 +1,1155 @@ +/* + * Copyright 2011 The LibYuv Project Authors. All rights reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#include "libyuv/convert_argb.h" + +#include "libyuv/cpu_id.h" +#ifdef HAVE_JPEG +#include "libyuv/mjpeg_decoder.h" +#endif +#include "libyuv/rotate_argb.h" +#include "libyuv/row.h" +#include "libyuv/video_common.h" + +#ifdef __cplusplus +namespace libyuv { +extern "C" { +#endif + +// Copy ARGB with optional flipping +LIBYUV_API +int ARGBCopy(const uint8* src_argb, int src_stride_argb, + uint8* dst_argb, int dst_stride_argb, + int width, int height) { + if (!src_argb || !dst_argb || + width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + src_argb = src_argb + (height - 1) * src_stride_argb; + src_stride_argb = -src_stride_argb; + } + + CopyPlane(src_argb, src_stride_argb, dst_argb, dst_stride_argb, + width * 4, height); + return 0; +} + +// Convert I444 to ARGB. +LIBYUV_API +int I444ToARGB(const uint8* src_y, int src_stride_y, + const uint8* src_u, int src_stride_u, + const uint8* src_v, int src_stride_v, + uint8* dst_argb, int dst_stride_argb, + int width, int height) { + int y; + void (*I444ToARGBRow)(const uint8* y_buf, + const uint8* u_buf, + const uint8* v_buf, + uint8* rgb_buf, + int width) = I444ToARGBRow_C; + if (!src_y || !src_u || !src_v || + !dst_argb || + width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + dst_argb = dst_argb + (height - 1) * dst_stride_argb; + dst_stride_argb = -dst_stride_argb; + } + // Coalesce rows. + if (src_stride_y == width && + src_stride_u == width && + src_stride_v == width && + dst_stride_argb == width * 4) { + width *= height; + height = 1; + src_stride_y = src_stride_u = src_stride_v = dst_stride_argb = 0; + } +#if defined(HAS_I444TOARGBROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + I444ToARGBRow = I444ToARGBRow_Any_SSSE3; + if (IS_ALIGNED(width, 8)) { + I444ToARGBRow = I444ToARGBRow_SSSE3; + } + } +#endif +#if defined(HAS_I444TOARGBROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + I444ToARGBRow = I444ToARGBRow_Any_AVX2; + if (IS_ALIGNED(width, 16)) { + I444ToARGBRow = I444ToARGBRow_AVX2; + } + } +#endif +#if defined(HAS_I444TOARGBROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + I444ToARGBRow = I444ToARGBRow_Any_NEON; + if (IS_ALIGNED(width, 8)) { + I444ToARGBRow = I444ToARGBRow_NEON; + } + } +#endif + + for (y = 0; y < height; ++y) { + I444ToARGBRow(src_y, src_u, src_v, dst_argb, width); + dst_argb += dst_stride_argb; + src_y += src_stride_y; + src_u += src_stride_u; + src_v += src_stride_v; + } + return 0; +} + +// Convert I422 to ARGB. +LIBYUV_API +int I422ToARGB(const uint8* src_y, int src_stride_y, + const uint8* src_u, int src_stride_u, + const uint8* src_v, int src_stride_v, + uint8* dst_argb, int dst_stride_argb, + int width, int height) { + int y; + void (*I422ToARGBRow)(const uint8* y_buf, + const uint8* u_buf, + const uint8* v_buf, + uint8* rgb_buf, + int width) = I422ToARGBRow_C; + if (!src_y || !src_u || !src_v || + !dst_argb || + width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + dst_argb = dst_argb + (height - 1) * dst_stride_argb; + dst_stride_argb = -dst_stride_argb; + } + // Coalesce rows. + if (src_stride_y == width && + src_stride_u * 2 == width && + src_stride_v * 2 == width && + dst_stride_argb == width * 4) { + width *= height; + height = 1; + src_stride_y = src_stride_u = src_stride_v = dst_stride_argb = 0; + } +#if defined(HAS_I422TOARGBROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + I422ToARGBRow = I422ToARGBRow_Any_SSSE3; + if (IS_ALIGNED(width, 8)) { + I422ToARGBRow = I422ToARGBRow_SSSE3; + } + } +#endif +#if defined(HAS_I422TOARGBROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + I422ToARGBRow = I422ToARGBRow_Any_AVX2; + if (IS_ALIGNED(width, 16)) { + I422ToARGBRow = I422ToARGBRow_AVX2; + } + } +#endif +#if defined(HAS_I422TOARGBROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + I422ToARGBRow = I422ToARGBRow_Any_NEON; + if (IS_ALIGNED(width, 8)) { + I422ToARGBRow = I422ToARGBRow_NEON; + } + } +#endif +#if defined(HAS_I422TOARGBROW_MIPS_DSPR2) + if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(width, 4) && + IS_ALIGNED(src_y, 4) && IS_ALIGNED(src_stride_y, 4) && + IS_ALIGNED(src_u, 2) && IS_ALIGNED(src_stride_u, 2) && + IS_ALIGNED(src_v, 2) && IS_ALIGNED(src_stride_v, 2) && + IS_ALIGNED(dst_argb, 4) && IS_ALIGNED(dst_stride_argb, 4)) { + I422ToARGBRow = I422ToARGBRow_MIPS_DSPR2; + } +#endif + + for (y = 0; y < height; ++y) { + I422ToARGBRow(src_y, src_u, src_v, dst_argb, width); + dst_argb += dst_stride_argb; + src_y += src_stride_y; + src_u += src_stride_u; + src_v += src_stride_v; + } + return 0; +} + +// Convert I411 to ARGB. +LIBYUV_API +int I411ToARGB(const uint8* src_y, int src_stride_y, + const uint8* src_u, int src_stride_u, + const uint8* src_v, int src_stride_v, + uint8* dst_argb, int dst_stride_argb, + int width, int height) { + int y; + void (*I411ToARGBRow)(const uint8* y_buf, + const uint8* u_buf, + const uint8* v_buf, + uint8* rgb_buf, + int width) = I411ToARGBRow_C; + if (!src_y || !src_u || !src_v || + !dst_argb || + width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + dst_argb = dst_argb + (height - 1) * dst_stride_argb; + dst_stride_argb = -dst_stride_argb; + } + // Coalesce rows. + if (src_stride_y == width && + src_stride_u * 4 == width && + src_stride_v * 4 == width && + dst_stride_argb == width * 4) { + width *= height; + height = 1; + src_stride_y = src_stride_u = src_stride_v = dst_stride_argb = 0; + } +#if defined(HAS_I411TOARGBROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + I411ToARGBRow = I411ToARGBRow_Any_SSSE3; + if (IS_ALIGNED(width, 8)) { + I411ToARGBRow = I411ToARGBRow_SSSE3; + } + } +#endif +#if defined(HAS_I411TOARGBROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + I411ToARGBRow = I411ToARGBRow_Any_AVX2; + if (IS_ALIGNED(width, 16)) { + I411ToARGBRow = I411ToARGBRow_AVX2; + } + } +#endif +#if defined(HAS_I411TOARGBROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + I411ToARGBRow = I411ToARGBRow_Any_NEON; + if (IS_ALIGNED(width, 8)) { + I411ToARGBRow = I411ToARGBRow_NEON; + } + } +#endif + + for (y = 0; y < height; ++y) { + I411ToARGBRow(src_y, src_u, src_v, dst_argb, width); + dst_argb += dst_stride_argb; + src_y += src_stride_y; + src_u += src_stride_u; + src_v += src_stride_v; + } + return 0; +} + +// Convert I400 to ARGB. +LIBYUV_API +int I400ToARGB(const uint8* src_y, int src_stride_y, + uint8* dst_argb, int dst_stride_argb, + int width, int height) { + int y; + void (*I400ToARGBRow)(const uint8* y_buf, + uint8* rgb_buf, + int width) = I400ToARGBRow_C; + if (!src_y || !dst_argb || + width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + dst_argb = dst_argb + (height - 1) * dst_stride_argb; + dst_stride_argb = -dst_stride_argb; + } + // Coalesce rows. + if (src_stride_y == width && + dst_stride_argb == width * 4) { + width *= height; + height = 1; + src_stride_y = dst_stride_argb = 0; + } +#if defined(HAS_I400TOARGBROW_SSE2) + if (TestCpuFlag(kCpuHasSSE2)) { + I400ToARGBRow = I400ToARGBRow_Any_SSE2; + if (IS_ALIGNED(width, 8)) { + I400ToARGBRow = I400ToARGBRow_SSE2; + } + } +#endif +#if defined(HAS_I400TOARGBROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + I400ToARGBRow = I400ToARGBRow_Any_AVX2; + if (IS_ALIGNED(width, 16)) { + I400ToARGBRow = I400ToARGBRow_AVX2; + } + } +#endif +#if defined(HAS_I400TOARGBROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + I400ToARGBRow = I400ToARGBRow_Any_NEON; + if (IS_ALIGNED(width, 8)) { + I400ToARGBRow = I400ToARGBRow_NEON; + } + } +#endif + + for (y = 0; y < height; ++y) { + I400ToARGBRow(src_y, dst_argb, width); + dst_argb += dst_stride_argb; + src_y += src_stride_y; + } + return 0; +} + +// Convert J400 to ARGB. +LIBYUV_API +int J400ToARGB(const uint8* src_y, int src_stride_y, + uint8* dst_argb, int dst_stride_argb, + int width, int height) { + int y; + void (*J400ToARGBRow)(const uint8* src_y, uint8* dst_argb, int pix) = + J400ToARGBRow_C; + if (!src_y || !dst_argb || + width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + src_y = src_y + (height - 1) * src_stride_y; + src_stride_y = -src_stride_y; + } + // Coalesce rows. + if (src_stride_y == width && + dst_stride_argb == width * 4) { + width *= height; + height = 1; + src_stride_y = dst_stride_argb = 0; + } +#if defined(HAS_J400TOARGBROW_SSE2) + if (TestCpuFlag(kCpuHasSSE2)) { + J400ToARGBRow = J400ToARGBRow_Any_SSE2; + if (IS_ALIGNED(width, 8)) { + J400ToARGBRow = J400ToARGBRow_SSE2; + } + } +#endif +#if defined(HAS_J400TOARGBROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + J400ToARGBRow = J400ToARGBRow_Any_AVX2; + if (IS_ALIGNED(width, 16)) { + J400ToARGBRow = J400ToARGBRow_AVX2; + } + } +#endif +#if defined(HAS_J400TOARGBROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + J400ToARGBRow = J400ToARGBRow_Any_NEON; + if (IS_ALIGNED(width, 8)) { + J400ToARGBRow = J400ToARGBRow_NEON; + } + } +#endif + for (y = 0; y < height; ++y) { + J400ToARGBRow(src_y, dst_argb, width); + src_y += src_stride_y; + dst_argb += dst_stride_argb; + } + return 0; +} + +// Shuffle table for converting BGRA to ARGB. +static uvec8 kShuffleMaskBGRAToARGB = { + 3u, 2u, 1u, 0u, 7u, 6u, 5u, 4u, 11u, 10u, 9u, 8u, 15u, 14u, 13u, 12u +}; + +// Shuffle table for converting ABGR to ARGB. +static uvec8 kShuffleMaskABGRToARGB = { + 2u, 1u, 0u, 3u, 6u, 5u, 4u, 7u, 10u, 9u, 8u, 11u, 14u, 13u, 12u, 15u +}; + +// Shuffle table for converting RGBA to ARGB. +static uvec8 kShuffleMaskRGBAToARGB = { + 1u, 2u, 3u, 0u, 5u, 6u, 7u, 4u, 9u, 10u, 11u, 8u, 13u, 14u, 15u, 12u +}; + +// Convert BGRA to ARGB. +LIBYUV_API +int BGRAToARGB(const uint8* src_bgra, int src_stride_bgra, + uint8* dst_argb, int dst_stride_argb, + int width, int height) { + return ARGBShuffle(src_bgra, src_stride_bgra, + dst_argb, dst_stride_argb, + (const uint8*)(&kShuffleMaskBGRAToARGB), + width, height); +} + +// Convert ARGB to BGRA (same as BGRAToARGB). +LIBYUV_API +int ARGBToBGRA(const uint8* src_bgra, int src_stride_bgra, + uint8* dst_argb, int dst_stride_argb, + int width, int height) { + return ARGBShuffle(src_bgra, src_stride_bgra, + dst_argb, dst_stride_argb, + (const uint8*)(&kShuffleMaskBGRAToARGB), + width, height); +} + +// Convert ABGR to ARGB. +LIBYUV_API +int ABGRToARGB(const uint8* src_abgr, int src_stride_abgr, + uint8* dst_argb, int dst_stride_argb, + int width, int height) { + return ARGBShuffle(src_abgr, src_stride_abgr, + dst_argb, dst_stride_argb, + (const uint8*)(&kShuffleMaskABGRToARGB), + width, height); +} + +// Convert ARGB to ABGR to (same as ABGRToARGB). +LIBYUV_API +int ARGBToABGR(const uint8* src_abgr, int src_stride_abgr, + uint8* dst_argb, int dst_stride_argb, + int width, int height) { + return ARGBShuffle(src_abgr, src_stride_abgr, + dst_argb, dst_stride_argb, + (const uint8*)(&kShuffleMaskABGRToARGB), + width, height); +} + +// Convert RGBA to ARGB. +LIBYUV_API +int RGBAToARGB(const uint8* src_rgba, int src_stride_rgba, + uint8* dst_argb, int dst_stride_argb, + int width, int height) { + return ARGBShuffle(src_rgba, src_stride_rgba, + dst_argb, dst_stride_argb, + (const uint8*)(&kShuffleMaskRGBAToARGB), + width, height); +} + +// Convert RGB24 to ARGB. +LIBYUV_API +int RGB24ToARGB(const uint8* src_rgb24, int src_stride_rgb24, + uint8* dst_argb, int dst_stride_argb, + int width, int height) { + int y; + void (*RGB24ToARGBRow)(const uint8* src_rgb, uint8* dst_argb, int pix) = + RGB24ToARGBRow_C; + if (!src_rgb24 || !dst_argb || + width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + src_rgb24 = src_rgb24 + (height - 1) * src_stride_rgb24; + src_stride_rgb24 = -src_stride_rgb24; + } + // Coalesce rows. + if (src_stride_rgb24 == width * 3 && + dst_stride_argb == width * 4) { + width *= height; + height = 1; + src_stride_rgb24 = dst_stride_argb = 0; + } +#if defined(HAS_RGB24TOARGBROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + RGB24ToARGBRow = RGB24ToARGBRow_Any_SSSE3; + if (IS_ALIGNED(width, 16)) { + RGB24ToARGBRow = RGB24ToARGBRow_SSSE3; + } + } +#endif +#if defined(HAS_RGB24TOARGBROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + RGB24ToARGBRow = RGB24ToARGBRow_Any_NEON; + if (IS_ALIGNED(width, 8)) { + RGB24ToARGBRow = RGB24ToARGBRow_NEON; + } + } +#endif + + for (y = 0; y < height; ++y) { + RGB24ToARGBRow(src_rgb24, dst_argb, width); + src_rgb24 += src_stride_rgb24; + dst_argb += dst_stride_argb; + } + return 0; +} + +// Convert RAW to ARGB. +LIBYUV_API +int RAWToARGB(const uint8* src_raw, int src_stride_raw, + uint8* dst_argb, int dst_stride_argb, + int width, int height) { + int y; + void (*RAWToARGBRow)(const uint8* src_rgb, uint8* dst_argb, int pix) = + RAWToARGBRow_C; + if (!src_raw || !dst_argb || + width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + src_raw = src_raw + (height - 1) * src_stride_raw; + src_stride_raw = -src_stride_raw; + } + // Coalesce rows. + if (src_stride_raw == width * 3 && + dst_stride_argb == width * 4) { + width *= height; + height = 1; + src_stride_raw = dst_stride_argb = 0; + } +#if defined(HAS_RAWTOARGBROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + RAWToARGBRow = RAWToARGBRow_Any_SSSE3; + if (IS_ALIGNED(width, 16)) { + RAWToARGBRow = RAWToARGBRow_SSSE3; + } + } +#endif +#if defined(HAS_RAWTOARGBROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + RAWToARGBRow = RAWToARGBRow_Any_NEON; + if (IS_ALIGNED(width, 8)) { + RAWToARGBRow = RAWToARGBRow_NEON; + } + } +#endif + + for (y = 0; y < height; ++y) { + RAWToARGBRow(src_raw, dst_argb, width); + src_raw += src_stride_raw; + dst_argb += dst_stride_argb; + } + return 0; +} + +// Convert RGB565 to ARGB. +LIBYUV_API +int RGB565ToARGB(const uint8* src_rgb565, int src_stride_rgb565, + uint8* dst_argb, int dst_stride_argb, + int width, int height) { + int y; + void (*RGB565ToARGBRow)(const uint8* src_rgb565, uint8* dst_argb, int pix) = + RGB565ToARGBRow_C; + if (!src_rgb565 || !dst_argb || + width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + src_rgb565 = src_rgb565 + (height - 1) * src_stride_rgb565; + src_stride_rgb565 = -src_stride_rgb565; + } + // Coalesce rows. + if (src_stride_rgb565 == width * 2 && + dst_stride_argb == width * 4) { + width *= height; + height = 1; + src_stride_rgb565 = dst_stride_argb = 0; + } +#if defined(HAS_RGB565TOARGBROW_SSE2) + if (TestCpuFlag(kCpuHasSSE2)) { + RGB565ToARGBRow = RGB565ToARGBRow_Any_SSE2; + if (IS_ALIGNED(width, 8)) { + RGB565ToARGBRow = RGB565ToARGBRow_SSE2; + } + } +#endif +#if defined(HAS_RGB565TOARGBROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + RGB565ToARGBRow = RGB565ToARGBRow_Any_AVX2; + if (IS_ALIGNED(width, 16)) { + RGB565ToARGBRow = RGB565ToARGBRow_AVX2; + } + } +#endif +#if defined(HAS_RGB565TOARGBROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + RGB565ToARGBRow = RGB565ToARGBRow_Any_NEON; + if (IS_ALIGNED(width, 8)) { + RGB565ToARGBRow = RGB565ToARGBRow_NEON; + } + } +#endif + + for (y = 0; y < height; ++y) { + RGB565ToARGBRow(src_rgb565, dst_argb, width); + src_rgb565 += src_stride_rgb565; + dst_argb += dst_stride_argb; + } + return 0; +} + +// Convert ARGB1555 to ARGB. +LIBYUV_API +int ARGB1555ToARGB(const uint8* src_argb1555, int src_stride_argb1555, + uint8* dst_argb, int dst_stride_argb, + int width, int height) { + int y; + void (*ARGB1555ToARGBRow)(const uint8* src_argb1555, uint8* dst_argb, + int pix) = ARGB1555ToARGBRow_C; + if (!src_argb1555 || !dst_argb || + width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + src_argb1555 = src_argb1555 + (height - 1) * src_stride_argb1555; + src_stride_argb1555 = -src_stride_argb1555; + } + // Coalesce rows. + if (src_stride_argb1555 == width * 2 && + dst_stride_argb == width * 4) { + width *= height; + height = 1; + src_stride_argb1555 = dst_stride_argb = 0; + } +#if defined(HAS_ARGB1555TOARGBROW_SSE2) + if (TestCpuFlag(kCpuHasSSE2)) { + ARGB1555ToARGBRow = ARGB1555ToARGBRow_Any_SSE2; + if (IS_ALIGNED(width, 8)) { + ARGB1555ToARGBRow = ARGB1555ToARGBRow_SSE2; + } + } +#endif +#if defined(HAS_ARGB1555TOARGBROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + ARGB1555ToARGBRow = ARGB1555ToARGBRow_Any_AVX2; + if (IS_ALIGNED(width, 16)) { + ARGB1555ToARGBRow = ARGB1555ToARGBRow_AVX2; + } + } +#endif +#if defined(HAS_ARGB1555TOARGBROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + ARGB1555ToARGBRow = ARGB1555ToARGBRow_Any_NEON; + if (IS_ALIGNED(width, 8)) { + ARGB1555ToARGBRow = ARGB1555ToARGBRow_NEON; + } + } +#endif + + for (y = 0; y < height; ++y) { + ARGB1555ToARGBRow(src_argb1555, dst_argb, width); + src_argb1555 += src_stride_argb1555; + dst_argb += dst_stride_argb; + } + return 0; +} + +// Convert ARGB4444 to ARGB. +LIBYUV_API +int ARGB4444ToARGB(const uint8* src_argb4444, int src_stride_argb4444, + uint8* dst_argb, int dst_stride_argb, + int width, int height) { + int y; + void (*ARGB4444ToARGBRow)(const uint8* src_argb4444, uint8* dst_argb, + int pix) = ARGB4444ToARGBRow_C; + if (!src_argb4444 || !dst_argb || + width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + src_argb4444 = src_argb4444 + (height - 1) * src_stride_argb4444; + src_stride_argb4444 = -src_stride_argb4444; + } + // Coalesce rows. + if (src_stride_argb4444 == width * 2 && + dst_stride_argb == width * 4) { + width *= height; + height = 1; + src_stride_argb4444 = dst_stride_argb = 0; + } +#if defined(HAS_ARGB4444TOARGBROW_SSE2) + if (TestCpuFlag(kCpuHasSSE2)) { + ARGB4444ToARGBRow = ARGB4444ToARGBRow_Any_SSE2; + if (IS_ALIGNED(width, 8)) { + ARGB4444ToARGBRow = ARGB4444ToARGBRow_SSE2; + } + } +#endif +#if defined(HAS_ARGB4444TOARGBROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + ARGB4444ToARGBRow = ARGB4444ToARGBRow_Any_AVX2; + if (IS_ALIGNED(width, 16)) { + ARGB4444ToARGBRow = ARGB4444ToARGBRow_AVX2; + } + } +#endif +#if defined(HAS_ARGB4444TOARGBROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + ARGB4444ToARGBRow = ARGB4444ToARGBRow_Any_NEON; + if (IS_ALIGNED(width, 8)) { + ARGB4444ToARGBRow = ARGB4444ToARGBRow_NEON; + } + } +#endif + + for (y = 0; y < height; ++y) { + ARGB4444ToARGBRow(src_argb4444, dst_argb, width); + src_argb4444 += src_stride_argb4444; + dst_argb += dst_stride_argb; + } + return 0; +} + +// Convert NV12 to ARGB. +LIBYUV_API +int NV12ToARGB(const uint8* src_y, int src_stride_y, + const uint8* src_uv, int src_stride_uv, + uint8* dst_argb, int dst_stride_argb, + int width, int height) { + int y; + void (*NV12ToARGBRow)(const uint8* y_buf, + const uint8* uv_buf, + uint8* rgb_buf, + int width) = NV12ToARGBRow_C; + if (!src_y || !src_uv || !dst_argb || + width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + dst_argb = dst_argb + (height - 1) * dst_stride_argb; + dst_stride_argb = -dst_stride_argb; + } +#if defined(HAS_NV12TOARGBROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + NV12ToARGBRow = NV12ToARGBRow_Any_SSSE3; + if (IS_ALIGNED(width, 8)) { + NV12ToARGBRow = NV12ToARGBRow_SSSE3; + } + } +#endif +#if defined(HAS_NV12TOARGBROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + NV12ToARGBRow = NV12ToARGBRow_Any_AVX2; + if (IS_ALIGNED(width, 16)) { + NV12ToARGBRow = NV12ToARGBRow_AVX2; + } + } +#endif +#if defined(HAS_NV12TOARGBROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + NV12ToARGBRow = NV12ToARGBRow_Any_NEON; + if (IS_ALIGNED(width, 8)) { + NV12ToARGBRow = NV12ToARGBRow_NEON; + } + } +#endif + + for (y = 0; y < height; ++y) { + NV12ToARGBRow(src_y, src_uv, dst_argb, width); + dst_argb += dst_stride_argb; + src_y += src_stride_y; + if (y & 1) { + src_uv += src_stride_uv; + } + } + return 0; +} + +// Convert NV21 to ARGB. +LIBYUV_API +int NV21ToARGB(const uint8* src_y, int src_stride_y, + const uint8* src_uv, int src_stride_uv, + uint8* dst_argb, int dst_stride_argb, + int width, int height) { + int y; + void (*NV21ToARGBRow)(const uint8* y_buf, + const uint8* uv_buf, + uint8* rgb_buf, + int width) = NV21ToARGBRow_C; + if (!src_y || !src_uv || !dst_argb || + width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + dst_argb = dst_argb + (height - 1) * dst_stride_argb; + dst_stride_argb = -dst_stride_argb; + } +#if defined(HAS_NV21TOARGBROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + NV21ToARGBRow = NV21ToARGBRow_Any_SSSE3; + if (IS_ALIGNED(width, 8)) { + NV21ToARGBRow = NV21ToARGBRow_SSSE3; + } + } +#endif +#if defined(HAS_NV21TOARGBROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + NV21ToARGBRow = NV21ToARGBRow_Any_AVX2; + if (IS_ALIGNED(width, 16)) { + NV21ToARGBRow = NV21ToARGBRow_AVX2; + } + } +#endif +#if defined(HAS_NV21TOARGBROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + NV21ToARGBRow = NV21ToARGBRow_Any_NEON; + if (IS_ALIGNED(width, 8)) { + NV21ToARGBRow = NV21ToARGBRow_NEON; + } + } +#endif + + for (y = 0; y < height; ++y) { + NV21ToARGBRow(src_y, src_uv, dst_argb, width); + dst_argb += dst_stride_argb; + src_y += src_stride_y; + if (y & 1) { + src_uv += src_stride_uv; + } + } + return 0; +} + +// Convert M420 to ARGB. +LIBYUV_API +int M420ToARGB(const uint8* src_m420, int src_stride_m420, + uint8* dst_argb, int dst_stride_argb, + int width, int height) { + int y; + void (*NV12ToARGBRow)(const uint8* y_buf, + const uint8* uv_buf, + uint8* rgb_buf, + int width) = NV12ToARGBRow_C; + if (!src_m420 || !dst_argb || + width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + dst_argb = dst_argb + (height - 1) * dst_stride_argb; + dst_stride_argb = -dst_stride_argb; + } +#if defined(HAS_NV12TOARGBROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + NV12ToARGBRow = NV12ToARGBRow_Any_SSSE3; + if (IS_ALIGNED(width, 8)) { + NV12ToARGBRow = NV12ToARGBRow_SSSE3; + } + } +#endif +#if defined(HAS_NV12TOARGBROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + NV12ToARGBRow = NV12ToARGBRow_Any_AVX2; + if (IS_ALIGNED(width, 16)) { + NV12ToARGBRow = NV12ToARGBRow_AVX2; + } + } +#endif +#if defined(HAS_NV12TOARGBROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + NV12ToARGBRow = NV12ToARGBRow_Any_NEON; + if (IS_ALIGNED(width, 8)) { + NV12ToARGBRow = NV12ToARGBRow_NEON; + } + } +#endif + + for (y = 0; y < height - 1; y += 2) { + NV12ToARGBRow(src_m420, src_m420 + src_stride_m420 * 2, dst_argb, width); + NV12ToARGBRow(src_m420 + src_stride_m420, src_m420 + src_stride_m420 * 2, + dst_argb + dst_stride_argb, width); + dst_argb += dst_stride_argb * 2; + src_m420 += src_stride_m420 * 3; + } + if (height & 1) { + NV12ToARGBRow(src_m420, src_m420 + src_stride_m420 * 2, dst_argb, width); + } + return 0; +} + +// Convert YUY2 to ARGB. +LIBYUV_API +int YUY2ToARGB(const uint8* src_yuy2, int src_stride_yuy2, + uint8* dst_argb, int dst_stride_argb, + int width, int height) { + int y; + void (*YUY2ToARGBRow)(const uint8* src_yuy2, uint8* dst_argb, int pix) = + YUY2ToARGBRow_C; + if (!src_yuy2 || !dst_argb || + width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + src_yuy2 = src_yuy2 + (height - 1) * src_stride_yuy2; + src_stride_yuy2 = -src_stride_yuy2; + } + // Coalesce rows. + if (src_stride_yuy2 == width * 2 && + dst_stride_argb == width * 4) { + width *= height; + height = 1; + src_stride_yuy2 = dst_stride_argb = 0; + } +#if defined(HAS_YUY2TOARGBROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + YUY2ToARGBRow = YUY2ToARGBRow_Any_SSSE3; + if (IS_ALIGNED(width, 16)) { + YUY2ToARGBRow = YUY2ToARGBRow_SSSE3; + } + } +#endif +#if defined(HAS_YUY2TOARGBROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + YUY2ToARGBRow = YUY2ToARGBRow_Any_AVX2; + if (IS_ALIGNED(width, 32)) { + YUY2ToARGBRow = YUY2ToARGBRow_AVX2; + } + } +#endif +#if defined(HAS_YUY2TOARGBROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + YUY2ToARGBRow = YUY2ToARGBRow_Any_NEON; + if (IS_ALIGNED(width, 8)) { + YUY2ToARGBRow = YUY2ToARGBRow_NEON; + } + } +#endif + for (y = 0; y < height; ++y) { + YUY2ToARGBRow(src_yuy2, dst_argb, width); + src_yuy2 += src_stride_yuy2; + dst_argb += dst_stride_argb; + } + return 0; +} + +// Convert UYVY to ARGB. +LIBYUV_API +int UYVYToARGB(const uint8* src_uyvy, int src_stride_uyvy, + uint8* dst_argb, int dst_stride_argb, + int width, int height) { + int y; + void (*UYVYToARGBRow)(const uint8* src_uyvy, uint8* dst_argb, int pix) = + UYVYToARGBRow_C; + if (!src_uyvy || !dst_argb || + width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + src_uyvy = src_uyvy + (height - 1) * src_stride_uyvy; + src_stride_uyvy = -src_stride_uyvy; + } + // Coalesce rows. + if (src_stride_uyvy == width * 2 && + dst_stride_argb == width * 4) { + width *= height; + height = 1; + src_stride_uyvy = dst_stride_argb = 0; + } +#if defined(HAS_UYVYTOARGBROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + UYVYToARGBRow = UYVYToARGBRow_Any_SSSE3; + if (IS_ALIGNED(width, 16)) { + UYVYToARGBRow = UYVYToARGBRow_SSSE3; + } + } +#endif +#if defined(HAS_UYVYTOARGBROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + UYVYToARGBRow = UYVYToARGBRow_Any_AVX2; + if (IS_ALIGNED(width, 32)) { + UYVYToARGBRow = UYVYToARGBRow_AVX2; + } + } +#endif +#if defined(HAS_UYVYTOARGBROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + UYVYToARGBRow = UYVYToARGBRow_Any_NEON; + if (IS_ALIGNED(width, 8)) { + UYVYToARGBRow = UYVYToARGBRow_NEON; + } + } +#endif + for (y = 0; y < height; ++y) { + UYVYToARGBRow(src_uyvy, dst_argb, width); + src_uyvy += src_stride_uyvy; + dst_argb += dst_stride_argb; + } + return 0; +} + +// Convert J420 to ARGB. +LIBYUV_API +int J420ToARGB(const uint8* src_y, int src_stride_y, + const uint8* src_u, int src_stride_u, + const uint8* src_v, int src_stride_v, + uint8* dst_argb, int dst_stride_argb, + int width, int height) { + int y; + void (*J422ToARGBRow)(const uint8* y_buf, + const uint8* u_buf, + const uint8* v_buf, + uint8* rgb_buf, + int width) = J422ToARGBRow_C; + if (!src_y || !src_u || !src_v || !dst_argb || + width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + dst_argb = dst_argb + (height - 1) * dst_stride_argb; + dst_stride_argb = -dst_stride_argb; + } +#if defined(HAS_J422TOARGBROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + J422ToARGBRow = J422ToARGBRow_Any_SSSE3; + if (IS_ALIGNED(width, 8)) { + J422ToARGBRow = J422ToARGBRow_SSSE3; + } + } +#endif +#if defined(HAS_J422TOARGBROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + J422ToARGBRow = J422ToARGBRow_Any_AVX2; + if (IS_ALIGNED(width, 16)) { + J422ToARGBRow = J422ToARGBRow_AVX2; + } + } +#endif +#if defined(HAS_J422TOARGBROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + J422ToARGBRow = J422ToARGBRow_Any_NEON; + if (IS_ALIGNED(width, 8)) { + J422ToARGBRow = J422ToARGBRow_NEON; + } + } +#endif +#if defined(HAS_J422TOARGBROW_MIPS_DSPR2) + if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(width, 4) && + IS_ALIGNED(src_y, 4) && IS_ALIGNED(src_stride_y, 4) && + IS_ALIGNED(src_u, 2) && IS_ALIGNED(src_stride_u, 2) && + IS_ALIGNED(src_v, 2) && IS_ALIGNED(src_stride_v, 2) && + IS_ALIGNED(dst_argb, 4) && IS_ALIGNED(dst_stride_argb, 4)) { + J422ToARGBRow = J422ToARGBRow_MIPS_DSPR2; + } +#endif + + for (y = 0; y < height; ++y) { + J422ToARGBRow(src_y, src_u, src_v, dst_argb, width); + dst_argb += dst_stride_argb; + src_y += src_stride_y; + if (y & 1) { + src_u += src_stride_u; + src_v += src_stride_v; + } + } + return 0; +} + +// Convert J422 to ARGB. +LIBYUV_API +int J422ToARGB(const uint8* src_y, int src_stride_y, + const uint8* src_u, int src_stride_u, + const uint8* src_v, int src_stride_v, + uint8* dst_argb, int dst_stride_argb, + int width, int height) { + int y; + void (*J422ToARGBRow)(const uint8* y_buf, + const uint8* u_buf, + const uint8* v_buf, + uint8* rgb_buf, + int width) = J422ToARGBRow_C; + if (!src_y || !src_u || !src_v || + !dst_argb || + width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + dst_argb = dst_argb + (height - 1) * dst_stride_argb; + dst_stride_argb = -dst_stride_argb; + } + // Coalesce rows. + if (src_stride_y == width && + src_stride_u * 2 == width && + src_stride_v * 2 == width && + dst_stride_argb == width * 4) { + width *= height; + height = 1; + src_stride_y = src_stride_u = src_stride_v = dst_stride_argb = 0; + } +#if defined(HAS_J422TOARGBROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + J422ToARGBRow = J422ToARGBRow_Any_SSSE3; + if (IS_ALIGNED(width, 8)) { + J422ToARGBRow = J422ToARGBRow_SSSE3; + } + } +#endif +#if defined(HAS_J422TOARGBROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + J422ToARGBRow = J422ToARGBRow_Any_AVX2; + if (IS_ALIGNED(width, 16)) { + J422ToARGBRow = J422ToARGBRow_AVX2; + } + } +#endif +#if defined(HAS_J422TOARGBROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + J422ToARGBRow = J422ToARGBRow_Any_NEON; + if (IS_ALIGNED(width, 8)) { + J422ToARGBRow = J422ToARGBRow_NEON; + } + } +#endif +#if defined(HAS_J422TOARGBROW_MIPS_DSPR2) + if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(width, 4) && + IS_ALIGNED(src_y, 4) && IS_ALIGNED(src_stride_y, 4) && + IS_ALIGNED(src_u, 2) && IS_ALIGNED(src_stride_u, 2) && + IS_ALIGNED(src_v, 2) && IS_ALIGNED(src_stride_v, 2) && + IS_ALIGNED(dst_argb, 4) && IS_ALIGNED(dst_stride_argb, 4)) { + J422ToARGBRow = J422ToARGBRow_MIPS_DSPR2; + } +#endif + + for (y = 0; y < height; ++y) { + J422ToARGBRow(src_y, src_u, src_v, dst_argb, width); + dst_argb += dst_stride_argb; + src_y += src_stride_y; + src_u += src_stride_u; + src_v += src_stride_v; + } + return 0; +} + +#ifdef __cplusplus +} // extern "C" +} // namespace libyuv +#endif diff --git a/media/libaom/src/third_party/libyuv/source/convert_from.cc b/media/libaom/src/third_party/libyuv/source/convert_from.cc new file mode 100644 index 000000000..31f1ac992 --- /dev/null +++ b/media/libaom/src/third_party/libyuv/source/convert_from.cc @@ -0,0 +1,1348 @@ +/* + * Copyright 2012 The LibYuv Project Authors. All rights reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#include "libyuv/convert_from.h" + +#include "libyuv/basic_types.h" +#include "libyuv/convert.h" // For I420Copy +#include "libyuv/cpu_id.h" +#include "libyuv/planar_functions.h" +#include "libyuv/rotate.h" +#include "libyuv/scale.h" // For ScalePlane() +#include "libyuv/video_common.h" +#include "libyuv/row.h" + +#ifdef __cplusplus +namespace libyuv { +extern "C" { +#endif + +#define SUBSAMPLE(v, a, s) (v < 0) ? (-((-v + a) >> s)) : ((v + a) >> s) +static __inline int Abs(int v) { + return v >= 0 ? v : -v; +} + +// I420 To any I4xx YUV format with mirroring. +static int I420ToI4xx(const uint8* src_y, int src_stride_y, + const uint8* src_u, int src_stride_u, + const uint8* src_v, int src_stride_v, + uint8* dst_y, int dst_stride_y, + uint8* dst_u, int dst_stride_u, + uint8* dst_v, int dst_stride_v, + int src_y_width, int src_y_height, + int dst_uv_width, int dst_uv_height) { + const int dst_y_width = Abs(src_y_width); + const int dst_y_height = Abs(src_y_height); + const int src_uv_width = SUBSAMPLE(src_y_width, 1, 1); + const int src_uv_height = SUBSAMPLE(src_y_height, 1, 1); + if (src_y_width == 0 || src_y_height == 0 || + dst_uv_width <= 0 || dst_uv_height <= 0) { + return -1; + } + ScalePlane(src_y, src_stride_y, src_y_width, src_y_height, + dst_y, dst_stride_y, dst_y_width, dst_y_height, + kFilterBilinear); + ScalePlane(src_u, src_stride_u, src_uv_width, src_uv_height, + dst_u, dst_stride_u, dst_uv_width, dst_uv_height, + kFilterBilinear); + ScalePlane(src_v, src_stride_v, src_uv_width, src_uv_height, + dst_v, dst_stride_v, dst_uv_width, dst_uv_height, + kFilterBilinear); + return 0; +} + +// 420 chroma is 1/2 width, 1/2 height +// 422 chroma is 1/2 width, 1x height +LIBYUV_API +int I420ToI422(const uint8* src_y, int src_stride_y, + const uint8* src_u, int src_stride_u, + const uint8* src_v, int src_stride_v, + uint8* dst_y, int dst_stride_y, + uint8* dst_u, int dst_stride_u, + uint8* dst_v, int dst_stride_v, + int width, int height) { + const int dst_uv_width = (Abs(width) + 1) >> 1; + const int dst_uv_height = Abs(height); + return I420ToI4xx(src_y, src_stride_y, + src_u, src_stride_u, + src_v, src_stride_v, + dst_y, dst_stride_y, + dst_u, dst_stride_u, + dst_v, dst_stride_v, + width, height, + dst_uv_width, dst_uv_height); +} + +// 420 chroma is 1/2 width, 1/2 height +// 444 chroma is 1x width, 1x height +LIBYUV_API +int I420ToI444(const uint8* src_y, int src_stride_y, + const uint8* src_u, int src_stride_u, + const uint8* src_v, int src_stride_v, + uint8* dst_y, int dst_stride_y, + uint8* dst_u, int dst_stride_u, + uint8* dst_v, int dst_stride_v, + int width, int height) { + const int dst_uv_width = Abs(width); + const int dst_uv_height = Abs(height); + return I420ToI4xx(src_y, src_stride_y, + src_u, src_stride_u, + src_v, src_stride_v, + dst_y, dst_stride_y, + dst_u, dst_stride_u, + dst_v, dst_stride_v, + width, height, + dst_uv_width, dst_uv_height); +} + +// 420 chroma is 1/2 width, 1/2 height +// 411 chroma is 1/4 width, 1x height +LIBYUV_API +int I420ToI411(const uint8* src_y, int src_stride_y, + const uint8* src_u, int src_stride_u, + const uint8* src_v, int src_stride_v, + uint8* dst_y, int dst_stride_y, + uint8* dst_u, int dst_stride_u, + uint8* dst_v, int dst_stride_v, + int width, int height) { + const int dst_uv_width = (Abs(width) + 3) >> 2; + const int dst_uv_height = Abs(height); + return I420ToI4xx(src_y, src_stride_y, + src_u, src_stride_u, + src_v, src_stride_v, + dst_y, dst_stride_y, + dst_u, dst_stride_u, + dst_v, dst_stride_v, + width, height, + dst_uv_width, dst_uv_height); +} + +// Copy to I400. Source can be I420,422,444,400,NV12,NV21 +LIBYUV_API +int I400Copy(const uint8* src_y, int src_stride_y, + uint8* dst_y, int dst_stride_y, + int width, int height) { + if (!src_y || !dst_y || + width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + src_y = src_y + (height - 1) * src_stride_y; + src_stride_y = -src_stride_y; + } + CopyPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height); + return 0; +} + +LIBYUV_API +int I422ToYUY2(const uint8* src_y, int src_stride_y, + const uint8* src_u, int src_stride_u, + const uint8* src_v, int src_stride_v, + uint8* dst_yuy2, int dst_stride_yuy2, + int width, int height) { + int y; + void (*I422ToYUY2Row)(const uint8* src_y, const uint8* src_u, + const uint8* src_v, uint8* dst_yuy2, int width) = + I422ToYUY2Row_C; + if (!src_y || !src_u || !src_v || !dst_yuy2 || + width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + dst_yuy2 = dst_yuy2 + (height - 1) * dst_stride_yuy2; + dst_stride_yuy2 = -dst_stride_yuy2; + } + // Coalesce rows. + if (src_stride_y == width && + src_stride_u * 2 == width && + src_stride_v * 2 == width && + dst_stride_yuy2 == width * 2) { + width *= height; + height = 1; + src_stride_y = src_stride_u = src_stride_v = dst_stride_yuy2 = 0; + } +#if defined(HAS_I422TOYUY2ROW_SSE2) + if (TestCpuFlag(kCpuHasSSE2)) { + I422ToYUY2Row = I422ToYUY2Row_Any_SSE2; + if (IS_ALIGNED(width, 16)) { + I422ToYUY2Row = I422ToYUY2Row_SSE2; + } + } +#endif +#if defined(HAS_I422TOYUY2ROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + I422ToYUY2Row = I422ToYUY2Row_Any_NEON; + if (IS_ALIGNED(width, 16)) { + I422ToYUY2Row = I422ToYUY2Row_NEON; + } + } +#endif + + for (y = 0; y < height; ++y) { + I422ToYUY2Row(src_y, src_u, src_v, dst_yuy2, width); + src_y += src_stride_y; + src_u += src_stride_u; + src_v += src_stride_v; + dst_yuy2 += dst_stride_yuy2; + } + return 0; +} + +LIBYUV_API +int I420ToYUY2(const uint8* src_y, int src_stride_y, + const uint8* src_u, int src_stride_u, + const uint8* src_v, int src_stride_v, + uint8* dst_yuy2, int dst_stride_yuy2, + int width, int height) { + int y; + void (*I422ToYUY2Row)(const uint8* src_y, const uint8* src_u, + const uint8* src_v, uint8* dst_yuy2, int width) = + I422ToYUY2Row_C; + if (!src_y || !src_u || !src_v || !dst_yuy2 || + width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + dst_yuy2 = dst_yuy2 + (height - 1) * dst_stride_yuy2; + dst_stride_yuy2 = -dst_stride_yuy2; + } +#if defined(HAS_I422TOYUY2ROW_SSE2) + if (TestCpuFlag(kCpuHasSSE2)) { + I422ToYUY2Row = I422ToYUY2Row_Any_SSE2; + if (IS_ALIGNED(width, 16)) { + I422ToYUY2Row = I422ToYUY2Row_SSE2; + } + } +#endif +#if defined(HAS_I422TOYUY2ROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + I422ToYUY2Row = I422ToYUY2Row_Any_NEON; + if (IS_ALIGNED(width, 16)) { + I422ToYUY2Row = I422ToYUY2Row_NEON; + } + } +#endif + + for (y = 0; y < height - 1; y += 2) { + I422ToYUY2Row(src_y, src_u, src_v, dst_yuy2, width); + I422ToYUY2Row(src_y + src_stride_y, src_u, src_v, + dst_yuy2 + dst_stride_yuy2, width); + src_y += src_stride_y * 2; + src_u += src_stride_u; + src_v += src_stride_v; + dst_yuy2 += dst_stride_yuy2 * 2; + } + if (height & 1) { + I422ToYUY2Row(src_y, src_u, src_v, dst_yuy2, width); + } + return 0; +} + +LIBYUV_API +int I422ToUYVY(const uint8* src_y, int src_stride_y, + const uint8* src_u, int src_stride_u, + const uint8* src_v, int src_stride_v, + uint8* dst_uyvy, int dst_stride_uyvy, + int width, int height) { + int y; + void (*I422ToUYVYRow)(const uint8* src_y, const uint8* src_u, + const uint8* src_v, uint8* dst_uyvy, int width) = + I422ToUYVYRow_C; + if (!src_y || !src_u || !src_v || !dst_uyvy || + width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + dst_uyvy = dst_uyvy + (height - 1) * dst_stride_uyvy; + dst_stride_uyvy = -dst_stride_uyvy; + } + // Coalesce rows. + if (src_stride_y == width && + src_stride_u * 2 == width && + src_stride_v * 2 == width && + dst_stride_uyvy == width * 2) { + width *= height; + height = 1; + src_stride_y = src_stride_u = src_stride_v = dst_stride_uyvy = 0; + } +#if defined(HAS_I422TOUYVYROW_SSE2) + if (TestCpuFlag(kCpuHasSSE2)) { + I422ToUYVYRow = I422ToUYVYRow_Any_SSE2; + if (IS_ALIGNED(width, 16)) { + I422ToUYVYRow = I422ToUYVYRow_SSE2; + } + } +#endif +#if defined(HAS_I422TOUYVYROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + I422ToUYVYRow = I422ToUYVYRow_Any_NEON; + if (IS_ALIGNED(width, 16)) { + I422ToUYVYRow = I422ToUYVYRow_NEON; + } + } +#endif + + for (y = 0; y < height; ++y) { + I422ToUYVYRow(src_y, src_u, src_v, dst_uyvy, width); + src_y += src_stride_y; + src_u += src_stride_u; + src_v += src_stride_v; + dst_uyvy += dst_stride_uyvy; + } + return 0; +} + +LIBYUV_API +int I420ToUYVY(const uint8* src_y, int src_stride_y, + const uint8* src_u, int src_stride_u, + const uint8* src_v, int src_stride_v, + uint8* dst_uyvy, int dst_stride_uyvy, + int width, int height) { + int y; + void (*I422ToUYVYRow)(const uint8* src_y, const uint8* src_u, + const uint8* src_v, uint8* dst_uyvy, int width) = + I422ToUYVYRow_C; + if (!src_y || !src_u || !src_v || !dst_uyvy || + width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + dst_uyvy = dst_uyvy + (height - 1) * dst_stride_uyvy; + dst_stride_uyvy = -dst_stride_uyvy; + } +#if defined(HAS_I422TOUYVYROW_SSE2) + if (TestCpuFlag(kCpuHasSSE2)) { + I422ToUYVYRow = I422ToUYVYRow_Any_SSE2; + if (IS_ALIGNED(width, 16)) { + I422ToUYVYRow = I422ToUYVYRow_SSE2; + } + } +#endif +#if defined(HAS_I422TOUYVYROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + I422ToUYVYRow = I422ToUYVYRow_Any_NEON; + if (IS_ALIGNED(width, 16)) { + I422ToUYVYRow = I422ToUYVYRow_NEON; + } + } +#endif + + for (y = 0; y < height - 1; y += 2) { + I422ToUYVYRow(src_y, src_u, src_v, dst_uyvy, width); + I422ToUYVYRow(src_y + src_stride_y, src_u, src_v, + dst_uyvy + dst_stride_uyvy, width); + src_y += src_stride_y * 2; + src_u += src_stride_u; + src_v += src_stride_v; + dst_uyvy += dst_stride_uyvy * 2; + } + if (height & 1) { + I422ToUYVYRow(src_y, src_u, src_v, dst_uyvy, width); + } + return 0; +} + +LIBYUV_API +int I420ToNV12(const uint8* src_y, int src_stride_y, + const uint8* src_u, int src_stride_u, + const uint8* src_v, int src_stride_v, + uint8* dst_y, int dst_stride_y, + uint8* dst_uv, int dst_stride_uv, + int width, int height) { + int y; + void (*MergeUVRow_)(const uint8* src_u, const uint8* src_v, uint8* dst_uv, + int width) = MergeUVRow_C; + // Coalesce rows. + int halfwidth = (width + 1) >> 1; + int halfheight = (height + 1) >> 1; + if (!src_y || !src_u || !src_v || !dst_y || !dst_uv || + width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + halfheight = (height + 1) >> 1; + dst_y = dst_y + (height - 1) * dst_stride_y; + dst_uv = dst_uv + (halfheight - 1) * dst_stride_uv; + dst_stride_y = -dst_stride_y; + dst_stride_uv = -dst_stride_uv; + } + if (src_stride_y == width && + dst_stride_y == width) { + width *= height; + height = 1; + src_stride_y = dst_stride_y = 0; + } + // Coalesce rows. + if (src_stride_u == halfwidth && + src_stride_v == halfwidth && + dst_stride_uv == halfwidth * 2) { + halfwidth *= halfheight; + halfheight = 1; + src_stride_u = src_stride_v = dst_stride_uv = 0; + } +#if defined(HAS_MERGEUVROW_SSE2) + if (TestCpuFlag(kCpuHasSSE2)) { + MergeUVRow_ = MergeUVRow_Any_SSE2; + if (IS_ALIGNED(halfwidth, 16)) { + MergeUVRow_ = MergeUVRow_SSE2; + } + } +#endif +#if defined(HAS_MERGEUVROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + MergeUVRow_ = MergeUVRow_Any_AVX2; + if (IS_ALIGNED(halfwidth, 32)) { + MergeUVRow_ = MergeUVRow_AVX2; + } + } +#endif +#if defined(HAS_MERGEUVROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + MergeUVRow_ = MergeUVRow_Any_NEON; + if (IS_ALIGNED(halfwidth, 16)) { + MergeUVRow_ = MergeUVRow_NEON; + } + } +#endif + + CopyPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height); + for (y = 0; y < halfheight; ++y) { + // Merge a row of U and V into a row of UV. + MergeUVRow_(src_u, src_v, dst_uv, halfwidth); + src_u += src_stride_u; + src_v += src_stride_v; + dst_uv += dst_stride_uv; + } + return 0; +} + +LIBYUV_API +int I420ToNV21(const uint8* src_y, int src_stride_y, + const uint8* src_u, int src_stride_u, + const uint8* src_v, int src_stride_v, + uint8* dst_y, int dst_stride_y, + uint8* dst_vu, int dst_stride_vu, + int width, int height) { + return I420ToNV12(src_y, src_stride_y, + src_v, src_stride_v, + src_u, src_stride_u, + dst_y, src_stride_y, + dst_vu, dst_stride_vu, + width, height); +} + +// Convert I420 to ARGB. +LIBYUV_API +int I420ToARGB(const uint8* src_y, int src_stride_y, + const uint8* src_u, int src_stride_u, + const uint8* src_v, int src_stride_v, + uint8* dst_argb, int dst_stride_argb, + int width, int height) { + int y; + void (*I422ToARGBRow)(const uint8* y_buf, + const uint8* u_buf, + const uint8* v_buf, + uint8* rgb_buf, + int width) = I422ToARGBRow_C; + if (!src_y || !src_u || !src_v || !dst_argb || + width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + dst_argb = dst_argb + (height - 1) * dst_stride_argb; + dst_stride_argb = -dst_stride_argb; + } +#if defined(HAS_I422TOARGBROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + I422ToARGBRow = I422ToARGBRow_Any_SSSE3; + if (IS_ALIGNED(width, 8)) { + I422ToARGBRow = I422ToARGBRow_SSSE3; + } + } +#endif +#if defined(HAS_I422TOARGBROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + I422ToARGBRow = I422ToARGBRow_Any_AVX2; + if (IS_ALIGNED(width, 16)) { + I422ToARGBRow = I422ToARGBRow_AVX2; + } + } +#endif +#if defined(HAS_I422TOARGBROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + I422ToARGBRow = I422ToARGBRow_Any_NEON; + if (IS_ALIGNED(width, 8)) { + I422ToARGBRow = I422ToARGBRow_NEON; + } + } +#endif +#if defined(HAS_I422TOARGBROW_MIPS_DSPR2) + if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(width, 4) && + IS_ALIGNED(src_y, 4) && IS_ALIGNED(src_stride_y, 4) && + IS_ALIGNED(src_u, 2) && IS_ALIGNED(src_stride_u, 2) && + IS_ALIGNED(src_v, 2) && IS_ALIGNED(src_stride_v, 2) && + IS_ALIGNED(dst_argb, 4) && IS_ALIGNED(dst_stride_argb, 4)) { + I422ToARGBRow = I422ToARGBRow_MIPS_DSPR2; + } +#endif + + for (y = 0; y < height; ++y) { + I422ToARGBRow(src_y, src_u, src_v, dst_argb, width); + dst_argb += dst_stride_argb; + src_y += src_stride_y; + if (y & 1) { + src_u += src_stride_u; + src_v += src_stride_v; + } + } + return 0; +} + +// Convert I420 to BGRA. +LIBYUV_API +int I420ToBGRA(const uint8* src_y, int src_stride_y, + const uint8* src_u, int src_stride_u, + const uint8* src_v, int src_stride_v, + uint8* dst_bgra, int dst_stride_bgra, + int width, int height) { + int y; + void (*I422ToBGRARow)(const uint8* y_buf, + const uint8* u_buf, + const uint8* v_buf, + uint8* rgb_buf, + int width) = I422ToBGRARow_C; + if (!src_y || !src_u || !src_v || !dst_bgra || + width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + dst_bgra = dst_bgra + (height - 1) * dst_stride_bgra; + dst_stride_bgra = -dst_stride_bgra; + } +#if defined(HAS_I422TOBGRAROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + I422ToBGRARow = I422ToBGRARow_Any_SSSE3; + if (IS_ALIGNED(width, 8)) { + I422ToBGRARow = I422ToBGRARow_SSSE3; + } + } +#endif +#if defined(HAS_I422TOBGRAROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + I422ToBGRARow = I422ToBGRARow_Any_AVX2; + if (IS_ALIGNED(width, 16)) { + I422ToBGRARow = I422ToBGRARow_AVX2; + } + } +#endif +#if defined(HAS_I422TOBGRAROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + I422ToBGRARow = I422ToBGRARow_Any_NEON; + if (IS_ALIGNED(width, 8)) { + I422ToBGRARow = I422ToBGRARow_NEON; + } + } +#endif +#if defined(HAS_I422TOBGRAROW_MIPS_DSPR2) + if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(width, 4) && + IS_ALIGNED(src_y, 4) && IS_ALIGNED(src_stride_y, 4) && + IS_ALIGNED(src_u, 2) && IS_ALIGNED(src_stride_u, 2) && + IS_ALIGNED(src_v, 2) && IS_ALIGNED(src_stride_v, 2) && + IS_ALIGNED(dst_bgra, 4) && IS_ALIGNED(dst_stride_bgra, 4)) { + I422ToBGRARow = I422ToBGRARow_MIPS_DSPR2; + } +#endif + + for (y = 0; y < height; ++y) { + I422ToBGRARow(src_y, src_u, src_v, dst_bgra, width); + dst_bgra += dst_stride_bgra; + src_y += src_stride_y; + if (y & 1) { + src_u += src_stride_u; + src_v += src_stride_v; + } + } + return 0; +} + +// Convert I420 to ABGR. +LIBYUV_API +int I420ToABGR(const uint8* src_y, int src_stride_y, + const uint8* src_u, int src_stride_u, + const uint8* src_v, int src_stride_v, + uint8* dst_abgr, int dst_stride_abgr, + int width, int height) { + int y; + void (*I422ToABGRRow)(const uint8* y_buf, + const uint8* u_buf, + const uint8* v_buf, + uint8* rgb_buf, + int width) = I422ToABGRRow_C; + if (!src_y || !src_u || !src_v || !dst_abgr || + width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + dst_abgr = dst_abgr + (height - 1) * dst_stride_abgr; + dst_stride_abgr = -dst_stride_abgr; + } +#if defined(HAS_I422TOABGRROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + I422ToABGRRow = I422ToABGRRow_Any_SSSE3; + if (IS_ALIGNED(width, 8)) { + I422ToABGRRow = I422ToABGRRow_SSSE3; + } + } +#endif +#if defined(HAS_I422TOABGRROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + I422ToABGRRow = I422ToABGRRow_Any_AVX2; + if (IS_ALIGNED(width, 16)) { + I422ToABGRRow = I422ToABGRRow_AVX2; + } + } +#endif +#if defined(HAS_I422TOABGRROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + I422ToABGRRow = I422ToABGRRow_Any_NEON; + if (IS_ALIGNED(width, 8)) { + I422ToABGRRow = I422ToABGRRow_NEON; + } + } +#endif + + for (y = 0; y < height; ++y) { + I422ToABGRRow(src_y, src_u, src_v, dst_abgr, width); + dst_abgr += dst_stride_abgr; + src_y += src_stride_y; + if (y & 1) { + src_u += src_stride_u; + src_v += src_stride_v; + } + } + return 0; +} + +// Convert I420 to RGBA. +LIBYUV_API +int I420ToRGBA(const uint8* src_y, int src_stride_y, + const uint8* src_u, int src_stride_u, + const uint8* src_v, int src_stride_v, + uint8* dst_rgba, int dst_stride_rgba, + int width, int height) { + int y; + void (*I422ToRGBARow)(const uint8* y_buf, + const uint8* u_buf, + const uint8* v_buf, + uint8* rgb_buf, + int width) = I422ToRGBARow_C; + if (!src_y || !src_u || !src_v || !dst_rgba || + width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + dst_rgba = dst_rgba + (height - 1) * dst_stride_rgba; + dst_stride_rgba = -dst_stride_rgba; + } +#if defined(HAS_I422TORGBAROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + I422ToRGBARow = I422ToRGBARow_Any_SSSE3; + if (IS_ALIGNED(width, 8)) { + I422ToRGBARow = I422ToRGBARow_SSSE3; + } + } +#endif +#if defined(HAS_I422TORGBAROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + I422ToRGBARow = I422ToRGBARow_Any_AVX2; + if (IS_ALIGNED(width, 16)) { + I422ToRGBARow = I422ToRGBARow_AVX2; + } + } +#endif +#if defined(HAS_I422TORGBAROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + I422ToRGBARow = I422ToRGBARow_Any_NEON; + if (IS_ALIGNED(width, 8)) { + I422ToRGBARow = I422ToRGBARow_NEON; + } + } +#endif + + for (y = 0; y < height; ++y) { + I422ToRGBARow(src_y, src_u, src_v, dst_rgba, width); + dst_rgba += dst_stride_rgba; + src_y += src_stride_y; + if (y & 1) { + src_u += src_stride_u; + src_v += src_stride_v; + } + } + return 0; +} + +// Convert I420 to RGB24. +LIBYUV_API +int I420ToRGB24(const uint8* src_y, int src_stride_y, + const uint8* src_u, int src_stride_u, + const uint8* src_v, int src_stride_v, + uint8* dst_rgb24, int dst_stride_rgb24, + int width, int height) { + int y; + void (*I422ToRGB24Row)(const uint8* y_buf, + const uint8* u_buf, + const uint8* v_buf, + uint8* rgb_buf, + int width) = I422ToRGB24Row_C; + if (!src_y || !src_u || !src_v || !dst_rgb24 || + width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + dst_rgb24 = dst_rgb24 + (height - 1) * dst_stride_rgb24; + dst_stride_rgb24 = -dst_stride_rgb24; + } +#if defined(HAS_I422TORGB24ROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + I422ToRGB24Row = I422ToRGB24Row_Any_SSSE3; + if (IS_ALIGNED(width, 8)) { + I422ToRGB24Row = I422ToRGB24Row_SSSE3; + } + } +#endif +#if defined(HAS_I422TORGB24ROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + I422ToRGB24Row = I422ToRGB24Row_Any_AVX2; + if (IS_ALIGNED(width, 16)) { + I422ToRGB24Row = I422ToRGB24Row_AVX2; + } + } +#endif +#if defined(HAS_I422TORGB24ROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + I422ToRGB24Row = I422ToRGB24Row_Any_NEON; + if (IS_ALIGNED(width, 8)) { + I422ToRGB24Row = I422ToRGB24Row_NEON; + } + } +#endif + + for (y = 0; y < height; ++y) { + I422ToRGB24Row(src_y, src_u, src_v, dst_rgb24, width); + dst_rgb24 += dst_stride_rgb24; + src_y += src_stride_y; + if (y & 1) { + src_u += src_stride_u; + src_v += src_stride_v; + } + } + return 0; +} + +// Convert I420 to RAW. +LIBYUV_API +int I420ToRAW(const uint8* src_y, int src_stride_y, + const uint8* src_u, int src_stride_u, + const uint8* src_v, int src_stride_v, + uint8* dst_raw, int dst_stride_raw, + int width, int height) { + int y; + void (*I422ToRAWRow)(const uint8* y_buf, + const uint8* u_buf, + const uint8* v_buf, + uint8* rgb_buf, + int width) = I422ToRAWRow_C; + if (!src_y || !src_u || !src_v || !dst_raw || + width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + dst_raw = dst_raw + (height - 1) * dst_stride_raw; + dst_stride_raw = -dst_stride_raw; + } +#if defined(HAS_I422TORAWROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + I422ToRAWRow = I422ToRAWRow_Any_SSSE3; + if (IS_ALIGNED(width, 8)) { + I422ToRAWRow = I422ToRAWRow_SSSE3; + } + } +#endif +#if defined(HAS_I422TORAWROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + I422ToRAWRow = I422ToRAWRow_Any_AVX2; + if (IS_ALIGNED(width, 16)) { + I422ToRAWRow = I422ToRAWRow_AVX2; + } + } +#endif +#if defined(HAS_I422TORAWROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + I422ToRAWRow = I422ToRAWRow_Any_NEON; + if (IS_ALIGNED(width, 8)) { + I422ToRAWRow = I422ToRAWRow_NEON; + } + } +#endif + + for (y = 0; y < height; ++y) { + I422ToRAWRow(src_y, src_u, src_v, dst_raw, width); + dst_raw += dst_stride_raw; + src_y += src_stride_y; + if (y & 1) { + src_u += src_stride_u; + src_v += src_stride_v; + } + } + return 0; +} + +// Convert I420 to ARGB1555. +LIBYUV_API +int I420ToARGB1555(const uint8* src_y, int src_stride_y, + const uint8* src_u, int src_stride_u, + const uint8* src_v, int src_stride_v, + uint8* dst_argb1555, int dst_stride_argb1555, + int width, int height) { + int y; + void (*I422ToARGB1555Row)(const uint8* y_buf, + const uint8* u_buf, + const uint8* v_buf, + uint8* rgb_buf, + int width) = I422ToARGB1555Row_C; + if (!src_y || !src_u || !src_v || !dst_argb1555 || + width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + dst_argb1555 = dst_argb1555 + (height - 1) * dst_stride_argb1555; + dst_stride_argb1555 = -dst_stride_argb1555; + } +#if defined(HAS_I422TOARGB1555ROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + I422ToARGB1555Row = I422ToARGB1555Row_Any_SSSE3; + if (IS_ALIGNED(width, 8)) { + I422ToARGB1555Row = I422ToARGB1555Row_SSSE3; + } + } +#endif +#if defined(HAS_I422TOARGB1555ROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + I422ToARGB1555Row = I422ToARGB1555Row_Any_AVX2; + if (IS_ALIGNED(width, 16)) { + I422ToARGB1555Row = I422ToARGB1555Row_AVX2; + } + } +#endif +#if defined(HAS_I422TOARGB1555ROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + I422ToARGB1555Row = I422ToARGB1555Row_Any_NEON; + if (IS_ALIGNED(width, 8)) { + I422ToARGB1555Row = I422ToARGB1555Row_NEON; + } + } +#endif + + for (y = 0; y < height; ++y) { + I422ToARGB1555Row(src_y, src_u, src_v, dst_argb1555, width); + dst_argb1555 += dst_stride_argb1555; + src_y += src_stride_y; + if (y & 1) { + src_u += src_stride_u; + src_v += src_stride_v; + } + } + return 0; +} + + +// Convert I420 to ARGB4444. +LIBYUV_API +int I420ToARGB4444(const uint8* src_y, int src_stride_y, + const uint8* src_u, int src_stride_u, + const uint8* src_v, int src_stride_v, + uint8* dst_argb4444, int dst_stride_argb4444, + int width, int height) { + int y; + void (*I422ToARGB4444Row)(const uint8* y_buf, + const uint8* u_buf, + const uint8* v_buf, + uint8* rgb_buf, + int width) = I422ToARGB4444Row_C; + if (!src_y || !src_u || !src_v || !dst_argb4444 || + width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + dst_argb4444 = dst_argb4444 + (height - 1) * dst_stride_argb4444; + dst_stride_argb4444 = -dst_stride_argb4444; + } +#if defined(HAS_I422TOARGB4444ROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + I422ToARGB4444Row = I422ToARGB4444Row_Any_SSSE3; + if (IS_ALIGNED(width, 8)) { + I422ToARGB4444Row = I422ToARGB4444Row_SSSE3; + } + } +#endif +#if defined(HAS_I422TOARGB4444ROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + I422ToARGB4444Row = I422ToARGB4444Row_Any_AVX2; + if (IS_ALIGNED(width, 16)) { + I422ToARGB4444Row = I422ToARGB4444Row_AVX2; + } + } +#endif +#if defined(HAS_I422TOARGB4444ROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + I422ToARGB4444Row = I422ToARGB4444Row_Any_NEON; + if (IS_ALIGNED(width, 8)) { + I422ToARGB4444Row = I422ToARGB4444Row_NEON; + } + } +#endif + + for (y = 0; y < height; ++y) { + I422ToARGB4444Row(src_y, src_u, src_v, dst_argb4444, width); + dst_argb4444 += dst_stride_argb4444; + src_y += src_stride_y; + if (y & 1) { + src_u += src_stride_u; + src_v += src_stride_v; + } + } + return 0; +} + +// Convert I420 to RGB565. +LIBYUV_API +int I420ToRGB565(const uint8* src_y, int src_stride_y, + const uint8* src_u, int src_stride_u, + const uint8* src_v, int src_stride_v, + uint8* dst_rgb565, int dst_stride_rgb565, + int width, int height) { + int y; + void (*I422ToRGB565Row)(const uint8* y_buf, + const uint8* u_buf, + const uint8* v_buf, + uint8* rgb_buf, + int width) = I422ToRGB565Row_C; + if (!src_y || !src_u || !src_v || !dst_rgb565 || + width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + dst_rgb565 = dst_rgb565 + (height - 1) * dst_stride_rgb565; + dst_stride_rgb565 = -dst_stride_rgb565; + } +#if defined(HAS_I422TORGB565ROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + I422ToRGB565Row = I422ToRGB565Row_Any_SSSE3; + if (IS_ALIGNED(width, 8)) { + I422ToRGB565Row = I422ToRGB565Row_SSSE3; + } + } +#endif +#if defined(HAS_I422TORGB565ROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + I422ToRGB565Row = I422ToRGB565Row_Any_AVX2; + if (IS_ALIGNED(width, 16)) { + I422ToRGB565Row = I422ToRGB565Row_AVX2; + } + } +#endif +#if defined(HAS_I422TORGB565ROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + I422ToRGB565Row = I422ToRGB565Row_Any_NEON; + if (IS_ALIGNED(width, 8)) { + I422ToRGB565Row = I422ToRGB565Row_NEON; + } + } +#endif + + for (y = 0; y < height; ++y) { + I422ToRGB565Row(src_y, src_u, src_v, dst_rgb565, width); + dst_rgb565 += dst_stride_rgb565; + src_y += src_stride_y; + if (y & 1) { + src_u += src_stride_u; + src_v += src_stride_v; + } + } + return 0; +} + +// Ordered 8x8 dither for 888 to 565. Values from 0 to 7. +static const uint8 kDither565_4x4[16] = { + 0, 4, 1, 5, + 6, 2, 7, 3, + 1, 5, 0, 4, + 7, 3, 6, 2, +}; + +// Convert I420 to RGB565 with dithering. +LIBYUV_API +int I420ToRGB565Dither(const uint8* src_y, int src_stride_y, + const uint8* src_u, int src_stride_u, + const uint8* src_v, int src_stride_v, + uint8* dst_rgb565, int dst_stride_rgb565, + const uint8* dither4x4, int width, int height) { + int y; + void (*I422ToARGBRow)(const uint8* y_buf, + const uint8* u_buf, + const uint8* v_buf, + uint8* rgb_buf, + int width) = I422ToARGBRow_C; + void (*ARGBToRGB565DitherRow)(const uint8* src_argb, uint8* dst_rgb, + const uint32 dither4, int pix) = ARGBToRGB565DitherRow_C; + if (!src_y || !src_u || !src_v || !dst_rgb565 || + width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + dst_rgb565 = dst_rgb565 + (height - 1) * dst_stride_rgb565; + dst_stride_rgb565 = -dst_stride_rgb565; + } + if (!dither4x4) { + dither4x4 = kDither565_4x4; + } +#if defined(HAS_I422TOARGBROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + I422ToARGBRow = I422ToARGBRow_Any_SSSE3; + if (IS_ALIGNED(width, 8)) { + I422ToARGBRow = I422ToARGBRow_SSSE3; + } + } +#endif +#if defined(HAS_I422TOARGBROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + I422ToARGBRow = I422ToARGBRow_Any_AVX2; + if (IS_ALIGNED(width, 16)) { + I422ToARGBRow = I422ToARGBRow_AVX2; + } + } +#endif +#if defined(HAS_I422TOARGBROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + I422ToARGBRow = I422ToARGBRow_Any_NEON; + if (IS_ALIGNED(width, 8)) { + I422ToARGBRow = I422ToARGBRow_NEON; + } + } +#endif +#if defined(HAS_I422TOARGBROW_MIPS_DSPR2) + if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(width, 4) && + IS_ALIGNED(src_y, 4) && IS_ALIGNED(src_stride_y, 4) && + IS_ALIGNED(src_u, 2) && IS_ALIGNED(src_stride_u, 2) && + IS_ALIGNED(src_v, 2) && IS_ALIGNED(src_stride_v, 2)) { + I422ToARGBRow = I422ToARGBRow_MIPS_DSPR2; + } +#endif +#if defined(HAS_ARGBTORGB565DITHERROW_SSE2) + if (TestCpuFlag(kCpuHasSSE2)) { + ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_Any_SSE2; + if (IS_ALIGNED(width, 4)) { + ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_SSE2; + } + } +#endif +#if defined(HAS_ARGBTORGB565DITHERROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_Any_AVX2; + if (IS_ALIGNED(width, 8)) { + ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_AVX2; + } + } +#endif +#if defined(HAS_ARGBTORGB565DITHERROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_Any_NEON; + if (IS_ALIGNED(width, 8)) { + ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_NEON; + } + } +#endif + { + // Allocate a row of argb. + align_buffer_64(row_argb, width * 4); + for (y = 0; y < height; ++y) { + I422ToARGBRow(src_y, src_u, src_v, row_argb, width); + ARGBToRGB565DitherRow(row_argb, dst_rgb565, + *(uint32*)(dither4x4 + ((y & 3) << 2)), width); + dst_rgb565 += dst_stride_rgb565; + src_y += src_stride_y; + if (y & 1) { + src_u += src_stride_u; + src_v += src_stride_v; + } + } + free_aligned_buffer_64(row_argb); + } + return 0; +} + +// Convert I420 to specified format +LIBYUV_API +int ConvertFromI420(const uint8* y, int y_stride, + const uint8* u, int u_stride, + const uint8* v, int v_stride, + uint8* dst_sample, int dst_sample_stride, + int width, int height, + uint32 fourcc) { + uint32 format = CanonicalFourCC(fourcc); + int r = 0; + if (!y || !u|| !v || !dst_sample || + width <= 0 || height == 0) { + return -1; + } + switch (format) { + // Single plane formats + case FOURCC_YUY2: + r = I420ToYUY2(y, y_stride, + u, u_stride, + v, v_stride, + dst_sample, + dst_sample_stride ? dst_sample_stride : width * 2, + width, height); + break; + case FOURCC_UYVY: + r = I420ToUYVY(y, y_stride, + u, u_stride, + v, v_stride, + dst_sample, + dst_sample_stride ? dst_sample_stride : width * 2, + width, height); + break; + case FOURCC_RGBP: + r = I420ToRGB565(y, y_stride, + u, u_stride, + v, v_stride, + dst_sample, + dst_sample_stride ? dst_sample_stride : width * 2, + width, height); + break; + case FOURCC_RGBO: + r = I420ToARGB1555(y, y_stride, + u, u_stride, + v, v_stride, + dst_sample, + dst_sample_stride ? dst_sample_stride : width * 2, + width, height); + break; + case FOURCC_R444: + r = I420ToARGB4444(y, y_stride, + u, u_stride, + v, v_stride, + dst_sample, + dst_sample_stride ? dst_sample_stride : width * 2, + width, height); + break; + case FOURCC_24BG: + r = I420ToRGB24(y, y_stride, + u, u_stride, + v, v_stride, + dst_sample, + dst_sample_stride ? dst_sample_stride : width * 3, + width, height); + break; + case FOURCC_RAW: + r = I420ToRAW(y, y_stride, + u, u_stride, + v, v_stride, + dst_sample, + dst_sample_stride ? dst_sample_stride : width * 3, + width, height); + break; + case FOURCC_ARGB: + r = I420ToARGB(y, y_stride, + u, u_stride, + v, v_stride, + dst_sample, + dst_sample_stride ? dst_sample_stride : width * 4, + width, height); + break; + case FOURCC_BGRA: + r = I420ToBGRA(y, y_stride, + u, u_stride, + v, v_stride, + dst_sample, + dst_sample_stride ? dst_sample_stride : width * 4, + width, height); + break; + case FOURCC_ABGR: + r = I420ToABGR(y, y_stride, + u, u_stride, + v, v_stride, + dst_sample, + dst_sample_stride ? dst_sample_stride : width * 4, + width, height); + break; + case FOURCC_RGBA: + r = I420ToRGBA(y, y_stride, + u, u_stride, + v, v_stride, + dst_sample, + dst_sample_stride ? dst_sample_stride : width * 4, + width, height); + break; + case FOURCC_I400: + r = I400Copy(y, y_stride, + dst_sample, + dst_sample_stride ? dst_sample_stride : width, + width, height); + break; + case FOURCC_NV12: { + uint8* dst_uv = dst_sample + width * height; + r = I420ToNV12(y, y_stride, + u, u_stride, + v, v_stride, + dst_sample, + dst_sample_stride ? dst_sample_stride : width, + dst_uv, + dst_sample_stride ? dst_sample_stride : width, + width, height); + break; + } + case FOURCC_NV21: { + uint8* dst_vu = dst_sample + width * height; + r = I420ToNV21(y, y_stride, + u, u_stride, + v, v_stride, + dst_sample, + dst_sample_stride ? dst_sample_stride : width, + dst_vu, + dst_sample_stride ? dst_sample_stride : width, + width, height); + break; + } + // TODO(fbarchard): Add M420. + // Triplanar formats + // TODO(fbarchard): halfstride instead of halfwidth + case FOURCC_I420: + case FOURCC_YU12: + case FOURCC_YV12: { + int halfwidth = (width + 1) / 2; + int halfheight = (height + 1) / 2; + uint8* dst_u; + uint8* dst_v; + if (format == FOURCC_YV12) { + dst_v = dst_sample + width * height; + dst_u = dst_v + halfwidth * halfheight; + } else { + dst_u = dst_sample + width * height; + dst_v = dst_u + halfwidth * halfheight; + } + r = I420Copy(y, y_stride, + u, u_stride, + v, v_stride, + dst_sample, width, + dst_u, halfwidth, + dst_v, halfwidth, + width, height); + break; + } + case FOURCC_I422: + case FOURCC_YV16: { + int halfwidth = (width + 1) / 2; + uint8* dst_u; + uint8* dst_v; + if (format == FOURCC_YV16) { + dst_v = dst_sample + width * height; + dst_u = dst_v + halfwidth * height; + } else { + dst_u = dst_sample + width * height; + dst_v = dst_u + halfwidth * height; + } + r = I420ToI422(y, y_stride, + u, u_stride, + v, v_stride, + dst_sample, width, + dst_u, halfwidth, + dst_v, halfwidth, + width, height); + break; + } + case FOURCC_I444: + case FOURCC_YV24: { + uint8* dst_u; + uint8* dst_v; + if (format == FOURCC_YV24) { + dst_v = dst_sample + width * height; + dst_u = dst_v + width * height; + } else { + dst_u = dst_sample + width * height; + dst_v = dst_u + width * height; + } + r = I420ToI444(y, y_stride, + u, u_stride, + v, v_stride, + dst_sample, width, + dst_u, width, + dst_v, width, + width, height); + break; + } + case FOURCC_I411: { + int quarterwidth = (width + 3) / 4; + uint8* dst_u = dst_sample + width * height; + uint8* dst_v = dst_u + quarterwidth * height; + r = I420ToI411(y, y_stride, + u, u_stride, + v, v_stride, + dst_sample, width, + dst_u, quarterwidth, + dst_v, quarterwidth, + width, height); + break; + } + + // Formats not supported - MJPG, biplanar, some rgb formats. + default: + return -1; // unknown fourcc - return failure code. + } + return r; +} + +#ifdef __cplusplus +} // extern "C" +} // namespace libyuv +#endif diff --git a/media/libaom/src/third_party/libyuv/source/convert_from_argb.cc b/media/libaom/src/third_party/libyuv/source/convert_from_argb.cc new file mode 100644 index 000000000..8d1e97aec --- /dev/null +++ b/media/libaom/src/third_party/libyuv/source/convert_from_argb.cc @@ -0,0 +1,1301 @@ +/* + * Copyright 2012 The LibYuv Project Authors. All rights reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#include "libyuv/convert_from_argb.h" + +#include "libyuv/basic_types.h" +#include "libyuv/cpu_id.h" +#include "libyuv/planar_functions.h" +#include "libyuv/row.h" + +#ifdef __cplusplus +namespace libyuv { +extern "C" { +#endif + +// ARGB little endian (bgra in memory) to I444 +LIBYUV_API +int ARGBToI444(const uint8* src_argb, int src_stride_argb, + uint8* dst_y, int dst_stride_y, + uint8* dst_u, int dst_stride_u, + uint8* dst_v, int dst_stride_v, + int width, int height) { + int y; + void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int pix) = + ARGBToYRow_C; + void (*ARGBToUV444Row)(const uint8* src_argb, uint8* dst_u, uint8* dst_v, + int pix) = ARGBToUV444Row_C; + if (!src_argb || !dst_y || !dst_u || !dst_v || width <= 0 || height == 0) { + return -1; + } + if (height < 0) { + height = -height; + src_argb = src_argb + (height - 1) * src_stride_argb; + src_stride_argb = -src_stride_argb; + } + // Coalesce rows. + if (src_stride_argb == width * 4 && + dst_stride_y == width && + dst_stride_u == width && + dst_stride_v == width) { + width *= height; + height = 1; + src_stride_argb = dst_stride_y = dst_stride_u = dst_stride_v = 0; + } +#if defined(HAS_ARGBTOUV444ROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + ARGBToUV444Row = ARGBToUV444Row_Any_SSSE3; + if (IS_ALIGNED(width, 16)) { + ARGBToUV444Row = ARGBToUV444Row_SSSE3; + } + } +#endif +#if defined(HAS_ARGBTOUV444ROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + ARGBToUV444Row = ARGBToUV444Row_Any_NEON; + if (IS_ALIGNED(width, 8)) { + ARGBToUV444Row = ARGBToUV444Row_NEON; + } + } +#endif +#if defined(HAS_ARGBTOYROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + ARGBToYRow = ARGBToYRow_Any_SSSE3; + if (IS_ALIGNED(width, 16)) { + ARGBToYRow = ARGBToYRow_SSSE3; + } + } +#endif +#if defined(HAS_ARGBTOYROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + ARGBToYRow = ARGBToYRow_Any_AVX2; + if (IS_ALIGNED(width, 32)) { + ARGBToYRow = ARGBToYRow_AVX2; + } + } +#endif +#if defined(HAS_ARGBTOYROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + ARGBToYRow = ARGBToYRow_Any_NEON; + if (IS_ALIGNED(width, 8)) { + ARGBToYRow = ARGBToYRow_NEON; + } + } +#endif + + for (y = 0; y < height; ++y) { + ARGBToUV444Row(src_argb, dst_u, dst_v, width); + ARGBToYRow(src_argb, dst_y, width); + src_argb += src_stride_argb; + dst_y += dst_stride_y; + dst_u += dst_stride_u; + dst_v += dst_stride_v; + } + return 0; +} + +// ARGB little endian (bgra in memory) to I422 +LIBYUV_API +int ARGBToI422(const uint8* src_argb, int src_stride_argb, + uint8* dst_y, int dst_stride_y, + uint8* dst_u, int dst_stride_u, + uint8* dst_v, int dst_stride_v, + int width, int height) { + int y; + void (*ARGBToUV422Row)(const uint8* src_argb, uint8* dst_u, uint8* dst_v, + int pix) = ARGBToUV422Row_C; + void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int pix) = + ARGBToYRow_C; + if (!src_argb || !dst_y || !dst_u || !dst_v || width <= 0 || height == 0) { + return -1; + } + if (height < 0) { + height = -height; + src_argb = src_argb + (height - 1) * src_stride_argb; + src_stride_argb = -src_stride_argb; + } + // Coalesce rows. + if (src_stride_argb == width * 4 && + dst_stride_y == width && + dst_stride_u * 2 == width && + dst_stride_v * 2 == width) { + width *= height; + height = 1; + src_stride_argb = dst_stride_y = dst_stride_u = dst_stride_v = 0; + } +#if defined(HAS_ARGBTOUV422ROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + ARGBToUV422Row = ARGBToUV422Row_Any_SSSE3; + if (IS_ALIGNED(width, 16)) { + ARGBToUV422Row = ARGBToUV422Row_SSSE3; + } + } +#endif +#if defined(HAS_ARGBTOUV422ROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + ARGBToUV422Row = ARGBToUV422Row_Any_NEON; + if (IS_ALIGNED(width, 16)) { + ARGBToUV422Row = ARGBToUV422Row_NEON; + } + } +#endif +#if defined(HAS_ARGBTOYROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + ARGBToYRow = ARGBToYRow_Any_SSSE3; + if (IS_ALIGNED(width, 16)) { + ARGBToYRow = ARGBToYRow_SSSE3; + } + } +#endif +#if defined(HAS_ARGBTOYROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + ARGBToYRow = ARGBToYRow_Any_AVX2; + if (IS_ALIGNED(width, 32)) { + ARGBToYRow = ARGBToYRow_AVX2; + } + } +#endif +#if defined(HAS_ARGBTOYROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + ARGBToYRow = ARGBToYRow_Any_NEON; + if (IS_ALIGNED(width, 8)) { + ARGBToYRow = ARGBToYRow_NEON; + } + } +#endif + + for (y = 0; y < height; ++y) { + ARGBToUV422Row(src_argb, dst_u, dst_v, width); + ARGBToYRow(src_argb, dst_y, width); + src_argb += src_stride_argb; + dst_y += dst_stride_y; + dst_u += dst_stride_u; + dst_v += dst_stride_v; + } + return 0; +} + +// ARGB little endian (bgra in memory) to I411 +LIBYUV_API +int ARGBToI411(const uint8* src_argb, int src_stride_argb, + uint8* dst_y, int dst_stride_y, + uint8* dst_u, int dst_stride_u, + uint8* dst_v, int dst_stride_v, + int width, int height) { + int y; + void (*ARGBToUV411Row)(const uint8* src_argb, uint8* dst_u, uint8* dst_v, + int pix) = ARGBToUV411Row_C; + void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int pix) = + ARGBToYRow_C; + if (!src_argb || !dst_y || !dst_u || !dst_v || width <= 0 || height == 0) { + return -1; + } + if (height < 0) { + height = -height; + src_argb = src_argb + (height - 1) * src_stride_argb; + src_stride_argb = -src_stride_argb; + } + // Coalesce rows. + if (src_stride_argb == width * 4 && + dst_stride_y == width && + dst_stride_u * 4 == width && + dst_stride_v * 4 == width) { + width *= height; + height = 1; + src_stride_argb = dst_stride_y = dst_stride_u = dst_stride_v = 0; + } +#if defined(HAS_ARGBTOYROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + ARGBToYRow = ARGBToYRow_Any_SSSE3; + if (IS_ALIGNED(width, 16)) { + ARGBToYRow = ARGBToYRow_SSSE3; + } + } +#endif +#if defined(HAS_ARGBTOYROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + ARGBToYRow = ARGBToYRow_Any_AVX2; + if (IS_ALIGNED(width, 32)) { + ARGBToYRow = ARGBToYRow_AVX2; + } + } +#endif +#if defined(HAS_ARGBTOYROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + ARGBToYRow = ARGBToYRow_Any_NEON; + if (IS_ALIGNED(width, 8)) { + ARGBToYRow = ARGBToYRow_NEON; + } + } +#endif +#if defined(HAS_ARGBTOUV411ROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + ARGBToUV411Row = ARGBToUV411Row_Any_NEON; + if (IS_ALIGNED(width, 32)) { + ARGBToUV411Row = ARGBToUV411Row_NEON; + } + } +#endif + + for (y = 0; y < height; ++y) { + ARGBToUV411Row(src_argb, dst_u, dst_v, width); + ARGBToYRow(src_argb, dst_y, width); + src_argb += src_stride_argb; + dst_y += dst_stride_y; + dst_u += dst_stride_u; + dst_v += dst_stride_v; + } + return 0; +} + +LIBYUV_API +int ARGBToNV12(const uint8* src_argb, int src_stride_argb, + uint8* dst_y, int dst_stride_y, + uint8* dst_uv, int dst_stride_uv, + int width, int height) { + int y; + int halfwidth = (width + 1) >> 1; + void (*ARGBToUVRow)(const uint8* src_argb0, int src_stride_argb, + uint8* dst_u, uint8* dst_v, int width) = ARGBToUVRow_C; + void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int pix) = + ARGBToYRow_C; + void (*MergeUVRow_)(const uint8* src_u, const uint8* src_v, uint8* dst_uv, + int width) = MergeUVRow_C; + if (!src_argb || + !dst_y || !dst_uv || + width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + src_argb = src_argb + (height - 1) * src_stride_argb; + src_stride_argb = -src_stride_argb; + } +#if defined(HAS_ARGBTOYROW_SSSE3) && defined(HAS_ARGBTOUVROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + ARGBToUVRow = ARGBToUVRow_Any_SSSE3; + ARGBToYRow = ARGBToYRow_Any_SSSE3; + if (IS_ALIGNED(width, 16)) { + ARGBToUVRow = ARGBToUVRow_SSSE3; + ARGBToYRow = ARGBToYRow_SSSE3; + } + } +#endif +#if defined(HAS_ARGBTOYROW_AVX2) && defined(HAS_ARGBTOUVROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + ARGBToUVRow = ARGBToUVRow_Any_AVX2; + ARGBToYRow = ARGBToYRow_Any_AVX2; + if (IS_ALIGNED(width, 32)) { + ARGBToUVRow = ARGBToUVRow_AVX2; + ARGBToYRow = ARGBToYRow_AVX2; + } + } +#endif +#if defined(HAS_ARGBTOYROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + ARGBToYRow = ARGBToYRow_Any_NEON; + if (IS_ALIGNED(width, 8)) { + ARGBToYRow = ARGBToYRow_NEON; + } + } +#endif +#if defined(HAS_ARGBTOUVROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + ARGBToUVRow = ARGBToUVRow_Any_NEON; + if (IS_ALIGNED(width, 16)) { + ARGBToUVRow = ARGBToUVRow_NEON; + } + } +#endif +#if defined(HAS_MERGEUVROW_SSE2) + if (TestCpuFlag(kCpuHasSSE2)) { + MergeUVRow_ = MergeUVRow_Any_SSE2; + if (IS_ALIGNED(halfwidth, 16)) { + MergeUVRow_ = MergeUVRow_SSE2; + } + } +#endif +#if defined(HAS_MERGEUVROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + MergeUVRow_ = MergeUVRow_Any_AVX2; + if (IS_ALIGNED(halfwidth, 32)) { + MergeUVRow_ = MergeUVRow_AVX2; + } + } +#endif +#if defined(HAS_MERGEUVROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + MergeUVRow_ = MergeUVRow_Any_NEON; + if (IS_ALIGNED(halfwidth, 16)) { + MergeUVRow_ = MergeUVRow_NEON; + } + } +#endif + { + // Allocate a rows of uv. + align_buffer_64(row_u, ((halfwidth + 31) & ~31) * 2); + uint8* row_v = row_u + ((halfwidth + 31) & ~31); + + for (y = 0; y < height - 1; y += 2) { + ARGBToUVRow(src_argb, src_stride_argb, row_u, row_v, width); + MergeUVRow_(row_u, row_v, dst_uv, halfwidth); + ARGBToYRow(src_argb, dst_y, width); + ARGBToYRow(src_argb + src_stride_argb, dst_y + dst_stride_y, width); + src_argb += src_stride_argb * 2; + dst_y += dst_stride_y * 2; + dst_uv += dst_stride_uv; + } + if (height & 1) { + ARGBToUVRow(src_argb, 0, row_u, row_v, width); + MergeUVRow_(row_u, row_v, dst_uv, halfwidth); + ARGBToYRow(src_argb, dst_y, width); + } + free_aligned_buffer_64(row_u); + } + return 0; +} + +// Same as NV12 but U and V swapped. +LIBYUV_API +int ARGBToNV21(const uint8* src_argb, int src_stride_argb, + uint8* dst_y, int dst_stride_y, + uint8* dst_uv, int dst_stride_uv, + int width, int height) { + int y; + int halfwidth = (width + 1) >> 1; + void (*ARGBToUVRow)(const uint8* src_argb0, int src_stride_argb, + uint8* dst_u, uint8* dst_v, int width) = ARGBToUVRow_C; + void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int pix) = + ARGBToYRow_C; + void (*MergeUVRow_)(const uint8* src_u, const uint8* src_v, uint8* dst_uv, + int width) = MergeUVRow_C; + if (!src_argb || + !dst_y || !dst_uv || + width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + src_argb = src_argb + (height - 1) * src_stride_argb; + src_stride_argb = -src_stride_argb; + } +#if defined(HAS_ARGBTOYROW_SSSE3) && defined(HAS_ARGBTOUVROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + ARGBToUVRow = ARGBToUVRow_Any_SSSE3; + ARGBToYRow = ARGBToYRow_Any_SSSE3; + if (IS_ALIGNED(width, 16)) { + ARGBToUVRow = ARGBToUVRow_SSSE3; + ARGBToYRow = ARGBToYRow_SSSE3; + } + } +#endif +#if defined(HAS_ARGBTOYROW_AVX2) && defined(HAS_ARGBTOUVROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + ARGBToUVRow = ARGBToUVRow_Any_AVX2; + ARGBToYRow = ARGBToYRow_Any_AVX2; + if (IS_ALIGNED(width, 32)) { + ARGBToUVRow = ARGBToUVRow_AVX2; + ARGBToYRow = ARGBToYRow_AVX2; + } + } +#endif +#if defined(HAS_ARGBTOYROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + ARGBToYRow = ARGBToYRow_Any_NEON; + if (IS_ALIGNED(width, 8)) { + ARGBToYRow = ARGBToYRow_NEON; + } + } +#endif +#if defined(HAS_ARGBTOUVROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + ARGBToUVRow = ARGBToUVRow_Any_NEON; + if (IS_ALIGNED(width, 16)) { + ARGBToUVRow = ARGBToUVRow_NEON; + } + } +#endif +#if defined(HAS_MERGEUVROW_SSE2) + if (TestCpuFlag(kCpuHasSSE2)) { + MergeUVRow_ = MergeUVRow_Any_SSE2; + if (IS_ALIGNED(halfwidth, 16)) { + MergeUVRow_ = MergeUVRow_SSE2; + } + } +#endif +#if defined(HAS_MERGEUVROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + MergeUVRow_ = MergeUVRow_Any_AVX2; + if (IS_ALIGNED(halfwidth, 32)) { + MergeUVRow_ = MergeUVRow_AVX2; + } + } +#endif +#if defined(HAS_MERGEUVROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + MergeUVRow_ = MergeUVRow_Any_NEON; + if (IS_ALIGNED(halfwidth, 16)) { + MergeUVRow_ = MergeUVRow_NEON; + } + } +#endif + { + // Allocate a rows of uv. + align_buffer_64(row_u, ((halfwidth + 31) & ~31) * 2); + uint8* row_v = row_u + ((halfwidth + 31) & ~31); + + for (y = 0; y < height - 1; y += 2) { + ARGBToUVRow(src_argb, src_stride_argb, row_u, row_v, width); + MergeUVRow_(row_v, row_u, dst_uv, halfwidth); + ARGBToYRow(src_argb, dst_y, width); + ARGBToYRow(src_argb + src_stride_argb, dst_y + dst_stride_y, width); + src_argb += src_stride_argb * 2; + dst_y += dst_stride_y * 2; + dst_uv += dst_stride_uv; + } + if (height & 1) { + ARGBToUVRow(src_argb, 0, row_u, row_v, width); + MergeUVRow_(row_v, row_u, dst_uv, halfwidth); + ARGBToYRow(src_argb, dst_y, width); + } + free_aligned_buffer_64(row_u); + } + return 0; +} + +// Convert ARGB to YUY2. +LIBYUV_API +int ARGBToYUY2(const uint8* src_argb, int src_stride_argb, + uint8* dst_yuy2, int dst_stride_yuy2, + int width, int height) { + int y; + void (*ARGBToUV422Row)(const uint8* src_argb, uint8* dst_u, uint8* dst_v, + int pix) = ARGBToUV422Row_C; + void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int pix) = + ARGBToYRow_C; + void (*I422ToYUY2Row)(const uint8* src_y, const uint8* src_u, + const uint8* src_v, uint8* dst_yuy2, int width) = I422ToYUY2Row_C; + + if (!src_argb || !dst_yuy2 || + width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + dst_yuy2 = dst_yuy2 + (height - 1) * dst_stride_yuy2; + dst_stride_yuy2 = -dst_stride_yuy2; + } + // Coalesce rows. + if (src_stride_argb == width * 4 && + dst_stride_yuy2 == width * 2) { + width *= height; + height = 1; + src_stride_argb = dst_stride_yuy2 = 0; + } +#if defined(HAS_ARGBTOUV422ROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + ARGBToUV422Row = ARGBToUV422Row_Any_SSSE3; + if (IS_ALIGNED(width, 16)) { + ARGBToUV422Row = ARGBToUV422Row_SSSE3; + } + } +#endif +#if defined(HAS_ARGBTOUV422ROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + ARGBToUV422Row = ARGBToUV422Row_Any_NEON; + if (IS_ALIGNED(width, 16)) { + ARGBToUV422Row = ARGBToUV422Row_NEON; + } + } +#endif +#if defined(HAS_ARGBTOYROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + ARGBToYRow = ARGBToYRow_Any_SSSE3; + if (IS_ALIGNED(width, 16)) { + ARGBToYRow = ARGBToYRow_SSSE3; + } + } +#endif +#if defined(HAS_ARGBTOYROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + ARGBToYRow = ARGBToYRow_Any_AVX2; + if (IS_ALIGNED(width, 32)) { + ARGBToYRow = ARGBToYRow_AVX2; + } + } +#endif +#if defined(HAS_ARGBTOYROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + ARGBToYRow = ARGBToYRow_Any_NEON; + if (IS_ALIGNED(width, 8)) { + ARGBToYRow = ARGBToYRow_NEON; + } + } +#endif + +#if defined(HAS_I422TOYUY2ROW_SSE2) + if (TestCpuFlag(kCpuHasSSE2)) { + I422ToYUY2Row = I422ToYUY2Row_Any_SSE2; + if (IS_ALIGNED(width, 16)) { + I422ToYUY2Row = I422ToYUY2Row_SSE2; + } + } +#endif +#if defined(HAS_I422TOYUY2ROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + I422ToYUY2Row = I422ToYUY2Row_Any_NEON; + if (IS_ALIGNED(width, 16)) { + I422ToYUY2Row = I422ToYUY2Row_NEON; + } + } +#endif + + { + // Allocate a rows of yuv. + align_buffer_64(row_y, ((width + 63) & ~63) * 2); + uint8* row_u = row_y + ((width + 63) & ~63); + uint8* row_v = row_u + ((width + 63) & ~63) / 2; + + for (y = 0; y < height; ++y) { + ARGBToUV422Row(src_argb, row_u, row_v, width); + ARGBToYRow(src_argb, row_y, width); + I422ToYUY2Row(row_y, row_u, row_v, dst_yuy2, width); + src_argb += src_stride_argb; + dst_yuy2 += dst_stride_yuy2; + } + + free_aligned_buffer_64(row_y); + } + return 0; +} + +// Convert ARGB to UYVY. +LIBYUV_API +int ARGBToUYVY(const uint8* src_argb, int src_stride_argb, + uint8* dst_uyvy, int dst_stride_uyvy, + int width, int height) { + int y; + void (*ARGBToUV422Row)(const uint8* src_argb, uint8* dst_u, uint8* dst_v, + int pix) = ARGBToUV422Row_C; + void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int pix) = + ARGBToYRow_C; + void (*I422ToUYVYRow)(const uint8* src_y, const uint8* src_u, + const uint8* src_v, uint8* dst_uyvy, int width) = I422ToUYVYRow_C; + + if (!src_argb || !dst_uyvy || + width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + dst_uyvy = dst_uyvy + (height - 1) * dst_stride_uyvy; + dst_stride_uyvy = -dst_stride_uyvy; + } + // Coalesce rows. + if (src_stride_argb == width * 4 && + dst_stride_uyvy == width * 2) { + width *= height; + height = 1; + src_stride_argb = dst_stride_uyvy = 0; + } +#if defined(HAS_ARGBTOUV422ROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + ARGBToUV422Row = ARGBToUV422Row_Any_SSSE3; + if (IS_ALIGNED(width, 16)) { + ARGBToUV422Row = ARGBToUV422Row_SSSE3; + } + } +#endif +#if defined(HAS_ARGBTOUV422ROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + ARGBToUV422Row = ARGBToUV422Row_Any_NEON; + if (IS_ALIGNED(width, 16)) { + ARGBToUV422Row = ARGBToUV422Row_NEON; + } + } +#endif +#if defined(HAS_ARGBTOYROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + ARGBToYRow = ARGBToYRow_Any_SSSE3; + if (IS_ALIGNED(width, 16)) { + ARGBToYRow = ARGBToYRow_SSSE3; + } + } +#endif +#if defined(HAS_ARGBTOYROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + ARGBToYRow = ARGBToYRow_Any_AVX2; + if (IS_ALIGNED(width, 32)) { + ARGBToYRow = ARGBToYRow_AVX2; + } + } +#endif +#if defined(HAS_ARGBTOYROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + ARGBToYRow = ARGBToYRow_Any_NEON; + if (IS_ALIGNED(width, 8)) { + ARGBToYRow = ARGBToYRow_NEON; + } + } +#endif + +#if defined(HAS_I422TOUYVYROW_SSE2) + if (TestCpuFlag(kCpuHasSSE2)) { + I422ToUYVYRow = I422ToUYVYRow_Any_SSE2; + if (IS_ALIGNED(width, 16)) { + I422ToUYVYRow = I422ToUYVYRow_SSE2; + } + } +#endif +#if defined(HAS_I422TOUYVYROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + I422ToUYVYRow = I422ToUYVYRow_Any_NEON; + if (IS_ALIGNED(width, 16)) { + I422ToUYVYRow = I422ToUYVYRow_NEON; + } + } +#endif + + { + // Allocate a rows of yuv. + align_buffer_64(row_y, ((width + 63) & ~63) * 2); + uint8* row_u = row_y + ((width + 63) & ~63); + uint8* row_v = row_u + ((width + 63) & ~63) / 2; + + for (y = 0; y < height; ++y) { + ARGBToUV422Row(src_argb, row_u, row_v, width); + ARGBToYRow(src_argb, row_y, width); + I422ToUYVYRow(row_y, row_u, row_v, dst_uyvy, width); + src_argb += src_stride_argb; + dst_uyvy += dst_stride_uyvy; + } + + free_aligned_buffer_64(row_y); + } + return 0; +} + +// Convert ARGB to I400. +LIBYUV_API +int ARGBToI400(const uint8* src_argb, int src_stride_argb, + uint8* dst_y, int dst_stride_y, + int width, int height) { + int y; + void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int pix) = + ARGBToYRow_C; + if (!src_argb || !dst_y || width <= 0 || height == 0) { + return -1; + } + if (height < 0) { + height = -height; + src_argb = src_argb + (height - 1) * src_stride_argb; + src_stride_argb = -src_stride_argb; + } + // Coalesce rows. + if (src_stride_argb == width * 4 && + dst_stride_y == width) { + width *= height; + height = 1; + src_stride_argb = dst_stride_y = 0; + } +#if defined(HAS_ARGBTOYROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + ARGBToYRow = ARGBToYRow_Any_SSSE3; + if (IS_ALIGNED(width, 16)) { + ARGBToYRow = ARGBToYRow_SSSE3; + } + } +#endif +#if defined(HAS_ARGBTOYROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + ARGBToYRow = ARGBToYRow_Any_AVX2; + if (IS_ALIGNED(width, 32)) { + ARGBToYRow = ARGBToYRow_AVX2; + } + } +#endif +#if defined(HAS_ARGBTOYROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + ARGBToYRow = ARGBToYRow_Any_NEON; + if (IS_ALIGNED(width, 8)) { + ARGBToYRow = ARGBToYRow_NEON; + } + } +#endif + + for (y = 0; y < height; ++y) { + ARGBToYRow(src_argb, dst_y, width); + src_argb += src_stride_argb; + dst_y += dst_stride_y; + } + return 0; +} + +// Shuffle table for converting ARGB to RGBA. +static uvec8 kShuffleMaskARGBToRGBA = { + 3u, 0u, 1u, 2u, 7u, 4u, 5u, 6u, 11u, 8u, 9u, 10u, 15u, 12u, 13u, 14u +}; + +// Convert ARGB to RGBA. +LIBYUV_API +int ARGBToRGBA(const uint8* src_argb, int src_stride_argb, + uint8* dst_rgba, int dst_stride_rgba, + int width, int height) { + return ARGBShuffle(src_argb, src_stride_argb, + dst_rgba, dst_stride_rgba, + (const uint8*)(&kShuffleMaskARGBToRGBA), + width, height); +} + +// Convert ARGB To RGB24. +LIBYUV_API +int ARGBToRGB24(const uint8* src_argb, int src_stride_argb, + uint8* dst_rgb24, int dst_stride_rgb24, + int width, int height) { + int y; + void (*ARGBToRGB24Row)(const uint8* src_argb, uint8* dst_rgb, int pix) = + ARGBToRGB24Row_C; + if (!src_argb || !dst_rgb24 || width <= 0 || height == 0) { + return -1; + } + if (height < 0) { + height = -height; + src_argb = src_argb + (height - 1) * src_stride_argb; + src_stride_argb = -src_stride_argb; + } + // Coalesce rows. + if (src_stride_argb == width * 4 && + dst_stride_rgb24 == width * 3) { + width *= height; + height = 1; + src_stride_argb = dst_stride_rgb24 = 0; + } +#if defined(HAS_ARGBTORGB24ROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + ARGBToRGB24Row = ARGBToRGB24Row_Any_SSSE3; + if (IS_ALIGNED(width, 16)) { + ARGBToRGB24Row = ARGBToRGB24Row_SSSE3; + } + } +#endif +#if defined(HAS_ARGBTORGB24ROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + ARGBToRGB24Row = ARGBToRGB24Row_Any_NEON; + if (IS_ALIGNED(width, 8)) { + ARGBToRGB24Row = ARGBToRGB24Row_NEON; + } + } +#endif + + for (y = 0; y < height; ++y) { + ARGBToRGB24Row(src_argb, dst_rgb24, width); + src_argb += src_stride_argb; + dst_rgb24 += dst_stride_rgb24; + } + return 0; +} + +// Convert ARGB To RAW. +LIBYUV_API +int ARGBToRAW(const uint8* src_argb, int src_stride_argb, + uint8* dst_raw, int dst_stride_raw, + int width, int height) { + int y; + void (*ARGBToRAWRow)(const uint8* src_argb, uint8* dst_rgb, int pix) = + ARGBToRAWRow_C; + if (!src_argb || !dst_raw || width <= 0 || height == 0) { + return -1; + } + if (height < 0) { + height = -height; + src_argb = src_argb + (height - 1) * src_stride_argb; + src_stride_argb = -src_stride_argb; + } + // Coalesce rows. + if (src_stride_argb == width * 4 && + dst_stride_raw == width * 3) { + width *= height; + height = 1; + src_stride_argb = dst_stride_raw = 0; + } +#if defined(HAS_ARGBTORAWROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + ARGBToRAWRow = ARGBToRAWRow_Any_SSSE3; + if (IS_ALIGNED(width, 16)) { + ARGBToRAWRow = ARGBToRAWRow_SSSE3; + } + } +#endif +#if defined(HAS_ARGBTORAWROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + ARGBToRAWRow = ARGBToRAWRow_Any_NEON; + if (IS_ALIGNED(width, 8)) { + ARGBToRAWRow = ARGBToRAWRow_NEON; + } + } +#endif + + for (y = 0; y < height; ++y) { + ARGBToRAWRow(src_argb, dst_raw, width); + src_argb += src_stride_argb; + dst_raw += dst_stride_raw; + } + return 0; +} + +// Ordered 8x8 dither for 888 to 565. Values from 0 to 7. +static const uint8 kDither565_4x4[16] = { + 0, 4, 1, 5, + 6, 2, 7, 3, + 1, 5, 0, 4, + 7, 3, 6, 2, +}; + +// Convert ARGB To RGB565 with 4x4 dither matrix (16 bytes). +LIBYUV_API +int ARGBToRGB565Dither(const uint8* src_argb, int src_stride_argb, + uint8* dst_rgb565, int dst_stride_rgb565, + const uint8* dither4x4, int width, int height) { + int y; + void (*ARGBToRGB565DitherRow)(const uint8* src_argb, uint8* dst_rgb, + const uint32 dither4, int pix) = ARGBToRGB565DitherRow_C; + if (!src_argb || !dst_rgb565 || width <= 0 || height == 0) { + return -1; + } + if (height < 0) { + height = -height; + src_argb = src_argb + (height - 1) * src_stride_argb; + src_stride_argb = -src_stride_argb; + } + if (!dither4x4) { + dither4x4 = kDither565_4x4; + } +#if defined(HAS_ARGBTORGB565DITHERROW_SSE2) + if (TestCpuFlag(kCpuHasSSE2)) { + ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_Any_SSE2; + if (IS_ALIGNED(width, 4)) { + ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_SSE2; + } + } +#endif +#if defined(HAS_ARGBTORGB565DITHERROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_Any_AVX2; + if (IS_ALIGNED(width, 8)) { + ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_AVX2; + } + } +#endif +#if defined(HAS_ARGBTORGB565DITHERROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_Any_NEON; + if (IS_ALIGNED(width, 8)) { + ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_NEON; + } + } +#endif + for (y = 0; y < height; ++y) { + ARGBToRGB565DitherRow(src_argb, dst_rgb565, + *(uint32*)(dither4x4 + ((y & 3) << 2)), width); + src_argb += src_stride_argb; + dst_rgb565 += dst_stride_rgb565; + } + return 0; +} + +// Convert ARGB To RGB565. +// TODO(fbarchard): Consider using dither function low level with zeros. +LIBYUV_API +int ARGBToRGB565(const uint8* src_argb, int src_stride_argb, + uint8* dst_rgb565, int dst_stride_rgb565, + int width, int height) { + int y; + void (*ARGBToRGB565Row)(const uint8* src_argb, uint8* dst_rgb, int pix) = + ARGBToRGB565Row_C; + if (!src_argb || !dst_rgb565 || width <= 0 || height == 0) { + return -1; + } + if (height < 0) { + height = -height; + src_argb = src_argb + (height - 1) * src_stride_argb; + src_stride_argb = -src_stride_argb; + } + // Coalesce rows. + if (src_stride_argb == width * 4 && + dst_stride_rgb565 == width * 2) { + width *= height; + height = 1; + src_stride_argb = dst_stride_rgb565 = 0; + } +#if defined(HAS_ARGBTORGB565ROW_SSE2) + if (TestCpuFlag(kCpuHasSSE2)) { + ARGBToRGB565Row = ARGBToRGB565Row_Any_SSE2; + if (IS_ALIGNED(width, 4)) { + ARGBToRGB565Row = ARGBToRGB565Row_SSE2; + } + } +#endif +#if defined(HAS_ARGBTORGB565ROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + ARGBToRGB565Row = ARGBToRGB565Row_Any_AVX2; + if (IS_ALIGNED(width, 8)) { + ARGBToRGB565Row = ARGBToRGB565Row_AVX2; + } + } +#endif +#if defined(HAS_ARGBTORGB565ROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + ARGBToRGB565Row = ARGBToRGB565Row_Any_NEON; + if (IS_ALIGNED(width, 8)) { + ARGBToRGB565Row = ARGBToRGB565Row_NEON; + } + } +#endif + + for (y = 0; y < height; ++y) { + ARGBToRGB565Row(src_argb, dst_rgb565, width); + src_argb += src_stride_argb; + dst_rgb565 += dst_stride_rgb565; + } + return 0; +} + +// Convert ARGB To ARGB1555. +LIBYUV_API +int ARGBToARGB1555(const uint8* src_argb, int src_stride_argb, + uint8* dst_argb1555, int dst_stride_argb1555, + int width, int height) { + int y; + void (*ARGBToARGB1555Row)(const uint8* src_argb, uint8* dst_rgb, int pix) = + ARGBToARGB1555Row_C; + if (!src_argb || !dst_argb1555 || width <= 0 || height == 0) { + return -1; + } + if (height < 0) { + height = -height; + src_argb = src_argb + (height - 1) * src_stride_argb; + src_stride_argb = -src_stride_argb; + } + // Coalesce rows. + if (src_stride_argb == width * 4 && + dst_stride_argb1555 == width * 2) { + width *= height; + height = 1; + src_stride_argb = dst_stride_argb1555 = 0; + } +#if defined(HAS_ARGBTOARGB1555ROW_SSE2) + if (TestCpuFlag(kCpuHasSSE2)) { + ARGBToARGB1555Row = ARGBToARGB1555Row_Any_SSE2; + if (IS_ALIGNED(width, 4)) { + ARGBToARGB1555Row = ARGBToARGB1555Row_SSE2; + } + } +#endif +#if defined(HAS_ARGBTOARGB1555ROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + ARGBToARGB1555Row = ARGBToARGB1555Row_Any_AVX2; + if (IS_ALIGNED(width, 8)) { + ARGBToARGB1555Row = ARGBToARGB1555Row_AVX2; + } + } +#endif +#if defined(HAS_ARGBTOARGB1555ROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + ARGBToARGB1555Row = ARGBToARGB1555Row_Any_NEON; + if (IS_ALIGNED(width, 8)) { + ARGBToARGB1555Row = ARGBToARGB1555Row_NEON; + } + } +#endif + + for (y = 0; y < height; ++y) { + ARGBToARGB1555Row(src_argb, dst_argb1555, width); + src_argb += src_stride_argb; + dst_argb1555 += dst_stride_argb1555; + } + return 0; +} + +// Convert ARGB To ARGB4444. +LIBYUV_API +int ARGBToARGB4444(const uint8* src_argb, int src_stride_argb, + uint8* dst_argb4444, int dst_stride_argb4444, + int width, int height) { + int y; + void (*ARGBToARGB4444Row)(const uint8* src_argb, uint8* dst_rgb, int pix) = + ARGBToARGB4444Row_C; + if (!src_argb || !dst_argb4444 || width <= 0 || height == 0) { + return -1; + } + if (height < 0) { + height = -height; + src_argb = src_argb + (height - 1) * src_stride_argb; + src_stride_argb = -src_stride_argb; + } + // Coalesce rows. + if (src_stride_argb == width * 4 && + dst_stride_argb4444 == width * 2) { + width *= height; + height = 1; + src_stride_argb = dst_stride_argb4444 = 0; + } +#if defined(HAS_ARGBTOARGB4444ROW_SSE2) + if (TestCpuFlag(kCpuHasSSE2)) { + ARGBToARGB4444Row = ARGBToARGB4444Row_Any_SSE2; + if (IS_ALIGNED(width, 4)) { + ARGBToARGB4444Row = ARGBToARGB4444Row_SSE2; + } + } +#endif +#if defined(HAS_ARGBTOARGB4444ROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + ARGBToARGB4444Row = ARGBToARGB4444Row_Any_AVX2; + if (IS_ALIGNED(width, 8)) { + ARGBToARGB4444Row = ARGBToARGB4444Row_AVX2; + } + } +#endif +#if defined(HAS_ARGBTOARGB4444ROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + ARGBToARGB4444Row = ARGBToARGB4444Row_Any_NEON; + if (IS_ALIGNED(width, 8)) { + ARGBToARGB4444Row = ARGBToARGB4444Row_NEON; + } + } +#endif + + for (y = 0; y < height; ++y) { + ARGBToARGB4444Row(src_argb, dst_argb4444, width); + src_argb += src_stride_argb; + dst_argb4444 += dst_stride_argb4444; + } + return 0; +} + +// Convert ARGB to J420. (JPeg full range I420). +LIBYUV_API +int ARGBToJ420(const uint8* src_argb, int src_stride_argb, + uint8* dst_yj, int dst_stride_yj, + uint8* dst_u, int dst_stride_u, + uint8* dst_v, int dst_stride_v, + int width, int height) { + int y; + void (*ARGBToUVJRow)(const uint8* src_argb0, int src_stride_argb, + uint8* dst_u, uint8* dst_v, int width) = ARGBToUVJRow_C; + void (*ARGBToYJRow)(const uint8* src_argb, uint8* dst_yj, int pix) = + ARGBToYJRow_C; + if (!src_argb || + !dst_yj || !dst_u || !dst_v || + width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + src_argb = src_argb + (height - 1) * src_stride_argb; + src_stride_argb = -src_stride_argb; + } +#if defined(HAS_ARGBTOYJROW_SSSE3) && defined(HAS_ARGBTOUVJROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + ARGBToUVJRow = ARGBToUVJRow_Any_SSSE3; + ARGBToYJRow = ARGBToYJRow_Any_SSSE3; + if (IS_ALIGNED(width, 16)) { + ARGBToUVJRow = ARGBToUVJRow_SSSE3; + ARGBToYJRow = ARGBToYJRow_SSSE3; + } + } +#endif +#if defined(HAS_ARGBTOYJROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + ARGBToYJRow = ARGBToYJRow_Any_AVX2; + if (IS_ALIGNED(width, 32)) { + ARGBToYJRow = ARGBToYJRow_AVX2; + } + } +#endif +#if defined(HAS_ARGBTOYJROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + ARGBToYJRow = ARGBToYJRow_Any_NEON; + if (IS_ALIGNED(width, 8)) { + ARGBToYJRow = ARGBToYJRow_NEON; + } + } +#endif +#if defined(HAS_ARGBTOUVJROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + ARGBToUVJRow = ARGBToUVJRow_Any_NEON; + if (IS_ALIGNED(width, 16)) { + ARGBToUVJRow = ARGBToUVJRow_NEON; + } + } +#endif + + for (y = 0; y < height - 1; y += 2) { + ARGBToUVJRow(src_argb, src_stride_argb, dst_u, dst_v, width); + ARGBToYJRow(src_argb, dst_yj, width); + ARGBToYJRow(src_argb + src_stride_argb, dst_yj + dst_stride_yj, width); + src_argb += src_stride_argb * 2; + dst_yj += dst_stride_yj * 2; + dst_u += dst_stride_u; + dst_v += dst_stride_v; + } + if (height & 1) { + ARGBToUVJRow(src_argb, 0, dst_u, dst_v, width); + ARGBToYJRow(src_argb, dst_yj, width); + } + return 0; +} + +// ARGB little endian (bgra in memory) to J422 +LIBYUV_API +int ARGBToJ422(const uint8* src_argb, int src_stride_argb, + uint8* dst_y, int dst_stride_y, + uint8* dst_u, int dst_stride_u, + uint8* dst_v, int dst_stride_v, + int width, int height) { + int y; + void (*ARGBToUVJ422Row)(const uint8* src_argb, uint8* dst_u, uint8* dst_v, + int pix) = ARGBToUVJ422Row_C; + void (*ARGBToYJRow)(const uint8* src_argb, uint8* dst_y, int pix) = + ARGBToYJRow_C; + if (!src_argb || !dst_y || !dst_u || !dst_v || width <= 0 || height == 0) { + return -1; + } + if (height < 0) { + height = -height; + src_argb = src_argb + (height - 1) * src_stride_argb; + src_stride_argb = -src_stride_argb; + } + // Coalesce rows. + if (src_stride_argb == width * 4 && + dst_stride_y == width && + dst_stride_u * 2 == width && + dst_stride_v * 2 == width) { + width *= height; + height = 1; + src_stride_argb = dst_stride_y = dst_stride_u = dst_stride_v = 0; + } +#if defined(HAS_ARGBTOUVJ422ROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + ARGBToUVJ422Row = ARGBToUVJ422Row_Any_SSSE3; + if (IS_ALIGNED(width, 16)) { + ARGBToUVJ422Row = ARGBToUVJ422Row_SSSE3; + } + } +#endif +#if defined(HAS_ARGBTOUVJ422ROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + ARGBToUVJ422Row = ARGBToUVJ422Row_Any_NEON; + if (IS_ALIGNED(width, 16)) { + ARGBToUVJ422Row = ARGBToUVJ422Row_NEON; + } + } +#endif + +#if defined(HAS_ARGBTOYJROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + ARGBToYJRow = ARGBToYJRow_Any_SSSE3; + if (IS_ALIGNED(width, 16)) { + ARGBToYJRow = ARGBToYJRow_SSSE3; + } + } +#endif +#if defined(HAS_ARGBTOYJROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + ARGBToYJRow = ARGBToYJRow_Any_AVX2; + if (IS_ALIGNED(width, 32)) { + ARGBToYJRow = ARGBToYJRow_AVX2; + } + } +#endif +#if defined(HAS_ARGBTOYJROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + ARGBToYJRow = ARGBToYJRow_Any_NEON; + if (IS_ALIGNED(width, 8)) { + ARGBToYJRow = ARGBToYJRow_NEON; + } + } +#endif + + for (y = 0; y < height; ++y) { + ARGBToUVJ422Row(src_argb, dst_u, dst_v, width); + ARGBToYJRow(src_argb, dst_y, width); + src_argb += src_stride_argb; + dst_y += dst_stride_y; + dst_u += dst_stride_u; + dst_v += dst_stride_v; + } + return 0; +} + +// Convert ARGB to J400. +LIBYUV_API +int ARGBToJ400(const uint8* src_argb, int src_stride_argb, + uint8* dst_yj, int dst_stride_yj, + int width, int height) { + int y; + void (*ARGBToYJRow)(const uint8* src_argb, uint8* dst_yj, int pix) = + ARGBToYJRow_C; + if (!src_argb || !dst_yj || width <= 0 || height == 0) { + return -1; + } + if (height < 0) { + height = -height; + src_argb = src_argb + (height - 1) * src_stride_argb; + src_stride_argb = -src_stride_argb; + } + // Coalesce rows. + if (src_stride_argb == width * 4 && + dst_stride_yj == width) { + width *= height; + height = 1; + src_stride_argb = dst_stride_yj = 0; + } +#if defined(HAS_ARGBTOYJROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + ARGBToYJRow = ARGBToYJRow_Any_SSSE3; + if (IS_ALIGNED(width, 16)) { + ARGBToYJRow = ARGBToYJRow_SSSE3; + } + } +#endif +#if defined(HAS_ARGBTOYJROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + ARGBToYJRow = ARGBToYJRow_Any_AVX2; + if (IS_ALIGNED(width, 32)) { + ARGBToYJRow = ARGBToYJRow_AVX2; + } + } +#endif +#if defined(HAS_ARGBTOYJROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + ARGBToYJRow = ARGBToYJRow_Any_NEON; + if (IS_ALIGNED(width, 8)) { + ARGBToYJRow = ARGBToYJRow_NEON; + } + } +#endif + + for (y = 0; y < height; ++y) { + ARGBToYJRow(src_argb, dst_yj, width); + src_argb += src_stride_argb; + dst_yj += dst_stride_yj; + } + return 0; +} + +#ifdef __cplusplus +} // extern "C" +} // namespace libyuv +#endif diff --git a/media/libaom/src/third_party/libyuv/source/convert_jpeg.cc b/media/libaom/src/third_party/libyuv/source/convert_jpeg.cc new file mode 100644 index 000000000..bcb980f7f --- /dev/null +++ b/media/libaom/src/third_party/libyuv/source/convert_jpeg.cc @@ -0,0 +1,392 @@ +/* + * Copyright 2011 The LibYuv Project Authors. All rights reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#include "libyuv/convert.h" + +#ifdef HAVE_JPEG +#include "libyuv/mjpeg_decoder.h" +#endif + +#ifdef __cplusplus +namespace libyuv { +extern "C" { +#endif + +#ifdef HAVE_JPEG +struct I420Buffers { + uint8* y; + int y_stride; + uint8* u; + int u_stride; + uint8* v; + int v_stride; + int w; + int h; +}; + +static void JpegCopyI420(void* opaque, + const uint8* const* data, + const int* strides, + int rows) { + I420Buffers* dest = (I420Buffers*)(opaque); + I420Copy(data[0], strides[0], + data[1], strides[1], + data[2], strides[2], + dest->y, dest->y_stride, + dest->u, dest->u_stride, + dest->v, dest->v_stride, + dest->w, rows); + dest->y += rows * dest->y_stride; + dest->u += ((rows + 1) >> 1) * dest->u_stride; + dest->v += ((rows + 1) >> 1) * dest->v_stride; + dest->h -= rows; +} + +static void JpegI422ToI420(void* opaque, + const uint8* const* data, + const int* strides, + int rows) { + I420Buffers* dest = (I420Buffers*)(opaque); + I422ToI420(data[0], strides[0], + data[1], strides[1], + data[2], strides[2], + dest->y, dest->y_stride, + dest->u, dest->u_stride, + dest->v, dest->v_stride, + dest->w, rows); + dest->y += rows * dest->y_stride; + dest->u += ((rows + 1) >> 1) * dest->u_stride; + dest->v += ((rows + 1) >> 1) * dest->v_stride; + dest->h -= rows; +} + +static void JpegI444ToI420(void* opaque, + const uint8* const* data, + const int* strides, + int rows) { + I420Buffers* dest = (I420Buffers*)(opaque); + I444ToI420(data[0], strides[0], + data[1], strides[1], + data[2], strides[2], + dest->y, dest->y_stride, + dest->u, dest->u_stride, + dest->v, dest->v_stride, + dest->w, rows); + dest->y += rows * dest->y_stride; + dest->u += ((rows + 1) >> 1) * dest->u_stride; + dest->v += ((rows + 1) >> 1) * dest->v_stride; + dest->h -= rows; +} + +static void JpegI411ToI420(void* opaque, + const uint8* const* data, + const int* strides, + int rows) { + I420Buffers* dest = (I420Buffers*)(opaque); + I411ToI420(data[0], strides[0], + data[1], strides[1], + data[2], strides[2], + dest->y, dest->y_stride, + dest->u, dest->u_stride, + dest->v, dest->v_stride, + dest->w, rows); + dest->y += rows * dest->y_stride; + dest->u += ((rows + 1) >> 1) * dest->u_stride; + dest->v += ((rows + 1) >> 1) * dest->v_stride; + dest->h -= rows; +} + +static void JpegI400ToI420(void* opaque, + const uint8* const* data, + const int* strides, + int rows) { + I420Buffers* dest = (I420Buffers*)(opaque); + I400ToI420(data[0], strides[0], + dest->y, dest->y_stride, + dest->u, dest->u_stride, + dest->v, dest->v_stride, + dest->w, rows); + dest->y += rows * dest->y_stride; + dest->u += ((rows + 1) >> 1) * dest->u_stride; + dest->v += ((rows + 1) >> 1) * dest->v_stride; + dest->h -= rows; +} + +// Query size of MJPG in pixels. +LIBYUV_API +int MJPGSize(const uint8* sample, size_t sample_size, + int* width, int* height) { + MJpegDecoder mjpeg_decoder; + LIBYUV_BOOL ret = mjpeg_decoder.LoadFrame(sample, sample_size); + if (ret) { + *width = mjpeg_decoder.GetWidth(); + *height = mjpeg_decoder.GetHeight(); + } + mjpeg_decoder.UnloadFrame(); + return ret ? 0 : -1; // -1 for runtime failure. +} + +// MJPG (Motion JPeg) to I420 +// TODO(fbarchard): review w and h requirement. dw and dh may be enough. +LIBYUV_API +int MJPGToI420(const uint8* sample, + size_t sample_size, + uint8* y, int y_stride, + uint8* u, int u_stride, + uint8* v, int v_stride, + int w, int h, + int dw, int dh) { + if (sample_size == kUnknownDataSize) { + // ERROR: MJPEG frame size unknown + return -1; + } + + // TODO(fbarchard): Port MJpeg to C. + MJpegDecoder mjpeg_decoder; + LIBYUV_BOOL ret = mjpeg_decoder.LoadFrame(sample, sample_size); + if (ret && (mjpeg_decoder.GetWidth() != w || + mjpeg_decoder.GetHeight() != h)) { + // ERROR: MJPEG frame has unexpected dimensions + mjpeg_decoder.UnloadFrame(); + return 1; // runtime failure + } + if (ret) { + I420Buffers bufs = { y, y_stride, u, u_stride, v, v_stride, dw, dh }; + // YUV420 + if (mjpeg_decoder.GetColorSpace() == + MJpegDecoder::kColorSpaceYCbCr && + mjpeg_decoder.GetNumComponents() == 3 && + mjpeg_decoder.GetVertSampFactor(0) == 2 && + mjpeg_decoder.GetHorizSampFactor(0) == 2 && + mjpeg_decoder.GetVertSampFactor(1) == 1 && + mjpeg_decoder.GetHorizSampFactor(1) == 1 && + mjpeg_decoder.GetVertSampFactor(2) == 1 && + mjpeg_decoder.GetHorizSampFactor(2) == 1) { + ret = mjpeg_decoder.DecodeToCallback(&JpegCopyI420, &bufs, dw, dh); + // YUV422 + } else if (mjpeg_decoder.GetColorSpace() == + MJpegDecoder::kColorSpaceYCbCr && + mjpeg_decoder.GetNumComponents() == 3 && + mjpeg_decoder.GetVertSampFactor(0) == 1 && + mjpeg_decoder.GetHorizSampFactor(0) == 2 && + mjpeg_decoder.GetVertSampFactor(1) == 1 && + mjpeg_decoder.GetHorizSampFactor(1) == 1 && + mjpeg_decoder.GetVertSampFactor(2) == 1 && + mjpeg_decoder.GetHorizSampFactor(2) == 1) { + ret = mjpeg_decoder.DecodeToCallback(&JpegI422ToI420, &bufs, dw, dh); + // YUV444 + } else if (mjpeg_decoder.GetColorSpace() == + MJpegDecoder::kColorSpaceYCbCr && + mjpeg_decoder.GetNumComponents() == 3 && + mjpeg_decoder.GetVertSampFactor(0) == 1 && + mjpeg_decoder.GetHorizSampFactor(0) == 1 && + mjpeg_decoder.GetVertSampFactor(1) == 1 && + mjpeg_decoder.GetHorizSampFactor(1) == 1 && + mjpeg_decoder.GetVertSampFactor(2) == 1 && + mjpeg_decoder.GetHorizSampFactor(2) == 1) { + ret = mjpeg_decoder.DecodeToCallback(&JpegI444ToI420, &bufs, dw, dh); + // YUV411 + } else if (mjpeg_decoder.GetColorSpace() == + MJpegDecoder::kColorSpaceYCbCr && + mjpeg_decoder.GetNumComponents() == 3 && + mjpeg_decoder.GetVertSampFactor(0) == 1 && + mjpeg_decoder.GetHorizSampFactor(0) == 4 && + mjpeg_decoder.GetVertSampFactor(1) == 1 && + mjpeg_decoder.GetHorizSampFactor(1) == 1 && + mjpeg_decoder.GetVertSampFactor(2) == 1 && + mjpeg_decoder.GetHorizSampFactor(2) == 1) { + ret = mjpeg_decoder.DecodeToCallback(&JpegI411ToI420, &bufs, dw, dh); + // YUV400 + } else if (mjpeg_decoder.GetColorSpace() == + MJpegDecoder::kColorSpaceGrayscale && + mjpeg_decoder.GetNumComponents() == 1 && + mjpeg_decoder.GetVertSampFactor(0) == 1 && + mjpeg_decoder.GetHorizSampFactor(0) == 1) { + ret = mjpeg_decoder.DecodeToCallback(&JpegI400ToI420, &bufs, dw, dh); + } else { + // TODO(fbarchard): Implement conversion for any other colorspace/sample + // factors that occur in practice. 411 is supported by libjpeg + // ERROR: Unable to convert MJPEG frame because format is not supported + mjpeg_decoder.UnloadFrame(); + return 1; + } + } + return ret ? 0 : 1; +} + +#ifdef HAVE_JPEG +struct ARGBBuffers { + uint8* argb; + int argb_stride; + int w; + int h; +}; + +static void JpegI420ToARGB(void* opaque, + const uint8* const* data, + const int* strides, + int rows) { + ARGBBuffers* dest = (ARGBBuffers*)(opaque); + I420ToARGB(data[0], strides[0], + data[1], strides[1], + data[2], strides[2], + dest->argb, dest->argb_stride, + dest->w, rows); + dest->argb += rows * dest->argb_stride; + dest->h -= rows; +} + +static void JpegI422ToARGB(void* opaque, + const uint8* const* data, + const int* strides, + int rows) { + ARGBBuffers* dest = (ARGBBuffers*)(opaque); + I422ToARGB(data[0], strides[0], + data[1], strides[1], + data[2], strides[2], + dest->argb, dest->argb_stride, + dest->w, rows); + dest->argb += rows * dest->argb_stride; + dest->h -= rows; +} + +static void JpegI444ToARGB(void* opaque, + const uint8* const* data, + const int* strides, + int rows) { + ARGBBuffers* dest = (ARGBBuffers*)(opaque); + I444ToARGB(data[0], strides[0], + data[1], strides[1], + data[2], strides[2], + dest->argb, dest->argb_stride, + dest->w, rows); + dest->argb += rows * dest->argb_stride; + dest->h -= rows; +} + +static void JpegI411ToARGB(void* opaque, + const uint8* const* data, + const int* strides, + int rows) { + ARGBBuffers* dest = (ARGBBuffers*)(opaque); + I411ToARGB(data[0], strides[0], + data[1], strides[1], + data[2], strides[2], + dest->argb, dest->argb_stride, + dest->w, rows); + dest->argb += rows * dest->argb_stride; + dest->h -= rows; +} + +static void JpegI400ToARGB(void* opaque, + const uint8* const* data, + const int* strides, + int rows) { + ARGBBuffers* dest = (ARGBBuffers*)(opaque); + I400ToARGB(data[0], strides[0], + dest->argb, dest->argb_stride, + dest->w, rows); + dest->argb += rows * dest->argb_stride; + dest->h -= rows; +} + +// MJPG (Motion JPeg) to ARGB +// TODO(fbarchard): review w and h requirement. dw and dh may be enough. +LIBYUV_API +int MJPGToARGB(const uint8* sample, + size_t sample_size, + uint8* argb, int argb_stride, + int w, int h, + int dw, int dh) { + if (sample_size == kUnknownDataSize) { + // ERROR: MJPEG frame size unknown + return -1; + } + + // TODO(fbarchard): Port MJpeg to C. + MJpegDecoder mjpeg_decoder; + LIBYUV_BOOL ret = mjpeg_decoder.LoadFrame(sample, sample_size); + if (ret && (mjpeg_decoder.GetWidth() != w || + mjpeg_decoder.GetHeight() != h)) { + // ERROR: MJPEG frame has unexpected dimensions + mjpeg_decoder.UnloadFrame(); + return 1; // runtime failure + } + if (ret) { + ARGBBuffers bufs = { argb, argb_stride, dw, dh }; + // YUV420 + if (mjpeg_decoder.GetColorSpace() == + MJpegDecoder::kColorSpaceYCbCr && + mjpeg_decoder.GetNumComponents() == 3 && + mjpeg_decoder.GetVertSampFactor(0) == 2 && + mjpeg_decoder.GetHorizSampFactor(0) == 2 && + mjpeg_decoder.GetVertSampFactor(1) == 1 && + mjpeg_decoder.GetHorizSampFactor(1) == 1 && + mjpeg_decoder.GetVertSampFactor(2) == 1 && + mjpeg_decoder.GetHorizSampFactor(2) == 1) { + ret = mjpeg_decoder.DecodeToCallback(&JpegI420ToARGB, &bufs, dw, dh); + // YUV422 + } else if (mjpeg_decoder.GetColorSpace() == + MJpegDecoder::kColorSpaceYCbCr && + mjpeg_decoder.GetNumComponents() == 3 && + mjpeg_decoder.GetVertSampFactor(0) == 1 && + mjpeg_decoder.GetHorizSampFactor(0) == 2 && + mjpeg_decoder.GetVertSampFactor(1) == 1 && + mjpeg_decoder.GetHorizSampFactor(1) == 1 && + mjpeg_decoder.GetVertSampFactor(2) == 1 && + mjpeg_decoder.GetHorizSampFactor(2) == 1) { + ret = mjpeg_decoder.DecodeToCallback(&JpegI422ToARGB, &bufs, dw, dh); + // YUV444 + } else if (mjpeg_decoder.GetColorSpace() == + MJpegDecoder::kColorSpaceYCbCr && + mjpeg_decoder.GetNumComponents() == 3 && + mjpeg_decoder.GetVertSampFactor(0) == 1 && + mjpeg_decoder.GetHorizSampFactor(0) == 1 && + mjpeg_decoder.GetVertSampFactor(1) == 1 && + mjpeg_decoder.GetHorizSampFactor(1) == 1 && + mjpeg_decoder.GetVertSampFactor(2) == 1 && + mjpeg_decoder.GetHorizSampFactor(2) == 1) { + ret = mjpeg_decoder.DecodeToCallback(&JpegI444ToARGB, &bufs, dw, dh); + // YUV411 + } else if (mjpeg_decoder.GetColorSpace() == + MJpegDecoder::kColorSpaceYCbCr && + mjpeg_decoder.GetNumComponents() == 3 && + mjpeg_decoder.GetVertSampFactor(0) == 1 && + mjpeg_decoder.GetHorizSampFactor(0) == 4 && + mjpeg_decoder.GetVertSampFactor(1) == 1 && + mjpeg_decoder.GetHorizSampFactor(1) == 1 && + mjpeg_decoder.GetVertSampFactor(2) == 1 && + mjpeg_decoder.GetHorizSampFactor(2) == 1) { + ret = mjpeg_decoder.DecodeToCallback(&JpegI411ToARGB, &bufs, dw, dh); + // YUV400 + } else if (mjpeg_decoder.GetColorSpace() == + MJpegDecoder::kColorSpaceGrayscale && + mjpeg_decoder.GetNumComponents() == 1 && + mjpeg_decoder.GetVertSampFactor(0) == 1 && + mjpeg_decoder.GetHorizSampFactor(0) == 1) { + ret = mjpeg_decoder.DecodeToCallback(&JpegI400ToARGB, &bufs, dw, dh); + } else { + // TODO(fbarchard): Implement conversion for any other colorspace/sample + // factors that occur in practice. 411 is supported by libjpeg + // ERROR: Unable to convert MJPEG frame because format is not supported + mjpeg_decoder.UnloadFrame(); + return 1; + } + } + return ret ? 0 : 1; +} +#endif + +#endif + +#ifdef __cplusplus +} // extern "C" +} // namespace libyuv +#endif diff --git a/media/libaom/src/third_party/libyuv/source/convert_to_argb.cc b/media/libaom/src/third_party/libyuv/source/convert_to_argb.cc new file mode 100644 index 000000000..af829fbd3 --- /dev/null +++ b/media/libaom/src/third_party/libyuv/source/convert_to_argb.cc @@ -0,0 +1,306 @@ +/* + * Copyright 2011 The LibYuv Project Authors. All rights reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#include "libyuv/convert_argb.h" + +#include "libyuv/cpu_id.h" +#ifdef HAVE_JPEG +#include "libyuv/mjpeg_decoder.h" +#endif +#include "libyuv/rotate_argb.h" +#include "libyuv/row.h" +#include "libyuv/video_common.h" + +#ifdef __cplusplus +namespace libyuv { +extern "C" { +#endif + +// Convert camera sample to I420 with cropping, rotation and vertical flip. +// src_width is used for source stride computation +// src_height is used to compute location of planes, and indicate inversion +// sample_size is measured in bytes and is the size of the frame. +// With MJPEG it is the compressed size of the frame. +LIBYUV_API +int ConvertToARGB(const uint8* sample, size_t sample_size, + uint8* crop_argb, int argb_stride, + int crop_x, int crop_y, + int src_width, int src_height, + int crop_width, int crop_height, + enum RotationMode rotation, + uint32 fourcc) { + uint32 format = CanonicalFourCC(fourcc); + int aligned_src_width = (src_width + 1) & ~1; + const uint8* src; + const uint8* src_uv; + int abs_src_height = (src_height < 0) ? -src_height : src_height; + int inv_crop_height = (crop_height < 0) ? -crop_height : crop_height; + int r = 0; + + // One pass rotation is available for some formats. For the rest, convert + // to I420 (with optional vertical flipping) into a temporary I420 buffer, + // and then rotate the I420 to the final destination buffer. + // For in-place conversion, if destination crop_argb is same as source sample, + // also enable temporary buffer. + LIBYUV_BOOL need_buf = (rotation && format != FOURCC_ARGB) || + crop_argb == sample; + uint8* tmp_argb = crop_argb; + int tmp_argb_stride = argb_stride; + uint8* rotate_buffer = NULL; + int abs_crop_height = (crop_height < 0) ? -crop_height : crop_height; + + if (crop_argb == NULL || sample == NULL || + src_width <= 0 || crop_width <= 0 || + src_height == 0 || crop_height == 0) { + return -1; + } + if (src_height < 0) { + inv_crop_height = -inv_crop_height; + } + + if (need_buf) { + int argb_size = crop_width * abs_crop_height * 4; + rotate_buffer = (uint8*)malloc(argb_size); + if (!rotate_buffer) { + return 1; // Out of memory runtime error. + } + crop_argb = rotate_buffer; + argb_stride = crop_width; + } + + switch (format) { + // Single plane formats + case FOURCC_YUY2: + src = sample + (aligned_src_width * crop_y + crop_x) * 2; + r = YUY2ToARGB(src, aligned_src_width * 2, + crop_argb, argb_stride, + crop_width, inv_crop_height); + break; + case FOURCC_UYVY: + src = sample + (aligned_src_width * crop_y + crop_x) * 2; + r = UYVYToARGB(src, aligned_src_width * 2, + crop_argb, argb_stride, + crop_width, inv_crop_height); + break; + case FOURCC_24BG: + src = sample + (src_width * crop_y + crop_x) * 3; + r = RGB24ToARGB(src, src_width * 3, + crop_argb, argb_stride, + crop_width, inv_crop_height); + break; + case FOURCC_RAW: + src = sample + (src_width * crop_y + crop_x) * 3; + r = RAWToARGB(src, src_width * 3, + crop_argb, argb_stride, + crop_width, inv_crop_height); + break; + case FOURCC_ARGB: + src = sample + (src_width * crop_y + crop_x) * 4; + r = ARGBToARGB(src, src_width * 4, + crop_argb, argb_stride, + crop_width, inv_crop_height); + break; + case FOURCC_BGRA: + src = sample + (src_width * crop_y + crop_x) * 4; + r = BGRAToARGB(src, src_width * 4, + crop_argb, argb_stride, + crop_width, inv_crop_height); + break; + case FOURCC_ABGR: + src = sample + (src_width * crop_y + crop_x) * 4; + r = ABGRToARGB(src, src_width * 4, + crop_argb, argb_stride, + crop_width, inv_crop_height); + break; + case FOURCC_RGBA: + src = sample + (src_width * crop_y + crop_x) * 4; + r = RGBAToARGB(src, src_width * 4, + crop_argb, argb_stride, + crop_width, inv_crop_height); + break; + case FOURCC_RGBP: + src = sample + (src_width * crop_y + crop_x) * 2; + r = RGB565ToARGB(src, src_width * 2, + crop_argb, argb_stride, + crop_width, inv_crop_height); + break; + case FOURCC_RGBO: + src = sample + (src_width * crop_y + crop_x) * 2; + r = ARGB1555ToARGB(src, src_width * 2, + crop_argb, argb_stride, + crop_width, inv_crop_height); + break; + case FOURCC_R444: + src = sample + (src_width * crop_y + crop_x) * 2; + r = ARGB4444ToARGB(src, src_width * 2, + crop_argb, argb_stride, + crop_width, inv_crop_height); + break; + case FOURCC_I400: + src = sample + src_width * crop_y + crop_x; + r = I400ToARGB(src, src_width, + crop_argb, argb_stride, + crop_width, inv_crop_height); + break; + + // Biplanar formats + case FOURCC_NV12: + src = sample + (src_width * crop_y + crop_x); + src_uv = sample + aligned_src_width * (src_height + crop_y / 2) + crop_x; + r = NV12ToARGB(src, src_width, + src_uv, aligned_src_width, + crop_argb, argb_stride, + crop_width, inv_crop_height); + break; + case FOURCC_NV21: + src = sample + (src_width * crop_y + crop_x); + src_uv = sample + aligned_src_width * (src_height + crop_y / 2) + crop_x; + // Call NV12 but with u and v parameters swapped. + r = NV21ToARGB(src, src_width, + src_uv, aligned_src_width, + crop_argb, argb_stride, + crop_width, inv_crop_height); + break; + case FOURCC_M420: + src = sample + (src_width * crop_y) * 12 / 8 + crop_x; + r = M420ToARGB(src, src_width, + crop_argb, argb_stride, + crop_width, inv_crop_height); + break; + // Triplanar formats + case FOURCC_I420: + case FOURCC_YU12: + case FOURCC_YV12: { + const uint8* src_y = sample + (src_width * crop_y + crop_x); + const uint8* src_u; + const uint8* src_v; + int halfwidth = (src_width + 1) / 2; + int halfheight = (abs_src_height + 1) / 2; + if (format == FOURCC_YV12) { + src_v = sample + src_width * abs_src_height + + (halfwidth * crop_y + crop_x) / 2; + src_u = sample + src_width * abs_src_height + + halfwidth * (halfheight + crop_y / 2) + crop_x / 2; + } else { + src_u = sample + src_width * abs_src_height + + (halfwidth * crop_y + crop_x) / 2; + src_v = sample + src_width * abs_src_height + + halfwidth * (halfheight + crop_y / 2) + crop_x / 2; + } + r = I420ToARGB(src_y, src_width, + src_u, halfwidth, + src_v, halfwidth, + crop_argb, argb_stride, + crop_width, inv_crop_height); + break; + } + + case FOURCC_J420: { + const uint8* src_y = sample + (src_width * crop_y + crop_x); + const uint8* src_u; + const uint8* src_v; + int halfwidth = (src_width + 1) / 2; + int halfheight = (abs_src_height + 1) / 2; + src_u = sample + src_width * abs_src_height + + (halfwidth * crop_y + crop_x) / 2; + src_v = sample + src_width * abs_src_height + + halfwidth * (halfheight + crop_y / 2) + crop_x / 2; + r = J420ToARGB(src_y, src_width, + src_u, halfwidth, + src_v, halfwidth, + crop_argb, argb_stride, + crop_width, inv_crop_height); + break; + } + + case FOURCC_I422: + case FOURCC_YV16: { + const uint8* src_y = sample + src_width * crop_y + crop_x; + const uint8* src_u; + const uint8* src_v; + int halfwidth = (src_width + 1) / 2; + if (format == FOURCC_YV16) { + src_v = sample + src_width * abs_src_height + + halfwidth * crop_y + crop_x / 2; + src_u = sample + src_width * abs_src_height + + halfwidth * (abs_src_height + crop_y) + crop_x / 2; + } else { + src_u = sample + src_width * abs_src_height + + halfwidth * crop_y + crop_x / 2; + src_v = sample + src_width * abs_src_height + + halfwidth * (abs_src_height + crop_y) + crop_x / 2; + } + r = I422ToARGB(src_y, src_width, + src_u, halfwidth, + src_v, halfwidth, + crop_argb, argb_stride, + crop_width, inv_crop_height); + break; + } + case FOURCC_I444: + case FOURCC_YV24: { + const uint8* src_y = sample + src_width * crop_y + crop_x; + const uint8* src_u; + const uint8* src_v; + if (format == FOURCC_YV24) { + src_v = sample + src_width * (abs_src_height + crop_y) + crop_x; + src_u = sample + src_width * (abs_src_height * 2 + crop_y) + crop_x; + } else { + src_u = sample + src_width * (abs_src_height + crop_y) + crop_x; + src_v = sample + src_width * (abs_src_height * 2 + crop_y) + crop_x; + } + r = I444ToARGB(src_y, src_width, + src_u, src_width, + src_v, src_width, + crop_argb, argb_stride, + crop_width, inv_crop_height); + break; + } + case FOURCC_I411: { + int quarterwidth = (src_width + 3) / 4; + const uint8* src_y = sample + src_width * crop_y + crop_x; + const uint8* src_u = sample + src_width * abs_src_height + + quarterwidth * crop_y + crop_x / 4; + const uint8* src_v = sample + src_width * abs_src_height + + quarterwidth * (abs_src_height + crop_y) + crop_x / 4; + r = I411ToARGB(src_y, src_width, + src_u, quarterwidth, + src_v, quarterwidth, + crop_argb, argb_stride, + crop_width, inv_crop_height); + break; + } +#ifdef HAVE_JPEG + case FOURCC_MJPG: + r = MJPGToARGB(sample, sample_size, + crop_argb, argb_stride, + src_width, abs_src_height, crop_width, inv_crop_height); + break; +#endif + default: + r = -1; // unknown fourcc - return failure code. + } + + if (need_buf) { + if (!r) { + r = ARGBRotate(crop_argb, argb_stride, + tmp_argb, tmp_argb_stride, + crop_width, abs_crop_height, rotation); + } + free(rotate_buffer); + } + + return r; +} + +#ifdef __cplusplus +} // extern "C" +} // namespace libyuv +#endif diff --git a/media/libaom/src/third_party/libyuv/source/convert_to_i420.cc b/media/libaom/src/third_party/libyuv/source/convert_to_i420.cc new file mode 100644 index 000000000..5e75369b5 --- /dev/null +++ b/media/libaom/src/third_party/libyuv/source/convert_to_i420.cc @@ -0,0 +1,339 @@ +/* + * Copyright 2011 The LibYuv Project Authors. All rights reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#include <stdlib.h> + +#include "libyuv/convert.h" + +#include "libyuv/video_common.h" + +#ifdef __cplusplus +namespace libyuv { +extern "C" { +#endif + +// Convert camera sample to I420 with cropping, rotation and vertical flip. +// src_width is used for source stride computation +// src_height is used to compute location of planes, and indicate inversion +// sample_size is measured in bytes and is the size of the frame. +// With MJPEG it is the compressed size of the frame. +LIBYUV_API +int ConvertToI420(const uint8* sample, + size_t sample_size, + uint8* y, int y_stride, + uint8* u, int u_stride, + uint8* v, int v_stride, + int crop_x, int crop_y, + int src_width, int src_height, + int crop_width, int crop_height, + enum RotationMode rotation, + uint32 fourcc) { + uint32 format = CanonicalFourCC(fourcc); + int aligned_src_width = (src_width + 1) & ~1; + const uint8* src; + const uint8* src_uv; + int abs_src_height = (src_height < 0) ? -src_height : src_height; + int inv_crop_height = (crop_height < 0) ? -crop_height : crop_height; + int r = 0; + LIBYUV_BOOL need_buf = (rotation && format != FOURCC_I420 && + format != FOURCC_NV12 && format != FOURCC_NV21 && + format != FOURCC_YU12 && format != FOURCC_YV12) || y == sample; + uint8* tmp_y = y; + uint8* tmp_u = u; + uint8* tmp_v = v; + int tmp_y_stride = y_stride; + int tmp_u_stride = u_stride; + int tmp_v_stride = v_stride; + uint8* rotate_buffer = NULL; + int abs_crop_height = (crop_height < 0) ? -crop_height : crop_height; + + if (!y || !u || !v || !sample || + src_width <= 0 || crop_width <= 0 || + src_height == 0 || crop_height == 0) { + return -1; + } + if (src_height < 0) { + inv_crop_height = -inv_crop_height; + } + + // One pass rotation is available for some formats. For the rest, convert + // to I420 (with optional vertical flipping) into a temporary I420 buffer, + // and then rotate the I420 to the final destination buffer. + // For in-place conversion, if destination y is same as source sample, + // also enable temporary buffer. + if (need_buf) { + int y_size = crop_width * abs_crop_height; + int uv_size = ((crop_width + 1) / 2) * ((abs_crop_height + 1) / 2); + rotate_buffer = (uint8*)malloc(y_size + uv_size * 2); + if (!rotate_buffer) { + return 1; // Out of memory runtime error. + } + y = rotate_buffer; + u = y + y_size; + v = u + uv_size; + y_stride = crop_width; + u_stride = v_stride = ((crop_width + 1) / 2); + } + + switch (format) { + // Single plane formats + case FOURCC_YUY2: + src = sample + (aligned_src_width * crop_y + crop_x) * 2; + r = YUY2ToI420(src, aligned_src_width * 2, + y, y_stride, + u, u_stride, + v, v_stride, + crop_width, inv_crop_height); + break; + case FOURCC_UYVY: + src = sample + (aligned_src_width * crop_y + crop_x) * 2; + r = UYVYToI420(src, aligned_src_width * 2, + y, y_stride, + u, u_stride, + v, v_stride, + crop_width, inv_crop_height); + break; + case FOURCC_RGBP: + src = sample + (src_width * crop_y + crop_x) * 2; + r = RGB565ToI420(src, src_width * 2, + y, y_stride, + u, u_stride, + v, v_stride, + crop_width, inv_crop_height); + break; + case FOURCC_RGBO: + src = sample + (src_width * crop_y + crop_x) * 2; + r = ARGB1555ToI420(src, src_width * 2, + y, y_stride, + u, u_stride, + v, v_stride, + crop_width, inv_crop_height); + break; + case FOURCC_R444: + src = sample + (src_width * crop_y + crop_x) * 2; + r = ARGB4444ToI420(src, src_width * 2, + y, y_stride, + u, u_stride, + v, v_stride, + crop_width, inv_crop_height); + break; + case FOURCC_24BG: + src = sample + (src_width * crop_y + crop_x) * 3; + r = RGB24ToI420(src, src_width * 3, + y, y_stride, + u, u_stride, + v, v_stride, + crop_width, inv_crop_height); + break; + case FOURCC_RAW: + src = sample + (src_width * crop_y + crop_x) * 3; + r = RAWToI420(src, src_width * 3, + y, y_stride, + u, u_stride, + v, v_stride, + crop_width, inv_crop_height); + break; + case FOURCC_ARGB: + src = sample + (src_width * crop_y + crop_x) * 4; + r = ARGBToI420(src, src_width * 4, + y, y_stride, + u, u_stride, + v, v_stride, + crop_width, inv_crop_height); + break; + case FOURCC_BGRA: + src = sample + (src_width * crop_y + crop_x) * 4; + r = BGRAToI420(src, src_width * 4, + y, y_stride, + u, u_stride, + v, v_stride, + crop_width, inv_crop_height); + break; + case FOURCC_ABGR: + src = sample + (src_width * crop_y + crop_x) * 4; + r = ABGRToI420(src, src_width * 4, + y, y_stride, + u, u_stride, + v, v_stride, + crop_width, inv_crop_height); + break; + case FOURCC_RGBA: + src = sample + (src_width * crop_y + crop_x) * 4; + r = RGBAToI420(src, src_width * 4, + y, y_stride, + u, u_stride, + v, v_stride, + crop_width, inv_crop_height); + break; + case FOURCC_I400: + src = sample + src_width * crop_y + crop_x; + r = I400ToI420(src, src_width, + y, y_stride, + u, u_stride, + v, v_stride, + crop_width, inv_crop_height); + break; + // Biplanar formats + case FOURCC_NV12: + src = sample + (src_width * crop_y + crop_x); + src_uv = sample + (src_width * src_height) + + ((crop_y / 2) * aligned_src_width) + ((crop_x / 2) * 2); + r = NV12ToI420Rotate(src, src_width, + src_uv, aligned_src_width, + y, y_stride, + u, u_stride, + v, v_stride, + crop_width, inv_crop_height, rotation); + break; + case FOURCC_NV21: + src = sample + (src_width * crop_y + crop_x); + src_uv = sample + (src_width * src_height) + + ((crop_y / 2) * aligned_src_width) + ((crop_x / 2) * 2); + // Call NV12 but with u and v parameters swapped. + r = NV12ToI420Rotate(src, src_width, + src_uv, aligned_src_width, + y, y_stride, + v, v_stride, + u, u_stride, + crop_width, inv_crop_height, rotation); + break; + case FOURCC_M420: + src = sample + (src_width * crop_y) * 12 / 8 + crop_x; + r = M420ToI420(src, src_width, + y, y_stride, + u, u_stride, + v, v_stride, + crop_width, inv_crop_height); + break; + // Triplanar formats + case FOURCC_I420: + case FOURCC_YU12: + case FOURCC_YV12: { + const uint8* src_y = sample + (src_width * crop_y + crop_x); + const uint8* src_u; + const uint8* src_v; + int halfwidth = (src_width + 1) / 2; + int halfheight = (abs_src_height + 1) / 2; + if (format == FOURCC_YV12) { + src_v = sample + src_width * abs_src_height + + (halfwidth * crop_y + crop_x) / 2; + src_u = sample + src_width * abs_src_height + + halfwidth * (halfheight + crop_y / 2) + crop_x / 2; + } else { + src_u = sample + src_width * abs_src_height + + (halfwidth * crop_y + crop_x) / 2; + src_v = sample + src_width * abs_src_height + + halfwidth * (halfheight + crop_y / 2) + crop_x / 2; + } + r = I420Rotate(src_y, src_width, + src_u, halfwidth, + src_v, halfwidth, + y, y_stride, + u, u_stride, + v, v_stride, + crop_width, inv_crop_height, rotation); + break; + } + case FOURCC_I422: + case FOURCC_YV16: { + const uint8* src_y = sample + src_width * crop_y + crop_x; + const uint8* src_u; + const uint8* src_v; + int halfwidth = (src_width + 1) / 2; + if (format == FOURCC_YV16) { + src_v = sample + src_width * abs_src_height + + halfwidth * crop_y + crop_x / 2; + src_u = sample + src_width * abs_src_height + + halfwidth * (abs_src_height + crop_y) + crop_x / 2; + } else { + src_u = sample + src_width * abs_src_height + + halfwidth * crop_y + crop_x / 2; + src_v = sample + src_width * abs_src_height + + halfwidth * (abs_src_height + crop_y) + crop_x / 2; + } + r = I422ToI420(src_y, src_width, + src_u, halfwidth, + src_v, halfwidth, + y, y_stride, + u, u_stride, + v, v_stride, + crop_width, inv_crop_height); + break; + } + case FOURCC_I444: + case FOURCC_YV24: { + const uint8* src_y = sample + src_width * crop_y + crop_x; + const uint8* src_u; + const uint8* src_v; + if (format == FOURCC_YV24) { + src_v = sample + src_width * (abs_src_height + crop_y) + crop_x; + src_u = sample + src_width * (abs_src_height * 2 + crop_y) + crop_x; + } else { + src_u = sample + src_width * (abs_src_height + crop_y) + crop_x; + src_v = sample + src_width * (abs_src_height * 2 + crop_y) + crop_x; + } + r = I444ToI420(src_y, src_width, + src_u, src_width, + src_v, src_width, + y, y_stride, + u, u_stride, + v, v_stride, + crop_width, inv_crop_height); + break; + } + case FOURCC_I411: { + int quarterwidth = (src_width + 3) / 4; + const uint8* src_y = sample + src_width * crop_y + crop_x; + const uint8* src_u = sample + src_width * abs_src_height + + quarterwidth * crop_y + crop_x / 4; + const uint8* src_v = sample + src_width * abs_src_height + + quarterwidth * (abs_src_height + crop_y) + crop_x / 4; + r = I411ToI420(src_y, src_width, + src_u, quarterwidth, + src_v, quarterwidth, + y, y_stride, + u, u_stride, + v, v_stride, + crop_width, inv_crop_height); + break; + } +#ifdef HAVE_JPEG + case FOURCC_MJPG: + r = MJPGToI420(sample, sample_size, + y, y_stride, + u, u_stride, + v, v_stride, + src_width, abs_src_height, crop_width, inv_crop_height); + break; +#endif + default: + r = -1; // unknown fourcc - return failure code. + } + + if (need_buf) { + if (!r) { + r = I420Rotate(y, y_stride, + u, u_stride, + v, v_stride, + tmp_y, tmp_y_stride, + tmp_u, tmp_u_stride, + tmp_v, tmp_v_stride, + crop_width, abs_crop_height, rotation); + } + free(rotate_buffer); + } + + return r; +} + +#ifdef __cplusplus +} // extern "C" +} // namespace libyuv +#endif diff --git a/media/libaom/src/third_party/libyuv/source/cpu_id.cc b/media/libaom/src/third_party/libyuv/source/cpu_id.cc new file mode 100644 index 000000000..72f686e3b --- /dev/null +++ b/media/libaom/src/third_party/libyuv/source/cpu_id.cc @@ -0,0 +1,307 @@ +/* + * Copyright 2011 The LibYuv Project Authors. All rights reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#include "libyuv/cpu_id.h" + +#if (defined(_MSC_VER) && !defined(__clang__)) && !defined(__clang__) +#include <intrin.h> // For __cpuidex() +#endif +#if !defined(__pnacl__) && !defined(__CLR_VER) && \ + !defined(__native_client__) && (defined(_M_IX86) || defined(_M_X64)) && \ + defined(_MSC_VER) && !defined(__clang__) && (_MSC_FULL_VER >= 160040219) +#include <immintrin.h> // For _xgetbv() +#endif + +#if !defined(__native_client__) +#include <stdlib.h> // For getenv() +#endif + +// For ArmCpuCaps() but unittested on all platforms +#include <stdio.h> +#include <string.h> + +#include "libyuv/basic_types.h" // For CPU_X86 + +#ifdef __cplusplus +namespace libyuv { +extern "C" { +#endif + +// For functions that use the stack and have runtime checks for overflow, +// use SAFEBUFFERS to avoid additional check. +#if (defined(_MSC_VER) && !defined(__clang__)) && (_MSC_FULL_VER >= 160040219) +#define SAFEBUFFERS __declspec(safebuffers) +#else +#define SAFEBUFFERS +#endif + +// Low level cpuid for X86. +#if (defined(_M_IX86) || defined(_M_X64) || \ + defined(__i386__) || defined(__x86_64__)) && \ + !defined(__pnacl__) && !defined(__CLR_VER) +LIBYUV_API +void CpuId(uint32 info_eax, uint32 info_ecx, uint32* cpu_info) { +#if (defined(_MSC_VER) && !defined(__clang__)) && !defined(__clang__) +// Visual C version uses intrinsic or inline x86 assembly. +#if (_MSC_FULL_VER >= 160040219) + __cpuidex((int*)(cpu_info), info_eax, info_ecx); +#elif defined(_M_IX86) + __asm { + mov eax, info_eax + mov ecx, info_ecx + mov edi, cpu_info + cpuid + mov [edi], eax + mov [edi + 4], ebx + mov [edi + 8], ecx + mov [edi + 12], edx + } +#else + if (info_ecx == 0) { + __cpuid((int*)(cpu_info), info_eax); + } else { + cpu_info[3] = cpu_info[2] = cpu_info[1] = cpu_info[0] = 0; + } +#endif +// GCC version uses inline x86 assembly. +#else // (defined(_MSC_VER) && !defined(__clang__)) && !defined(__clang__) + uint32 info_ebx, info_edx; + asm volatile ( // NOLINT +#if defined( __i386__) && defined(__PIC__) + // Preserve ebx for fpic 32 bit. + "mov %%ebx, %%edi \n" + "cpuid \n" + "xchg %%edi, %%ebx \n" + : "=D" (info_ebx), +#else + "cpuid \n" + : "=b" (info_ebx), +#endif // defined( __i386__) && defined(__PIC__) + "+a" (info_eax), "+c" (info_ecx), "=d" (info_edx)); + cpu_info[0] = info_eax; + cpu_info[1] = info_ebx; + cpu_info[2] = info_ecx; + cpu_info[3] = info_edx; +#endif // (defined(_MSC_VER) && !defined(__clang__)) && !defined(__clang__) +} +#else // (defined(_M_IX86) || defined(_M_X64) ... +LIBYUV_API +void CpuId(uint32 eax, uint32 ecx, uint32* cpu_info) { + cpu_info[0] = cpu_info[1] = cpu_info[2] = cpu_info[3] = 0; +} +#endif + +// TODO(fbarchard): Enable xgetbv when validator supports it. +#if (defined(_M_IX86) || defined(_M_X64) || \ + defined(__i386__) || defined(__x86_64__)) && \ + !defined(__pnacl__) && !defined(__CLR_VER) && !defined(__native_client__) +#define HAS_XGETBV +// X86 CPUs have xgetbv to detect OS saves high parts of ymm registers. +int TestOsSaveYmm() { + uint32 xcr0 = 0u; +#if (defined(_MSC_VER) && !defined(__clang__)) && (_MSC_FULL_VER >= 160040219) + xcr0 = (uint32)(_xgetbv(0)); // VS2010 SP1 required. +#elif defined(_M_IX86) && defined(_MSC_VER) && !defined(__clang__) + __asm { + xor ecx, ecx // xcr 0 + _asm _emit 0x0f _asm _emit 0x01 _asm _emit 0xd0 // For VS2010 and earlier. + mov xcr0, eax + } +#elif defined(__i386__) || defined(__x86_64__) + asm(".byte 0x0f, 0x01, 0xd0" : "=a" (xcr0) : "c" (0) : "%edx"); +#endif // defined(__i386__) || defined(__x86_64__) + return((xcr0 & 6) == 6); // Is ymm saved? +} +#endif // defined(_M_IX86) || defined(_M_X64) .. + +// based on libaom arm_cpudetect.c +// For Arm, but public to allow testing on any CPU +LIBYUV_API SAFEBUFFERS +int ArmCpuCaps(const char* cpuinfo_name) { + char cpuinfo_line[512]; + FILE* f = fopen(cpuinfo_name, "r"); + if (!f) { + // Assume Neon if /proc/cpuinfo is unavailable. + // This will occur for Chrome sandbox for Pepper or Render process. + return kCpuHasNEON; + } + while (fgets(cpuinfo_line, sizeof(cpuinfo_line) - 1, f)) { + if (memcmp(cpuinfo_line, "Features", 8) == 0) { + char* p = strstr(cpuinfo_line, " neon"); + if (p && (p[5] == ' ' || p[5] == '\n')) { + fclose(f); + return kCpuHasNEON; + } + // aarch64 uses asimd for Neon. + p = strstr(cpuinfo_line, " asimd"); + if (p && (p[6] == ' ' || p[6] == '\n')) { + fclose(f); + return kCpuHasNEON; + } + } + } + fclose(f); + return 0; +} + +#if defined(__mips__) && defined(__linux__) +static int MipsCpuCaps(const char* search_string) { + char cpuinfo_line[512]; + const char* file_name = "/proc/cpuinfo"; + FILE* f = fopen(file_name, "r"); + if (!f) { + // Assume DSP if /proc/cpuinfo is unavailable. + // This will occur for Chrome sandbox for Pepper or Render process. + return kCpuHasMIPS_DSP; + } + while (fgets(cpuinfo_line, sizeof(cpuinfo_line) - 1, f) != NULL) { + if (strstr(cpuinfo_line, search_string) != NULL) { + fclose(f); + return kCpuHasMIPS_DSP; + } + } + fclose(f); + return 0; +} +#endif + +// CPU detect function for SIMD instruction sets. +LIBYUV_API +int cpu_info_ = kCpuInit; // cpu_info is not initialized yet. + +// Test environment variable for disabling CPU features. Any non-zero value +// to disable. Zero ignored to make it easy to set the variable on/off. +#if !defined(__native_client__) && !defined(_M_ARM) + +static LIBYUV_BOOL TestEnv(const char* name) { + const char* var = getenv(name); + if (var) { + if (var[0] != '0') { + return LIBYUV_TRUE; + } + } + return LIBYUV_FALSE; +} +#else // nacl does not support getenv(). +static LIBYUV_BOOL TestEnv(const char*) { + return LIBYUV_FALSE; +} +#endif + +LIBYUV_API SAFEBUFFERS +int InitCpuFlags(void) { +#if !defined(__pnacl__) && !defined(__CLR_VER) && defined(CPU_X86) + + uint32 cpu_info0[4] = { 0, 0, 0, 0 }; + uint32 cpu_info1[4] = { 0, 0, 0, 0 }; + uint32 cpu_info7[4] = { 0, 0, 0, 0 }; + CpuId(0, 0, cpu_info0); + CpuId(1, 0, cpu_info1); + if (cpu_info0[0] >= 7) { + CpuId(7, 0, cpu_info7); + } + cpu_info_ = ((cpu_info1[3] & 0x04000000) ? kCpuHasSSE2 : 0) | + ((cpu_info1[2] & 0x00000200) ? kCpuHasSSSE3 : 0) | + ((cpu_info1[2] & 0x00080000) ? kCpuHasSSE41 : 0) | + ((cpu_info1[2] & 0x00100000) ? kCpuHasSSE42 : 0) | + ((cpu_info7[1] & 0x00000200) ? kCpuHasERMS : 0) | + ((cpu_info1[2] & 0x00001000) ? kCpuHasFMA3 : 0) | + kCpuHasX86; + +#ifdef HAS_XGETBV + if ((cpu_info1[2] & 0x18000000) == 0x18000000 && // AVX and OSSave + TestOsSaveYmm()) { // Saves YMM. + cpu_info_ |= ((cpu_info7[1] & 0x00000020) ? kCpuHasAVX2 : 0) | + kCpuHasAVX; + } +#endif + // Environment variable overrides for testing. + if (TestEnv("LIBYUV_DISABLE_X86")) { + cpu_info_ &= ~kCpuHasX86; + } + if (TestEnv("LIBYUV_DISABLE_SSE2")) { + cpu_info_ &= ~kCpuHasSSE2; + } + if (TestEnv("LIBYUV_DISABLE_SSSE3")) { + cpu_info_ &= ~kCpuHasSSSE3; + } + if (TestEnv("LIBYUV_DISABLE_SSE41")) { + cpu_info_ &= ~kCpuHasSSE41; + } + if (TestEnv("LIBYUV_DISABLE_SSE42")) { + cpu_info_ &= ~kCpuHasSSE42; + } + if (TestEnv("LIBYUV_DISABLE_AVX")) { + cpu_info_ &= ~kCpuHasAVX; + } + if (TestEnv("LIBYUV_DISABLE_AVX2")) { + cpu_info_ &= ~kCpuHasAVX2; + } + if (TestEnv("LIBYUV_DISABLE_ERMS")) { + cpu_info_ &= ~kCpuHasERMS; + } + if (TestEnv("LIBYUV_DISABLE_FMA3")) { + cpu_info_ &= ~kCpuHasFMA3; + } +#endif +#if defined(__mips__) && defined(__linux__) + // Linux mips parse text file for dsp detect. + cpu_info_ = MipsCpuCaps("dsp"); // set kCpuHasMIPS_DSP. +#if defined(__mips_dspr2) + cpu_info_ |= kCpuHasMIPS_DSPR2; +#endif + cpu_info_ |= kCpuHasMIPS; + + if (getenv("LIBYUV_DISABLE_MIPS")) { + cpu_info_ &= ~kCpuHasMIPS; + } + if (getenv("LIBYUV_DISABLE_MIPS_DSP")) { + cpu_info_ &= ~kCpuHasMIPS_DSP; + } + if (getenv("LIBYUV_DISABLE_MIPS_DSPR2")) { + cpu_info_ &= ~kCpuHasMIPS_DSPR2; + } +#endif +#if defined(__arm__) || defined(__aarch64__) +// gcc -mfpu=neon defines __ARM_NEON__ +// __ARM_NEON__ generates code that requires Neon. NaCL also requires Neon. +// For Linux, /proc/cpuinfo can be tested but without that assume Neon. +#if defined(__ARM_NEON__) || defined(__native_client__) || !defined(__linux__) + cpu_info_ = kCpuHasNEON; +// For aarch64(arm64), /proc/cpuinfo's feature is not complete, e.g. no neon +// flag in it. +// So for aarch64, neon enabling is hard coded here. +#endif +#if defined(__aarch64__) + cpu_info_ = kCpuHasNEON; +#else + // Linux arm parse text file for neon detect. + cpu_info_ = ArmCpuCaps("/proc/cpuinfo"); +#endif + cpu_info_ |= kCpuHasARM; + if (TestEnv("LIBYUV_DISABLE_NEON")) { + cpu_info_ &= ~kCpuHasNEON; + } +#endif // __arm__ + if (TestEnv("LIBYUV_DISABLE_ASM")) { + cpu_info_ = 0; + } + return cpu_info_; +} + +LIBYUV_API +void MaskCpuFlags(int enable_flags) { + cpu_info_ = InitCpuFlags() & enable_flags; +} + +#ifdef __cplusplus +} // extern "C" +} // namespace libyuv +#endif diff --git a/media/libaom/src/third_party/libyuv/source/mjpeg_decoder.cc b/media/libaom/src/third_party/libyuv/source/mjpeg_decoder.cc new file mode 100644 index 000000000..75f8a610e --- /dev/null +++ b/media/libaom/src/third_party/libyuv/source/mjpeg_decoder.cc @@ -0,0 +1,572 @@ +/* + * Copyright 2012 The LibYuv Project Authors. All rights reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#include "libyuv/mjpeg_decoder.h" + +#ifdef HAVE_JPEG +#include <assert.h> + +#if !defined(__pnacl__) && !defined(__CLR_VER) && \ + !defined(COVERAGE_ENABLED) && !defined(TARGET_IPHONE_SIMULATOR) +// Must be included before jpeglib. +#include <setjmp.h> +#define HAVE_SETJMP + +#if defined(_MSC_VER) +// disable warning 4324: structure was padded due to __declspec(align()) +#pragma warning(disable:4324) +#endif + +#endif +struct FILE; // For jpeglib.h. + +// C++ build requires extern C for jpeg internals. +#ifdef __cplusplus +extern "C" { +#endif + +#include <jpeglib.h> + +#ifdef __cplusplus +} // extern "C" +#endif + +#include "libyuv/planar_functions.h" // For CopyPlane(). + +namespace libyuv { + +#ifdef HAVE_SETJMP +struct SetJmpErrorMgr { + jpeg_error_mgr base; // Must be at the top + jmp_buf setjmp_buffer; +}; +#endif + +const int MJpegDecoder::kColorSpaceUnknown = JCS_UNKNOWN; +const int MJpegDecoder::kColorSpaceGrayscale = JCS_GRAYSCALE; +const int MJpegDecoder::kColorSpaceRgb = JCS_RGB; +const int MJpegDecoder::kColorSpaceYCbCr = JCS_YCbCr; +const int MJpegDecoder::kColorSpaceCMYK = JCS_CMYK; +const int MJpegDecoder::kColorSpaceYCCK = JCS_YCCK; + +// Methods that are passed to jpeglib. +boolean fill_input_buffer(jpeg_decompress_struct* cinfo); +void init_source(jpeg_decompress_struct* cinfo); +void skip_input_data(jpeg_decompress_struct* cinfo, + long num_bytes); // NOLINT +void term_source(jpeg_decompress_struct* cinfo); +void ErrorHandler(jpeg_common_struct* cinfo); + +MJpegDecoder::MJpegDecoder() + : has_scanline_padding_(LIBYUV_FALSE), + num_outbufs_(0), + scanlines_(NULL), + scanlines_sizes_(NULL), + databuf_(NULL), + databuf_strides_(NULL) { + decompress_struct_ = new jpeg_decompress_struct; + source_mgr_ = new jpeg_source_mgr; +#ifdef HAVE_SETJMP + error_mgr_ = new SetJmpErrorMgr; + decompress_struct_->err = jpeg_std_error(&error_mgr_->base); + // Override standard exit()-based error handler. + error_mgr_->base.error_exit = &ErrorHandler; +#endif + decompress_struct_->client_data = NULL; + source_mgr_->init_source = &init_source; + source_mgr_->fill_input_buffer = &fill_input_buffer; + source_mgr_->skip_input_data = &skip_input_data; + source_mgr_->resync_to_restart = &jpeg_resync_to_restart; + source_mgr_->term_source = &term_source; + jpeg_create_decompress(decompress_struct_); + decompress_struct_->src = source_mgr_; + buf_vec_.buffers = &buf_; + buf_vec_.len = 1; +} + +MJpegDecoder::~MJpegDecoder() { + jpeg_destroy_decompress(decompress_struct_); + delete decompress_struct_; + delete source_mgr_; +#ifdef HAVE_SETJMP + delete error_mgr_; +#endif + DestroyOutputBuffers(); +} + +LIBYUV_BOOL MJpegDecoder::LoadFrame(const uint8* src, size_t src_len) { + if (!ValidateJpeg(src, src_len)) { + return LIBYUV_FALSE; + } + + buf_.data = src; + buf_.len = static_cast<int>(src_len); + buf_vec_.pos = 0; + decompress_struct_->client_data = &buf_vec_; +#ifdef HAVE_SETJMP + if (setjmp(error_mgr_->setjmp_buffer)) { + // We called jpeg_read_header, it experienced an error, and we called + // longjmp() and rewound the stack to here. Return error. + return LIBYUV_FALSE; + } +#endif + if (jpeg_read_header(decompress_struct_, TRUE) != JPEG_HEADER_OK) { + // ERROR: Bad MJPEG header + return LIBYUV_FALSE; + } + AllocOutputBuffers(GetNumComponents()); + for (int i = 0; i < num_outbufs_; ++i) { + int scanlines_size = GetComponentScanlinesPerImcuRow(i); + if (scanlines_sizes_[i] != scanlines_size) { + if (scanlines_[i]) { + delete scanlines_[i]; + } + scanlines_[i] = new uint8* [scanlines_size]; + scanlines_sizes_[i] = scanlines_size; + } + + // We allocate padding for the final scanline to pad it up to DCTSIZE bytes + // to avoid memory errors, since jpeglib only reads full MCUs blocks. For + // the preceding scanlines, the padding is not needed/wanted because the + // following addresses will already be valid (they are the initial bytes of + // the next scanline) and will be overwritten when jpeglib writes out that + // next scanline. + int databuf_stride = GetComponentStride(i); + int databuf_size = scanlines_size * databuf_stride; + if (databuf_strides_[i] != databuf_stride) { + if (databuf_[i]) { + delete databuf_[i]; + } + databuf_[i] = new uint8[databuf_size]; + databuf_strides_[i] = databuf_stride; + } + + if (GetComponentStride(i) != GetComponentWidth(i)) { + has_scanline_padding_ = LIBYUV_TRUE; + } + } + return LIBYUV_TRUE; +} + +static int DivideAndRoundUp(int numerator, int denominator) { + return (numerator + denominator - 1) / denominator; +} + +static int DivideAndRoundDown(int numerator, int denominator) { + return numerator / denominator; +} + +// Returns width of the last loaded frame. +int MJpegDecoder::GetWidth() { + return decompress_struct_->image_width; +} + +// Returns height of the last loaded frame. +int MJpegDecoder::GetHeight() { + return decompress_struct_->image_height; +} + +// Returns format of the last loaded frame. The return value is one of the +// kColorSpace* constants. +int MJpegDecoder::GetColorSpace() { + return decompress_struct_->jpeg_color_space; +} + +// Number of color components in the color space. +int MJpegDecoder::GetNumComponents() { + return decompress_struct_->num_components; +} + +// Sample factors of the n-th component. +int MJpegDecoder::GetHorizSampFactor(int component) { + return decompress_struct_->comp_info[component].h_samp_factor; +} + +int MJpegDecoder::GetVertSampFactor(int component) { + return decompress_struct_->comp_info[component].v_samp_factor; +} + +int MJpegDecoder::GetHorizSubSampFactor(int component) { + return decompress_struct_->max_h_samp_factor / + GetHorizSampFactor(component); +} + +int MJpegDecoder::GetVertSubSampFactor(int component) { + return decompress_struct_->max_v_samp_factor / + GetVertSampFactor(component); +} + +int MJpegDecoder::GetImageScanlinesPerImcuRow() { + return decompress_struct_->max_v_samp_factor * DCTSIZE; +} + +int MJpegDecoder::GetComponentScanlinesPerImcuRow(int component) { + int vs = GetVertSubSampFactor(component); + return DivideAndRoundUp(GetImageScanlinesPerImcuRow(), vs); +} + +int MJpegDecoder::GetComponentWidth(int component) { + int hs = GetHorizSubSampFactor(component); + return DivideAndRoundUp(GetWidth(), hs); +} + +int MJpegDecoder::GetComponentHeight(int component) { + int vs = GetVertSubSampFactor(component); + return DivideAndRoundUp(GetHeight(), vs); +} + +// Get width in bytes padded out to a multiple of DCTSIZE +int MJpegDecoder::GetComponentStride(int component) { + return (GetComponentWidth(component) + DCTSIZE - 1) & ~(DCTSIZE - 1); +} + +int MJpegDecoder::GetComponentSize(int component) { + return GetComponentWidth(component) * GetComponentHeight(component); +} + +LIBYUV_BOOL MJpegDecoder::UnloadFrame() { +#ifdef HAVE_SETJMP + if (setjmp(error_mgr_->setjmp_buffer)) { + // We called jpeg_abort_decompress, it experienced an error, and we called + // longjmp() and rewound the stack to here. Return error. + return LIBYUV_FALSE; + } +#endif + jpeg_abort_decompress(decompress_struct_); + return LIBYUV_TRUE; +} + +// TODO(fbarchard): Allow rectangle to be specified: x, y, width, height. +LIBYUV_BOOL MJpegDecoder::DecodeToBuffers( + uint8** planes, int dst_width, int dst_height) { + if (dst_width != GetWidth() || + dst_height > GetHeight()) { + // ERROR: Bad dimensions + return LIBYUV_FALSE; + } +#ifdef HAVE_SETJMP + if (setjmp(error_mgr_->setjmp_buffer)) { + // We called into jpeglib, it experienced an error sometime during this + // function call, and we called longjmp() and rewound the stack to here. + // Return error. + return LIBYUV_FALSE; + } +#endif + if (!StartDecode()) { + return LIBYUV_FALSE; + } + SetScanlinePointers(databuf_); + int lines_left = dst_height; + // Compute amount of lines to skip to implement vertical crop. + // TODO(fbarchard): Ensure skip is a multiple of maximum component + // subsample. ie 2 + int skip = (GetHeight() - dst_height) / 2; + if (skip > 0) { + // There is no API to skip lines in the output data, so we read them + // into the temp buffer. + while (skip >= GetImageScanlinesPerImcuRow()) { + if (!DecodeImcuRow()) { + FinishDecode(); + return LIBYUV_FALSE; + } + skip -= GetImageScanlinesPerImcuRow(); + } + if (skip > 0) { + // Have a partial iMCU row left over to skip. Must read it and then + // copy the parts we want into the destination. + if (!DecodeImcuRow()) { + FinishDecode(); + return LIBYUV_FALSE; + } + for (int i = 0; i < num_outbufs_; ++i) { + // TODO(fbarchard): Compute skip to avoid this + assert(skip % GetVertSubSampFactor(i) == 0); + int rows_to_skip = + DivideAndRoundDown(skip, GetVertSubSampFactor(i)); + int scanlines_to_copy = GetComponentScanlinesPerImcuRow(i) - + rows_to_skip; + int data_to_skip = rows_to_skip * GetComponentStride(i); + CopyPlane(databuf_[i] + data_to_skip, GetComponentStride(i), + planes[i], GetComponentWidth(i), + GetComponentWidth(i), scanlines_to_copy); + planes[i] += scanlines_to_copy * GetComponentWidth(i); + } + lines_left -= (GetImageScanlinesPerImcuRow() - skip); + } + } + + // Read full MCUs but cropped horizontally + for (; lines_left > GetImageScanlinesPerImcuRow(); + lines_left -= GetImageScanlinesPerImcuRow()) { + if (!DecodeImcuRow()) { + FinishDecode(); + return LIBYUV_FALSE; + } + for (int i = 0; i < num_outbufs_; ++i) { + int scanlines_to_copy = GetComponentScanlinesPerImcuRow(i); + CopyPlane(databuf_[i], GetComponentStride(i), + planes[i], GetComponentWidth(i), + GetComponentWidth(i), scanlines_to_copy); + planes[i] += scanlines_to_copy * GetComponentWidth(i); + } + } + + if (lines_left > 0) { + // Have a partial iMCU row left over to decode. + if (!DecodeImcuRow()) { + FinishDecode(); + return LIBYUV_FALSE; + } + for (int i = 0; i < num_outbufs_; ++i) { + int scanlines_to_copy = + DivideAndRoundUp(lines_left, GetVertSubSampFactor(i)); + CopyPlane(databuf_[i], GetComponentStride(i), + planes[i], GetComponentWidth(i), + GetComponentWidth(i), scanlines_to_copy); + planes[i] += scanlines_to_copy * GetComponentWidth(i); + } + } + return FinishDecode(); +} + +LIBYUV_BOOL MJpegDecoder::DecodeToCallback(CallbackFunction fn, void* opaque, + int dst_width, int dst_height) { + if (dst_width != GetWidth() || + dst_height > GetHeight()) { + // ERROR: Bad dimensions + return LIBYUV_FALSE; + } +#ifdef HAVE_SETJMP + if (setjmp(error_mgr_->setjmp_buffer)) { + // We called into jpeglib, it experienced an error sometime during this + // function call, and we called longjmp() and rewound the stack to here. + // Return error. + return LIBYUV_FALSE; + } +#endif + if (!StartDecode()) { + return LIBYUV_FALSE; + } + SetScanlinePointers(databuf_); + int lines_left = dst_height; + // TODO(fbarchard): Compute amount of lines to skip to implement vertical crop + int skip = (GetHeight() - dst_height) / 2; + if (skip > 0) { + while (skip >= GetImageScanlinesPerImcuRow()) { + if (!DecodeImcuRow()) { + FinishDecode(); + return LIBYUV_FALSE; + } + skip -= GetImageScanlinesPerImcuRow(); + } + if (skip > 0) { + // Have a partial iMCU row left over to skip. + if (!DecodeImcuRow()) { + FinishDecode(); + return LIBYUV_FALSE; + } + for (int i = 0; i < num_outbufs_; ++i) { + // TODO(fbarchard): Compute skip to avoid this + assert(skip % GetVertSubSampFactor(i) == 0); + int rows_to_skip = DivideAndRoundDown(skip, GetVertSubSampFactor(i)); + int data_to_skip = rows_to_skip * GetComponentStride(i); + // Change our own data buffer pointers so we can pass them to the + // callback. + databuf_[i] += data_to_skip; + } + int scanlines_to_copy = GetImageScanlinesPerImcuRow() - skip; + (*fn)(opaque, databuf_, databuf_strides_, scanlines_to_copy); + // Now change them back. + for (int i = 0; i < num_outbufs_; ++i) { + int rows_to_skip = DivideAndRoundDown(skip, GetVertSubSampFactor(i)); + int data_to_skip = rows_to_skip * GetComponentStride(i); + databuf_[i] -= data_to_skip; + } + lines_left -= scanlines_to_copy; + } + } + // Read full MCUs until we get to the crop point. + for (; lines_left >= GetImageScanlinesPerImcuRow(); + lines_left -= GetImageScanlinesPerImcuRow()) { + if (!DecodeImcuRow()) { + FinishDecode(); + return LIBYUV_FALSE; + } + (*fn)(opaque, databuf_, databuf_strides_, GetImageScanlinesPerImcuRow()); + } + if (lines_left > 0) { + // Have a partial iMCU row left over to decode. + if (!DecodeImcuRow()) { + FinishDecode(); + return LIBYUV_FALSE; + } + (*fn)(opaque, databuf_, databuf_strides_, lines_left); + } + return FinishDecode(); +} + +void init_source(j_decompress_ptr cinfo) { + fill_input_buffer(cinfo); +} + +boolean fill_input_buffer(j_decompress_ptr cinfo) { + BufferVector* buf_vec = reinterpret_cast<BufferVector*>(cinfo->client_data); + if (buf_vec->pos >= buf_vec->len) { + assert(0 && "No more data"); + // ERROR: No more data + return FALSE; + } + cinfo->src->next_input_byte = buf_vec->buffers[buf_vec->pos].data; + cinfo->src->bytes_in_buffer = buf_vec->buffers[buf_vec->pos].len; + ++buf_vec->pos; + return TRUE; +} + +void skip_input_data(j_decompress_ptr cinfo, + long num_bytes) { // NOLINT + cinfo->src->next_input_byte += num_bytes; +} + +void term_source(j_decompress_ptr cinfo) { + // Nothing to do. +} + +#ifdef HAVE_SETJMP +void ErrorHandler(j_common_ptr cinfo) { + // This is called when a jpeglib command experiences an error. Unfortunately + // jpeglib's error handling model is not very flexible, because it expects the + // error handler to not return--i.e., it wants the program to terminate. To + // recover from errors we use setjmp() as shown in their example. setjmp() is + // C's implementation for the "call with current continuation" functionality + // seen in some functional programming languages. + // A formatted message can be output, but is unsafe for release. +#ifdef DEBUG + char buf[JMSG_LENGTH_MAX]; + (*cinfo->err->format_message)(cinfo, buf); + // ERROR: Error in jpeglib: buf +#endif + + SetJmpErrorMgr* mgr = reinterpret_cast<SetJmpErrorMgr*>(cinfo->err); + // This rewinds the call stack to the point of the corresponding setjmp() + // and causes it to return (for a second time) with value 1. + longjmp(mgr->setjmp_buffer, 1); +} +#endif + +void MJpegDecoder::AllocOutputBuffers(int num_outbufs) { + if (num_outbufs != num_outbufs_) { + // We could perhaps optimize this case to resize the output buffers without + // necessarily having to delete and recreate each one, but it's not worth + // it. + DestroyOutputBuffers(); + + scanlines_ = new uint8** [num_outbufs]; + scanlines_sizes_ = new int[num_outbufs]; + databuf_ = new uint8* [num_outbufs]; + databuf_strides_ = new int[num_outbufs]; + + for (int i = 0; i < num_outbufs; ++i) { + scanlines_[i] = NULL; + scanlines_sizes_[i] = 0; + databuf_[i] = NULL; + databuf_strides_[i] = 0; + } + + num_outbufs_ = num_outbufs; + } +} + +void MJpegDecoder::DestroyOutputBuffers() { + for (int i = 0; i < num_outbufs_; ++i) { + delete [] scanlines_[i]; + delete [] databuf_[i]; + } + delete [] scanlines_; + delete [] databuf_; + delete [] scanlines_sizes_; + delete [] databuf_strides_; + scanlines_ = NULL; + databuf_ = NULL; + scanlines_sizes_ = NULL; + databuf_strides_ = NULL; + num_outbufs_ = 0; +} + +// JDCT_IFAST and do_block_smoothing improve performance substantially. +LIBYUV_BOOL MJpegDecoder::StartDecode() { + decompress_struct_->raw_data_out = TRUE; + decompress_struct_->dct_method = JDCT_IFAST; // JDCT_ISLOW is default + decompress_struct_->dither_mode = JDITHER_NONE; + // Not applicable to 'raw': + decompress_struct_->do_fancy_upsampling = (boolean)(LIBYUV_FALSE); + // Only for buffered mode: + decompress_struct_->enable_2pass_quant = (boolean)(LIBYUV_FALSE); + // Blocky but fast: + decompress_struct_->do_block_smoothing = (boolean)(LIBYUV_FALSE); + + if (!jpeg_start_decompress(decompress_struct_)) { + // ERROR: Couldn't start JPEG decompressor"; + return LIBYUV_FALSE; + } + return LIBYUV_TRUE; +} + +LIBYUV_BOOL MJpegDecoder::FinishDecode() { + // jpeglib considers it an error if we finish without decoding the whole + // image, so we call "abort" rather than "finish". + jpeg_abort_decompress(decompress_struct_); + return LIBYUV_TRUE; +} + +void MJpegDecoder::SetScanlinePointers(uint8** data) { + for (int i = 0; i < num_outbufs_; ++i) { + uint8* data_i = data[i]; + for (int j = 0; j < scanlines_sizes_[i]; ++j) { + scanlines_[i][j] = data_i; + data_i += GetComponentStride(i); + } + } +} + +inline LIBYUV_BOOL MJpegDecoder::DecodeImcuRow() { + return (unsigned int)(GetImageScanlinesPerImcuRow()) == + jpeg_read_raw_data(decompress_struct_, + scanlines_, + GetImageScanlinesPerImcuRow()); +} + +// The helper function which recognizes the jpeg sub-sampling type. +JpegSubsamplingType MJpegDecoder::JpegSubsamplingTypeHelper( + int* subsample_x, int* subsample_y, int number_of_components) { + if (number_of_components == 3) { // Color images. + if (subsample_x[0] == 1 && subsample_y[0] == 1 && + subsample_x[1] == 2 && subsample_y[1] == 2 && + subsample_x[2] == 2 && subsample_y[2] == 2) { + return kJpegYuv420; + } else if (subsample_x[0] == 1 && subsample_y[0] == 1 && + subsample_x[1] == 2 && subsample_y[1] == 1 && + subsample_x[2] == 2 && subsample_y[2] == 1) { + return kJpegYuv422; + } else if (subsample_x[0] == 1 && subsample_y[0] == 1 && + subsample_x[1] == 1 && subsample_y[1] == 1 && + subsample_x[2] == 1 && subsample_y[2] == 1) { + return kJpegYuv444; + } + } else if (number_of_components == 1) { // Grey-scale images. + if (subsample_x[0] == 1 && subsample_y[0] == 1) { + return kJpegYuv400; + } + } + return kJpegUnknown; +} + +} // namespace libyuv +#endif // HAVE_JPEG + diff --git a/media/libaom/src/third_party/libyuv/source/mjpeg_validate.cc b/media/libaom/src/third_party/libyuv/source/mjpeg_validate.cc new file mode 100644 index 000000000..8edfbe1e7 --- /dev/null +++ b/media/libaom/src/third_party/libyuv/source/mjpeg_validate.cc @@ -0,0 +1,101 @@ +/* + * Copyright 2012 The LibYuv Project Authors. All rights reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#include "libyuv/mjpeg_decoder.h" + +#include <string.h> // For memchr. + +#ifdef __cplusplus +namespace libyuv { +extern "C" { +#endif + +// Enable this to try scasb implementation. +// #define ENABLE_SCASB 1 + +#ifdef ENABLE_SCASB + +// Multiple of 1. +__declspec(naked) +const uint8* ScanRow_ERMS(const uint8* src, uint32 val, int count) { + __asm { + mov edx, edi + mov edi, [esp + 4] // src + mov eax, [esp + 8] // val + mov ecx, [esp + 12] // count + repne scasb + jne sr99 + mov eax, edi + sub eax, 1 + mov edi, edx + ret + + sr99: + mov eax, 0 + mov edi, edx + ret + } +} +#endif + +// Helper function to scan for EOI marker. +static LIBYUV_BOOL ScanEOI(const uint8* sample, size_t sample_size) { + const uint8* end = sample + sample_size - 1; + const uint8* it = sample; + for (;;) { +#ifdef ENABLE_SCASB + it = ScanRow_ERMS(it, 0xff, end - it); +#else + it = static_cast<const uint8*>(memchr(it, 0xff, end - it)); +#endif + if (it == NULL) { + break; + } + if (it[1] == 0xd9) { + return LIBYUV_TRUE; // Success: Valid jpeg. + } + ++it; // Skip over current 0xff. + } + // ERROR: Invalid jpeg end code not found. Size sample_size + return LIBYUV_FALSE; +} + +// Helper function to validate the jpeg appears intact. +LIBYUV_BOOL ValidateJpeg(const uint8* sample, size_t sample_size) { + const size_t kBackSearchSize = 1024; + if (sample_size < 64) { + // ERROR: Invalid jpeg size: sample_size + return LIBYUV_FALSE; + } + if (sample[0] != 0xff || sample[1] != 0xd8) { // Start Of Image + // ERROR: Invalid jpeg initial start code + return LIBYUV_FALSE; + } + // Step over SOI marker. + sample += 2; + sample_size -= 2; + + // Look for the End Of Image (EOI) marker in the end kilobyte of the buffer. + if (sample_size > kBackSearchSize) { + if (ScanEOI(sample + sample_size - kBackSearchSize, kBackSearchSize)) { + return LIBYUV_TRUE; // Success: Valid jpeg. + } + // Reduce search size for forward search. + sample_size = sample_size - kBackSearchSize + 1; + } + return ScanEOI(sample, sample_size); + +} + +#ifdef __cplusplus +} // extern "C" +} // namespace libyuv +#endif + diff --git a/media/libaom/src/third_party/libyuv/source/planar_functions.cc b/media/libaom/src/third_party/libyuv/source/planar_functions.cc new file mode 100644 index 000000000..b96bd5020 --- /dev/null +++ b/media/libaom/src/third_party/libyuv/source/planar_functions.cc @@ -0,0 +1,2555 @@ +/* + * Copyright 2011 The LibYuv Project Authors. All rights reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#include "libyuv/planar_functions.h" + +#include <string.h> // for memset() + +#include "libyuv/cpu_id.h" +#ifdef HAVE_JPEG +#include "libyuv/mjpeg_decoder.h" +#endif +#include "libyuv/row.h" + +#ifdef __cplusplus +namespace libyuv { +extern "C" { +#endif + +// Copy a plane of data +LIBYUV_API +void CopyPlane(const uint8* src_y, int src_stride_y, + uint8* dst_y, int dst_stride_y, + int width, int height) { + int y; + void (*CopyRow)(const uint8* src, uint8* dst, int width) = CopyRow_C; + // Coalesce rows. + if (src_stride_y == width && + dst_stride_y == width) { + width *= height; + height = 1; + src_stride_y = dst_stride_y = 0; + } + // Nothing to do. + if (src_y == dst_y && src_stride_y == dst_stride_y) { + return; + } +#if defined(HAS_COPYROW_SSE2) + if (TestCpuFlag(kCpuHasSSE2)) { + CopyRow = IS_ALIGNED(width, 32) ? CopyRow_SSE2 : CopyRow_Any_SSE2; + } +#endif +#if defined(HAS_COPYROW_AVX) + if (TestCpuFlag(kCpuHasAVX)) { + CopyRow = IS_ALIGNED(width, 64) ? CopyRow_AVX : CopyRow_Any_AVX; + } +#endif +#if defined(HAS_COPYROW_ERMS) + if (TestCpuFlag(kCpuHasERMS)) { + CopyRow = CopyRow_ERMS; + } +#endif +#if defined(HAS_COPYROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + CopyRow = IS_ALIGNED(width, 32) ? CopyRow_NEON : CopyRow_Any_NEON; + } +#endif +#if defined(HAS_COPYROW_MIPS) + if (TestCpuFlag(kCpuHasMIPS)) { + CopyRow = CopyRow_MIPS; + } +#endif + + // Copy plane + for (y = 0; y < height; ++y) { + CopyRow(src_y, dst_y, width); + src_y += src_stride_y; + dst_y += dst_stride_y; + } +} + +LIBYUV_API +void CopyPlane_16(const uint16* src_y, int src_stride_y, + uint16* dst_y, int dst_stride_y, + int width, int height) { + int y; + void (*CopyRow)(const uint16* src, uint16* dst, int width) = CopyRow_16_C; + // Coalesce rows. + if (src_stride_y == width && + dst_stride_y == width) { + width *= height; + height = 1; + src_stride_y = dst_stride_y = 0; + } +#if defined(HAS_COPYROW_16_SSE2) + if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 32)) { + CopyRow = CopyRow_16_SSE2; + } +#endif +#if defined(HAS_COPYROW_16_ERMS) + if (TestCpuFlag(kCpuHasERMS)) { + CopyRow = CopyRow_16_ERMS; + } +#endif +#if defined(HAS_COPYROW_16_NEON) + if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 32)) { + CopyRow = CopyRow_16_NEON; + } +#endif +#if defined(HAS_COPYROW_16_MIPS) + if (TestCpuFlag(kCpuHasMIPS)) { + CopyRow = CopyRow_16_MIPS; + } +#endif + + // Copy plane + for (y = 0; y < height; ++y) { + CopyRow(src_y, dst_y, width); + src_y += src_stride_y; + dst_y += dst_stride_y; + } +} + +// Copy I422. +LIBYUV_API +int I422Copy(const uint8* src_y, int src_stride_y, + const uint8* src_u, int src_stride_u, + const uint8* src_v, int src_stride_v, + uint8* dst_y, int dst_stride_y, + uint8* dst_u, int dst_stride_u, + uint8* dst_v, int dst_stride_v, + int width, int height) { + int halfwidth = (width + 1) >> 1; + if (!src_y || !src_u || !src_v || + !dst_y || !dst_u || !dst_v || + width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + src_y = src_y + (height - 1) * src_stride_y; + src_u = src_u + (height - 1) * src_stride_u; + src_v = src_v + (height - 1) * src_stride_v; + src_stride_y = -src_stride_y; + src_stride_u = -src_stride_u; + src_stride_v = -src_stride_v; + } + CopyPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height); + CopyPlane(src_u, src_stride_u, dst_u, dst_stride_u, halfwidth, height); + CopyPlane(src_v, src_stride_v, dst_v, dst_stride_v, halfwidth, height); + return 0; +} + +// Copy I444. +LIBYUV_API +int I444Copy(const uint8* src_y, int src_stride_y, + const uint8* src_u, int src_stride_u, + const uint8* src_v, int src_stride_v, + uint8* dst_y, int dst_stride_y, + uint8* dst_u, int dst_stride_u, + uint8* dst_v, int dst_stride_v, + int width, int height) { + if (!src_y || !src_u || !src_v || + !dst_y || !dst_u || !dst_v || + width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + src_y = src_y + (height - 1) * src_stride_y; + src_u = src_u + (height - 1) * src_stride_u; + src_v = src_v + (height - 1) * src_stride_v; + src_stride_y = -src_stride_y; + src_stride_u = -src_stride_u; + src_stride_v = -src_stride_v; + } + + CopyPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height); + CopyPlane(src_u, src_stride_u, dst_u, dst_stride_u, width, height); + CopyPlane(src_v, src_stride_v, dst_v, dst_stride_v, width, height); + return 0; +} + +// Copy I400. +LIBYUV_API +int I400ToI400(const uint8* src_y, int src_stride_y, + uint8* dst_y, int dst_stride_y, + int width, int height) { + if (!src_y || !dst_y || width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + src_y = src_y + (height - 1) * src_stride_y; + src_stride_y = -src_stride_y; + } + CopyPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height); + return 0; +} + +// Convert I420 to I400. +LIBYUV_API +int I420ToI400(const uint8* src_y, int src_stride_y, + const uint8* src_u, int src_stride_u, + const uint8* src_v, int src_stride_v, + uint8* dst_y, int dst_stride_y, + int width, int height) { + if (!src_y || !dst_y || width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + src_y = src_y + (height - 1) * src_stride_y; + src_stride_y = -src_stride_y; + } + CopyPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height); + return 0; +} + +// Mirror a plane of data. +void MirrorPlane(const uint8* src_y, int src_stride_y, + uint8* dst_y, int dst_stride_y, + int width, int height) { + int y; + void (*MirrorRow)(const uint8* src, uint8* dst, int width) = MirrorRow_C; + // Negative height means invert the image. + if (height < 0) { + height = -height; + src_y = src_y + (height - 1) * src_stride_y; + src_stride_y = -src_stride_y; + } +#if defined(HAS_MIRRORROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + MirrorRow = MirrorRow_Any_NEON; + if (IS_ALIGNED(width, 16)) { + MirrorRow = MirrorRow_NEON; + } + } +#endif +#if defined(HAS_MIRRORROW_SSE2) + if (TestCpuFlag(kCpuHasSSE2)) { + MirrorRow = MirrorRow_Any_SSE2; + if (IS_ALIGNED(width, 16)) { + MirrorRow = MirrorRow_SSE2; + } + } +#endif +#if defined(HAS_MIRRORROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + MirrorRow = MirrorRow_Any_SSSE3; + if (IS_ALIGNED(width, 16)) { + MirrorRow = MirrorRow_SSSE3; + } + } +#endif +#if defined(HAS_MIRRORROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + MirrorRow = MirrorRow_Any_AVX2; + if (IS_ALIGNED(width, 32)) { + MirrorRow = MirrorRow_AVX2; + } + } +#endif +// TODO(fbarchard): Mirror on mips handle unaligned memory. +#if defined(HAS_MIRRORROW_MIPS_DSPR2) + if (TestCpuFlag(kCpuHasMIPS_DSPR2) && + IS_ALIGNED(src_y, 4) && IS_ALIGNED(src_stride_y, 4) && + IS_ALIGNED(dst_y, 4) && IS_ALIGNED(dst_stride_y, 4)) { + MirrorRow = MirrorRow_MIPS_DSPR2; + } +#endif + + // Mirror plane + for (y = 0; y < height; ++y) { + MirrorRow(src_y, dst_y, width); + src_y += src_stride_y; + dst_y += dst_stride_y; + } +} + +// Convert YUY2 to I422. +LIBYUV_API +int YUY2ToI422(const uint8* src_yuy2, int src_stride_yuy2, + uint8* dst_y, int dst_stride_y, + uint8* dst_u, int dst_stride_u, + uint8* dst_v, int dst_stride_v, + int width, int height) { + int y; + void (*YUY2ToUV422Row)(const uint8* src_yuy2, + uint8* dst_u, uint8* dst_v, int pix) = + YUY2ToUV422Row_C; + void (*YUY2ToYRow)(const uint8* src_yuy2, uint8* dst_y, int pix) = + YUY2ToYRow_C; + // Negative height means invert the image. + if (height < 0) { + height = -height; + src_yuy2 = src_yuy2 + (height - 1) * src_stride_yuy2; + src_stride_yuy2 = -src_stride_yuy2; + } + // Coalesce rows. + if (src_stride_yuy2 == width * 2 && + dst_stride_y == width && + dst_stride_u * 2 == width && + dst_stride_v * 2 == width) { + width *= height; + height = 1; + src_stride_yuy2 = dst_stride_y = dst_stride_u = dst_stride_v = 0; + } +#if defined(HAS_YUY2TOYROW_SSE2) + if (TestCpuFlag(kCpuHasSSE2)) { + YUY2ToUV422Row = YUY2ToUV422Row_Any_SSE2; + YUY2ToYRow = YUY2ToYRow_Any_SSE2; + if (IS_ALIGNED(width, 16)) { + YUY2ToUV422Row = YUY2ToUV422Row_SSE2; + YUY2ToYRow = YUY2ToYRow_SSE2; + } + } +#endif +#if defined(HAS_YUY2TOYROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + YUY2ToUV422Row = YUY2ToUV422Row_Any_AVX2; + YUY2ToYRow = YUY2ToYRow_Any_AVX2; + if (IS_ALIGNED(width, 32)) { + YUY2ToUV422Row = YUY2ToUV422Row_AVX2; + YUY2ToYRow = YUY2ToYRow_AVX2; + } + } +#endif +#if defined(HAS_YUY2TOYROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + YUY2ToYRow = YUY2ToYRow_Any_NEON; + if (width >= 16) { + YUY2ToUV422Row = YUY2ToUV422Row_Any_NEON; + } + if (IS_ALIGNED(width, 16)) { + YUY2ToYRow = YUY2ToYRow_NEON; + YUY2ToUV422Row = YUY2ToUV422Row_NEON; + } + } +#endif + + for (y = 0; y < height; ++y) { + YUY2ToUV422Row(src_yuy2, dst_u, dst_v, width); + YUY2ToYRow(src_yuy2, dst_y, width); + src_yuy2 += src_stride_yuy2; + dst_y += dst_stride_y; + dst_u += dst_stride_u; + dst_v += dst_stride_v; + } + return 0; +} + +// Convert UYVY to I422. +LIBYUV_API +int UYVYToI422(const uint8* src_uyvy, int src_stride_uyvy, + uint8* dst_y, int dst_stride_y, + uint8* dst_u, int dst_stride_u, + uint8* dst_v, int dst_stride_v, + int width, int height) { + int y; + void (*UYVYToUV422Row)(const uint8* src_uyvy, + uint8* dst_u, uint8* dst_v, int pix) = + UYVYToUV422Row_C; + void (*UYVYToYRow)(const uint8* src_uyvy, + uint8* dst_y, int pix) = UYVYToYRow_C; + // Negative height means invert the image. + if (height < 0) { + height = -height; + src_uyvy = src_uyvy + (height - 1) * src_stride_uyvy; + src_stride_uyvy = -src_stride_uyvy; + } + // Coalesce rows. + if (src_stride_uyvy == width * 2 && + dst_stride_y == width && + dst_stride_u * 2 == width && + dst_stride_v * 2 == width) { + width *= height; + height = 1; + src_stride_uyvy = dst_stride_y = dst_stride_u = dst_stride_v = 0; + } +#if defined(HAS_UYVYTOYROW_SSE2) + if (TestCpuFlag(kCpuHasSSE2)) { + UYVYToUV422Row = UYVYToUV422Row_Any_SSE2; + UYVYToYRow = UYVYToYRow_Any_SSE2; + if (IS_ALIGNED(width, 16)) { + UYVYToUV422Row = UYVYToUV422Row_SSE2; + UYVYToYRow = UYVYToYRow_SSE2; + } + } +#endif +#if defined(HAS_UYVYTOYROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + UYVYToUV422Row = UYVYToUV422Row_Any_AVX2; + UYVYToYRow = UYVYToYRow_Any_AVX2; + if (IS_ALIGNED(width, 32)) { + UYVYToUV422Row = UYVYToUV422Row_AVX2; + UYVYToYRow = UYVYToYRow_AVX2; + } + } +#endif +#if defined(HAS_UYVYTOYROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + UYVYToYRow = UYVYToYRow_Any_NEON; + if (width >= 16) { + UYVYToUV422Row = UYVYToUV422Row_Any_NEON; + } + if (IS_ALIGNED(width, 16)) { + UYVYToYRow = UYVYToYRow_NEON; + UYVYToUV422Row = UYVYToUV422Row_NEON; + } + } +#endif + + for (y = 0; y < height; ++y) { + UYVYToUV422Row(src_uyvy, dst_u, dst_v, width); + UYVYToYRow(src_uyvy, dst_y, width); + src_uyvy += src_stride_uyvy; + dst_y += dst_stride_y; + dst_u += dst_stride_u; + dst_v += dst_stride_v; + } + return 0; +} + +// Mirror I400 with optional flipping +LIBYUV_API +int I400Mirror(const uint8* src_y, int src_stride_y, + uint8* dst_y, int dst_stride_y, + int width, int height) { + if (!src_y || !dst_y || + width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + src_y = src_y + (height - 1) * src_stride_y; + src_stride_y = -src_stride_y; + } + + MirrorPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height); + return 0; +} + +// Mirror I420 with optional flipping +LIBYUV_API +int I420Mirror(const uint8* src_y, int src_stride_y, + const uint8* src_u, int src_stride_u, + const uint8* src_v, int src_stride_v, + uint8* dst_y, int dst_stride_y, + uint8* dst_u, int dst_stride_u, + uint8* dst_v, int dst_stride_v, + int width, int height) { + int halfwidth = (width + 1) >> 1; + int halfheight = (height + 1) >> 1; + if (!src_y || !src_u || !src_v || !dst_y || !dst_u || !dst_v || + width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + halfheight = (height + 1) >> 1; + src_y = src_y + (height - 1) * src_stride_y; + src_u = src_u + (halfheight - 1) * src_stride_u; + src_v = src_v + (halfheight - 1) * src_stride_v; + src_stride_y = -src_stride_y; + src_stride_u = -src_stride_u; + src_stride_v = -src_stride_v; + } + + if (dst_y) { + MirrorPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height); + } + MirrorPlane(src_u, src_stride_u, dst_u, dst_stride_u, halfwidth, halfheight); + MirrorPlane(src_v, src_stride_v, dst_v, dst_stride_v, halfwidth, halfheight); + return 0; +} + +// ARGB mirror. +LIBYUV_API +int ARGBMirror(const uint8* src_argb, int src_stride_argb, + uint8* dst_argb, int dst_stride_argb, + int width, int height) { + int y; + void (*ARGBMirrorRow)(const uint8* src, uint8* dst, int width) = + ARGBMirrorRow_C; + if (!src_argb || !dst_argb || width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + src_argb = src_argb + (height - 1) * src_stride_argb; + src_stride_argb = -src_stride_argb; + } +#if defined(HAS_ARGBMIRRORROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + ARGBMirrorRow = ARGBMirrorRow_Any_NEON; + if (IS_ALIGNED(width, 4)) { + ARGBMirrorRow = ARGBMirrorRow_NEON; + } + } +#endif +#if defined(HAS_ARGBMIRRORROW_SSE2) + if (TestCpuFlag(kCpuHasSSE2)) { + ARGBMirrorRow = ARGBMirrorRow_Any_SSE2; + if (IS_ALIGNED(width, 4)) { + ARGBMirrorRow = ARGBMirrorRow_SSE2; + } + } +#endif +#if defined(HAS_ARGBMIRRORROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + ARGBMirrorRow = ARGBMirrorRow_Any_AVX2; + if (IS_ALIGNED(width, 8)) { + ARGBMirrorRow = ARGBMirrorRow_AVX2; + } + } +#endif + + // Mirror plane + for (y = 0; y < height; ++y) { + ARGBMirrorRow(src_argb, dst_argb, width); + src_argb += src_stride_argb; + dst_argb += dst_stride_argb; + } + return 0; +} + +// Get a blender that optimized for the CPU and pixel count. +// As there are 6 blenders to choose from, the caller should try to use +// the same blend function for all pixels if possible. +LIBYUV_API +ARGBBlendRow GetARGBBlend() { + void (*ARGBBlendRow)(const uint8* src_argb, const uint8* src_argb1, + uint8* dst_argb, int width) = ARGBBlendRow_C; +#if defined(HAS_ARGBBLENDROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + ARGBBlendRow = ARGBBlendRow_SSSE3; + return ARGBBlendRow; + } +#endif +#if defined(HAS_ARGBBLENDROW_SSE2) + if (TestCpuFlag(kCpuHasSSE2)) { + ARGBBlendRow = ARGBBlendRow_SSE2; + } +#endif +#if defined(HAS_ARGBBLENDROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + ARGBBlendRow = ARGBBlendRow_NEON; + } +#endif + return ARGBBlendRow; +} + +// Alpha Blend 2 ARGB images and store to destination. +LIBYUV_API +int ARGBBlend(const uint8* src_argb0, int src_stride_argb0, + const uint8* src_argb1, int src_stride_argb1, + uint8* dst_argb, int dst_stride_argb, + int width, int height) { + int y; + void (*ARGBBlendRow)(const uint8* src_argb, const uint8* src_argb1, + uint8* dst_argb, int width) = GetARGBBlend(); + if (!src_argb0 || !src_argb1 || !dst_argb || width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + dst_argb = dst_argb + (height - 1) * dst_stride_argb; + dst_stride_argb = -dst_stride_argb; + } + // Coalesce rows. + if (src_stride_argb0 == width * 4 && + src_stride_argb1 == width * 4 && + dst_stride_argb == width * 4) { + width *= height; + height = 1; + src_stride_argb0 = src_stride_argb1 = dst_stride_argb = 0; + } + + for (y = 0; y < height; ++y) { + ARGBBlendRow(src_argb0, src_argb1, dst_argb, width); + src_argb0 += src_stride_argb0; + src_argb1 += src_stride_argb1; + dst_argb += dst_stride_argb; + } + return 0; +} + +// Multiply 2 ARGB images and store to destination. +LIBYUV_API +int ARGBMultiply(const uint8* src_argb0, int src_stride_argb0, + const uint8* src_argb1, int src_stride_argb1, + uint8* dst_argb, int dst_stride_argb, + int width, int height) { + int y; + void (*ARGBMultiplyRow)(const uint8* src0, const uint8* src1, uint8* dst, + int width) = ARGBMultiplyRow_C; + if (!src_argb0 || !src_argb1 || !dst_argb || width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + dst_argb = dst_argb + (height - 1) * dst_stride_argb; + dst_stride_argb = -dst_stride_argb; + } + // Coalesce rows. + if (src_stride_argb0 == width * 4 && + src_stride_argb1 == width * 4 && + dst_stride_argb == width * 4) { + width *= height; + height = 1; + src_stride_argb0 = src_stride_argb1 = dst_stride_argb = 0; + } +#if defined(HAS_ARGBMULTIPLYROW_SSE2) + if (TestCpuFlag(kCpuHasSSE2)) { + ARGBMultiplyRow = ARGBMultiplyRow_Any_SSE2; + if (IS_ALIGNED(width, 4)) { + ARGBMultiplyRow = ARGBMultiplyRow_SSE2; + } + } +#endif +#if defined(HAS_ARGBMULTIPLYROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + ARGBMultiplyRow = ARGBMultiplyRow_Any_AVX2; + if (IS_ALIGNED(width, 8)) { + ARGBMultiplyRow = ARGBMultiplyRow_AVX2; + } + } +#endif +#if defined(HAS_ARGBMULTIPLYROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + ARGBMultiplyRow = ARGBMultiplyRow_Any_NEON; + if (IS_ALIGNED(width, 8)) { + ARGBMultiplyRow = ARGBMultiplyRow_NEON; + } + } +#endif + + // Multiply plane + for (y = 0; y < height; ++y) { + ARGBMultiplyRow(src_argb0, src_argb1, dst_argb, width); + src_argb0 += src_stride_argb0; + src_argb1 += src_stride_argb1; + dst_argb += dst_stride_argb; + } + return 0; +} + +// Add 2 ARGB images and store to destination. +LIBYUV_API +int ARGBAdd(const uint8* src_argb0, int src_stride_argb0, + const uint8* src_argb1, int src_stride_argb1, + uint8* dst_argb, int dst_stride_argb, + int width, int height) { + int y; + void (*ARGBAddRow)(const uint8* src0, const uint8* src1, uint8* dst, + int width) = ARGBAddRow_C; + if (!src_argb0 || !src_argb1 || !dst_argb || width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + dst_argb = dst_argb + (height - 1) * dst_stride_argb; + dst_stride_argb = -dst_stride_argb; + } + // Coalesce rows. + if (src_stride_argb0 == width * 4 && + src_stride_argb1 == width * 4 && + dst_stride_argb == width * 4) { + width *= height; + height = 1; + src_stride_argb0 = src_stride_argb1 = dst_stride_argb = 0; + } +#if defined(HAS_ARGBADDROW_SSE2) && (defined(_MSC_VER) && !defined(__clang__)) + if (TestCpuFlag(kCpuHasSSE2)) { + ARGBAddRow = ARGBAddRow_SSE2; + } +#endif +#if defined(HAS_ARGBADDROW_SSE2) && !(defined(_MSC_VER) && !defined(__clang__)) + if (TestCpuFlag(kCpuHasSSE2)) { + ARGBAddRow = ARGBAddRow_Any_SSE2; + if (IS_ALIGNED(width, 4)) { + ARGBAddRow = ARGBAddRow_SSE2; + } + } +#endif +#if defined(HAS_ARGBADDROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + ARGBAddRow = ARGBAddRow_Any_AVX2; + if (IS_ALIGNED(width, 8)) { + ARGBAddRow = ARGBAddRow_AVX2; + } + } +#endif +#if defined(HAS_ARGBADDROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + ARGBAddRow = ARGBAddRow_Any_NEON; + if (IS_ALIGNED(width, 8)) { + ARGBAddRow = ARGBAddRow_NEON; + } + } +#endif + + // Add plane + for (y = 0; y < height; ++y) { + ARGBAddRow(src_argb0, src_argb1, dst_argb, width); + src_argb0 += src_stride_argb0; + src_argb1 += src_stride_argb1; + dst_argb += dst_stride_argb; + } + return 0; +} + +// Subtract 2 ARGB images and store to destination. +LIBYUV_API +int ARGBSubtract(const uint8* src_argb0, int src_stride_argb0, + const uint8* src_argb1, int src_stride_argb1, + uint8* dst_argb, int dst_stride_argb, + int width, int height) { + int y; + void (*ARGBSubtractRow)(const uint8* src0, const uint8* src1, uint8* dst, + int width) = ARGBSubtractRow_C; + if (!src_argb0 || !src_argb1 || !dst_argb || width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + dst_argb = dst_argb + (height - 1) * dst_stride_argb; + dst_stride_argb = -dst_stride_argb; + } + // Coalesce rows. + if (src_stride_argb0 == width * 4 && + src_stride_argb1 == width * 4 && + dst_stride_argb == width * 4) { + width *= height; + height = 1; + src_stride_argb0 = src_stride_argb1 = dst_stride_argb = 0; + } +#if defined(HAS_ARGBSUBTRACTROW_SSE2) + if (TestCpuFlag(kCpuHasSSE2)) { + ARGBSubtractRow = ARGBSubtractRow_Any_SSE2; + if (IS_ALIGNED(width, 4)) { + ARGBSubtractRow = ARGBSubtractRow_SSE2; + } + } +#endif +#if defined(HAS_ARGBSUBTRACTROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + ARGBSubtractRow = ARGBSubtractRow_Any_AVX2; + if (IS_ALIGNED(width, 8)) { + ARGBSubtractRow = ARGBSubtractRow_AVX2; + } + } +#endif +#if defined(HAS_ARGBSUBTRACTROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + ARGBSubtractRow = ARGBSubtractRow_Any_NEON; + if (IS_ALIGNED(width, 8)) { + ARGBSubtractRow = ARGBSubtractRow_NEON; + } + } +#endif + + // Subtract plane + for (y = 0; y < height; ++y) { + ARGBSubtractRow(src_argb0, src_argb1, dst_argb, width); + src_argb0 += src_stride_argb0; + src_argb1 += src_stride_argb1; + dst_argb += dst_stride_argb; + } + return 0; +} + +// Convert I422 to BGRA. +LIBYUV_API +int I422ToBGRA(const uint8* src_y, int src_stride_y, + const uint8* src_u, int src_stride_u, + const uint8* src_v, int src_stride_v, + uint8* dst_bgra, int dst_stride_bgra, + int width, int height) { + int y; + void (*I422ToBGRARow)(const uint8* y_buf, + const uint8* u_buf, + const uint8* v_buf, + uint8* rgb_buf, + int width) = I422ToBGRARow_C; + if (!src_y || !src_u || !src_v || + !dst_bgra || + width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + dst_bgra = dst_bgra + (height - 1) * dst_stride_bgra; + dst_stride_bgra = -dst_stride_bgra; + } + // Coalesce rows. + if (src_stride_y == width && + src_stride_u * 2 == width && + src_stride_v * 2 == width && + dst_stride_bgra == width * 4) { + width *= height; + height = 1; + src_stride_y = src_stride_u = src_stride_v = dst_stride_bgra = 0; + } +#if defined(HAS_I422TOBGRAROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + I422ToBGRARow = I422ToBGRARow_Any_SSSE3; + if (IS_ALIGNED(width, 8)) { + I422ToBGRARow = I422ToBGRARow_SSSE3; + } + } +#endif +#if defined(HAS_I422TOBGRAROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + I422ToBGRARow = I422ToBGRARow_Any_AVX2; + if (IS_ALIGNED(width, 16)) { + I422ToBGRARow = I422ToBGRARow_AVX2; + } + } +#endif +#if defined(HAS_I422TOBGRAROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + I422ToBGRARow = I422ToBGRARow_Any_NEON; + if (IS_ALIGNED(width, 8)) { + I422ToBGRARow = I422ToBGRARow_NEON; + } + } +#endif +#if defined(HAS_I422TOBGRAROW_MIPS_DSPR2) + if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(width, 4) && + IS_ALIGNED(src_y, 4) && IS_ALIGNED(src_stride_y, 4) && + IS_ALIGNED(src_u, 2) && IS_ALIGNED(src_stride_u, 2) && + IS_ALIGNED(src_v, 2) && IS_ALIGNED(src_stride_v, 2) && + IS_ALIGNED(dst_bgra, 4) && IS_ALIGNED(dst_stride_bgra, 4)) { + I422ToBGRARow = I422ToBGRARow_MIPS_DSPR2; + } +#endif + + for (y = 0; y < height; ++y) { + I422ToBGRARow(src_y, src_u, src_v, dst_bgra, width); + dst_bgra += dst_stride_bgra; + src_y += src_stride_y; + src_u += src_stride_u; + src_v += src_stride_v; + } + return 0; +} + +// Convert I422 to ABGR. +LIBYUV_API +int I422ToABGR(const uint8* src_y, int src_stride_y, + const uint8* src_u, int src_stride_u, + const uint8* src_v, int src_stride_v, + uint8* dst_abgr, int dst_stride_abgr, + int width, int height) { + int y; + void (*I422ToABGRRow)(const uint8* y_buf, + const uint8* u_buf, + const uint8* v_buf, + uint8* rgb_buf, + int width) = I422ToABGRRow_C; + if (!src_y || !src_u || !src_v || + !dst_abgr || + width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + dst_abgr = dst_abgr + (height - 1) * dst_stride_abgr; + dst_stride_abgr = -dst_stride_abgr; + } + // Coalesce rows. + if (src_stride_y == width && + src_stride_u * 2 == width && + src_stride_v * 2 == width && + dst_stride_abgr == width * 4) { + width *= height; + height = 1; + src_stride_y = src_stride_u = src_stride_v = dst_stride_abgr = 0; + } +#if defined(HAS_I422TOABGRROW_NEON) + if (TestCpuFlag(kCpuHasNEON) && width >= 8) { + I422ToABGRRow = I422ToABGRRow_Any_NEON; + if (IS_ALIGNED(width, 8)) { + I422ToABGRRow = I422ToABGRRow_NEON; + } + } +#endif +#if defined(HAS_I422TOABGRROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + I422ToABGRRow = I422ToABGRRow_Any_SSSE3; + if (IS_ALIGNED(width, 8)) { + I422ToABGRRow = I422ToABGRRow_SSSE3; + } + } +#endif +#if defined(HAS_I422TOABGRROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + I422ToABGRRow = I422ToABGRRow_Any_AVX2; + if (IS_ALIGNED(width, 16)) { + I422ToABGRRow = I422ToABGRRow_AVX2; + } + } +#endif + + for (y = 0; y < height; ++y) { + I422ToABGRRow(src_y, src_u, src_v, dst_abgr, width); + dst_abgr += dst_stride_abgr; + src_y += src_stride_y; + src_u += src_stride_u; + src_v += src_stride_v; + } + return 0; +} + +// Convert I422 to RGBA. +LIBYUV_API +int I422ToRGBA(const uint8* src_y, int src_stride_y, + const uint8* src_u, int src_stride_u, + const uint8* src_v, int src_stride_v, + uint8* dst_rgba, int dst_stride_rgba, + int width, int height) { + int y; + void (*I422ToRGBARow)(const uint8* y_buf, + const uint8* u_buf, + const uint8* v_buf, + uint8* rgb_buf, + int width) = I422ToRGBARow_C; + if (!src_y || !src_u || !src_v || + !dst_rgba || + width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + dst_rgba = dst_rgba + (height - 1) * dst_stride_rgba; + dst_stride_rgba = -dst_stride_rgba; + } + // Coalesce rows. + if (src_stride_y == width && + src_stride_u * 2 == width && + src_stride_v * 2 == width && + dst_stride_rgba == width * 4) { + width *= height; + height = 1; + src_stride_y = src_stride_u = src_stride_v = dst_stride_rgba = 0; + } +#if defined(HAS_I422TORGBAROW_NEON) + if (TestCpuFlag(kCpuHasNEON) && width >= 8) { + I422ToRGBARow = I422ToRGBARow_Any_NEON; + if (IS_ALIGNED(width, 8)) { + I422ToRGBARow = I422ToRGBARow_NEON; + } + } +#endif +#if defined(HAS_I422TORGBAROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + I422ToRGBARow = I422ToRGBARow_Any_SSSE3; + if (IS_ALIGNED(width, 8)) { + I422ToRGBARow = I422ToRGBARow_SSSE3; + } + } +#endif +#if defined(HAS_I422TORGBAROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + I422ToRGBARow = I422ToRGBARow_Any_AVX2; + if (IS_ALIGNED(width, 16)) { + I422ToRGBARow = I422ToRGBARow_AVX2; + } + } +#endif + + for (y = 0; y < height; ++y) { + I422ToRGBARow(src_y, src_u, src_v, dst_rgba, width); + dst_rgba += dst_stride_rgba; + src_y += src_stride_y; + src_u += src_stride_u; + src_v += src_stride_v; + } + return 0; +} + +// Convert NV12 to RGB565. +LIBYUV_API +int NV12ToRGB565(const uint8* src_y, int src_stride_y, + const uint8* src_uv, int src_stride_uv, + uint8* dst_rgb565, int dst_stride_rgb565, + int width, int height) { + int y; + void (*NV12ToRGB565Row)(const uint8* y_buf, + const uint8* uv_buf, + uint8* rgb_buf, + int width) = NV12ToRGB565Row_C; + if (!src_y || !src_uv || !dst_rgb565 || + width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + dst_rgb565 = dst_rgb565 + (height - 1) * dst_stride_rgb565; + dst_stride_rgb565 = -dst_stride_rgb565; + } +#if defined(HAS_NV12TORGB565ROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + NV12ToRGB565Row = NV12ToRGB565Row_Any_SSSE3; + if (IS_ALIGNED(width, 8)) { + NV12ToRGB565Row = NV12ToRGB565Row_SSSE3; + } + } +#endif +#if defined(HAS_NV12TORGB565ROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + NV12ToRGB565Row = NV12ToRGB565Row_Any_AVX2; + if (IS_ALIGNED(width, 16)) { + NV12ToRGB565Row = NV12ToRGB565Row_AVX2; + } + } +#endif +#if defined(HAS_NV12TORGB565ROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + NV12ToRGB565Row = NV12ToRGB565Row_Any_NEON; + if (IS_ALIGNED(width, 8)) { + NV12ToRGB565Row = NV12ToRGB565Row_NEON; + } + } +#endif + + for (y = 0; y < height; ++y) { + NV12ToRGB565Row(src_y, src_uv, dst_rgb565, width); + dst_rgb565 += dst_stride_rgb565; + src_y += src_stride_y; + if (y & 1) { + src_uv += src_stride_uv; + } + } + return 0; +} + +// Convert NV21 to RGB565. +LIBYUV_API +int NV21ToRGB565(const uint8* src_y, int src_stride_y, + const uint8* src_vu, int src_stride_vu, + uint8* dst_rgb565, int dst_stride_rgb565, + int width, int height) { + int y; + void (*NV21ToRGB565Row)(const uint8* y_buf, + const uint8* src_vu, + uint8* rgb_buf, + int width) = NV21ToRGB565Row_C; + if (!src_y || !src_vu || !dst_rgb565 || + width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + dst_rgb565 = dst_rgb565 + (height - 1) * dst_stride_rgb565; + dst_stride_rgb565 = -dst_stride_rgb565; + } +#if defined(HAS_NV21TORGB565ROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + NV21ToRGB565Row = NV21ToRGB565Row_Any_SSSE3; + if (IS_ALIGNED(width, 8)) { + NV21ToRGB565Row = NV21ToRGB565Row_SSSE3; + } + } +#endif +#if defined(HAS_NV21TORGB565ROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + NV21ToRGB565Row = NV21ToRGB565Row_Any_AVX2; + if (IS_ALIGNED(width, 16)) { + NV21ToRGB565Row = NV21ToRGB565Row_AVX2; + } + } +#endif +#if defined(HAS_NV21TORGB565ROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + NV21ToRGB565Row = NV21ToRGB565Row_Any_NEON; + if (IS_ALIGNED(width, 8)) { + NV21ToRGB565Row = NV21ToRGB565Row_NEON; + } + } +#endif + + for (y = 0; y < height; ++y) { + NV21ToRGB565Row(src_y, src_vu, dst_rgb565, width); + dst_rgb565 += dst_stride_rgb565; + src_y += src_stride_y; + if (y & 1) { + src_vu += src_stride_vu; + } + } + return 0; +} + +LIBYUV_API +void SetPlane(uint8* dst_y, int dst_stride_y, + int width, int height, + uint32 value) { + int y; + void (*SetRow)(uint8* dst, uint8 value, int pix) = SetRow_C; + if (height < 0) { + height = -height; + dst_y = dst_y + (height - 1) * dst_stride_y; + dst_stride_y = -dst_stride_y; + } + // Coalesce rows. + if (dst_stride_y == width) { + width *= height; + height = 1; + dst_stride_y = 0; + } +#if defined(HAS_SETROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + SetRow = SetRow_Any_NEON; + if (IS_ALIGNED(width, 16)) { + SetRow = SetRow_NEON; + } + } +#endif +#if defined(HAS_SETROW_X86) + if (TestCpuFlag(kCpuHasX86)) { + SetRow = SetRow_Any_X86; + if (IS_ALIGNED(width, 4)) { + SetRow = SetRow_X86; + } + } +#endif +#if defined(HAS_SETROW_ERMS) + if (TestCpuFlag(kCpuHasERMS)) { + SetRow = SetRow_ERMS; + } +#endif + + // Set plane + for (y = 0; y < height; ++y) { + SetRow(dst_y, value, width); + dst_y += dst_stride_y; + } +} + +// Draw a rectangle into I420 +LIBYUV_API +int I420Rect(uint8* dst_y, int dst_stride_y, + uint8* dst_u, int dst_stride_u, + uint8* dst_v, int dst_stride_v, + int x, int y, + int width, int height, + int value_y, int value_u, int value_v) { + int halfwidth = (width + 1) >> 1; + int halfheight = (height + 1) >> 1; + uint8* start_y = dst_y + y * dst_stride_y + x; + uint8* start_u = dst_u + (y / 2) * dst_stride_u + (x / 2); + uint8* start_v = dst_v + (y / 2) * dst_stride_v + (x / 2); + if (!dst_y || !dst_u || !dst_v || + width <= 0 || height == 0 || + x < 0 || y < 0 || + value_y < 0 || value_y > 255 || + value_u < 0 || value_u > 255 || + value_v < 0 || value_v > 255) { + return -1; + } + + SetPlane(start_y, dst_stride_y, width, height, value_y); + SetPlane(start_u, dst_stride_u, halfwidth, halfheight, value_u); + SetPlane(start_v, dst_stride_v, halfwidth, halfheight, value_v); + return 0; +} + +// Draw a rectangle into ARGB +LIBYUV_API +int ARGBRect(uint8* dst_argb, int dst_stride_argb, + int dst_x, int dst_y, + int width, int height, + uint32 value) { + int y; + void (*ARGBSetRow)(uint8* dst_argb, uint32 value, int pix) = ARGBSetRow_C; + if (!dst_argb || + width <= 0 || height == 0 || + dst_x < 0 || dst_y < 0) { + return -1; + } + if (height < 0) { + height = -height; + dst_argb = dst_argb + (height - 1) * dst_stride_argb; + dst_stride_argb = -dst_stride_argb; + } + dst_argb += dst_y * dst_stride_argb + dst_x * 4; + // Coalesce rows. + if (dst_stride_argb == width * 4) { + width *= height; + height = 1; + dst_stride_argb = 0; + } + +#if defined(HAS_ARGBSETROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + ARGBSetRow = ARGBSetRow_Any_NEON; + if (IS_ALIGNED(width, 4)) { + ARGBSetRow = ARGBSetRow_NEON; + } + } +#endif +#if defined(HAS_ARGBSETROW_X86) + if (TestCpuFlag(kCpuHasX86)) { + ARGBSetRow = ARGBSetRow_X86; + } +#endif + + // Set plane + for (y = 0; y < height; ++y) { + ARGBSetRow(dst_argb, value, width); + dst_argb += dst_stride_argb; + } + return 0; +} + +// Convert unattentuated ARGB to preattenuated ARGB. +// An unattenutated ARGB alpha blend uses the formula +// p = a * f + (1 - a) * b +// where +// p is output pixel +// f is foreground pixel +// b is background pixel +// a is alpha value from foreground pixel +// An preattenutated ARGB alpha blend uses the formula +// p = f + (1 - a) * b +// where +// f is foreground pixel premultiplied by alpha + +LIBYUV_API +int ARGBAttenuate(const uint8* src_argb, int src_stride_argb, + uint8* dst_argb, int dst_stride_argb, + int width, int height) { + int y; + void (*ARGBAttenuateRow)(const uint8* src_argb, uint8* dst_argb, + int width) = ARGBAttenuateRow_C; + if (!src_argb || !dst_argb || width <= 0 || height == 0) { + return -1; + } + if (height < 0) { + height = -height; + src_argb = src_argb + (height - 1) * src_stride_argb; + src_stride_argb = -src_stride_argb; + } + // Coalesce rows. + if (src_stride_argb == width * 4 && + dst_stride_argb == width * 4) { + width *= height; + height = 1; + src_stride_argb = dst_stride_argb = 0; + } +#if defined(HAS_ARGBATTENUATEROW_SSE2) + if (TestCpuFlag(kCpuHasSSE2)) { + ARGBAttenuateRow = ARGBAttenuateRow_Any_SSE2; + if (IS_ALIGNED(width, 4)) { + ARGBAttenuateRow = ARGBAttenuateRow_SSE2; + } + } +#endif +#if defined(HAS_ARGBATTENUATEROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + ARGBAttenuateRow = ARGBAttenuateRow_Any_SSSE3; + if (IS_ALIGNED(width, 4)) { + ARGBAttenuateRow = ARGBAttenuateRow_SSSE3; + } + } +#endif +#if defined(HAS_ARGBATTENUATEROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + ARGBAttenuateRow = ARGBAttenuateRow_Any_AVX2; + if (IS_ALIGNED(width, 8)) { + ARGBAttenuateRow = ARGBAttenuateRow_AVX2; + } + } +#endif +#if defined(HAS_ARGBATTENUATEROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + ARGBAttenuateRow = ARGBAttenuateRow_Any_NEON; + if (IS_ALIGNED(width, 8)) { + ARGBAttenuateRow = ARGBAttenuateRow_NEON; + } + } +#endif + + for (y = 0; y < height; ++y) { + ARGBAttenuateRow(src_argb, dst_argb, width); + src_argb += src_stride_argb; + dst_argb += dst_stride_argb; + } + return 0; +} + +// Convert preattentuated ARGB to unattenuated ARGB. +LIBYUV_API +int ARGBUnattenuate(const uint8* src_argb, int src_stride_argb, + uint8* dst_argb, int dst_stride_argb, + int width, int height) { + int y; + void (*ARGBUnattenuateRow)(const uint8* src_argb, uint8* dst_argb, + int width) = ARGBUnattenuateRow_C; + if (!src_argb || !dst_argb || width <= 0 || height == 0) { + return -1; + } + if (height < 0) { + height = -height; + src_argb = src_argb + (height - 1) * src_stride_argb; + src_stride_argb = -src_stride_argb; + } + // Coalesce rows. + if (src_stride_argb == width * 4 && + dst_stride_argb == width * 4) { + width *= height; + height = 1; + src_stride_argb = dst_stride_argb = 0; + } +#if defined(HAS_ARGBUNATTENUATEROW_SSE2) + if (TestCpuFlag(kCpuHasSSE2)) { + ARGBUnattenuateRow = ARGBUnattenuateRow_Any_SSE2; + if (IS_ALIGNED(width, 4)) { + ARGBUnattenuateRow = ARGBUnattenuateRow_SSE2; + } + } +#endif +#if defined(HAS_ARGBUNATTENUATEROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + ARGBUnattenuateRow = ARGBUnattenuateRow_Any_AVX2; + if (IS_ALIGNED(width, 8)) { + ARGBUnattenuateRow = ARGBUnattenuateRow_AVX2; + } + } +#endif +// TODO(fbarchard): Neon version. + + for (y = 0; y < height; ++y) { + ARGBUnattenuateRow(src_argb, dst_argb, width); + src_argb += src_stride_argb; + dst_argb += dst_stride_argb; + } + return 0; +} + +// Convert ARGB to Grayed ARGB. +LIBYUV_API +int ARGBGrayTo(const uint8* src_argb, int src_stride_argb, + uint8* dst_argb, int dst_stride_argb, + int width, int height) { + int y; + void (*ARGBGrayRow)(const uint8* src_argb, uint8* dst_argb, + int width) = ARGBGrayRow_C; + if (!src_argb || !dst_argb || width <= 0 || height == 0) { + return -1; + } + if (height < 0) { + height = -height; + src_argb = src_argb + (height - 1) * src_stride_argb; + src_stride_argb = -src_stride_argb; + } + // Coalesce rows. + if (src_stride_argb == width * 4 && + dst_stride_argb == width * 4) { + width *= height; + height = 1; + src_stride_argb = dst_stride_argb = 0; + } +#if defined(HAS_ARGBGRAYROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3) && IS_ALIGNED(width, 8)) { + ARGBGrayRow = ARGBGrayRow_SSSE3; + } +#endif +#if defined(HAS_ARGBGRAYROW_NEON) + if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 8)) { + ARGBGrayRow = ARGBGrayRow_NEON; + } +#endif + + for (y = 0; y < height; ++y) { + ARGBGrayRow(src_argb, dst_argb, width); + src_argb += src_stride_argb; + dst_argb += dst_stride_argb; + } + return 0; +} + +// Make a rectangle of ARGB gray scale. +LIBYUV_API +int ARGBGray(uint8* dst_argb, int dst_stride_argb, + int dst_x, int dst_y, + int width, int height) { + int y; + void (*ARGBGrayRow)(const uint8* src_argb, uint8* dst_argb, + int width) = ARGBGrayRow_C; + uint8* dst = dst_argb + dst_y * dst_stride_argb + dst_x * 4; + if (!dst_argb || width <= 0 || height <= 0 || dst_x < 0 || dst_y < 0) { + return -1; + } + // Coalesce rows. + if (dst_stride_argb == width * 4) { + width *= height; + height = 1; + dst_stride_argb = 0; + } +#if defined(HAS_ARGBGRAYROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3) && IS_ALIGNED(width, 8)) { + ARGBGrayRow = ARGBGrayRow_SSSE3; + } +#endif +#if defined(HAS_ARGBGRAYROW_NEON) + if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 8)) { + ARGBGrayRow = ARGBGrayRow_NEON; + } +#endif + for (y = 0; y < height; ++y) { + ARGBGrayRow(dst, dst, width); + dst += dst_stride_argb; + } + return 0; +} + +// Make a rectangle of ARGB Sepia tone. +LIBYUV_API +int ARGBSepia(uint8* dst_argb, int dst_stride_argb, + int dst_x, int dst_y, int width, int height) { + int y; + void (*ARGBSepiaRow)(uint8* dst_argb, int width) = ARGBSepiaRow_C; + uint8* dst = dst_argb + dst_y * dst_stride_argb + dst_x * 4; + if (!dst_argb || width <= 0 || height <= 0 || dst_x < 0 || dst_y < 0) { + return -1; + } + // Coalesce rows. + if (dst_stride_argb == width * 4) { + width *= height; + height = 1; + dst_stride_argb = 0; + } +#if defined(HAS_ARGBSEPIAROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3) && IS_ALIGNED(width, 8)) { + ARGBSepiaRow = ARGBSepiaRow_SSSE3; + } +#endif +#if defined(HAS_ARGBSEPIAROW_NEON) + if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 8)) { + ARGBSepiaRow = ARGBSepiaRow_NEON; + } +#endif + for (y = 0; y < height; ++y) { + ARGBSepiaRow(dst, width); + dst += dst_stride_argb; + } + return 0; +} + +// Apply a 4x4 matrix to each ARGB pixel. +// Note: Normally for shading, but can be used to swizzle or invert. +LIBYUV_API +int ARGBColorMatrix(const uint8* src_argb, int src_stride_argb, + uint8* dst_argb, int dst_stride_argb, + const int8* matrix_argb, + int width, int height) { + int y; + void (*ARGBColorMatrixRow)(const uint8* src_argb, uint8* dst_argb, + const int8* matrix_argb, int width) = ARGBColorMatrixRow_C; + if (!src_argb || !dst_argb || !matrix_argb || width <= 0 || height == 0) { + return -1; + } + if (height < 0) { + height = -height; + src_argb = src_argb + (height - 1) * src_stride_argb; + src_stride_argb = -src_stride_argb; + } + // Coalesce rows. + if (src_stride_argb == width * 4 && + dst_stride_argb == width * 4) { + width *= height; + height = 1; + src_stride_argb = dst_stride_argb = 0; + } +#if defined(HAS_ARGBCOLORMATRIXROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3) && IS_ALIGNED(width, 8)) { + ARGBColorMatrixRow = ARGBColorMatrixRow_SSSE3; + } +#endif +#if defined(HAS_ARGBCOLORMATRIXROW_NEON) + if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 8)) { + ARGBColorMatrixRow = ARGBColorMatrixRow_NEON; + } +#endif + for (y = 0; y < height; ++y) { + ARGBColorMatrixRow(src_argb, dst_argb, matrix_argb, width); + src_argb += src_stride_argb; + dst_argb += dst_stride_argb; + } + return 0; +} + +// Apply a 4x3 matrix to each ARGB pixel. +// Deprecated. +LIBYUV_API +int RGBColorMatrix(uint8* dst_argb, int dst_stride_argb, + const int8* matrix_rgb, + int dst_x, int dst_y, int width, int height) { + SIMD_ALIGNED(int8 matrix_argb[16]); + uint8* dst = dst_argb + dst_y * dst_stride_argb + dst_x * 4; + if (!dst_argb || !matrix_rgb || width <= 0 || height <= 0 || + dst_x < 0 || dst_y < 0) { + return -1; + } + + // Convert 4x3 7 bit matrix to 4x4 6 bit matrix. + matrix_argb[0] = matrix_rgb[0] / 2; + matrix_argb[1] = matrix_rgb[1] / 2; + matrix_argb[2] = matrix_rgb[2] / 2; + matrix_argb[3] = matrix_rgb[3] / 2; + matrix_argb[4] = matrix_rgb[4] / 2; + matrix_argb[5] = matrix_rgb[5] / 2; + matrix_argb[6] = matrix_rgb[6] / 2; + matrix_argb[7] = matrix_rgb[7] / 2; + matrix_argb[8] = matrix_rgb[8] / 2; + matrix_argb[9] = matrix_rgb[9] / 2; + matrix_argb[10] = matrix_rgb[10] / 2; + matrix_argb[11] = matrix_rgb[11] / 2; + matrix_argb[14] = matrix_argb[13] = matrix_argb[12] = 0; + matrix_argb[15] = 64; // 1.0 + + return ARGBColorMatrix((const uint8*)(dst), dst_stride_argb, + dst, dst_stride_argb, + &matrix_argb[0], width, height); +} + +// Apply a color table each ARGB pixel. +// Table contains 256 ARGB values. +LIBYUV_API +int ARGBColorTable(uint8* dst_argb, int dst_stride_argb, + const uint8* table_argb, + int dst_x, int dst_y, int width, int height) { + int y; + void (*ARGBColorTableRow)(uint8* dst_argb, const uint8* table_argb, + int width) = ARGBColorTableRow_C; + uint8* dst = dst_argb + dst_y * dst_stride_argb + dst_x * 4; + if (!dst_argb || !table_argb || width <= 0 || height <= 0 || + dst_x < 0 || dst_y < 0) { + return -1; + } + // Coalesce rows. + if (dst_stride_argb == width * 4) { + width *= height; + height = 1; + dst_stride_argb = 0; + } +#if defined(HAS_ARGBCOLORTABLEROW_X86) + if (TestCpuFlag(kCpuHasX86)) { + ARGBColorTableRow = ARGBColorTableRow_X86; + } +#endif + for (y = 0; y < height; ++y) { + ARGBColorTableRow(dst, table_argb, width); + dst += dst_stride_argb; + } + return 0; +} + +// Apply a color table each ARGB pixel but preserve destination alpha. +// Table contains 256 ARGB values. +LIBYUV_API +int RGBColorTable(uint8* dst_argb, int dst_stride_argb, + const uint8* table_argb, + int dst_x, int dst_y, int width, int height) { + int y; + void (*RGBColorTableRow)(uint8* dst_argb, const uint8* table_argb, + int width) = RGBColorTableRow_C; + uint8* dst = dst_argb + dst_y * dst_stride_argb + dst_x * 4; + if (!dst_argb || !table_argb || width <= 0 || height <= 0 || + dst_x < 0 || dst_y < 0) { + return -1; + } + // Coalesce rows. + if (dst_stride_argb == width * 4) { + width *= height; + height = 1; + dst_stride_argb = 0; + } +#if defined(HAS_RGBCOLORTABLEROW_X86) + if (TestCpuFlag(kCpuHasX86)) { + RGBColorTableRow = RGBColorTableRow_X86; + } +#endif + for (y = 0; y < height; ++y) { + RGBColorTableRow(dst, table_argb, width); + dst += dst_stride_argb; + } + return 0; +} + +// ARGBQuantize is used to posterize art. +// e.g. rgb / qvalue * qvalue + qvalue / 2 +// But the low levels implement efficiently with 3 parameters, and could be +// used for other high level operations. +// dst_argb[0] = (b * scale >> 16) * interval_size + interval_offset; +// where scale is 1 / interval_size as a fixed point value. +// The divide is replaces with a multiply by reciprocal fixed point multiply. +// Caveat - although SSE2 saturates, the C function does not and should be used +// with care if doing anything but quantization. +LIBYUV_API +int ARGBQuantize(uint8* dst_argb, int dst_stride_argb, + int scale, int interval_size, int interval_offset, + int dst_x, int dst_y, int width, int height) { + int y; + void (*ARGBQuantizeRow)(uint8* dst_argb, int scale, int interval_size, + int interval_offset, int width) = ARGBQuantizeRow_C; + uint8* dst = dst_argb + dst_y * dst_stride_argb + dst_x * 4; + if (!dst_argb || width <= 0 || height <= 0 || dst_x < 0 || dst_y < 0 || + interval_size < 1 || interval_size > 255) { + return -1; + } + // Coalesce rows. + if (dst_stride_argb == width * 4) { + width *= height; + height = 1; + dst_stride_argb = 0; + } +#if defined(HAS_ARGBQUANTIZEROW_SSE2) + if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 4)) { + ARGBQuantizeRow = ARGBQuantizeRow_SSE2; + } +#endif +#if defined(HAS_ARGBQUANTIZEROW_NEON) + if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 8)) { + ARGBQuantizeRow = ARGBQuantizeRow_NEON; + } +#endif + for (y = 0; y < height; ++y) { + ARGBQuantizeRow(dst, scale, interval_size, interval_offset, width); + dst += dst_stride_argb; + } + return 0; +} + +// Computes table of cumulative sum for image where the value is the sum +// of all values above and to the left of the entry. Used by ARGBBlur. +LIBYUV_API +int ARGBComputeCumulativeSum(const uint8* src_argb, int src_stride_argb, + int32* dst_cumsum, int dst_stride32_cumsum, + int width, int height) { + int y; + void (*ComputeCumulativeSumRow)(const uint8* row, int32* cumsum, + const int32* previous_cumsum, int width) = ComputeCumulativeSumRow_C; + int32* previous_cumsum = dst_cumsum; + if (!dst_cumsum || !src_argb || width <= 0 || height <= 0) { + return -1; + } +#if defined(HAS_CUMULATIVESUMTOAVERAGEROW_SSE2) + if (TestCpuFlag(kCpuHasSSE2)) { + ComputeCumulativeSumRow = ComputeCumulativeSumRow_SSE2; + } +#endif + memset(dst_cumsum, 0, width * sizeof(dst_cumsum[0]) * 4); // 4 int per pixel. + for (y = 0; y < height; ++y) { + ComputeCumulativeSumRow(src_argb, dst_cumsum, previous_cumsum, width); + previous_cumsum = dst_cumsum; + dst_cumsum += dst_stride32_cumsum; + src_argb += src_stride_argb; + } + return 0; +} + +// Blur ARGB image. +// Caller should allocate CumulativeSum table of width * height * 16 bytes +// aligned to 16 byte boundary. height can be radius * 2 + 2 to save memory +// as the buffer is treated as circular. +LIBYUV_API +int ARGBBlur(const uint8* src_argb, int src_stride_argb, + uint8* dst_argb, int dst_stride_argb, + int32* dst_cumsum, int dst_stride32_cumsum, + int width, int height, int radius) { + int y; + void (*ComputeCumulativeSumRow)(const uint8 *row, int32 *cumsum, + const int32* previous_cumsum, int width) = ComputeCumulativeSumRow_C; + void (*CumulativeSumToAverageRow)(const int32* topleft, const int32* botleft, + int width, int area, uint8* dst, int count) = CumulativeSumToAverageRow_C; + int32* cumsum_bot_row; + int32* max_cumsum_bot_row; + int32* cumsum_top_row; + + if (!src_argb || !dst_argb || width <= 0 || height == 0) { + return -1; + } + if (height < 0) { + height = -height; + src_argb = src_argb + (height - 1) * src_stride_argb; + src_stride_argb = -src_stride_argb; + } + if (radius > height) { + radius = height; + } + if (radius > (width / 2 - 1)) { + radius = width / 2 - 1; + } + if (radius <= 0) { + return -1; + } +#if defined(HAS_CUMULATIVESUMTOAVERAGEROW_SSE2) + if (TestCpuFlag(kCpuHasSSE2)) { + ComputeCumulativeSumRow = ComputeCumulativeSumRow_SSE2; + CumulativeSumToAverageRow = CumulativeSumToAverageRow_SSE2; + } +#endif + // Compute enough CumulativeSum for first row to be blurred. After this + // one row of CumulativeSum is updated at a time. + ARGBComputeCumulativeSum(src_argb, src_stride_argb, + dst_cumsum, dst_stride32_cumsum, + width, radius); + + src_argb = src_argb + radius * src_stride_argb; + cumsum_bot_row = &dst_cumsum[(radius - 1) * dst_stride32_cumsum]; + + max_cumsum_bot_row = &dst_cumsum[(radius * 2 + 2) * dst_stride32_cumsum]; + cumsum_top_row = &dst_cumsum[0]; + + for (y = 0; y < height; ++y) { + int top_y = ((y - radius - 1) >= 0) ? (y - radius - 1) : 0; + int bot_y = ((y + radius) < height) ? (y + radius) : (height - 1); + int area = radius * (bot_y - top_y); + int boxwidth = radius * 4; + int x; + int n; + + // Increment cumsum_top_row pointer with circular buffer wrap around. + if (top_y) { + cumsum_top_row += dst_stride32_cumsum; + if (cumsum_top_row >= max_cumsum_bot_row) { + cumsum_top_row = dst_cumsum; + } + } + // Increment cumsum_bot_row pointer with circular buffer wrap around and + // then fill in a row of CumulativeSum. + if ((y + radius) < height) { + const int32* prev_cumsum_bot_row = cumsum_bot_row; + cumsum_bot_row += dst_stride32_cumsum; + if (cumsum_bot_row >= max_cumsum_bot_row) { + cumsum_bot_row = dst_cumsum; + } + ComputeCumulativeSumRow(src_argb, cumsum_bot_row, prev_cumsum_bot_row, + width); + src_argb += src_stride_argb; + } + + // Left clipped. + for (x = 0; x < radius + 1; ++x) { + CumulativeSumToAverageRow(cumsum_top_row, cumsum_bot_row, + boxwidth, area, &dst_argb[x * 4], 1); + area += (bot_y - top_y); + boxwidth += 4; + } + + // Middle unclipped. + n = (width - 1) - radius - x + 1; + CumulativeSumToAverageRow(cumsum_top_row, cumsum_bot_row, + boxwidth, area, &dst_argb[x * 4], n); + + // Right clipped. + for (x += n; x <= width - 1; ++x) { + area -= (bot_y - top_y); + boxwidth -= 4; + CumulativeSumToAverageRow(cumsum_top_row + (x - radius - 1) * 4, + cumsum_bot_row + (x - radius - 1) * 4, + boxwidth, area, &dst_argb[x * 4], 1); + } + dst_argb += dst_stride_argb; + } + return 0; +} + +// Multiply ARGB image by a specified ARGB value. +LIBYUV_API +int ARGBShade(const uint8* src_argb, int src_stride_argb, + uint8* dst_argb, int dst_stride_argb, + int width, int height, uint32 value) { + int y; + void (*ARGBShadeRow)(const uint8* src_argb, uint8* dst_argb, + int width, uint32 value) = ARGBShadeRow_C; + if (!src_argb || !dst_argb || width <= 0 || height == 0 || value == 0u) { + return -1; + } + if (height < 0) { + height = -height; + src_argb = src_argb + (height - 1) * src_stride_argb; + src_stride_argb = -src_stride_argb; + } + // Coalesce rows. + if (src_stride_argb == width * 4 && + dst_stride_argb == width * 4) { + width *= height; + height = 1; + src_stride_argb = dst_stride_argb = 0; + } +#if defined(HAS_ARGBSHADEROW_SSE2) + if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 4)) { + ARGBShadeRow = ARGBShadeRow_SSE2; + } +#endif +#if defined(HAS_ARGBSHADEROW_NEON) + if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 8)) { + ARGBShadeRow = ARGBShadeRow_NEON; + } +#endif + + for (y = 0; y < height; ++y) { + ARGBShadeRow(src_argb, dst_argb, width, value); + src_argb += src_stride_argb; + dst_argb += dst_stride_argb; + } + return 0; +} + +// Interpolate 2 ARGB images by specified amount (0 to 255). +LIBYUV_API +int ARGBInterpolate(const uint8* src_argb0, int src_stride_argb0, + const uint8* src_argb1, int src_stride_argb1, + uint8* dst_argb, int dst_stride_argb, + int width, int height, int interpolation) { + int y; + void (*InterpolateRow)(uint8* dst_ptr, const uint8* src_ptr, + ptrdiff_t src_stride, int dst_width, + int source_y_fraction) = InterpolateRow_C; + if (!src_argb0 || !src_argb1 || !dst_argb || width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + dst_argb = dst_argb + (height - 1) * dst_stride_argb; + dst_stride_argb = -dst_stride_argb; + } + // Coalesce rows. + if (src_stride_argb0 == width * 4 && + src_stride_argb1 == width * 4 && + dst_stride_argb == width * 4) { + width *= height; + height = 1; + src_stride_argb0 = src_stride_argb1 = dst_stride_argb = 0; + } +#if defined(HAS_INTERPOLATEROW_SSE2) + if (TestCpuFlag(kCpuHasSSE2)) { + InterpolateRow = InterpolateRow_Any_SSE2; + if (IS_ALIGNED(width, 4)) { + InterpolateRow = InterpolateRow_SSE2; + } + } +#endif +#if defined(HAS_INTERPOLATEROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + InterpolateRow = InterpolateRow_Any_SSSE3; + if (IS_ALIGNED(width, 4)) { + InterpolateRow = InterpolateRow_SSSE3; + } + } +#endif +#if defined(HAS_INTERPOLATEROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + InterpolateRow = InterpolateRow_Any_AVX2; + if (IS_ALIGNED(width, 8)) { + InterpolateRow = InterpolateRow_AVX2; + } + } +#endif +#if defined(HAS_INTERPOLATEROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + InterpolateRow = InterpolateRow_Any_NEON; + if (IS_ALIGNED(width, 4)) { + InterpolateRow = InterpolateRow_NEON; + } + } +#endif +#if defined(HAS_INTERPOLATEROW_MIPS_DSPR2) + if (TestCpuFlag(kCpuHasMIPS_DSPR2) && + IS_ALIGNED(src_argb0, 4) && IS_ALIGNED(src_stride_argb0, 4) && + IS_ALIGNED(src_argb1, 4) && IS_ALIGNED(src_stride_argb1, 4) && + IS_ALIGNED(dst_argb, 4) && IS_ALIGNED(dst_stride_argb, 4)) { + InterpolateRow = InterpolateRow_MIPS_DSPR2; + } +#endif + + for (y = 0; y < height; ++y) { + InterpolateRow(dst_argb, src_argb0, src_argb1 - src_argb0, + width * 4, interpolation); + src_argb0 += src_stride_argb0; + src_argb1 += src_stride_argb1; + dst_argb += dst_stride_argb; + } + return 0; +} + +// Shuffle ARGB channel order. e.g. BGRA to ARGB. +LIBYUV_API +int ARGBShuffle(const uint8* src_bgra, int src_stride_bgra, + uint8* dst_argb, int dst_stride_argb, + const uint8* shuffler, int width, int height) { + int y; + void (*ARGBShuffleRow)(const uint8* src_bgra, uint8* dst_argb, + const uint8* shuffler, int pix) = ARGBShuffleRow_C; + if (!src_bgra || !dst_argb || + width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + src_bgra = src_bgra + (height - 1) * src_stride_bgra; + src_stride_bgra = -src_stride_bgra; + } + // Coalesce rows. + if (src_stride_bgra == width * 4 && + dst_stride_argb == width * 4) { + width *= height; + height = 1; + src_stride_bgra = dst_stride_argb = 0; + } +#if defined(HAS_ARGBSHUFFLEROW_SSE2) + if (TestCpuFlag(kCpuHasSSE2)) { + ARGBShuffleRow = ARGBShuffleRow_Any_SSE2; + if (IS_ALIGNED(width, 4)) { + ARGBShuffleRow = ARGBShuffleRow_SSE2; + } + } +#endif +#if defined(HAS_ARGBSHUFFLEROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + ARGBShuffleRow = ARGBShuffleRow_Any_SSSE3; + if (IS_ALIGNED(width, 8)) { + ARGBShuffleRow = ARGBShuffleRow_SSSE3; + } + } +#endif +#if defined(HAS_ARGBSHUFFLEROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + ARGBShuffleRow = ARGBShuffleRow_Any_AVX2; + if (IS_ALIGNED(width, 16)) { + ARGBShuffleRow = ARGBShuffleRow_AVX2; + } + } +#endif +#if defined(HAS_ARGBSHUFFLEROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + ARGBShuffleRow = ARGBShuffleRow_Any_NEON; + if (IS_ALIGNED(width, 4)) { + ARGBShuffleRow = ARGBShuffleRow_NEON; + } + } +#endif + + for (y = 0; y < height; ++y) { + ARGBShuffleRow(src_bgra, dst_argb, shuffler, width); + src_bgra += src_stride_bgra; + dst_argb += dst_stride_argb; + } + return 0; +} + +// Sobel ARGB effect. +static int ARGBSobelize(const uint8* src_argb, int src_stride_argb, + uint8* dst_argb, int dst_stride_argb, + int width, int height, + void (*SobelRow)(const uint8* src_sobelx, + const uint8* src_sobely, + uint8* dst, int width)) { + int y; + void (*ARGBToYJRow)(const uint8* src_argb, uint8* dst_g, int pix) = + ARGBToYJRow_C; + void (*SobelYRow)(const uint8* src_y0, const uint8* src_y1, + uint8* dst_sobely, int width) = SobelYRow_C; + void (*SobelXRow)(const uint8* src_y0, const uint8* src_y1, + const uint8* src_y2, uint8* dst_sobely, int width) = + SobelXRow_C; + const int kEdge = 16; // Extra pixels at start of row for extrude/align. + if (!src_argb || !dst_argb || width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + src_argb = src_argb + (height - 1) * src_stride_argb; + src_stride_argb = -src_stride_argb; + } + +#if defined(HAS_ARGBTOYJROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + ARGBToYJRow = ARGBToYJRow_Any_SSSE3; + if (IS_ALIGNED(width, 16)) { + ARGBToYJRow = ARGBToYJRow_SSSE3; + } + } +#endif +#if defined(HAS_ARGBTOYJROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + ARGBToYJRow = ARGBToYJRow_Any_AVX2; + if (IS_ALIGNED(width, 32)) { + ARGBToYJRow = ARGBToYJRow_AVX2; + } + } +#endif +#if defined(HAS_ARGBTOYJROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + ARGBToYJRow = ARGBToYJRow_Any_NEON; + if (IS_ALIGNED(width, 8)) { + ARGBToYJRow = ARGBToYJRow_NEON; + } + } +#endif + +#if defined(HAS_SOBELYROW_SSE2) + if (TestCpuFlag(kCpuHasSSE2)) { + SobelYRow = SobelYRow_SSE2; + } +#endif +#if defined(HAS_SOBELYROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + SobelYRow = SobelYRow_NEON; + } +#endif +#if defined(HAS_SOBELXROW_SSE2) + if (TestCpuFlag(kCpuHasSSE2)) { + SobelXRow = SobelXRow_SSE2; + } +#endif +#if defined(HAS_SOBELXROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + SobelXRow = SobelXRow_NEON; + } +#endif + { + // 3 rows with edges before/after. + const int kRowSize = (width + kEdge + 31) & ~31; + align_buffer_64(rows, kRowSize * 2 + (kEdge + kRowSize * 3 + kEdge)); + uint8* row_sobelx = rows; + uint8* row_sobely = rows + kRowSize; + uint8* row_y = rows + kRowSize * 2; + + // Convert first row. + uint8* row_y0 = row_y + kEdge; + uint8* row_y1 = row_y0 + kRowSize; + uint8* row_y2 = row_y1 + kRowSize; + ARGBToYJRow(src_argb, row_y0, width); + row_y0[-1] = row_y0[0]; + memset(row_y0 + width, row_y0[width - 1], 16); // Extrude 16 for valgrind. + ARGBToYJRow(src_argb, row_y1, width); + row_y1[-1] = row_y1[0]; + memset(row_y1 + width, row_y1[width - 1], 16); + memset(row_y2 + width, 0, 16); + + for (y = 0; y < height; ++y) { + // Convert next row of ARGB to G. + if (y < (height - 1)) { + src_argb += src_stride_argb; + } + ARGBToYJRow(src_argb, row_y2, width); + row_y2[-1] = row_y2[0]; + row_y2[width] = row_y2[width - 1]; + + SobelXRow(row_y0 - 1, row_y1 - 1, row_y2 - 1, row_sobelx, width); + SobelYRow(row_y0 - 1, row_y2 - 1, row_sobely, width); + SobelRow(row_sobelx, row_sobely, dst_argb, width); + + // Cycle thru circular queue of 3 row_y buffers. + { + uint8* row_yt = row_y0; + row_y0 = row_y1; + row_y1 = row_y2; + row_y2 = row_yt; + } + + dst_argb += dst_stride_argb; + } + free_aligned_buffer_64(rows); + } + return 0; +} + +// Sobel ARGB effect. +LIBYUV_API +int ARGBSobel(const uint8* src_argb, int src_stride_argb, + uint8* dst_argb, int dst_stride_argb, + int width, int height) { + void (*SobelRow)(const uint8* src_sobelx, const uint8* src_sobely, + uint8* dst_argb, int width) = SobelRow_C; +#if defined(HAS_SOBELROW_SSE2) + if (TestCpuFlag(kCpuHasSSE2)) { + SobelRow = SobelRow_Any_SSE2; + if (IS_ALIGNED(width, 16)) { + SobelRow = SobelRow_SSE2; + } + } +#endif +#if defined(HAS_SOBELROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + SobelRow = SobelRow_Any_NEON; + if (IS_ALIGNED(width, 8)) { + SobelRow = SobelRow_NEON; + } + } +#endif + return ARGBSobelize(src_argb, src_stride_argb, dst_argb, dst_stride_argb, + width, height, SobelRow); +} + +// Sobel ARGB effect with planar output. +LIBYUV_API +int ARGBSobelToPlane(const uint8* src_argb, int src_stride_argb, + uint8* dst_y, int dst_stride_y, + int width, int height) { + void (*SobelToPlaneRow)(const uint8* src_sobelx, const uint8* src_sobely, + uint8* dst_, int width) = SobelToPlaneRow_C; +#if defined(HAS_SOBELTOPLANEROW_SSE2) + if (TestCpuFlag(kCpuHasSSE2)) { + SobelToPlaneRow = SobelToPlaneRow_Any_SSE2; + if (IS_ALIGNED(width, 16)) { + SobelToPlaneRow = SobelToPlaneRow_SSE2; + } + } +#endif +#if defined(HAS_SOBELTOPLANEROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + SobelToPlaneRow = SobelToPlaneRow_Any_NEON; + if (IS_ALIGNED(width, 16)) { + SobelToPlaneRow = SobelToPlaneRow_NEON; + } + } +#endif + return ARGBSobelize(src_argb, src_stride_argb, dst_y, dst_stride_y, + width, height, SobelToPlaneRow); +} + +// SobelXY ARGB effect. +// Similar to Sobel, but also stores Sobel X in R and Sobel Y in B. G = Sobel. +LIBYUV_API +int ARGBSobelXY(const uint8* src_argb, int src_stride_argb, + uint8* dst_argb, int dst_stride_argb, + int width, int height) { + void (*SobelXYRow)(const uint8* src_sobelx, const uint8* src_sobely, + uint8* dst_argb, int width) = SobelXYRow_C; +#if defined(HAS_SOBELXYROW_SSE2) + if (TestCpuFlag(kCpuHasSSE2)) { + SobelXYRow = SobelXYRow_Any_SSE2; + if (IS_ALIGNED(width, 16)) { + SobelXYRow = SobelXYRow_SSE2; + } + } +#endif +#if defined(HAS_SOBELXYROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + SobelXYRow = SobelXYRow_Any_NEON; + if (IS_ALIGNED(width, 8)) { + SobelXYRow = SobelXYRow_NEON; + } + } +#endif + return ARGBSobelize(src_argb, src_stride_argb, dst_argb, dst_stride_argb, + width, height, SobelXYRow); +} + +// Apply a 4x4 polynomial to each ARGB pixel. +LIBYUV_API +int ARGBPolynomial(const uint8* src_argb, int src_stride_argb, + uint8* dst_argb, int dst_stride_argb, + const float* poly, + int width, int height) { + int y; + void (*ARGBPolynomialRow)(const uint8* src_argb, + uint8* dst_argb, const float* poly, + int width) = ARGBPolynomialRow_C; + if (!src_argb || !dst_argb || !poly || width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + src_argb = src_argb + (height - 1) * src_stride_argb; + src_stride_argb = -src_stride_argb; + } + // Coalesce rows. + if (src_stride_argb == width * 4 && + dst_stride_argb == width * 4) { + width *= height; + height = 1; + src_stride_argb = dst_stride_argb = 0; + } +#if defined(HAS_ARGBPOLYNOMIALROW_SSE2) + if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 2)) { + ARGBPolynomialRow = ARGBPolynomialRow_SSE2; + } +#endif +#if defined(HAS_ARGBPOLYNOMIALROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2) && TestCpuFlag(kCpuHasFMA3) && + IS_ALIGNED(width, 2)) { + ARGBPolynomialRow = ARGBPolynomialRow_AVX2; + } +#endif + + for (y = 0; y < height; ++y) { + ARGBPolynomialRow(src_argb, dst_argb, poly, width); + src_argb += src_stride_argb; + dst_argb += dst_stride_argb; + } + return 0; +} + +// Apply a lumacolortable to each ARGB pixel. +LIBYUV_API +int ARGBLumaColorTable(const uint8* src_argb, int src_stride_argb, + uint8* dst_argb, int dst_stride_argb, + const uint8* luma, + int width, int height) { + int y; + void (*ARGBLumaColorTableRow)(const uint8* src_argb, uint8* dst_argb, + int width, const uint8* luma, const uint32 lumacoeff) = + ARGBLumaColorTableRow_C; + if (!src_argb || !dst_argb || !luma || width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + src_argb = src_argb + (height - 1) * src_stride_argb; + src_stride_argb = -src_stride_argb; + } + // Coalesce rows. + if (src_stride_argb == width * 4 && + dst_stride_argb == width * 4) { + width *= height; + height = 1; + src_stride_argb = dst_stride_argb = 0; + } +#if defined(HAS_ARGBLUMACOLORTABLEROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3) && IS_ALIGNED(width, 4)) { + ARGBLumaColorTableRow = ARGBLumaColorTableRow_SSSE3; + } +#endif + + for (y = 0; y < height; ++y) { + ARGBLumaColorTableRow(src_argb, dst_argb, width, luma, 0x00264b0f); + src_argb += src_stride_argb; + dst_argb += dst_stride_argb; + } + return 0; +} + +// Copy Alpha from one ARGB image to another. +LIBYUV_API +int ARGBCopyAlpha(const uint8* src_argb, int src_stride_argb, + uint8* dst_argb, int dst_stride_argb, + int width, int height) { + int y; + void (*ARGBCopyAlphaRow)(const uint8* src_argb, uint8* dst_argb, int width) = + ARGBCopyAlphaRow_C; + if (!src_argb || !dst_argb || width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + src_argb = src_argb + (height - 1) * src_stride_argb; + src_stride_argb = -src_stride_argb; + } + // Coalesce rows. + if (src_stride_argb == width * 4 && + dst_stride_argb == width * 4) { + width *= height; + height = 1; + src_stride_argb = dst_stride_argb = 0; + } +#if defined(HAS_ARGBCOPYALPHAROW_SSE2) + if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 8)) { + ARGBCopyAlphaRow = ARGBCopyAlphaRow_SSE2; + } +#endif +#if defined(HAS_ARGBCOPYALPHAROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2) && IS_ALIGNED(width, 16)) { + ARGBCopyAlphaRow = ARGBCopyAlphaRow_AVX2; + } +#endif + + for (y = 0; y < height; ++y) { + ARGBCopyAlphaRow(src_argb, dst_argb, width); + src_argb += src_stride_argb; + dst_argb += dst_stride_argb; + } + return 0; +} + +// Copy a planar Y channel to the alpha channel of a destination ARGB image. +LIBYUV_API +int ARGBCopyYToAlpha(const uint8* src_y, int src_stride_y, + uint8* dst_argb, int dst_stride_argb, + int width, int height) { + int y; + void (*ARGBCopyYToAlphaRow)(const uint8* src_y, uint8* dst_argb, int width) = + ARGBCopyYToAlphaRow_C; + if (!src_y || !dst_argb || width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + src_y = src_y + (height - 1) * src_stride_y; + src_stride_y = -src_stride_y; + } + // Coalesce rows. + if (src_stride_y == width && + dst_stride_argb == width * 4) { + width *= height; + height = 1; + src_stride_y = dst_stride_argb = 0; + } +#if defined(HAS_ARGBCOPYYTOALPHAROW_SSE2) + if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 8)) { + ARGBCopyYToAlphaRow = ARGBCopyYToAlphaRow_SSE2; + } +#endif +#if defined(HAS_ARGBCOPYYTOALPHAROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2) && IS_ALIGNED(width, 16)) { + ARGBCopyYToAlphaRow = ARGBCopyYToAlphaRow_AVX2; + } +#endif + + for (y = 0; y < height; ++y) { + ARGBCopyYToAlphaRow(src_y, dst_argb, width); + src_y += src_stride_y; + dst_argb += dst_stride_argb; + } + return 0; +} + +LIBYUV_API +int YUY2ToNV12(const uint8* src_yuy2, int src_stride_yuy2, + uint8* dst_y, int dst_stride_y, + uint8* dst_uv, int dst_stride_uv, + int width, int height) { + int y; + int halfwidth = (width + 1) >> 1; + void (*SplitUVRow)(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int pix) = + SplitUVRow_C; + void (*InterpolateRow)(uint8* dst_ptr, const uint8* src_ptr, + ptrdiff_t src_stride, int dst_width, + int source_y_fraction) = InterpolateRow_C; + if (!src_yuy2 || + !dst_y || !dst_uv || + width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + src_yuy2 = src_yuy2 + (height - 1) * src_stride_yuy2; + src_stride_yuy2 = -src_stride_yuy2; + } +#if defined(HAS_SPLITUVROW_SSE2) + if (TestCpuFlag(kCpuHasSSE2)) { + SplitUVRow = SplitUVRow_Any_SSE2; + if (IS_ALIGNED(width, 16)) { + SplitUVRow = SplitUVRow_SSE2; + } + } +#endif +#if defined(HAS_SPLITUVROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + SplitUVRow = SplitUVRow_Any_AVX2; + if (IS_ALIGNED(width, 32)) { + SplitUVRow = SplitUVRow_AVX2; + } + } +#endif +#if defined(HAS_SPLITUVROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + SplitUVRow = SplitUVRow_Any_NEON; + if (IS_ALIGNED(width, 16)) { + SplitUVRow = SplitUVRow_NEON; + } + } +#endif +#if defined(HAS_INTERPOLATEROW_SSE2) + if (TestCpuFlag(kCpuHasSSE2)) { + InterpolateRow = InterpolateRow_Any_SSE2; + if (IS_ALIGNED(width, 16)) { + InterpolateRow = InterpolateRow_SSE2; + } + } +#endif +#if defined(HAS_INTERPOLATEROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + InterpolateRow = InterpolateRow_Any_SSSE3; + if (IS_ALIGNED(width, 16)) { + InterpolateRow = InterpolateRow_SSSE3; + } + } +#endif +#if defined(HAS_INTERPOLATEROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + InterpolateRow = InterpolateRow_Any_AVX2; + if (IS_ALIGNED(width, 32)) { + InterpolateRow = InterpolateRow_AVX2; + } + } +#endif +#if defined(HAS_INTERPOLATEROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + InterpolateRow = InterpolateRow_Any_NEON; + if (IS_ALIGNED(width, 16)) { + InterpolateRow = InterpolateRow_NEON; + } + } +#endif + + { + int awidth = halfwidth * 2; + // 2 rows of uv + align_buffer_64(rows, awidth * 2); + + for (y = 0; y < height - 1; y += 2) { + // Split Y from UV. + SplitUVRow(src_yuy2, dst_y, rows, awidth); + SplitUVRow(src_yuy2 + src_stride_yuy2, dst_y + dst_stride_y, + rows + awidth, awidth); + InterpolateRow(dst_uv, rows, awidth, awidth, 128); + src_yuy2 += src_stride_yuy2 * 2; + dst_y += dst_stride_y * 2; + dst_uv += dst_stride_uv; + } + if (height & 1) { + // Split Y from UV. + SplitUVRow(src_yuy2, dst_y, dst_uv, width); + } + free_aligned_buffer_64(rows); + } + return 0; +} + +LIBYUV_API +int UYVYToNV12(const uint8* src_uyvy, int src_stride_uyvy, + uint8* dst_y, int dst_stride_y, + uint8* dst_uv, int dst_stride_uv, + int width, int height) { + int y; + int halfwidth = (width + 1) >> 1; + void (*SplitUVRow)(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int pix) = + SplitUVRow_C; + void (*InterpolateRow)(uint8* dst_ptr, const uint8* src_ptr, + ptrdiff_t src_stride, int dst_width, + int source_y_fraction) = InterpolateRow_C; + if (!src_uyvy || + !dst_y || !dst_uv || + width <= 0 || height == 0) { + return -1; + } + // Negative height means invert the image. + if (height < 0) { + height = -height; + src_uyvy = src_uyvy + (height - 1) * src_stride_uyvy; + src_stride_uyvy = -src_stride_uyvy; + } +#if defined(HAS_SPLITUVROW_SSE2) + if (TestCpuFlag(kCpuHasSSE2)) { + SplitUVRow = SplitUVRow_Any_SSE2; + if (IS_ALIGNED(width, 16)) { + SplitUVRow = SplitUVRow_SSE2; + } + } +#endif +#if defined(HAS_SPLITUVROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + SplitUVRow = SplitUVRow_Any_AVX2; + if (IS_ALIGNED(width, 32)) { + SplitUVRow = SplitUVRow_AVX2; + } + } +#endif +#if defined(HAS_SPLITUVROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + SplitUVRow = SplitUVRow_Any_NEON; + if (IS_ALIGNED(width, 16)) { + SplitUVRow = SplitUVRow_NEON; + } + } +#endif +#if defined(HAS_INTERPOLATEROW_SSE2) + if (TestCpuFlag(kCpuHasSSE2)) { + InterpolateRow = InterpolateRow_Any_SSE2; + if (IS_ALIGNED(width, 16)) { + InterpolateRow = InterpolateRow_SSE2; + } + } +#endif +#if defined(HAS_INTERPOLATEROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + InterpolateRow = InterpolateRow_Any_SSSE3; + if (IS_ALIGNED(width, 16)) { + InterpolateRow = InterpolateRow_SSSE3; + } + } +#endif +#if defined(HAS_INTERPOLATEROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + InterpolateRow = InterpolateRow_Any_AVX2; + if (IS_ALIGNED(width, 32)) { + InterpolateRow = InterpolateRow_AVX2; + } + } +#endif +#if defined(HAS_INTERPOLATEROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + InterpolateRow = InterpolateRow_Any_NEON; + if (IS_ALIGNED(width, 16)) { + InterpolateRow = InterpolateRow_NEON; + } + } +#endif + + { + int awidth = halfwidth * 2; + // 2 rows of uv + align_buffer_64(rows, awidth * 2); + + for (y = 0; y < height - 1; y += 2) { + // Split Y from UV. + SplitUVRow(src_uyvy, rows, dst_y, awidth); + SplitUVRow(src_uyvy + src_stride_uyvy, rows + awidth, + dst_y + dst_stride_y, awidth); + InterpolateRow(dst_uv, rows, awidth, awidth, 128); + src_uyvy += src_stride_uyvy * 2; + dst_y += dst_stride_y * 2; + dst_uv += dst_stride_uv; + } + if (height & 1) { + // Split Y from UV. + SplitUVRow(src_uyvy, dst_y, dst_uv, width); + } + free_aligned_buffer_64(rows); + } + return 0; +} + +#ifdef __cplusplus +} // extern "C" +} // namespace libyuv +#endif diff --git a/media/libaom/src/third_party/libyuv/source/rotate.cc b/media/libaom/src/third_party/libyuv/source/rotate.cc new file mode 100644 index 000000000..be3d58920 --- /dev/null +++ b/media/libaom/src/third_party/libyuv/source/rotate.cc @@ -0,0 +1,496 @@ +/* + * Copyright 2011 The LibYuv Project Authors. All rights reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#include "libyuv/rotate.h" + +#include "libyuv/cpu_id.h" +#include "libyuv/convert.h" +#include "libyuv/planar_functions.h" +#include "libyuv/rotate_row.h" +#include "libyuv/row.h" + +#ifdef __cplusplus +namespace libyuv { +extern "C" { +#endif + +LIBYUV_API +void TransposePlane(const uint8* src, int src_stride, + uint8* dst, int dst_stride, + int width, int height) { + int i = height; + void (*TransposeWx8)(const uint8* src, int src_stride, + uint8* dst, int dst_stride, int width) = TransposeWx8_C; +#if defined(HAS_TRANSPOSEWX8_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + TransposeWx8 = TransposeWx8_NEON; + } +#endif +#if defined(HAS_TRANSPOSEWX8_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + TransposeWx8 = TransposeWx8_Any_SSSE3; + if (IS_ALIGNED(width, 8)) { + TransposeWx8 = TransposeWx8_SSSE3; + } + } +#endif +#if defined(HAS_TRANSPOSEWX8_FAST_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + TransposeWx8 = TransposeWx8_Fast_Any_SSSE3; + if (IS_ALIGNED(width, 16)) { + TransposeWx8 = TransposeWx8_Fast_SSSE3; + } + } +#endif +#if defined(HAS_TRANSPOSEWX8_MIPS_DSPR2) + if (TestCpuFlag(kCpuHasMIPS_DSPR2)) { + if (IS_ALIGNED(width, 4) && + IS_ALIGNED(src, 4) && IS_ALIGNED(src_stride, 4)) { + TransposeWx8 = TransposeWx8_Fast_MIPS_DSPR2; + } else { + TransposeWx8 = TransposeWx8_MIPS_DSPR2; + } + } +#endif + + // Work across the source in 8x8 tiles + while (i >= 8) { + TransposeWx8(src, src_stride, dst, dst_stride, width); + src += 8 * src_stride; // Go down 8 rows. + dst += 8; // Move over 8 columns. + i -= 8; + } + + if (i > 0) { + TransposeWxH_C(src, src_stride, dst, dst_stride, width, i); + } +} + +LIBYUV_API +void RotatePlane90(const uint8* src, int src_stride, + uint8* dst, int dst_stride, + int width, int height) { + // Rotate by 90 is a transpose with the source read + // from bottom to top. So set the source pointer to the end + // of the buffer and flip the sign of the source stride. + src += src_stride * (height - 1); + src_stride = -src_stride; + TransposePlane(src, src_stride, dst, dst_stride, width, height); +} + +LIBYUV_API +void RotatePlane270(const uint8* src, int src_stride, + uint8* dst, int dst_stride, + int width, int height) { + // Rotate by 270 is a transpose with the destination written + // from bottom to top. So set the destination pointer to the end + // of the buffer and flip the sign of the destination stride. + dst += dst_stride * (width - 1); + dst_stride = -dst_stride; + TransposePlane(src, src_stride, dst, dst_stride, width, height); +} + +LIBYUV_API +void RotatePlane180(const uint8* src, int src_stride, + uint8* dst, int dst_stride, + int width, int height) { + // Swap first and last row and mirror the content. Uses a temporary row. + align_buffer_64(row, width); + const uint8* src_bot = src + src_stride * (height - 1); + uint8* dst_bot = dst + dst_stride * (height - 1); + int half_height = (height + 1) >> 1; + int y; + void (*MirrorRow)(const uint8* src, uint8* dst, int width) = MirrorRow_C; + void (*CopyRow)(const uint8* src, uint8* dst, int width) = CopyRow_C; +#if defined(HAS_MIRRORROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + MirrorRow = MirrorRow_Any_NEON; + if (IS_ALIGNED(width, 16)) { + MirrorRow = MirrorRow_NEON; + } + } +#endif +#if defined(HAS_MIRRORROW_SSE2) + if (TestCpuFlag(kCpuHasSSE2)) { + MirrorRow = MirrorRow_Any_SSE2; + if (IS_ALIGNED(width, 16)) { + MirrorRow = MirrorRow_SSE2; + } + } +#endif +#if defined(HAS_MIRRORROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + MirrorRow = MirrorRow_Any_SSSE3; + if (IS_ALIGNED(width, 16)) { + MirrorRow = MirrorRow_SSSE3; + } + } +#endif +#if defined(HAS_MIRRORROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + MirrorRow = MirrorRow_Any_AVX2; + if (IS_ALIGNED(width, 32)) { + MirrorRow = MirrorRow_AVX2; + } + } +#endif +// TODO(fbarchard): Mirror on mips handle unaligned memory. +#if defined(HAS_MIRRORROW_MIPS_DSPR2) + if (TestCpuFlag(kCpuHasMIPS_DSPR2) && + IS_ALIGNED(src, 4) && IS_ALIGNED(src_stride, 4) && + IS_ALIGNED(dst, 4) && IS_ALIGNED(dst_stride, 4)) { + MirrorRow = MirrorRow_MIPS_DSPR2; + } +#endif +#if defined(HAS_COPYROW_SSE2) + if (TestCpuFlag(kCpuHasSSE2)) { + CopyRow = IS_ALIGNED(width, 32) ? CopyRow_SSE2 : CopyRow_Any_SSE2; + } +#endif +#if defined(HAS_COPYROW_AVX) + if (TestCpuFlag(kCpuHasAVX)) { + CopyRow = IS_ALIGNED(width, 64) ? CopyRow_AVX : CopyRow_Any_AVX; + } +#endif +#if defined(HAS_COPYROW_ERMS) + if (TestCpuFlag(kCpuHasERMS)) { + CopyRow = CopyRow_ERMS; + } +#endif +#if defined(HAS_COPYROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + CopyRow = IS_ALIGNED(width, 32) ? CopyRow_NEON : CopyRow_Any_NEON; + } +#endif +#if defined(HAS_COPYROW_MIPS) + if (TestCpuFlag(kCpuHasMIPS)) { + CopyRow = CopyRow_MIPS; + } +#endif + + // Odd height will harmlessly mirror the middle row twice. + for (y = 0; y < half_height; ++y) { + MirrorRow(src, row, width); // Mirror first row into a buffer + src += src_stride; + MirrorRow(src_bot, dst, width); // Mirror last row into first row + dst += dst_stride; + CopyRow(row, dst_bot, width); // Copy first mirrored row into last + src_bot -= src_stride; + dst_bot -= dst_stride; + } + free_aligned_buffer_64(row); +} + +LIBYUV_API +void TransposeUV(const uint8* src, int src_stride, + uint8* dst_a, int dst_stride_a, + uint8* dst_b, int dst_stride_b, + int width, int height) { + int i = height; + void (*TransposeUVWx8)(const uint8* src, int src_stride, + uint8* dst_a, int dst_stride_a, + uint8* dst_b, int dst_stride_b, + int width) = TransposeUVWx8_C; +#if defined(HAS_TRANSPOSEUVWX8_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + TransposeUVWx8 = TransposeUVWx8_NEON; + } +#endif +#if defined(HAS_TRANSPOSEUVWX8_SSE2) + if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 8)) { + TransposeUVWx8 = TransposeUVWx8_SSE2; + } +#endif +#if defined(HAS_TRANSPOSEUVWx8_MIPS_DSPR2) + if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(width, 2) && + IS_ALIGNED(src, 4) && IS_ALIGNED(src_stride, 4)) { + TransposeUVWx8 = TransposeUVWx8_MIPS_DSPR2; + } +#endif + + // Work through the source in 8x8 tiles. + while (i >= 8) { + TransposeUVWx8(src, src_stride, + dst_a, dst_stride_a, + dst_b, dst_stride_b, + width); + src += 8 * src_stride; // Go down 8 rows. + dst_a += 8; // Move over 8 columns. + dst_b += 8; // Move over 8 columns. + i -= 8; + } + + if (i > 0) { + TransposeUVWxH_C(src, src_stride, + dst_a, dst_stride_a, + dst_b, dst_stride_b, + width, i); + } +} + +LIBYUV_API +void RotateUV90(const uint8* src, int src_stride, + uint8* dst_a, int dst_stride_a, + uint8* dst_b, int dst_stride_b, + int width, int height) { + src += src_stride * (height - 1); + src_stride = -src_stride; + + TransposeUV(src, src_stride, + dst_a, dst_stride_a, + dst_b, dst_stride_b, + width, height); +} + +LIBYUV_API +void RotateUV270(const uint8* src, int src_stride, + uint8* dst_a, int dst_stride_a, + uint8* dst_b, int dst_stride_b, + int width, int height) { + dst_a += dst_stride_a * (width - 1); + dst_b += dst_stride_b * (width - 1); + dst_stride_a = -dst_stride_a; + dst_stride_b = -dst_stride_b; + + TransposeUV(src, src_stride, + dst_a, dst_stride_a, + dst_b, dst_stride_b, + width, height); +} + +// Rotate 180 is a horizontal and vertical flip. +LIBYUV_API +void RotateUV180(const uint8* src, int src_stride, + uint8* dst_a, int dst_stride_a, + uint8* dst_b, int dst_stride_b, + int width, int height) { + int i; + void (*MirrorRowUV)(const uint8* src, uint8* dst_u, uint8* dst_v, int width) = + MirrorUVRow_C; +#if defined(HAS_MIRRORUVROW_NEON) + if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 8)) { + MirrorRowUV = MirrorUVRow_NEON; + } +#endif +#if defined(HAS_MIRRORROW_UV_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3) && IS_ALIGNED(width, 16)) { + MirrorRowUV = MirrorUVRow_SSSE3; + } +#endif +#if defined(HAS_MIRRORUVROW_MIPS_DSPR2) + if (TestCpuFlag(kCpuHasMIPS_DSPR2) && + IS_ALIGNED(src, 4) && IS_ALIGNED(src_stride, 4)) { + MirrorRowUV = MirrorUVRow_MIPS_DSPR2; + } +#endif + + dst_a += dst_stride_a * (height - 1); + dst_b += dst_stride_b * (height - 1); + + for (i = 0; i < height; ++i) { + MirrorRowUV(src, dst_a, dst_b, width); + src += src_stride; + dst_a -= dst_stride_a; + dst_b -= dst_stride_b; + } +} + +LIBYUV_API +int RotatePlane(const uint8* src, int src_stride, + uint8* dst, int dst_stride, + int width, int height, + enum RotationMode mode) { + if (!src || width <= 0 || height == 0 || !dst) { + return -1; + } + + // Negative height means invert the image. + if (height < 0) { + height = -height; + src = src + (height - 1) * src_stride; + src_stride = -src_stride; + } + + switch (mode) { + case kRotate0: + // copy frame + CopyPlane(src, src_stride, + dst, dst_stride, + width, height); + return 0; + case kRotate90: + RotatePlane90(src, src_stride, + dst, dst_stride, + width, height); + return 0; + case kRotate270: + RotatePlane270(src, src_stride, + dst, dst_stride, + width, height); + return 0; + case kRotate180: + RotatePlane180(src, src_stride, + dst, dst_stride, + width, height); + return 0; + default: + break; + } + return -1; +} + +LIBYUV_API +int I420Rotate(const uint8* src_y, int src_stride_y, + const uint8* src_u, int src_stride_u, + const uint8* src_v, int src_stride_v, + uint8* dst_y, int dst_stride_y, + uint8* dst_u, int dst_stride_u, + uint8* dst_v, int dst_stride_v, + int width, int height, + enum RotationMode mode) { + int halfwidth = (width + 1) >> 1; + int halfheight = (height + 1) >> 1; + if (!src_y || !src_u || !src_v || width <= 0 || height == 0 || + !dst_y || !dst_u || !dst_v) { + return -1; + } + + // Negative height means invert the image. + if (height < 0) { + height = -height; + halfheight = (height + 1) >> 1; + src_y = src_y + (height - 1) * src_stride_y; + src_u = src_u + (halfheight - 1) * src_stride_u; + src_v = src_v + (halfheight - 1) * src_stride_v; + src_stride_y = -src_stride_y; + src_stride_u = -src_stride_u; + src_stride_v = -src_stride_v; + } + + switch (mode) { + case kRotate0: + // copy frame + return I420Copy(src_y, src_stride_y, + src_u, src_stride_u, + src_v, src_stride_v, + dst_y, dst_stride_y, + dst_u, dst_stride_u, + dst_v, dst_stride_v, + width, height); + case kRotate90: + RotatePlane90(src_y, src_stride_y, + dst_y, dst_stride_y, + width, height); + RotatePlane90(src_u, src_stride_u, + dst_u, dst_stride_u, + halfwidth, halfheight); + RotatePlane90(src_v, src_stride_v, + dst_v, dst_stride_v, + halfwidth, halfheight); + return 0; + case kRotate270: + RotatePlane270(src_y, src_stride_y, + dst_y, dst_stride_y, + width, height); + RotatePlane270(src_u, src_stride_u, + dst_u, dst_stride_u, + halfwidth, halfheight); + RotatePlane270(src_v, src_stride_v, + dst_v, dst_stride_v, + halfwidth, halfheight); + return 0; + case kRotate180: + RotatePlane180(src_y, src_stride_y, + dst_y, dst_stride_y, + width, height); + RotatePlane180(src_u, src_stride_u, + dst_u, dst_stride_u, + halfwidth, halfheight); + RotatePlane180(src_v, src_stride_v, + dst_v, dst_stride_v, + halfwidth, halfheight); + return 0; + default: + break; + } + return -1; +} + +LIBYUV_API +int NV12ToI420Rotate(const uint8* src_y, int src_stride_y, + const uint8* src_uv, int src_stride_uv, + uint8* dst_y, int dst_stride_y, + uint8* dst_u, int dst_stride_u, + uint8* dst_v, int dst_stride_v, + int width, int height, + enum RotationMode mode) { + int halfwidth = (width + 1) >> 1; + int halfheight = (height + 1) >> 1; + if (!src_y || !src_uv || width <= 0 || height == 0 || + !dst_y || !dst_u || !dst_v) { + return -1; + } + + // Negative height means invert the image. + if (height < 0) { + height = -height; + halfheight = (height + 1) >> 1; + src_y = src_y + (height - 1) * src_stride_y; + src_uv = src_uv + (halfheight - 1) * src_stride_uv; + src_stride_y = -src_stride_y; + src_stride_uv = -src_stride_uv; + } + + switch (mode) { + case kRotate0: + // copy frame + return NV12ToI420(src_y, src_stride_y, + src_uv, src_stride_uv, + dst_y, dst_stride_y, + dst_u, dst_stride_u, + dst_v, dst_stride_v, + width, height); + case kRotate90: + RotatePlane90(src_y, src_stride_y, + dst_y, dst_stride_y, + width, height); + RotateUV90(src_uv, src_stride_uv, + dst_u, dst_stride_u, + dst_v, dst_stride_v, + halfwidth, halfheight); + return 0; + case kRotate270: + RotatePlane270(src_y, src_stride_y, + dst_y, dst_stride_y, + width, height); + RotateUV270(src_uv, src_stride_uv, + dst_u, dst_stride_u, + dst_v, dst_stride_v, + halfwidth, halfheight); + return 0; + case kRotate180: + RotatePlane180(src_y, src_stride_y, + dst_y, dst_stride_y, + width, height); + RotateUV180(src_uv, src_stride_uv, + dst_u, dst_stride_u, + dst_v, dst_stride_v, + halfwidth, halfheight); + return 0; + default: + break; + } + return -1; +} + +#ifdef __cplusplus +} // extern "C" +} // namespace libyuv +#endif diff --git a/media/libaom/src/third_party/libyuv/source/rotate_any.cc b/media/libaom/src/third_party/libyuv/source/rotate_any.cc new file mode 100644 index 000000000..4d6eb34e1 --- /dev/null +++ b/media/libaom/src/third_party/libyuv/source/rotate_any.cc @@ -0,0 +1,55 @@ +/* + * Copyright 2015 The LibYuv Project Authors. All rights reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#include "libyuv/rotate.h" +#include "libyuv/rotate_row.h" + +#include "libyuv/basic_types.h" + +#ifdef __cplusplus +namespace libyuv { +extern "C" { +#endif + +#define TANY(NAMEANY, TPOS_SIMD, TPOS_C, MASK) \ + void NAMEANY(const uint8* src, int src_stride, \ + uint8* dst, int dst_stride, int width) { \ + int r = width & MASK; \ + int n = width - r; \ + if (n > 0) { \ + TPOS_SIMD(src, src_stride, dst, dst_stride, n); \ + } \ + TPOS_C(src + n, src_stride, dst + n * dst_stride, dst_stride, r); \ + } + +#ifdef HAS_TRANSPOSEWX8_NEON +TANY(TransposeWx8_Any_NEON, TransposeWx8_NEON, TransposeWx8_C, 7) +#endif +#ifdef HAS_TRANSPOSEWX8_SSSE3 +TANY(TransposeWx8_Any_SSSE3, TransposeWx8_SSSE3, TransposeWx8_C, 7) +#endif +#ifdef HAS_TRANSPOSEWX8_FAST_SSSE3 +TANY(TransposeWx8_Fast_Any_SSSE3, TransposeWx8_Fast_SSSE3, TransposeWx8_C, 15) +#endif +#ifdef HAS_TRANSPOSEWX8_MIPS_DSPR2 +TANY(TransposeWx8_Any_MIPS_DSPR2, TransposeWx8_MIPS_DSPR2, TransposeWx8_C, 7) +#endif + +#undef TANY + +#ifdef __cplusplus +} // extern "C" +} // namespace libyuv +#endif + + + + + diff --git a/media/libaom/src/third_party/libyuv/source/rotate_argb.cc b/media/libaom/src/third_party/libyuv/source/rotate_argb.cc new file mode 100644 index 000000000..787c0ad1b --- /dev/null +++ b/media/libaom/src/third_party/libyuv/source/rotate_argb.cc @@ -0,0 +1,205 @@ +/* + * Copyright 2012 The LibYuv Project Authors. All rights reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#include "libyuv/rotate.h" + +#include "libyuv/cpu_id.h" +#include "libyuv/convert.h" +#include "libyuv/planar_functions.h" +#include "libyuv/row.h" + +#ifdef __cplusplus +namespace libyuv { +extern "C" { +#endif + +// ARGBScale has a function to copy pixels to a row, striding each source +// pixel by a constant. +#if !defined(LIBYUV_DISABLE_X86) && \ + (defined(_M_IX86) || \ + (defined(__x86_64__) && !defined(__native_client__)) || defined(__i386__)) +#define HAS_SCALEARGBROWDOWNEVEN_SSE2 +void ScaleARGBRowDownEven_SSE2(const uint8* src_ptr, int src_stride, + int src_stepx, uint8* dst_ptr, int dst_width); +#endif +#if !defined(LIBYUV_DISABLE_NEON) && !defined(__native_client__) && \ + (defined(__ARM_NEON__) || defined(LIBYUV_NEON) || defined(__aarch64__)) +#define HAS_SCALEARGBROWDOWNEVEN_NEON +void ScaleARGBRowDownEven_NEON(const uint8* src_ptr, int src_stride, + int src_stepx, uint8* dst_ptr, int dst_width); +#endif + +void ScaleARGBRowDownEven_C(const uint8* src_ptr, int, + int src_stepx, uint8* dst_ptr, int dst_width); + +static void ARGBTranspose(const uint8* src, int src_stride, + uint8* dst, int dst_stride, int width, int height) { + int i; + int src_pixel_step = src_stride >> 2; + void (*ScaleARGBRowDownEven)(const uint8* src_ptr, int src_stride, + int src_step, uint8* dst_ptr, int dst_width) = ScaleARGBRowDownEven_C; +#if defined(HAS_SCALEARGBROWDOWNEVEN_SSE2) + if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(height, 4)) { // Width of dest. + ScaleARGBRowDownEven = ScaleARGBRowDownEven_SSE2; + } +#endif +#if defined(HAS_SCALEARGBROWDOWNEVEN_NEON) + if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(height, 4)) { // Width of dest. + ScaleARGBRowDownEven = ScaleARGBRowDownEven_NEON; + } +#endif + + for (i = 0; i < width; ++i) { // column of source to row of dest. + ScaleARGBRowDownEven(src, 0, src_pixel_step, dst, height); + dst += dst_stride; + src += 4; + } +} + +void ARGBRotate90(const uint8* src, int src_stride, + uint8* dst, int dst_stride, int width, int height) { + // Rotate by 90 is a ARGBTranspose with the source read + // from bottom to top. So set the source pointer to the end + // of the buffer and flip the sign of the source stride. + src += src_stride * (height - 1); + src_stride = -src_stride; + ARGBTranspose(src, src_stride, dst, dst_stride, width, height); +} + +void ARGBRotate270(const uint8* src, int src_stride, + uint8* dst, int dst_stride, int width, int height) { + // Rotate by 270 is a ARGBTranspose with the destination written + // from bottom to top. So set the destination pointer to the end + // of the buffer and flip the sign of the destination stride. + dst += dst_stride * (width - 1); + dst_stride = -dst_stride; + ARGBTranspose(src, src_stride, dst, dst_stride, width, height); +} + +void ARGBRotate180(const uint8* src, int src_stride, + uint8* dst, int dst_stride, int width, int height) { + // Swap first and last row and mirror the content. Uses a temporary row. + align_buffer_64(row, width * 4); + const uint8* src_bot = src + src_stride * (height - 1); + uint8* dst_bot = dst + dst_stride * (height - 1); + int half_height = (height + 1) >> 1; + int y; + void (*ARGBMirrorRow)(const uint8* src, uint8* dst, int width) = + ARGBMirrorRow_C; + void (*CopyRow)(const uint8* src, uint8* dst, int width) = CopyRow_C; +#if defined(HAS_ARGBMIRRORROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + ARGBMirrorRow = ARGBMirrorRow_Any_NEON; + if (IS_ALIGNED(width, 4)) { + ARGBMirrorRow = ARGBMirrorRow_NEON; + } + } +#endif +#if defined(HAS_ARGBMIRRORROW_SSE2) + if (TestCpuFlag(kCpuHasSSE2)) { + ARGBMirrorRow = ARGBMirrorRow_Any_SSE2; + if (IS_ALIGNED(width, 4)) { + ARGBMirrorRow = ARGBMirrorRow_SSE2; + } + } +#endif +#if defined(HAS_ARGBMIRRORROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + ARGBMirrorRow = ARGBMirrorRow_Any_AVX2; + if (IS_ALIGNED(width, 8)) { + ARGBMirrorRow = ARGBMirrorRow_AVX2; + } + } +#endif +#if defined(HAS_COPYROW_SSE2) + if (TestCpuFlag(kCpuHasSSE2)) { + CopyRow = IS_ALIGNED(width * 4, 32) ? CopyRow_SSE2 : CopyRow_Any_SSE2; + } +#endif +#if defined(HAS_COPYROW_AVX) + if (TestCpuFlag(kCpuHasAVX)) { + CopyRow = IS_ALIGNED(width * 4, 64) ? CopyRow_AVX : CopyRow_Any_AVX; + } +#endif +#if defined(HAS_COPYROW_ERMS) + if (TestCpuFlag(kCpuHasERMS)) { + CopyRow = CopyRow_ERMS; + } +#endif +#if defined(HAS_COPYROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + CopyRow = IS_ALIGNED(width * 4, 32) ? CopyRow_NEON : CopyRow_Any_NEON; + } +#endif +#if defined(HAS_COPYROW_MIPS) + if (TestCpuFlag(kCpuHasMIPS)) { + CopyRow = CopyRow_MIPS; + } +#endif + + // Odd height will harmlessly mirror the middle row twice. + for (y = 0; y < half_height; ++y) { + ARGBMirrorRow(src, row, width); // Mirror first row into a buffer + ARGBMirrorRow(src_bot, dst, width); // Mirror last row into first row + CopyRow(row, dst_bot, width * 4); // Copy first mirrored row into last + src += src_stride; + dst += dst_stride; + src_bot -= src_stride; + dst_bot -= dst_stride; + } + free_aligned_buffer_64(row); +} + +LIBYUV_API +int ARGBRotate(const uint8* src_argb, int src_stride_argb, + uint8* dst_argb, int dst_stride_argb, int width, int height, + enum RotationMode mode) { + if (!src_argb || width <= 0 || height == 0 || !dst_argb) { + return -1; + } + + // Negative height means invert the image. + if (height < 0) { + height = -height; + src_argb = src_argb + (height - 1) * src_stride_argb; + src_stride_argb = -src_stride_argb; + } + + switch (mode) { + case kRotate0: + // copy frame + return ARGBCopy(src_argb, src_stride_argb, + dst_argb, dst_stride_argb, + width, height); + case kRotate90: + ARGBRotate90(src_argb, src_stride_argb, + dst_argb, dst_stride_argb, + width, height); + return 0; + case kRotate270: + ARGBRotate270(src_argb, src_stride_argb, + dst_argb, dst_stride_argb, + width, height); + return 0; + case kRotate180: + ARGBRotate180(src_argb, src_stride_argb, + dst_argb, dst_stride_argb, + width, height); + return 0; + default: + break; + } + return -1; +} + +#ifdef __cplusplus +} // extern "C" +} // namespace libyuv +#endif diff --git a/media/libaom/src/third_party/libyuv/source/rotate_common.cc b/media/libaom/src/third_party/libyuv/source/rotate_common.cc new file mode 100644 index 000000000..b33a9a0c6 --- /dev/null +++ b/media/libaom/src/third_party/libyuv/source/rotate_common.cc @@ -0,0 +1,92 @@ +/* + * Copyright 2011 The LibYuv Project Authors. All rights reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#include "libyuv/row.h" +#include "libyuv/rotate_row.h" + +#ifdef __cplusplus +namespace libyuv { +extern "C" { +#endif + +void TransposeWx8_C(const uint8* src, int src_stride, + uint8* dst, int dst_stride, int width) { + int i; + for (i = 0; i < width; ++i) { + dst[0] = src[0 * src_stride]; + dst[1] = src[1 * src_stride]; + dst[2] = src[2 * src_stride]; + dst[3] = src[3 * src_stride]; + dst[4] = src[4 * src_stride]; + dst[5] = src[5 * src_stride]; + dst[6] = src[6 * src_stride]; + dst[7] = src[7 * src_stride]; + ++src; + dst += dst_stride; + } +} + +void TransposeUVWx8_C(const uint8* src, int src_stride, + uint8* dst_a, int dst_stride_a, + uint8* dst_b, int dst_stride_b, int width) { + int i; + for (i = 0; i < width; ++i) { + dst_a[0] = src[0 * src_stride + 0]; + dst_b[0] = src[0 * src_stride + 1]; + dst_a[1] = src[1 * src_stride + 0]; + dst_b[1] = src[1 * src_stride + 1]; + dst_a[2] = src[2 * src_stride + 0]; + dst_b[2] = src[2 * src_stride + 1]; + dst_a[3] = src[3 * src_stride + 0]; + dst_b[3] = src[3 * src_stride + 1]; + dst_a[4] = src[4 * src_stride + 0]; + dst_b[4] = src[4 * src_stride + 1]; + dst_a[5] = src[5 * src_stride + 0]; + dst_b[5] = src[5 * src_stride + 1]; + dst_a[6] = src[6 * src_stride + 0]; + dst_b[6] = src[6 * src_stride + 1]; + dst_a[7] = src[7 * src_stride + 0]; + dst_b[7] = src[7 * src_stride + 1]; + src += 2; + dst_a += dst_stride_a; + dst_b += dst_stride_b; + } +} + +void TransposeWxH_C(const uint8* src, int src_stride, + uint8* dst, int dst_stride, + int width, int height) { + int i; + for (i = 0; i < width; ++i) { + int j; + for (j = 0; j < height; ++j) { + dst[i * dst_stride + j] = src[j * src_stride + i]; + } + } +} + +void TransposeUVWxH_C(const uint8* src, int src_stride, + uint8* dst_a, int dst_stride_a, + uint8* dst_b, int dst_stride_b, + int width, int height) { + int i; + for (i = 0; i < width * 2; i += 2) { + int j; + for (j = 0; j < height; ++j) { + dst_a[j + ((i >> 1) * dst_stride_a)] = src[i + (j * src_stride)]; + dst_b[j + ((i >> 1) * dst_stride_b)] = src[i + (j * src_stride) + 1]; + } + } +} + +#ifdef __cplusplus +} // extern "C" +} // namespace libyuv +#endif diff --git a/media/libaom/src/third_party/libyuv/source/rotate_gcc.cc b/media/libaom/src/third_party/libyuv/source/rotate_gcc.cc new file mode 100644 index 000000000..fd385bcd3 --- /dev/null +++ b/media/libaom/src/third_party/libyuv/source/rotate_gcc.cc @@ -0,0 +1,493 @@ +/* + * Copyright 2015 The LibYuv Project Authors. All rights reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#include "libyuv/row.h" +#include "libyuv/rotate_row.h" + +#ifdef __cplusplus +namespace libyuv { +extern "C" { +#endif + +// This module is for GCC x86 and x64. +#if !defined(LIBYUV_DISABLE_X86) && (defined(__x86_64__) || defined(__i386__)) + +#if !defined(LIBYUV_DISABLE_X86) && \ + (defined(__i386__) || (defined(__x86_64__) && !defined(__native_client__))) +void TransposeWx8_SSSE3(const uint8* src, int src_stride, + uint8* dst, int dst_stride, int width) { + asm volatile ( + // Read in the data from the source pointer. + // First round of bit swap. + ".p2align 2 \n" + "1: \n" + "movq (%0),%%xmm0 \n" + "movq (%0,%3),%%xmm1 \n" + "lea (%0,%3,2),%0 \n" + "punpcklbw %%xmm1,%%xmm0 \n" + "movq (%0),%%xmm2 \n" + "movdqa %%xmm0,%%xmm1 \n" + "palignr $0x8,%%xmm1,%%xmm1 \n" + "movq (%0,%3),%%xmm3 \n" + "lea (%0,%3,2),%0 \n" + "punpcklbw %%xmm3,%%xmm2 \n" + "movdqa %%xmm2,%%xmm3 \n" + "movq (%0),%%xmm4 \n" + "palignr $0x8,%%xmm3,%%xmm3 \n" + "movq (%0,%3),%%xmm5 \n" + "lea (%0,%3,2),%0 \n" + "punpcklbw %%xmm5,%%xmm4 \n" + "movdqa %%xmm4,%%xmm5 \n" + "movq (%0),%%xmm6 \n" + "palignr $0x8,%%xmm5,%%xmm5 \n" + "movq (%0,%3),%%xmm7 \n" + "lea (%0,%3,2),%0 \n" + "punpcklbw %%xmm7,%%xmm6 \n" + "neg %3 \n" + "movdqa %%xmm6,%%xmm7 \n" + "lea 0x8(%0,%3,8),%0 \n" + "palignr $0x8,%%xmm7,%%xmm7 \n" + "neg %3 \n" + // Second round of bit swap. + "punpcklwd %%xmm2,%%xmm0 \n" + "punpcklwd %%xmm3,%%xmm1 \n" + "movdqa %%xmm0,%%xmm2 \n" + "movdqa %%xmm1,%%xmm3 \n" + "palignr $0x8,%%xmm2,%%xmm2 \n" + "palignr $0x8,%%xmm3,%%xmm3 \n" + "punpcklwd %%xmm6,%%xmm4 \n" + "punpcklwd %%xmm7,%%xmm5 \n" + "movdqa %%xmm4,%%xmm6 \n" + "movdqa %%xmm5,%%xmm7 \n" + "palignr $0x8,%%xmm6,%%xmm6 \n" + "palignr $0x8,%%xmm7,%%xmm7 \n" + // Third round of bit swap. + // Write to the destination pointer. + "punpckldq %%xmm4,%%xmm0 \n" + "movq %%xmm0,(%1) \n" + "movdqa %%xmm0,%%xmm4 \n" + "palignr $0x8,%%xmm4,%%xmm4 \n" + "movq %%xmm4,(%1,%4) \n" + "lea (%1,%4,2),%1 \n" + "punpckldq %%xmm6,%%xmm2 \n" + "movdqa %%xmm2,%%xmm6 \n" + "movq %%xmm2,(%1) \n" + "palignr $0x8,%%xmm6,%%xmm6 \n" + "punpckldq %%xmm5,%%xmm1 \n" + "movq %%xmm6,(%1,%4) \n" + "lea (%1,%4,2),%1 \n" + "movdqa %%xmm1,%%xmm5 \n" + "movq %%xmm1,(%1) \n" + "palignr $0x8,%%xmm5,%%xmm5 \n" + "movq %%xmm5,(%1,%4) \n" + "lea (%1,%4,2),%1 \n" + "punpckldq %%xmm7,%%xmm3 \n" + "movq %%xmm3,(%1) \n" + "movdqa %%xmm3,%%xmm7 \n" + "palignr $0x8,%%xmm7,%%xmm7 \n" + "sub $0x8,%2 \n" + "movq %%xmm7,(%1,%4) \n" + "lea (%1,%4,2),%1 \n" + "jg 1b \n" + : "+r"(src), // %0 + "+r"(dst), // %1 + "+r"(width) // %2 + : "r"((intptr_t)(src_stride)), // %3 + "r"((intptr_t)(dst_stride)) // %4 + : "memory", "cc", + "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7" + ); +} + +#if !defined(LIBYUV_DISABLE_X86) && defined(__i386__) && !defined(__clang__) +void TransposeUVWx8_SSE2(const uint8* src, int src_stride, + uint8* dst_a, int dst_stride_a, + uint8* dst_b, int dst_stride_b, int width); + asm ( + DECLARE_FUNCTION(TransposeUVWx8_SSE2) + "push %ebx \n" + "push %esi \n" + "push %edi \n" + "push %ebp \n" + "mov 0x14(%esp),%eax \n" + "mov 0x18(%esp),%edi \n" + "mov 0x1c(%esp),%edx \n" + "mov 0x20(%esp),%esi \n" + "mov 0x24(%esp),%ebx \n" + "mov 0x28(%esp),%ebp \n" + "mov %esp,%ecx \n" + "sub $0x14,%esp \n" + "and $0xfffffff0,%esp \n" + "mov %ecx,0x10(%esp) \n" + "mov 0x2c(%ecx),%ecx \n" + +"1: \n" + "movdqu (%eax),%xmm0 \n" + "movdqu (%eax,%edi,1),%xmm1 \n" + "lea (%eax,%edi,2),%eax \n" + "movdqa %xmm0,%xmm7 \n" + "punpcklbw %xmm1,%xmm0 \n" + "punpckhbw %xmm1,%xmm7 \n" + "movdqa %xmm7,%xmm1 \n" + "movdqu (%eax),%xmm2 \n" + "movdqu (%eax,%edi,1),%xmm3 \n" + "lea (%eax,%edi,2),%eax \n" + "movdqa %xmm2,%xmm7 \n" + "punpcklbw %xmm3,%xmm2 \n" + "punpckhbw %xmm3,%xmm7 \n" + "movdqa %xmm7,%xmm3 \n" + "movdqu (%eax),%xmm4 \n" + "movdqu (%eax,%edi,1),%xmm5 \n" + "lea (%eax,%edi,2),%eax \n" + "movdqa %xmm4,%xmm7 \n" + "punpcklbw %xmm5,%xmm4 \n" + "punpckhbw %xmm5,%xmm7 \n" + "movdqa %xmm7,%xmm5 \n" + "movdqu (%eax),%xmm6 \n" + "movdqu (%eax,%edi,1),%xmm7 \n" + "lea (%eax,%edi,2),%eax \n" + "movdqu %xmm5,(%esp) \n" + "neg %edi \n" + "movdqa %xmm6,%xmm5 \n" + "punpcklbw %xmm7,%xmm6 \n" + "punpckhbw %xmm7,%xmm5 \n" + "movdqa %xmm5,%xmm7 \n" + "lea 0x10(%eax,%edi,8),%eax \n" + "neg %edi \n" + "movdqa %xmm0,%xmm5 \n" + "punpcklwd %xmm2,%xmm0 \n" + "punpckhwd %xmm2,%xmm5 \n" + "movdqa %xmm5,%xmm2 \n" + "movdqa %xmm1,%xmm5 \n" + "punpcklwd %xmm3,%xmm1 \n" + "punpckhwd %xmm3,%xmm5 \n" + "movdqa %xmm5,%xmm3 \n" + "movdqa %xmm4,%xmm5 \n" + "punpcklwd %xmm6,%xmm4 \n" + "punpckhwd %xmm6,%xmm5 \n" + "movdqa %xmm5,%xmm6 \n" + "movdqu (%esp),%xmm5 \n" + "movdqu %xmm6,(%esp) \n" + "movdqa %xmm5,%xmm6 \n" + "punpcklwd %xmm7,%xmm5 \n" + "punpckhwd %xmm7,%xmm6 \n" + "movdqa %xmm6,%xmm7 \n" + "movdqa %xmm0,%xmm6 \n" + "punpckldq %xmm4,%xmm0 \n" + "punpckhdq %xmm4,%xmm6 \n" + "movdqa %xmm6,%xmm4 \n" + "movdqu (%esp),%xmm6 \n" + "movlpd %xmm0,(%edx) \n" + "movhpd %xmm0,(%ebx) \n" + "movlpd %xmm4,(%edx,%esi,1) \n" + "lea (%edx,%esi,2),%edx \n" + "movhpd %xmm4,(%ebx,%ebp,1) \n" + "lea (%ebx,%ebp,2),%ebx \n" + "movdqa %xmm2,%xmm0 \n" + "punpckldq %xmm6,%xmm2 \n" + "movlpd %xmm2,(%edx) \n" + "movhpd %xmm2,(%ebx) \n" + "punpckhdq %xmm6,%xmm0 \n" + "movlpd %xmm0,(%edx,%esi,1) \n" + "lea (%edx,%esi,2),%edx \n" + "movhpd %xmm0,(%ebx,%ebp,1) \n" + "lea (%ebx,%ebp,2),%ebx \n" + "movdqa %xmm1,%xmm0 \n" + "punpckldq %xmm5,%xmm1 \n" + "movlpd %xmm1,(%edx) \n" + "movhpd %xmm1,(%ebx) \n" + "punpckhdq %xmm5,%xmm0 \n" + "movlpd %xmm0,(%edx,%esi,1) \n" + "lea (%edx,%esi,2),%edx \n" + "movhpd %xmm0,(%ebx,%ebp,1) \n" + "lea (%ebx,%ebp,2),%ebx \n" + "movdqa %xmm3,%xmm0 \n" + "punpckldq %xmm7,%xmm3 \n" + "movlpd %xmm3,(%edx) \n" + "movhpd %xmm3,(%ebx) \n" + "punpckhdq %xmm7,%xmm0 \n" + "sub $0x8,%ecx \n" + "movlpd %xmm0,(%edx,%esi,1) \n" + "lea (%edx,%esi,2),%edx \n" + "movhpd %xmm0,(%ebx,%ebp,1) \n" + "lea (%ebx,%ebp,2),%ebx \n" + "jg 1b \n" + "mov 0x10(%esp),%esp \n" + "pop %ebp \n" + "pop %edi \n" + "pop %esi \n" + "pop %ebx \n" +#if defined(__native_client__) + "pop %ecx \n" + "and $0xffffffe0,%ecx \n" + "jmp *%ecx \n" +#else + "ret \n" +#endif +); +#endif +#if !defined(LIBYUV_DISABLE_X86) && !defined(__native_client__) && \ + defined(__x86_64__) +// 64 bit version has enough registers to do 16x8 to 8x16 at a time. +void TransposeWx8_Fast_SSSE3(const uint8* src, int src_stride, + uint8* dst, int dst_stride, int width) { + asm volatile ( + // Read in the data from the source pointer. + // First round of bit swap. + ".p2align 2 \n" +"1: \n" + "movdqu (%0),%%xmm0 \n" + "movdqu (%0,%3),%%xmm1 \n" + "lea (%0,%3,2),%0 \n" + "movdqa %%xmm0,%%xmm8 \n" + "punpcklbw %%xmm1,%%xmm0 \n" + "punpckhbw %%xmm1,%%xmm8 \n" + "movdqu (%0),%%xmm2 \n" + "movdqa %%xmm0,%%xmm1 \n" + "movdqa %%xmm8,%%xmm9 \n" + "palignr $0x8,%%xmm1,%%xmm1 \n" + "palignr $0x8,%%xmm9,%%xmm9 \n" + "movdqu (%0,%3),%%xmm3 \n" + "lea (%0,%3,2),%0 \n" + "movdqa %%xmm2,%%xmm10 \n" + "punpcklbw %%xmm3,%%xmm2 \n" + "punpckhbw %%xmm3,%%xmm10 \n" + "movdqa %%xmm2,%%xmm3 \n" + "movdqa %%xmm10,%%xmm11 \n" + "movdqu (%0),%%xmm4 \n" + "palignr $0x8,%%xmm3,%%xmm3 \n" + "palignr $0x8,%%xmm11,%%xmm11 \n" + "movdqu (%0,%3),%%xmm5 \n" + "lea (%0,%3,2),%0 \n" + "movdqa %%xmm4,%%xmm12 \n" + "punpcklbw %%xmm5,%%xmm4 \n" + "punpckhbw %%xmm5,%%xmm12 \n" + "movdqa %%xmm4,%%xmm5 \n" + "movdqa %%xmm12,%%xmm13 \n" + "movdqu (%0),%%xmm6 \n" + "palignr $0x8,%%xmm5,%%xmm5 \n" + "palignr $0x8,%%xmm13,%%xmm13 \n" + "movdqu (%0,%3),%%xmm7 \n" + "lea (%0,%3,2),%0 \n" + "movdqa %%xmm6,%%xmm14 \n" + "punpcklbw %%xmm7,%%xmm6 \n" + "punpckhbw %%xmm7,%%xmm14 \n" + "neg %3 \n" + "movdqa %%xmm6,%%xmm7 \n" + "movdqa %%xmm14,%%xmm15 \n" + "lea 0x10(%0,%3,8),%0 \n" + "palignr $0x8,%%xmm7,%%xmm7 \n" + "palignr $0x8,%%xmm15,%%xmm15 \n" + "neg %3 \n" + // Second round of bit swap. + "punpcklwd %%xmm2,%%xmm0 \n" + "punpcklwd %%xmm3,%%xmm1 \n" + "movdqa %%xmm0,%%xmm2 \n" + "movdqa %%xmm1,%%xmm3 \n" + "palignr $0x8,%%xmm2,%%xmm2 \n" + "palignr $0x8,%%xmm3,%%xmm3 \n" + "punpcklwd %%xmm6,%%xmm4 \n" + "punpcklwd %%xmm7,%%xmm5 \n" + "movdqa %%xmm4,%%xmm6 \n" + "movdqa %%xmm5,%%xmm7 \n" + "palignr $0x8,%%xmm6,%%xmm6 \n" + "palignr $0x8,%%xmm7,%%xmm7 \n" + "punpcklwd %%xmm10,%%xmm8 \n" + "punpcklwd %%xmm11,%%xmm9 \n" + "movdqa %%xmm8,%%xmm10 \n" + "movdqa %%xmm9,%%xmm11 \n" + "palignr $0x8,%%xmm10,%%xmm10 \n" + "palignr $0x8,%%xmm11,%%xmm11 \n" + "punpcklwd %%xmm14,%%xmm12 \n" + "punpcklwd %%xmm15,%%xmm13 \n" + "movdqa %%xmm12,%%xmm14 \n" + "movdqa %%xmm13,%%xmm15 \n" + "palignr $0x8,%%xmm14,%%xmm14 \n" + "palignr $0x8,%%xmm15,%%xmm15 \n" + // Third round of bit swap. + // Write to the destination pointer. + "punpckldq %%xmm4,%%xmm0 \n" + "movq %%xmm0,(%1) \n" + "movdqa %%xmm0,%%xmm4 \n" + "palignr $0x8,%%xmm4,%%xmm4 \n" + "movq %%xmm4,(%1,%4) \n" + "lea (%1,%4,2),%1 \n" + "punpckldq %%xmm6,%%xmm2 \n" + "movdqa %%xmm2,%%xmm6 \n" + "movq %%xmm2,(%1) \n" + "palignr $0x8,%%xmm6,%%xmm6 \n" + "punpckldq %%xmm5,%%xmm1 \n" + "movq %%xmm6,(%1,%4) \n" + "lea (%1,%4,2),%1 \n" + "movdqa %%xmm1,%%xmm5 \n" + "movq %%xmm1,(%1) \n" + "palignr $0x8,%%xmm5,%%xmm5 \n" + "movq %%xmm5,(%1,%4) \n" + "lea (%1,%4,2),%1 \n" + "punpckldq %%xmm7,%%xmm3 \n" + "movq %%xmm3,(%1) \n" + "movdqa %%xmm3,%%xmm7 \n" + "palignr $0x8,%%xmm7,%%xmm7 \n" + "movq %%xmm7,(%1,%4) \n" + "lea (%1,%4,2),%1 \n" + "punpckldq %%xmm12,%%xmm8 \n" + "movq %%xmm8,(%1) \n" + "movdqa %%xmm8,%%xmm12 \n" + "palignr $0x8,%%xmm12,%%xmm12 \n" + "movq %%xmm12,(%1,%4) \n" + "lea (%1,%4,2),%1 \n" + "punpckldq %%xmm14,%%xmm10 \n" + "movdqa %%xmm10,%%xmm14 \n" + "movq %%xmm10,(%1) \n" + "palignr $0x8,%%xmm14,%%xmm14 \n" + "punpckldq %%xmm13,%%xmm9 \n" + "movq %%xmm14,(%1,%4) \n" + "lea (%1,%4,2),%1 \n" + "movdqa %%xmm9,%%xmm13 \n" + "movq %%xmm9,(%1) \n" + "palignr $0x8,%%xmm13,%%xmm13 \n" + "movq %%xmm13,(%1,%4) \n" + "lea (%1,%4,2),%1 \n" + "punpckldq %%xmm15,%%xmm11 \n" + "movq %%xmm11,(%1) \n" + "movdqa %%xmm11,%%xmm15 \n" + "palignr $0x8,%%xmm15,%%xmm15 \n" + "sub $0x10,%2 \n" + "movq %%xmm15,(%1,%4) \n" + "lea (%1,%4,2),%1 \n" + "jg 1b \n" + : "+r"(src), // %0 + "+r"(dst), // %1 + "+r"(width) // %2 + : "r"((intptr_t)(src_stride)), // %3 + "r"((intptr_t)(dst_stride)) // %4 + : "memory", "cc", + "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7", + "xmm8", "xmm9", "xmm10", "xmm11", "xmm12", "xmm13", "xmm14", "xmm15" +); +} + +void TransposeUVWx8_SSE2(const uint8* src, int src_stride, + uint8* dst_a, int dst_stride_a, + uint8* dst_b, int dst_stride_b, int width) { + asm volatile ( + // Read in the data from the source pointer. + // First round of bit swap. + ".p2align 2 \n" +"1: \n" + "movdqu (%0),%%xmm0 \n" + "movdqu (%0,%4),%%xmm1 \n" + "lea (%0,%4,2),%0 \n" + "movdqa %%xmm0,%%xmm8 \n" + "punpcklbw %%xmm1,%%xmm0 \n" + "punpckhbw %%xmm1,%%xmm8 \n" + "movdqa %%xmm8,%%xmm1 \n" + "movdqu (%0),%%xmm2 \n" + "movdqu (%0,%4),%%xmm3 \n" + "lea (%0,%4,2),%0 \n" + "movdqa %%xmm2,%%xmm8 \n" + "punpcklbw %%xmm3,%%xmm2 \n" + "punpckhbw %%xmm3,%%xmm8 \n" + "movdqa %%xmm8,%%xmm3 \n" + "movdqu (%0),%%xmm4 \n" + "movdqu (%0,%4),%%xmm5 \n" + "lea (%0,%4,2),%0 \n" + "movdqa %%xmm4,%%xmm8 \n" + "punpcklbw %%xmm5,%%xmm4 \n" + "punpckhbw %%xmm5,%%xmm8 \n" + "movdqa %%xmm8,%%xmm5 \n" + "movdqu (%0),%%xmm6 \n" + "movdqu (%0,%4),%%xmm7 \n" + "lea (%0,%4,2),%0 \n" + "movdqa %%xmm6,%%xmm8 \n" + "punpcklbw %%xmm7,%%xmm6 \n" + "neg %4 \n" + "lea 0x10(%0,%4,8),%0 \n" + "punpckhbw %%xmm7,%%xmm8 \n" + "movdqa %%xmm8,%%xmm7 \n" + "neg %4 \n" + // Second round of bit swap. + "movdqa %%xmm0,%%xmm8 \n" + "movdqa %%xmm1,%%xmm9 \n" + "punpckhwd %%xmm2,%%xmm8 \n" + "punpckhwd %%xmm3,%%xmm9 \n" + "punpcklwd %%xmm2,%%xmm0 \n" + "punpcklwd %%xmm3,%%xmm1 \n" + "movdqa %%xmm8,%%xmm2 \n" + "movdqa %%xmm9,%%xmm3 \n" + "movdqa %%xmm4,%%xmm8 \n" + "movdqa %%xmm5,%%xmm9 \n" + "punpckhwd %%xmm6,%%xmm8 \n" + "punpckhwd %%xmm7,%%xmm9 \n" + "punpcklwd %%xmm6,%%xmm4 \n" + "punpcklwd %%xmm7,%%xmm5 \n" + "movdqa %%xmm8,%%xmm6 \n" + "movdqa %%xmm9,%%xmm7 \n" + // Third round of bit swap. + // Write to the destination pointer. + "movdqa %%xmm0,%%xmm8 \n" + "punpckldq %%xmm4,%%xmm0 \n" + "movlpd %%xmm0,(%1) \n" // Write back U channel + "movhpd %%xmm0,(%2) \n" // Write back V channel + "punpckhdq %%xmm4,%%xmm8 \n" + "movlpd %%xmm8,(%1,%5) \n" + "lea (%1,%5,2),%1 \n" + "movhpd %%xmm8,(%2,%6) \n" + "lea (%2,%6,2),%2 \n" + "movdqa %%xmm2,%%xmm8 \n" + "punpckldq %%xmm6,%%xmm2 \n" + "movlpd %%xmm2,(%1) \n" + "movhpd %%xmm2,(%2) \n" + "punpckhdq %%xmm6,%%xmm8 \n" + "movlpd %%xmm8,(%1,%5) \n" + "lea (%1,%5,2),%1 \n" + "movhpd %%xmm8,(%2,%6) \n" + "lea (%2,%6,2),%2 \n" + "movdqa %%xmm1,%%xmm8 \n" + "punpckldq %%xmm5,%%xmm1 \n" + "movlpd %%xmm1,(%1) \n" + "movhpd %%xmm1,(%2) \n" + "punpckhdq %%xmm5,%%xmm8 \n" + "movlpd %%xmm8,(%1,%5) \n" + "lea (%1,%5,2),%1 \n" + "movhpd %%xmm8,(%2,%6) \n" + "lea (%2,%6,2),%2 \n" + "movdqa %%xmm3,%%xmm8 \n" + "punpckldq %%xmm7,%%xmm3 \n" + "movlpd %%xmm3,(%1) \n" + "movhpd %%xmm3,(%2) \n" + "punpckhdq %%xmm7,%%xmm8 \n" + "sub $0x8,%3 \n" + "movlpd %%xmm8,(%1,%5) \n" + "lea (%1,%5,2),%1 \n" + "movhpd %%xmm8,(%2,%6) \n" + "lea (%2,%6,2),%2 \n" + "jg 1b \n" + : "+r"(src), // %0 + "+r"(dst_a), // %1 + "+r"(dst_b), // %2 + "+r"(width) // %3 + : "r"((intptr_t)(src_stride)), // %4 + "r"((intptr_t)(dst_stride_a)), // %5 + "r"((intptr_t)(dst_stride_b)) // %6 + : "memory", "cc", + "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7", + "xmm8", "xmm9" +); +} +#endif +#endif + +#endif // defined(__x86_64__) || defined(__i386__) + +#ifdef __cplusplus +} // extern "C" +} // namespace libyuv +#endif diff --git a/media/libaom/src/third_party/libyuv/source/rotate_mips.cc b/media/libaom/src/third_party/libyuv/source/rotate_mips.cc new file mode 100644 index 000000000..efe6bd909 --- /dev/null +++ b/media/libaom/src/third_party/libyuv/source/rotate_mips.cc @@ -0,0 +1,484 @@ +/* + * Copyright 2011 The LibYuv Project Authors. All rights reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#include "libyuv/row.h" +#include "libyuv/rotate_row.h" + +#include "libyuv/basic_types.h" + +#ifdef __cplusplus +namespace libyuv { +extern "C" { +#endif + +#if !defined(LIBYUV_DISABLE_MIPS) && \ + defined(__mips_dsp) && (__mips_dsp_rev >= 2) && \ + (_MIPS_SIM == _MIPS_SIM_ABI32) + +void TransposeWx8_MIPS_DSPR2(const uint8* src, int src_stride, + uint8* dst, int dst_stride, int width) { + __asm__ __volatile__ ( + ".set push \n" + ".set noreorder \n" + "sll $t2, %[src_stride], 0x1 \n" // src_stride x 2 + "sll $t4, %[src_stride], 0x2 \n" // src_stride x 4 + "sll $t9, %[src_stride], 0x3 \n" // src_stride x 8 + "addu $t3, $t2, %[src_stride] \n" + "addu $t5, $t4, %[src_stride] \n" + "addu $t6, $t2, $t4 \n" + "andi $t0, %[dst], 0x3 \n" + "andi $t1, %[dst_stride], 0x3 \n" + "or $t0, $t0, $t1 \n" + "bnez $t0, 11f \n" + " subu $t7, $t9, %[src_stride] \n" +//dst + dst_stride word aligned + "1: \n" + "lbu $t0, 0(%[src]) \n" + "lbux $t1, %[src_stride](%[src]) \n" + "lbux $t8, $t2(%[src]) \n" + "lbux $t9, $t3(%[src]) \n" + "sll $t1, $t1, 16 \n" + "sll $t9, $t9, 16 \n" + "or $t0, $t0, $t1 \n" + "or $t8, $t8, $t9 \n" + "precr.qb.ph $s0, $t8, $t0 \n" + "lbux $t0, $t4(%[src]) \n" + "lbux $t1, $t5(%[src]) \n" + "lbux $t8, $t6(%[src]) \n" + "lbux $t9, $t7(%[src]) \n" + "sll $t1, $t1, 16 \n" + "sll $t9, $t9, 16 \n" + "or $t0, $t0, $t1 \n" + "or $t8, $t8, $t9 \n" + "precr.qb.ph $s1, $t8, $t0 \n" + "sw $s0, 0(%[dst]) \n" + "addiu %[width], -1 \n" + "addiu %[src], 1 \n" + "sw $s1, 4(%[dst]) \n" + "bnez %[width], 1b \n" + " addu %[dst], %[dst], %[dst_stride] \n" + "b 2f \n" +//dst + dst_stride unaligned + "11: \n" + "lbu $t0, 0(%[src]) \n" + "lbux $t1, %[src_stride](%[src]) \n" + "lbux $t8, $t2(%[src]) \n" + "lbux $t9, $t3(%[src]) \n" + "sll $t1, $t1, 16 \n" + "sll $t9, $t9, 16 \n" + "or $t0, $t0, $t1 \n" + "or $t8, $t8, $t9 \n" + "precr.qb.ph $s0, $t8, $t0 \n" + "lbux $t0, $t4(%[src]) \n" + "lbux $t1, $t5(%[src]) \n" + "lbux $t8, $t6(%[src]) \n" + "lbux $t9, $t7(%[src]) \n" + "sll $t1, $t1, 16 \n" + "sll $t9, $t9, 16 \n" + "or $t0, $t0, $t1 \n" + "or $t8, $t8, $t9 \n" + "precr.qb.ph $s1, $t8, $t0 \n" + "swr $s0, 0(%[dst]) \n" + "swl $s0, 3(%[dst]) \n" + "addiu %[width], -1 \n" + "addiu %[src], 1 \n" + "swr $s1, 4(%[dst]) \n" + "swl $s1, 7(%[dst]) \n" + "bnez %[width], 11b \n" + "addu %[dst], %[dst], %[dst_stride] \n" + "2: \n" + ".set pop \n" + :[src] "+r" (src), + [dst] "+r" (dst), + [width] "+r" (width) + :[src_stride] "r" (src_stride), + [dst_stride] "r" (dst_stride) + : "t0", "t1", "t2", "t3", "t4", "t5", + "t6", "t7", "t8", "t9", + "s0", "s1" + ); +} + +void TransposeWx8_Fast_MIPS_DSPR2(const uint8* src, int src_stride, + uint8* dst, int dst_stride, int width) { + __asm__ __volatile__ ( + ".set noat \n" + ".set push \n" + ".set noreorder \n" + "beqz %[width], 2f \n" + " sll $t2, %[src_stride], 0x1 \n" // src_stride x 2 + "sll $t4, %[src_stride], 0x2 \n" // src_stride x 4 + "sll $t9, %[src_stride], 0x3 \n" // src_stride x 8 + "addu $t3, $t2, %[src_stride] \n" + "addu $t5, $t4, %[src_stride] \n" + "addu $t6, $t2, $t4 \n" + + "srl $AT, %[width], 0x2 \n" + "andi $t0, %[dst], 0x3 \n" + "andi $t1, %[dst_stride], 0x3 \n" + "or $t0, $t0, $t1 \n" + "bnez $t0, 11f \n" + " subu $t7, $t9, %[src_stride] \n" +//dst + dst_stride word aligned + "1: \n" + "lw $t0, 0(%[src]) \n" + "lwx $t1, %[src_stride](%[src]) \n" + "lwx $t8, $t2(%[src]) \n" + "lwx $t9, $t3(%[src]) \n" + +// t0 = | 30 | 20 | 10 | 00 | +// t1 = | 31 | 21 | 11 | 01 | +// t8 = | 32 | 22 | 12 | 02 | +// t9 = | 33 | 23 | 13 | 03 | + + "precr.qb.ph $s0, $t1, $t0 \n" + "precr.qb.ph $s1, $t9, $t8 \n" + "precrq.qb.ph $s2, $t1, $t0 \n" + "precrq.qb.ph $s3, $t9, $t8 \n" + + // s0 = | 21 | 01 | 20 | 00 | + // s1 = | 23 | 03 | 22 | 02 | + // s2 = | 31 | 11 | 30 | 10 | + // s3 = | 33 | 13 | 32 | 12 | + + "precr.qb.ph $s4, $s1, $s0 \n" + "precrq.qb.ph $s5, $s1, $s0 \n" + "precr.qb.ph $s6, $s3, $s2 \n" + "precrq.qb.ph $s7, $s3, $s2 \n" + + // s4 = | 03 | 02 | 01 | 00 | + // s5 = | 23 | 22 | 21 | 20 | + // s6 = | 13 | 12 | 11 | 10 | + // s7 = | 33 | 32 | 31 | 30 | + + "lwx $t0, $t4(%[src]) \n" + "lwx $t1, $t5(%[src]) \n" + "lwx $t8, $t6(%[src]) \n" + "lwx $t9, $t7(%[src]) \n" + +// t0 = | 34 | 24 | 14 | 04 | +// t1 = | 35 | 25 | 15 | 05 | +// t8 = | 36 | 26 | 16 | 06 | +// t9 = | 37 | 27 | 17 | 07 | + + "precr.qb.ph $s0, $t1, $t0 \n" + "precr.qb.ph $s1, $t9, $t8 \n" + "precrq.qb.ph $s2, $t1, $t0 \n" + "precrq.qb.ph $s3, $t9, $t8 \n" + + // s0 = | 25 | 05 | 24 | 04 | + // s1 = | 27 | 07 | 26 | 06 | + // s2 = | 35 | 15 | 34 | 14 | + // s3 = | 37 | 17 | 36 | 16 | + + "precr.qb.ph $t0, $s1, $s0 \n" + "precrq.qb.ph $t1, $s1, $s0 \n" + "precr.qb.ph $t8, $s3, $s2 \n" + "precrq.qb.ph $t9, $s3, $s2 \n" + + // t0 = | 07 | 06 | 05 | 04 | + // t1 = | 27 | 26 | 25 | 24 | + // t8 = | 17 | 16 | 15 | 14 | + // t9 = | 37 | 36 | 35 | 34 | + + "addu $s0, %[dst], %[dst_stride] \n" + "addu $s1, $s0, %[dst_stride] \n" + "addu $s2, $s1, %[dst_stride] \n" + + "sw $s4, 0(%[dst]) \n" + "sw $t0, 4(%[dst]) \n" + "sw $s6, 0($s0) \n" + "sw $t8, 4($s0) \n" + "sw $s5, 0($s1) \n" + "sw $t1, 4($s1) \n" + "sw $s7, 0($s2) \n" + "sw $t9, 4($s2) \n" + + "addiu $AT, -1 \n" + "addiu %[src], 4 \n" + + "bnez $AT, 1b \n" + " addu %[dst], $s2, %[dst_stride] \n" + "b 2f \n" +//dst + dst_stride unaligned + "11: \n" + "lw $t0, 0(%[src]) \n" + "lwx $t1, %[src_stride](%[src]) \n" + "lwx $t8, $t2(%[src]) \n" + "lwx $t9, $t3(%[src]) \n" + +// t0 = | 30 | 20 | 10 | 00 | +// t1 = | 31 | 21 | 11 | 01 | +// t8 = | 32 | 22 | 12 | 02 | +// t9 = | 33 | 23 | 13 | 03 | + + "precr.qb.ph $s0, $t1, $t0 \n" + "precr.qb.ph $s1, $t9, $t8 \n" + "precrq.qb.ph $s2, $t1, $t0 \n" + "precrq.qb.ph $s3, $t9, $t8 \n" + + // s0 = | 21 | 01 | 20 | 00 | + // s1 = | 23 | 03 | 22 | 02 | + // s2 = | 31 | 11 | 30 | 10 | + // s3 = | 33 | 13 | 32 | 12 | + + "precr.qb.ph $s4, $s1, $s0 \n" + "precrq.qb.ph $s5, $s1, $s0 \n" + "precr.qb.ph $s6, $s3, $s2 \n" + "precrq.qb.ph $s7, $s3, $s2 \n" + + // s4 = | 03 | 02 | 01 | 00 | + // s5 = | 23 | 22 | 21 | 20 | + // s6 = | 13 | 12 | 11 | 10 | + // s7 = | 33 | 32 | 31 | 30 | + + "lwx $t0, $t4(%[src]) \n" + "lwx $t1, $t5(%[src]) \n" + "lwx $t8, $t6(%[src]) \n" + "lwx $t9, $t7(%[src]) \n" + +// t0 = | 34 | 24 | 14 | 04 | +// t1 = | 35 | 25 | 15 | 05 | +// t8 = | 36 | 26 | 16 | 06 | +// t9 = | 37 | 27 | 17 | 07 | + + "precr.qb.ph $s0, $t1, $t0 \n" + "precr.qb.ph $s1, $t9, $t8 \n" + "precrq.qb.ph $s2, $t1, $t0 \n" + "precrq.qb.ph $s3, $t9, $t8 \n" + + // s0 = | 25 | 05 | 24 | 04 | + // s1 = | 27 | 07 | 26 | 06 | + // s2 = | 35 | 15 | 34 | 14 | + // s3 = | 37 | 17 | 36 | 16 | + + "precr.qb.ph $t0, $s1, $s0 \n" + "precrq.qb.ph $t1, $s1, $s0 \n" + "precr.qb.ph $t8, $s3, $s2 \n" + "precrq.qb.ph $t9, $s3, $s2 \n" + + // t0 = | 07 | 06 | 05 | 04 | + // t1 = | 27 | 26 | 25 | 24 | + // t8 = | 17 | 16 | 15 | 14 | + // t9 = | 37 | 36 | 35 | 34 | + + "addu $s0, %[dst], %[dst_stride] \n" + "addu $s1, $s0, %[dst_stride] \n" + "addu $s2, $s1, %[dst_stride] \n" + + "swr $s4, 0(%[dst]) \n" + "swl $s4, 3(%[dst]) \n" + "swr $t0, 4(%[dst]) \n" + "swl $t0, 7(%[dst]) \n" + "swr $s6, 0($s0) \n" + "swl $s6, 3($s0) \n" + "swr $t8, 4($s0) \n" + "swl $t8, 7($s0) \n" + "swr $s5, 0($s1) \n" + "swl $s5, 3($s1) \n" + "swr $t1, 4($s1) \n" + "swl $t1, 7($s1) \n" + "swr $s7, 0($s2) \n" + "swl $s7, 3($s2) \n" + "swr $t9, 4($s2) \n" + "swl $t9, 7($s2) \n" + + "addiu $AT, -1 \n" + "addiu %[src], 4 \n" + + "bnez $AT, 11b \n" + " addu %[dst], $s2, %[dst_stride] \n" + "2: \n" + ".set pop \n" + ".set at \n" + :[src] "+r" (src), + [dst] "+r" (dst), + [width] "+r" (width) + :[src_stride] "r" (src_stride), + [dst_stride] "r" (dst_stride) + : "t0", "t1", "t2", "t3", "t4", "t5", "t6", "t7", "t8", "t9", + "s0", "s1", "s2", "s3", "s4", "s5", "s6", "s7" + ); +} + +void TransposeUVWx8_MIPS_DSPR2(const uint8* src, int src_stride, + uint8* dst_a, int dst_stride_a, + uint8* dst_b, int dst_stride_b, + int width) { + __asm__ __volatile__ ( + ".set push \n" + ".set noreorder \n" + "beqz %[width], 2f \n" + " sll $t2, %[src_stride], 0x1 \n" // src_stride x 2 + "sll $t4, %[src_stride], 0x2 \n" // src_stride x 4 + "sll $t9, %[src_stride], 0x3 \n" // src_stride x 8 + "addu $t3, $t2, %[src_stride] \n" + "addu $t5, $t4, %[src_stride] \n" + "addu $t6, $t2, $t4 \n" + "subu $t7, $t9, %[src_stride] \n" + "srl $t1, %[width], 1 \n" + +// check word aligment for dst_a, dst_b, dst_stride_a and dst_stride_b + "andi $t0, %[dst_a], 0x3 \n" + "andi $t8, %[dst_b], 0x3 \n" + "or $t0, $t0, $t8 \n" + "andi $t8, %[dst_stride_a], 0x3 \n" + "andi $s5, %[dst_stride_b], 0x3 \n" + "or $t8, $t8, $s5 \n" + "or $t0, $t0, $t8 \n" + "bnez $t0, 11f \n" + " nop \n" +// dst + dst_stride word aligned (both, a & b dst addresses) + "1: \n" + "lw $t0, 0(%[src]) \n" // |B0|A0|b0|a0| + "lwx $t8, %[src_stride](%[src]) \n" // |B1|A1|b1|a1| + "addu $s5, %[dst_a], %[dst_stride_a] \n" + "lwx $t9, $t2(%[src]) \n" // |B2|A2|b2|a2| + "lwx $s0, $t3(%[src]) \n" // |B3|A3|b3|a3| + "addu $s6, %[dst_b], %[dst_stride_b] \n" + + "precrq.ph.w $s1, $t8, $t0 \n" // |B1|A1|B0|A0| + "precrq.ph.w $s2, $s0, $t9 \n" // |B3|A3|B2|A2| + "precr.qb.ph $s3, $s2, $s1 \n" // |A3|A2|A1|A0| + "precrq.qb.ph $s4, $s2, $s1 \n" // |B3|B2|B1|B0| + + "sll $t0, $t0, 16 \n" + "packrl.ph $s1, $t8, $t0 \n" // |b1|a1|b0|a0| + "sll $t9, $t9, 16 \n" + "packrl.ph $s2, $s0, $t9 \n" // |b3|a3|b2|a2| + + "sw $s3, 0($s5) \n" + "sw $s4, 0($s6) \n" + + "precr.qb.ph $s3, $s2, $s1 \n" // |a3|a2|a1|a0| + "precrq.qb.ph $s4, $s2, $s1 \n" // |b3|b2|b1|b0| + + "lwx $t0, $t4(%[src]) \n" // |B4|A4|b4|a4| + "lwx $t8, $t5(%[src]) \n" // |B5|A5|b5|a5| + "lwx $t9, $t6(%[src]) \n" // |B6|A6|b6|a6| + "lwx $s0, $t7(%[src]) \n" // |B7|A7|b7|a7| + "sw $s3, 0(%[dst_a]) \n" + "sw $s4, 0(%[dst_b]) \n" + + "precrq.ph.w $s1, $t8, $t0 \n" // |B5|A5|B4|A4| + "precrq.ph.w $s2, $s0, $t9 \n" // |B6|A6|B7|A7| + "precr.qb.ph $s3, $s2, $s1 \n" // |A7|A6|A5|A4| + "precrq.qb.ph $s4, $s2, $s1 \n" // |B7|B6|B5|B4| + + "sll $t0, $t0, 16 \n" + "packrl.ph $s1, $t8, $t0 \n" // |b5|a5|b4|a4| + "sll $t9, $t9, 16 \n" + "packrl.ph $s2, $s0, $t9 \n" // |b7|a7|b6|a6| + "sw $s3, 4($s5) \n" + "sw $s4, 4($s6) \n" + + "precr.qb.ph $s3, $s2, $s1 \n" // |a7|a6|a5|a4| + "precrq.qb.ph $s4, $s2, $s1 \n" // |b7|b6|b5|b4| + + "addiu %[src], 4 \n" + "addiu $t1, -1 \n" + "sll $t0, %[dst_stride_a], 1 \n" + "sll $t8, %[dst_stride_b], 1 \n" + "sw $s3, 4(%[dst_a]) \n" + "sw $s4, 4(%[dst_b]) \n" + "addu %[dst_a], %[dst_a], $t0 \n" + "bnez $t1, 1b \n" + " addu %[dst_b], %[dst_b], $t8 \n" + "b 2f \n" + " nop \n" + +// dst_a or dst_b or dst_stride_a or dst_stride_b not word aligned + "11: \n" + "lw $t0, 0(%[src]) \n" // |B0|A0|b0|a0| + "lwx $t8, %[src_stride](%[src]) \n" // |B1|A1|b1|a1| + "addu $s5, %[dst_a], %[dst_stride_a] \n" + "lwx $t9, $t2(%[src]) \n" // |B2|A2|b2|a2| + "lwx $s0, $t3(%[src]) \n" // |B3|A3|b3|a3| + "addu $s6, %[dst_b], %[dst_stride_b] \n" + + "precrq.ph.w $s1, $t8, $t0 \n" // |B1|A1|B0|A0| + "precrq.ph.w $s2, $s0, $t9 \n" // |B3|A3|B2|A2| + "precr.qb.ph $s3, $s2, $s1 \n" // |A3|A2|A1|A0| + "precrq.qb.ph $s4, $s2, $s1 \n" // |B3|B2|B1|B0| + + "sll $t0, $t0, 16 \n" + "packrl.ph $s1, $t8, $t0 \n" // |b1|a1|b0|a0| + "sll $t9, $t9, 16 \n" + "packrl.ph $s2, $s0, $t9 \n" // |b3|a3|b2|a2| + + "swr $s3, 0($s5) \n" + "swl $s3, 3($s5) \n" + "swr $s4, 0($s6) \n" + "swl $s4, 3($s6) \n" + + "precr.qb.ph $s3, $s2, $s1 \n" // |a3|a2|a1|a0| + "precrq.qb.ph $s4, $s2, $s1 \n" // |b3|b2|b1|b0| + + "lwx $t0, $t4(%[src]) \n" // |B4|A4|b4|a4| + "lwx $t8, $t5(%[src]) \n" // |B5|A5|b5|a5| + "lwx $t9, $t6(%[src]) \n" // |B6|A6|b6|a6| + "lwx $s0, $t7(%[src]) \n" // |B7|A7|b7|a7| + "swr $s3, 0(%[dst_a]) \n" + "swl $s3, 3(%[dst_a]) \n" + "swr $s4, 0(%[dst_b]) \n" + "swl $s4, 3(%[dst_b]) \n" + + "precrq.ph.w $s1, $t8, $t0 \n" // |B5|A5|B4|A4| + "precrq.ph.w $s2, $s0, $t9 \n" // |B6|A6|B7|A7| + "precr.qb.ph $s3, $s2, $s1 \n" // |A7|A6|A5|A4| + "precrq.qb.ph $s4, $s2, $s1 \n" // |B7|B6|B5|B4| + + "sll $t0, $t0, 16 \n" + "packrl.ph $s1, $t8, $t0 \n" // |b5|a5|b4|a4| + "sll $t9, $t9, 16 \n" + "packrl.ph $s2, $s0, $t9 \n" // |b7|a7|b6|a6| + + "swr $s3, 4($s5) \n" + "swl $s3, 7($s5) \n" + "swr $s4, 4($s6) \n" + "swl $s4, 7($s6) \n" + + "precr.qb.ph $s3, $s2, $s1 \n" // |a7|a6|a5|a4| + "precrq.qb.ph $s4, $s2, $s1 \n" // |b7|b6|b5|b4| + + "addiu %[src], 4 \n" + "addiu $t1, -1 \n" + "sll $t0, %[dst_stride_a], 1 \n" + "sll $t8, %[dst_stride_b], 1 \n" + "swr $s3, 4(%[dst_a]) \n" + "swl $s3, 7(%[dst_a]) \n" + "swr $s4, 4(%[dst_b]) \n" + "swl $s4, 7(%[dst_b]) \n" + "addu %[dst_a], %[dst_a], $t0 \n" + "bnez $t1, 11b \n" + " addu %[dst_b], %[dst_b], $t8 \n" + + "2: \n" + ".set pop \n" + : [src] "+r" (src), + [dst_a] "+r" (dst_a), + [dst_b] "+r" (dst_b), + [width] "+r" (width), + [src_stride] "+r" (src_stride) + : [dst_stride_a] "r" (dst_stride_a), + [dst_stride_b] "r" (dst_stride_b) + : "t0", "t1", "t2", "t3", "t4", "t5", + "t6", "t7", "t8", "t9", + "s0", "s1", "s2", "s3", + "s4", "s5", "s6" + ); +} + +#endif // defined(__mips_dsp) && (__mips_dsp_rev >= 2) + +#ifdef __cplusplus +} // extern "C" +} // namespace libyuv +#endif diff --git a/media/libaom/src/third_party/libyuv/source/rotate_neon.cc b/media/libaom/src/third_party/libyuv/source/rotate_neon.cc new file mode 100644 index 000000000..76043b3b3 --- /dev/null +++ b/media/libaom/src/third_party/libyuv/source/rotate_neon.cc @@ -0,0 +1,535 @@ +/* + * Copyright 2011 The LibYuv Project Authors. All rights reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#include "libyuv/row.h" +#include "libyuv/rotate_row.h" + +#include "libyuv/basic_types.h" + +#ifdef __cplusplus +namespace libyuv { +extern "C" { +#endif + +#if !defined(LIBYUV_DISABLE_NEON) && defined(__ARM_NEON__) && \ + !defined(__aarch64__) + +static uvec8 kVTbl4x4Transpose = + { 0, 4, 8, 12, 1, 5, 9, 13, 2, 6, 10, 14, 3, 7, 11, 15 }; + +void TransposeWx8_NEON(const uint8* src, int src_stride, + uint8* dst, int dst_stride, + int width) { + const uint8* src_temp = NULL; + asm volatile ( + // loops are on blocks of 8. loop will stop when + // counter gets to or below 0. starting the counter + // at w-8 allow for this + "sub %5, #8 \n" + + // handle 8x8 blocks. this should be the majority of the plane + ".p2align 2 \n" + "1: \n" + "mov %0, %1 \n" + + MEMACCESS(0) + "vld1.8 {d0}, [%0], %2 \n" + MEMACCESS(0) + "vld1.8 {d1}, [%0], %2 \n" + MEMACCESS(0) + "vld1.8 {d2}, [%0], %2 \n" + MEMACCESS(0) + "vld1.8 {d3}, [%0], %2 \n" + MEMACCESS(0) + "vld1.8 {d4}, [%0], %2 \n" + MEMACCESS(0) + "vld1.8 {d5}, [%0], %2 \n" + MEMACCESS(0) + "vld1.8 {d6}, [%0], %2 \n" + MEMACCESS(0) + "vld1.8 {d7}, [%0] \n" + + "vtrn.8 d1, d0 \n" + "vtrn.8 d3, d2 \n" + "vtrn.8 d5, d4 \n" + "vtrn.8 d7, d6 \n" + + "vtrn.16 d1, d3 \n" + "vtrn.16 d0, d2 \n" + "vtrn.16 d5, d7 \n" + "vtrn.16 d4, d6 \n" + + "vtrn.32 d1, d5 \n" + "vtrn.32 d0, d4 \n" + "vtrn.32 d3, d7 \n" + "vtrn.32 d2, d6 \n" + + "vrev16.8 q0, q0 \n" + "vrev16.8 q1, q1 \n" + "vrev16.8 q2, q2 \n" + "vrev16.8 q3, q3 \n" + + "mov %0, %3 \n" + + MEMACCESS(0) + "vst1.8 {d1}, [%0], %4 \n" + MEMACCESS(0) + "vst1.8 {d0}, [%0], %4 \n" + MEMACCESS(0) + "vst1.8 {d3}, [%0], %4 \n" + MEMACCESS(0) + "vst1.8 {d2}, [%0], %4 \n" + MEMACCESS(0) + "vst1.8 {d5}, [%0], %4 \n" + MEMACCESS(0) + "vst1.8 {d4}, [%0], %4 \n" + MEMACCESS(0) + "vst1.8 {d7}, [%0], %4 \n" + MEMACCESS(0) + "vst1.8 {d6}, [%0] \n" + + "add %1, #8 \n" // src += 8 + "add %3, %3, %4, lsl #3 \n" // dst += 8 * dst_stride + "subs %5, #8 \n" // w -= 8 + "bge 1b \n" + + // add 8 back to counter. if the result is 0 there are + // no residuals. + "adds %5, #8 \n" + "beq 4f \n" + + // some residual, so between 1 and 7 lines left to transpose + "cmp %5, #2 \n" + "blt 3f \n" + + "cmp %5, #4 \n" + "blt 2f \n" + + // 4x8 block + "mov %0, %1 \n" + MEMACCESS(0) + "vld1.32 {d0[0]}, [%0], %2 \n" + MEMACCESS(0) + "vld1.32 {d0[1]}, [%0], %2 \n" + MEMACCESS(0) + "vld1.32 {d1[0]}, [%0], %2 \n" + MEMACCESS(0) + "vld1.32 {d1[1]}, [%0], %2 \n" + MEMACCESS(0) + "vld1.32 {d2[0]}, [%0], %2 \n" + MEMACCESS(0) + "vld1.32 {d2[1]}, [%0], %2 \n" + MEMACCESS(0) + "vld1.32 {d3[0]}, [%0], %2 \n" + MEMACCESS(0) + "vld1.32 {d3[1]}, [%0] \n" + + "mov %0, %3 \n" + + MEMACCESS(6) + "vld1.8 {q3}, [%6] \n" + + "vtbl.8 d4, {d0, d1}, d6 \n" + "vtbl.8 d5, {d0, d1}, d7 \n" + "vtbl.8 d0, {d2, d3}, d6 \n" + "vtbl.8 d1, {d2, d3}, d7 \n" + + // TODO(frkoenig): Rework shuffle above to + // write out with 4 instead of 8 writes. + MEMACCESS(0) + "vst1.32 {d4[0]}, [%0], %4 \n" + MEMACCESS(0) + "vst1.32 {d4[1]}, [%0], %4 \n" + MEMACCESS(0) + "vst1.32 {d5[0]}, [%0], %4 \n" + MEMACCESS(0) + "vst1.32 {d5[1]}, [%0] \n" + + "add %0, %3, #4 \n" + MEMACCESS(0) + "vst1.32 {d0[0]}, [%0], %4 \n" + MEMACCESS(0) + "vst1.32 {d0[1]}, [%0], %4 \n" + MEMACCESS(0) + "vst1.32 {d1[0]}, [%0], %4 \n" + MEMACCESS(0) + "vst1.32 {d1[1]}, [%0] \n" + + "add %1, #4 \n" // src += 4 + "add %3, %3, %4, lsl #2 \n" // dst += 4 * dst_stride + "subs %5, #4 \n" // w -= 4 + "beq 4f \n" + + // some residual, check to see if it includes a 2x8 block, + // or less + "cmp %5, #2 \n" + "blt 3f \n" + + // 2x8 block + "2: \n" + "mov %0, %1 \n" + MEMACCESS(0) + "vld1.16 {d0[0]}, [%0], %2 \n" + MEMACCESS(0) + "vld1.16 {d1[0]}, [%0], %2 \n" + MEMACCESS(0) + "vld1.16 {d0[1]}, [%0], %2 \n" + MEMACCESS(0) + "vld1.16 {d1[1]}, [%0], %2 \n" + MEMACCESS(0) + "vld1.16 {d0[2]}, [%0], %2 \n" + MEMACCESS(0) + "vld1.16 {d1[2]}, [%0], %2 \n" + MEMACCESS(0) + "vld1.16 {d0[3]}, [%0], %2 \n" + MEMACCESS(0) + "vld1.16 {d1[3]}, [%0] \n" + + "vtrn.8 d0, d1 \n" + + "mov %0, %3 \n" + + MEMACCESS(0) + "vst1.64 {d0}, [%0], %4 \n" + MEMACCESS(0) + "vst1.64 {d1}, [%0] \n" + + "add %1, #2 \n" // src += 2 + "add %3, %3, %4, lsl #1 \n" // dst += 2 * dst_stride + "subs %5, #2 \n" // w -= 2 + "beq 4f \n" + + // 1x8 block + "3: \n" + MEMACCESS(1) + "vld1.8 {d0[0]}, [%1], %2 \n" + MEMACCESS(1) + "vld1.8 {d0[1]}, [%1], %2 \n" + MEMACCESS(1) + "vld1.8 {d0[2]}, [%1], %2 \n" + MEMACCESS(1) + "vld1.8 {d0[3]}, [%1], %2 \n" + MEMACCESS(1) + "vld1.8 {d0[4]}, [%1], %2 \n" + MEMACCESS(1) + "vld1.8 {d0[5]}, [%1], %2 \n" + MEMACCESS(1) + "vld1.8 {d0[6]}, [%1], %2 \n" + MEMACCESS(1) + "vld1.8 {d0[7]}, [%1] \n" + + MEMACCESS(3) + "vst1.64 {d0}, [%3] \n" + + "4: \n" + + : "+r"(src_temp), // %0 + "+r"(src), // %1 + "+r"(src_stride), // %2 + "+r"(dst), // %3 + "+r"(dst_stride), // %4 + "+r"(width) // %5 + : "r"(&kVTbl4x4Transpose) // %6 + : "memory", "cc", "q0", "q1", "q2", "q3" + ); +} + +static uvec8 kVTbl4x4TransposeDi = + { 0, 8, 1, 9, 2, 10, 3, 11, 4, 12, 5, 13, 6, 14, 7, 15 }; + +void TransposeUVWx8_NEON(const uint8* src, int src_stride, + uint8* dst_a, int dst_stride_a, + uint8* dst_b, int dst_stride_b, + int width) { + const uint8* src_temp = NULL; + asm volatile ( + // loops are on blocks of 8. loop will stop when + // counter gets to or below 0. starting the counter + // at w-8 allow for this + "sub %7, #8 \n" + + // handle 8x8 blocks. this should be the majority of the plane + ".p2align 2 \n" + "1: \n" + "mov %0, %1 \n" + + MEMACCESS(0) + "vld2.8 {d0, d1}, [%0], %2 \n" + MEMACCESS(0) + "vld2.8 {d2, d3}, [%0], %2 \n" + MEMACCESS(0) + "vld2.8 {d4, d5}, [%0], %2 \n" + MEMACCESS(0) + "vld2.8 {d6, d7}, [%0], %2 \n" + MEMACCESS(0) + "vld2.8 {d16, d17}, [%0], %2 \n" + MEMACCESS(0) + "vld2.8 {d18, d19}, [%0], %2 \n" + MEMACCESS(0) + "vld2.8 {d20, d21}, [%0], %2 \n" + MEMACCESS(0) + "vld2.8 {d22, d23}, [%0] \n" + + "vtrn.8 q1, q0 \n" + "vtrn.8 q3, q2 \n" + "vtrn.8 q9, q8 \n" + "vtrn.8 q11, q10 \n" + + "vtrn.16 q1, q3 \n" + "vtrn.16 q0, q2 \n" + "vtrn.16 q9, q11 \n" + "vtrn.16 q8, q10 \n" + + "vtrn.32 q1, q9 \n" + "vtrn.32 q0, q8 \n" + "vtrn.32 q3, q11 \n" + "vtrn.32 q2, q10 \n" + + "vrev16.8 q0, q0 \n" + "vrev16.8 q1, q1 \n" + "vrev16.8 q2, q2 \n" + "vrev16.8 q3, q3 \n" + "vrev16.8 q8, q8 \n" + "vrev16.8 q9, q9 \n" + "vrev16.8 q10, q10 \n" + "vrev16.8 q11, q11 \n" + + "mov %0, %3 \n" + + MEMACCESS(0) + "vst1.8 {d2}, [%0], %4 \n" + MEMACCESS(0) + "vst1.8 {d0}, [%0], %4 \n" + MEMACCESS(0) + "vst1.8 {d6}, [%0], %4 \n" + MEMACCESS(0) + "vst1.8 {d4}, [%0], %4 \n" + MEMACCESS(0) + "vst1.8 {d18}, [%0], %4 \n" + MEMACCESS(0) + "vst1.8 {d16}, [%0], %4 \n" + MEMACCESS(0) + "vst1.8 {d22}, [%0], %4 \n" + MEMACCESS(0) + "vst1.8 {d20}, [%0] \n" + + "mov %0, %5 \n" + + MEMACCESS(0) + "vst1.8 {d3}, [%0], %6 \n" + MEMACCESS(0) + "vst1.8 {d1}, [%0], %6 \n" + MEMACCESS(0) + "vst1.8 {d7}, [%0], %6 \n" + MEMACCESS(0) + "vst1.8 {d5}, [%0], %6 \n" + MEMACCESS(0) + "vst1.8 {d19}, [%0], %6 \n" + MEMACCESS(0) + "vst1.8 {d17}, [%0], %6 \n" + MEMACCESS(0) + "vst1.8 {d23}, [%0], %6 \n" + MEMACCESS(0) + "vst1.8 {d21}, [%0] \n" + + "add %1, #8*2 \n" // src += 8*2 + "add %3, %3, %4, lsl #3 \n" // dst_a += 8 * dst_stride_a + "add %5, %5, %6, lsl #3 \n" // dst_b += 8 * dst_stride_b + "subs %7, #8 \n" // w -= 8 + "bge 1b \n" + + // add 8 back to counter. if the result is 0 there are + // no residuals. + "adds %7, #8 \n" + "beq 4f \n" + + // some residual, so between 1 and 7 lines left to transpose + "cmp %7, #2 \n" + "blt 3f \n" + + "cmp %7, #4 \n" + "blt 2f \n" + + // TODO(frkoenig): Clean this up + // 4x8 block + "mov %0, %1 \n" + MEMACCESS(0) + "vld1.64 {d0}, [%0], %2 \n" + MEMACCESS(0) + "vld1.64 {d1}, [%0], %2 \n" + MEMACCESS(0) + "vld1.64 {d2}, [%0], %2 \n" + MEMACCESS(0) + "vld1.64 {d3}, [%0], %2 \n" + MEMACCESS(0) + "vld1.64 {d4}, [%0], %2 \n" + MEMACCESS(0) + "vld1.64 {d5}, [%0], %2 \n" + MEMACCESS(0) + "vld1.64 {d6}, [%0], %2 \n" + MEMACCESS(0) + "vld1.64 {d7}, [%0] \n" + + MEMACCESS(8) + "vld1.8 {q15}, [%8] \n" + + "vtrn.8 q0, q1 \n" + "vtrn.8 q2, q3 \n" + + "vtbl.8 d16, {d0, d1}, d30 \n" + "vtbl.8 d17, {d0, d1}, d31 \n" + "vtbl.8 d18, {d2, d3}, d30 \n" + "vtbl.8 d19, {d2, d3}, d31 \n" + "vtbl.8 d20, {d4, d5}, d30 \n" + "vtbl.8 d21, {d4, d5}, d31 \n" + "vtbl.8 d22, {d6, d7}, d30 \n" + "vtbl.8 d23, {d6, d7}, d31 \n" + + "mov %0, %3 \n" + + MEMACCESS(0) + "vst1.32 {d16[0]}, [%0], %4 \n" + MEMACCESS(0) + "vst1.32 {d16[1]}, [%0], %4 \n" + MEMACCESS(0) + "vst1.32 {d17[0]}, [%0], %4 \n" + MEMACCESS(0) + "vst1.32 {d17[1]}, [%0], %4 \n" + + "add %0, %3, #4 \n" + MEMACCESS(0) + "vst1.32 {d20[0]}, [%0], %4 \n" + MEMACCESS(0) + "vst1.32 {d20[1]}, [%0], %4 \n" + MEMACCESS(0) + "vst1.32 {d21[0]}, [%0], %4 \n" + MEMACCESS(0) + "vst1.32 {d21[1]}, [%0] \n" + + "mov %0, %5 \n" + + MEMACCESS(0) + "vst1.32 {d18[0]}, [%0], %6 \n" + MEMACCESS(0) + "vst1.32 {d18[1]}, [%0], %6 \n" + MEMACCESS(0) + "vst1.32 {d19[0]}, [%0], %6 \n" + MEMACCESS(0) + "vst1.32 {d19[1]}, [%0], %6 \n" + + "add %0, %5, #4 \n" + MEMACCESS(0) + "vst1.32 {d22[0]}, [%0], %6 \n" + MEMACCESS(0) + "vst1.32 {d22[1]}, [%0], %6 \n" + MEMACCESS(0) + "vst1.32 {d23[0]}, [%0], %6 \n" + MEMACCESS(0) + "vst1.32 {d23[1]}, [%0] \n" + + "add %1, #4*2 \n" // src += 4 * 2 + "add %3, %3, %4, lsl #2 \n" // dst_a += 4 * dst_stride_a + "add %5, %5, %6, lsl #2 \n" // dst_b += 4 * dst_stride_b + "subs %7, #4 \n" // w -= 4 + "beq 4f \n" + + // some residual, check to see if it includes a 2x8 block, + // or less + "cmp %7, #2 \n" + "blt 3f \n" + + // 2x8 block + "2: \n" + "mov %0, %1 \n" + MEMACCESS(0) + "vld2.16 {d0[0], d2[0]}, [%0], %2 \n" + MEMACCESS(0) + "vld2.16 {d1[0], d3[0]}, [%0], %2 \n" + MEMACCESS(0) + "vld2.16 {d0[1], d2[1]}, [%0], %2 \n" + MEMACCESS(0) + "vld2.16 {d1[1], d3[1]}, [%0], %2 \n" + MEMACCESS(0) + "vld2.16 {d0[2], d2[2]}, [%0], %2 \n" + MEMACCESS(0) + "vld2.16 {d1[2], d3[2]}, [%0], %2 \n" + MEMACCESS(0) + "vld2.16 {d0[3], d2[3]}, [%0], %2 \n" + MEMACCESS(0) + "vld2.16 {d1[3], d3[3]}, [%0] \n" + + "vtrn.8 d0, d1 \n" + "vtrn.8 d2, d3 \n" + + "mov %0, %3 \n" + + MEMACCESS(0) + "vst1.64 {d0}, [%0], %4 \n" + MEMACCESS(0) + "vst1.64 {d2}, [%0] \n" + + "mov %0, %5 \n" + + MEMACCESS(0) + "vst1.64 {d1}, [%0], %6 \n" + MEMACCESS(0) + "vst1.64 {d3}, [%0] \n" + + "add %1, #2*2 \n" // src += 2 * 2 + "add %3, %3, %4, lsl #1 \n" // dst_a += 2 * dst_stride_a + "add %5, %5, %6, lsl #1 \n" // dst_b += 2 * dst_stride_b + "subs %7, #2 \n" // w -= 2 + "beq 4f \n" + + // 1x8 block + "3: \n" + MEMACCESS(1) + "vld2.8 {d0[0], d1[0]}, [%1], %2 \n" + MEMACCESS(1) + "vld2.8 {d0[1], d1[1]}, [%1], %2 \n" + MEMACCESS(1) + "vld2.8 {d0[2], d1[2]}, [%1], %2 \n" + MEMACCESS(1) + "vld2.8 {d0[3], d1[3]}, [%1], %2 \n" + MEMACCESS(1) + "vld2.8 {d0[4], d1[4]}, [%1], %2 \n" + MEMACCESS(1) + "vld2.8 {d0[5], d1[5]}, [%1], %2 \n" + MEMACCESS(1) + "vld2.8 {d0[6], d1[6]}, [%1], %2 \n" + MEMACCESS(1) + "vld2.8 {d0[7], d1[7]}, [%1] \n" + + MEMACCESS(3) + "vst1.64 {d0}, [%3] \n" + MEMACCESS(5) + "vst1.64 {d1}, [%5] \n" + + "4: \n" + + : "+r"(src_temp), // %0 + "+r"(src), // %1 + "+r"(src_stride), // %2 + "+r"(dst_a), // %3 + "+r"(dst_stride_a), // %4 + "+r"(dst_b), // %5 + "+r"(dst_stride_b), // %6 + "+r"(width) // %7 + : "r"(&kVTbl4x4TransposeDi) // %8 + : "memory", "cc", + "q0", "q1", "q2", "q3", "q8", "q9", "q10", "q11" + ); +} +#endif // defined(__ARM_NEON__) && !defined(__aarch64__) + +#ifdef __cplusplus +} // extern "C" +} // namespace libyuv +#endif diff --git a/media/libaom/src/third_party/libyuv/source/rotate_neon64.cc b/media/libaom/src/third_party/libyuv/source/rotate_neon64.cc new file mode 100644 index 000000000..f52c082b3 --- /dev/null +++ b/media/libaom/src/third_party/libyuv/source/rotate_neon64.cc @@ -0,0 +1,543 @@ +/* + * Copyright 2014 The LibYuv Project Authors. All rights reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#include "libyuv/row.h" +#include "libyuv/rotate_row.h" + +#include "libyuv/basic_types.h" + +#ifdef __cplusplus +namespace libyuv { +extern "C" { +#endif + +// This module is for GCC Neon armv8 64 bit. +#if !defined(LIBYUV_DISABLE_NEON) && defined(__aarch64__) + +static uvec8 kVTbl4x4Transpose = + { 0, 4, 8, 12, 1, 5, 9, 13, 2, 6, 10, 14, 3, 7, 11, 15 }; + +void TransposeWx8_NEON(const uint8* src, int src_stride, + uint8* dst, int dst_stride, int width) { + const uint8* src_temp = NULL; + int64 width64 = (int64) width; // Work around clang 3.4 warning. + asm volatile ( + // loops are on blocks of 8. loop will stop when + // counter gets to or below 0. starting the counter + // at w-8 allow for this + "sub %3, %3, #8 \n" + + // handle 8x8 blocks. this should be the majority of the plane + "1: \n" + "mov %0, %1 \n" + + MEMACCESS(0) + "ld1 {v0.8b}, [%0], %5 \n" + MEMACCESS(0) + "ld1 {v1.8b}, [%0], %5 \n" + MEMACCESS(0) + "ld1 {v2.8b}, [%0], %5 \n" + MEMACCESS(0) + "ld1 {v3.8b}, [%0], %5 \n" + MEMACCESS(0) + "ld1 {v4.8b}, [%0], %5 \n" + MEMACCESS(0) + "ld1 {v5.8b}, [%0], %5 \n" + MEMACCESS(0) + "ld1 {v6.8b}, [%0], %5 \n" + MEMACCESS(0) + "ld1 {v7.8b}, [%0] \n" + + "trn2 v16.8b, v0.8b, v1.8b \n" + "trn1 v17.8b, v0.8b, v1.8b \n" + "trn2 v18.8b, v2.8b, v3.8b \n" + "trn1 v19.8b, v2.8b, v3.8b \n" + "trn2 v20.8b, v4.8b, v5.8b \n" + "trn1 v21.8b, v4.8b, v5.8b \n" + "trn2 v22.8b, v6.8b, v7.8b \n" + "trn1 v23.8b, v6.8b, v7.8b \n" + + "trn2 v3.4h, v17.4h, v19.4h \n" + "trn1 v1.4h, v17.4h, v19.4h \n" + "trn2 v2.4h, v16.4h, v18.4h \n" + "trn1 v0.4h, v16.4h, v18.4h \n" + "trn2 v7.4h, v21.4h, v23.4h \n" + "trn1 v5.4h, v21.4h, v23.4h \n" + "trn2 v6.4h, v20.4h, v22.4h \n" + "trn1 v4.4h, v20.4h, v22.4h \n" + + "trn2 v21.2s, v1.2s, v5.2s \n" + "trn1 v17.2s, v1.2s, v5.2s \n" + "trn2 v20.2s, v0.2s, v4.2s \n" + "trn1 v16.2s, v0.2s, v4.2s \n" + "trn2 v23.2s, v3.2s, v7.2s \n" + "trn1 v19.2s, v3.2s, v7.2s \n" + "trn2 v22.2s, v2.2s, v6.2s \n" + "trn1 v18.2s, v2.2s, v6.2s \n" + + "mov %0, %2 \n" + + MEMACCESS(0) + "st1 {v17.8b}, [%0], %6 \n" + MEMACCESS(0) + "st1 {v16.8b}, [%0], %6 \n" + MEMACCESS(0) + "st1 {v19.8b}, [%0], %6 \n" + MEMACCESS(0) + "st1 {v18.8b}, [%0], %6 \n" + MEMACCESS(0) + "st1 {v21.8b}, [%0], %6 \n" + MEMACCESS(0) + "st1 {v20.8b}, [%0], %6 \n" + MEMACCESS(0) + "st1 {v23.8b}, [%0], %6 \n" + MEMACCESS(0) + "st1 {v22.8b}, [%0] \n" + + "add %1, %1, #8 \n" // src += 8 + "add %2, %2, %6, lsl #3 \n" // dst += 8 * dst_stride + "subs %3, %3, #8 \n" // w -= 8 + "b.ge 1b \n" + + // add 8 back to counter. if the result is 0 there are + // no residuals. + "adds %3, %3, #8 \n" + "b.eq 4f \n" + + // some residual, so between 1 and 7 lines left to transpose + "cmp %3, #2 \n" + "b.lt 3f \n" + + "cmp %3, #4 \n" + "b.lt 2f \n" + + // 4x8 block + "mov %0, %1 \n" + MEMACCESS(0) + "ld1 {v0.s}[0], [%0], %5 \n" + MEMACCESS(0) + "ld1 {v0.s}[1], [%0], %5 \n" + MEMACCESS(0) + "ld1 {v0.s}[2], [%0], %5 \n" + MEMACCESS(0) + "ld1 {v0.s}[3], [%0], %5 \n" + MEMACCESS(0) + "ld1 {v1.s}[0], [%0], %5 \n" + MEMACCESS(0) + "ld1 {v1.s}[1], [%0], %5 \n" + MEMACCESS(0) + "ld1 {v1.s}[2], [%0], %5 \n" + MEMACCESS(0) + "ld1 {v1.s}[3], [%0] \n" + + "mov %0, %2 \n" + + MEMACCESS(4) + "ld1 {v2.16b}, [%4] \n" + + "tbl v3.16b, {v0.16b}, v2.16b \n" + "tbl v0.16b, {v1.16b}, v2.16b \n" + + // TODO(frkoenig): Rework shuffle above to + // write out with 4 instead of 8 writes. + MEMACCESS(0) + "st1 {v3.s}[0], [%0], %6 \n" + MEMACCESS(0) + "st1 {v3.s}[1], [%0], %6 \n" + MEMACCESS(0) + "st1 {v3.s}[2], [%0], %6 \n" + MEMACCESS(0) + "st1 {v3.s}[3], [%0] \n" + + "add %0, %2, #4 \n" + MEMACCESS(0) + "st1 {v0.s}[0], [%0], %6 \n" + MEMACCESS(0) + "st1 {v0.s}[1], [%0], %6 \n" + MEMACCESS(0) + "st1 {v0.s}[2], [%0], %6 \n" + MEMACCESS(0) + "st1 {v0.s}[3], [%0] \n" + + "add %1, %1, #4 \n" // src += 4 + "add %2, %2, %6, lsl #2 \n" // dst += 4 * dst_stride + "subs %3, %3, #4 \n" // w -= 4 + "b.eq 4f \n" + + // some residual, check to see if it includes a 2x8 block, + // or less + "cmp %3, #2 \n" + "b.lt 3f \n" + + // 2x8 block + "2: \n" + "mov %0, %1 \n" + MEMACCESS(0) + "ld1 {v0.h}[0], [%0], %5 \n" + MEMACCESS(0) + "ld1 {v1.h}[0], [%0], %5 \n" + MEMACCESS(0) + "ld1 {v0.h}[1], [%0], %5 \n" + MEMACCESS(0) + "ld1 {v1.h}[1], [%0], %5 \n" + MEMACCESS(0) + "ld1 {v0.h}[2], [%0], %5 \n" + MEMACCESS(0) + "ld1 {v1.h}[2], [%0], %5 \n" + MEMACCESS(0) + "ld1 {v0.h}[3], [%0], %5 \n" + MEMACCESS(0) + "ld1 {v1.h}[3], [%0] \n" + + "trn2 v2.8b, v0.8b, v1.8b \n" + "trn1 v3.8b, v0.8b, v1.8b \n" + + "mov %0, %2 \n" + + MEMACCESS(0) + "st1 {v3.8b}, [%0], %6 \n" + MEMACCESS(0) + "st1 {v2.8b}, [%0] \n" + + "add %1, %1, #2 \n" // src += 2 + "add %2, %2, %6, lsl #1 \n" // dst += 2 * dst_stride + "subs %3, %3, #2 \n" // w -= 2 + "b.eq 4f \n" + + // 1x8 block + "3: \n" + MEMACCESS(1) + "ld1 {v0.b}[0], [%1], %5 \n" + MEMACCESS(1) + "ld1 {v0.b}[1], [%1], %5 \n" + MEMACCESS(1) + "ld1 {v0.b}[2], [%1], %5 \n" + MEMACCESS(1) + "ld1 {v0.b}[3], [%1], %5 \n" + MEMACCESS(1) + "ld1 {v0.b}[4], [%1], %5 \n" + MEMACCESS(1) + "ld1 {v0.b}[5], [%1], %5 \n" + MEMACCESS(1) + "ld1 {v0.b}[6], [%1], %5 \n" + MEMACCESS(1) + "ld1 {v0.b}[7], [%1] \n" + + MEMACCESS(2) + "st1 {v0.8b}, [%2] \n" + + "4: \n" + + : "+r"(src_temp), // %0 + "+r"(src), // %1 + "+r"(dst), // %2 + "+r"(width64) // %3 + : "r"(&kVTbl4x4Transpose), // %4 + "r"(static_cast<ptrdiff_t>(src_stride)), // %5 + "r"(static_cast<ptrdiff_t>(dst_stride)) // %6 + : "memory", "cc", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16", + "v17", "v18", "v19", "v20", "v21", "v22", "v23" + ); +} + +static uint8 kVTbl4x4TransposeDi[32] = + { 0, 16, 32, 48, 2, 18, 34, 50, 4, 20, 36, 52, 6, 22, 38, 54, + 1, 17, 33, 49, 3, 19, 35, 51, 5, 21, 37, 53, 7, 23, 39, 55}; + +void TransposeUVWx8_NEON(const uint8* src, int src_stride, + uint8* dst_a, int dst_stride_a, + uint8* dst_b, int dst_stride_b, + int width) { + const uint8* src_temp = NULL; + int64 width64 = (int64) width; // Work around clang 3.4 warning. + asm volatile ( + // loops are on blocks of 8. loop will stop when + // counter gets to or below 0. starting the counter + // at w-8 allow for this + "sub %4, %4, #8 \n" + + // handle 8x8 blocks. this should be the majority of the plane + "1: \n" + "mov %0, %1 \n" + + MEMACCESS(0) + "ld1 {v0.16b}, [%0], %5 \n" + MEMACCESS(0) + "ld1 {v1.16b}, [%0], %5 \n" + MEMACCESS(0) + "ld1 {v2.16b}, [%0], %5 \n" + MEMACCESS(0) + "ld1 {v3.16b}, [%0], %5 \n" + MEMACCESS(0) + "ld1 {v4.16b}, [%0], %5 \n" + MEMACCESS(0) + "ld1 {v5.16b}, [%0], %5 \n" + MEMACCESS(0) + "ld1 {v6.16b}, [%0], %5 \n" + MEMACCESS(0) + "ld1 {v7.16b}, [%0] \n" + + "trn1 v16.16b, v0.16b, v1.16b \n" + "trn2 v17.16b, v0.16b, v1.16b \n" + "trn1 v18.16b, v2.16b, v3.16b \n" + "trn2 v19.16b, v2.16b, v3.16b \n" + "trn1 v20.16b, v4.16b, v5.16b \n" + "trn2 v21.16b, v4.16b, v5.16b \n" + "trn1 v22.16b, v6.16b, v7.16b \n" + "trn2 v23.16b, v6.16b, v7.16b \n" + + "trn1 v0.8h, v16.8h, v18.8h \n" + "trn2 v1.8h, v16.8h, v18.8h \n" + "trn1 v2.8h, v20.8h, v22.8h \n" + "trn2 v3.8h, v20.8h, v22.8h \n" + "trn1 v4.8h, v17.8h, v19.8h \n" + "trn2 v5.8h, v17.8h, v19.8h \n" + "trn1 v6.8h, v21.8h, v23.8h \n" + "trn2 v7.8h, v21.8h, v23.8h \n" + + "trn1 v16.4s, v0.4s, v2.4s \n" + "trn2 v17.4s, v0.4s, v2.4s \n" + "trn1 v18.4s, v1.4s, v3.4s \n" + "trn2 v19.4s, v1.4s, v3.4s \n" + "trn1 v20.4s, v4.4s, v6.4s \n" + "trn2 v21.4s, v4.4s, v6.4s \n" + "trn1 v22.4s, v5.4s, v7.4s \n" + "trn2 v23.4s, v5.4s, v7.4s \n" + + "mov %0, %2 \n" + + MEMACCESS(0) + "st1 {v16.d}[0], [%0], %6 \n" + MEMACCESS(0) + "st1 {v18.d}[0], [%0], %6 \n" + MEMACCESS(0) + "st1 {v17.d}[0], [%0], %6 \n" + MEMACCESS(0) + "st1 {v19.d}[0], [%0], %6 \n" + MEMACCESS(0) + "st1 {v16.d}[1], [%0], %6 \n" + MEMACCESS(0) + "st1 {v18.d}[1], [%0], %6 \n" + MEMACCESS(0) + "st1 {v17.d}[1], [%0], %6 \n" + MEMACCESS(0) + "st1 {v19.d}[1], [%0] \n" + + "mov %0, %3 \n" + + MEMACCESS(0) + "st1 {v20.d}[0], [%0], %7 \n" + MEMACCESS(0) + "st1 {v22.d}[0], [%0], %7 \n" + MEMACCESS(0) + "st1 {v21.d}[0], [%0], %7 \n" + MEMACCESS(0) + "st1 {v23.d}[0], [%0], %7 \n" + MEMACCESS(0) + "st1 {v20.d}[1], [%0], %7 \n" + MEMACCESS(0) + "st1 {v22.d}[1], [%0], %7 \n" + MEMACCESS(0) + "st1 {v21.d}[1], [%0], %7 \n" + MEMACCESS(0) + "st1 {v23.d}[1], [%0] \n" + + "add %1, %1, #16 \n" // src += 8*2 + "add %2, %2, %6, lsl #3 \n" // dst_a += 8 * dst_stride_a + "add %3, %3, %7, lsl #3 \n" // dst_b += 8 * dst_stride_b + "subs %4, %4, #8 \n" // w -= 8 + "b.ge 1b \n" + + // add 8 back to counter. if the result is 0 there are + // no residuals. + "adds %4, %4, #8 \n" + "b.eq 4f \n" + + // some residual, so between 1 and 7 lines left to transpose + "cmp %4, #2 \n" + "b.lt 3f \n" + + "cmp %4, #4 \n" + "b.lt 2f \n" + + // TODO(frkoenig): Clean this up + // 4x8 block + "mov %0, %1 \n" + MEMACCESS(0) + "ld1 {v0.8b}, [%0], %5 \n" + MEMACCESS(0) + "ld1 {v1.8b}, [%0], %5 \n" + MEMACCESS(0) + "ld1 {v2.8b}, [%0], %5 \n" + MEMACCESS(0) + "ld1 {v3.8b}, [%0], %5 \n" + MEMACCESS(0) + "ld1 {v4.8b}, [%0], %5 \n" + MEMACCESS(0) + "ld1 {v5.8b}, [%0], %5 \n" + MEMACCESS(0) + "ld1 {v6.8b}, [%0], %5 \n" + MEMACCESS(0) + "ld1 {v7.8b}, [%0] \n" + + MEMACCESS(8) + "ld1 {v30.16b}, [%8], #16 \n" + "ld1 {v31.16b}, [%8] \n" + + "tbl v16.16b, {v0.16b, v1.16b, v2.16b, v3.16b}, v30.16b \n" + "tbl v17.16b, {v0.16b, v1.16b, v2.16b, v3.16b}, v31.16b \n" + "tbl v18.16b, {v4.16b, v5.16b, v6.16b, v7.16b}, v30.16b \n" + "tbl v19.16b, {v4.16b, v5.16b, v6.16b, v7.16b}, v31.16b \n" + + "mov %0, %2 \n" + + MEMACCESS(0) + "st1 {v16.s}[0], [%0], %6 \n" + MEMACCESS(0) + "st1 {v16.s}[1], [%0], %6 \n" + MEMACCESS(0) + "st1 {v16.s}[2], [%0], %6 \n" + MEMACCESS(0) + "st1 {v16.s}[3], [%0], %6 \n" + + "add %0, %2, #4 \n" + MEMACCESS(0) + "st1 {v18.s}[0], [%0], %6 \n" + MEMACCESS(0) + "st1 {v18.s}[1], [%0], %6 \n" + MEMACCESS(0) + "st1 {v18.s}[2], [%0], %6 \n" + MEMACCESS(0) + "st1 {v18.s}[3], [%0] \n" + + "mov %0, %3 \n" + + MEMACCESS(0) + "st1 {v17.s}[0], [%0], %7 \n" + MEMACCESS(0) + "st1 {v17.s}[1], [%0], %7 \n" + MEMACCESS(0) + "st1 {v17.s}[2], [%0], %7 \n" + MEMACCESS(0) + "st1 {v17.s}[3], [%0], %7 \n" + + "add %0, %3, #4 \n" + MEMACCESS(0) + "st1 {v19.s}[0], [%0], %7 \n" + MEMACCESS(0) + "st1 {v19.s}[1], [%0], %7 \n" + MEMACCESS(0) + "st1 {v19.s}[2], [%0], %7 \n" + MEMACCESS(0) + "st1 {v19.s}[3], [%0] \n" + + "add %1, %1, #8 \n" // src += 4 * 2 + "add %2, %2, %6, lsl #2 \n" // dst_a += 4 * dst_stride_a + "add %3, %3, %7, lsl #2 \n" // dst_b += 4 * dst_stride_b + "subs %4, %4, #4 \n" // w -= 4 + "b.eq 4f \n" + + // some residual, check to see if it includes a 2x8 block, + // or less + "cmp %4, #2 \n" + "b.lt 3f \n" + + // 2x8 block + "2: \n" + "mov %0, %1 \n" + MEMACCESS(0) + "ld2 {v0.h, v1.h}[0], [%0], %5 \n" + MEMACCESS(0) + "ld2 {v2.h, v3.h}[0], [%0], %5 \n" + MEMACCESS(0) + "ld2 {v0.h, v1.h}[1], [%0], %5 \n" + MEMACCESS(0) + "ld2 {v2.h, v3.h}[1], [%0], %5 \n" + MEMACCESS(0) + "ld2 {v0.h, v1.h}[2], [%0], %5 \n" + MEMACCESS(0) + "ld2 {v2.h, v3.h}[2], [%0], %5 \n" + MEMACCESS(0) + "ld2 {v0.h, v1.h}[3], [%0], %5 \n" + MEMACCESS(0) + "ld2 {v2.h, v3.h}[3], [%0] \n" + + "trn1 v4.8b, v0.8b, v2.8b \n" + "trn2 v5.8b, v0.8b, v2.8b \n" + "trn1 v6.8b, v1.8b, v3.8b \n" + "trn2 v7.8b, v1.8b, v3.8b \n" + + "mov %0, %2 \n" + + MEMACCESS(0) + "st1 {v4.d}[0], [%0], %6 \n" + MEMACCESS(0) + "st1 {v6.d}[0], [%0] \n" + + "mov %0, %3 \n" + + MEMACCESS(0) + "st1 {v5.d}[0], [%0], %7 \n" + MEMACCESS(0) + "st1 {v7.d}[0], [%0] \n" + + "add %1, %1, #4 \n" // src += 2 * 2 + "add %2, %2, %6, lsl #1 \n" // dst_a += 2 * dst_stride_a + "add %3, %3, %7, lsl #1 \n" // dst_b += 2 * dst_stride_b + "subs %4, %4, #2 \n" // w -= 2 + "b.eq 4f \n" + + // 1x8 block + "3: \n" + MEMACCESS(1) + "ld2 {v0.b, v1.b}[0], [%1], %5 \n" + MEMACCESS(1) + "ld2 {v0.b, v1.b}[1], [%1], %5 \n" + MEMACCESS(1) + "ld2 {v0.b, v1.b}[2], [%1], %5 \n" + MEMACCESS(1) + "ld2 {v0.b, v1.b}[3], [%1], %5 \n" + MEMACCESS(1) + "ld2 {v0.b, v1.b}[4], [%1], %5 \n" + MEMACCESS(1) + "ld2 {v0.b, v1.b}[5], [%1], %5 \n" + MEMACCESS(1) + "ld2 {v0.b, v1.b}[6], [%1], %5 \n" + MEMACCESS(1) + "ld2 {v0.b, v1.b}[7], [%1] \n" + + MEMACCESS(2) + "st1 {v0.d}[0], [%2] \n" + MEMACCESS(3) + "st1 {v1.d}[0], [%3] \n" + + "4: \n" + + : "+r"(src_temp), // %0 + "+r"(src), // %1 + "+r"(dst_a), // %2 + "+r"(dst_b), // %3 + "+r"(width64) // %4 + : "r"(static_cast<ptrdiff_t>(src_stride)), // %5 + "r"(static_cast<ptrdiff_t>(dst_stride_a)), // %6 + "r"(static_cast<ptrdiff_t>(dst_stride_b)), // %7 + "r"(&kVTbl4x4TransposeDi) // %8 + : "memory", "cc", + "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", + "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", + "v30", "v31" + ); +} +#endif // !defined(LIBYUV_DISABLE_NEON) && defined(__aarch64__) + +#ifdef __cplusplus +} // extern "C" +} // namespace libyuv +#endif diff --git a/media/libaom/src/third_party/libyuv/source/rotate_win.cc b/media/libaom/src/third_party/libyuv/source/rotate_win.cc new file mode 100644 index 000000000..2760066df --- /dev/null +++ b/media/libaom/src/third_party/libyuv/source/rotate_win.cc @@ -0,0 +1,248 @@ +/* + * Copyright 2013 The LibYuv Project Authors. All rights reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#include "libyuv/row.h" +#include "libyuv/rotate_row.h" + +#ifdef __cplusplus +namespace libyuv { +extern "C" { +#endif + +// This module is for Visual C x86. +#if !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86) && \ + defined(_MSC_VER) && !defined(__clang__) + +__declspec(naked) +void TransposeWx8_SSSE3(const uint8* src, int src_stride, + uint8* dst, int dst_stride, int width) { + __asm { + push edi + push esi + push ebp + mov eax, [esp + 12 + 4] // src + mov edi, [esp + 12 + 8] // src_stride + mov edx, [esp + 12 + 12] // dst + mov esi, [esp + 12 + 16] // dst_stride + mov ecx, [esp + 12 + 20] // width + + // Read in the data from the source pointer. + // First round of bit swap. + align 4 + convertloop: + movq xmm0, qword ptr [eax] + lea ebp, [eax + 8] + movq xmm1, qword ptr [eax + edi] + lea eax, [eax + 2 * edi] + punpcklbw xmm0, xmm1 + movq xmm2, qword ptr [eax] + movdqa xmm1, xmm0 + palignr xmm1, xmm1, 8 + movq xmm3, qword ptr [eax + edi] + lea eax, [eax + 2 * edi] + punpcklbw xmm2, xmm3 + movdqa xmm3, xmm2 + movq xmm4, qword ptr [eax] + palignr xmm3, xmm3, 8 + movq xmm5, qword ptr [eax + edi] + punpcklbw xmm4, xmm5 + lea eax, [eax + 2 * edi] + movdqa xmm5, xmm4 + movq xmm6, qword ptr [eax] + palignr xmm5, xmm5, 8 + movq xmm7, qword ptr [eax + edi] + punpcklbw xmm6, xmm7 + mov eax, ebp + movdqa xmm7, xmm6 + palignr xmm7, xmm7, 8 + // Second round of bit swap. + punpcklwd xmm0, xmm2 + punpcklwd xmm1, xmm3 + movdqa xmm2, xmm0 + movdqa xmm3, xmm1 + palignr xmm2, xmm2, 8 + palignr xmm3, xmm3, 8 + punpcklwd xmm4, xmm6 + punpcklwd xmm5, xmm7 + movdqa xmm6, xmm4 + movdqa xmm7, xmm5 + palignr xmm6, xmm6, 8 + palignr xmm7, xmm7, 8 + // Third round of bit swap. + // Write to the destination pointer. + punpckldq xmm0, xmm4 + movq qword ptr [edx], xmm0 + movdqa xmm4, xmm0 + palignr xmm4, xmm4, 8 + movq qword ptr [edx + esi], xmm4 + lea edx, [edx + 2 * esi] + punpckldq xmm2, xmm6 + movdqa xmm6, xmm2 + palignr xmm6, xmm6, 8 + movq qword ptr [edx], xmm2 + punpckldq xmm1, xmm5 + movq qword ptr [edx + esi], xmm6 + lea edx, [edx + 2 * esi] + movdqa xmm5, xmm1 + movq qword ptr [edx], xmm1 + palignr xmm5, xmm5, 8 + punpckldq xmm3, xmm7 + movq qword ptr [edx + esi], xmm5 + lea edx, [edx + 2 * esi] + movq qword ptr [edx], xmm3 + movdqa xmm7, xmm3 + palignr xmm7, xmm7, 8 + sub ecx, 8 + movq qword ptr [edx + esi], xmm7 + lea edx, [edx + 2 * esi] + jg convertloop + + pop ebp + pop esi + pop edi + ret + } +} + +__declspec(naked) +void TransposeUVWx8_SSE2(const uint8* src, int src_stride, + uint8* dst_a, int dst_stride_a, + uint8* dst_b, int dst_stride_b, + int w) { + __asm { + push ebx + push esi + push edi + push ebp + mov eax, [esp + 16 + 4] // src + mov edi, [esp + 16 + 8] // src_stride + mov edx, [esp + 16 + 12] // dst_a + mov esi, [esp + 16 + 16] // dst_stride_a + mov ebx, [esp + 16 + 20] // dst_b + mov ebp, [esp + 16 + 24] // dst_stride_b + mov ecx, esp + sub esp, 4 + 16 + and esp, ~15 + mov [esp + 16], ecx + mov ecx, [ecx + 16 + 28] // w + + align 4 + convertloop: + // Read in the data from the source pointer. + // First round of bit swap. + movdqu xmm0, [eax] + movdqu xmm1, [eax + edi] + lea eax, [eax + 2 * edi] + movdqa xmm7, xmm0 // use xmm7 as temp register. + punpcklbw xmm0, xmm1 + punpckhbw xmm7, xmm1 + movdqa xmm1, xmm7 + movdqu xmm2, [eax] + movdqu xmm3, [eax + edi] + lea eax, [eax + 2 * edi] + movdqa xmm7, xmm2 + punpcklbw xmm2, xmm3 + punpckhbw xmm7, xmm3 + movdqa xmm3, xmm7 + movdqu xmm4, [eax] + movdqu xmm5, [eax + edi] + lea eax, [eax + 2 * edi] + movdqa xmm7, xmm4 + punpcklbw xmm4, xmm5 + punpckhbw xmm7, xmm5 + movdqa xmm5, xmm7 + movdqu xmm6, [eax] + movdqu xmm7, [eax + edi] + lea eax, [eax + 2 * edi] + movdqu [esp], xmm5 // backup xmm5 + neg edi + movdqa xmm5, xmm6 // use xmm5 as temp register. + punpcklbw xmm6, xmm7 + punpckhbw xmm5, xmm7 + movdqa xmm7, xmm5 + lea eax, [eax + 8 * edi + 16] + neg edi + // Second round of bit swap. + movdqa xmm5, xmm0 + punpcklwd xmm0, xmm2 + punpckhwd xmm5, xmm2 + movdqa xmm2, xmm5 + movdqa xmm5, xmm1 + punpcklwd xmm1, xmm3 + punpckhwd xmm5, xmm3 + movdqa xmm3, xmm5 + movdqa xmm5, xmm4 + punpcklwd xmm4, xmm6 + punpckhwd xmm5, xmm6 + movdqa xmm6, xmm5 + movdqu xmm5, [esp] // restore xmm5 + movdqu [esp], xmm6 // backup xmm6 + movdqa xmm6, xmm5 // use xmm6 as temp register. + punpcklwd xmm5, xmm7 + punpckhwd xmm6, xmm7 + movdqa xmm7, xmm6 + // Third round of bit swap. + // Write to the destination pointer. + movdqa xmm6, xmm0 + punpckldq xmm0, xmm4 + punpckhdq xmm6, xmm4 + movdqa xmm4, xmm6 + movdqu xmm6, [esp] // restore xmm6 + movlpd qword ptr [edx], xmm0 + movhpd qword ptr [ebx], xmm0 + movlpd qword ptr [edx + esi], xmm4 + lea edx, [edx + 2 * esi] + movhpd qword ptr [ebx + ebp], xmm4 + lea ebx, [ebx + 2 * ebp] + movdqa xmm0, xmm2 // use xmm0 as the temp register. + punpckldq xmm2, xmm6 + movlpd qword ptr [edx], xmm2 + movhpd qword ptr [ebx], xmm2 + punpckhdq xmm0, xmm6 + movlpd qword ptr [edx + esi], xmm0 + lea edx, [edx + 2 * esi] + movhpd qword ptr [ebx + ebp], xmm0 + lea ebx, [ebx + 2 * ebp] + movdqa xmm0, xmm1 // use xmm0 as the temp register. + punpckldq xmm1, xmm5 + movlpd qword ptr [edx], xmm1 + movhpd qword ptr [ebx], xmm1 + punpckhdq xmm0, xmm5 + movlpd qword ptr [edx + esi], xmm0 + lea edx, [edx + 2 * esi] + movhpd qword ptr [ebx + ebp], xmm0 + lea ebx, [ebx + 2 * ebp] + movdqa xmm0, xmm3 // use xmm0 as the temp register. + punpckldq xmm3, xmm7 + movlpd qword ptr [edx], xmm3 + movhpd qword ptr [ebx], xmm3 + punpckhdq xmm0, xmm7 + sub ecx, 8 + movlpd qword ptr [edx + esi], xmm0 + lea edx, [edx + 2 * esi] + movhpd qword ptr [ebx + ebp], xmm0 + lea ebx, [ebx + 2 * ebp] + jg convertloop + + mov esp, [esp + 16] + pop ebp + pop edi + pop esi + pop ebx + ret + } +} + +#endif // !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86) + +#ifdef __cplusplus +} // extern "C" +} // namespace libyuv +#endif diff --git a/media/libaom/src/third_party/libyuv/source/row_any.cc b/media/libaom/src/third_party/libyuv/source/row_any.cc new file mode 100644 index 000000000..1cb1f6b93 --- /dev/null +++ b/media/libaom/src/third_party/libyuv/source/row_any.cc @@ -0,0 +1,680 @@ +/* + * Copyright 2012 The LibYuv Project Authors. All rights reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#include "libyuv/row.h" + +#include <string.h> // For memset. + +#include "libyuv/basic_types.h" + +#ifdef __cplusplus +namespace libyuv { +extern "C" { +#endif + +// Subsampled source needs to be increase by 1 of not even. +#define SS(width, shift) (((width) + (1 << (shift)) - 1) >> (shift)) + +// Any 3 planes to 1. +#define ANY31(NAMEANY, ANY_SIMD, UVSHIFT, DUVSHIFT, BPP, MASK) \ + void NAMEANY(const uint8* y_buf, const uint8* u_buf, const uint8* v_buf, \ + uint8* dst_ptr, int width) { \ + SIMD_ALIGNED(uint8 temp[64 * 4]); \ + memset(temp, 0, 64 * 3); /* for YUY2 and msan */ \ + int r = width & MASK; \ + int n = width & ~MASK; \ + if (n > 0) { \ + ANY_SIMD(y_buf, u_buf, v_buf, dst_ptr, n); \ + } \ + memcpy(temp, y_buf + n, r); \ + memcpy(temp + 64, u_buf + (n >> UVSHIFT), SS(r, UVSHIFT)); \ + memcpy(temp + 128, v_buf + (n >> UVSHIFT), SS(r, UVSHIFT)); \ + ANY_SIMD(temp, temp + 64, temp + 128, temp + 192, MASK + 1); \ + memcpy(dst_ptr + (n >> DUVSHIFT) * BPP, temp + 192, \ + SS(r, DUVSHIFT) * BPP); \ + } + +#ifdef HAS_I422TOARGBROW_SSSE3 +ANY31(I422ToARGBRow_Any_SSSE3, I422ToARGBRow_SSSE3, 1, 0, 4, 7) +#endif +#ifdef HAS_I444TOARGBROW_SSSE3 +ANY31(I444ToARGBRow_Any_SSSE3, I444ToARGBRow_SSSE3, 0, 0, 4, 7) +ANY31(I411ToARGBRow_Any_SSSE3, I411ToARGBRow_SSSE3, 2, 0, 4, 7) +ANY31(I422ToBGRARow_Any_SSSE3, I422ToBGRARow_SSSE3, 1, 0, 4, 7) +ANY31(I422ToABGRRow_Any_SSSE3, I422ToABGRRow_SSSE3, 1, 0, 4, 7) +ANY31(I422ToRGBARow_Any_SSSE3, I422ToRGBARow_SSSE3, 1, 0, 4, 7) +ANY31(I422ToARGB4444Row_Any_SSSE3, I422ToARGB4444Row_SSSE3, 1, 0, 2, 7) +ANY31(I422ToARGB1555Row_Any_SSSE3, I422ToARGB1555Row_SSSE3, 1, 0, 2, 7) +ANY31(I422ToRGB565Row_Any_SSSE3, I422ToRGB565Row_SSSE3, 1, 0, 2, 7) +ANY31(I422ToRGB24Row_Any_SSSE3, I422ToRGB24Row_SSSE3, 1, 0, 3, 7) +ANY31(I422ToRAWRow_Any_SSSE3, I422ToRAWRow_SSSE3, 1, 0, 3, 7) +ANY31(I422ToYUY2Row_Any_SSE2, I422ToYUY2Row_SSE2, 1, 1, 4, 15) +ANY31(I422ToUYVYRow_Any_SSE2, I422ToUYVYRow_SSE2, 1, 1, 4, 15) +#endif // HAS_I444TOARGBROW_SSSE3 +#ifdef HAS_I422TORGB24ROW_AVX2 +ANY31(I422ToRGB24Row_Any_AVX2, I422ToRGB24Row_AVX2, 1, 0, 3, 15) +#endif +#ifdef HAS_I422TORAWROW_AVX2 +ANY31(I422ToRAWRow_Any_AVX2, I422ToRAWRow_AVX2, 1, 0, 3, 15) +#endif +#ifdef HAS_J422TOARGBROW_SSSE3 +ANY31(J422ToARGBRow_Any_SSSE3, J422ToARGBRow_SSSE3, 1, 0, 4, 7) +#endif +#ifdef HAS_J422TOARGBROW_AVX2 +ANY31(J422ToARGBRow_Any_AVX2, J422ToARGBRow_AVX2, 1, 0, 4, 15) +#endif +#ifdef HAS_I422TOARGBROW_AVX2 +ANY31(I422ToARGBRow_Any_AVX2, I422ToARGBRow_AVX2, 1, 0, 4, 15) +#endif +#ifdef HAS_I422TOBGRAROW_AVX2 +ANY31(I422ToBGRARow_Any_AVX2, I422ToBGRARow_AVX2, 1, 0, 4, 15) +#endif +#ifdef HAS_I422TORGBAROW_AVX2 +ANY31(I422ToRGBARow_Any_AVX2, I422ToRGBARow_AVX2, 1, 0, 4, 15) +#endif +#ifdef HAS_I422TOABGRROW_AVX2 +ANY31(I422ToABGRRow_Any_AVX2, I422ToABGRRow_AVX2, 1, 0, 4, 15) +#endif +#ifdef HAS_I444TOARGBROW_AVX2 +ANY31(I444ToARGBRow_Any_AVX2, I444ToARGBRow_AVX2, 0, 0, 4, 15) +#endif +#ifdef HAS_I411TOARGBROW_AVX2 +ANY31(I411ToARGBRow_Any_AVX2, I411ToARGBRow_AVX2, 2, 0, 4, 15) +#endif +#ifdef HAS_I422TOARGB4444ROW_AVX2 +ANY31(I422ToARGB4444Row_Any_AVX2, I422ToARGB4444Row_AVX2, 1, 0, 2, 7) +#endif +#ifdef HAS_I422TOARGB1555ROW_AVX2 +ANY31(I422ToARGB1555Row_Any_AVX2, I422ToARGB1555Row_AVX2, 1, 0, 2, 7) +#endif +#ifdef HAS_I422TORGB565ROW_AVX2 +ANY31(I422ToRGB565Row_Any_AVX2, I422ToRGB565Row_AVX2, 1, 0, 2, 7) +#endif +#ifdef HAS_I422TOARGBROW_NEON +ANY31(I444ToARGBRow_Any_NEON, I444ToARGBRow_NEON, 0, 0, 4, 7) +ANY31(I422ToARGBRow_Any_NEON, I422ToARGBRow_NEON, 1, 0, 4, 7) +ANY31(I411ToARGBRow_Any_NEON, I411ToARGBRow_NEON, 2, 0, 4, 7) +ANY31(I422ToBGRARow_Any_NEON, I422ToBGRARow_NEON, 1, 0, 4, 7) +ANY31(I422ToABGRRow_Any_NEON, I422ToABGRRow_NEON, 1, 0, 4, 7) +ANY31(I422ToRGBARow_Any_NEON, I422ToRGBARow_NEON, 1, 0, 4, 7) +ANY31(I422ToRGB24Row_Any_NEON, I422ToRGB24Row_NEON, 1, 0, 3, 7) +ANY31(I422ToRAWRow_Any_NEON, I422ToRAWRow_NEON, 1, 0, 3, 7) +ANY31(I422ToARGB4444Row_Any_NEON, I422ToARGB4444Row_NEON, 1, 0, 2, 7) +ANY31(I422ToARGB1555Row_Any_NEON, I422ToARGB1555Row_NEON, 1, 0, 2, 7) +ANY31(I422ToRGB565Row_Any_NEON, I422ToRGB565Row_NEON, 1, 0, 2, 7) +#endif +#ifdef HAS_I422TOYUY2ROW_NEON +ANY31(I422ToYUY2Row_Any_NEON, I422ToYUY2Row_NEON, 1, 1, 4, 15) +#endif +#ifdef HAS_I422TOUYVYROW_NEON +ANY31(I422ToUYVYRow_Any_NEON, I422ToUYVYRow_NEON, 1, 1, 4, 15) +#endif +#undef ANY31 + +// Any 2 planes to 1. +#define ANY21(NAMEANY, ANY_SIMD, UVSHIFT, SBPP, SBPP2, BPP, MASK) \ + void NAMEANY(const uint8* y_buf, const uint8* uv_buf, \ + uint8* dst_ptr, int width) { \ + SIMD_ALIGNED(uint8 temp[64 * 3]); \ + memset(temp, 0, 64 * 2); /* for msan */ \ + int r = width & MASK; \ + int n = width & ~MASK; \ + if (n > 0) { \ + ANY_SIMD(y_buf, uv_buf, dst_ptr, n); \ + } \ + memcpy(temp, y_buf + n * SBPP, r * SBPP); \ + memcpy(temp + 64, uv_buf + (n >> UVSHIFT) * SBPP2, \ + SS(r, UVSHIFT) * SBPP2); \ + ANY_SIMD(temp, temp + 64, temp + 128, MASK + 1); \ + memcpy(dst_ptr + n * BPP, temp + 128, r * BPP); \ + } + +// Biplanar to RGB. +#ifdef HAS_NV12TOARGBROW_SSSE3 +ANY21(NV12ToARGBRow_Any_SSSE3, NV12ToARGBRow_SSSE3, 1, 1, 2, 4, 7) +ANY21(NV21ToARGBRow_Any_SSSE3, NV21ToARGBRow_SSSE3, 1, 1, 2, 4, 7) +#endif +#ifdef HAS_NV12TOARGBROW_AVX2 +ANY21(NV12ToARGBRow_Any_AVX2, NV12ToARGBRow_AVX2, 1, 1, 2, 4, 15) +ANY21(NV21ToARGBRow_Any_AVX2, NV21ToARGBRow_AVX2, 1, 1, 2, 4, 15) +#endif +#ifdef HAS_NV12TOARGBROW_NEON +ANY21(NV12ToARGBRow_Any_NEON, NV12ToARGBRow_NEON, 1, 1, 2, 4, 7) +ANY21(NV21ToARGBRow_Any_NEON, NV21ToARGBRow_NEON, 1, 1, 2, 4, 7) +#endif +#ifdef HAS_NV12TORGB565ROW_SSSE3 +ANY21(NV12ToRGB565Row_Any_SSSE3, NV12ToRGB565Row_SSSE3, 1, 1, 2, 2, 7) +ANY21(NV21ToRGB565Row_Any_SSSE3, NV21ToRGB565Row_SSSE3, 1, 1, 2, 2, 7) +#endif +#ifdef HAS_NV12TORGB565ROW_AVX2 +ANY21(NV12ToRGB565Row_Any_AVX2, NV12ToRGB565Row_AVX2, 1, 1, 2, 2, 15) +ANY21(NV21ToRGB565Row_Any_AVX2, NV21ToRGB565Row_AVX2, 1, 1, 2, 2, 15) +#endif +#ifdef HAS_NV12TORGB565ROW_NEON +ANY21(NV12ToRGB565Row_Any_NEON, NV12ToRGB565Row_NEON, 1, 1, 2, 2, 7) +ANY21(NV21ToRGB565Row_Any_NEON, NV21ToRGB565Row_NEON, 1, 1, 2, 2, 7) +#endif + +// Merge functions. +#ifdef HAS_MERGEUVROW_SSE2 +ANY21(MergeUVRow_Any_SSE2, MergeUVRow_SSE2, 0, 1, 1, 2, 15) +#endif +#ifdef HAS_MERGEUVROW_AVX2 +ANY21(MergeUVRow_Any_AVX2, MergeUVRow_AVX2, 0, 1, 1, 2, 31) +#endif +#ifdef HAS_MERGEUVROW_NEON +ANY21(MergeUVRow_Any_NEON, MergeUVRow_NEON, 0, 1, 1, 2, 15) +#endif + +// Math functions. +#ifdef HAS_ARGBMULTIPLYROW_SSE2 +ANY21(ARGBMultiplyRow_Any_SSE2, ARGBMultiplyRow_SSE2, 0, 4, 4, 4, 3) +#endif +#ifdef HAS_ARGBADDROW_SSE2 +ANY21(ARGBAddRow_Any_SSE2, ARGBAddRow_SSE2, 0, 4, 4, 4, 3) +#endif +#ifdef HAS_ARGBSUBTRACTROW_SSE2 +ANY21(ARGBSubtractRow_Any_SSE2, ARGBSubtractRow_SSE2, 0, 4, 4, 4, 3) +#endif +#ifdef HAS_ARGBMULTIPLYROW_AVX2 +ANY21(ARGBMultiplyRow_Any_AVX2, ARGBMultiplyRow_AVX2, 0, 4, 4, 4, 7) +#endif +#ifdef HAS_ARGBADDROW_AVX2 +ANY21(ARGBAddRow_Any_AVX2, ARGBAddRow_AVX2, 0, 4, 4, 4, 7) +#endif +#ifdef HAS_ARGBSUBTRACTROW_AVX2 +ANY21(ARGBSubtractRow_Any_AVX2, ARGBSubtractRow_AVX2, 0, 4, 4, 4, 7) +#endif +#ifdef HAS_ARGBMULTIPLYROW_NEON +ANY21(ARGBMultiplyRow_Any_NEON, ARGBMultiplyRow_NEON, 0, 4, 4, 4, 7) +#endif +#ifdef HAS_ARGBADDROW_NEON +ANY21(ARGBAddRow_Any_NEON, ARGBAddRow_NEON, 0, 4, 4, 4, 7) +#endif +#ifdef HAS_ARGBSUBTRACTROW_NEON +ANY21(ARGBSubtractRow_Any_NEON, ARGBSubtractRow_NEON, 0, 4, 4, 4, 7) +#endif +#ifdef HAS_SOBELROW_SSE2 +ANY21(SobelRow_Any_SSE2, SobelRow_SSE2, 0, 1, 1, 4, 15) +#endif +#ifdef HAS_SOBELROW_NEON +ANY21(SobelRow_Any_NEON, SobelRow_NEON, 0, 1, 1, 4, 7) +#endif +#ifdef HAS_SOBELTOPLANEROW_SSE2 +ANY21(SobelToPlaneRow_Any_SSE2, SobelToPlaneRow_SSE2, 0, 1, 1, 1, 15) +#endif +#ifdef HAS_SOBELTOPLANEROW_NEON +ANY21(SobelToPlaneRow_Any_NEON, SobelToPlaneRow_NEON, 0, 1, 1, 1, 15) +#endif +#ifdef HAS_SOBELXYROW_SSE2 +ANY21(SobelXYRow_Any_SSE2, SobelXYRow_SSE2, 0, 1, 1, 4, 15) +#endif +#ifdef HAS_SOBELXYROW_NEON +ANY21(SobelXYRow_Any_NEON, SobelXYRow_NEON, 0, 1, 1, 4, 7) +#endif +#undef ANY21 + +// Any 1 to 1. +#define ANY11(NAMEANY, ANY_SIMD, UVSHIFT, SBPP, BPP, MASK) \ + void NAMEANY(const uint8* src_ptr, uint8* dst_ptr, int width) { \ + SIMD_ALIGNED(uint8 temp[128 * 2]); \ + memset(temp, 0, 128); /* for YUY2 and msan */ \ + int r = width & MASK; \ + int n = width & ~MASK; \ + if (n > 0) { \ + ANY_SIMD(src_ptr, dst_ptr, n); \ + } \ + memcpy(temp, src_ptr + (n >> UVSHIFT) * SBPP, SS(r, UVSHIFT) * SBPP); \ + ANY_SIMD(temp, temp + 128, MASK + 1); \ + memcpy(dst_ptr + n * BPP, temp + 128, r * BPP); \ + } + +#ifdef HAS_COPYROW_AVX +ANY11(CopyRow_Any_AVX, CopyRow_AVX, 0, 1, 1, 63) +#endif +#ifdef HAS_COPYROW_SSE2 +ANY11(CopyRow_Any_SSE2, CopyRow_SSE2, 0, 1, 1, 31) +#endif +#ifdef HAS_COPYROW_NEON +ANY11(CopyRow_Any_NEON, CopyRow_NEON, 0, 1, 1, 31) +#endif +#if defined(HAS_ARGBTORGB24ROW_SSSE3) +ANY11(ARGBToRGB24Row_Any_SSSE3, ARGBToRGB24Row_SSSE3, 0, 4, 3, 15) +ANY11(ARGBToRAWRow_Any_SSSE3, ARGBToRAWRow_SSSE3, 0, 4, 3, 15) +ANY11(ARGBToRGB565Row_Any_SSE2, ARGBToRGB565Row_SSE2, 0, 4, 2, 3) +ANY11(ARGBToARGB1555Row_Any_SSE2, ARGBToARGB1555Row_SSE2, 0, 4, 2, 3) +ANY11(ARGBToARGB4444Row_Any_SSE2, ARGBToARGB4444Row_SSE2, 0, 4, 2, 3) +#endif +#if defined(HAS_ARGBTOARGB4444ROW_AVX2) +ANY11(ARGBToRGB565Row_Any_AVX2, ARGBToRGB565Row_AVX2, 0, 4, 2, 7) +ANY11(ARGBToARGB1555Row_Any_AVX2, ARGBToARGB1555Row_AVX2, 0, 4, 2, 7) +ANY11(ARGBToARGB4444Row_Any_AVX2, ARGBToARGB4444Row_AVX2, 0, 4, 2, 7) +#endif +#if defined(HAS_J400TOARGBROW_SSE2) +ANY11(J400ToARGBRow_Any_SSE2, J400ToARGBRow_SSE2, 0, 1, 4, 7) +#endif +#if defined(HAS_J400TOARGBROW_AVX2) +ANY11(J400ToARGBRow_Any_AVX2, J400ToARGBRow_AVX2, 0, 1, 4, 15) +#endif +#if defined(HAS_I400TOARGBROW_SSE2) +ANY11(I400ToARGBRow_Any_SSE2, I400ToARGBRow_SSE2, 0, 1, 4, 7) +#endif +#if defined(HAS_I400TOARGBROW_AVX2) +ANY11(I400ToARGBRow_Any_AVX2, I400ToARGBRow_AVX2, 0, 1, 4, 15) +#endif +#if defined(HAS_YUY2TOARGBROW_SSSE3) +ANY11(YUY2ToARGBRow_Any_SSSE3, YUY2ToARGBRow_SSSE3, 1, 4, 4, 15) +ANY11(UYVYToARGBRow_Any_SSSE3, UYVYToARGBRow_SSSE3, 1, 4, 4, 15) +ANY11(RGB24ToARGBRow_Any_SSSE3, RGB24ToARGBRow_SSSE3, 0, 3, 4, 15) +ANY11(RAWToARGBRow_Any_SSSE3, RAWToARGBRow_SSSE3, 0, 3, 4, 15) +ANY11(RGB565ToARGBRow_Any_SSE2, RGB565ToARGBRow_SSE2, 0, 2, 4, 7) +ANY11(ARGB1555ToARGBRow_Any_SSE2, ARGB1555ToARGBRow_SSE2, 0, 2, 4, 7) +ANY11(ARGB4444ToARGBRow_Any_SSE2, ARGB4444ToARGBRow_SSE2, 0, 2, 4, 7) +#endif +#if defined(HAS_RGB565TOARGBROW_AVX2) +ANY11(RGB565ToARGBRow_Any_AVX2, RGB565ToARGBRow_AVX2, 0, 2, 4, 15) +#endif +#if defined(HAS_ARGB1555TOARGBROW_AVX2) +ANY11(ARGB1555ToARGBRow_Any_AVX2, ARGB1555ToARGBRow_AVX2, 0, 2, 4, 15) +#endif +#if defined(HAS_ARGB4444TOARGBROW_AVX2) +ANY11(ARGB4444ToARGBRow_Any_AVX2, ARGB4444ToARGBRow_AVX2, 0, 2, 4, 15) +#endif +#if defined(HAS_YUY2TOARGBROW_AVX2) +ANY11(YUY2ToARGBRow_Any_AVX2, YUY2ToARGBRow_AVX2, 1, 4, 4, 31) +ANY11(UYVYToARGBRow_Any_AVX2, UYVYToARGBRow_AVX2, 1, 4, 4, 31) +#endif +#if defined(HAS_ARGBTORGB24ROW_NEON) +ANY11(ARGBToRGB24Row_Any_NEON, ARGBToRGB24Row_NEON, 0, 4, 3, 7) +ANY11(ARGBToRAWRow_Any_NEON, ARGBToRAWRow_NEON, 0, 4, 3, 7) +ANY11(ARGBToRGB565Row_Any_NEON, ARGBToRGB565Row_NEON, 0, 4, 2, 7) +ANY11(ARGBToARGB1555Row_Any_NEON, ARGBToARGB1555Row_NEON, 0, 4, 2, 7) +ANY11(ARGBToARGB4444Row_Any_NEON, ARGBToARGB4444Row_NEON, 0, 4, 2, 7) +ANY11(J400ToARGBRow_Any_NEON, J400ToARGBRow_NEON, 0, 1, 4, 7) +ANY11(I400ToARGBRow_Any_NEON, I400ToARGBRow_NEON, 0, 1, 4, 7) +ANY11(YUY2ToARGBRow_Any_NEON, YUY2ToARGBRow_NEON, 1, 4, 4, 7) +ANY11(UYVYToARGBRow_Any_NEON, UYVYToARGBRow_NEON, 1, 4, 4, 7) +#endif +#ifdef HAS_ARGBTOYROW_AVX2 +ANY11(ARGBToYRow_Any_AVX2, ARGBToYRow_AVX2, 0, 4, 1, 31) +#endif +#ifdef HAS_ARGBTOYJROW_AVX2 +ANY11(ARGBToYJRow_Any_AVX2, ARGBToYJRow_AVX2, 0, 4, 1, 31) +#endif +#ifdef HAS_UYVYTOYROW_AVX2 +ANY11(UYVYToYRow_Any_AVX2, UYVYToYRow_AVX2, 0, 2, 1, 31) +#endif +#ifdef HAS_YUY2TOYROW_AVX2 +ANY11(YUY2ToYRow_Any_AVX2, YUY2ToYRow_AVX2, 1, 4, 1, 31) +#endif +#ifdef HAS_ARGBTOYROW_SSSE3 +ANY11(ARGBToYRow_Any_SSSE3, ARGBToYRow_SSSE3, 0, 4, 1, 15) +#endif +#ifdef HAS_BGRATOYROW_SSSE3 +ANY11(BGRAToYRow_Any_SSSE3, BGRAToYRow_SSSE3, 0, 4, 1, 15) +ANY11(ABGRToYRow_Any_SSSE3, ABGRToYRow_SSSE3, 0, 4, 1, 15) +ANY11(RGBAToYRow_Any_SSSE3, RGBAToYRow_SSSE3, 0, 4, 1, 15) +ANY11(YUY2ToYRow_Any_SSE2, YUY2ToYRow_SSE2, 1, 4, 1, 15) +ANY11(UYVYToYRow_Any_SSE2, UYVYToYRow_SSE2, 1, 4, 1, 15) +#endif +#ifdef HAS_ARGBTOYJROW_SSSE3 +ANY11(ARGBToYJRow_Any_SSSE3, ARGBToYJRow_SSSE3, 0, 4, 1, 15) +#endif +#ifdef HAS_ARGBTOYROW_NEON +ANY11(ARGBToYRow_Any_NEON, ARGBToYRow_NEON, 0, 4, 1, 7) +#endif +#ifdef HAS_ARGBTOYJROW_NEON +ANY11(ARGBToYJRow_Any_NEON, ARGBToYJRow_NEON, 0, 4, 1, 7) +#endif +#ifdef HAS_BGRATOYROW_NEON +ANY11(BGRAToYRow_Any_NEON, BGRAToYRow_NEON, 0, 4, 1, 7) +#endif +#ifdef HAS_ABGRTOYROW_NEON +ANY11(ABGRToYRow_Any_NEON, ABGRToYRow_NEON, 0, 4, 1, 7) +#endif +#ifdef HAS_RGBATOYROW_NEON +ANY11(RGBAToYRow_Any_NEON, RGBAToYRow_NEON, 0, 4, 1, 7) +#endif +#ifdef HAS_RGB24TOYROW_NEON +ANY11(RGB24ToYRow_Any_NEON, RGB24ToYRow_NEON, 0, 3, 1, 7) +#endif +#ifdef HAS_RAWTOYROW_NEON +ANY11(RAWToYRow_Any_NEON, RAWToYRow_NEON, 0, 3, 1, 7) +#endif +#ifdef HAS_RGB565TOYROW_NEON +ANY11(RGB565ToYRow_Any_NEON, RGB565ToYRow_NEON, 0, 2, 1, 7) +#endif +#ifdef HAS_ARGB1555TOYROW_NEON +ANY11(ARGB1555ToYRow_Any_NEON, ARGB1555ToYRow_NEON, 0, 2, 1, 7) +#endif +#ifdef HAS_ARGB4444TOYROW_NEON +ANY11(ARGB4444ToYRow_Any_NEON, ARGB4444ToYRow_NEON, 0, 2, 1, 7) +#endif +#ifdef HAS_YUY2TOYROW_NEON +ANY11(YUY2ToYRow_Any_NEON, YUY2ToYRow_NEON, 1, 4, 1, 15) +#endif +#ifdef HAS_UYVYTOYROW_NEON +ANY11(UYVYToYRow_Any_NEON, UYVYToYRow_NEON, 0, 2, 1, 15) +#endif +#ifdef HAS_RGB24TOARGBROW_NEON +ANY11(RGB24ToARGBRow_Any_NEON, RGB24ToARGBRow_NEON, 0, 3, 4, 7) +#endif +#ifdef HAS_RAWTOARGBROW_NEON +ANY11(RAWToARGBRow_Any_NEON, RAWToARGBRow_NEON, 0, 3, 4, 7) +#endif +#ifdef HAS_RGB565TOARGBROW_NEON +ANY11(RGB565ToARGBRow_Any_NEON, RGB565ToARGBRow_NEON, 0, 2, 4, 7) +#endif +#ifdef HAS_ARGB1555TOARGBROW_NEON +ANY11(ARGB1555ToARGBRow_Any_NEON, ARGB1555ToARGBRow_NEON, 0, 2, 4, 7) +#endif +#ifdef HAS_ARGB4444TOARGBROW_NEON +ANY11(ARGB4444ToARGBRow_Any_NEON, ARGB4444ToARGBRow_NEON, 0, 2, 4, 7) +#endif +#ifdef HAS_ARGBATTENUATEROW_SSSE3 +ANY11(ARGBAttenuateRow_Any_SSSE3, ARGBAttenuateRow_SSSE3, 0, 4, 4, 3) +#endif +#ifdef HAS_ARGBATTENUATEROW_SSE2 +ANY11(ARGBAttenuateRow_Any_SSE2, ARGBAttenuateRow_SSE2, 0, 4, 4, 3) +#endif +#ifdef HAS_ARGBUNATTENUATEROW_SSE2 +ANY11(ARGBUnattenuateRow_Any_SSE2, ARGBUnattenuateRow_SSE2, 0, 4, 4, 3) +#endif +#ifdef HAS_ARGBATTENUATEROW_AVX2 +ANY11(ARGBAttenuateRow_Any_AVX2, ARGBAttenuateRow_AVX2, 0, 4, 4, 7) +#endif +#ifdef HAS_ARGBUNATTENUATEROW_AVX2 +ANY11(ARGBUnattenuateRow_Any_AVX2, ARGBUnattenuateRow_AVX2, 0, 4, 4, 7) +#endif +#ifdef HAS_ARGBATTENUATEROW_NEON +ANY11(ARGBAttenuateRow_Any_NEON, ARGBAttenuateRow_NEON, 0, 4, 4, 7) +#endif +#undef ANY11 + +// Any 1 to 1 with parameter. +#define ANY11P(NAMEANY, ANY_SIMD, T, SBPP, BPP, MASK) \ + void NAMEANY(const uint8* src_ptr, uint8* dst_ptr, \ + T shuffler, int width) { \ + SIMD_ALIGNED(uint8 temp[64 * 2]); \ + memset(temp, 0, 64); /* for msan */ \ + int r = width & MASK; \ + int n = width & ~MASK; \ + if (n > 0) { \ + ANY_SIMD(src_ptr, dst_ptr, shuffler, n); \ + } \ + memcpy(temp, src_ptr + n * SBPP, r * SBPP); \ + ANY_SIMD(temp, temp + 64, shuffler, MASK + 1); \ + memcpy(dst_ptr + n * BPP, temp + 64, r * BPP); \ + } + +#if defined(HAS_ARGBTORGB565DITHERROW_SSE2) +ANY11P(ARGBToRGB565DitherRow_Any_SSE2, ARGBToRGB565DitherRow_SSE2, + const uint32, 4, 2, 3) +#endif +#if defined(HAS_ARGBTORGB565DITHERROW_AVX2) +ANY11P(ARGBToRGB565DitherRow_Any_AVX2, ARGBToRGB565DitherRow_AVX2, + const uint32, 4, 2, 7) +#endif +#if defined(HAS_ARGBTORGB565DITHERROW_NEON) +ANY11P(ARGBToRGB565DitherRow_Any_NEON, ARGBToRGB565DitherRow_NEON, + const uint32, 4, 2, 7) +#endif +#ifdef HAS_ARGBSHUFFLEROW_SSE2 +ANY11P(ARGBShuffleRow_Any_SSE2, ARGBShuffleRow_SSE2, const uint8*, 4, 4, 3) +#endif +#ifdef HAS_ARGBSHUFFLEROW_SSSE3 +ANY11P(ARGBShuffleRow_Any_SSSE3, ARGBShuffleRow_SSSE3, const uint8*, 4, 4, 7) +#endif +#ifdef HAS_ARGBSHUFFLEROW_AVX2 +ANY11P(ARGBShuffleRow_Any_AVX2, ARGBShuffleRow_AVX2, const uint8*, 4, 4, 15) +#endif +#ifdef HAS_ARGBSHUFFLEROW_NEON +ANY11P(ARGBShuffleRow_Any_NEON, ARGBShuffleRow_NEON, const uint8*, 4, 4, 3) +#endif +#undef ANY11P + +// Any 1 to 1 interpolate. Takes 2 rows of source via stride. +#define ANY11T(NAMEANY, ANY_SIMD, SBPP, BPP, MASK) \ + void NAMEANY(uint8* dst_ptr, const uint8* src_ptr, \ + ptrdiff_t src_stride_ptr, int width, \ + int source_y_fraction) { \ + SIMD_ALIGNED(uint8 temp[64 * 3]); \ + memset(temp, 0, 64 * 2); /* for msan */ \ + int r = width & MASK; \ + int n = width & ~MASK; \ + if (n > 0) { \ + ANY_SIMD(dst_ptr, src_ptr, src_stride_ptr, n, source_y_fraction); \ + } \ + memcpy(temp, src_ptr + n * SBPP, r * SBPP); \ + memcpy(temp + 64, src_ptr + src_stride_ptr + n * SBPP, r * SBPP); \ + ANY_SIMD(temp + 128, temp, 64, MASK + 1, source_y_fraction); \ + memcpy(dst_ptr + n * BPP, temp + 128, r * BPP); \ + } + +#ifdef HAS_INTERPOLATEROW_AVX2 +ANY11T(InterpolateRow_Any_AVX2, InterpolateRow_AVX2, 1, 1, 31) +#endif +#ifdef HAS_INTERPOLATEROW_SSSE3 +ANY11T(InterpolateRow_Any_SSSE3, InterpolateRow_SSSE3, 1, 1, 15) +#endif +#ifdef HAS_INTERPOLATEROW_SSE2 +ANY11T(InterpolateRow_Any_SSE2, InterpolateRow_SSE2, 1, 1, 15) +#endif +#ifdef HAS_INTERPOLATEROW_NEON +ANY11T(InterpolateRow_Any_NEON, InterpolateRow_NEON, 1, 1, 15) +#endif +#ifdef HAS_INTERPOLATEROW_MIPS_DSPR2 +ANY11T(InterpolateRow_Any_MIPS_DSPR2, InterpolateRow_MIPS_DSPR2, 1, 1, 3) +#endif +#undef ANY11T + +// Any 1 to 1 mirror. +#define ANY11M(NAMEANY, ANY_SIMD, BPP, MASK) \ + void NAMEANY(const uint8* src_ptr, uint8* dst_ptr, int width) { \ + SIMD_ALIGNED(uint8 temp[64 * 2]); \ + memset(temp, 0, 64); /* for msan */ \ + int r = width & MASK; \ + int n = width & ~MASK; \ + if (n > 0) { \ + ANY_SIMD(src_ptr + r * BPP, dst_ptr, n); \ + } \ + memcpy(temp, src_ptr, r * BPP); \ + ANY_SIMD(temp, temp + 64, MASK + 1); \ + memcpy(dst_ptr + n * BPP, temp + 64 + (MASK + 1 - r) * BPP, r * BPP); \ + } + +#ifdef HAS_MIRRORROW_AVX2 +ANY11M(MirrorRow_Any_AVX2, MirrorRow_AVX2, 1, 31) +#endif +#ifdef HAS_MIRRORROW_SSSE3 +ANY11M(MirrorRow_Any_SSSE3, MirrorRow_SSSE3, 1, 15) +#endif +#ifdef HAS_MIRRORROW_SSE2 +ANY11M(MirrorRow_Any_SSE2, MirrorRow_SSE2, 1, 15) +#endif +#ifdef HAS_MIRRORROW_NEON +ANY11M(MirrorRow_Any_NEON, MirrorRow_NEON, 1, 15) +#endif +#ifdef HAS_ARGBMIRRORROW_AVX2 +ANY11M(ARGBMirrorRow_Any_AVX2, ARGBMirrorRow_AVX2, 4, 7) +#endif +#ifdef HAS_ARGBMIRRORROW_SSE2 +ANY11M(ARGBMirrorRow_Any_SSE2, ARGBMirrorRow_SSE2, 4, 3) +#endif +#ifdef HAS_ARGBMIRRORROW_NEON +ANY11M(ARGBMirrorRow_Any_NEON, ARGBMirrorRow_NEON, 4, 3) +#endif +#undef ANY11M + +// Any 1 plane. (memset) +#define ANY1(NAMEANY, ANY_SIMD, T, BPP, MASK) \ + void NAMEANY(uint8* dst_ptr, T v32, int width) { \ + SIMD_ALIGNED(uint8 temp[64]); \ + int r = width & MASK; \ + int n = width & ~MASK; \ + if (n > 0) { \ + ANY_SIMD(dst_ptr, v32, n); \ + } \ + ANY_SIMD(temp, v32, MASK + 1); \ + memcpy(dst_ptr + n * BPP, temp, r * BPP); \ + } + +#ifdef HAS_SETROW_X86 +ANY1(SetRow_Any_X86, SetRow_X86, uint8, 1, 3) +#endif +#ifdef HAS_SETROW_NEON +ANY1(SetRow_Any_NEON, SetRow_NEON, uint8, 1, 15) +#endif +#ifdef HAS_ARGBSETROW_NEON +ANY1(ARGBSetRow_Any_NEON, ARGBSetRow_NEON, uint32, 4, 3) +#endif +#undef ANY1 + +// Any 1 to 2. Outputs UV planes. +#define ANY12(NAMEANY, ANY_SIMD, UVSHIFT, BPP, DUVSHIFT, MASK) \ + void NAMEANY(const uint8* src_ptr, uint8* dst_u, uint8* dst_v, int width) {\ + SIMD_ALIGNED(uint8 temp[128 * 3]); \ + memset(temp, 0, 128); /* for msan */ \ + int r = width & MASK; \ + int n = width & ~MASK; \ + if (n > 0) { \ + ANY_SIMD(src_ptr, dst_u, dst_v, n); \ + } \ + memcpy(temp, src_ptr + (n >> UVSHIFT) * BPP, SS(r, UVSHIFT) * BPP); \ + if ((width & 1) && BPP == 4) { /* repeat last 4 bytes for subsampler */ \ + memcpy(temp + SS(r, UVSHIFT) * BPP, \ + temp + SS(r, UVSHIFT) * BPP - BPP, 4); \ + } \ + ANY_SIMD(temp, temp + 128, temp + 256, MASK + 1); \ + memcpy(dst_u + (n >> DUVSHIFT), temp + 128, SS(r, DUVSHIFT)); \ + memcpy(dst_v + (n >> DUVSHIFT), temp + 256, SS(r, DUVSHIFT)); \ + } + +#ifdef HAS_SPLITUVROW_SSE2 +ANY12(SplitUVRow_Any_SSE2, SplitUVRow_SSE2, 0, 2, 0, 15) +#endif +#ifdef HAS_SPLITUVROW_AVX2 +ANY12(SplitUVRow_Any_AVX2, SplitUVRow_AVX2, 0, 2, 0, 31) +#endif +#ifdef HAS_SPLITUVROW_NEON +ANY12(SplitUVRow_Any_NEON, SplitUVRow_NEON, 0, 2, 0, 15) +#endif +#ifdef HAS_SPLITUVROW_MIPS_DSPR2 +ANY12(SplitUVRow_Any_MIPS_DSPR2, SplitUVRow_MIPS_DSPR2, 0, 2, 0, 15) +#endif +#ifdef HAS_ARGBTOUV444ROW_SSSE3 +ANY12(ARGBToUV444Row_Any_SSSE3, ARGBToUV444Row_SSSE3, 0, 4, 0, 15) +#endif +#ifdef HAS_YUY2TOUV422ROW_AVX2 +ANY12(YUY2ToUV422Row_Any_AVX2, YUY2ToUV422Row_AVX2, 1, 4, 1, 31) +ANY12(UYVYToUV422Row_Any_AVX2, UYVYToUV422Row_AVX2, 1, 4, 1, 31) +#endif +#ifdef HAS_ARGBTOUV422ROW_SSSE3 +ANY12(ARGBToUV422Row_Any_SSSE3, ARGBToUV422Row_SSSE3, 0, 4, 1, 15) +#endif +#ifdef HAS_YUY2TOUV422ROW_SSE2 +ANY12(YUY2ToUV422Row_Any_SSE2, YUY2ToUV422Row_SSE2, 1, 4, 1, 15) +ANY12(UYVYToUV422Row_Any_SSE2, UYVYToUV422Row_SSE2, 1, 4, 1, 15) +#endif +#ifdef HAS_YUY2TOUV422ROW_NEON +ANY12(ARGBToUV444Row_Any_NEON, ARGBToUV444Row_NEON, 0, 4, 0, 7) +ANY12(ARGBToUV422Row_Any_NEON, ARGBToUV422Row_NEON, 0, 4, 1, 15) +ANY12(ARGBToUV411Row_Any_NEON, ARGBToUV411Row_NEON, 0, 4, 2, 31) +ANY12(YUY2ToUV422Row_Any_NEON, YUY2ToUV422Row_NEON, 1, 4, 1, 15) +ANY12(UYVYToUV422Row_Any_NEON, UYVYToUV422Row_NEON, 1, 4, 1, 15) +#endif +#undef ANY12 + +// Any 1 to 2 with source stride (2 rows of source). Outputs UV planes. +// 128 byte row allows for 32 avx ARGB pixels. +#define ANY12S(NAMEANY, ANY_SIMD, UVSHIFT, BPP, MASK) \ + void NAMEANY(const uint8* src_ptr, int src_stride_ptr, \ + uint8* dst_u, uint8* dst_v, int width) { \ + SIMD_ALIGNED(uint8 temp[128 * 4]); \ + memset(temp, 0, 128 * 2); /* for msan */ \ + int r = width & MASK; \ + int n = width & ~MASK; \ + if (n > 0) { \ + ANY_SIMD(src_ptr, src_stride_ptr, dst_u, dst_v, n); \ + } \ + memcpy(temp, src_ptr + (n >> UVSHIFT) * BPP, SS(r, UVSHIFT) * BPP); \ + memcpy(temp + 128, src_ptr + src_stride_ptr + (n >> UVSHIFT) * BPP, \ + SS(r, UVSHIFT) * BPP); \ + if ((width & 1) && BPP == 4) { /* repeat last 4 bytes for subsampler */ \ + memcpy(temp + SS(r, UVSHIFT) * BPP, \ + temp + SS(r, UVSHIFT) * BPP - BPP, 4); \ + memcpy(temp + 128 + SS(r, UVSHIFT) * BPP, \ + temp + 128 + SS(r, UVSHIFT) * BPP - BPP, 4); \ + } \ + ANY_SIMD(temp, 128, temp + 256, temp + 384, MASK + 1); \ + memcpy(dst_u + (n >> 1), temp + 256, SS(r, 1)); \ + memcpy(dst_v + (n >> 1), temp + 384, SS(r, 1)); \ + } + +#ifdef HAS_ARGBTOUVROW_AVX2 +ANY12S(ARGBToUVRow_Any_AVX2, ARGBToUVRow_AVX2, 0, 4, 31) +#endif +#ifdef HAS_ARGBTOUVROW_SSSE3 +ANY12S(ARGBToUVRow_Any_SSSE3, ARGBToUVRow_SSSE3, 0, 4, 15) +ANY12S(ARGBToUVJRow_Any_SSSE3, ARGBToUVJRow_SSSE3, 0, 4, 15) +ANY12S(BGRAToUVRow_Any_SSSE3, BGRAToUVRow_SSSE3, 0, 4, 15) +ANY12S(ABGRToUVRow_Any_SSSE3, ABGRToUVRow_SSSE3, 0, 4, 15) +ANY12S(RGBAToUVRow_Any_SSSE3, RGBAToUVRow_SSSE3, 0, 4, 15) +#endif +#ifdef HAS_YUY2TOUVROW_AVX2 +ANY12S(YUY2ToUVRow_Any_AVX2, YUY2ToUVRow_AVX2, 1, 4, 31) +ANY12S(UYVYToUVRow_Any_AVX2, UYVYToUVRow_AVX2, 1, 4, 31) +#endif +#ifdef HAS_YUY2TOUVROW_SSE2 +ANY12S(YUY2ToUVRow_Any_SSE2, YUY2ToUVRow_SSE2, 1, 4, 15) +ANY12S(UYVYToUVRow_Any_SSE2, UYVYToUVRow_SSE2, 1, 4, 15) +#endif +#ifdef HAS_ARGBTOUVROW_NEON +ANY12S(ARGBToUVRow_Any_NEON, ARGBToUVRow_NEON, 0, 4, 15) +#endif +#ifdef HAS_ARGBTOUVJROW_NEON +ANY12S(ARGBToUVJRow_Any_NEON, ARGBToUVJRow_NEON, 0, 4, 15) +#endif +#ifdef HAS_BGRATOUVROW_NEON +ANY12S(BGRAToUVRow_Any_NEON, BGRAToUVRow_NEON, 0, 4, 15) +#endif +#ifdef HAS_ABGRTOUVROW_NEON +ANY12S(ABGRToUVRow_Any_NEON, ABGRToUVRow_NEON, 0, 4, 15) +#endif +#ifdef HAS_RGBATOUVROW_NEON +ANY12S(RGBAToUVRow_Any_NEON, RGBAToUVRow_NEON, 0, 4, 15) +#endif +#ifdef HAS_RGB24TOUVROW_NEON +ANY12S(RGB24ToUVRow_Any_NEON, RGB24ToUVRow_NEON, 0, 3, 15) +#endif +#ifdef HAS_RAWTOUVROW_NEON +ANY12S(RAWToUVRow_Any_NEON, RAWToUVRow_NEON, 0, 3, 15) +#endif +#ifdef HAS_RGB565TOUVROW_NEON +ANY12S(RGB565ToUVRow_Any_NEON, RGB565ToUVRow_NEON, 0, 2, 15) +#endif +#ifdef HAS_ARGB1555TOUVROW_NEON +ANY12S(ARGB1555ToUVRow_Any_NEON, ARGB1555ToUVRow_NEON, 0, 2, 15) +#endif +#ifdef HAS_ARGB4444TOUVROW_NEON +ANY12S(ARGB4444ToUVRow_Any_NEON, ARGB4444ToUVRow_NEON, 0, 2, 15) +#endif +#ifdef HAS_YUY2TOUVROW_NEON +ANY12S(YUY2ToUVRow_Any_NEON, YUY2ToUVRow_NEON, 1, 4, 15) +#endif +#ifdef HAS_UYVYTOUVROW_NEON +ANY12S(UYVYToUVRow_Any_NEON, UYVYToUVRow_NEON, 1, 4, 15) +#endif +#undef ANY12S + +#ifdef __cplusplus +} // extern "C" +} // namespace libyuv +#endif diff --git a/media/libaom/src/third_party/libyuv/source/row_common.cc b/media/libaom/src/third_party/libyuv/source/row_common.cc new file mode 100644 index 000000000..49875894f --- /dev/null +++ b/media/libaom/src/third_party/libyuv/source/row_common.cc @@ -0,0 +1,2576 @@ +/* + * Copyright 2011 The LibYuv Project Authors. All rights reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#include "libyuv/row.h" + +#include <string.h> // For memcpy and memset. + +#include "libyuv/basic_types.h" + +#ifdef __cplusplus +namespace libyuv { +extern "C" { +#endif + +// llvm x86 is poor at ternary operator, so use branchless min/max. + +#define USE_BRANCHLESS 1 +#if USE_BRANCHLESS +static __inline int32 clamp0(int32 v) { + return ((-(v) >> 31) & (v)); +} + +static __inline int32 clamp255(int32 v) { + return (((255 - (v)) >> 31) | (v)) & 255; +} + +static __inline uint32 Clamp(int32 val) { + int v = clamp0(val); + return (uint32)(clamp255(v)); +} + +static __inline uint32 Abs(int32 v) { + int m = v >> 31; + return (v + m) ^ m; +} +#else // USE_BRANCHLESS +static __inline int32 clamp0(int32 v) { + return (v < 0) ? 0 : v; +} + +static __inline int32 clamp255(int32 v) { + return (v > 255) ? 255 : v; +} + +static __inline uint32 Clamp(int32 val) { + int v = clamp0(val); + return (uint32)(clamp255(v)); +} + +static __inline uint32 Abs(int32 v) { + return (v < 0) ? -v : v; +} +#endif // USE_BRANCHLESS + +#ifdef LIBYUV_LITTLE_ENDIAN +#define WRITEWORD(p, v) *(uint32*)(p) = v +#else +static inline void WRITEWORD(uint8* p, uint32 v) { + p[0] = (uint8)(v & 255); + p[1] = (uint8)((v >> 8) & 255); + p[2] = (uint8)((v >> 16) & 255); + p[3] = (uint8)((v >> 24) & 255); +} +#endif + +void RGB24ToARGBRow_C(const uint8* src_rgb24, uint8* dst_argb, int width) { + int x; + for (x = 0; x < width; ++x) { + uint8 b = src_rgb24[0]; + uint8 g = src_rgb24[1]; + uint8 r = src_rgb24[2]; + dst_argb[0] = b; + dst_argb[1] = g; + dst_argb[2] = r; + dst_argb[3] = 255u; + dst_argb += 4; + src_rgb24 += 3; + } +} + +void RAWToARGBRow_C(const uint8* src_raw, uint8* dst_argb, int width) { + int x; + for (x = 0; x < width; ++x) { + uint8 r = src_raw[0]; + uint8 g = src_raw[1]; + uint8 b = src_raw[2]; + dst_argb[0] = b; + dst_argb[1] = g; + dst_argb[2] = r; + dst_argb[3] = 255u; + dst_argb += 4; + src_raw += 3; + } +} + +void RGB565ToARGBRow_C(const uint8* src_rgb565, uint8* dst_argb, int width) { + int x; + for (x = 0; x < width; ++x) { + uint8 b = src_rgb565[0] & 0x1f; + uint8 g = (src_rgb565[0] >> 5) | ((src_rgb565[1] & 0x07) << 3); + uint8 r = src_rgb565[1] >> 3; + dst_argb[0] = (b << 3) | (b >> 2); + dst_argb[1] = (g << 2) | (g >> 4); + dst_argb[2] = (r << 3) | (r >> 2); + dst_argb[3] = 255u; + dst_argb += 4; + src_rgb565 += 2; + } +} + +void ARGB1555ToARGBRow_C(const uint8* src_argb1555, uint8* dst_argb, + int width) { + int x; + for (x = 0; x < width; ++x) { + uint8 b = src_argb1555[0] & 0x1f; + uint8 g = (src_argb1555[0] >> 5) | ((src_argb1555[1] & 0x03) << 3); + uint8 r = (src_argb1555[1] & 0x7c) >> 2; + uint8 a = src_argb1555[1] >> 7; + dst_argb[0] = (b << 3) | (b >> 2); + dst_argb[1] = (g << 3) | (g >> 2); + dst_argb[2] = (r << 3) | (r >> 2); + dst_argb[3] = -a; + dst_argb += 4; + src_argb1555 += 2; + } +} + +void ARGB4444ToARGBRow_C(const uint8* src_argb4444, uint8* dst_argb, + int width) { + int x; + for (x = 0; x < width; ++x) { + uint8 b = src_argb4444[0] & 0x0f; + uint8 g = src_argb4444[0] >> 4; + uint8 r = src_argb4444[1] & 0x0f; + uint8 a = src_argb4444[1] >> 4; + dst_argb[0] = (b << 4) | b; + dst_argb[1] = (g << 4) | g; + dst_argb[2] = (r << 4) | r; + dst_argb[3] = (a << 4) | a; + dst_argb += 4; + src_argb4444 += 2; + } +} + +void ARGBToRGB24Row_C(const uint8* src_argb, uint8* dst_rgb, int width) { + int x; + for (x = 0; x < width; ++x) { + uint8 b = src_argb[0]; + uint8 g = src_argb[1]; + uint8 r = src_argb[2]; + dst_rgb[0] = b; + dst_rgb[1] = g; + dst_rgb[2] = r; + dst_rgb += 3; + src_argb += 4; + } +} + +void ARGBToRAWRow_C(const uint8* src_argb, uint8* dst_rgb, int width) { + int x; + for (x = 0; x < width; ++x) { + uint8 b = src_argb[0]; + uint8 g = src_argb[1]; + uint8 r = src_argb[2]; + dst_rgb[0] = r; + dst_rgb[1] = g; + dst_rgb[2] = b; + dst_rgb += 3; + src_argb += 4; + } +} + +void ARGBToRGB565Row_C(const uint8* src_argb, uint8* dst_rgb, int width) { + int x; + for (x = 0; x < width - 1; x += 2) { + uint8 b0 = src_argb[0] >> 3; + uint8 g0 = src_argb[1] >> 2; + uint8 r0 = src_argb[2] >> 3; + uint8 b1 = src_argb[4] >> 3; + uint8 g1 = src_argb[5] >> 2; + uint8 r1 = src_argb[6] >> 3; + WRITEWORD(dst_rgb, b0 | (g0 << 5) | (r0 << 11) | + (b1 << 16) | (g1 << 21) | (r1 << 27)); + dst_rgb += 4; + src_argb += 8; + } + if (width & 1) { + uint8 b0 = src_argb[0] >> 3; + uint8 g0 = src_argb[1] >> 2; + uint8 r0 = src_argb[2] >> 3; + *(uint16*)(dst_rgb) = b0 | (g0 << 5) | (r0 << 11); + } +} + +// dither4 is a row of 4 values from 4x4 dither matrix. +// The 4x4 matrix contains values to increase RGB. When converting to +// fewer bits (565) this provides an ordered dither. +// The order in the 4x4 matrix in first byte is upper left. +// The 4 values are passed as an int, then referenced as an array, so +// endian will not affect order of the original matrix. But the dither4 +// will containing the first pixel in the lower byte for little endian +// or the upper byte for big endian. +void ARGBToRGB565DitherRow_C(const uint8* src_argb, uint8* dst_rgb, + const uint32 dither4, int width) { + int x; + for (x = 0; x < width - 1; x += 2) { + int dither0 = ((const unsigned char*)(&dither4))[x & 3]; + int dither1 = ((const unsigned char*)(&dither4))[(x + 1) & 3]; + uint8 b0 = clamp255(src_argb[0] + dither0) >> 3; + uint8 g0 = clamp255(src_argb[1] + dither0) >> 2; + uint8 r0 = clamp255(src_argb[2] + dither0) >> 3; + uint8 b1 = clamp255(src_argb[4] + dither1) >> 3; + uint8 g1 = clamp255(src_argb[5] + dither1) >> 2; + uint8 r1 = clamp255(src_argb[6] + dither1) >> 3; + WRITEWORD(dst_rgb, b0 | (g0 << 5) | (r0 << 11) | + (b1 << 16) | (g1 << 21) | (r1 << 27)); + dst_rgb += 4; + src_argb += 8; + } + if (width & 1) { + int dither0 = ((const unsigned char*)(&dither4))[(width - 1) & 3]; + uint8 b0 = clamp255(src_argb[0] + dither0) >> 3; + uint8 g0 = clamp255(src_argb[1] + dither0) >> 2; + uint8 r0 = clamp255(src_argb[2] + dither0) >> 3; + *(uint16*)(dst_rgb) = b0 | (g0 << 5) | (r0 << 11); + } +} + +void ARGBToARGB1555Row_C(const uint8* src_argb, uint8* dst_rgb, int width) { + int x; + for (x = 0; x < width - 1; x += 2) { + uint8 b0 = src_argb[0] >> 3; + uint8 g0 = src_argb[1] >> 3; + uint8 r0 = src_argb[2] >> 3; + uint8 a0 = src_argb[3] >> 7; + uint8 b1 = src_argb[4] >> 3; + uint8 g1 = src_argb[5] >> 3; + uint8 r1 = src_argb[6] >> 3; + uint8 a1 = src_argb[7] >> 7; + *(uint32*)(dst_rgb) = + b0 | (g0 << 5) | (r0 << 10) | (a0 << 15) | + (b1 << 16) | (g1 << 21) | (r1 << 26) | (a1 << 31); + dst_rgb += 4; + src_argb += 8; + } + if (width & 1) { + uint8 b0 = src_argb[0] >> 3; + uint8 g0 = src_argb[1] >> 3; + uint8 r0 = src_argb[2] >> 3; + uint8 a0 = src_argb[3] >> 7; + *(uint16*)(dst_rgb) = + b0 | (g0 << 5) | (r0 << 10) | (a0 << 15); + } +} + +void ARGBToARGB4444Row_C(const uint8* src_argb, uint8* dst_rgb, int width) { + int x; + for (x = 0; x < width - 1; x += 2) { + uint8 b0 = src_argb[0] >> 4; + uint8 g0 = src_argb[1] >> 4; + uint8 r0 = src_argb[2] >> 4; + uint8 a0 = src_argb[3] >> 4; + uint8 b1 = src_argb[4] >> 4; + uint8 g1 = src_argb[5] >> 4; + uint8 r1 = src_argb[6] >> 4; + uint8 a1 = src_argb[7] >> 4; + *(uint32*)(dst_rgb) = + b0 | (g0 << 4) | (r0 << 8) | (a0 << 12) | + (b1 << 16) | (g1 << 20) | (r1 << 24) | (a1 << 28); + dst_rgb += 4; + src_argb += 8; + } + if (width & 1) { + uint8 b0 = src_argb[0] >> 4; + uint8 g0 = src_argb[1] >> 4; + uint8 r0 = src_argb[2] >> 4; + uint8 a0 = src_argb[3] >> 4; + *(uint16*)(dst_rgb) = + b0 | (g0 << 4) | (r0 << 8) | (a0 << 12); + } +} + +static __inline int RGBToY(uint8 r, uint8 g, uint8 b) { + return (66 * r + 129 * g + 25 * b + 0x1080) >> 8; +} + +static __inline int RGBToU(uint8 r, uint8 g, uint8 b) { + return (112 * b - 74 * g - 38 * r + 0x8080) >> 8; +} +static __inline int RGBToV(uint8 r, uint8 g, uint8 b) { + return (112 * r - 94 * g - 18 * b + 0x8080) >> 8; +} + +#define MAKEROWY(NAME, R, G, B, BPP) \ +void NAME ## ToYRow_C(const uint8* src_argb0, uint8* dst_y, int width) { \ + int x; \ + for (x = 0; x < width; ++x) { \ + dst_y[0] = RGBToY(src_argb0[R], src_argb0[G], src_argb0[B]); \ + src_argb0 += BPP; \ + dst_y += 1; \ + } \ +} \ +void NAME ## ToUVRow_C(const uint8* src_rgb0, int src_stride_rgb, \ + uint8* dst_u, uint8* dst_v, int width) { \ + const uint8* src_rgb1 = src_rgb0 + src_stride_rgb; \ + int x; \ + for (x = 0; x < width - 1; x += 2) { \ + uint8 ab = (src_rgb0[B] + src_rgb0[B + BPP] + \ + src_rgb1[B] + src_rgb1[B + BPP]) >> 2; \ + uint8 ag = (src_rgb0[G] + src_rgb0[G + BPP] + \ + src_rgb1[G] + src_rgb1[G + BPP]) >> 2; \ + uint8 ar = (src_rgb0[R] + src_rgb0[R + BPP] + \ + src_rgb1[R] + src_rgb1[R + BPP]) >> 2; \ + dst_u[0] = RGBToU(ar, ag, ab); \ + dst_v[0] = RGBToV(ar, ag, ab); \ + src_rgb0 += BPP * 2; \ + src_rgb1 += BPP * 2; \ + dst_u += 1; \ + dst_v += 1; \ + } \ + if (width & 1) { \ + uint8 ab = (src_rgb0[B] + src_rgb1[B]) >> 1; \ + uint8 ag = (src_rgb0[G] + src_rgb1[G]) >> 1; \ + uint8 ar = (src_rgb0[R] + src_rgb1[R]) >> 1; \ + dst_u[0] = RGBToU(ar, ag, ab); \ + dst_v[0] = RGBToV(ar, ag, ab); \ + } \ +} + +MAKEROWY(ARGB, 2, 1, 0, 4) +MAKEROWY(BGRA, 1, 2, 3, 4) +MAKEROWY(ABGR, 0, 1, 2, 4) +MAKEROWY(RGBA, 3, 2, 1, 4) +MAKEROWY(RGB24, 2, 1, 0, 3) +MAKEROWY(RAW, 0, 1, 2, 3) +#undef MAKEROWY + +// JPeg uses a variation on BT.601-1 full range +// y = 0.29900 * r + 0.58700 * g + 0.11400 * b +// u = -0.16874 * r - 0.33126 * g + 0.50000 * b + center +// v = 0.50000 * r - 0.41869 * g - 0.08131 * b + center +// BT.601 Mpeg range uses: +// b 0.1016 * 255 = 25.908 = 25 +// g 0.5078 * 255 = 129.489 = 129 +// r 0.2578 * 255 = 65.739 = 66 +// JPeg 8 bit Y (not used): +// b 0.11400 * 256 = 29.184 = 29 +// g 0.58700 * 256 = 150.272 = 150 +// r 0.29900 * 256 = 76.544 = 77 +// JPeg 7 bit Y: +// b 0.11400 * 128 = 14.592 = 15 +// g 0.58700 * 128 = 75.136 = 75 +// r 0.29900 * 128 = 38.272 = 38 +// JPeg 8 bit U: +// b 0.50000 * 255 = 127.5 = 127 +// g -0.33126 * 255 = -84.4713 = -84 +// r -0.16874 * 255 = -43.0287 = -43 +// JPeg 8 bit V: +// b -0.08131 * 255 = -20.73405 = -20 +// g -0.41869 * 255 = -106.76595 = -107 +// r 0.50000 * 255 = 127.5 = 127 + +static __inline int RGBToYJ(uint8 r, uint8 g, uint8 b) { + return (38 * r + 75 * g + 15 * b + 64) >> 7; +} + +static __inline int RGBToUJ(uint8 r, uint8 g, uint8 b) { + return (127 * b - 84 * g - 43 * r + 0x8080) >> 8; +} +static __inline int RGBToVJ(uint8 r, uint8 g, uint8 b) { + return (127 * r - 107 * g - 20 * b + 0x8080) >> 8; +} + +#define AVGB(a, b) (((a) + (b) + 1) >> 1) + +#define MAKEROWYJ(NAME, R, G, B, BPP) \ +void NAME ## ToYJRow_C(const uint8* src_argb0, uint8* dst_y, int width) { \ + int x; \ + for (x = 0; x < width; ++x) { \ + dst_y[0] = RGBToYJ(src_argb0[R], src_argb0[G], src_argb0[B]); \ + src_argb0 += BPP; \ + dst_y += 1; \ + } \ +} \ +void NAME ## ToUVJRow_C(const uint8* src_rgb0, int src_stride_rgb, \ + uint8* dst_u, uint8* dst_v, int width) { \ + const uint8* src_rgb1 = src_rgb0 + src_stride_rgb; \ + int x; \ + for (x = 0; x < width - 1; x += 2) { \ + uint8 ab = AVGB(AVGB(src_rgb0[B], src_rgb1[B]), \ + AVGB(src_rgb0[B + BPP], src_rgb1[B + BPP])); \ + uint8 ag = AVGB(AVGB(src_rgb0[G], src_rgb1[G]), \ + AVGB(src_rgb0[G + BPP], src_rgb1[G + BPP])); \ + uint8 ar = AVGB(AVGB(src_rgb0[R], src_rgb1[R]), \ + AVGB(src_rgb0[R + BPP], src_rgb1[R + BPP])); \ + dst_u[0] = RGBToUJ(ar, ag, ab); \ + dst_v[0] = RGBToVJ(ar, ag, ab); \ + src_rgb0 += BPP * 2; \ + src_rgb1 += BPP * 2; \ + dst_u += 1; \ + dst_v += 1; \ + } \ + if (width & 1) { \ + uint8 ab = AVGB(src_rgb0[B], src_rgb1[B]); \ + uint8 ag = AVGB(src_rgb0[G], src_rgb1[G]); \ + uint8 ar = AVGB(src_rgb0[R], src_rgb1[R]); \ + dst_u[0] = RGBToUJ(ar, ag, ab); \ + dst_v[0] = RGBToVJ(ar, ag, ab); \ + } \ +} + +MAKEROWYJ(ARGB, 2, 1, 0, 4) +#undef MAKEROWYJ + +void ARGBToUVJ422Row_C(const uint8* src_argb, + uint8* dst_u, uint8* dst_v, int width) { + int x; + for (x = 0; x < width - 1; x += 2) { + uint8 ab = (src_argb[0] + src_argb[4]) >> 1; + uint8 ag = (src_argb[1] + src_argb[5]) >> 1; + uint8 ar = (src_argb[2] + src_argb[6]) >> 1; + dst_u[0] = RGBToUJ(ar, ag, ab); + dst_v[0] = RGBToVJ(ar, ag, ab); + src_argb += 8; + dst_u += 1; + dst_v += 1; + } + if (width & 1) { + uint8 ab = src_argb[0]; + uint8 ag = src_argb[1]; + uint8 ar = src_argb[2]; + dst_u[0] = RGBToUJ(ar, ag, ab); + dst_v[0] = RGBToVJ(ar, ag, ab); + } +} + +void RGB565ToYRow_C(const uint8* src_rgb565, uint8* dst_y, int width) { + int x; + for (x = 0; x < width; ++x) { + uint8 b = src_rgb565[0] & 0x1f; + uint8 g = (src_rgb565[0] >> 5) | ((src_rgb565[1] & 0x07) << 3); + uint8 r = src_rgb565[1] >> 3; + b = (b << 3) | (b >> 2); + g = (g << 2) | (g >> 4); + r = (r << 3) | (r >> 2); + dst_y[0] = RGBToY(r, g, b); + src_rgb565 += 2; + dst_y += 1; + } +} + +void ARGB1555ToYRow_C(const uint8* src_argb1555, uint8* dst_y, int width) { + int x; + for (x = 0; x < width; ++x) { + uint8 b = src_argb1555[0] & 0x1f; + uint8 g = (src_argb1555[0] >> 5) | ((src_argb1555[1] & 0x03) << 3); + uint8 r = (src_argb1555[1] & 0x7c) >> 2; + b = (b << 3) | (b >> 2); + g = (g << 3) | (g >> 2); + r = (r << 3) | (r >> 2); + dst_y[0] = RGBToY(r, g, b); + src_argb1555 += 2; + dst_y += 1; + } +} + +void ARGB4444ToYRow_C(const uint8* src_argb4444, uint8* dst_y, int width) { + int x; + for (x = 0; x < width; ++x) { + uint8 b = src_argb4444[0] & 0x0f; + uint8 g = src_argb4444[0] >> 4; + uint8 r = src_argb4444[1] & 0x0f; + b = (b << 4) | b; + g = (g << 4) | g; + r = (r << 4) | r; + dst_y[0] = RGBToY(r, g, b); + src_argb4444 += 2; + dst_y += 1; + } +} + +void RGB565ToUVRow_C(const uint8* src_rgb565, int src_stride_rgb565, + uint8* dst_u, uint8* dst_v, int width) { + const uint8* next_rgb565 = src_rgb565 + src_stride_rgb565; + int x; + for (x = 0; x < width - 1; x += 2) { + uint8 b0 = src_rgb565[0] & 0x1f; + uint8 g0 = (src_rgb565[0] >> 5) | ((src_rgb565[1] & 0x07) << 3); + uint8 r0 = src_rgb565[1] >> 3; + uint8 b1 = src_rgb565[2] & 0x1f; + uint8 g1 = (src_rgb565[2] >> 5) | ((src_rgb565[3] & 0x07) << 3); + uint8 r1 = src_rgb565[3] >> 3; + uint8 b2 = next_rgb565[0] & 0x1f; + uint8 g2 = (next_rgb565[0] >> 5) | ((next_rgb565[1] & 0x07) << 3); + uint8 r2 = next_rgb565[1] >> 3; + uint8 b3 = next_rgb565[2] & 0x1f; + uint8 g3 = (next_rgb565[2] >> 5) | ((next_rgb565[3] & 0x07) << 3); + uint8 r3 = next_rgb565[3] >> 3; + uint8 b = (b0 + b1 + b2 + b3); // 565 * 4 = 787. + uint8 g = (g0 + g1 + g2 + g3); + uint8 r = (r0 + r1 + r2 + r3); + b = (b << 1) | (b >> 6); // 787 -> 888. + r = (r << 1) | (r >> 6); + dst_u[0] = RGBToU(r, g, b); + dst_v[0] = RGBToV(r, g, b); + src_rgb565 += 4; + next_rgb565 += 4; + dst_u += 1; + dst_v += 1; + } + if (width & 1) { + uint8 b0 = src_rgb565[0] & 0x1f; + uint8 g0 = (src_rgb565[0] >> 5) | ((src_rgb565[1] & 0x07) << 3); + uint8 r0 = src_rgb565[1] >> 3; + uint8 b2 = next_rgb565[0] & 0x1f; + uint8 g2 = (next_rgb565[0] >> 5) | ((next_rgb565[1] & 0x07) << 3); + uint8 r2 = next_rgb565[1] >> 3; + uint8 b = (b0 + b2); // 565 * 2 = 676. + uint8 g = (g0 + g2); + uint8 r = (r0 + r2); + b = (b << 2) | (b >> 4); // 676 -> 888 + g = (g << 1) | (g >> 6); + r = (r << 2) | (r >> 4); + dst_u[0] = RGBToU(r, g, b); + dst_v[0] = RGBToV(r, g, b); + } +} + +void ARGB1555ToUVRow_C(const uint8* src_argb1555, int src_stride_argb1555, + uint8* dst_u, uint8* dst_v, int width) { + const uint8* next_argb1555 = src_argb1555 + src_stride_argb1555; + int x; + for (x = 0; x < width - 1; x += 2) { + uint8 b0 = src_argb1555[0] & 0x1f; + uint8 g0 = (src_argb1555[0] >> 5) | ((src_argb1555[1] & 0x03) << 3); + uint8 r0 = (src_argb1555[1] & 0x7c) >> 2; + uint8 b1 = src_argb1555[2] & 0x1f; + uint8 g1 = (src_argb1555[2] >> 5) | ((src_argb1555[3] & 0x03) << 3); + uint8 r1 = (src_argb1555[3] & 0x7c) >> 2; + uint8 b2 = next_argb1555[0] & 0x1f; + uint8 g2 = (next_argb1555[0] >> 5) | ((next_argb1555[1] & 0x03) << 3); + uint8 r2 = (next_argb1555[1] & 0x7c) >> 2; + uint8 b3 = next_argb1555[2] & 0x1f; + uint8 g3 = (next_argb1555[2] >> 5) | ((next_argb1555[3] & 0x03) << 3); + uint8 r3 = (next_argb1555[3] & 0x7c) >> 2; + uint8 b = (b0 + b1 + b2 + b3); // 555 * 4 = 777. + uint8 g = (g0 + g1 + g2 + g3); + uint8 r = (r0 + r1 + r2 + r3); + b = (b << 1) | (b >> 6); // 777 -> 888. + g = (g << 1) | (g >> 6); + r = (r << 1) | (r >> 6); + dst_u[0] = RGBToU(r, g, b); + dst_v[0] = RGBToV(r, g, b); + src_argb1555 += 4; + next_argb1555 += 4; + dst_u += 1; + dst_v += 1; + } + if (width & 1) { + uint8 b0 = src_argb1555[0] & 0x1f; + uint8 g0 = (src_argb1555[0] >> 5) | ((src_argb1555[1] & 0x03) << 3); + uint8 r0 = (src_argb1555[1] & 0x7c) >> 2; + uint8 b2 = next_argb1555[0] & 0x1f; + uint8 g2 = (next_argb1555[0] >> 5) | ((next_argb1555[1] & 0x03) << 3); + uint8 r2 = next_argb1555[1] >> 3; + uint8 b = (b0 + b2); // 555 * 2 = 666. + uint8 g = (g0 + g2); + uint8 r = (r0 + r2); + b = (b << 2) | (b >> 4); // 666 -> 888. + g = (g << 2) | (g >> 4); + r = (r << 2) | (r >> 4); + dst_u[0] = RGBToU(r, g, b); + dst_v[0] = RGBToV(r, g, b); + } +} + +void ARGB4444ToUVRow_C(const uint8* src_argb4444, int src_stride_argb4444, + uint8* dst_u, uint8* dst_v, int width) { + const uint8* next_argb4444 = src_argb4444 + src_stride_argb4444; + int x; + for (x = 0; x < width - 1; x += 2) { + uint8 b0 = src_argb4444[0] & 0x0f; + uint8 g0 = src_argb4444[0] >> 4; + uint8 r0 = src_argb4444[1] & 0x0f; + uint8 b1 = src_argb4444[2] & 0x0f; + uint8 g1 = src_argb4444[2] >> 4; + uint8 r1 = src_argb4444[3] & 0x0f; + uint8 b2 = next_argb4444[0] & 0x0f; + uint8 g2 = next_argb4444[0] >> 4; + uint8 r2 = next_argb4444[1] & 0x0f; + uint8 b3 = next_argb4444[2] & 0x0f; + uint8 g3 = next_argb4444[2] >> 4; + uint8 r3 = next_argb4444[3] & 0x0f; + uint8 b = (b0 + b1 + b2 + b3); // 444 * 4 = 666. + uint8 g = (g0 + g1 + g2 + g3); + uint8 r = (r0 + r1 + r2 + r3); + b = (b << 2) | (b >> 4); // 666 -> 888. + g = (g << 2) | (g >> 4); + r = (r << 2) | (r >> 4); + dst_u[0] = RGBToU(r, g, b); + dst_v[0] = RGBToV(r, g, b); + src_argb4444 += 4; + next_argb4444 += 4; + dst_u += 1; + dst_v += 1; + } + if (width & 1) { + uint8 b0 = src_argb4444[0] & 0x0f; + uint8 g0 = src_argb4444[0] >> 4; + uint8 r0 = src_argb4444[1] & 0x0f; + uint8 b2 = next_argb4444[0] & 0x0f; + uint8 g2 = next_argb4444[0] >> 4; + uint8 r2 = next_argb4444[1] & 0x0f; + uint8 b = (b0 + b2); // 444 * 2 = 555. + uint8 g = (g0 + g2); + uint8 r = (r0 + r2); + b = (b << 3) | (b >> 2); // 555 -> 888. + g = (g << 3) | (g >> 2); + r = (r << 3) | (r >> 2); + dst_u[0] = RGBToU(r, g, b); + dst_v[0] = RGBToV(r, g, b); + } +} + +void ARGBToUV444Row_C(const uint8* src_argb, + uint8* dst_u, uint8* dst_v, int width) { + int x; + for (x = 0; x < width; ++x) { + uint8 ab = src_argb[0]; + uint8 ag = src_argb[1]; + uint8 ar = src_argb[2]; + dst_u[0] = RGBToU(ar, ag, ab); + dst_v[0] = RGBToV(ar, ag, ab); + src_argb += 4; + dst_u += 1; + dst_v += 1; + } +} + +void ARGBToUV422Row_C(const uint8* src_argb, + uint8* dst_u, uint8* dst_v, int width) { + int x; + for (x = 0; x < width - 1; x += 2) { + uint8 ab = (src_argb[0] + src_argb[4]) >> 1; + uint8 ag = (src_argb[1] + src_argb[5]) >> 1; + uint8 ar = (src_argb[2] + src_argb[6]) >> 1; + dst_u[0] = RGBToU(ar, ag, ab); + dst_v[0] = RGBToV(ar, ag, ab); + src_argb += 8; + dst_u += 1; + dst_v += 1; + } + if (width & 1) { + uint8 ab = src_argb[0]; + uint8 ag = src_argb[1]; + uint8 ar = src_argb[2]; + dst_u[0] = RGBToU(ar, ag, ab); + dst_v[0] = RGBToV(ar, ag, ab); + } +} + +void ARGBToUV411Row_C(const uint8* src_argb, + uint8* dst_u, uint8* dst_v, int width) { + int x; + for (x = 0; x < width - 3; x += 4) { + uint8 ab = (src_argb[0] + src_argb[4] + src_argb[8] + src_argb[12]) >> 2; + uint8 ag = (src_argb[1] + src_argb[5] + src_argb[9] + src_argb[13]) >> 2; + uint8 ar = (src_argb[2] + src_argb[6] + src_argb[10] + src_argb[14]) >> 2; + dst_u[0] = RGBToU(ar, ag, ab); + dst_v[0] = RGBToV(ar, ag, ab); + src_argb += 16; + dst_u += 1; + dst_v += 1; + } + if ((width & 3) == 3) { + uint8 ab = (src_argb[0] + src_argb[4] + src_argb[8]) / 3; + uint8 ag = (src_argb[1] + src_argb[5] + src_argb[9]) / 3; + uint8 ar = (src_argb[2] + src_argb[6] + src_argb[10]) / 3; + dst_u[0] = RGBToU(ar, ag, ab); + dst_v[0] = RGBToV(ar, ag, ab); + } else if ((width & 3) == 2) { + uint8 ab = (src_argb[0] + src_argb[4]) >> 1; + uint8 ag = (src_argb[1] + src_argb[5]) >> 1; + uint8 ar = (src_argb[2] + src_argb[6]) >> 1; + dst_u[0] = RGBToU(ar, ag, ab); + dst_v[0] = RGBToV(ar, ag, ab); + } else if ((width & 3) == 1) { + uint8 ab = src_argb[0]; + uint8 ag = src_argb[1]; + uint8 ar = src_argb[2]; + dst_u[0] = RGBToU(ar, ag, ab); + dst_v[0] = RGBToV(ar, ag, ab); + } +} + +void ARGBGrayRow_C(const uint8* src_argb, uint8* dst_argb, int width) { + int x; + for (x = 0; x < width; ++x) { + uint8 y = RGBToYJ(src_argb[2], src_argb[1], src_argb[0]); + dst_argb[2] = dst_argb[1] = dst_argb[0] = y; + dst_argb[3] = src_argb[3]; + dst_argb += 4; + src_argb += 4; + } +} + +// Convert a row of image to Sepia tone. +void ARGBSepiaRow_C(uint8* dst_argb, int width) { + int x; + for (x = 0; x < width; ++x) { + int b = dst_argb[0]; + int g = dst_argb[1]; + int r = dst_argb[2]; + int sb = (b * 17 + g * 68 + r * 35) >> 7; + int sg = (b * 22 + g * 88 + r * 45) >> 7; + int sr = (b * 24 + g * 98 + r * 50) >> 7; + // b does not over flow. a is preserved from original. + dst_argb[0] = sb; + dst_argb[1] = clamp255(sg); + dst_argb[2] = clamp255(sr); + dst_argb += 4; + } +} + +// Apply color matrix to a row of image. Matrix is signed. +// TODO(fbarchard): Consider adding rounding (+32). +void ARGBColorMatrixRow_C(const uint8* src_argb, uint8* dst_argb, + const int8* matrix_argb, int width) { + int x; + for (x = 0; x < width; ++x) { + int b = src_argb[0]; + int g = src_argb[1]; + int r = src_argb[2]; + int a = src_argb[3]; + int sb = (b * matrix_argb[0] + g * matrix_argb[1] + + r * matrix_argb[2] + a * matrix_argb[3]) >> 6; + int sg = (b * matrix_argb[4] + g * matrix_argb[5] + + r * matrix_argb[6] + a * matrix_argb[7]) >> 6; + int sr = (b * matrix_argb[8] + g * matrix_argb[9] + + r * matrix_argb[10] + a * matrix_argb[11]) >> 6; + int sa = (b * matrix_argb[12] + g * matrix_argb[13] + + r * matrix_argb[14] + a * matrix_argb[15]) >> 6; + dst_argb[0] = Clamp(sb); + dst_argb[1] = Clamp(sg); + dst_argb[2] = Clamp(sr); + dst_argb[3] = Clamp(sa); + src_argb += 4; + dst_argb += 4; + } +} + +// Apply color table to a row of image. +void ARGBColorTableRow_C(uint8* dst_argb, const uint8* table_argb, int width) { + int x; + for (x = 0; x < width; ++x) { + int b = dst_argb[0]; + int g = dst_argb[1]; + int r = dst_argb[2]; + int a = dst_argb[3]; + dst_argb[0] = table_argb[b * 4 + 0]; + dst_argb[1] = table_argb[g * 4 + 1]; + dst_argb[2] = table_argb[r * 4 + 2]; + dst_argb[3] = table_argb[a * 4 + 3]; + dst_argb += 4; + } +} + +// Apply color table to a row of image. +void RGBColorTableRow_C(uint8* dst_argb, const uint8* table_argb, int width) { + int x; + for (x = 0; x < width; ++x) { + int b = dst_argb[0]; + int g = dst_argb[1]; + int r = dst_argb[2]; + dst_argb[0] = table_argb[b * 4 + 0]; + dst_argb[1] = table_argb[g * 4 + 1]; + dst_argb[2] = table_argb[r * 4 + 2]; + dst_argb += 4; + } +} + +void ARGBQuantizeRow_C(uint8* dst_argb, int scale, int interval_size, + int interval_offset, int width) { + int x; + for (x = 0; x < width; ++x) { + int b = dst_argb[0]; + int g = dst_argb[1]; + int r = dst_argb[2]; + dst_argb[0] = (b * scale >> 16) * interval_size + interval_offset; + dst_argb[1] = (g * scale >> 16) * interval_size + interval_offset; + dst_argb[2] = (r * scale >> 16) * interval_size + interval_offset; + dst_argb += 4; + } +} + +#define REPEAT8(v) (v) | ((v) << 8) +#define SHADE(f, v) v * f >> 24 + +void ARGBShadeRow_C(const uint8* src_argb, uint8* dst_argb, int width, + uint32 value) { + const uint32 b_scale = REPEAT8(value & 0xff); + const uint32 g_scale = REPEAT8((value >> 8) & 0xff); + const uint32 r_scale = REPEAT8((value >> 16) & 0xff); + const uint32 a_scale = REPEAT8(value >> 24); + + int i; + for (i = 0; i < width; ++i) { + const uint32 b = REPEAT8(src_argb[0]); + const uint32 g = REPEAT8(src_argb[1]); + const uint32 r = REPEAT8(src_argb[2]); + const uint32 a = REPEAT8(src_argb[3]); + dst_argb[0] = SHADE(b, b_scale); + dst_argb[1] = SHADE(g, g_scale); + dst_argb[2] = SHADE(r, r_scale); + dst_argb[3] = SHADE(a, a_scale); + src_argb += 4; + dst_argb += 4; + } +} +#undef REPEAT8 +#undef SHADE + +#define REPEAT8(v) (v) | ((v) << 8) +#define SHADE(f, v) v * f >> 16 + +void ARGBMultiplyRow_C(const uint8* src_argb0, const uint8* src_argb1, + uint8* dst_argb, int width) { + int i; + for (i = 0; i < width; ++i) { + const uint32 b = REPEAT8(src_argb0[0]); + const uint32 g = REPEAT8(src_argb0[1]); + const uint32 r = REPEAT8(src_argb0[2]); + const uint32 a = REPEAT8(src_argb0[3]); + const uint32 b_scale = src_argb1[0]; + const uint32 g_scale = src_argb1[1]; + const uint32 r_scale = src_argb1[2]; + const uint32 a_scale = src_argb1[3]; + dst_argb[0] = SHADE(b, b_scale); + dst_argb[1] = SHADE(g, g_scale); + dst_argb[2] = SHADE(r, r_scale); + dst_argb[3] = SHADE(a, a_scale); + src_argb0 += 4; + src_argb1 += 4; + dst_argb += 4; + } +} +#undef REPEAT8 +#undef SHADE + +#define SHADE(f, v) clamp255(v + f) + +void ARGBAddRow_C(const uint8* src_argb0, const uint8* src_argb1, + uint8* dst_argb, int width) { + int i; + for (i = 0; i < width; ++i) { + const int b = src_argb0[0]; + const int g = src_argb0[1]; + const int r = src_argb0[2]; + const int a = src_argb0[3]; + const int b_add = src_argb1[0]; + const int g_add = src_argb1[1]; + const int r_add = src_argb1[2]; + const int a_add = src_argb1[3]; + dst_argb[0] = SHADE(b, b_add); + dst_argb[1] = SHADE(g, g_add); + dst_argb[2] = SHADE(r, r_add); + dst_argb[3] = SHADE(a, a_add); + src_argb0 += 4; + src_argb1 += 4; + dst_argb += 4; + } +} +#undef SHADE + +#define SHADE(f, v) clamp0(f - v) + +void ARGBSubtractRow_C(const uint8* src_argb0, const uint8* src_argb1, + uint8* dst_argb, int width) { + int i; + for (i = 0; i < width; ++i) { + const int b = src_argb0[0]; + const int g = src_argb0[1]; + const int r = src_argb0[2]; + const int a = src_argb0[3]; + const int b_sub = src_argb1[0]; + const int g_sub = src_argb1[1]; + const int r_sub = src_argb1[2]; + const int a_sub = src_argb1[3]; + dst_argb[0] = SHADE(b, b_sub); + dst_argb[1] = SHADE(g, g_sub); + dst_argb[2] = SHADE(r, r_sub); + dst_argb[3] = SHADE(a, a_sub); + src_argb0 += 4; + src_argb1 += 4; + dst_argb += 4; + } +} +#undef SHADE + +// Sobel functions which mimics SSSE3. +void SobelXRow_C(const uint8* src_y0, const uint8* src_y1, const uint8* src_y2, + uint8* dst_sobelx, int width) { + int i; + for (i = 0; i < width; ++i) { + int a = src_y0[i]; + int b = src_y1[i]; + int c = src_y2[i]; + int a_sub = src_y0[i + 2]; + int b_sub = src_y1[i + 2]; + int c_sub = src_y2[i + 2]; + int a_diff = a - a_sub; + int b_diff = b - b_sub; + int c_diff = c - c_sub; + int sobel = Abs(a_diff + b_diff * 2 + c_diff); + dst_sobelx[i] = (uint8)(clamp255(sobel)); + } +} + +void SobelYRow_C(const uint8* src_y0, const uint8* src_y1, + uint8* dst_sobely, int width) { + int i; + for (i = 0; i < width; ++i) { + int a = src_y0[i + 0]; + int b = src_y0[i + 1]; + int c = src_y0[i + 2]; + int a_sub = src_y1[i + 0]; + int b_sub = src_y1[i + 1]; + int c_sub = src_y1[i + 2]; + int a_diff = a - a_sub; + int b_diff = b - b_sub; + int c_diff = c - c_sub; + int sobel = Abs(a_diff + b_diff * 2 + c_diff); + dst_sobely[i] = (uint8)(clamp255(sobel)); + } +} + +void SobelRow_C(const uint8* src_sobelx, const uint8* src_sobely, + uint8* dst_argb, int width) { + int i; + for (i = 0; i < width; ++i) { + int r = src_sobelx[i]; + int b = src_sobely[i]; + int s = clamp255(r + b); + dst_argb[0] = (uint8)(s); + dst_argb[1] = (uint8)(s); + dst_argb[2] = (uint8)(s); + dst_argb[3] = (uint8)(255u); + dst_argb += 4; + } +} + +void SobelToPlaneRow_C(const uint8* src_sobelx, const uint8* src_sobely, + uint8* dst_y, int width) { + int i; + for (i = 0; i < width; ++i) { + int r = src_sobelx[i]; + int b = src_sobely[i]; + int s = clamp255(r + b); + dst_y[i] = (uint8)(s); + } +} + +void SobelXYRow_C(const uint8* src_sobelx, const uint8* src_sobely, + uint8* dst_argb, int width) { + int i; + for (i = 0; i < width; ++i) { + int r = src_sobelx[i]; + int b = src_sobely[i]; + int g = clamp255(r + b); + dst_argb[0] = (uint8)(b); + dst_argb[1] = (uint8)(g); + dst_argb[2] = (uint8)(r); + dst_argb[3] = (uint8)(255u); + dst_argb += 4; + } +} + +void J400ToARGBRow_C(const uint8* src_y, uint8* dst_argb, int width) { + // Copy a Y to RGB. + int x; + for (x = 0; x < width; ++x) { + uint8 y = src_y[0]; + dst_argb[2] = dst_argb[1] = dst_argb[0] = y; + dst_argb[3] = 255u; + dst_argb += 4; + ++src_y; + } +} + +// BT.601 YUV to RGB reference +// R = (Y - 16) * 1.164 - V * -1.596 +// G = (Y - 16) * 1.164 - U * 0.391 - V * 0.813 +// B = (Y - 16) * 1.164 - U * -2.018 + +// Y contribution to R,G,B. Scale and bias. +// TODO(fbarchard): Consider moving constants into a common header. +#define YG 18997 /* round(1.164 * 64 * 256 * 256 / 257) */ +#define YGB -1160 /* 1.164 * 64 * -16 + 64 / 2 */ + +// U and V contributions to R,G,B. +#define UB -128 /* max(-128, round(-2.018 * 64)) */ +#define UG 25 /* round(0.391 * 64) */ +#define VG 52 /* round(0.813 * 64) */ +#define VR -102 /* round(-1.596 * 64) */ + +// Bias values to subtract 16 from Y and 128 from U and V. +#define BB (UB * 128 + YGB) +#define BG (UG * 128 + VG * 128 + YGB) +#define BR (VR * 128 + YGB) + +// C reference code that mimics the YUV assembly. +static __inline void YuvPixel(uint8 y, uint8 u, uint8 v, + uint8* b, uint8* g, uint8* r) { + uint32 y1 = (uint32)(y * 0x0101 * YG) >> 16; + *b = Clamp((int32)(-(u * UB) + y1 + BB) >> 6); + *g = Clamp((int32)(-(v * VG + u * UG) + y1 + BG) >> 6); + *r = Clamp((int32)(-(v * VR)+ y1 + BR) >> 6); +} + +// C reference code that mimics the YUV assembly. +static __inline void YPixel(uint8 y, uint8* b, uint8* g, uint8* r) { + uint32 y1 = (uint32)(y * 0x0101 * YG) >> 16; + *b = Clamp((int32)(y1 + YGB) >> 6); + *g = Clamp((int32)(y1 + YGB) >> 6); + *r = Clamp((int32)(y1 + YGB) >> 6); +} + +#undef YG +#undef YGB +#undef UB +#undef UG +#undef VG +#undef VR +#undef BB +#undef BG +#undef BR + +// JPEG YUV to RGB reference +// * R = Y - V * -1.40200 +// * G = Y - U * 0.34414 - V * 0.71414 +// * B = Y - U * -1.77200 + +// Y contribution to R,G,B. Scale and bias. +// TODO(fbarchard): Consider moving constants into a common header. +#define YGJ 16320 /* round(1.000 * 64 * 256 * 256 / 257) */ +#define YGBJ 32 /* 64 / 2 */ + +// U and V contributions to R,G,B. +#define UBJ -113 /* round(-1.77200 * 64) */ +#define UGJ 22 /* round(0.34414 * 64) */ +#define VGJ 46 /* round(0.71414 * 64) */ +#define VRJ -90 /* round(-1.40200 * 64) */ + +// Bias values to subtract 16 from Y and 128 from U and V. +#define BBJ (UBJ * 128 + YGBJ) +#define BGJ (UGJ * 128 + VGJ * 128 + YGBJ) +#define BRJ (VRJ * 128 + YGBJ) + +// C reference code that mimics the YUV assembly. +static __inline void YuvJPixel(uint8 y, uint8 u, uint8 v, + uint8* b, uint8* g, uint8* r) { + uint32 y1 = (uint32)(y * 0x0101 * YGJ) >> 16; + *b = Clamp((int32)(-(u * UBJ) + y1 + BBJ) >> 6); + *g = Clamp((int32)(-(v * VGJ + u * UGJ) + y1 + BGJ) >> 6); + *r = Clamp((int32)(-(v * VRJ) + y1 + BRJ) >> 6); +} + +#undef YGJ +#undef YGBJ +#undef UBJ +#undef UGJ +#undef VGJ +#undef VRJ +#undef BBJ +#undef BGJ +#undef BRJ + +#if !defined(LIBYUV_DISABLE_NEON) && \ + (defined(__ARM_NEON__) || defined(__aarch64__) || defined(LIBYUV_NEON)) +// C mimic assembly. +// TODO(fbarchard): Remove subsampling from Neon. +void I444ToARGBRow_C(const uint8* src_y, + const uint8* src_u, + const uint8* src_v, + uint8* rgb_buf, + int width) { + int x; + for (x = 0; x < width - 1; x += 2) { + uint8 u = (src_u[0] + src_u[1] + 1) >> 1; + uint8 v = (src_v[0] + src_v[1] + 1) >> 1; + YuvPixel(src_y[0], u, v, rgb_buf + 0, rgb_buf + 1, rgb_buf + 2); + rgb_buf[3] = 255; + YuvPixel(src_y[1], u, v, rgb_buf + 4, rgb_buf + 5, rgb_buf + 6); + rgb_buf[7] = 255; + src_y += 2; + src_u += 2; + src_v += 2; + rgb_buf += 8; // Advance 2 pixels. + } + if (width & 1) { + YuvPixel(src_y[0], src_u[0], src_v[0], + rgb_buf + 0, rgb_buf + 1, rgb_buf + 2); + } +} +#else +void I444ToARGBRow_C(const uint8* src_y, + const uint8* src_u, + const uint8* src_v, + uint8* rgb_buf, + int width) { + int x; + for (x = 0; x < width; ++x) { + YuvPixel(src_y[0], src_u[0], src_v[0], + rgb_buf + 0, rgb_buf + 1, rgb_buf + 2); + rgb_buf[3] = 255; + src_y += 1; + src_u += 1; + src_v += 1; + rgb_buf += 4; // Advance 1 pixel. + } +} +#endif + +// Also used for 420 +void I422ToARGBRow_C(const uint8* src_y, + const uint8* src_u, + const uint8* src_v, + uint8* rgb_buf, + int width) { + int x; + for (x = 0; x < width - 1; x += 2) { + YuvPixel(src_y[0], src_u[0], src_v[0], + rgb_buf + 0, rgb_buf + 1, rgb_buf + 2); + rgb_buf[3] = 255; + YuvPixel(src_y[1], src_u[0], src_v[0], + rgb_buf + 4, rgb_buf + 5, rgb_buf + 6); + rgb_buf[7] = 255; + src_y += 2; + src_u += 1; + src_v += 1; + rgb_buf += 8; // Advance 2 pixels. + } + if (width & 1) { + YuvPixel(src_y[0], src_u[0], src_v[0], + rgb_buf + 0, rgb_buf + 1, rgb_buf + 2); + rgb_buf[3] = 255; + } +} + +void J422ToARGBRow_C(const uint8* src_y, + const uint8* src_u, + const uint8* src_v, + uint8* rgb_buf, + int width) { + int x; + for (x = 0; x < width - 1; x += 2) { + YuvJPixel(src_y[0], src_u[0], src_v[0], + rgb_buf + 0, rgb_buf + 1, rgb_buf + 2); + rgb_buf[3] = 255; + YuvJPixel(src_y[1], src_u[0], src_v[0], + rgb_buf + 4, rgb_buf + 5, rgb_buf + 6); + rgb_buf[7] = 255; + src_y += 2; + src_u += 1; + src_v += 1; + rgb_buf += 8; // Advance 2 pixels. + } + if (width & 1) { + YuvJPixel(src_y[0], src_u[0], src_v[0], + rgb_buf + 0, rgb_buf + 1, rgb_buf + 2); + rgb_buf[3] = 255; + } +} + +void I422ToRGB24Row_C(const uint8* src_y, + const uint8* src_u, + const uint8* src_v, + uint8* rgb_buf, + int width) { + int x; + for (x = 0; x < width - 1; x += 2) { + YuvPixel(src_y[0], src_u[0], src_v[0], + rgb_buf + 0, rgb_buf + 1, rgb_buf + 2); + YuvPixel(src_y[1], src_u[0], src_v[0], + rgb_buf + 3, rgb_buf + 4, rgb_buf + 5); + src_y += 2; + src_u += 1; + src_v += 1; + rgb_buf += 6; // Advance 2 pixels. + } + if (width & 1) { + YuvPixel(src_y[0], src_u[0], src_v[0], + rgb_buf + 0, rgb_buf + 1, rgb_buf + 2); + } +} + +void I422ToRAWRow_C(const uint8* src_y, + const uint8* src_u, + const uint8* src_v, + uint8* rgb_buf, + int width) { + int x; + for (x = 0; x < width - 1; x += 2) { + YuvPixel(src_y[0], src_u[0], src_v[0], + rgb_buf + 2, rgb_buf + 1, rgb_buf + 0); + YuvPixel(src_y[1], src_u[0], src_v[0], + rgb_buf + 5, rgb_buf + 4, rgb_buf + 3); + src_y += 2; + src_u += 1; + src_v += 1; + rgb_buf += 6; // Advance 2 pixels. + } + if (width & 1) { + YuvPixel(src_y[0], src_u[0], src_v[0], + rgb_buf + 2, rgb_buf + 1, rgb_buf + 0); + } +} + +void I422ToARGB4444Row_C(const uint8* src_y, + const uint8* src_u, + const uint8* src_v, + uint8* dst_argb4444, + int width) { + uint8 b0; + uint8 g0; + uint8 r0; + uint8 b1; + uint8 g1; + uint8 r1; + int x; + for (x = 0; x < width - 1; x += 2) { + YuvPixel(src_y[0], src_u[0], src_v[0], &b0, &g0, &r0); + YuvPixel(src_y[1], src_u[0], src_v[0], &b1, &g1, &r1); + b0 = b0 >> 4; + g0 = g0 >> 4; + r0 = r0 >> 4; + b1 = b1 >> 4; + g1 = g1 >> 4; + r1 = r1 >> 4; + *(uint32*)(dst_argb4444) = b0 | (g0 << 4) | (r0 << 8) | + (b1 << 16) | (g1 << 20) | (r1 << 24) | 0xf000f000; + src_y += 2; + src_u += 1; + src_v += 1; + dst_argb4444 += 4; // Advance 2 pixels. + } + if (width & 1) { + YuvPixel(src_y[0], src_u[0], src_v[0], &b0, &g0, &r0); + b0 = b0 >> 4; + g0 = g0 >> 4; + r0 = r0 >> 4; + *(uint16*)(dst_argb4444) = b0 | (g0 << 4) | (r0 << 8) | + 0xf000; + } +} + +void I422ToARGB1555Row_C(const uint8* src_y, + const uint8* src_u, + const uint8* src_v, + uint8* dst_argb1555, + int width) { + uint8 b0; + uint8 g0; + uint8 r0; + uint8 b1; + uint8 g1; + uint8 r1; + int x; + for (x = 0; x < width - 1; x += 2) { + YuvPixel(src_y[0], src_u[0], src_v[0], &b0, &g0, &r0); + YuvPixel(src_y[1], src_u[0], src_v[0], &b1, &g1, &r1); + b0 = b0 >> 3; + g0 = g0 >> 3; + r0 = r0 >> 3; + b1 = b1 >> 3; + g1 = g1 >> 3; + r1 = r1 >> 3; + *(uint32*)(dst_argb1555) = b0 | (g0 << 5) | (r0 << 10) | + (b1 << 16) | (g1 << 21) | (r1 << 26) | 0x80008000; + src_y += 2; + src_u += 1; + src_v += 1; + dst_argb1555 += 4; // Advance 2 pixels. + } + if (width & 1) { + YuvPixel(src_y[0], src_u[0], src_v[0], &b0, &g0, &r0); + b0 = b0 >> 3; + g0 = g0 >> 3; + r0 = r0 >> 3; + *(uint16*)(dst_argb1555) = b0 | (g0 << 5) | (r0 << 10) | + 0x8000; + } +} + +void I422ToRGB565Row_C(const uint8* src_y, + const uint8* src_u, + const uint8* src_v, + uint8* dst_rgb565, + int width) { + uint8 b0; + uint8 g0; + uint8 r0; + uint8 b1; + uint8 g1; + uint8 r1; + int x; + for (x = 0; x < width - 1; x += 2) { + YuvPixel(src_y[0], src_u[0], src_v[0], &b0, &g0, &r0); + YuvPixel(src_y[1], src_u[0], src_v[0], &b1, &g1, &r1); + b0 = b0 >> 3; + g0 = g0 >> 2; + r0 = r0 >> 3; + b1 = b1 >> 3; + g1 = g1 >> 2; + r1 = r1 >> 3; + *(uint32*)(dst_rgb565) = b0 | (g0 << 5) | (r0 << 11) | + (b1 << 16) | (g1 << 21) | (r1 << 27); + src_y += 2; + src_u += 1; + src_v += 1; + dst_rgb565 += 4; // Advance 2 pixels. + } + if (width & 1) { + YuvPixel(src_y[0], src_u[0], src_v[0], &b0, &g0, &r0); + b0 = b0 >> 3; + g0 = g0 >> 2; + r0 = r0 >> 3; + *(uint16*)(dst_rgb565) = b0 | (g0 << 5) | (r0 << 11); + } +} + +void I411ToARGBRow_C(const uint8* src_y, + const uint8* src_u, + const uint8* src_v, + uint8* rgb_buf, + int width) { + int x; + for (x = 0; x < width - 3; x += 4) { + YuvPixel(src_y[0], src_u[0], src_v[0], + rgb_buf + 0, rgb_buf + 1, rgb_buf + 2); + rgb_buf[3] = 255; + YuvPixel(src_y[1], src_u[0], src_v[0], + rgb_buf + 4, rgb_buf + 5, rgb_buf + 6); + rgb_buf[7] = 255; + YuvPixel(src_y[2], src_u[0], src_v[0], + rgb_buf + 8, rgb_buf + 9, rgb_buf + 10); + rgb_buf[11] = 255; + YuvPixel(src_y[3], src_u[0], src_v[0], + rgb_buf + 12, rgb_buf + 13, rgb_buf + 14); + rgb_buf[15] = 255; + src_y += 4; + src_u += 1; + src_v += 1; + rgb_buf += 16; // Advance 4 pixels. + } + if (width & 2) { + YuvPixel(src_y[0], src_u[0], src_v[0], + rgb_buf + 0, rgb_buf + 1, rgb_buf + 2); + rgb_buf[3] = 255; + YuvPixel(src_y[1], src_u[0], src_v[0], + rgb_buf + 4, rgb_buf + 5, rgb_buf + 6); + rgb_buf[7] = 255; + src_y += 2; + rgb_buf += 8; // Advance 2 pixels. + } + if (width & 1) { + YuvPixel(src_y[0], src_u[0], src_v[0], + rgb_buf + 0, rgb_buf + 1, rgb_buf + 2); + rgb_buf[3] = 255; + } +} + +void NV12ToARGBRow_C(const uint8* src_y, + const uint8* src_uv, + uint8* rgb_buf, + int width) { + int x; + for (x = 0; x < width - 1; x += 2) { + YuvPixel(src_y[0], src_uv[0], src_uv[1], + rgb_buf + 0, rgb_buf + 1, rgb_buf + 2); + rgb_buf[3] = 255; + YuvPixel(src_y[1], src_uv[0], src_uv[1], + rgb_buf + 4, rgb_buf + 5, rgb_buf + 6); + rgb_buf[7] = 255; + src_y += 2; + src_uv += 2; + rgb_buf += 8; // Advance 2 pixels. + } + if (width & 1) { + YuvPixel(src_y[0], src_uv[0], src_uv[1], + rgb_buf + 0, rgb_buf + 1, rgb_buf + 2); + rgb_buf[3] = 255; + } +} + +void NV21ToARGBRow_C(const uint8* src_y, + const uint8* src_vu, + uint8* rgb_buf, + int width) { + int x; + for (x = 0; x < width - 1; x += 2) { + YuvPixel(src_y[0], src_vu[1], src_vu[0], + rgb_buf + 0, rgb_buf + 1, rgb_buf + 2); + rgb_buf[3] = 255; + + YuvPixel(src_y[1], src_vu[1], src_vu[0], + rgb_buf + 4, rgb_buf + 5, rgb_buf + 6); + rgb_buf[7] = 255; + + src_y += 2; + src_vu += 2; + rgb_buf += 8; // Advance 2 pixels. + } + if (width & 1) { + YuvPixel(src_y[0], src_vu[1], src_vu[0], + rgb_buf + 0, rgb_buf + 1, rgb_buf + 2); + rgb_buf[3] = 255; + } +} + +void NV12ToRGB565Row_C(const uint8* src_y, + const uint8* src_uv, + uint8* dst_rgb565, + int width) { + uint8 b0; + uint8 g0; + uint8 r0; + uint8 b1; + uint8 g1; + uint8 r1; + int x; + for (x = 0; x < width - 1; x += 2) { + YuvPixel(src_y[0], src_uv[0], src_uv[1], &b0, &g0, &r0); + YuvPixel(src_y[1], src_uv[0], src_uv[1], &b1, &g1, &r1); + b0 = b0 >> 3; + g0 = g0 >> 2; + r0 = r0 >> 3; + b1 = b1 >> 3; + g1 = g1 >> 2; + r1 = r1 >> 3; + *(uint32*)(dst_rgb565) = b0 | (g0 << 5) | (r0 << 11) | + (b1 << 16) | (g1 << 21) | (r1 << 27); + src_y += 2; + src_uv += 2; + dst_rgb565 += 4; // Advance 2 pixels. + } + if (width & 1) { + YuvPixel(src_y[0], src_uv[0], src_uv[1], &b0, &g0, &r0); + b0 = b0 >> 3; + g0 = g0 >> 2; + r0 = r0 >> 3; + *(uint16*)(dst_rgb565) = b0 | (g0 << 5) | (r0 << 11); + } +} + +void NV21ToRGB565Row_C(const uint8* src_y, + const uint8* vsrc_u, + uint8* dst_rgb565, + int width) { + uint8 b0; + uint8 g0; + uint8 r0; + uint8 b1; + uint8 g1; + uint8 r1; + int x; + for (x = 0; x < width - 1; x += 2) { + YuvPixel(src_y[0], vsrc_u[1], vsrc_u[0], &b0, &g0, &r0); + YuvPixel(src_y[1], vsrc_u[1], vsrc_u[0], &b1, &g1, &r1); + b0 = b0 >> 3; + g0 = g0 >> 2; + r0 = r0 >> 3; + b1 = b1 >> 3; + g1 = g1 >> 2; + r1 = r1 >> 3; + *(uint32*)(dst_rgb565) = b0 | (g0 << 5) | (r0 << 11) | + (b1 << 16) | (g1 << 21) | (r1 << 27); + src_y += 2; + vsrc_u += 2; + dst_rgb565 += 4; // Advance 2 pixels. + } + if (width & 1) { + YuvPixel(src_y[0], vsrc_u[1], vsrc_u[0], &b0, &g0, &r0); + b0 = b0 >> 3; + g0 = g0 >> 2; + r0 = r0 >> 3; + *(uint16*)(dst_rgb565) = b0 | (g0 << 5) | (r0 << 11); + } +} + +void YUY2ToARGBRow_C(const uint8* src_yuy2, + uint8* rgb_buf, + int width) { + int x; + for (x = 0; x < width - 1; x += 2) { + YuvPixel(src_yuy2[0], src_yuy2[1], src_yuy2[3], + rgb_buf + 0, rgb_buf + 1, rgb_buf + 2); + rgb_buf[3] = 255; + YuvPixel(src_yuy2[2], src_yuy2[1], src_yuy2[3], + rgb_buf + 4, rgb_buf + 5, rgb_buf + 6); + rgb_buf[7] = 255; + src_yuy2 += 4; + rgb_buf += 8; // Advance 2 pixels. + } + if (width & 1) { + YuvPixel(src_yuy2[0], src_yuy2[1], src_yuy2[3], + rgb_buf + 0, rgb_buf + 1, rgb_buf + 2); + rgb_buf[3] = 255; + } +} + +void UYVYToARGBRow_C(const uint8* src_uyvy, + uint8* rgb_buf, + int width) { + int x; + for (x = 0; x < width - 1; x += 2) { + YuvPixel(src_uyvy[1], src_uyvy[0], src_uyvy[2], + rgb_buf + 0, rgb_buf + 1, rgb_buf + 2); + rgb_buf[3] = 255; + YuvPixel(src_uyvy[3], src_uyvy[0], src_uyvy[2], + rgb_buf + 4, rgb_buf + 5, rgb_buf + 6); + rgb_buf[7] = 255; + src_uyvy += 4; + rgb_buf += 8; // Advance 2 pixels. + } + if (width & 1) { + YuvPixel(src_uyvy[1], src_uyvy[0], src_uyvy[2], + rgb_buf + 0, rgb_buf + 1, rgb_buf + 2); + rgb_buf[3] = 255; + } +} + +void I422ToBGRARow_C(const uint8* src_y, + const uint8* src_u, + const uint8* src_v, + uint8* rgb_buf, + int width) { + int x; + for (x = 0; x < width - 1; x += 2) { + YuvPixel(src_y[0], src_u[0], src_v[0], + rgb_buf + 3, rgb_buf + 2, rgb_buf + 1); + rgb_buf[0] = 255; + YuvPixel(src_y[1], src_u[0], src_v[0], + rgb_buf + 7, rgb_buf + 6, rgb_buf + 5); + rgb_buf[4] = 255; + src_y += 2; + src_u += 1; + src_v += 1; + rgb_buf += 8; // Advance 2 pixels. + } + if (width & 1) { + YuvPixel(src_y[0], src_u[0], src_v[0], + rgb_buf + 3, rgb_buf + 2, rgb_buf + 1); + rgb_buf[0] = 255; + } +} + +void I422ToABGRRow_C(const uint8* src_y, + const uint8* src_u, + const uint8* src_v, + uint8* rgb_buf, + int width) { + int x; + for (x = 0; x < width - 1; x += 2) { + YuvPixel(src_y[0], src_u[0], src_v[0], + rgb_buf + 2, rgb_buf + 1, rgb_buf + 0); + rgb_buf[3] = 255; + YuvPixel(src_y[1], src_u[0], src_v[0], + rgb_buf + 6, rgb_buf + 5, rgb_buf + 4); + rgb_buf[7] = 255; + src_y += 2; + src_u += 1; + src_v += 1; + rgb_buf += 8; // Advance 2 pixels. + } + if (width & 1) { + YuvPixel(src_y[0], src_u[0], src_v[0], + rgb_buf + 2, rgb_buf + 1, rgb_buf + 0); + rgb_buf[3] = 255; + } +} + +void I422ToRGBARow_C(const uint8* src_y, + const uint8* src_u, + const uint8* src_v, + uint8* rgb_buf, + int width) { + int x; + for (x = 0; x < width - 1; x += 2) { + YuvPixel(src_y[0], src_u[0], src_v[0], + rgb_buf + 1, rgb_buf + 2, rgb_buf + 3); + rgb_buf[0] = 255; + YuvPixel(src_y[1], src_u[0], src_v[0], + rgb_buf + 5, rgb_buf + 6, rgb_buf + 7); + rgb_buf[4] = 255; + src_y += 2; + src_u += 1; + src_v += 1; + rgb_buf += 8; // Advance 2 pixels. + } + if (width & 1) { + YuvPixel(src_y[0], src_u[0], src_v[0], + rgb_buf + 1, rgb_buf + 2, rgb_buf + 3); + rgb_buf[0] = 255; + } +} + +void I400ToARGBRow_C(const uint8* src_y, uint8* rgb_buf, int width) { + int x; + for (x = 0; x < width - 1; x += 2) { + YPixel(src_y[0], rgb_buf + 0, rgb_buf + 1, rgb_buf + 2); + rgb_buf[3] = 255; + YPixel(src_y[1], rgb_buf + 4, rgb_buf + 5, rgb_buf + 6); + rgb_buf[7] = 255; + src_y += 2; + rgb_buf += 8; // Advance 2 pixels. + } + if (width & 1) { + YPixel(src_y[0], rgb_buf + 0, rgb_buf + 1, rgb_buf + 2); + rgb_buf[3] = 255; + } +} + +void MirrorRow_C(const uint8* src, uint8* dst, int width) { + int x; + src += width - 1; + for (x = 0; x < width - 1; x += 2) { + dst[x] = src[0]; + dst[x + 1] = src[-1]; + src -= 2; + } + if (width & 1) { + dst[width - 1] = src[0]; + } +} + +void MirrorUVRow_C(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int width) { + int x; + src_uv += (width - 1) << 1; + for (x = 0; x < width - 1; x += 2) { + dst_u[x] = src_uv[0]; + dst_u[x + 1] = src_uv[-2]; + dst_v[x] = src_uv[1]; + dst_v[x + 1] = src_uv[-2 + 1]; + src_uv -= 4; + } + if (width & 1) { + dst_u[width - 1] = src_uv[0]; + dst_v[width - 1] = src_uv[1]; + } +} + +void ARGBMirrorRow_C(const uint8* src, uint8* dst, int width) { + int x; + const uint32* src32 = (const uint32*)(src); + uint32* dst32 = (uint32*)(dst); + src32 += width - 1; + for (x = 0; x < width - 1; x += 2) { + dst32[x] = src32[0]; + dst32[x + 1] = src32[-1]; + src32 -= 2; + } + if (width & 1) { + dst32[width - 1] = src32[0]; + } +} + +void SplitUVRow_C(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int width) { + int x; + for (x = 0; x < width - 1; x += 2) { + dst_u[x] = src_uv[0]; + dst_u[x + 1] = src_uv[2]; + dst_v[x] = src_uv[1]; + dst_v[x + 1] = src_uv[3]; + src_uv += 4; + } + if (width & 1) { + dst_u[width - 1] = src_uv[0]; + dst_v[width - 1] = src_uv[1]; + } +} + +void MergeUVRow_C(const uint8* src_u, const uint8* src_v, uint8* dst_uv, + int width) { + int x; + for (x = 0; x < width - 1; x += 2) { + dst_uv[0] = src_u[x]; + dst_uv[1] = src_v[x]; + dst_uv[2] = src_u[x + 1]; + dst_uv[3] = src_v[x + 1]; + dst_uv += 4; + } + if (width & 1) { + dst_uv[0] = src_u[width - 1]; + dst_uv[1] = src_v[width - 1]; + } +} + +void CopyRow_C(const uint8* src, uint8* dst, int count) { + memcpy(dst, src, count); +} + +void CopyRow_16_C(const uint16* src, uint16* dst, int count) { + memcpy(dst, src, count * 2); +} + +void SetRow_C(uint8* dst, uint8 v8, int width) { + memset(dst, v8, width); +} + +void ARGBSetRow_C(uint8* dst_argb, uint32 v32, int width) { + uint32* d = (uint32*)(dst_argb); + int x; + for (x = 0; x < width; ++x) { + d[x] = v32; + } +} + +// Filter 2 rows of YUY2 UV's (422) into U and V (420). +void YUY2ToUVRow_C(const uint8* src_yuy2, int src_stride_yuy2, + uint8* dst_u, uint8* dst_v, int width) { + // Output a row of UV values, filtering 2 rows of YUY2. + int x; + for (x = 0; x < width; x += 2) { + dst_u[0] = (src_yuy2[1] + src_yuy2[src_stride_yuy2 + 1] + 1) >> 1; + dst_v[0] = (src_yuy2[3] + src_yuy2[src_stride_yuy2 + 3] + 1) >> 1; + src_yuy2 += 4; + dst_u += 1; + dst_v += 1; + } +} + +// Copy row of YUY2 UV's (422) into U and V (422). +void YUY2ToUV422Row_C(const uint8* src_yuy2, + uint8* dst_u, uint8* dst_v, int width) { + // Output a row of UV values. + int x; + for (x = 0; x < width; x += 2) { + dst_u[0] = src_yuy2[1]; + dst_v[0] = src_yuy2[3]; + src_yuy2 += 4; + dst_u += 1; + dst_v += 1; + } +} + +// Copy row of YUY2 Y's (422) into Y (420/422). +void YUY2ToYRow_C(const uint8* src_yuy2, uint8* dst_y, int width) { + // Output a row of Y values. + int x; + for (x = 0; x < width - 1; x += 2) { + dst_y[x] = src_yuy2[0]; + dst_y[x + 1] = src_yuy2[2]; + src_yuy2 += 4; + } + if (width & 1) { + dst_y[width - 1] = src_yuy2[0]; + } +} + +// Filter 2 rows of UYVY UV's (422) into U and V (420). +void UYVYToUVRow_C(const uint8* src_uyvy, int src_stride_uyvy, + uint8* dst_u, uint8* dst_v, int width) { + // Output a row of UV values. + int x; + for (x = 0; x < width; x += 2) { + dst_u[0] = (src_uyvy[0] + src_uyvy[src_stride_uyvy + 0] + 1) >> 1; + dst_v[0] = (src_uyvy[2] + src_uyvy[src_stride_uyvy + 2] + 1) >> 1; + src_uyvy += 4; + dst_u += 1; + dst_v += 1; + } +} + +// Copy row of UYVY UV's (422) into U and V (422). +void UYVYToUV422Row_C(const uint8* src_uyvy, + uint8* dst_u, uint8* dst_v, int width) { + // Output a row of UV values. + int x; + for (x = 0; x < width; x += 2) { + dst_u[0] = src_uyvy[0]; + dst_v[0] = src_uyvy[2]; + src_uyvy += 4; + dst_u += 1; + dst_v += 1; + } +} + +// Copy row of UYVY Y's (422) into Y (420/422). +void UYVYToYRow_C(const uint8* src_uyvy, uint8* dst_y, int width) { + // Output a row of Y values. + int x; + for (x = 0; x < width - 1; x += 2) { + dst_y[x] = src_uyvy[1]; + dst_y[x + 1] = src_uyvy[3]; + src_uyvy += 4; + } + if (width & 1) { + dst_y[width - 1] = src_uyvy[1]; + } +} + +#define BLEND(f, b, a) (((256 - a) * b) >> 8) + f + +// Blend src_argb0 over src_argb1 and store to dst_argb. +// dst_argb may be src_argb0 or src_argb1. +// This code mimics the SSSE3 version for better testability. +void ARGBBlendRow_C(const uint8* src_argb0, const uint8* src_argb1, + uint8* dst_argb, int width) { + int x; + for (x = 0; x < width - 1; x += 2) { + uint32 fb = src_argb0[0]; + uint32 fg = src_argb0[1]; + uint32 fr = src_argb0[2]; + uint32 a = src_argb0[3]; + uint32 bb = src_argb1[0]; + uint32 bg = src_argb1[1]; + uint32 br = src_argb1[2]; + dst_argb[0] = BLEND(fb, bb, a); + dst_argb[1] = BLEND(fg, bg, a); + dst_argb[2] = BLEND(fr, br, a); + dst_argb[3] = 255u; + + fb = src_argb0[4 + 0]; + fg = src_argb0[4 + 1]; + fr = src_argb0[4 + 2]; + a = src_argb0[4 + 3]; + bb = src_argb1[4 + 0]; + bg = src_argb1[4 + 1]; + br = src_argb1[4 + 2]; + dst_argb[4 + 0] = BLEND(fb, bb, a); + dst_argb[4 + 1] = BLEND(fg, bg, a); + dst_argb[4 + 2] = BLEND(fr, br, a); + dst_argb[4 + 3] = 255u; + src_argb0 += 8; + src_argb1 += 8; + dst_argb += 8; + } + + if (width & 1) { + uint32 fb = src_argb0[0]; + uint32 fg = src_argb0[1]; + uint32 fr = src_argb0[2]; + uint32 a = src_argb0[3]; + uint32 bb = src_argb1[0]; + uint32 bg = src_argb1[1]; + uint32 br = src_argb1[2]; + dst_argb[0] = BLEND(fb, bb, a); + dst_argb[1] = BLEND(fg, bg, a); + dst_argb[2] = BLEND(fr, br, a); + dst_argb[3] = 255u; + } +} +#undef BLEND +#define ATTENUATE(f, a) (a | (a << 8)) * (f | (f << 8)) >> 24 + +// Multiply source RGB by alpha and store to destination. +// This code mimics the SSSE3 version for better testability. +void ARGBAttenuateRow_C(const uint8* src_argb, uint8* dst_argb, int width) { + int i; + for (i = 0; i < width - 1; i += 2) { + uint32 b = src_argb[0]; + uint32 g = src_argb[1]; + uint32 r = src_argb[2]; + uint32 a = src_argb[3]; + dst_argb[0] = ATTENUATE(b, a); + dst_argb[1] = ATTENUATE(g, a); + dst_argb[2] = ATTENUATE(r, a); + dst_argb[3] = a; + b = src_argb[4]; + g = src_argb[5]; + r = src_argb[6]; + a = src_argb[7]; + dst_argb[4] = ATTENUATE(b, a); + dst_argb[5] = ATTENUATE(g, a); + dst_argb[6] = ATTENUATE(r, a); + dst_argb[7] = a; + src_argb += 8; + dst_argb += 8; + } + + if (width & 1) { + const uint32 b = src_argb[0]; + const uint32 g = src_argb[1]; + const uint32 r = src_argb[2]; + const uint32 a = src_argb[3]; + dst_argb[0] = ATTENUATE(b, a); + dst_argb[1] = ATTENUATE(g, a); + dst_argb[2] = ATTENUATE(r, a); + dst_argb[3] = a; + } +} +#undef ATTENUATE + +// Divide source RGB by alpha and store to destination. +// b = (b * 255 + (a / 2)) / a; +// g = (g * 255 + (a / 2)) / a; +// r = (r * 255 + (a / 2)) / a; +// Reciprocal method is off by 1 on some values. ie 125 +// 8.8 fixed point inverse table with 1.0 in upper short and 1 / a in lower. +#define T(a) 0x01000000 + (0x10000 / a) +const uint32 fixed_invtbl8[256] = { + 0x01000000, 0x0100ffff, T(0x02), T(0x03), T(0x04), T(0x05), T(0x06), T(0x07), + T(0x08), T(0x09), T(0x0a), T(0x0b), T(0x0c), T(0x0d), T(0x0e), T(0x0f), + T(0x10), T(0x11), T(0x12), T(0x13), T(0x14), T(0x15), T(0x16), T(0x17), + T(0x18), T(0x19), T(0x1a), T(0x1b), T(0x1c), T(0x1d), T(0x1e), T(0x1f), + T(0x20), T(0x21), T(0x22), T(0x23), T(0x24), T(0x25), T(0x26), T(0x27), + T(0x28), T(0x29), T(0x2a), T(0x2b), T(0x2c), T(0x2d), T(0x2e), T(0x2f), + T(0x30), T(0x31), T(0x32), T(0x33), T(0x34), T(0x35), T(0x36), T(0x37), + T(0x38), T(0x39), T(0x3a), T(0x3b), T(0x3c), T(0x3d), T(0x3e), T(0x3f), + T(0x40), T(0x41), T(0x42), T(0x43), T(0x44), T(0x45), T(0x46), T(0x47), + T(0x48), T(0x49), T(0x4a), T(0x4b), T(0x4c), T(0x4d), T(0x4e), T(0x4f), + T(0x50), T(0x51), T(0x52), T(0x53), T(0x54), T(0x55), T(0x56), T(0x57), + T(0x58), T(0x59), T(0x5a), T(0x5b), T(0x5c), T(0x5d), T(0x5e), T(0x5f), + T(0x60), T(0x61), T(0x62), T(0x63), T(0x64), T(0x65), T(0x66), T(0x67), + T(0x68), T(0x69), T(0x6a), T(0x6b), T(0x6c), T(0x6d), T(0x6e), T(0x6f), + T(0x70), T(0x71), T(0x72), T(0x73), T(0x74), T(0x75), T(0x76), T(0x77), + T(0x78), T(0x79), T(0x7a), T(0x7b), T(0x7c), T(0x7d), T(0x7e), T(0x7f), + T(0x80), T(0x81), T(0x82), T(0x83), T(0x84), T(0x85), T(0x86), T(0x87), + T(0x88), T(0x89), T(0x8a), T(0x8b), T(0x8c), T(0x8d), T(0x8e), T(0x8f), + T(0x90), T(0x91), T(0x92), T(0x93), T(0x94), T(0x95), T(0x96), T(0x97), + T(0x98), T(0x99), T(0x9a), T(0x9b), T(0x9c), T(0x9d), T(0x9e), T(0x9f), + T(0xa0), T(0xa1), T(0xa2), T(0xa3), T(0xa4), T(0xa5), T(0xa6), T(0xa7), + T(0xa8), T(0xa9), T(0xaa), T(0xab), T(0xac), T(0xad), T(0xae), T(0xaf), + T(0xb0), T(0xb1), T(0xb2), T(0xb3), T(0xb4), T(0xb5), T(0xb6), T(0xb7), + T(0xb8), T(0xb9), T(0xba), T(0xbb), T(0xbc), T(0xbd), T(0xbe), T(0xbf), + T(0xc0), T(0xc1), T(0xc2), T(0xc3), T(0xc4), T(0xc5), T(0xc6), T(0xc7), + T(0xc8), T(0xc9), T(0xca), T(0xcb), T(0xcc), T(0xcd), T(0xce), T(0xcf), + T(0xd0), T(0xd1), T(0xd2), T(0xd3), T(0xd4), T(0xd5), T(0xd6), T(0xd7), + T(0xd8), T(0xd9), T(0xda), T(0xdb), T(0xdc), T(0xdd), T(0xde), T(0xdf), + T(0xe0), T(0xe1), T(0xe2), T(0xe3), T(0xe4), T(0xe5), T(0xe6), T(0xe7), + T(0xe8), T(0xe9), T(0xea), T(0xeb), T(0xec), T(0xed), T(0xee), T(0xef), + T(0xf0), T(0xf1), T(0xf2), T(0xf3), T(0xf4), T(0xf5), T(0xf6), T(0xf7), + T(0xf8), T(0xf9), T(0xfa), T(0xfb), T(0xfc), T(0xfd), T(0xfe), 0x01000100 }; +#undef T + +void ARGBUnattenuateRow_C(const uint8* src_argb, uint8* dst_argb, int width) { + int i; + for (i = 0; i < width; ++i) { + uint32 b = src_argb[0]; + uint32 g = src_argb[1]; + uint32 r = src_argb[2]; + const uint32 a = src_argb[3]; + const uint32 ia = fixed_invtbl8[a] & 0xffff; // 8.8 fixed point + b = (b * ia) >> 8; + g = (g * ia) >> 8; + r = (r * ia) >> 8; + // Clamping should not be necessary but is free in assembly. + dst_argb[0] = clamp255(b); + dst_argb[1] = clamp255(g); + dst_argb[2] = clamp255(r); + dst_argb[3] = a; + src_argb += 4; + dst_argb += 4; + } +} + +void ComputeCumulativeSumRow_C(const uint8* row, int32* cumsum, + const int32* previous_cumsum, int width) { + int32 row_sum[4] = {0, 0, 0, 0}; + int x; + for (x = 0; x < width; ++x) { + row_sum[0] += row[x * 4 + 0]; + row_sum[1] += row[x * 4 + 1]; + row_sum[2] += row[x * 4 + 2]; + row_sum[3] += row[x * 4 + 3]; + cumsum[x * 4 + 0] = row_sum[0] + previous_cumsum[x * 4 + 0]; + cumsum[x * 4 + 1] = row_sum[1] + previous_cumsum[x * 4 + 1]; + cumsum[x * 4 + 2] = row_sum[2] + previous_cumsum[x * 4 + 2]; + cumsum[x * 4 + 3] = row_sum[3] + previous_cumsum[x * 4 + 3]; + } +} + +void CumulativeSumToAverageRow_C(const int32* tl, const int32* bl, + int w, int area, uint8* dst, int count) { + float ooa = 1.0f / area; + int i; + for (i = 0; i < count; ++i) { + dst[0] = (uint8)((bl[w + 0] + tl[0] - bl[0] - tl[w + 0]) * ooa); + dst[1] = (uint8)((bl[w + 1] + tl[1] - bl[1] - tl[w + 1]) * ooa); + dst[2] = (uint8)((bl[w + 2] + tl[2] - bl[2] - tl[w + 2]) * ooa); + dst[3] = (uint8)((bl[w + 3] + tl[3] - bl[3] - tl[w + 3]) * ooa); + dst += 4; + tl += 4; + bl += 4; + } +} + +// Copy pixels from rotated source to destination row with a slope. +LIBYUV_API +void ARGBAffineRow_C(const uint8* src_argb, int src_argb_stride, + uint8* dst_argb, const float* uv_dudv, int width) { + int i; + // Render a row of pixels from source into a buffer. + float uv[2]; + uv[0] = uv_dudv[0]; + uv[1] = uv_dudv[1]; + for (i = 0; i < width; ++i) { + int x = (int)(uv[0]); + int y = (int)(uv[1]); + *(uint32*)(dst_argb) = + *(const uint32*)(src_argb + y * src_argb_stride + + x * 4); + dst_argb += 4; + uv[0] += uv_dudv[2]; + uv[1] += uv_dudv[3]; + } +} + +// Blend 2 rows into 1. +static void HalfRow_C(const uint8* src_uv, int src_uv_stride, + uint8* dst_uv, int pix) { + int x; + for (x = 0; x < pix; ++x) { + dst_uv[x] = (src_uv[x] + src_uv[src_uv_stride + x] + 1) >> 1; + } +} + +static void HalfRow_16_C(const uint16* src_uv, int src_uv_stride, + uint16* dst_uv, int pix) { + int x; + for (x = 0; x < pix; ++x) { + dst_uv[x] = (src_uv[x] + src_uv[src_uv_stride + x] + 1) >> 1; + } +} + +// C version 2x2 -> 2x1. +void InterpolateRow_C(uint8* dst_ptr, const uint8* src_ptr, + ptrdiff_t src_stride, + int width, int source_y_fraction) { + int y1_fraction = source_y_fraction; + int y0_fraction = 256 - y1_fraction; + const uint8* src_ptr1 = src_ptr + src_stride; + int x; + if (source_y_fraction == 0) { + memcpy(dst_ptr, src_ptr, width); + return; + } + if (source_y_fraction == 128) { + HalfRow_C(src_ptr, (int)(src_stride), dst_ptr, width); + return; + } + for (x = 0; x < width - 1; x += 2) { + dst_ptr[0] = (src_ptr[0] * y0_fraction + src_ptr1[0] * y1_fraction) >> 8; + dst_ptr[1] = (src_ptr[1] * y0_fraction + src_ptr1[1] * y1_fraction) >> 8; + src_ptr += 2; + src_ptr1 += 2; + dst_ptr += 2; + } + if (width & 1) { + dst_ptr[0] = (src_ptr[0] * y0_fraction + src_ptr1[0] * y1_fraction) >> 8; + } +} + +void InterpolateRow_16_C(uint16* dst_ptr, const uint16* src_ptr, + ptrdiff_t src_stride, + int width, int source_y_fraction) { + int y1_fraction = source_y_fraction; + int y0_fraction = 256 - y1_fraction; + const uint16* src_ptr1 = src_ptr + src_stride; + int x; + if (source_y_fraction == 0) { + memcpy(dst_ptr, src_ptr, width * 2); + return; + } + if (source_y_fraction == 128) { + HalfRow_16_C(src_ptr, (int)(src_stride), dst_ptr, width); + return; + } + for (x = 0; x < width - 1; x += 2) { + dst_ptr[0] = (src_ptr[0] * y0_fraction + src_ptr1[0] * y1_fraction) >> 8; + dst_ptr[1] = (src_ptr[1] * y0_fraction + src_ptr1[1] * y1_fraction) >> 8; + src_ptr += 2; + src_ptr1 += 2; + dst_ptr += 2; + } + if (width & 1) { + dst_ptr[0] = (src_ptr[0] * y0_fraction + src_ptr1[0] * y1_fraction) >> 8; + } +} + +// Use first 4 shuffler values to reorder ARGB channels. +void ARGBShuffleRow_C(const uint8* src_argb, uint8* dst_argb, + const uint8* shuffler, int pix) { + int index0 = shuffler[0]; + int index1 = shuffler[1]; + int index2 = shuffler[2]; + int index3 = shuffler[3]; + // Shuffle a row of ARGB. + int x; + for (x = 0; x < pix; ++x) { + // To support in-place conversion. + uint8 b = src_argb[index0]; + uint8 g = src_argb[index1]; + uint8 r = src_argb[index2]; + uint8 a = src_argb[index3]; + dst_argb[0] = b; + dst_argb[1] = g; + dst_argb[2] = r; + dst_argb[3] = a; + src_argb += 4; + dst_argb += 4; + } +} + +void I422ToYUY2Row_C(const uint8* src_y, + const uint8* src_u, + const uint8* src_v, + uint8* dst_frame, int width) { + int x; + for (x = 0; x < width - 1; x += 2) { + dst_frame[0] = src_y[0]; + dst_frame[1] = src_u[0]; + dst_frame[2] = src_y[1]; + dst_frame[3] = src_v[0]; + dst_frame += 4; + src_y += 2; + src_u += 1; + src_v += 1; + } + if (width & 1) { + dst_frame[0] = src_y[0]; + dst_frame[1] = src_u[0]; + dst_frame[2] = 0; + dst_frame[3] = src_v[0]; + } +} + +void I422ToUYVYRow_C(const uint8* src_y, + const uint8* src_u, + const uint8* src_v, + uint8* dst_frame, int width) { + int x; + for (x = 0; x < width - 1; x += 2) { + dst_frame[0] = src_u[0]; + dst_frame[1] = src_y[0]; + dst_frame[2] = src_v[0]; + dst_frame[3] = src_y[1]; + dst_frame += 4; + src_y += 2; + src_u += 1; + src_v += 1; + } + if (width & 1) { + dst_frame[0] = src_u[0]; + dst_frame[1] = src_y[0]; + dst_frame[2] = src_v[0]; + dst_frame[3] = 0; + } +} + +// Maximum temporary width for wrappers to process at a time, in pixels. +#define MAXTWIDTH 2048 + +#if !(defined(_MSC_VER) && !defined(__clang__)) && \ + defined(HAS_I422TORGB565ROW_SSSE3) +// row_win.cc has asm version, but GCC uses 2 step wrapper. +void I422ToRGB565Row_SSSE3(const uint8* src_y, + const uint8* src_u, + const uint8* src_v, + uint8* dst_rgb565, + int width) { + SIMD_ALIGNED(uint8 row[MAXTWIDTH * 4]); + while (width > 0) { + int twidth = width > MAXTWIDTH ? MAXTWIDTH : width; + I422ToARGBRow_SSSE3(src_y, src_u, src_v, row, twidth); + ARGBToRGB565Row_SSE2(row, dst_rgb565, twidth); + src_y += twidth; + src_u += twidth / 2; + src_v += twidth / 2; + dst_rgb565 += twidth * 2; + width -= twidth; + } +} +#endif + +#if defined(HAS_I422TOARGB1555ROW_SSSE3) +void I422ToARGB1555Row_SSSE3(const uint8* src_y, + const uint8* src_u, + const uint8* src_v, + uint8* dst_argb1555, + int width) { + // Row buffer for intermediate ARGB pixels. + SIMD_ALIGNED(uint8 row[MAXTWIDTH * 4]); + while (width > 0) { + int twidth = width > MAXTWIDTH ? MAXTWIDTH : width; + I422ToARGBRow_SSSE3(src_y, src_u, src_v, row, twidth); + ARGBToARGB1555Row_SSE2(row, dst_argb1555, twidth); + src_y += twidth; + src_u += twidth / 2; + src_v += twidth / 2; + dst_argb1555 += twidth * 2; + width -= twidth; + } +} +#endif + +#if defined(HAS_I422TOARGB4444ROW_SSSE3) +void I422ToARGB4444Row_SSSE3(const uint8* src_y, + const uint8* src_u, + const uint8* src_v, + uint8* dst_argb4444, + int width) { + // Row buffer for intermediate ARGB pixels. + SIMD_ALIGNED(uint8 row[MAXTWIDTH * 4]); + while (width > 0) { + int twidth = width > MAXTWIDTH ? MAXTWIDTH : width; + I422ToARGBRow_SSSE3(src_y, src_u, src_v, row, twidth); + ARGBToARGB4444Row_SSE2(row, dst_argb4444, twidth); + src_y += twidth; + src_u += twidth / 2; + src_v += twidth / 2; + dst_argb4444 += twidth * 2; + width -= twidth; + } +} +#endif + +#if defined(HAS_NV12TORGB565ROW_SSSE3) +void NV12ToRGB565Row_SSSE3(const uint8* src_y, const uint8* src_uv, + uint8* dst_rgb565, int width) { + // Row buffer for intermediate ARGB pixels. + SIMD_ALIGNED(uint8 row[MAXTWIDTH * 4]); + while (width > 0) { + int twidth = width > MAXTWIDTH ? MAXTWIDTH : width; + NV12ToARGBRow_SSSE3(src_y, src_uv, row, twidth); + ARGBToRGB565Row_SSE2(row, dst_rgb565, twidth); + src_y += twidth; + src_uv += twidth; + dst_rgb565 += twidth * 2; + width -= twidth; + } +} +#endif + +#if defined(HAS_NV21TORGB565ROW_SSSE3) +void NV21ToRGB565Row_SSSE3(const uint8* src_y, const uint8* src_vu, + uint8* dst_rgb565, int width) { + // Row buffer for intermediate ARGB pixels. + SIMD_ALIGNED(uint8 row[MAXTWIDTH * 4]); + while (width > 0) { + int twidth = width > MAXTWIDTH ? MAXTWIDTH : width; + NV21ToARGBRow_SSSE3(src_y, src_vu, row, twidth); + ARGBToRGB565Row_SSE2(row, dst_rgb565, twidth); + src_y += twidth; + src_vu += twidth; + dst_rgb565 += twidth * 2; + width -= twidth; + } +} +#endif + +#if defined(HAS_YUY2TOARGBROW_SSSE3) +void YUY2ToARGBRow_SSSE3(const uint8* src_yuy2, uint8* dst_argb, int width) { + // Row buffers for intermediate YUV pixels. + SIMD_ALIGNED(uint8 row_y[MAXTWIDTH]); + SIMD_ALIGNED(uint8 row_u[MAXTWIDTH / 2]); + SIMD_ALIGNED(uint8 row_v[MAXTWIDTH / 2]); + while (width > 0) { + int twidth = width > MAXTWIDTH ? MAXTWIDTH : width; + YUY2ToUV422Row_SSE2(src_yuy2, row_u, row_v, twidth); + YUY2ToYRow_SSE2(src_yuy2, row_y, twidth); + I422ToARGBRow_SSSE3(row_y, row_u, row_v, dst_argb, twidth); + src_yuy2 += twidth * 2; + dst_argb += twidth * 4; + width -= twidth; + } +} +#endif + +#if defined(HAS_UYVYTOARGBROW_SSSE3) +void UYVYToARGBRow_SSSE3(const uint8* src_uyvy, uint8* dst_argb, int width) { + // Row buffers for intermediate YUV pixels. + SIMD_ALIGNED(uint8 row_y[MAXTWIDTH]); + SIMD_ALIGNED(uint8 row_u[MAXTWIDTH / 2]); + SIMD_ALIGNED(uint8 row_v[MAXTWIDTH / 2]); + while (width > 0) { + int twidth = width > MAXTWIDTH ? MAXTWIDTH : width; + UYVYToUV422Row_SSE2(src_uyvy, row_u, row_v, twidth); + UYVYToYRow_SSE2(src_uyvy, row_y, twidth); + I422ToARGBRow_SSSE3(row_y, row_u, row_v, dst_argb, twidth); + src_uyvy += twidth * 2; + dst_argb += twidth * 4; + width -= twidth; + } +} +#endif // !defined(LIBYUV_DISABLE_X86) + +#if defined(HAS_I422TORGB565ROW_AVX2) +void I422ToRGB565Row_AVX2(const uint8* src_y, + const uint8* src_u, + const uint8* src_v, + uint8* dst_rgb565, + int width) { + SIMD_ALIGNED32(uint8 row[MAXTWIDTH * 4]); + while (width > 0) { + int twidth = width > MAXTWIDTH ? MAXTWIDTH : width; + I422ToARGBRow_AVX2(src_y, src_u, src_v, row, twidth); + ARGBToRGB565Row_AVX2(row, dst_rgb565, twidth); + src_y += twidth; + src_u += twidth / 2; + src_v += twidth / 2; + dst_rgb565 += twidth * 2; + width -= twidth; + } +} +#endif + +#if defined(HAS_I422TOARGB1555ROW_AVX2) +void I422ToARGB1555Row_AVX2(const uint8* src_y, + const uint8* src_u, + const uint8* src_v, + uint8* dst_argb1555, + int width) { + // Row buffer for intermediate ARGB pixels. + SIMD_ALIGNED32(uint8 row[MAXTWIDTH * 4]); + while (width > 0) { + int twidth = width > MAXTWIDTH ? MAXTWIDTH : width; + I422ToARGBRow_AVX2(src_y, src_u, src_v, row, twidth); + ARGBToARGB1555Row_AVX2(row, dst_argb1555, twidth); + src_y += twidth; + src_u += twidth / 2; + src_v += twidth / 2; + dst_argb1555 += twidth * 2; + width -= twidth; + } +} +#endif + +#if defined(HAS_I422TOARGB4444ROW_AVX2) +void I422ToARGB4444Row_AVX2(const uint8* src_y, + const uint8* src_u, + const uint8* src_v, + uint8* dst_argb4444, + int width) { + // Row buffer for intermediate ARGB pixels. + SIMD_ALIGNED32(uint8 row[MAXTWIDTH * 4]); + while (width > 0) { + int twidth = width > MAXTWIDTH ? MAXTWIDTH : width; + I422ToARGBRow_AVX2(src_y, src_u, src_v, row, twidth); + ARGBToARGB4444Row_AVX2(row, dst_argb4444, twidth); + src_y += twidth; + src_u += twidth / 2; + src_v += twidth / 2; + dst_argb4444 += twidth * 2; + width -= twidth; + } +} +#endif + +#if defined(HAS_I422TORGB24ROW_AVX2) +void I422ToRGB24Row_AVX2(const uint8* src_y, + const uint8* src_u, + const uint8* src_v, + uint8* dst_rgb24, + int width) { + // Row buffer for intermediate ARGB pixels. + SIMD_ALIGNED32(uint8 row[MAXTWIDTH * 4]); + while (width > 0) { + int twidth = width > MAXTWIDTH ? MAXTWIDTH : width; + I422ToARGBRow_AVX2(src_y, src_u, src_v, row, twidth); + // TODO(fbarchard): ARGBToRGB24Row_AVX2 + ARGBToRGB24Row_SSSE3(row, dst_rgb24, twidth); + src_y += twidth; + src_u += twidth / 2; + src_v += twidth / 2; + dst_rgb24 += twidth * 3; + width -= twidth; + } +} +#endif + +#if defined(HAS_I422TORAWROW_AVX2) +void I422ToRAWRow_AVX2(const uint8* src_y, + const uint8* src_u, + const uint8* src_v, + uint8* dst_raw, + int width) { + // Row buffer for intermediate ARGB pixels. + SIMD_ALIGNED32(uint8 row[MAXTWIDTH * 4]); + while (width > 0) { + int twidth = width > MAXTWIDTH ? MAXTWIDTH : width; + I422ToARGBRow_AVX2(src_y, src_u, src_v, row, twidth); + // TODO(fbarchard): ARGBToRAWRow_AVX2 + ARGBToRAWRow_SSSE3(row, dst_raw, twidth); + src_y += twidth; + src_u += twidth / 2; + src_v += twidth / 2; + dst_raw += twidth * 3; + width -= twidth; + } +} +#endif + +#if defined(HAS_NV12TORGB565ROW_AVX2) +void NV12ToRGB565Row_AVX2(const uint8* src_y, const uint8* src_uv, + uint8* dst_rgb565, int width) { + // Row buffer for intermediate ARGB pixels. + SIMD_ALIGNED32(uint8 row[MAXTWIDTH * 4]); + while (width > 0) { + int twidth = width > MAXTWIDTH ? MAXTWIDTH : width; + NV12ToARGBRow_AVX2(src_y, src_uv, row, twidth); + ARGBToRGB565Row_AVX2(row, dst_rgb565, twidth); + src_y += twidth; + src_uv += twidth; + dst_rgb565 += twidth * 2; + width -= twidth; + } +} +#endif + +#if defined(HAS_NV21TORGB565ROW_AVX2) +void NV21ToRGB565Row_AVX2(const uint8* src_y, const uint8* src_vu, + uint8* dst_rgb565, int width) { + // Row buffer for intermediate ARGB pixels. + SIMD_ALIGNED32(uint8 row[MAXTWIDTH * 4]); + while (width > 0) { + int twidth = width > MAXTWIDTH ? MAXTWIDTH : width; + NV21ToARGBRow_AVX2(src_y, src_vu, row, twidth); + ARGBToRGB565Row_AVX2(row, dst_rgb565, twidth); + src_y += twidth; + src_vu += twidth; + dst_rgb565 += twidth * 2; + width -= twidth; + } +} +#endif + +#if defined(HAS_YUY2TOARGBROW_AVX2) +void YUY2ToARGBRow_AVX2(const uint8* src_yuy2, uint8* dst_argb, int width) { + // Row buffers for intermediate YUV pixels. + SIMD_ALIGNED32(uint8 row_y[MAXTWIDTH]); + SIMD_ALIGNED32(uint8 row_u[MAXTWIDTH / 2]); + SIMD_ALIGNED32(uint8 row_v[MAXTWIDTH / 2]); + while (width > 0) { + int twidth = width > MAXTWIDTH ? MAXTWIDTH : width; + YUY2ToUV422Row_AVX2(src_yuy2, row_u, row_v, twidth); + YUY2ToYRow_AVX2(src_yuy2, row_y, twidth); + I422ToARGBRow_AVX2(row_y, row_u, row_v, dst_argb, twidth); + src_yuy2 += twidth * 2; + dst_argb += twidth * 4; + width -= twidth; + } +} +#endif + +#if defined(HAS_UYVYTOARGBROW_AVX2) +void UYVYToARGBRow_AVX2(const uint8* src_uyvy, uint8* dst_argb, int width) { + // Row buffers for intermediate YUV pixels. + SIMD_ALIGNED32(uint8 row_y[MAXTWIDTH]); + SIMD_ALIGNED32(uint8 row_u[MAXTWIDTH / 2]); + SIMD_ALIGNED32(uint8 row_v[MAXTWIDTH / 2]); + while (width > 0) { + int twidth = width > MAXTWIDTH ? MAXTWIDTH : width; + UYVYToUV422Row_AVX2(src_uyvy, row_u, row_v, twidth); + UYVYToYRow_AVX2(src_uyvy, row_y, twidth); + I422ToARGBRow_AVX2(row_y, row_u, row_v, dst_argb, twidth); + src_uyvy += twidth * 2; + dst_argb += twidth * 4; + width -= twidth; + } +} +#endif // !defined(LIBYUV_DISABLE_X86) + +void ARGBPolynomialRow_C(const uint8* src_argb, + uint8* dst_argb, const float* poly, + int width) { + int i; + for (i = 0; i < width; ++i) { + float b = (float)(src_argb[0]); + float g = (float)(src_argb[1]); + float r = (float)(src_argb[2]); + float a = (float)(src_argb[3]); + float b2 = b * b; + float g2 = g * g; + float r2 = r * r; + float a2 = a * a; + float db = poly[0] + poly[4] * b; + float dg = poly[1] + poly[5] * g; + float dr = poly[2] + poly[6] * r; + float da = poly[3] + poly[7] * a; + float b3 = b2 * b; + float g3 = g2 * g; + float r3 = r2 * r; + float a3 = a2 * a; + db += poly[8] * b2; + dg += poly[9] * g2; + dr += poly[10] * r2; + da += poly[11] * a2; + db += poly[12] * b3; + dg += poly[13] * g3; + dr += poly[14] * r3; + da += poly[15] * a3; + + dst_argb[0] = Clamp((int32)(db)); + dst_argb[1] = Clamp((int32)(dg)); + dst_argb[2] = Clamp((int32)(dr)); + dst_argb[3] = Clamp((int32)(da)); + src_argb += 4; + dst_argb += 4; + } +} + +void ARGBLumaColorTableRow_C(const uint8* src_argb, uint8* dst_argb, int width, + const uint8* luma, uint32 lumacoeff) { + uint32 bc = lumacoeff & 0xff; + uint32 gc = (lumacoeff >> 8) & 0xff; + uint32 rc = (lumacoeff >> 16) & 0xff; + + int i; + for (i = 0; i < width - 1; i += 2) { + // Luminance in rows, color values in columns. + const uint8* luma0 = ((src_argb[0] * bc + src_argb[1] * gc + + src_argb[2] * rc) & 0x7F00u) + luma; + const uint8* luma1; + dst_argb[0] = luma0[src_argb[0]]; + dst_argb[1] = luma0[src_argb[1]]; + dst_argb[2] = luma0[src_argb[2]]; + dst_argb[3] = src_argb[3]; + luma1 = ((src_argb[4] * bc + src_argb[5] * gc + + src_argb[6] * rc) & 0x7F00u) + luma; + dst_argb[4] = luma1[src_argb[4]]; + dst_argb[5] = luma1[src_argb[5]]; + dst_argb[6] = luma1[src_argb[6]]; + dst_argb[7] = src_argb[7]; + src_argb += 8; + dst_argb += 8; + } + if (width & 1) { + // Luminance in rows, color values in columns. + const uint8* luma0 = ((src_argb[0] * bc + src_argb[1] * gc + + src_argb[2] * rc) & 0x7F00u) + luma; + dst_argb[0] = luma0[src_argb[0]]; + dst_argb[1] = luma0[src_argb[1]]; + dst_argb[2] = luma0[src_argb[2]]; + dst_argb[3] = src_argb[3]; + } +} + +void ARGBCopyAlphaRow_C(const uint8* src, uint8* dst, int width) { + int i; + for (i = 0; i < width - 1; i += 2) { + dst[3] = src[3]; + dst[7] = src[7]; + dst += 8; + src += 8; + } + if (width & 1) { + dst[3] = src[3]; + } +} + +void ARGBCopyYToAlphaRow_C(const uint8* src, uint8* dst, int width) { + int i; + for (i = 0; i < width - 1; i += 2) { + dst[3] = src[0]; + dst[7] = src[1]; + dst += 8; + src += 2; + } + if (width & 1) { + dst[3] = src[0]; + } +} + +#ifdef __cplusplus +} // extern "C" +} // namespace libyuv +#endif diff --git a/media/libaom/src/third_party/libyuv/source/row_gcc.cc b/media/libaom/src/third_party/libyuv/source/row_gcc.cc new file mode 100644 index 000000000..820de0a1c --- /dev/null +++ b/media/libaom/src/third_party/libyuv/source/row_gcc.cc @@ -0,0 +1,5475 @@ +// VERSION 2 +/* + * Copyright 2011 The LibYuv Project Authors. All rights reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#include "libyuv/row.h" + +#ifdef __cplusplus +namespace libyuv { +extern "C" { +#endif + +// This module is for GCC x86 and x64. +#if !defined(LIBYUV_DISABLE_X86) && (defined(__x86_64__) || defined(__i386__)) + +#if defined(HAS_ARGBTOYROW_SSSE3) || defined(HAS_ARGBGRAYROW_SSSE3) + +// Constants for ARGB +static vec8 kARGBToY = { + 13, 65, 33, 0, 13, 65, 33, 0, 13, 65, 33, 0, 13, 65, 33, 0 +}; + +// JPeg full range. +static vec8 kARGBToYJ = { + 15, 75, 38, 0, 15, 75, 38, 0, 15, 75, 38, 0, 15, 75, 38, 0 +}; +#endif // defined(HAS_ARGBTOYROW_SSSE3) || defined(HAS_ARGBGRAYROW_SSSE3) + +#if defined(HAS_ARGBTOYROW_SSSE3) || defined(HAS_I422TOARGBROW_SSSE3) + +static vec8 kARGBToU = { + 112, -74, -38, 0, 112, -74, -38, 0, 112, -74, -38, 0, 112, -74, -38, 0 +}; + +static vec8 kARGBToUJ = { + 127, -84, -43, 0, 127, -84, -43, 0, 127, -84, -43, 0, 127, -84, -43, 0 +}; + +static vec8 kARGBToV = { + -18, -94, 112, 0, -18, -94, 112, 0, -18, -94, 112, 0, -18, -94, 112, 0, +}; + +static vec8 kARGBToVJ = { + -20, -107, 127, 0, -20, -107, 127, 0, -20, -107, 127, 0, -20, -107, 127, 0 +}; + +// Constants for BGRA +static vec8 kBGRAToY = { + 0, 33, 65, 13, 0, 33, 65, 13, 0, 33, 65, 13, 0, 33, 65, 13 +}; + +static vec8 kBGRAToU = { + 0, -38, -74, 112, 0, -38, -74, 112, 0, -38, -74, 112, 0, -38, -74, 112 +}; + +static vec8 kBGRAToV = { + 0, 112, -94, -18, 0, 112, -94, -18, 0, 112, -94, -18, 0, 112, -94, -18 +}; + +// Constants for ABGR +static vec8 kABGRToY = { + 33, 65, 13, 0, 33, 65, 13, 0, 33, 65, 13, 0, 33, 65, 13, 0 +}; + +static vec8 kABGRToU = { + -38, -74, 112, 0, -38, -74, 112, 0, -38, -74, 112, 0, -38, -74, 112, 0 +}; + +static vec8 kABGRToV = { + 112, -94, -18, 0, 112, -94, -18, 0, 112, -94, -18, 0, 112, -94, -18, 0 +}; + +// Constants for RGBA. +static vec8 kRGBAToY = { + 0, 13, 65, 33, 0, 13, 65, 33, 0, 13, 65, 33, 0, 13, 65, 33 +}; + +static vec8 kRGBAToU = { + 0, 112, -74, -38, 0, 112, -74, -38, 0, 112, -74, -38, 0, 112, -74, -38 +}; + +static vec8 kRGBAToV = { + 0, -18, -94, 112, 0, -18, -94, 112, 0, -18, -94, 112, 0, -18, -94, 112 +}; + +static uvec8 kAddY16 = { + 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u +}; + +// 7 bit fixed point 0.5. +static vec16 kAddYJ64 = { + 64, 64, 64, 64, 64, 64, 64, 64 +}; + +static uvec8 kAddUV128 = { + 128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u, + 128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u +}; + +static uvec16 kAddUVJ128 = { + 0x8080u, 0x8080u, 0x8080u, 0x8080u, 0x8080u, 0x8080u, 0x8080u, 0x8080u +}; +#endif // defined(HAS_ARGBTOYROW_SSSE3) || defined(HAS_I422TOARGBROW_SSSE3) + +#ifdef HAS_RGB24TOARGBROW_SSSE3 + +// Shuffle table for converting RGB24 to ARGB. +static uvec8 kShuffleMaskRGB24ToARGB = { + 0u, 1u, 2u, 12u, 3u, 4u, 5u, 13u, 6u, 7u, 8u, 14u, 9u, 10u, 11u, 15u +}; + +// Shuffle table for converting RAW to ARGB. +static uvec8 kShuffleMaskRAWToARGB = { + 2u, 1u, 0u, 12u, 5u, 4u, 3u, 13u, 8u, 7u, 6u, 14u, 11u, 10u, 9u, 15u +}; + +// Shuffle table for converting ARGB to RGB24. +static uvec8 kShuffleMaskARGBToRGB24 = { + 0u, 1u, 2u, 4u, 5u, 6u, 8u, 9u, 10u, 12u, 13u, 14u, 128u, 128u, 128u, 128u +}; + +// Shuffle table for converting ARGB to RAW. +static uvec8 kShuffleMaskARGBToRAW = { + 2u, 1u, 0u, 6u, 5u, 4u, 10u, 9u, 8u, 14u, 13u, 12u, 128u, 128u, 128u, 128u +}; + +// Shuffle table for converting ARGBToRGB24 for I422ToRGB24. First 8 + next 4 +static uvec8 kShuffleMaskARGBToRGB24_0 = { + 0u, 1u, 2u, 4u, 5u, 6u, 8u, 9u, 128u, 128u, 128u, 128u, 10u, 12u, 13u, 14u +}; + +// Shuffle table for converting ARGB to RAW. +static uvec8 kShuffleMaskARGBToRAW_0 = { + 2u, 1u, 0u, 6u, 5u, 4u, 10u, 9u, 128u, 128u, 128u, 128u, 8u, 14u, 13u, 12u +}; +#endif // HAS_RGB24TOARGBROW_SSSE3 + +#if defined(TESTING) && defined(__x86_64__) +void TestRow_SSE2(const uint8* src_y, uint8* dst_argb, int pix) { + asm volatile ( + ".p2align 5 \n" + "mov %%eax,%%eax \n" + "mov %%ebx,%%ebx \n" + "mov %%ecx,%%ecx \n" + "mov %%edx,%%edx \n" + "mov %%esi,%%esi \n" + "mov %%edi,%%edi \n" + "mov %%ebp,%%ebp \n" + "mov %%esp,%%esp \n" + ".p2align 5 \n" + "mov %%r8d,%%r8d \n" + "mov %%r9d,%%r9d \n" + "mov %%r10d,%%r10d \n" + "mov %%r11d,%%r11d \n" + "mov %%r12d,%%r12d \n" + "mov %%r13d,%%r13d \n" + "mov %%r14d,%%r14d \n" + "mov %%r15d,%%r15d \n" + ".p2align 5 \n" + "lea (%%rax),%%eax \n" + "lea (%%rbx),%%ebx \n" + "lea (%%rcx),%%ecx \n" + "lea (%%rdx),%%edx \n" + "lea (%%rsi),%%esi \n" + "lea (%%rdi),%%edi \n" + "lea (%%rbp),%%ebp \n" + "lea (%%rsp),%%esp \n" + ".p2align 5 \n" + "lea (%%r8),%%r8d \n" + "lea (%%r9),%%r9d \n" + "lea (%%r10),%%r10d \n" + "lea (%%r11),%%r11d \n" + "lea (%%r12),%%r12d \n" + "lea (%%r13),%%r13d \n" + "lea (%%r14),%%r14d \n" + "lea (%%r15),%%r15d \n" + + ".p2align 5 \n" + "lea 0x10(%%rax),%%eax \n" + "lea 0x10(%%rbx),%%ebx \n" + "lea 0x10(%%rcx),%%ecx \n" + "lea 0x10(%%rdx),%%edx \n" + "lea 0x10(%%rsi),%%esi \n" + "lea 0x10(%%rdi),%%edi \n" + "lea 0x10(%%rbp),%%ebp \n" + "lea 0x10(%%rsp),%%esp \n" + ".p2align 5 \n" + "lea 0x10(%%r8),%%r8d \n" + "lea 0x10(%%r9),%%r9d \n" + "lea 0x10(%%r10),%%r10d \n" + "lea 0x10(%%r11),%%r11d \n" + "lea 0x10(%%r12),%%r12d \n" + "lea 0x10(%%r13),%%r13d \n" + "lea 0x10(%%r14),%%r14d \n" + "lea 0x10(%%r15),%%r15d \n" + + ".p2align 5 \n" + "add 0x10,%%eax \n" + "add 0x10,%%ebx \n" + "add 0x10,%%ecx \n" + "add 0x10,%%edx \n" + "add 0x10,%%esi \n" + "add 0x10,%%edi \n" + "add 0x10,%%ebp \n" + "add 0x10,%%esp \n" + ".p2align 5 \n" + "add 0x10,%%r8d \n" + "add 0x10,%%r9d \n" + "add 0x10,%%r10d \n" + "add 0x10,%%r11d \n" + "add 0x10,%%r12d \n" + "add 0x10,%%r13d \n" + "add 0x10,%%r14d \n" + "add 0x10,%%r15d \n" + + ".p2align 2 \n" + "1: \n" + "movq " MEMACCESS(0) ",%%xmm0 \n" + "lea " MEMLEA(0x8,0) ",%0 \n" + "movdqu %%xmm0," MEMACCESS(1) " \n" + "lea " MEMLEA(0x20,1) ",%1 \n" + "sub $0x8,%2 \n" + "jg 1b \n" + : "+r"(src_y), // %0 + "+r"(dst_argb), // %1 + "+r"(pix) // %2 + : + : "memory", "cc", "xmm0", "xmm1", "xmm5" + ); +} +#endif // TESTING + +#ifdef HAS_J400TOARGBROW_SSE2 +void J400ToARGBRow_SSE2(const uint8* src_y, uint8* dst_argb, int pix) { + asm volatile ( + "pcmpeqb %%xmm5,%%xmm5 \n" + "pslld $0x18,%%xmm5 \n" + LABELALIGN + "1: \n" + "movq " MEMACCESS(0) ",%%xmm0 \n" + "lea " MEMLEA(0x8,0) ",%0 \n" + "punpcklbw %%xmm0,%%xmm0 \n" + "movdqa %%xmm0,%%xmm1 \n" + "punpcklwd %%xmm0,%%xmm0 \n" + "punpckhwd %%xmm1,%%xmm1 \n" + "por %%xmm5,%%xmm0 \n" + "por %%xmm5,%%xmm1 \n" + "movdqu %%xmm0," MEMACCESS(1) " \n" + "movdqu %%xmm1," MEMACCESS2(0x10,1) " \n" + "lea " MEMLEA(0x20,1) ",%1 \n" + "sub $0x8,%2 \n" + "jg 1b \n" + : "+r"(src_y), // %0 + "+r"(dst_argb), // %1 + "+r"(pix) // %2 + :: "memory", "cc", "xmm0", "xmm1", "xmm5" + ); +} +#endif // HAS_J400TOARGBROW_SSE2 + +#ifdef HAS_RGB24TOARGBROW_SSSE3 +void RGB24ToARGBRow_SSSE3(const uint8* src_rgb24, uint8* dst_argb, int pix) { + asm volatile ( + "pcmpeqb %%xmm5,%%xmm5 \n" // generate mask 0xff000000 + "pslld $0x18,%%xmm5 \n" + "movdqa %3,%%xmm4 \n" + LABELALIGN + "1: \n" + "movdqu " MEMACCESS(0) ",%%xmm0 \n" + "movdqu " MEMACCESS2(0x10,0) ",%%xmm1 \n" + "movdqu " MEMACCESS2(0x20,0) ",%%xmm3 \n" + "lea " MEMLEA(0x30,0) ",%0 \n" + "movdqa %%xmm3,%%xmm2 \n" + "palignr $0x8,%%xmm1,%%xmm2 \n" + "pshufb %%xmm4,%%xmm2 \n" + "por %%xmm5,%%xmm2 \n" + "palignr $0xc,%%xmm0,%%xmm1 \n" + "pshufb %%xmm4,%%xmm0 \n" + "movdqu %%xmm2," MEMACCESS2(0x20,1) " \n" + "por %%xmm5,%%xmm0 \n" + "pshufb %%xmm4,%%xmm1 \n" + "movdqu %%xmm0," MEMACCESS(1) " \n" + "por %%xmm5,%%xmm1 \n" + "palignr $0x4,%%xmm3,%%xmm3 \n" + "pshufb %%xmm4,%%xmm3 \n" + "movdqu %%xmm1," MEMACCESS2(0x10,1) " \n" + "por %%xmm5,%%xmm3 \n" + "movdqu %%xmm3," MEMACCESS2(0x30,1) " \n" + "lea " MEMLEA(0x40,1) ",%1 \n" + "sub $0x10,%2 \n" + "jg 1b \n" + : "+r"(src_rgb24), // %0 + "+r"(dst_argb), // %1 + "+r"(pix) // %2 + : "m"(kShuffleMaskRGB24ToARGB) // %3 + : "memory", "cc" , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5" + ); +} + +void RAWToARGBRow_SSSE3(const uint8* src_raw, uint8* dst_argb, int pix) { + asm volatile ( + "pcmpeqb %%xmm5,%%xmm5 \n" // generate mask 0xff000000 + "pslld $0x18,%%xmm5 \n" + "movdqa %3,%%xmm4 \n" + LABELALIGN + "1: \n" + "movdqu " MEMACCESS(0) ",%%xmm0 \n" + "movdqu " MEMACCESS2(0x10,0) ",%%xmm1 \n" + "movdqu " MEMACCESS2(0x20,0) ",%%xmm3 \n" + "lea " MEMLEA(0x30,0) ",%0 \n" + "movdqa %%xmm3,%%xmm2 \n" + "palignr $0x8,%%xmm1,%%xmm2 \n" + "pshufb %%xmm4,%%xmm2 \n" + "por %%xmm5,%%xmm2 \n" + "palignr $0xc,%%xmm0,%%xmm1 \n" + "pshufb %%xmm4,%%xmm0 \n" + "movdqu %%xmm2," MEMACCESS2(0x20,1) " \n" + "por %%xmm5,%%xmm0 \n" + "pshufb %%xmm4,%%xmm1 \n" + "movdqu %%xmm0," MEMACCESS(1) " \n" + "por %%xmm5,%%xmm1 \n" + "palignr $0x4,%%xmm3,%%xmm3 \n" + "pshufb %%xmm4,%%xmm3 \n" + "movdqu %%xmm1," MEMACCESS2(0x10,1) " \n" + "por %%xmm5,%%xmm3 \n" + "movdqu %%xmm3," MEMACCESS2(0x30,1) " \n" + "lea " MEMLEA(0x40,1) ",%1 \n" + "sub $0x10,%2 \n" + "jg 1b \n" + : "+r"(src_raw), // %0 + "+r"(dst_argb), // %1 + "+r"(pix) // %2 + : "m"(kShuffleMaskRAWToARGB) // %3 + : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5" + ); +} + +void RGB565ToARGBRow_SSE2(const uint8* src, uint8* dst, int pix) { + asm volatile ( + "mov $0x1080108,%%eax \n" + "movd %%eax,%%xmm5 \n" + "pshufd $0x0,%%xmm5,%%xmm5 \n" + "mov $0x20802080,%%eax \n" + "movd %%eax,%%xmm6 \n" + "pshufd $0x0,%%xmm6,%%xmm6 \n" + "pcmpeqb %%xmm3,%%xmm3 \n" + "psllw $0xb,%%xmm3 \n" + "pcmpeqb %%xmm4,%%xmm4 \n" + "psllw $0xa,%%xmm4 \n" + "psrlw $0x5,%%xmm4 \n" + "pcmpeqb %%xmm7,%%xmm7 \n" + "psllw $0x8,%%xmm7 \n" + "sub %0,%1 \n" + "sub %0,%1 \n" + LABELALIGN + "1: \n" + "movdqu " MEMACCESS(0) ",%%xmm0 \n" + "movdqa %%xmm0,%%xmm1 \n" + "movdqa %%xmm0,%%xmm2 \n" + "pand %%xmm3,%%xmm1 \n" + "psllw $0xb,%%xmm2 \n" + "pmulhuw %%xmm5,%%xmm1 \n" + "pmulhuw %%xmm5,%%xmm2 \n" + "psllw $0x8,%%xmm1 \n" + "por %%xmm2,%%xmm1 \n" + "pand %%xmm4,%%xmm0 \n" + "pmulhuw %%xmm6,%%xmm0 \n" + "por %%xmm7,%%xmm0 \n" + "movdqa %%xmm1,%%xmm2 \n" + "punpcklbw %%xmm0,%%xmm1 \n" + "punpckhbw %%xmm0,%%xmm2 \n" + MEMOPMEM(movdqu,xmm1,0x00,1,0,2) // movdqu %%xmm1,(%1,%0,2) + MEMOPMEM(movdqu,xmm2,0x10,1,0,2) // movdqu %%xmm2,0x10(%1,%0,2) + "lea " MEMLEA(0x10,0) ",%0 \n" + "sub $0x8,%2 \n" + "jg 1b \n" + : "+r"(src), // %0 + "+r"(dst), // %1 + "+r"(pix) // %2 + : + : "memory", "cc", "eax", NACL_R14 + "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7" + ); +} + +void ARGB1555ToARGBRow_SSE2(const uint8* src, uint8* dst, int pix) { + asm volatile ( + "mov $0x1080108,%%eax \n" + "movd %%eax,%%xmm5 \n" + "pshufd $0x0,%%xmm5,%%xmm5 \n" + "mov $0x42004200,%%eax \n" + "movd %%eax,%%xmm6 \n" + "pshufd $0x0,%%xmm6,%%xmm6 \n" + "pcmpeqb %%xmm3,%%xmm3 \n" + "psllw $0xb,%%xmm3 \n" + "movdqa %%xmm3,%%xmm4 \n" + "psrlw $0x6,%%xmm4 \n" + "pcmpeqb %%xmm7,%%xmm7 \n" + "psllw $0x8,%%xmm7 \n" + "sub %0,%1 \n" + "sub %0,%1 \n" + LABELALIGN + "1: \n" + "movdqu " MEMACCESS(0) ",%%xmm0 \n" + "movdqa %%xmm0,%%xmm1 \n" + "movdqa %%xmm0,%%xmm2 \n" + "psllw $0x1,%%xmm1 \n" + "psllw $0xb,%%xmm2 \n" + "pand %%xmm3,%%xmm1 \n" + "pmulhuw %%xmm5,%%xmm2 \n" + "pmulhuw %%xmm5,%%xmm1 \n" + "psllw $0x8,%%xmm1 \n" + "por %%xmm2,%%xmm1 \n" + "movdqa %%xmm0,%%xmm2 \n" + "pand %%xmm4,%%xmm0 \n" + "psraw $0x8,%%xmm2 \n" + "pmulhuw %%xmm6,%%xmm0 \n" + "pand %%xmm7,%%xmm2 \n" + "por %%xmm2,%%xmm0 \n" + "movdqa %%xmm1,%%xmm2 \n" + "punpcklbw %%xmm0,%%xmm1 \n" + "punpckhbw %%xmm0,%%xmm2 \n" + MEMOPMEM(movdqu,xmm1,0x00,1,0,2) // movdqu %%xmm1,(%1,%0,2) + MEMOPMEM(movdqu,xmm2,0x10,1,0,2) // movdqu %%xmm2,0x10(%1,%0,2) + "lea " MEMLEA(0x10,0) ",%0 \n" + "sub $0x8,%2 \n" + "jg 1b \n" + : "+r"(src), // %0 + "+r"(dst), // %1 + "+r"(pix) // %2 + : + : "memory", "cc", "eax", NACL_R14 + "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7" + ); +} + +void ARGB4444ToARGBRow_SSE2(const uint8* src, uint8* dst, int pix) { + asm volatile ( + "mov $0xf0f0f0f,%%eax \n" + "movd %%eax,%%xmm4 \n" + "pshufd $0x0,%%xmm4,%%xmm4 \n" + "movdqa %%xmm4,%%xmm5 \n" + "pslld $0x4,%%xmm5 \n" + "sub %0,%1 \n" + "sub %0,%1 \n" + LABELALIGN + "1: \n" + "movdqu " MEMACCESS(0) ",%%xmm0 \n" + "movdqa %%xmm0,%%xmm2 \n" + "pand %%xmm4,%%xmm0 \n" + "pand %%xmm5,%%xmm2 \n" + "movdqa %%xmm0,%%xmm1 \n" + "movdqa %%xmm2,%%xmm3 \n" + "psllw $0x4,%%xmm1 \n" + "psrlw $0x4,%%xmm3 \n" + "por %%xmm1,%%xmm0 \n" + "por %%xmm3,%%xmm2 \n" + "movdqa %%xmm0,%%xmm1 \n" + "punpcklbw %%xmm2,%%xmm0 \n" + "punpckhbw %%xmm2,%%xmm1 \n" + MEMOPMEM(movdqu,xmm0,0x00,1,0,2) // movdqu %%xmm0,(%1,%0,2) + MEMOPMEM(movdqu,xmm1,0x10,1,0,2) // movdqu %%xmm1,0x10(%1,%0,2) + "lea " MEMLEA(0x10,0) ",%0 \n" + "sub $0x8,%2 \n" + "jg 1b \n" + : "+r"(src), // %0 + "+r"(dst), // %1 + "+r"(pix) // %2 + : + : "memory", "cc", "eax", NACL_R14 + "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5" + ); +} + +void ARGBToRGB24Row_SSSE3(const uint8* src, uint8* dst, int pix) { + asm volatile ( + "movdqa %3,%%xmm6 \n" + LABELALIGN + "1: \n" + "movdqu " MEMACCESS(0) ",%%xmm0 \n" + "movdqu " MEMACCESS2(0x10,0) ",%%xmm1 \n" + "movdqu " MEMACCESS2(0x20,0) ",%%xmm2 \n" + "movdqu " MEMACCESS2(0x30,0) ",%%xmm3 \n" + "lea " MEMLEA(0x40,0) ",%0 \n" + "pshufb %%xmm6,%%xmm0 \n" + "pshufb %%xmm6,%%xmm1 \n" + "pshufb %%xmm6,%%xmm2 \n" + "pshufb %%xmm6,%%xmm3 \n" + "movdqa %%xmm1,%%xmm4 \n" + "psrldq $0x4,%%xmm1 \n" + "pslldq $0xc,%%xmm4 \n" + "movdqa %%xmm2,%%xmm5 \n" + "por %%xmm4,%%xmm0 \n" + "pslldq $0x8,%%xmm5 \n" + "movdqu %%xmm0," MEMACCESS(1) " \n" + "por %%xmm5,%%xmm1 \n" + "psrldq $0x8,%%xmm2 \n" + "pslldq $0x4,%%xmm3 \n" + "por %%xmm3,%%xmm2 \n" + "movdqu %%xmm1," MEMACCESS2(0x10,1) " \n" + "movdqu %%xmm2," MEMACCESS2(0x20,1) " \n" + "lea " MEMLEA(0x30,1) ",%1 \n" + "sub $0x10,%2 \n" + "jg 1b \n" + : "+r"(src), // %0 + "+r"(dst), // %1 + "+r"(pix) // %2 + : "m"(kShuffleMaskARGBToRGB24) // %3 + : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6" + ); +} + +void ARGBToRAWRow_SSSE3(const uint8* src, uint8* dst, int pix) { + asm volatile ( + "movdqa %3,%%xmm6 \n" + LABELALIGN + "1: \n" + "movdqu " MEMACCESS(0) ",%%xmm0 \n" + "movdqu " MEMACCESS2(0x10,0) ",%%xmm1 \n" + "movdqu " MEMACCESS2(0x20,0) ",%%xmm2 \n" + "movdqu " MEMACCESS2(0x30,0) ",%%xmm3 \n" + "lea " MEMLEA(0x40,0) ",%0 \n" + "pshufb %%xmm6,%%xmm0 \n" + "pshufb %%xmm6,%%xmm1 \n" + "pshufb %%xmm6,%%xmm2 \n" + "pshufb %%xmm6,%%xmm3 \n" + "movdqa %%xmm1,%%xmm4 \n" + "psrldq $0x4,%%xmm1 \n" + "pslldq $0xc,%%xmm4 \n" + "movdqa %%xmm2,%%xmm5 \n" + "por %%xmm4,%%xmm0 \n" + "pslldq $0x8,%%xmm5 \n" + "movdqu %%xmm0," MEMACCESS(1) " \n" + "por %%xmm5,%%xmm1 \n" + "psrldq $0x8,%%xmm2 \n" + "pslldq $0x4,%%xmm3 \n" + "por %%xmm3,%%xmm2 \n" + "movdqu %%xmm1," MEMACCESS2(0x10,1) " \n" + "movdqu %%xmm2," MEMACCESS2(0x20,1) " \n" + "lea " MEMLEA(0x30,1) ",%1 \n" + "sub $0x10,%2 \n" + "jg 1b \n" + : "+r"(src), // %0 + "+r"(dst), // %1 + "+r"(pix) // %2 + : "m"(kShuffleMaskARGBToRAW) // %3 + : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6" + ); +} + +void ARGBToRGB565Row_SSE2(const uint8* src, uint8* dst, int pix) { + asm volatile ( + "pcmpeqb %%xmm3,%%xmm3 \n" + "psrld $0x1b,%%xmm3 \n" + "pcmpeqb %%xmm4,%%xmm4 \n" + "psrld $0x1a,%%xmm4 \n" + "pslld $0x5,%%xmm4 \n" + "pcmpeqb %%xmm5,%%xmm5 \n" + "pslld $0xb,%%xmm5 \n" + LABELALIGN + "1: \n" + "movdqu " MEMACCESS(0) ",%%xmm0 \n" + "movdqa %%xmm0,%%xmm1 \n" + "movdqa %%xmm0,%%xmm2 \n" + "pslld $0x8,%%xmm0 \n" + "psrld $0x3,%%xmm1 \n" + "psrld $0x5,%%xmm2 \n" + "psrad $0x10,%%xmm0 \n" + "pand %%xmm3,%%xmm1 \n" + "pand %%xmm4,%%xmm2 \n" + "pand %%xmm5,%%xmm0 \n" + "por %%xmm2,%%xmm1 \n" + "por %%xmm1,%%xmm0 \n" + "packssdw %%xmm0,%%xmm0 \n" + "lea " MEMLEA(0x10,0) ",%0 \n" + "movq %%xmm0," MEMACCESS(1) " \n" + "lea " MEMLEA(0x8,1) ",%1 \n" + "sub $0x4,%2 \n" + "jg 1b \n" + : "+r"(src), // %0 + "+r"(dst), // %1 + "+r"(pix) // %2 + :: "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5" + ); +} + +void ARGBToARGB1555Row_SSE2(const uint8* src, uint8* dst, int pix) { + asm volatile ( + "pcmpeqb %%xmm4,%%xmm4 \n" + "psrld $0x1b,%%xmm4 \n" + "movdqa %%xmm4,%%xmm5 \n" + "pslld $0x5,%%xmm5 \n" + "movdqa %%xmm4,%%xmm6 \n" + "pslld $0xa,%%xmm6 \n" + "pcmpeqb %%xmm7,%%xmm7 \n" + "pslld $0xf,%%xmm7 \n" + LABELALIGN + "1: \n" + "movdqu " MEMACCESS(0) ",%%xmm0 \n" + "movdqa %%xmm0,%%xmm1 \n" + "movdqa %%xmm0,%%xmm2 \n" + "movdqa %%xmm0,%%xmm3 \n" + "psrad $0x10,%%xmm0 \n" + "psrld $0x3,%%xmm1 \n" + "psrld $0x6,%%xmm2 \n" + "psrld $0x9,%%xmm3 \n" + "pand %%xmm7,%%xmm0 \n" + "pand %%xmm4,%%xmm1 \n" + "pand %%xmm5,%%xmm2 \n" + "pand %%xmm6,%%xmm3 \n" + "por %%xmm1,%%xmm0 \n" + "por %%xmm3,%%xmm2 \n" + "por %%xmm2,%%xmm0 \n" + "packssdw %%xmm0,%%xmm0 \n" + "lea " MEMLEA(0x10,0) ",%0 \n" + "movq %%xmm0," MEMACCESS(1) " \n" + "lea " MEMLEA(0x8,1) ",%1 \n" + "sub $0x4,%2 \n" + "jg 1b \n" + : "+r"(src), // %0 + "+r"(dst), // %1 + "+r"(pix) // %2 + :: "memory", "cc", + "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7" + ); +} + +void ARGBToARGB4444Row_SSE2(const uint8* src, uint8* dst, int pix) { + asm volatile ( + "pcmpeqb %%xmm4,%%xmm4 \n" + "psllw $0xc,%%xmm4 \n" + "movdqa %%xmm4,%%xmm3 \n" + "psrlw $0x8,%%xmm3 \n" + LABELALIGN + "1: \n" + "movdqu " MEMACCESS(0) ",%%xmm0 \n" + "movdqa %%xmm0,%%xmm1 \n" + "pand %%xmm3,%%xmm0 \n" + "pand %%xmm4,%%xmm1 \n" + "psrlq $0x4,%%xmm0 \n" + "psrlq $0x8,%%xmm1 \n" + "por %%xmm1,%%xmm0 \n" + "packuswb %%xmm0,%%xmm0 \n" + "lea " MEMLEA(0x10,0) ",%0 \n" + "movq %%xmm0," MEMACCESS(1) " \n" + "lea " MEMLEA(0x8,1) ",%1 \n" + "sub $0x4,%2 \n" + "jg 1b \n" + : "+r"(src), // %0 + "+r"(dst), // %1 + "+r"(pix) // %2 + :: "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4" + ); +} +#endif // HAS_RGB24TOARGBROW_SSSE3 + +#ifdef HAS_ARGBTOYROW_SSSE3 +// Convert 16 ARGB pixels (64 bytes) to 16 Y values. +void ARGBToYRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) { + asm volatile ( + "movdqa %3,%%xmm4 \n" + "movdqa %4,%%xmm5 \n" + LABELALIGN + "1: \n" + "movdqu " MEMACCESS(0) ",%%xmm0 \n" + "movdqu " MEMACCESS2(0x10,0) ",%%xmm1 \n" + "movdqu " MEMACCESS2(0x20,0) ",%%xmm2 \n" + "movdqu " MEMACCESS2(0x30,0) ",%%xmm3 \n" + "pmaddubsw %%xmm4,%%xmm0 \n" + "pmaddubsw %%xmm4,%%xmm1 \n" + "pmaddubsw %%xmm4,%%xmm2 \n" + "pmaddubsw %%xmm4,%%xmm3 \n" + "lea " MEMLEA(0x40,0) ",%0 \n" + "phaddw %%xmm1,%%xmm0 \n" + "phaddw %%xmm3,%%xmm2 \n" + "psrlw $0x7,%%xmm0 \n" + "psrlw $0x7,%%xmm2 \n" + "packuswb %%xmm2,%%xmm0 \n" + "paddb %%xmm5,%%xmm0 \n" + "movdqu %%xmm0," MEMACCESS(1) " \n" + "lea " MEMLEA(0x10,1) ",%1 \n" + "sub $0x10,%2 \n" + "jg 1b \n" + : "+r"(src_argb), // %0 + "+r"(dst_y), // %1 + "+r"(pix) // %2 + : "m"(kARGBToY), // %3 + "m"(kAddY16) // %4 + : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5" + ); +} +#endif // HAS_ARGBTOYROW_SSSE3 + +#ifdef HAS_ARGBTOYJROW_SSSE3 +// Convert 16 ARGB pixels (64 bytes) to 16 YJ values. +// Same as ARGBToYRow but different coefficients, no add 16, but do rounding. +void ARGBToYJRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) { + asm volatile ( + "movdqa %3,%%xmm4 \n" + "movdqa %4,%%xmm5 \n" + LABELALIGN + "1: \n" + "movdqu " MEMACCESS(0) ",%%xmm0 \n" + "movdqu " MEMACCESS2(0x10,0) ",%%xmm1 \n" + "movdqu " MEMACCESS2(0x20,0) ",%%xmm2 \n" + "movdqu " MEMACCESS2(0x30,0) ",%%xmm3 \n" + "pmaddubsw %%xmm4,%%xmm0 \n" + "pmaddubsw %%xmm4,%%xmm1 \n" + "pmaddubsw %%xmm4,%%xmm2 \n" + "pmaddubsw %%xmm4,%%xmm3 \n" + "lea " MEMLEA(0x40,0) ",%0 \n" + "phaddw %%xmm1,%%xmm0 \n" + "phaddw %%xmm3,%%xmm2 \n" + "paddw %%xmm5,%%xmm0 \n" + "paddw %%xmm5,%%xmm2 \n" + "psrlw $0x7,%%xmm0 \n" + "psrlw $0x7,%%xmm2 \n" + "packuswb %%xmm2,%%xmm0 \n" + "movdqu %%xmm0," MEMACCESS(1) " \n" + "lea " MEMLEA(0x10,1) ",%1 \n" + "sub $0x10,%2 \n" + "jg 1b \n" + : "+r"(src_argb), // %0 + "+r"(dst_y), // %1 + "+r"(pix) // %2 + : "m"(kARGBToYJ), // %3 + "m"(kAddYJ64) // %4 + : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5" + ); +} +#endif // HAS_ARGBTOYJROW_SSSE3 + +#ifdef HAS_ARGBTOYROW_AVX2 +// vpermd for vphaddw + vpackuswb vpermd. +static const lvec32 kPermdARGBToY_AVX = { + 0, 4, 1, 5, 2, 6, 3, 7 +}; + +// Convert 32 ARGB pixels (128 bytes) to 32 Y values. +void ARGBToYRow_AVX2(const uint8* src_argb, uint8* dst_y, int pix) { + asm volatile ( + "vbroadcastf128 %3,%%ymm4 \n" + "vbroadcastf128 %4,%%ymm5 \n" + "vmovdqu %5,%%ymm6 \n" + LABELALIGN + "1: \n" + "vmovdqu " MEMACCESS(0) ",%%ymm0 \n" + "vmovdqu " MEMACCESS2(0x20,0) ",%%ymm1 \n" + "vmovdqu " MEMACCESS2(0x40,0) ",%%ymm2 \n" + "vmovdqu " MEMACCESS2(0x60,0) ",%%ymm3 \n" + "vpmaddubsw %%ymm4,%%ymm0,%%ymm0 \n" + "vpmaddubsw %%ymm4,%%ymm1,%%ymm1 \n" + "vpmaddubsw %%ymm4,%%ymm2,%%ymm2 \n" + "vpmaddubsw %%ymm4,%%ymm3,%%ymm3 \n" + "lea " MEMLEA(0x80,0) ",%0 \n" + "vphaddw %%ymm1,%%ymm0,%%ymm0 \n" // mutates. + "vphaddw %%ymm3,%%ymm2,%%ymm2 \n" + "vpsrlw $0x7,%%ymm0,%%ymm0 \n" + "vpsrlw $0x7,%%ymm2,%%ymm2 \n" + "vpackuswb %%ymm2,%%ymm0,%%ymm0 \n" // mutates. + "vpermd %%ymm0,%%ymm6,%%ymm0 \n" // unmutate. + "vpaddb %%ymm5,%%ymm0,%%ymm0 \n" // add 16 for Y + "vmovdqu %%ymm0," MEMACCESS(1) " \n" + "lea " MEMLEA(0x20,1) ",%1 \n" + "sub $0x20,%2 \n" + "jg 1b \n" + "vzeroupper \n" + : "+r"(src_argb), // %0 + "+r"(dst_y), // %1 + "+r"(pix) // %2 + : "m"(kARGBToY), // %3 + "m"(kAddY16), // %4 + "m"(kPermdARGBToY_AVX) // %5 + : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6" + ); +} +#endif // HAS_ARGBTOYROW_AVX2 + +#ifdef HAS_ARGBTOYJROW_AVX2 +// Convert 32 ARGB pixels (128 bytes) to 32 Y values. +void ARGBToYJRow_AVX2(const uint8* src_argb, uint8* dst_y, int pix) { + asm volatile ( + "vbroadcastf128 %3,%%ymm4 \n" + "vbroadcastf128 %4,%%ymm5 \n" + "vmovdqu %5,%%ymm6 \n" + LABELALIGN + "1: \n" + "vmovdqu " MEMACCESS(0) ",%%ymm0 \n" + "vmovdqu " MEMACCESS2(0x20,0) ",%%ymm1 \n" + "vmovdqu " MEMACCESS2(0x40,0) ",%%ymm2 \n" + "vmovdqu " MEMACCESS2(0x60,0) ",%%ymm3 \n" + "vpmaddubsw %%ymm4,%%ymm0,%%ymm0 \n" + "vpmaddubsw %%ymm4,%%ymm1,%%ymm1 \n" + "vpmaddubsw %%ymm4,%%ymm2,%%ymm2 \n" + "vpmaddubsw %%ymm4,%%ymm3,%%ymm3 \n" + "lea " MEMLEA(0x80,0) ",%0 \n" + "vphaddw %%ymm1,%%ymm0,%%ymm0 \n" // mutates. + "vphaddw %%ymm3,%%ymm2,%%ymm2 \n" + "vpaddw %%ymm5,%%ymm0,%%ymm0 \n" // Add .5 for rounding. + "vpaddw %%ymm5,%%ymm2,%%ymm2 \n" + "vpsrlw $0x7,%%ymm0,%%ymm0 \n" + "vpsrlw $0x7,%%ymm2,%%ymm2 \n" + "vpackuswb %%ymm2,%%ymm0,%%ymm0 \n" // mutates. + "vpermd %%ymm0,%%ymm6,%%ymm0 \n" // unmutate. + "vmovdqu %%ymm0," MEMACCESS(1) " \n" + "lea " MEMLEA(0x20,1) ",%1 \n" + "sub $0x20,%2 \n" + "jg 1b \n" + "vzeroupper \n" + : "+r"(src_argb), // %0 + "+r"(dst_y), // %1 + "+r"(pix) // %2 + : "m"(kARGBToYJ), // %3 + "m"(kAddYJ64), // %4 + "m"(kPermdARGBToY_AVX) // %5 + : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6" + ); +} +#endif // HAS_ARGBTOYJROW_AVX2 + +#ifdef HAS_ARGBTOUVROW_SSSE3 +void ARGBToUVRow_SSSE3(const uint8* src_argb0, int src_stride_argb, + uint8* dst_u, uint8* dst_v, int width) { + asm volatile ( + "movdqa %5,%%xmm3 \n" + "movdqa %6,%%xmm4 \n" + "movdqa %7,%%xmm5 \n" + "sub %1,%2 \n" + LABELALIGN + "1: \n" + "movdqu " MEMACCESS(0) ",%%xmm0 \n" + MEMOPREG(movdqu,0x00,0,4,1,xmm7) // movdqu (%0,%4,1),%%xmm7 + "pavgb %%xmm7,%%xmm0 \n" + "movdqu " MEMACCESS2(0x10,0) ",%%xmm1 \n" + MEMOPREG(movdqu,0x10,0,4,1,xmm7) // movdqu 0x10(%0,%4,1),%%xmm7 + "pavgb %%xmm7,%%xmm1 \n" + "movdqu " MEMACCESS2(0x20,0) ",%%xmm2 \n" + MEMOPREG(movdqu,0x20,0,4,1,xmm7) // movdqu 0x20(%0,%4,1),%%xmm7 + "pavgb %%xmm7,%%xmm2 \n" + "movdqu " MEMACCESS2(0x30,0) ",%%xmm6 \n" + MEMOPREG(movdqu,0x30,0,4,1,xmm7) // movdqu 0x30(%0,%4,1),%%xmm7 + "pavgb %%xmm7,%%xmm6 \n" + + "lea " MEMLEA(0x40,0) ",%0 \n" + "movdqa %%xmm0,%%xmm7 \n" + "shufps $0x88,%%xmm1,%%xmm0 \n" + "shufps $0xdd,%%xmm1,%%xmm7 \n" + "pavgb %%xmm7,%%xmm0 \n" + "movdqa %%xmm2,%%xmm7 \n" + "shufps $0x88,%%xmm6,%%xmm2 \n" + "shufps $0xdd,%%xmm6,%%xmm7 \n" + "pavgb %%xmm7,%%xmm2 \n" + "movdqa %%xmm0,%%xmm1 \n" + "movdqa %%xmm2,%%xmm6 \n" + "pmaddubsw %%xmm4,%%xmm0 \n" + "pmaddubsw %%xmm4,%%xmm2 \n" + "pmaddubsw %%xmm3,%%xmm1 \n" + "pmaddubsw %%xmm3,%%xmm6 \n" + "phaddw %%xmm2,%%xmm0 \n" + "phaddw %%xmm6,%%xmm1 \n" + "psraw $0x8,%%xmm0 \n" + "psraw $0x8,%%xmm1 \n" + "packsswb %%xmm1,%%xmm0 \n" + "paddb %%xmm5,%%xmm0 \n" + "movlps %%xmm0," MEMACCESS(1) " \n" + MEMOPMEM(movhps,xmm0,0x00,1,2,1) // movhps %%xmm0,(%1,%2,1) + "lea " MEMLEA(0x8,1) ",%1 \n" + "sub $0x10,%3 \n" + "jg 1b \n" + : "+r"(src_argb0), // %0 + "+r"(dst_u), // %1 + "+r"(dst_v), // %2 + "+rm"(width) // %3 + : "r"((intptr_t)(src_stride_argb)), // %4 + "m"(kARGBToV), // %5 + "m"(kARGBToU), // %6 + "m"(kAddUV128) // %7 + : "memory", "cc", NACL_R14 + "xmm0", "xmm1", "xmm2", "xmm6", "xmm7" + ); +} +#endif // HAS_ARGBTOUVROW_SSSE3 + +#ifdef HAS_ARGBTOUVROW_AVX2 +// vpshufb for vphaddw + vpackuswb packed to shorts. +static const lvec8 kShufARGBToUV_AVX = { + 0, 1, 8, 9, 2, 3, 10, 11, 4, 5, 12, 13, 6, 7, 14, 15, + 0, 1, 8, 9, 2, 3, 10, 11, 4, 5, 12, 13, 6, 7, 14, 15 +}; +void ARGBToUVRow_AVX2(const uint8* src_argb0, int src_stride_argb, + uint8* dst_u, uint8* dst_v, int width) { + asm volatile ( + "vbroadcastf128 %5,%%ymm5 \n" + "vbroadcastf128 %6,%%ymm6 \n" + "vbroadcastf128 %7,%%ymm7 \n" + "sub %1,%2 \n" + LABELALIGN + "1: \n" + "vmovdqu " MEMACCESS(0) ",%%ymm0 \n" + "vmovdqu " MEMACCESS2(0x20,0) ",%%ymm1 \n" + "vmovdqu " MEMACCESS2(0x40,0) ",%%ymm2 \n" + "vmovdqu " MEMACCESS2(0x60,0) ",%%ymm3 \n" + VMEMOPREG(vpavgb,0x00,0,4,1,ymm0,ymm0) // vpavgb (%0,%4,1),%%ymm0,%%ymm0 + VMEMOPREG(vpavgb,0x20,0,4,1,ymm1,ymm1) + VMEMOPREG(vpavgb,0x40,0,4,1,ymm2,ymm2) + VMEMOPREG(vpavgb,0x60,0,4,1,ymm3,ymm3) + "lea " MEMLEA(0x80,0) ",%0 \n" + "vshufps $0x88,%%ymm1,%%ymm0,%%ymm4 \n" + "vshufps $0xdd,%%ymm1,%%ymm0,%%ymm0 \n" + "vpavgb %%ymm4,%%ymm0,%%ymm0 \n" + "vshufps $0x88,%%ymm3,%%ymm2,%%ymm4 \n" + "vshufps $0xdd,%%ymm3,%%ymm2,%%ymm2 \n" + "vpavgb %%ymm4,%%ymm2,%%ymm2 \n" + + "vpmaddubsw %%ymm7,%%ymm0,%%ymm1 \n" + "vpmaddubsw %%ymm7,%%ymm2,%%ymm3 \n" + "vpmaddubsw %%ymm6,%%ymm0,%%ymm0 \n" + "vpmaddubsw %%ymm6,%%ymm2,%%ymm2 \n" + "vphaddw %%ymm3,%%ymm1,%%ymm1 \n" + "vphaddw %%ymm2,%%ymm0,%%ymm0 \n" + "vpsraw $0x8,%%ymm1,%%ymm1 \n" + "vpsraw $0x8,%%ymm0,%%ymm0 \n" + "vpacksswb %%ymm0,%%ymm1,%%ymm0 \n" + "vpermq $0xd8,%%ymm0,%%ymm0 \n" + "vpshufb %8,%%ymm0,%%ymm0 \n" + "vpaddb %%ymm5,%%ymm0,%%ymm0 \n" + + "vextractf128 $0x0,%%ymm0," MEMACCESS(1) " \n" + VEXTOPMEM(vextractf128,1,ymm0,0x0,1,2,1) // vextractf128 $1,%%ymm0,(%1,%2,1) + "lea " MEMLEA(0x10,1) ",%1 \n" + "sub $0x20,%3 \n" + "jg 1b \n" + "vzeroupper \n" + : "+r"(src_argb0), // %0 + "+r"(dst_u), // %1 + "+r"(dst_v), // %2 + "+rm"(width) // %3 + : "r"((intptr_t)(src_stride_argb)), // %4 + "m"(kAddUV128), // %5 + "m"(kARGBToV), // %6 + "m"(kARGBToU), // %7 + "m"(kShufARGBToUV_AVX) // %8 + : "memory", "cc", NACL_R14 + "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7" + ); +} +#endif // HAS_ARGBTOUVROW_AVX2 + +#ifdef HAS_ARGBTOUVJROW_SSSE3 +void ARGBToUVJRow_SSSE3(const uint8* src_argb0, int src_stride_argb, + uint8* dst_u, uint8* dst_v, int width) { + asm volatile ( + "movdqa %5,%%xmm3 \n" + "movdqa %6,%%xmm4 \n" + "movdqa %7,%%xmm5 \n" + "sub %1,%2 \n" + LABELALIGN + "1: \n" + "movdqu " MEMACCESS(0) ",%%xmm0 \n" + MEMOPREG(movdqu,0x00,0,4,1,xmm7) // movdqu (%0,%4,1),%%xmm7 + "pavgb %%xmm7,%%xmm0 \n" + "movdqu " MEMACCESS2(0x10,0) ",%%xmm1 \n" + MEMOPREG(movdqu,0x10,0,4,1,xmm7) // movdqu 0x10(%0,%4,1),%%xmm7 + "pavgb %%xmm7,%%xmm1 \n" + "movdqu " MEMACCESS2(0x20,0) ",%%xmm2 \n" + MEMOPREG(movdqu,0x20,0,4,1,xmm7) // movdqu 0x20(%0,%4,1),%%xmm7 + "pavgb %%xmm7,%%xmm2 \n" + "movdqu " MEMACCESS2(0x30,0) ",%%xmm6 \n" + MEMOPREG(movdqu,0x30,0,4,1,xmm7) // movdqu 0x30(%0,%4,1),%%xmm7 + "pavgb %%xmm7,%%xmm6 \n" + + "lea " MEMLEA(0x40,0) ",%0 \n" + "movdqa %%xmm0,%%xmm7 \n" + "shufps $0x88,%%xmm1,%%xmm0 \n" + "shufps $0xdd,%%xmm1,%%xmm7 \n" + "pavgb %%xmm7,%%xmm0 \n" + "movdqa %%xmm2,%%xmm7 \n" + "shufps $0x88,%%xmm6,%%xmm2 \n" + "shufps $0xdd,%%xmm6,%%xmm7 \n" + "pavgb %%xmm7,%%xmm2 \n" + "movdqa %%xmm0,%%xmm1 \n" + "movdqa %%xmm2,%%xmm6 \n" + "pmaddubsw %%xmm4,%%xmm0 \n" + "pmaddubsw %%xmm4,%%xmm2 \n" + "pmaddubsw %%xmm3,%%xmm1 \n" + "pmaddubsw %%xmm3,%%xmm6 \n" + "phaddw %%xmm2,%%xmm0 \n" + "phaddw %%xmm6,%%xmm1 \n" + "paddw %%xmm5,%%xmm0 \n" + "paddw %%xmm5,%%xmm1 \n" + "psraw $0x8,%%xmm0 \n" + "psraw $0x8,%%xmm1 \n" + "packsswb %%xmm1,%%xmm0 \n" + "movlps %%xmm0," MEMACCESS(1) " \n" + MEMOPMEM(movhps,xmm0,0x00,1,2,1) // movhps %%xmm0,(%1,%2,1) + "lea " MEMLEA(0x8,1) ",%1 \n" + "sub $0x10,%3 \n" + "jg 1b \n" + : "+r"(src_argb0), // %0 + "+r"(dst_u), // %1 + "+r"(dst_v), // %2 + "+rm"(width) // %3 + : "r"((intptr_t)(src_stride_argb)), // %4 + "m"(kARGBToVJ), // %5 + "m"(kARGBToUJ), // %6 + "m"(kAddUVJ128) // %7 + : "memory", "cc", NACL_R14 + "xmm0", "xmm1", "xmm2", "xmm6", "xmm7" + ); +} +#endif // HAS_ARGBTOUVJROW_SSSE3 + +#ifdef HAS_ARGBTOUV444ROW_SSSE3 +void ARGBToUV444Row_SSSE3(const uint8* src_argb, uint8* dst_u, uint8* dst_v, + int width) { + asm volatile ( + "movdqa %4,%%xmm3 \n" + "movdqa %5,%%xmm4 \n" + "movdqa %6,%%xmm5 \n" + "sub %1,%2 \n" + LABELALIGN + "1: \n" + "movdqu " MEMACCESS(0) ",%%xmm0 \n" + "movdqu " MEMACCESS2(0x10,0) ",%%xmm1 \n" + "movdqu " MEMACCESS2(0x20,0) ",%%xmm2 \n" + "movdqu " MEMACCESS2(0x30,0) ",%%xmm6 \n" + "pmaddubsw %%xmm4,%%xmm0 \n" + "pmaddubsw %%xmm4,%%xmm1 \n" + "pmaddubsw %%xmm4,%%xmm2 \n" + "pmaddubsw %%xmm4,%%xmm6 \n" + "phaddw %%xmm1,%%xmm0 \n" + "phaddw %%xmm6,%%xmm2 \n" + "psraw $0x8,%%xmm0 \n" + "psraw $0x8,%%xmm2 \n" + "packsswb %%xmm2,%%xmm0 \n" + "paddb %%xmm5,%%xmm0 \n" + "movdqu %%xmm0," MEMACCESS(1) " \n" + "movdqu " MEMACCESS(0) ",%%xmm0 \n" + "movdqu " MEMACCESS2(0x10,0) ",%%xmm1 \n" + "movdqu " MEMACCESS2(0x20,0) ",%%xmm2 \n" + "movdqu " MEMACCESS2(0x30,0) ",%%xmm6 \n" + "pmaddubsw %%xmm3,%%xmm0 \n" + "pmaddubsw %%xmm3,%%xmm1 \n" + "pmaddubsw %%xmm3,%%xmm2 \n" + "pmaddubsw %%xmm3,%%xmm6 \n" + "phaddw %%xmm1,%%xmm0 \n" + "phaddw %%xmm6,%%xmm2 \n" + "psraw $0x8,%%xmm0 \n" + "psraw $0x8,%%xmm2 \n" + "packsswb %%xmm2,%%xmm0 \n" + "paddb %%xmm5,%%xmm0 \n" + "lea " MEMLEA(0x40,0) ",%0 \n" + MEMOPMEM(movdqu,xmm0,0x00,1,2,1) // movdqu %%xmm0,(%1,%2,1) + "lea " MEMLEA(0x10,1) ",%1 \n" + "sub $0x10,%3 \n" + "jg 1b \n" + : "+r"(src_argb), // %0 + "+r"(dst_u), // %1 + "+r"(dst_v), // %2 + "+rm"(width) // %3 + : "m"(kARGBToV), // %4 + "m"(kARGBToU), // %5 + "m"(kAddUV128) // %6 + : "memory", "cc", NACL_R14 + "xmm0", "xmm1", "xmm2", "xmm6" + ); +} +#endif // HAS_ARGBTOUV444ROW_SSSE3 + +#ifdef HAS_ARGBTOUV422ROW_SSSE3 +void ARGBToUV422Row_SSSE3(const uint8* src_argb0, + uint8* dst_u, uint8* dst_v, int width) { + asm volatile ( + "movdqa %4,%%xmm3 \n" + "movdqa %5,%%xmm4 \n" + "movdqa %6,%%xmm5 \n" + "sub %1,%2 \n" + LABELALIGN + "1: \n" + "movdqu " MEMACCESS(0) ",%%xmm0 \n" + "movdqu " MEMACCESS2(0x10,0) ",%%xmm1 \n" + "movdqu " MEMACCESS2(0x20,0) ",%%xmm2 \n" + "movdqu " MEMACCESS2(0x30,0) ",%%xmm6 \n" + "lea " MEMLEA(0x40,0) ",%0 \n" + "movdqa %%xmm0,%%xmm7 \n" + "shufps $0x88,%%xmm1,%%xmm0 \n" + "shufps $0xdd,%%xmm1,%%xmm7 \n" + "pavgb %%xmm7,%%xmm0 \n" + "movdqa %%xmm2,%%xmm7 \n" + "shufps $0x88,%%xmm6,%%xmm2 \n" + "shufps $0xdd,%%xmm6,%%xmm7 \n" + "pavgb %%xmm7,%%xmm2 \n" + "movdqa %%xmm0,%%xmm1 \n" + "movdqa %%xmm2,%%xmm6 \n" + "pmaddubsw %%xmm4,%%xmm0 \n" + "pmaddubsw %%xmm4,%%xmm2 \n" + "pmaddubsw %%xmm3,%%xmm1 \n" + "pmaddubsw %%xmm3,%%xmm6 \n" + "phaddw %%xmm2,%%xmm0 \n" + "phaddw %%xmm6,%%xmm1 \n" + "psraw $0x8,%%xmm0 \n" + "psraw $0x8,%%xmm1 \n" + "packsswb %%xmm1,%%xmm0 \n" + "paddb %%xmm5,%%xmm0 \n" + "movlps %%xmm0," MEMACCESS(1) " \n" + MEMOPMEM(movhps,xmm0,0x00,1,2,1) // movhps %%xmm0,(%1,%2,1) + "lea " MEMLEA(0x8,1) ",%1 \n" + "sub $0x10,%3 \n" + "jg 1b \n" + : "+r"(src_argb0), // %0 + "+r"(dst_u), // %1 + "+r"(dst_v), // %2 + "+rm"(width) // %3 + : "m"(kARGBToV), // %4 + "m"(kARGBToU), // %5 + "m"(kAddUV128) // %6 + : "memory", "cc", NACL_R14 + "xmm0", "xmm1", "xmm2", "xmm6", "xmm7" + ); +} +#endif // HAS_ARGBTOUV422ROW_SSSE3 + +void BGRAToYRow_SSSE3(const uint8* src_bgra, uint8* dst_y, int pix) { + asm volatile ( + "movdqa %4,%%xmm5 \n" + "movdqa %3,%%xmm4 \n" + LABELALIGN + "1: \n" + "movdqu " MEMACCESS(0) ",%%xmm0 \n" + "movdqu " MEMACCESS2(0x10,0) ",%%xmm1 \n" + "movdqu " MEMACCESS2(0x20,0) ",%%xmm2 \n" + "movdqu " MEMACCESS2(0x30,0) ",%%xmm3 \n" + "pmaddubsw %%xmm4,%%xmm0 \n" + "pmaddubsw %%xmm4,%%xmm1 \n" + "pmaddubsw %%xmm4,%%xmm2 \n" + "pmaddubsw %%xmm4,%%xmm3 \n" + "lea " MEMLEA(0x40,0) ",%0 \n" + "phaddw %%xmm1,%%xmm0 \n" + "phaddw %%xmm3,%%xmm2 \n" + "psrlw $0x7,%%xmm0 \n" + "psrlw $0x7,%%xmm2 \n" + "packuswb %%xmm2,%%xmm0 \n" + "paddb %%xmm5,%%xmm0 \n" + "movdqu %%xmm0," MEMACCESS(1) " \n" + "lea " MEMLEA(0x10,1) ",%1 \n" + "sub $0x10,%2 \n" + "jg 1b \n" + : "+r"(src_bgra), // %0 + "+r"(dst_y), // %1 + "+r"(pix) // %2 + : "m"(kBGRAToY), // %3 + "m"(kAddY16) // %4 + : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5" + ); +} + +void BGRAToUVRow_SSSE3(const uint8* src_bgra0, int src_stride_bgra, + uint8* dst_u, uint8* dst_v, int width) { + asm volatile ( + "movdqa %5,%%xmm3 \n" + "movdqa %6,%%xmm4 \n" + "movdqa %7,%%xmm5 \n" + "sub %1,%2 \n" + LABELALIGN + "1: \n" + "movdqu " MEMACCESS(0) ",%%xmm0 \n" + MEMOPREG(movdqu,0x00,0,4,1,xmm7) // movdqu (%0,%4,1),%%xmm7 + "pavgb %%xmm7,%%xmm0 \n" + "movdqu " MEMACCESS2(0x10,0) ",%%xmm1 \n" + MEMOPREG(movdqu,0x10,0,4,1,xmm7) // movdqu 0x10(%0,%4,1),%%xmm7 + "pavgb %%xmm7,%%xmm1 \n" + "movdqu " MEMACCESS2(0x20,0) ",%%xmm2 \n" + MEMOPREG(movdqu,0x20,0,4,1,xmm7) // movdqu 0x20(%0,%4,1),%%xmm7 + "pavgb %%xmm7,%%xmm2 \n" + "movdqu " MEMACCESS2(0x30,0) ",%%xmm6 \n" + MEMOPREG(movdqu,0x30,0,4,1,xmm7) // movdqu 0x30(%0,%4,1),%%xmm7 + "pavgb %%xmm7,%%xmm6 \n" + + "lea " MEMLEA(0x40,0) ",%0 \n" + "movdqa %%xmm0,%%xmm7 \n" + "shufps $0x88,%%xmm1,%%xmm0 \n" + "shufps $0xdd,%%xmm1,%%xmm7 \n" + "pavgb %%xmm7,%%xmm0 \n" + "movdqa %%xmm2,%%xmm7 \n" + "shufps $0x88,%%xmm6,%%xmm2 \n" + "shufps $0xdd,%%xmm6,%%xmm7 \n" + "pavgb %%xmm7,%%xmm2 \n" + "movdqa %%xmm0,%%xmm1 \n" + "movdqa %%xmm2,%%xmm6 \n" + "pmaddubsw %%xmm4,%%xmm0 \n" + "pmaddubsw %%xmm4,%%xmm2 \n" + "pmaddubsw %%xmm3,%%xmm1 \n" + "pmaddubsw %%xmm3,%%xmm6 \n" + "phaddw %%xmm2,%%xmm0 \n" + "phaddw %%xmm6,%%xmm1 \n" + "psraw $0x8,%%xmm0 \n" + "psraw $0x8,%%xmm1 \n" + "packsswb %%xmm1,%%xmm0 \n" + "paddb %%xmm5,%%xmm0 \n" + "movlps %%xmm0," MEMACCESS(1) " \n" + MEMOPMEM(movhps,xmm0,0x00,1,2,1) // movhps %%xmm0,(%1,%2,1) + "lea " MEMLEA(0x8,1) ",%1 \n" + "sub $0x10,%3 \n" + "jg 1b \n" + : "+r"(src_bgra0), // %0 + "+r"(dst_u), // %1 + "+r"(dst_v), // %2 + "+rm"(width) // %3 + : "r"((intptr_t)(src_stride_bgra)), // %4 + "m"(kBGRAToV), // %5 + "m"(kBGRAToU), // %6 + "m"(kAddUV128) // %7 + : "memory", "cc", NACL_R14 + "xmm0", "xmm1", "xmm2", "xmm6", "xmm7" + ); +} + +void ABGRToYRow_SSSE3(const uint8* src_abgr, uint8* dst_y, int pix) { + asm volatile ( + "movdqa %4,%%xmm5 \n" + "movdqa %3,%%xmm4 \n" + LABELALIGN + "1: \n" + "movdqu " MEMACCESS(0) ",%%xmm0 \n" + "movdqu " MEMACCESS2(0x10,0) ",%%xmm1 \n" + "movdqu " MEMACCESS2(0x20,0) ",%%xmm2 \n" + "movdqu " MEMACCESS2(0x30,0) ",%%xmm3 \n" + "pmaddubsw %%xmm4,%%xmm0 \n" + "pmaddubsw %%xmm4,%%xmm1 \n" + "pmaddubsw %%xmm4,%%xmm2 \n" + "pmaddubsw %%xmm4,%%xmm3 \n" + "lea " MEMLEA(0x40,0) ",%0 \n" + "phaddw %%xmm1,%%xmm0 \n" + "phaddw %%xmm3,%%xmm2 \n" + "psrlw $0x7,%%xmm0 \n" + "psrlw $0x7,%%xmm2 \n" + "packuswb %%xmm2,%%xmm0 \n" + "paddb %%xmm5,%%xmm0 \n" + "movdqu %%xmm0," MEMACCESS(1) " \n" + "lea " MEMLEA(0x10,1) ",%1 \n" + "sub $0x10,%2 \n" + "jg 1b \n" + : "+r"(src_abgr), // %0 + "+r"(dst_y), // %1 + "+r"(pix) // %2 + : "m"(kABGRToY), // %3 + "m"(kAddY16) // %4 + : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5" + ); +} + +void RGBAToYRow_SSSE3(const uint8* src_rgba, uint8* dst_y, int pix) { + asm volatile ( + "movdqa %4,%%xmm5 \n" + "movdqa %3,%%xmm4 \n" + LABELALIGN + "1: \n" + "movdqu " MEMACCESS(0) ",%%xmm0 \n" + "movdqu " MEMACCESS2(0x10,0) ",%%xmm1 \n" + "movdqu " MEMACCESS2(0x20,0) ",%%xmm2 \n" + "movdqu " MEMACCESS2(0x30,0) ",%%xmm3 \n" + "pmaddubsw %%xmm4,%%xmm0 \n" + "pmaddubsw %%xmm4,%%xmm1 \n" + "pmaddubsw %%xmm4,%%xmm2 \n" + "pmaddubsw %%xmm4,%%xmm3 \n" + "lea " MEMLEA(0x40,0) ",%0 \n" + "phaddw %%xmm1,%%xmm0 \n" + "phaddw %%xmm3,%%xmm2 \n" + "psrlw $0x7,%%xmm0 \n" + "psrlw $0x7,%%xmm2 \n" + "packuswb %%xmm2,%%xmm0 \n" + "paddb %%xmm5,%%xmm0 \n" + "movdqu %%xmm0," MEMACCESS(1) " \n" + "lea " MEMLEA(0x10,1) ",%1 \n" + "sub $0x10,%2 \n" + "jg 1b \n" + : "+r"(src_rgba), // %0 + "+r"(dst_y), // %1 + "+r"(pix) // %2 + : "m"(kRGBAToY), // %3 + "m"(kAddY16) // %4 + : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5" + ); +} + +void ABGRToUVRow_SSSE3(const uint8* src_abgr0, int src_stride_abgr, + uint8* dst_u, uint8* dst_v, int width) { + asm volatile ( + "movdqa %5,%%xmm3 \n" + "movdqa %6,%%xmm4 \n" + "movdqa %7,%%xmm5 \n" + "sub %1,%2 \n" + LABELALIGN + "1: \n" + "movdqu " MEMACCESS(0) ",%%xmm0 \n" + MEMOPREG(movdqu,0x00,0,4,1,xmm7) // movdqu (%0,%4,1),%%xmm7 + "pavgb %%xmm7,%%xmm0 \n" + "movdqu " MEMACCESS2(0x10,0) ",%%xmm1 \n" + MEMOPREG(movdqu,0x10,0,4,1,xmm7) // movdqu 0x10(%0,%4,1),%%xmm7 + "pavgb %%xmm7,%%xmm1 \n" + "movdqu " MEMACCESS2(0x20,0) ",%%xmm2 \n" + MEMOPREG(movdqu,0x20,0,4,1,xmm7) // movdqu 0x20(%0,%4,1),%%xmm7 + "pavgb %%xmm7,%%xmm2 \n" + "movdqu " MEMACCESS2(0x30,0) ",%%xmm6 \n" + MEMOPREG(movdqu,0x30,0,4,1,xmm7) // movdqu 0x30(%0,%4,1),%%xmm7 + "pavgb %%xmm7,%%xmm6 \n" + + "lea " MEMLEA(0x40,0) ",%0 \n" + "movdqa %%xmm0,%%xmm7 \n" + "shufps $0x88,%%xmm1,%%xmm0 \n" + "shufps $0xdd,%%xmm1,%%xmm7 \n" + "pavgb %%xmm7,%%xmm0 \n" + "movdqa %%xmm2,%%xmm7 \n" + "shufps $0x88,%%xmm6,%%xmm2 \n" + "shufps $0xdd,%%xmm6,%%xmm7 \n" + "pavgb %%xmm7,%%xmm2 \n" + "movdqa %%xmm0,%%xmm1 \n" + "movdqa %%xmm2,%%xmm6 \n" + "pmaddubsw %%xmm4,%%xmm0 \n" + "pmaddubsw %%xmm4,%%xmm2 \n" + "pmaddubsw %%xmm3,%%xmm1 \n" + "pmaddubsw %%xmm3,%%xmm6 \n" + "phaddw %%xmm2,%%xmm0 \n" + "phaddw %%xmm6,%%xmm1 \n" + "psraw $0x8,%%xmm0 \n" + "psraw $0x8,%%xmm1 \n" + "packsswb %%xmm1,%%xmm0 \n" + "paddb %%xmm5,%%xmm0 \n" + "movlps %%xmm0," MEMACCESS(1) " \n" + MEMOPMEM(movhps,xmm0,0x00,1,2,1) // movhps %%xmm0,(%1,%2,1) + "lea " MEMLEA(0x8,1) ",%1 \n" + "sub $0x10,%3 \n" + "jg 1b \n" + : "+r"(src_abgr0), // %0 + "+r"(dst_u), // %1 + "+r"(dst_v), // %2 + "+rm"(width) // %3 + : "r"((intptr_t)(src_stride_abgr)), // %4 + "m"(kABGRToV), // %5 + "m"(kABGRToU), // %6 + "m"(kAddUV128) // %7 + : "memory", "cc", NACL_R14 + "xmm0", "xmm1", "xmm2", "xmm6", "xmm7" + ); +} + +void RGBAToUVRow_SSSE3(const uint8* src_rgba0, int src_stride_rgba, + uint8* dst_u, uint8* dst_v, int width) { + asm volatile ( + "movdqa %5,%%xmm3 \n" + "movdqa %6,%%xmm4 \n" + "movdqa %7,%%xmm5 \n" + "sub %1,%2 \n" + LABELALIGN + "1: \n" + "movdqu " MEMACCESS(0) ",%%xmm0 \n" + MEMOPREG(movdqu,0x00,0,4,1,xmm7) // movdqu (%0,%4,1),%%xmm7 + "pavgb %%xmm7,%%xmm0 \n" + "movdqu " MEMACCESS2(0x10,0) ",%%xmm1 \n" + MEMOPREG(movdqu,0x10,0,4,1,xmm7) // movdqu 0x10(%0,%4,1),%%xmm7 + "pavgb %%xmm7,%%xmm1 \n" + "movdqu " MEMACCESS2(0x20,0) ",%%xmm2 \n" + MEMOPREG(movdqu,0x20,0,4,1,xmm7) // movdqu 0x20(%0,%4,1),%%xmm7 + "pavgb %%xmm7,%%xmm2 \n" + "movdqu " MEMACCESS2(0x30,0) ",%%xmm6 \n" + MEMOPREG(movdqu,0x30,0,4,1,xmm7) // movdqu 0x30(%0,%4,1),%%xmm7 + "pavgb %%xmm7,%%xmm6 \n" + + "lea " MEMLEA(0x40,0) ",%0 \n" + "movdqa %%xmm0,%%xmm7 \n" + "shufps $0x88,%%xmm1,%%xmm0 \n" + "shufps $0xdd,%%xmm1,%%xmm7 \n" + "pavgb %%xmm7,%%xmm0 \n" + "movdqa %%xmm2,%%xmm7 \n" + "shufps $0x88,%%xmm6,%%xmm2 \n" + "shufps $0xdd,%%xmm6,%%xmm7 \n" + "pavgb %%xmm7,%%xmm2 \n" + "movdqa %%xmm0,%%xmm1 \n" + "movdqa %%xmm2,%%xmm6 \n" + "pmaddubsw %%xmm4,%%xmm0 \n" + "pmaddubsw %%xmm4,%%xmm2 \n" + "pmaddubsw %%xmm3,%%xmm1 \n" + "pmaddubsw %%xmm3,%%xmm6 \n" + "phaddw %%xmm2,%%xmm0 \n" + "phaddw %%xmm6,%%xmm1 \n" + "psraw $0x8,%%xmm0 \n" + "psraw $0x8,%%xmm1 \n" + "packsswb %%xmm1,%%xmm0 \n" + "paddb %%xmm5,%%xmm0 \n" + "movlps %%xmm0," MEMACCESS(1) " \n" + MEMOPMEM(movhps,xmm0,0x00,1,2,1) // movhps %%xmm0,(%1,%2,1) + "lea " MEMLEA(0x8,1) ",%1 \n" + "sub $0x10,%3 \n" + "jg 1b \n" + : "+r"(src_rgba0), // %0 + "+r"(dst_u), // %1 + "+r"(dst_v), // %2 + "+rm"(width) // %3 + : "r"((intptr_t)(src_stride_rgba)), // %4 + "m"(kRGBAToV), // %5 + "m"(kRGBAToU), // %6 + "m"(kAddUV128) // %7 + : "memory", "cc", NACL_R14 + "xmm0", "xmm1", "xmm2", "xmm6", "xmm7" + ); +} + +#if defined(HAS_I422TOARGBROW_SSSE3) || defined(HAS_I422TOARGBROW_AVX2) + +struct YuvConstants { + lvec8 kUVToB; // 0 + lvec8 kUVToG; // 32 + lvec8 kUVToR; // 64 + lvec16 kUVBiasB; // 96 + lvec16 kUVBiasG; // 128 + lvec16 kUVBiasR; // 160 + lvec16 kYToRgb; // 192 +}; + +// BT.601 YUV to RGB reference +// R = (Y - 16) * 1.164 - V * -1.596 +// G = (Y - 16) * 1.164 - U * 0.391 - V * 0.813 +// B = (Y - 16) * 1.164 - U * -2.018 + +// Y contribution to R,G,B. Scale and bias. +// TODO(fbarchard): Consider moving constants into a common header. +#define YG 18997 /* round(1.164 * 64 * 256 * 256 / 257) */ +#define YGB -1160 /* 1.164 * 64 * -16 + 64 / 2 */ + +// U and V contributions to R,G,B. +#define UB -128 /* max(-128, round(-2.018 * 64)) */ +#define UG 25 /* round(0.391 * 64) */ +#define VG 52 /* round(0.813 * 64) */ +#define VR -102 /* round(-1.596 * 64) */ + +// Bias values to subtract 16 from Y and 128 from U and V. +#define BB (UB * 128 + YGB) +#define BG (UG * 128 + VG * 128 + YGB) +#define BR (VR * 128 + YGB) + +// BT601 constants for YUV to RGB. +static YuvConstants SIMD_ALIGNED(kYuvConstants) = { + { UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, + UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0 }, + { UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, + UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG }, + { 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, + 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR }, + { BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB }, + { BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG }, + { BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR }, + { YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG } +}; + +// BT601 constants for NV21 where chroma plane is VU instead of UV. +static YuvConstants SIMD_ALIGNED(kYvuConstants) = { + { 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, + 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB }, + { VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, + VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG }, + { VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, + VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0 }, + { BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB }, + { BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG }, + { BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR }, + { YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG } +}; + +#undef YG +#undef YGB +#undef UB +#undef UG +#undef VG +#undef VR +#undef BB +#undef BG +#undef BR + +// JPEG YUV to RGB reference +// * R = Y - V * -1.40200 +// * G = Y - U * 0.34414 - V * 0.71414 +// * B = Y - U * -1.77200 + +// Y contribution to R,G,B. Scale and bias. +// TODO(fbarchard): Consider moving constants into a common header. +#define YGJ 16320 /* round(1.000 * 64 * 256 * 256 / 257) */ +#define YGBJ 32 /* 64 / 2 */ + +// U and V contributions to R,G,B. +#define UBJ -113 /* round(-1.77200 * 64) */ +#define UGJ 22 /* round(0.34414 * 64) */ +#define VGJ 46 /* round(0.71414 * 64) */ +#define VRJ -90 /* round(-1.40200 * 64) */ + +// Bias values to subtract 16 from Y and 128 from U and V. +#define BBJ (UBJ * 128 + YGBJ) +#define BGJ (UGJ * 128 + VGJ * 128 + YGBJ) +#define BRJ (VRJ * 128 + YGBJ) + +// JPEG constants for YUV to RGB. +YuvConstants SIMD_ALIGNED(kYuvJConstants) = { + { UBJ, 0, UBJ, 0, UBJ, 0, UBJ, 0, UBJ, 0, UBJ, 0, UBJ, 0, UBJ, 0, + UBJ, 0, UBJ, 0, UBJ, 0, UBJ, 0, UBJ, 0, UBJ, 0, UBJ, 0, UBJ, 0 }, + { UGJ, VGJ, UGJ, VGJ, UGJ, VGJ, UGJ, VGJ, + UGJ, VGJ, UGJ, VGJ, UGJ, VGJ, UGJ, VGJ, + UGJ, VGJ, UGJ, VGJ, UGJ, VGJ, UGJ, VGJ, + UGJ, VGJ, UGJ, VGJ, UGJ, VGJ, UGJ, VGJ }, + { 0, VRJ, 0, VRJ, 0, VRJ, 0, VRJ, 0, VRJ, 0, VRJ, 0, VRJ, 0, VRJ, + 0, VRJ, 0, VRJ, 0, VRJ, 0, VRJ, 0, VRJ, 0, VRJ, 0, VRJ, 0, VRJ }, + { BBJ, BBJ, BBJ, BBJ, BBJ, BBJ, BBJ, BBJ, + BBJ, BBJ, BBJ, BBJ, BBJ, BBJ, BBJ, BBJ }, + { BGJ, BGJ, BGJ, BGJ, BGJ, BGJ, BGJ, BGJ, + BGJ, BGJ, BGJ, BGJ, BGJ, BGJ, BGJ, BGJ }, + { BRJ, BRJ, BRJ, BRJ, BRJ, BRJ, BRJ, BRJ, + BRJ, BRJ, BRJ, BRJ, BRJ, BRJ, BRJ, BRJ }, + { YGJ, YGJ, YGJ, YGJ, YGJ, YGJ, YGJ, YGJ, + YGJ, YGJ, YGJ, YGJ, YGJ, YGJ, YGJ, YGJ } +}; + +#undef YGJ +#undef YGBJ +#undef UBJ +#undef UGJ +#undef VGJ +#undef VRJ +#undef BBJ +#undef BGJ +#undef BRJ + +// Read 8 UV from 411 +#define READYUV444 \ + "movq " MEMACCESS([u_buf]) ",%%xmm0 \n" \ + MEMOPREG(movq, 0x00, [u_buf], [v_buf], 1, xmm1) \ + "lea " MEMLEA(0x8, [u_buf]) ",%[u_buf] \n" \ + "punpcklbw %%xmm1,%%xmm0 \n" + +// Read 4 UV from 422, upsample to 8 UV +#define READYUV422 \ + "movd " MEMACCESS([u_buf]) ",%%xmm0 \n" \ + MEMOPREG(movd, 0x00, [u_buf], [v_buf], 1, xmm1) \ + "lea " MEMLEA(0x4, [u_buf]) ",%[u_buf] \n" \ + "punpcklbw %%xmm1,%%xmm0 \n" \ + "punpcklwd %%xmm0,%%xmm0 \n" + +// Read 2 UV from 411, upsample to 8 UV +#define READYUV411 \ + "movd " MEMACCESS([u_buf]) ",%%xmm0 \n" \ + MEMOPREG(movd, 0x00, [u_buf], [v_buf], 1, xmm1) \ + "lea " MEMLEA(0x2, [u_buf]) ",%[u_buf] \n" \ + "punpcklbw %%xmm1,%%xmm0 \n" \ + "punpcklwd %%xmm0,%%xmm0 \n" \ + "punpckldq %%xmm0,%%xmm0 \n" + +// Read 4 UV from NV12, upsample to 8 UV +#define READNV12 \ + "movq " MEMACCESS([uv_buf]) ",%%xmm0 \n" \ + "lea " MEMLEA(0x8, [uv_buf]) ",%[uv_buf] \n" \ + "punpcklwd %%xmm0,%%xmm0 \n" + +// Convert 8 pixels: 8 UV and 8 Y +#define YUVTORGB(YuvConstants) \ + "movdqa %%xmm0,%%xmm1 \n" \ + "movdqa %%xmm0,%%xmm2 \n" \ + "movdqa %%xmm0,%%xmm3 \n" \ + "movdqa " MEMACCESS2(96, [YuvConstants]) ",%%xmm0 \n" \ + "pmaddubsw " MEMACCESS([YuvConstants]) ",%%xmm1 \n" \ + "psubw %%xmm1,%%xmm0 \n" \ + "movdqa " MEMACCESS2(128, [YuvConstants]) ",%%xmm1 \n" \ + "pmaddubsw " MEMACCESS2(32, [YuvConstants]) ",%%xmm2 \n" \ + "psubw %%xmm2,%%xmm1 \n" \ + "movdqa " MEMACCESS2(160, [YuvConstants]) ",%%xmm2 \n" \ + "pmaddubsw " MEMACCESS2(64, [YuvConstants]) ",%%xmm3 \n" \ + "psubw %%xmm3,%%xmm2 \n" \ + "movq " MEMACCESS([y_buf]) ",%%xmm3 \n" \ + "lea " MEMLEA(0x8, [y_buf]) ",%[y_buf] \n" \ + "punpcklbw %%xmm3,%%xmm3 \n" \ + "pmulhuw " MEMACCESS2(192, [YuvConstants]) ",%%xmm3 \n" \ + "paddsw %%xmm3,%%xmm0 \n" \ + "paddsw %%xmm3,%%xmm1 \n" \ + "paddsw %%xmm3,%%xmm2 \n" \ + "psraw $0x6,%%xmm0 \n" \ + "psraw $0x6,%%xmm1 \n" \ + "psraw $0x6,%%xmm2 \n" \ + "packuswb %%xmm0,%%xmm0 \n" \ + "packuswb %%xmm1,%%xmm1 \n" \ + "packuswb %%xmm2,%%xmm2 \n" + +// Store 8 ARGB values. Assumes XMM5 is zero. +#define STOREARGB \ + "punpcklbw %%xmm1,%%xmm0 \n" \ + "punpcklbw %%xmm5,%%xmm2 \n" \ + "movdqa %%xmm0,%%xmm1 \n" \ + "punpcklwd %%xmm2,%%xmm0 \n" \ + "punpckhwd %%xmm2,%%xmm1 \n" \ + "movdqu %%xmm0," MEMACCESS([dst_argb]) " \n" \ + "movdqu %%xmm1," MEMACCESS2(0x10, [dst_argb]) " \n" \ + "lea " MEMLEA(0x20, [dst_argb]) ", %[dst_argb] \n" + +// Store 8 BGRA values. Assumes XMM5 is zero. +#define STOREBGRA \ + "pcmpeqb %%xmm5,%%xmm5 \n" \ + "punpcklbw %%xmm0,%%xmm1 \n" \ + "punpcklbw %%xmm2,%%xmm5 \n" \ + "movdqa %%xmm5,%%xmm0 \n" \ + "punpcklwd %%xmm1,%%xmm5 \n" \ + "punpckhwd %%xmm1,%%xmm0 \n" \ + "movdqu %%xmm5," MEMACCESS([dst_bgra]) " \n" \ + "movdqu %%xmm0," MEMACCESS2(0x10, [dst_bgra]) " \n" \ + "lea " MEMLEA(0x20, [dst_bgra]) ", %[dst_bgra] \n" + +// Store 8 ABGR values. Assumes XMM5 is zero. +#define STOREABGR \ + "punpcklbw %%xmm1,%%xmm2 \n" \ + "punpcklbw %%xmm5,%%xmm0 \n" \ + "movdqa %%xmm2,%%xmm1 \n" \ + "punpcklwd %%xmm0,%%xmm2 \n" \ + "punpckhwd %%xmm0,%%xmm1 \n" \ + "movdqu %%xmm2," MEMACCESS([dst_abgr]) " \n" \ + "movdqu %%xmm1," MEMACCESS2(0x10, [dst_abgr]) " \n" \ + "lea " MEMLEA(0x20, [dst_abgr]) ", %[dst_abgr] \n" + +// Store 8 RGBA values. Assumes XMM5 is zero. +#define STORERGBA \ + "pcmpeqb %%xmm5,%%xmm5 \n" \ + "punpcklbw %%xmm2,%%xmm1 \n" \ + "punpcklbw %%xmm0,%%xmm5 \n" \ + "movdqa %%xmm5,%%xmm0 \n" \ + "punpcklwd %%xmm1,%%xmm5 \n" \ + "punpckhwd %%xmm1,%%xmm0 \n" \ + "movdqu %%xmm5," MEMACCESS([dst_rgba]) " \n" \ + "movdqu %%xmm0," MEMACCESS2(0x10, [dst_rgba]) " \n" \ + "lea " MEMLEA(0x20, [dst_rgba]) ",%[dst_rgba] \n" + +void OMITFP I444ToARGBRow_SSSE3(const uint8* y_buf, + const uint8* u_buf, + const uint8* v_buf, + uint8* dst_argb, + int width) { + asm volatile ( + "sub %[u_buf],%[v_buf] \n" + "pcmpeqb %%xmm5,%%xmm5 \n" + LABELALIGN + "1: \n" + READYUV444 + YUVTORGB(kYuvConstants) + STOREARGB + "sub $0x8,%[width] \n" + "jg 1b \n" + : [y_buf]"+r"(y_buf), // %[y_buf] + [u_buf]"+r"(u_buf), // %[u_buf] + [v_buf]"+r"(v_buf), // %[v_buf] + [dst_argb]"+r"(dst_argb), // %[dst_argb] + [width]"+rm"(width) // %[width] + : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants] + : "memory", "cc", NACL_R14 + "xmm0", "xmm1", "xmm2", "xmm3", "xmm5" + ); +} + +// TODO(fbarchard): Consider putting masks into constants. +void OMITFP I422ToRGB24Row_SSSE3(const uint8* y_buf, + const uint8* u_buf, + const uint8* v_buf, + uint8* dst_rgb24, + int width) { + asm volatile ( + "movdqa %[kShuffleMaskARGBToRGB24_0],%%xmm5 \n" + "movdqa %[kShuffleMaskARGBToRGB24],%%xmm6 \n" + "sub %[u_buf],%[v_buf] \n" + LABELALIGN + "1: \n" + READYUV422 + YUVTORGB(kYuvConstants) + "punpcklbw %%xmm1,%%xmm0 \n" + "punpcklbw %%xmm2,%%xmm2 \n" + "movdqa %%xmm0,%%xmm1 \n" + "punpcklwd %%xmm2,%%xmm0 \n" + "punpckhwd %%xmm2,%%xmm1 \n" + "pshufb %%xmm5,%%xmm0 \n" + "pshufb %%xmm6,%%xmm1 \n" + "palignr $0xc,%%xmm0,%%xmm1 \n" + "movq %%xmm0," MEMACCESS([dst_rgb24]) "\n" + "movdqu %%xmm1," MEMACCESS2(0x8,[dst_rgb24]) "\n" + "lea " MEMLEA(0x18,[dst_rgb24]) ",%[dst_rgb24] \n" + "subl $0x8,%[width] \n" + "jg 1b \n" + : [y_buf]"+r"(y_buf), // %[y_buf] + [u_buf]"+r"(u_buf), // %[u_buf] + [v_buf]"+r"(v_buf), // %[v_buf] + [dst_rgb24]"+r"(dst_rgb24), // %[dst_rgb24] +// TODO(fbarchard): Make width a register for 32 bit. +#if defined(__i386__) && defined(__pic__) + [width]"+m"(width) // %[width] +#else + [width]"+rm"(width) // %[width] +#endif + : [kYuvConstants]"r"(&kYuvConstants.kUVToB), + [kShuffleMaskARGBToRGB24_0]"m"(kShuffleMaskARGBToRGB24_0), + [kShuffleMaskARGBToRGB24]"m"(kShuffleMaskARGBToRGB24) + : "memory", "cc", NACL_R14 + "xmm0", "xmm1", "xmm2", "xmm3", "xmm5", "xmm6" + ); +} + +void OMITFP I422ToRAWRow_SSSE3(const uint8* y_buf, + const uint8* u_buf, + const uint8* v_buf, + uint8* dst_raw, + int width) { + asm volatile ( + "movdqa %[kShuffleMaskARGBToRAW_0],%%xmm5 \n" + "movdqa %[kShuffleMaskARGBToRAW],%%xmm6 \n" + "sub %[u_buf],%[v_buf] \n" + LABELALIGN + "1: \n" + READYUV422 + YUVTORGB(kYuvConstants) + "punpcklbw %%xmm1,%%xmm0 \n" + "punpcklbw %%xmm2,%%xmm2 \n" + "movdqa %%xmm0,%%xmm1 \n" + "punpcklwd %%xmm2,%%xmm0 \n" + "punpckhwd %%xmm2,%%xmm1 \n" + "pshufb %%xmm5,%%xmm0 \n" + "pshufb %%xmm6,%%xmm1 \n" + "palignr $0xc,%%xmm0,%%xmm1 \n" + "movq %%xmm0," MEMACCESS([dst_raw]) " \n" + "movdqu %%xmm1," MEMACCESS2(0x8,[dst_raw]) "\n" + "lea " MEMLEA(0x18,[dst_raw]) ",%[dst_raw] \n" + "subl $0x8,%[width] \n" + "jg 1b \n" + : [y_buf]"+r"(y_buf), // %[y_buf] + [u_buf]"+r"(u_buf), // %[u_buf] + [v_buf]"+r"(v_buf), // %[v_buf] + [dst_raw]"+r"(dst_raw), // %[dst_raw] +// TODO(fbarchard): Make width a register for 32 bit. +#if defined(__i386__) && defined(__pic__) + [width]"+m"(width) // %[width] +#else + [width]"+rm"(width) // %[width] +#endif + : [kYuvConstants]"r"(&kYuvConstants.kUVToB), + [kShuffleMaskARGBToRAW_0]"m"(kShuffleMaskARGBToRAW_0), + [kShuffleMaskARGBToRAW]"m"(kShuffleMaskARGBToRAW) + : "memory", "cc", NACL_R14 + "xmm0", "xmm1", "xmm2", "xmm3", "xmm5", "xmm6" + ); +} + +void OMITFP I422ToARGBRow_SSSE3(const uint8* y_buf, + const uint8* u_buf, + const uint8* v_buf, + uint8* dst_argb, + int width) { + asm volatile ( + "sub %[u_buf],%[v_buf] \n" + "pcmpeqb %%xmm5,%%xmm5 \n" + LABELALIGN + "1: \n" + READYUV422 + YUVTORGB(kYuvConstants) + STOREARGB + "sub $0x8,%[width] \n" + "jg 1b \n" + : [y_buf]"+r"(y_buf), // %[y_buf] + [u_buf]"+r"(u_buf), // %[u_buf] + [v_buf]"+r"(v_buf), // %[v_buf] + [dst_argb]"+r"(dst_argb), // %[dst_argb] + [width]"+rm"(width) // %[width] + : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants] + : "memory", "cc", NACL_R14 + "xmm0", "xmm1", "xmm2", "xmm3", "xmm5" + ); +} + +void OMITFP J422ToARGBRow_SSSE3(const uint8* y_buf, + const uint8* u_buf, + const uint8* v_buf, + uint8* dst_argb, + int width) { + asm volatile ( + "sub %[u_buf],%[v_buf] \n" + "pcmpeqb %%xmm5,%%xmm5 \n" + LABELALIGN + "1: \n" + READYUV422 + YUVTORGB(kYuvConstants) + STOREARGB + "sub $0x8,%[width] \n" + "jg 1b \n" + : [y_buf]"+r"(y_buf), // %[y_buf] + [u_buf]"+r"(u_buf), // %[u_buf] + [v_buf]"+r"(v_buf), // %[v_buf] + [dst_argb]"+r"(dst_argb), // %[dst_argb] + [width]"+rm"(width) // %[width] + : [kYuvConstants]"r"(&kYuvJConstants.kUVToB) // %[kYuvConstants] + : "memory", "cc", NACL_R14 + "xmm0", "xmm1", "xmm2", "xmm3", "xmm5" + ); +} + +void OMITFP I411ToARGBRow_SSSE3(const uint8* y_buf, + const uint8* u_buf, + const uint8* v_buf, + uint8* dst_argb, + int width) { + asm volatile ( + "sub %[u_buf],%[v_buf] \n" + "pcmpeqb %%xmm5,%%xmm5 \n" + LABELALIGN + "1: \n" + READYUV411 + YUVTORGB(kYuvConstants) + STOREARGB + "sub $0x8,%[width] \n" + "jg 1b \n" + : [y_buf]"+r"(y_buf), // %[y_buf] + [u_buf]"+r"(u_buf), // %[u_buf] + [v_buf]"+r"(v_buf), // %[v_buf] + [dst_argb]"+r"(dst_argb), // %[dst_argb] + [width]"+rm"(width) // %[width] + : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants] + : "memory", "cc", NACL_R14 + "xmm0", "xmm1", "xmm2", "xmm3", "xmm5" + ); +} + +void OMITFP NV12ToARGBRow_SSSE3(const uint8* y_buf, + const uint8* uv_buf, + uint8* dst_argb, + int width) { + asm volatile ( + "pcmpeqb %%xmm5,%%xmm5 \n" + LABELALIGN + "1: \n" + READNV12 + YUVTORGB(kYuvConstants) + STOREARGB + "sub $0x8,%[width] \n" + "jg 1b \n" + : [y_buf]"+r"(y_buf), // %[y_buf] + [uv_buf]"+r"(uv_buf), // %[uv_buf] + [dst_argb]"+r"(dst_argb), // %[dst_argb] + [width]"+rm"(width) // %[width] + : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants] + // Does not use r14. + : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm5" + ); +} + +void OMITFP NV21ToARGBRow_SSSE3(const uint8* y_buf, + const uint8* uv_buf, + uint8* dst_argb, + int width) { + asm volatile ( + "pcmpeqb %%xmm5,%%xmm5 \n" + LABELALIGN + "1: \n" + READNV12 + YUVTORGB(kYuvConstants) + STOREARGB + "sub $0x8,%[width] \n" + "jg 1b \n" + : [y_buf]"+r"(y_buf), // %[y_buf] + [uv_buf]"+r"(uv_buf), // %[uv_buf] + [dst_argb]"+r"(dst_argb), // %[dst_argb] + [width]"+rm"(width) // %[width] + : [kYuvConstants]"r"(&kYvuConstants.kUVToB) // %[kYuvConstants] + // Does not use r14. + : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm5" + ); +} + +void OMITFP I422ToBGRARow_SSSE3(const uint8* y_buf, + const uint8* u_buf, + const uint8* v_buf, + uint8* dst_bgra, + int width) { + asm volatile ( + "sub %[u_buf],%[v_buf] \n" + "pcmpeqb %%xmm5,%%xmm5 \n" + LABELALIGN + "1: \n" + READYUV422 + YUVTORGB(kYuvConstants) + STOREBGRA + "sub $0x8,%[width] \n" + "jg 1b \n" + : [y_buf]"+r"(y_buf), // %[y_buf] + [u_buf]"+r"(u_buf), // %[u_buf] + [v_buf]"+r"(v_buf), // %[v_buf] + [dst_bgra]"+r"(dst_bgra), // %[dst_bgra] + [width]"+rm"(width) // %[width] + : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants] + : "memory", "cc", NACL_R14 + "xmm0", "xmm1", "xmm2", "xmm3", "xmm5" + ); +} + +void OMITFP I422ToABGRRow_SSSE3(const uint8* y_buf, + const uint8* u_buf, + const uint8* v_buf, + uint8* dst_abgr, + int width) { + asm volatile ( + "sub %[u_buf],%[v_buf] \n" + "pcmpeqb %%xmm5,%%xmm5 \n" + LABELALIGN + "1: \n" + READYUV422 + YUVTORGB(kYuvConstants) + STOREABGR + "sub $0x8,%[width] \n" + "jg 1b \n" + : [y_buf]"+r"(y_buf), // %[y_buf] + [u_buf]"+r"(u_buf), // %[u_buf] + [v_buf]"+r"(v_buf), // %[v_buf] + [dst_abgr]"+r"(dst_abgr), // %[dst_abgr] + [width]"+rm"(width) // %[width] + : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants] + : "memory", "cc", NACL_R14 + "xmm0", "xmm1", "xmm2", "xmm3", "xmm5" + ); +} + +void OMITFP I422ToRGBARow_SSSE3(const uint8* y_buf, + const uint8* u_buf, + const uint8* v_buf, + uint8* dst_rgba, + int width) { + asm volatile ( + "sub %[u_buf],%[v_buf] \n" + "pcmpeqb %%xmm5,%%xmm5 \n" + LABELALIGN + "1: \n" + READYUV422 + YUVTORGB(kYuvConstants) + STORERGBA + "sub $0x8,%[width] \n" + "jg 1b \n" + : [y_buf]"+r"(y_buf), // %[y_buf] + [u_buf]"+r"(u_buf), // %[u_buf] + [v_buf]"+r"(v_buf), // %[v_buf] + [dst_rgba]"+r"(dst_rgba), // %[dst_rgba] + [width]"+rm"(width) // %[width] + : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants] + : "memory", "cc", NACL_R14 + "xmm0", "xmm1", "xmm2", "xmm3", "xmm5" + ); +} + +#endif // HAS_I422TOARGBROW_SSSE3 + +// Read 8 UV from 422, upsample to 16 UV. +#define READYUV422_AVX2 \ + "vmovq " MEMACCESS([u_buf]) ",%%xmm0 \n" \ + MEMOPREG(vmovq, 0x00, [u_buf], [v_buf], 1, xmm1) \ + "lea " MEMLEA(0x8, [u_buf]) ",%[u_buf] \n" \ + "vpunpcklbw %%ymm1,%%ymm0,%%ymm0 \n" \ + "vpermq $0xd8,%%ymm0,%%ymm0 \n" \ + "vpunpcklwd %%ymm0,%%ymm0,%%ymm0 \n" + +// Convert 16 pixels: 16 UV and 16 Y. +#define YUVTORGB_AVX2(YuvConstants) \ + "vpmaddubsw " MEMACCESS2(64, [YuvConstants]) ",%%ymm0,%%ymm2 \n" \ + "vpmaddubsw " MEMACCESS2(32, [YuvConstants]) ",%%ymm0,%%ymm1 \n" \ + "vpmaddubsw " MEMACCESS([YuvConstants]) ",%%ymm0,%%ymm0 \n" \ + "vmovdqu " MEMACCESS2(160, [YuvConstants]) ",%%ymm3 \n" \ + "vpsubw %%ymm2,%%ymm3,%%ymm2 \n" \ + "vmovdqu " MEMACCESS2(128, [YuvConstants]) ",%%ymm3 \n" \ + "vpsubw %%ymm1,%%ymm3,%%ymm1 \n" \ + "vmovdqu " MEMACCESS2(96, [YuvConstants]) ",%%ymm3 \n" \ + "vpsubw %%ymm0,%%ymm3,%%ymm0 \n" \ + "vmovdqu " MEMACCESS([y_buf]) ",%%xmm3 \n" \ + "lea " MEMLEA(0x10, [y_buf]) ",%[y_buf] \n" \ + "vpermq $0xd8,%%ymm3,%%ymm3 \n" \ + "vpunpcklbw %%ymm3,%%ymm3,%%ymm3 \n" \ + "vpmulhuw " MEMACCESS2(192, [YuvConstants]) ",%%ymm3,%%ymm3 \n" \ + "vpaddsw %%ymm3,%%ymm0,%%ymm0 \n" \ + "vpaddsw %%ymm3,%%ymm1,%%ymm1 \n" \ + "vpaddsw %%ymm3,%%ymm2,%%ymm2 \n" \ + "vpsraw $0x6,%%ymm0,%%ymm0 \n" \ + "vpsraw $0x6,%%ymm1,%%ymm1 \n" \ + "vpsraw $0x6,%%ymm2,%%ymm2 \n" \ + "vpackuswb %%ymm0,%%ymm0,%%ymm0 \n" \ + "vpackuswb %%ymm1,%%ymm1,%%ymm1 \n" \ + "vpackuswb %%ymm2,%%ymm2,%%ymm2 \n" + +#if defined(HAS_I422TOBGRAROW_AVX2) +// 16 pixels +// 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 BGRA (64 bytes). +void OMITFP I422ToBGRARow_AVX2(const uint8* y_buf, + const uint8* u_buf, + const uint8* v_buf, + uint8* dst_bgra, + int width) { + asm volatile ( + "sub %[u_buf],%[v_buf] \n" + "vpcmpeqb %%ymm5,%%ymm5,%%ymm5 \n" + LABELALIGN + "1: \n" + READYUV422_AVX2 + YUVTORGB_AVX2(kYuvConstants) + + // Step 3: Weave into BGRA + "vpunpcklbw %%ymm0,%%ymm1,%%ymm1 \n" // GB + "vpermq $0xd8,%%ymm1,%%ymm1 \n" + "vpunpcklbw %%ymm2,%%ymm5,%%ymm2 \n" // AR + "vpermq $0xd8,%%ymm2,%%ymm2 \n" + "vpunpcklwd %%ymm1,%%ymm2,%%ymm0 \n" // ARGB first 8 pixels + "vpunpckhwd %%ymm1,%%ymm2,%%ymm2 \n" // ARGB next 8 pixels + + "vmovdqu %%ymm0," MEMACCESS([dst_bgra]) "\n" + "vmovdqu %%ymm2," MEMACCESS2(0x20,[dst_bgra]) "\n" + "lea " MEMLEA(0x40,[dst_bgra]) ",%[dst_bgra] \n" + "sub $0x10,%[width] \n" + "jg 1b \n" + "vzeroupper \n" + : [y_buf]"+r"(y_buf), // %[y_buf] + [u_buf]"+r"(u_buf), // %[u_buf] + [v_buf]"+r"(v_buf), // %[v_buf] + [dst_bgra]"+r"(dst_bgra), // %[dst_bgra] + [width]"+rm"(width) // %[width] + : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants] + : "memory", "cc", NACL_R14 + "xmm0", "xmm1", "xmm2", "xmm3", "xmm5" + ); +} +#endif // HAS_I422TOBGRAROW_AVX2 + +#if defined(HAS_I422TOARGBROW_AVX2) +// 16 pixels +// 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 ARGB (64 bytes). +void OMITFP I422ToARGBRow_AVX2(const uint8* y_buf, + const uint8* u_buf, + const uint8* v_buf, + uint8* dst_argb, + int width) { + asm volatile ( + "sub %[u_buf],%[v_buf] \n" + "vpcmpeqb %%ymm5,%%ymm5,%%ymm5 \n" + LABELALIGN + "1: \n" + READYUV422_AVX2 + YUVTORGB_AVX2(kYuvConstants) + + // Step 3: Weave into ARGB + "vpunpcklbw %%ymm1,%%ymm0,%%ymm0 \n" // BG + "vpermq $0xd8,%%ymm0,%%ymm0 \n" + "vpunpcklbw %%ymm5,%%ymm2,%%ymm2 \n" // RA + "vpermq $0xd8,%%ymm2,%%ymm2 \n" + "vpunpcklwd %%ymm2,%%ymm0,%%ymm1 \n" // BGRA first 8 pixels + "vpunpckhwd %%ymm2,%%ymm0,%%ymm0 \n" // BGRA next 8 pixels + + "vmovdqu %%ymm1," MEMACCESS([dst_argb]) "\n" + "vmovdqu %%ymm0," MEMACCESS2(0x20,[dst_argb]) "\n" + "lea " MEMLEA(0x40,[dst_argb]) ",%[dst_argb] \n" + "sub $0x10,%[width] \n" + "jg 1b \n" + "vzeroupper \n" + : [y_buf]"+r"(y_buf), // %[y_buf] + [u_buf]"+r"(u_buf), // %[u_buf] + [v_buf]"+r"(v_buf), // %[v_buf] + [dst_argb]"+r"(dst_argb), // %[dst_argb] + [width]"+rm"(width) // %[width] + : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants] + : "memory", "cc", NACL_R14 + "xmm0", "xmm1", "xmm2", "xmm3", "xmm5" + ); +} +#endif // HAS_I422TOARGBROW_AVX2 + +#if defined(HAS_J422TOARGBROW_AVX2) +// 16 pixels +// 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 ARGB (64 bytes). +void OMITFP J422ToARGBRow_AVX2(const uint8* y_buf, + const uint8* u_buf, + const uint8* v_buf, + uint8* dst_argb, + int width) { + asm volatile ( + "sub %[u_buf],%[v_buf] \n" + "vpcmpeqb %%ymm5,%%ymm5,%%ymm5 \n" + LABELALIGN + "1: \n" + READYUV422_AVX2 + YUVTORGB_AVX2(kYuvConstants) + + // Step 3: Weave into ARGB + "vpunpcklbw %%ymm1,%%ymm0,%%ymm0 \n" // BG + "vpermq $0xd8,%%ymm0,%%ymm0 \n" + "vpunpcklbw %%ymm5,%%ymm2,%%ymm2 \n" // RA + "vpermq $0xd8,%%ymm2,%%ymm2 \n" + "vpunpcklwd %%ymm2,%%ymm0,%%ymm1 \n" // BGRA first 8 pixels + "vpunpckhwd %%ymm2,%%ymm0,%%ymm0 \n" // BGRA next 8 pixels + + "vmovdqu %%ymm1," MEMACCESS([dst_argb]) "\n" + "vmovdqu %%ymm0," MEMACCESS2(0x20,[dst_argb]) "\n" + "lea " MEMLEA(0x40,[dst_argb]) ",%[dst_argb] \n" + "sub $0x10,%[width] \n" + "jg 1b \n" + "vzeroupper \n" + : [y_buf]"+r"(y_buf), // %[y_buf] + [u_buf]"+r"(u_buf), // %[u_buf] + [v_buf]"+r"(v_buf), // %[v_buf] + [dst_argb]"+r"(dst_argb), // %[dst_argb] + [width]"+rm"(width) // %[width] + : [kYuvConstants]"r"(&kYuvJConstants.kUVToB) // %[kYuvConstants] + : "memory", "cc", NACL_R14 + "xmm0", "xmm1", "xmm2", "xmm3", "xmm5" + ); +} +#endif // HAS_J422TOARGBROW_AVX2 + +#if defined(HAS_I422TOABGRROW_AVX2) +// 16 pixels +// 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 ABGR (64 bytes). +void OMITFP I422ToABGRRow_AVX2(const uint8* y_buf, + const uint8* u_buf, + const uint8* v_buf, + uint8* dst_argb, + int width) { + asm volatile ( + "sub %[u_buf],%[v_buf] \n" + "vpcmpeqb %%ymm5,%%ymm5,%%ymm5 \n" + LABELALIGN + "1: \n" + READYUV422_AVX2 + YUVTORGB_AVX2(kYuvConstants) + + // Step 3: Weave into ABGR + "vpunpcklbw %%ymm1,%%ymm2,%%ymm1 \n" // RG + "vpermq $0xd8,%%ymm1,%%ymm1 \n" + "vpunpcklbw %%ymm5,%%ymm0,%%ymm2 \n" // BA + "vpermq $0xd8,%%ymm2,%%ymm2 \n" + "vpunpcklwd %%ymm2,%%ymm1,%%ymm0 \n" // RGBA first 8 pixels + "vpunpckhwd %%ymm2,%%ymm1,%%ymm1 \n" // RGBA next 8 pixels + "vmovdqu %%ymm0," MEMACCESS([dst_argb]) "\n" + "vmovdqu %%ymm1," MEMACCESS2(0x20,[dst_argb]) "\n" + "lea " MEMLEA(0x40,[dst_argb]) ",%[dst_argb] \n" + "sub $0x10,%[width] \n" + "jg 1b \n" + "vzeroupper \n" + : [y_buf]"+r"(y_buf), // %[y_buf] + [u_buf]"+r"(u_buf), // %[u_buf] + [v_buf]"+r"(v_buf), // %[v_buf] + [dst_argb]"+r"(dst_argb), // %[dst_argb] + [width]"+rm"(width) // %[width] + : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants] + : "memory", "cc", NACL_R14 + "xmm0", "xmm1", "xmm2", "xmm3", "xmm5" + ); +} +#endif // HAS_I422TOABGRROW_AVX2 + +#if defined(HAS_I422TORGBAROW_AVX2) +// 16 pixels +// 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 RGBA (64 bytes). +void OMITFP I422ToRGBARow_AVX2(const uint8* y_buf, + const uint8* u_buf, + const uint8* v_buf, + uint8* dst_argb, + int width) { + asm volatile ( + "sub %[u_buf],%[v_buf] \n" + "vpcmpeqb %%ymm5,%%ymm5,%%ymm5 \n" + LABELALIGN + "1: \n" + READYUV422_AVX2 + YUVTORGB_AVX2(kYuvConstants) + + // Step 3: Weave into RGBA + "vpunpcklbw %%ymm2,%%ymm1,%%ymm1 \n" + "vpermq $0xd8,%%ymm1,%%ymm1 \n" + "vpunpcklbw %%ymm0,%%ymm5,%%ymm2 \n" + "vpermq $0xd8,%%ymm2,%%ymm2 \n" + "vpunpcklwd %%ymm1,%%ymm2,%%ymm0 \n" + "vpunpckhwd %%ymm1,%%ymm2,%%ymm1 \n" + "vmovdqu %%ymm0," MEMACCESS([dst_argb]) "\n" + "vmovdqu %%ymm1," MEMACCESS2(0x20,[dst_argb]) "\n" + "lea " MEMLEA(0x40,[dst_argb]) ",%[dst_argb] \n" + "sub $0x10,%[width] \n" + "jg 1b \n" + "vzeroupper \n" + : [y_buf]"+r"(y_buf), // %[y_buf] + [u_buf]"+r"(u_buf), // %[u_buf] + [v_buf]"+r"(v_buf), // %[v_buf] + [dst_argb]"+r"(dst_argb), // %[dst_argb] + [width]"+rm"(width) // %[width] + : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants] + : "memory", "cc", NACL_R14 + "xmm0", "xmm1", "xmm2", "xmm3", "xmm5" + ); +} +#endif // HAS_I422TORGBAROW_AVX2 + +#ifdef HAS_I400TOARGBROW_SSE2 +void I400ToARGBRow_SSE2(const uint8* y_buf, uint8* dst_argb, int width) { + asm volatile ( + "mov $0x4a354a35,%%eax \n" // 4a35 = 18997 = 1.164 + "movd %%eax,%%xmm2 \n" + "pshufd $0x0,%%xmm2,%%xmm2 \n" + "mov $0x04880488,%%eax \n" // 0488 = 1160 = 1.164 * 16 + "movd %%eax,%%xmm3 \n" + "pshufd $0x0,%%xmm3,%%xmm3 \n" + "pcmpeqb %%xmm4,%%xmm4 \n" + "pslld $0x18,%%xmm4 \n" + LABELALIGN + "1: \n" + // Step 1: Scale Y contribution to 8 G values. G = (y - 16) * 1.164 + "movq " MEMACCESS(0) ",%%xmm0 \n" + "lea " MEMLEA(0x8,0) ",%0 \n" + "punpcklbw %%xmm0,%%xmm0 \n" + "pmulhuw %%xmm2,%%xmm0 \n" + "psubusw %%xmm3,%%xmm0 \n" + "psrlw $6, %%xmm0 \n" + "packuswb %%xmm0,%%xmm0 \n" + + // Step 2: Weave into ARGB + "punpcklbw %%xmm0,%%xmm0 \n" + "movdqa %%xmm0,%%xmm1 \n" + "punpcklwd %%xmm0,%%xmm0 \n" + "punpckhwd %%xmm1,%%xmm1 \n" + "por %%xmm4,%%xmm0 \n" + "por %%xmm4,%%xmm1 \n" + "movdqu %%xmm0," MEMACCESS(1) " \n" + "movdqu %%xmm1," MEMACCESS2(0x10,1) " \n" + "lea " MEMLEA(0x20,1) ",%1 \n" + + "sub $0x8,%2 \n" + "jg 1b \n" + : "+r"(y_buf), // %0 + "+r"(dst_argb), // %1 + "+rm"(width) // %2 + : + : "memory", "cc", "eax" + , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4" + ); +} +#endif // HAS_I400TOARGBROW_SSE2 + +#ifdef HAS_I400TOARGBROW_AVX2 +// 16 pixels of Y converted to 16 pixels of ARGB (64 bytes). +// note: vpunpcklbw mutates and vpackuswb unmutates. +void I400ToARGBRow_AVX2(const uint8* y_buf, uint8* dst_argb, int width) { + asm volatile ( + "mov $0x4a354a35,%%eax \n" // 0488 = 1160 = 1.164 * 16 + "vmovd %%eax,%%xmm2 \n" + "vbroadcastss %%xmm2,%%ymm2 \n" + "mov $0x4880488,%%eax \n" // 4a35 = 18997 = 1.164 + "vmovd %%eax,%%xmm3 \n" + "vbroadcastss %%xmm3,%%ymm3 \n" + "vpcmpeqb %%ymm4,%%ymm4,%%ymm4 \n" + "vpslld $0x18,%%ymm4,%%ymm4 \n" + + LABELALIGN + "1: \n" + // Step 1: Scale Y contribution to 16 G values. G = (y - 16) * 1.164 + "vmovdqu " MEMACCESS(0) ",%%xmm0 \n" + "lea " MEMLEA(0x10,0) ",%0 \n" + "vpermq $0xd8,%%ymm0,%%ymm0 \n" + "vpunpcklbw %%ymm0,%%ymm0,%%ymm0 \n" + "vpmulhuw %%ymm2,%%ymm0,%%ymm0 \n" + "vpsubusw %%ymm3,%%ymm0,%%ymm0 \n" + "vpsrlw $0x6,%%ymm0,%%ymm0 \n" + "vpackuswb %%ymm0,%%ymm0,%%ymm0 \n" + "vpunpcklbw %%ymm0,%%ymm0,%%ymm1 \n" + "vpermq $0xd8,%%ymm1,%%ymm1 \n" + "vpunpcklwd %%ymm1,%%ymm1,%%ymm0 \n" + "vpunpckhwd %%ymm1,%%ymm1,%%ymm1 \n" + "vpor %%ymm4,%%ymm0,%%ymm0 \n" + "vpor %%ymm4,%%ymm1,%%ymm1 \n" + "vmovdqu %%ymm0," MEMACCESS(1) " \n" + "vmovdqu %%ymm1," MEMACCESS2(0x20,1) " \n" + "lea " MEMLEA(0x40,1) ",%1 \n" + "sub $0x10,%2 \n" + "jg 1b \n" + "vzeroupper \n" + : "+r"(y_buf), // %0 + "+r"(dst_argb), // %1 + "+rm"(width) // %2 + : + : "memory", "cc", "eax" + , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4" + ); +} +#endif // HAS_I400TOARGBROW_AVX2 + +#ifdef HAS_MIRRORROW_SSSE3 +// Shuffle table for reversing the bytes. +static uvec8 kShuffleMirror = { + 15u, 14u, 13u, 12u, 11u, 10u, 9u, 8u, 7u, 6u, 5u, 4u, 3u, 2u, 1u, 0u +}; + +void MirrorRow_SSSE3(const uint8* src, uint8* dst, int width) { + intptr_t temp_width = (intptr_t)(width); + asm volatile ( + "movdqa %3,%%xmm5 \n" + LABELALIGN + "1: \n" + MEMOPREG(movdqu,-0x10,0,2,1,xmm0) // movdqu -0x10(%0,%2),%%xmm0 + "pshufb %%xmm5,%%xmm0 \n" + "movdqu %%xmm0," MEMACCESS(1) " \n" + "lea " MEMLEA(0x10,1) ",%1 \n" + "sub $0x10,%2 \n" + "jg 1b \n" + : "+r"(src), // %0 + "+r"(dst), // %1 + "+r"(temp_width) // %2 + : "m"(kShuffleMirror) // %3 + : "memory", "cc", NACL_R14 + "xmm0", "xmm5" + ); +} +#endif // HAS_MIRRORROW_SSSE3 + +#ifdef HAS_MIRRORROW_AVX2 +void MirrorRow_AVX2(const uint8* src, uint8* dst, int width) { + intptr_t temp_width = (intptr_t)(width); + asm volatile ( + "vbroadcastf128 %3,%%ymm5 \n" + LABELALIGN + "1: \n" + MEMOPREG(vmovdqu,-0x20,0,2,1,ymm0) // vmovdqu -0x20(%0,%2),%%ymm0 + "vpshufb %%ymm5,%%ymm0,%%ymm0 \n" + "vpermq $0x4e,%%ymm0,%%ymm0 \n" + "vmovdqu %%ymm0," MEMACCESS(1) " \n" + "lea " MEMLEA(0x20,1) ",%1 \n" + "sub $0x20,%2 \n" + "jg 1b \n" + "vzeroupper \n" + : "+r"(src), // %0 + "+r"(dst), // %1 + "+r"(temp_width) // %2 + : "m"(kShuffleMirror) // %3 + : "memory", "cc", NACL_R14 + "xmm0", "xmm5" + ); +} +#endif // HAS_MIRRORROW_AVX2 + +#ifdef HAS_MIRRORROW_SSE2 +void MirrorRow_SSE2(const uint8* src, uint8* dst, int width) { + intptr_t temp_width = (intptr_t)(width); + asm volatile ( + LABELALIGN + "1: \n" + MEMOPREG(movdqu,-0x10,0,2,1,xmm0) // movdqu -0x10(%0,%2),%%xmm0 + "movdqa %%xmm0,%%xmm1 \n" + "psllw $0x8,%%xmm0 \n" + "psrlw $0x8,%%xmm1 \n" + "por %%xmm1,%%xmm0 \n" + "pshuflw $0x1b,%%xmm0,%%xmm0 \n" + "pshufhw $0x1b,%%xmm0,%%xmm0 \n" + "pshufd $0x4e,%%xmm0,%%xmm0 \n" + "movdqu %%xmm0," MEMACCESS(1) " \n" + "lea " MEMLEA(0x10,1)",%1 \n" + "sub $0x10,%2 \n" + "jg 1b \n" + : "+r"(src), // %0 + "+r"(dst), // %1 + "+r"(temp_width) // %2 + : + : "memory", "cc", NACL_R14 + "xmm0", "xmm1" + ); +} +#endif // HAS_MIRRORROW_SSE2 + +#ifdef HAS_MIRRORROW_UV_SSSE3 +// Shuffle table for reversing the bytes of UV channels. +static uvec8 kShuffleMirrorUV = { + 14u, 12u, 10u, 8u, 6u, 4u, 2u, 0u, 15u, 13u, 11u, 9u, 7u, 5u, 3u, 1u +}; +void MirrorUVRow_SSSE3(const uint8* src, uint8* dst_u, uint8* dst_v, + int width) { + intptr_t temp_width = (intptr_t)(width); + asm volatile ( + "movdqa %4,%%xmm1 \n" + "lea " MEMLEA4(-0x10,0,3,2) ",%0 \n" + "sub %1,%2 \n" + LABELALIGN + "1: \n" + "movdqu " MEMACCESS(0) ",%%xmm0 \n" + "lea " MEMLEA(-0x10,0) ",%0 \n" + "pshufb %%xmm1,%%xmm0 \n" + "movlpd %%xmm0," MEMACCESS(1) " \n" + MEMOPMEM(movhpd,xmm0,0x00,1,2,1) // movhpd %%xmm0,(%1,%2) + "lea " MEMLEA(0x8,1) ",%1 \n" + "sub $8,%3 \n" + "jg 1b \n" + : "+r"(src), // %0 + "+r"(dst_u), // %1 + "+r"(dst_v), // %2 + "+r"(temp_width) // %3 + : "m"(kShuffleMirrorUV) // %4 + : "memory", "cc", NACL_R14 + "xmm0", "xmm1" + ); +} +#endif // HAS_MIRRORROW_UV_SSSE3 + +#ifdef HAS_ARGBMIRRORROW_SSE2 + +void ARGBMirrorRow_SSE2(const uint8* src, uint8* dst, int width) { + intptr_t temp_width = (intptr_t)(width); + asm volatile ( + "lea " MEMLEA4(-0x10,0,2,4) ",%0 \n" + LABELALIGN + "1: \n" + "movdqu " MEMACCESS(0) ",%%xmm0 \n" + "pshufd $0x1b,%%xmm0,%%xmm0 \n" + "lea " MEMLEA(-0x10,0) ",%0 \n" + "movdqu %%xmm0," MEMACCESS(1) " \n" + "lea " MEMLEA(0x10,1) ",%1 \n" + "sub $0x4,%2 \n" + "jg 1b \n" + : "+r"(src), // %0 + "+r"(dst), // %1 + "+r"(temp_width) // %2 + : + : "memory", "cc" + , "xmm0" + ); +} +#endif // HAS_ARGBMIRRORROW_SSE2 + +#ifdef HAS_ARGBMIRRORROW_AVX2 +// Shuffle table for reversing the bytes. +static const ulvec32 kARGBShuffleMirror_AVX2 = { + 7u, 6u, 5u, 4u, 3u, 2u, 1u, 0u +}; +void ARGBMirrorRow_AVX2(const uint8* src, uint8* dst, int width) { + intptr_t temp_width = (intptr_t)(width); + asm volatile ( + "vmovdqu %3,%%ymm5 \n" + LABELALIGN + "1: \n" + VMEMOPREG(vpermd,-0x20,0,2,4,ymm5,ymm0) // vpermd -0x20(%0,%2,4),ymm5,ymm0 + "vmovdqu %%ymm0," MEMACCESS(1) " \n" + "lea " MEMLEA(0x20,1) ",%1 \n" + "sub $0x8,%2 \n" + "jg 1b \n" + "vzeroupper \n" + : "+r"(src), // %0 + "+r"(dst), // %1 + "+r"(temp_width) // %2 + : "m"(kARGBShuffleMirror_AVX2) // %3 + : "memory", "cc", NACL_R14 + "xmm0", "xmm5" + ); +} +#endif // HAS_ARGBMIRRORROW_AVX2 + +#ifdef HAS_SPLITUVROW_AVX2 +void SplitUVRow_AVX2(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int pix) { + asm volatile ( + "vpcmpeqb %%ymm5,%%ymm5,%%ymm5 \n" + "vpsrlw $0x8,%%ymm5,%%ymm5 \n" + "sub %1,%2 \n" + LABELALIGN + "1: \n" + "vmovdqu " MEMACCESS(0) ",%%ymm0 \n" + "vmovdqu " MEMACCESS2(0x20,0) ",%%ymm1 \n" + "lea " MEMLEA(0x40,0) ",%0 \n" + "vpsrlw $0x8,%%ymm0,%%ymm2 \n" + "vpsrlw $0x8,%%ymm1,%%ymm3 \n" + "vpand %%ymm5,%%ymm0,%%ymm0 \n" + "vpand %%ymm5,%%ymm1,%%ymm1 \n" + "vpackuswb %%ymm1,%%ymm0,%%ymm0 \n" + "vpackuswb %%ymm3,%%ymm2,%%ymm2 \n" + "vpermq $0xd8,%%ymm0,%%ymm0 \n" + "vpermq $0xd8,%%ymm2,%%ymm2 \n" + "vmovdqu %%ymm0," MEMACCESS(1) " \n" + MEMOPMEM(vmovdqu,ymm2,0x00,1,2,1) // vmovdqu %%ymm2,(%1,%2) + "lea " MEMLEA(0x20,1) ",%1 \n" + "sub $0x20,%3 \n" + "jg 1b \n" + "vzeroupper \n" + : "+r"(src_uv), // %0 + "+r"(dst_u), // %1 + "+r"(dst_v), // %2 + "+r"(pix) // %3 + : + : "memory", "cc", NACL_R14 + "xmm0", "xmm1", "xmm2", "xmm3", "xmm5" + ); +} +#endif // HAS_SPLITUVROW_AVX2 + +#ifdef HAS_SPLITUVROW_SSE2 +void SplitUVRow_SSE2(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int pix) { + asm volatile ( + "pcmpeqb %%xmm5,%%xmm5 \n" + "psrlw $0x8,%%xmm5 \n" + "sub %1,%2 \n" + LABELALIGN + "1: \n" + "movdqu " MEMACCESS(0) ",%%xmm0 \n" + "movdqu " MEMACCESS2(0x10,0) ",%%xmm1 \n" + "lea " MEMLEA(0x20,0) ",%0 \n" + "movdqa %%xmm0,%%xmm2 \n" + "movdqa %%xmm1,%%xmm3 \n" + "pand %%xmm5,%%xmm0 \n" + "pand %%xmm5,%%xmm1 \n" + "packuswb %%xmm1,%%xmm0 \n" + "psrlw $0x8,%%xmm2 \n" + "psrlw $0x8,%%xmm3 \n" + "packuswb %%xmm3,%%xmm2 \n" + "movdqu %%xmm0," MEMACCESS(1) " \n" + MEMOPMEM(movdqu,xmm2,0x00,1,2,1) // movdqu %%xmm2,(%1,%2) + "lea " MEMLEA(0x10,1) ",%1 \n" + "sub $0x10,%3 \n" + "jg 1b \n" + : "+r"(src_uv), // %0 + "+r"(dst_u), // %1 + "+r"(dst_v), // %2 + "+r"(pix) // %3 + : + : "memory", "cc", NACL_R14 + "xmm0", "xmm1", "xmm2", "xmm3", "xmm5" + ); +} +#endif // HAS_SPLITUVROW_SSE2 + +#ifdef HAS_MERGEUVROW_AVX2 +void MergeUVRow_AVX2(const uint8* src_u, const uint8* src_v, uint8* dst_uv, + int width) { + asm volatile ( + "sub %0,%1 \n" + LABELALIGN + "1: \n" + "vmovdqu " MEMACCESS(0) ",%%ymm0 \n" + MEMOPREG(vmovdqu,0x00,0,1,1,ymm1) // vmovdqu (%0,%1,1),%%ymm1 + "lea " MEMLEA(0x20,0) ",%0 \n" + "vpunpcklbw %%ymm1,%%ymm0,%%ymm2 \n" + "vpunpckhbw %%ymm1,%%ymm0,%%ymm0 \n" + "vextractf128 $0x0,%%ymm2," MEMACCESS(2) " \n" + "vextractf128 $0x0,%%ymm0," MEMACCESS2(0x10,2) "\n" + "vextractf128 $0x1,%%ymm2," MEMACCESS2(0x20,2) "\n" + "vextractf128 $0x1,%%ymm0," MEMACCESS2(0x30,2) "\n" + "lea " MEMLEA(0x40,2) ",%2 \n" + "sub $0x20,%3 \n" + "jg 1b \n" + "vzeroupper \n" + : "+r"(src_u), // %0 + "+r"(src_v), // %1 + "+r"(dst_uv), // %2 + "+r"(width) // %3 + : + : "memory", "cc", NACL_R14 + "xmm0", "xmm1", "xmm2" + ); +} +#endif // HAS_MERGEUVROW_AVX2 + +#ifdef HAS_MERGEUVROW_SSE2 +void MergeUVRow_SSE2(const uint8* src_u, const uint8* src_v, uint8* dst_uv, + int width) { + asm volatile ( + "sub %0,%1 \n" + LABELALIGN + "1: \n" + "movdqu " MEMACCESS(0) ",%%xmm0 \n" + MEMOPREG(movdqu,0x00,0,1,1,xmm1) // movdqu (%0,%1,1),%%xmm1 + "lea " MEMLEA(0x10,0) ",%0 \n" + "movdqa %%xmm0,%%xmm2 \n" + "punpcklbw %%xmm1,%%xmm0 \n" + "punpckhbw %%xmm1,%%xmm2 \n" + "movdqu %%xmm0," MEMACCESS(2) " \n" + "movdqu %%xmm2," MEMACCESS2(0x10,2) " \n" + "lea " MEMLEA(0x20,2) ",%2 \n" + "sub $0x10,%3 \n" + "jg 1b \n" + : "+r"(src_u), // %0 + "+r"(src_v), // %1 + "+r"(dst_uv), // %2 + "+r"(width) // %3 + : + : "memory", "cc", NACL_R14 + "xmm0", "xmm1", "xmm2" + ); +} +#endif // HAS_MERGEUVROW_SSE2 + +#ifdef HAS_COPYROW_SSE2 +void CopyRow_SSE2(const uint8* src, uint8* dst, int count) { + asm volatile ( + LABELALIGN + "1: \n" + "movdqu " MEMACCESS(0) ",%%xmm0 \n" + "movdqu " MEMACCESS2(0x10,0) ",%%xmm1 \n" + "lea " MEMLEA(0x20,0) ",%0 \n" + "movdqu %%xmm0," MEMACCESS(1) " \n" + "movdqu %%xmm1," MEMACCESS2(0x10,1) " \n" + "lea " MEMLEA(0x20,1) ",%1 \n" + "sub $0x20,%2 \n" + "jg 1b \n" + : "+r"(src), // %0 + "+r"(dst), // %1 + "+r"(count) // %2 + : + : "memory", "cc" + , "xmm0", "xmm1" + ); +} +#endif // HAS_COPYROW_SSE2 + +#ifdef HAS_COPYROW_AVX +void CopyRow_AVX(const uint8* src, uint8* dst, int count) { + asm volatile ( + LABELALIGN + "1: \n" + "vmovdqu " MEMACCESS(0) ",%%ymm0 \n" + "vmovdqu " MEMACCESS2(0x20,0) ",%%ymm1 \n" + "lea " MEMLEA(0x40,0) ",%0 \n" + "vmovdqu %%ymm0," MEMACCESS(1) " \n" + "vmovdqu %%ymm1," MEMACCESS2(0x20,1) " \n" + "lea " MEMLEA(0x40,1) ",%1 \n" + "sub $0x40,%2 \n" + "jg 1b \n" + : "+r"(src), // %0 + "+r"(dst), // %1 + "+r"(count) // %2 + : + : "memory", "cc" + , "xmm0", "xmm1" + ); +} +#endif // HAS_COPYROW_AVX + +#ifdef HAS_COPYROW_ERMS +// Multiple of 1. +void CopyRow_ERMS(const uint8* src, uint8* dst, int width) { + size_t width_tmp = (size_t)(width); + asm volatile ( + "rep movsb " MEMMOVESTRING(0,1) " \n" + : "+S"(src), // %0 + "+D"(dst), // %1 + "+c"(width_tmp) // %2 + : + : "memory", "cc" + ); +} +#endif // HAS_COPYROW_ERMS + +#ifdef HAS_ARGBCOPYALPHAROW_SSE2 +// width in pixels +void ARGBCopyAlphaRow_SSE2(const uint8* src, uint8* dst, int width) { + asm volatile ( + "pcmpeqb %%xmm0,%%xmm0 \n" + "pslld $0x18,%%xmm0 \n" + "pcmpeqb %%xmm1,%%xmm1 \n" + "psrld $0x8,%%xmm1 \n" + LABELALIGN + "1: \n" + "movdqu " MEMACCESS(0) ",%%xmm2 \n" + "movdqu " MEMACCESS2(0x10,0) ",%%xmm3 \n" + "lea " MEMLEA(0x20,0) ",%0 \n" + "movdqu " MEMACCESS(1) ",%%xmm4 \n" + "movdqu " MEMACCESS2(0x10,1) ",%%xmm5 \n" + "pand %%xmm0,%%xmm2 \n" + "pand %%xmm0,%%xmm3 \n" + "pand %%xmm1,%%xmm4 \n" + "pand %%xmm1,%%xmm5 \n" + "por %%xmm4,%%xmm2 \n" + "por %%xmm5,%%xmm3 \n" + "movdqu %%xmm2," MEMACCESS(1) " \n" + "movdqu %%xmm3," MEMACCESS2(0x10,1) " \n" + "lea " MEMLEA(0x20,1) ",%1 \n" + "sub $0x8,%2 \n" + "jg 1b \n" + : "+r"(src), // %0 + "+r"(dst), // %1 + "+r"(width) // %2 + : + : "memory", "cc" + , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5" + ); +} +#endif // HAS_ARGBCOPYALPHAROW_SSE2 + +#ifdef HAS_ARGBCOPYALPHAROW_AVX2 +// width in pixels +void ARGBCopyAlphaRow_AVX2(const uint8* src, uint8* dst, int width) { + asm volatile ( + "vpcmpeqb %%ymm0,%%ymm0,%%ymm0 \n" + "vpsrld $0x8,%%ymm0,%%ymm0 \n" + LABELALIGN + "1: \n" + "vmovdqu " MEMACCESS(0) ",%%ymm1 \n" + "vmovdqu " MEMACCESS2(0x20,0) ",%%ymm2 \n" + "lea " MEMLEA(0x40,0) ",%0 \n" + "vpblendvb %%ymm0," MEMACCESS(1) ",%%ymm1,%%ymm1 \n" + "vpblendvb %%ymm0," MEMACCESS2(0x20,1) ",%%ymm2,%%ymm2 \n" + "vmovdqu %%ymm1," MEMACCESS(1) " \n" + "vmovdqu %%ymm2," MEMACCESS2(0x20,1) " \n" + "lea " MEMLEA(0x40,1) ",%1 \n" + "sub $0x10,%2 \n" + "jg 1b \n" + "vzeroupper \n" + : "+r"(src), // %0 + "+r"(dst), // %1 + "+r"(width) // %2 + : + : "memory", "cc" + , "xmm0", "xmm1", "xmm2" + ); +} +#endif // HAS_ARGBCOPYALPHAROW_AVX2 + +#ifdef HAS_ARGBCOPYYTOALPHAROW_SSE2 +// width in pixels +void ARGBCopyYToAlphaRow_SSE2(const uint8* src, uint8* dst, int width) { + asm volatile ( + "pcmpeqb %%xmm0,%%xmm0 \n" + "pslld $0x18,%%xmm0 \n" + "pcmpeqb %%xmm1,%%xmm1 \n" + "psrld $0x8,%%xmm1 \n" + LABELALIGN + "1: \n" + "movq " MEMACCESS(0) ",%%xmm2 \n" + "lea " MEMLEA(0x8,0) ",%0 \n" + "punpcklbw %%xmm2,%%xmm2 \n" + "punpckhwd %%xmm2,%%xmm3 \n" + "punpcklwd %%xmm2,%%xmm2 \n" + "movdqu " MEMACCESS(1) ",%%xmm4 \n" + "movdqu " MEMACCESS2(0x10,1) ",%%xmm5 \n" + "pand %%xmm0,%%xmm2 \n" + "pand %%xmm0,%%xmm3 \n" + "pand %%xmm1,%%xmm4 \n" + "pand %%xmm1,%%xmm5 \n" + "por %%xmm4,%%xmm2 \n" + "por %%xmm5,%%xmm3 \n" + "movdqu %%xmm2," MEMACCESS(1) " \n" + "movdqu %%xmm3," MEMACCESS2(0x10,1) " \n" + "lea " MEMLEA(0x20,1) ",%1 \n" + "sub $0x8,%2 \n" + "jg 1b \n" + : "+r"(src), // %0 + "+r"(dst), // %1 + "+r"(width) // %2 + : + : "memory", "cc" + , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5" + ); +} +#endif // HAS_ARGBCOPYYTOALPHAROW_SSE2 + +#ifdef HAS_ARGBCOPYYTOALPHAROW_AVX2 +// width in pixels +void ARGBCopyYToAlphaRow_AVX2(const uint8* src, uint8* dst, int width) { + asm volatile ( + "vpcmpeqb %%ymm0,%%ymm0,%%ymm0 \n" + "vpsrld $0x8,%%ymm0,%%ymm0 \n" + LABELALIGN + "1: \n" + "vpmovzxbd " MEMACCESS(0) ",%%ymm1 \n" + "vpmovzxbd " MEMACCESS2(0x8,0) ",%%ymm2 \n" + "lea " MEMLEA(0x10,0) ",%0 \n" + "vpslld $0x18,%%ymm1,%%ymm1 \n" + "vpslld $0x18,%%ymm2,%%ymm2 \n" + "vpblendvb %%ymm0," MEMACCESS(1) ",%%ymm1,%%ymm1 \n" + "vpblendvb %%ymm0," MEMACCESS2(0x20,1) ",%%ymm2,%%ymm2 \n" + "vmovdqu %%ymm1," MEMACCESS(1) " \n" + "vmovdqu %%ymm2," MEMACCESS2(0x20,1) " \n" + "lea " MEMLEA(0x40,1) ",%1 \n" + "sub $0x10,%2 \n" + "jg 1b \n" + "vzeroupper \n" + : "+r"(src), // %0 + "+r"(dst), // %1 + "+r"(width) // %2 + : + : "memory", "cc" + , "xmm0", "xmm1", "xmm2" + ); +} +#endif // HAS_ARGBCOPYYTOALPHAROW_AVX2 + +#ifdef HAS_SETROW_X86 +void SetRow_X86(uint8* dst, uint8 v8, int width) { + size_t width_tmp = (size_t)(width >> 2); + const uint32 v32 = v8 * 0x01010101; // Duplicate byte to all bytes. + asm volatile ( + "rep stosl " MEMSTORESTRING(eax,0) " \n" + : "+D"(dst), // %0 + "+c"(width_tmp) // %1 + : "a"(v32) // %2 + : "memory", "cc"); +} + +void SetRow_ERMS(uint8* dst, uint8 v8, int width) { + size_t width_tmp = (size_t)(width); + asm volatile ( + "rep stosb " MEMSTORESTRING(al,0) " \n" + : "+D"(dst), // %0 + "+c"(width_tmp) // %1 + : "a"(v8) // %2 + : "memory", "cc"); +} + +void ARGBSetRow_X86(uint8* dst_argb, uint32 v32, int width) { + size_t width_tmp = (size_t)(width); + asm volatile ( + "rep stosl " MEMSTORESTRING(eax,0) " \n" + : "+D"(dst_argb), // %0 + "+c"(width_tmp) // %1 + : "a"(v32) // %2 + : "memory", "cc"); +} +#endif // HAS_SETROW_X86 + +#ifdef HAS_YUY2TOYROW_SSE2 +void YUY2ToYRow_SSE2(const uint8* src_yuy2, uint8* dst_y, int pix) { + asm volatile ( + "pcmpeqb %%xmm5,%%xmm5 \n" + "psrlw $0x8,%%xmm5 \n" + LABELALIGN + "1: \n" + "movdqu " MEMACCESS(0) ",%%xmm0 \n" + "movdqu " MEMACCESS2(0x10,0) ",%%xmm1 \n" + "lea " MEMLEA(0x20,0) ",%0 \n" + "pand %%xmm5,%%xmm0 \n" + "pand %%xmm5,%%xmm1 \n" + "packuswb %%xmm1,%%xmm0 \n" + "movdqu %%xmm0," MEMACCESS(1) " \n" + "lea " MEMLEA(0x10,1) ",%1 \n" + "sub $0x10,%2 \n" + "jg 1b \n" + : "+r"(src_yuy2), // %0 + "+r"(dst_y), // %1 + "+r"(pix) // %2 + : + : "memory", "cc" + , "xmm0", "xmm1", "xmm5" + ); +} + +void YUY2ToUVRow_SSE2(const uint8* src_yuy2, int stride_yuy2, + uint8* dst_u, uint8* dst_v, int pix) { + asm volatile ( + "pcmpeqb %%xmm5,%%xmm5 \n" + "psrlw $0x8,%%xmm5 \n" + "sub %1,%2 \n" + LABELALIGN + "1: \n" + "movdqu " MEMACCESS(0) ",%%xmm0 \n" + "movdqu " MEMACCESS2(0x10,0) ",%%xmm1 \n" + MEMOPREG(movdqu,0x00,0,4,1,xmm2) // movdqu (%0,%4,1),%%xmm2 + MEMOPREG(movdqu,0x10,0,4,1,xmm3) // movdqu 0x10(%0,%4,1),%%xmm3 + "lea " MEMLEA(0x20,0) ",%0 \n" + "pavgb %%xmm2,%%xmm0 \n" + "pavgb %%xmm3,%%xmm1 \n" + "psrlw $0x8,%%xmm0 \n" + "psrlw $0x8,%%xmm1 \n" + "packuswb %%xmm1,%%xmm0 \n" + "movdqa %%xmm0,%%xmm1 \n" + "pand %%xmm5,%%xmm0 \n" + "packuswb %%xmm0,%%xmm0 \n" + "psrlw $0x8,%%xmm1 \n" + "packuswb %%xmm1,%%xmm1 \n" + "movq %%xmm0," MEMACCESS(1) " \n" + MEMOPMEM(movq,xmm1,0x00,1,2,1) // movq %%xmm1,(%1,%2) + "lea " MEMLEA(0x8,1) ",%1 \n" + "sub $0x10,%3 \n" + "jg 1b \n" + : "+r"(src_yuy2), // %0 + "+r"(dst_u), // %1 + "+r"(dst_v), // %2 + "+r"(pix) // %3 + : "r"((intptr_t)(stride_yuy2)) // %4 + : "memory", "cc", NACL_R14 + "xmm0", "xmm1", "xmm2", "xmm3", "xmm5" + ); +} + +void YUY2ToUV422Row_SSE2(const uint8* src_yuy2, + uint8* dst_u, uint8* dst_v, int pix) { + asm volatile ( + "pcmpeqb %%xmm5,%%xmm5 \n" + "psrlw $0x8,%%xmm5 \n" + "sub %1,%2 \n" + LABELALIGN + "1: \n" + "movdqu " MEMACCESS(0) ",%%xmm0 \n" + "movdqu " MEMACCESS2(0x10,0) ",%%xmm1 \n" + "lea " MEMLEA(0x20,0) ",%0 \n" + "psrlw $0x8,%%xmm0 \n" + "psrlw $0x8,%%xmm1 \n" + "packuswb %%xmm1,%%xmm0 \n" + "movdqa %%xmm0,%%xmm1 \n" + "pand %%xmm5,%%xmm0 \n" + "packuswb %%xmm0,%%xmm0 \n" + "psrlw $0x8,%%xmm1 \n" + "packuswb %%xmm1,%%xmm1 \n" + "movq %%xmm0," MEMACCESS(1) " \n" + MEMOPMEM(movq,xmm1,0x00,1,2,1) // movq %%xmm1,(%1,%2) + "lea " MEMLEA(0x8,1) ",%1 \n" + "sub $0x10,%3 \n" + "jg 1b \n" + : "+r"(src_yuy2), // %0 + "+r"(dst_u), // %1 + "+r"(dst_v), // %2 + "+r"(pix) // %3 + : + : "memory", "cc", NACL_R14 + "xmm0", "xmm1", "xmm5" + ); +} + +void UYVYToYRow_SSE2(const uint8* src_uyvy, uint8* dst_y, int pix) { + asm volatile ( + LABELALIGN + "1: \n" + "movdqu " MEMACCESS(0) ",%%xmm0 \n" + "movdqu " MEMACCESS2(0x10,0) ",%%xmm1 \n" + "lea " MEMLEA(0x20,0) ",%0 \n" + "psrlw $0x8,%%xmm0 \n" + "psrlw $0x8,%%xmm1 \n" + "packuswb %%xmm1,%%xmm0 \n" + "movdqu %%xmm0," MEMACCESS(1) " \n" + "lea " MEMLEA(0x10,1) ",%1 \n" + "sub $0x10,%2 \n" + "jg 1b \n" + : "+r"(src_uyvy), // %0 + "+r"(dst_y), // %1 + "+r"(pix) // %2 + : + : "memory", "cc" + , "xmm0", "xmm1" + ); +} + +void UYVYToUVRow_SSE2(const uint8* src_uyvy, int stride_uyvy, + uint8* dst_u, uint8* dst_v, int pix) { + asm volatile ( + "pcmpeqb %%xmm5,%%xmm5 \n" + "psrlw $0x8,%%xmm5 \n" + "sub %1,%2 \n" + LABELALIGN + "1: \n" + "movdqu " MEMACCESS(0) ",%%xmm0 \n" + "movdqu " MEMACCESS2(0x10,0) ",%%xmm1 \n" + MEMOPREG(movdqu,0x00,0,4,1,xmm2) // movdqu (%0,%4,1),%%xmm2 + MEMOPREG(movdqu,0x10,0,4,1,xmm3) // movdqu 0x10(%0,%4,1),%%xmm3 + "lea " MEMLEA(0x20,0) ",%0 \n" + "pavgb %%xmm2,%%xmm0 \n" + "pavgb %%xmm3,%%xmm1 \n" + "pand %%xmm5,%%xmm0 \n" + "pand %%xmm5,%%xmm1 \n" + "packuswb %%xmm1,%%xmm0 \n" + "movdqa %%xmm0,%%xmm1 \n" + "pand %%xmm5,%%xmm0 \n" + "packuswb %%xmm0,%%xmm0 \n" + "psrlw $0x8,%%xmm1 \n" + "packuswb %%xmm1,%%xmm1 \n" + "movq %%xmm0," MEMACCESS(1) " \n" + MEMOPMEM(movq,xmm1,0x00,1,2,1) // movq %%xmm1,(%1,%2) + "lea " MEMLEA(0x8,1) ",%1 \n" + "sub $0x10,%3 \n" + "jg 1b \n" + : "+r"(src_uyvy), // %0 + "+r"(dst_u), // %1 + "+r"(dst_v), // %2 + "+r"(pix) // %3 + : "r"((intptr_t)(stride_uyvy)) // %4 + : "memory", "cc", NACL_R14 + "xmm0", "xmm1", "xmm2", "xmm3", "xmm5" + ); +} + +void UYVYToUV422Row_SSE2(const uint8* src_uyvy, + uint8* dst_u, uint8* dst_v, int pix) { + asm volatile ( + "pcmpeqb %%xmm5,%%xmm5 \n" + "psrlw $0x8,%%xmm5 \n" + "sub %1,%2 \n" + LABELALIGN + "1: \n" + "movdqu " MEMACCESS(0) ",%%xmm0 \n" + "movdqu " MEMACCESS2(0x10,0) ",%%xmm1 \n" + "lea " MEMLEA(0x20,0) ",%0 \n" + "pand %%xmm5,%%xmm0 \n" + "pand %%xmm5,%%xmm1 \n" + "packuswb %%xmm1,%%xmm0 \n" + "movdqa %%xmm0,%%xmm1 \n" + "pand %%xmm5,%%xmm0 \n" + "packuswb %%xmm0,%%xmm0 \n" + "psrlw $0x8,%%xmm1 \n" + "packuswb %%xmm1,%%xmm1 \n" + "movq %%xmm0," MEMACCESS(1) " \n" + MEMOPMEM(movq,xmm1,0x00,1,2,1) // movq %%xmm1,(%1,%2) + "lea " MEMLEA(0x8,1) ",%1 \n" + "sub $0x10,%3 \n" + "jg 1b \n" + : "+r"(src_uyvy), // %0 + "+r"(dst_u), // %1 + "+r"(dst_v), // %2 + "+r"(pix) // %3 + : + : "memory", "cc", NACL_R14 + "xmm0", "xmm1", "xmm5" + ); +} +#endif // HAS_YUY2TOYROW_SSE2 + +#ifdef HAS_YUY2TOYROW_AVX2 +void YUY2ToYRow_AVX2(const uint8* src_yuy2, uint8* dst_y, int pix) { + asm volatile ( + "vpcmpeqb %%ymm5,%%ymm5,%%ymm5 \n" + "vpsrlw $0x8,%%ymm5,%%ymm5 \n" + LABELALIGN + "1: \n" + "vmovdqu " MEMACCESS(0) ",%%ymm0 \n" + "vmovdqu " MEMACCESS2(0x20,0) ",%%ymm1 \n" + "lea " MEMLEA(0x40,0) ",%0 \n" + "vpand %%ymm5,%%ymm0,%%ymm0 \n" + "vpand %%ymm5,%%ymm1,%%ymm1 \n" + "vpackuswb %%ymm1,%%ymm0,%%ymm0 \n" + "vpermq $0xd8,%%ymm0,%%ymm0 \n" + "vmovdqu %%ymm0," MEMACCESS(1) " \n" + "lea " MEMLEA(0x20,1) ",%1 \n" + "sub $0x20,%2 \n" + "jg 1b \n" + "vzeroupper \n" + : "+r"(src_yuy2), // %0 + "+r"(dst_y), // %1 + "+r"(pix) // %2 + : + : "memory", "cc" + , "xmm0", "xmm1", "xmm5" + ); +} + +void YUY2ToUVRow_AVX2(const uint8* src_yuy2, int stride_yuy2, + uint8* dst_u, uint8* dst_v, int pix) { + asm volatile ( + "vpcmpeqb %%ymm5,%%ymm5,%%ymm5 \n" + "vpsrlw $0x8,%%ymm5,%%ymm5 \n" + "sub %1,%2 \n" + LABELALIGN + "1: \n" + "vmovdqu " MEMACCESS(0) ",%%ymm0 \n" + "vmovdqu " MEMACCESS2(0x20,0) ",%%ymm1 \n" + VMEMOPREG(vpavgb,0x00,0,4,1,ymm0,ymm0) // vpavgb (%0,%4,1),%%ymm0,%%ymm0 + VMEMOPREG(vpavgb,0x20,0,4,1,ymm1,ymm1) + "lea " MEMLEA(0x40,0) ",%0 \n" + "vpsrlw $0x8,%%ymm0,%%ymm0 \n" + "vpsrlw $0x8,%%ymm1,%%ymm1 \n" + "vpackuswb %%ymm1,%%ymm0,%%ymm0 \n" + "vpermq $0xd8,%%ymm0,%%ymm0 \n" + "vpand %%ymm5,%%ymm0,%%ymm1 \n" + "vpsrlw $0x8,%%ymm0,%%ymm0 \n" + "vpackuswb %%ymm1,%%ymm1,%%ymm1 \n" + "vpackuswb %%ymm0,%%ymm0,%%ymm0 \n" + "vpermq $0xd8,%%ymm1,%%ymm1 \n" + "vpermq $0xd8,%%ymm0,%%ymm0 \n" + "vextractf128 $0x0,%%ymm1," MEMACCESS(1) " \n" + VEXTOPMEM(vextractf128,0,ymm0,0x00,1,2,1) // vextractf128 $0x0,%%ymm0,(%1,%2,1) + "lea " MEMLEA(0x10,1) ",%1 \n" + "sub $0x20,%3 \n" + "jg 1b \n" + "vzeroupper \n" + : "+r"(src_yuy2), // %0 + "+r"(dst_u), // %1 + "+r"(dst_v), // %2 + "+r"(pix) // %3 + : "r"((intptr_t)(stride_yuy2)) // %4 + : "memory", "cc", NACL_R14 + "xmm0", "xmm1", "xmm5" + ); +} + +void YUY2ToUV422Row_AVX2(const uint8* src_yuy2, + uint8* dst_u, uint8* dst_v, int pix) { + asm volatile ( + "vpcmpeqb %%ymm5,%%ymm5,%%ymm5 \n" + "vpsrlw $0x8,%%ymm5,%%ymm5 \n" + "sub %1,%2 \n" + LABELALIGN + "1: \n" + "vmovdqu " MEMACCESS(0) ",%%ymm0 \n" + "vmovdqu " MEMACCESS2(0x20,0) ",%%ymm1 \n" + "lea " MEMLEA(0x40,0) ",%0 \n" + "vpsrlw $0x8,%%ymm0,%%ymm0 \n" + "vpsrlw $0x8,%%ymm1,%%ymm1 \n" + "vpackuswb %%ymm1,%%ymm0,%%ymm0 \n" + "vpermq $0xd8,%%ymm0,%%ymm0 \n" + "vpand %%ymm5,%%ymm0,%%ymm1 \n" + "vpsrlw $0x8,%%ymm0,%%ymm0 \n" + "vpackuswb %%ymm1,%%ymm1,%%ymm1 \n" + "vpackuswb %%ymm0,%%ymm0,%%ymm0 \n" + "vpermq $0xd8,%%ymm1,%%ymm1 \n" + "vpermq $0xd8,%%ymm0,%%ymm0 \n" + "vextractf128 $0x0,%%ymm1," MEMACCESS(1) " \n" + VEXTOPMEM(vextractf128,0,ymm0,0x00,1,2,1) // vextractf128 $0x0,%%ymm0,(%1,%2,1) + "lea " MEMLEA(0x10,1) ",%1 \n" + "sub $0x20,%3 \n" + "jg 1b \n" + "vzeroupper \n" + : "+r"(src_yuy2), // %0 + "+r"(dst_u), // %1 + "+r"(dst_v), // %2 + "+r"(pix) // %3 + : + : "memory", "cc", NACL_R14 + "xmm0", "xmm1", "xmm5" + ); +} + +void UYVYToYRow_AVX2(const uint8* src_uyvy, uint8* dst_y, int pix) { + asm volatile ( + LABELALIGN + "1: \n" + "vmovdqu " MEMACCESS(0) ",%%ymm0 \n" + "vmovdqu " MEMACCESS2(0x20,0) ",%%ymm1 \n" + "lea " MEMLEA(0x40,0) ",%0 \n" + "vpsrlw $0x8,%%ymm0,%%ymm0 \n" + "vpsrlw $0x8,%%ymm1,%%ymm1 \n" + "vpackuswb %%ymm1,%%ymm0,%%ymm0 \n" + "vpermq $0xd8,%%ymm0,%%ymm0 \n" + "vmovdqu %%ymm0," MEMACCESS(1) " \n" + "lea " MEMLEA(0x20,1) ",%1 \n" + "sub $0x20,%2 \n" + "jg 1b \n" + "vzeroupper \n" + : "+r"(src_uyvy), // %0 + "+r"(dst_y), // %1 + "+r"(pix) // %2 + : + : "memory", "cc" + , "xmm0", "xmm1", "xmm5" + ); +} +void UYVYToUVRow_AVX2(const uint8* src_uyvy, int stride_uyvy, + uint8* dst_u, uint8* dst_v, int pix) { + asm volatile ( + "vpcmpeqb %%ymm5,%%ymm5,%%ymm5 \n" + "vpsrlw $0x8,%%ymm5,%%ymm5 \n" + "sub %1,%2 \n" + + LABELALIGN + "1: \n" + "vmovdqu " MEMACCESS(0) ",%%ymm0 \n" + "vmovdqu " MEMACCESS2(0x20,0) ",%%ymm1 \n" + VMEMOPREG(vpavgb,0x00,0,4,1,ymm0,ymm0) // vpavgb (%0,%4,1),%%ymm0,%%ymm0 + VMEMOPREG(vpavgb,0x20,0,4,1,ymm1,ymm1) + "lea " MEMLEA(0x40,0) ",%0 \n" + "vpand %%ymm5,%%ymm0,%%ymm0 \n" + "vpand %%ymm5,%%ymm1,%%ymm1 \n" + "vpackuswb %%ymm1,%%ymm0,%%ymm0 \n" + "vpermq $0xd8,%%ymm0,%%ymm0 \n" + "vpand %%ymm5,%%ymm0,%%ymm1 \n" + "vpsrlw $0x8,%%ymm0,%%ymm0 \n" + "vpackuswb %%ymm1,%%ymm1,%%ymm1 \n" + "vpackuswb %%ymm0,%%ymm0,%%ymm0 \n" + "vpermq $0xd8,%%ymm1,%%ymm1 \n" + "vpermq $0xd8,%%ymm0,%%ymm0 \n" + "vextractf128 $0x0,%%ymm1," MEMACCESS(1) " \n" + VEXTOPMEM(vextractf128,0,ymm0,0x00,1,2,1) // vextractf128 $0x0,%%ymm0,(%1,%2,1) + "lea " MEMLEA(0x10,1) ",%1 \n" + "sub $0x20,%3 \n" + "jg 1b \n" + "vzeroupper \n" + : "+r"(src_uyvy), // %0 + "+r"(dst_u), // %1 + "+r"(dst_v), // %2 + "+r"(pix) // %3 + : "r"((intptr_t)(stride_uyvy)) // %4 + : "memory", "cc", NACL_R14 + "xmm0", "xmm1", "xmm5" + ); +} + +void UYVYToUV422Row_AVX2(const uint8* src_uyvy, + uint8* dst_u, uint8* dst_v, int pix) { + asm volatile ( + "vpcmpeqb %%ymm5,%%ymm5,%%ymm5 \n" + "vpsrlw $0x8,%%ymm5,%%ymm5 \n" + "sub %1,%2 \n" + LABELALIGN + "1: \n" + "vmovdqu " MEMACCESS(0) ",%%ymm0 \n" + "vmovdqu " MEMACCESS2(0x20,0) ",%%ymm1 \n" + "lea " MEMLEA(0x40,0) ",%0 \n" + "vpand %%ymm5,%%ymm0,%%ymm0 \n" + "vpand %%ymm5,%%ymm1,%%ymm1 \n" + "vpackuswb %%ymm1,%%ymm0,%%ymm0 \n" + "vpermq $0xd8,%%ymm0,%%ymm0 \n" + "vpand %%ymm5,%%ymm0,%%ymm1 \n" + "vpsrlw $0x8,%%ymm0,%%ymm0 \n" + "vpackuswb %%ymm1,%%ymm1,%%ymm1 \n" + "vpackuswb %%ymm0,%%ymm0,%%ymm0 \n" + "vpermq $0xd8,%%ymm1,%%ymm1 \n" + "vpermq $0xd8,%%ymm0,%%ymm0 \n" + "vextractf128 $0x0,%%ymm1," MEMACCESS(1) " \n" + VEXTOPMEM(vextractf128,0,ymm0,0x00,1,2,1) // vextractf128 $0x0,%%ymm0,(%1,%2,1) + "lea " MEMLEA(0x10,1) ",%1 \n" + "sub $0x20,%3 \n" + "jg 1b \n" + "vzeroupper \n" + : "+r"(src_uyvy), // %0 + "+r"(dst_u), // %1 + "+r"(dst_v), // %2 + "+r"(pix) // %3 + : + : "memory", "cc", NACL_R14 + "xmm0", "xmm1", "xmm5" + ); +} +#endif // HAS_YUY2TOYROW_AVX2 + +#ifdef HAS_ARGBBLENDROW_SSE2 +// Blend 8 pixels at a time. +void ARGBBlendRow_SSE2(const uint8* src_argb0, const uint8* src_argb1, + uint8* dst_argb, int width) { + asm volatile ( + "pcmpeqb %%xmm7,%%xmm7 \n" + "psrlw $0xf,%%xmm7 \n" + "pcmpeqb %%xmm6,%%xmm6 \n" + "psrlw $0x8,%%xmm6 \n" + "pcmpeqb %%xmm5,%%xmm5 \n" + "psllw $0x8,%%xmm5 \n" + "pcmpeqb %%xmm4,%%xmm4 \n" + "pslld $0x18,%%xmm4 \n" + "sub $0x4,%3 \n" + "jl 49f \n" + + // 4 pixel loop. + LABELALIGN + "41: \n" + "movdqu " MEMACCESS(0) ",%%xmm3 \n" + "lea " MEMLEA(0x10,0) ",%0 \n" + "movdqa %%xmm3,%%xmm0 \n" + "pxor %%xmm4,%%xmm3 \n" + "movdqu " MEMACCESS(1) ",%%xmm2 \n" + "psrlw $0x8,%%xmm3 \n" + "pshufhw $0xf5,%%xmm3,%%xmm3 \n" + "pshuflw $0xf5,%%xmm3,%%xmm3 \n" + "pand %%xmm6,%%xmm2 \n" + "paddw %%xmm7,%%xmm3 \n" + "pmullw %%xmm3,%%xmm2 \n" + "movdqu " MEMACCESS(1) ",%%xmm1 \n" + "lea " MEMLEA(0x10,1) ",%1 \n" + "psrlw $0x8,%%xmm1 \n" + "por %%xmm4,%%xmm0 \n" + "pmullw %%xmm3,%%xmm1 \n" + "psrlw $0x8,%%xmm2 \n" + "paddusb %%xmm2,%%xmm0 \n" + "pand %%xmm5,%%xmm1 \n" + "paddusb %%xmm1,%%xmm0 \n" + "movdqu %%xmm0," MEMACCESS(2) " \n" + "lea " MEMLEA(0x10,2) ",%2 \n" + "sub $0x4,%3 \n" + "jge 41b \n" + + "49: \n" + "add $0x3,%3 \n" + "jl 99f \n" + + // 1 pixel loop. + "91: \n" + "movd " MEMACCESS(0) ",%%xmm3 \n" + "lea " MEMLEA(0x4,0) ",%0 \n" + "movdqa %%xmm3,%%xmm0 \n" + "pxor %%xmm4,%%xmm3 \n" + "movd " MEMACCESS(1) ",%%xmm2 \n" + "psrlw $0x8,%%xmm3 \n" + "pshufhw $0xf5,%%xmm3,%%xmm3 \n" + "pshuflw $0xf5,%%xmm3,%%xmm3 \n" + "pand %%xmm6,%%xmm2 \n" + "paddw %%xmm7,%%xmm3 \n" + "pmullw %%xmm3,%%xmm2 \n" + "movd " MEMACCESS(1) ",%%xmm1 \n" + "lea " MEMLEA(0x4,1) ",%1 \n" + "psrlw $0x8,%%xmm1 \n" + "por %%xmm4,%%xmm0 \n" + "pmullw %%xmm3,%%xmm1 \n" + "psrlw $0x8,%%xmm2 \n" + "paddusb %%xmm2,%%xmm0 \n" + "pand %%xmm5,%%xmm1 \n" + "paddusb %%xmm1,%%xmm0 \n" + "movd %%xmm0," MEMACCESS(2) " \n" + "lea " MEMLEA(0x4,2) ",%2 \n" + "sub $0x1,%3 \n" + "jge 91b \n" + "99: \n" + : "+r"(src_argb0), // %0 + "+r"(src_argb1), // %1 + "+r"(dst_argb), // %2 + "+r"(width) // %3 + : + : "memory", "cc" + , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7" + ); +} +#endif // HAS_ARGBBLENDROW_SSE2 + +#ifdef HAS_ARGBBLENDROW_SSSE3 +// Shuffle table for isolating alpha. +static uvec8 kShuffleAlpha = { + 3u, 0x80, 3u, 0x80, 7u, 0x80, 7u, 0x80, + 11u, 0x80, 11u, 0x80, 15u, 0x80, 15u, 0x80 +}; + +// Blend 8 pixels at a time +// Shuffle table for reversing the bytes. + +// Same as SSE2, but replaces +// psrlw xmm3, 8 // alpha +// pshufhw xmm3, xmm3,0F5h // 8 alpha words +// pshuflw xmm3, xmm3,0F5h +// with.. +// pshufb xmm3, kShuffleAlpha // alpha + +void ARGBBlendRow_SSSE3(const uint8* src_argb0, const uint8* src_argb1, + uint8* dst_argb, int width) { + asm volatile ( + "pcmpeqb %%xmm7,%%xmm7 \n" + "psrlw $0xf,%%xmm7 \n" + "pcmpeqb %%xmm6,%%xmm6 \n" + "psrlw $0x8,%%xmm6 \n" + "pcmpeqb %%xmm5,%%xmm5 \n" + "psllw $0x8,%%xmm5 \n" + "pcmpeqb %%xmm4,%%xmm4 \n" + "pslld $0x18,%%xmm4 \n" + "sub $0x4,%3 \n" + "jl 49f \n" + + // 4 pixel loop. + LABELALIGN + "40: \n" + "movdqu " MEMACCESS(0) ",%%xmm3 \n" + "lea " MEMLEA(0x10,0) ",%0 \n" + "movdqa %%xmm3,%%xmm0 \n" + "pxor %%xmm4,%%xmm3 \n" + "movdqu " MEMACCESS(1) ",%%xmm2 \n" + "pshufb %4,%%xmm3 \n" + "pand %%xmm6,%%xmm2 \n" + "paddw %%xmm7,%%xmm3 \n" + "pmullw %%xmm3,%%xmm2 \n" + "movdqu " MEMACCESS(1) ",%%xmm1 \n" + "lea " MEMLEA(0x10,1) ",%1 \n" + "psrlw $0x8,%%xmm1 \n" + "por %%xmm4,%%xmm0 \n" + "pmullw %%xmm3,%%xmm1 \n" + "psrlw $0x8,%%xmm2 \n" + "paddusb %%xmm2,%%xmm0 \n" + "pand %%xmm5,%%xmm1 \n" + "paddusb %%xmm1,%%xmm0 \n" + "movdqu %%xmm0," MEMACCESS(2) " \n" + "lea " MEMLEA(0x10,2) ",%2 \n" + "sub $0x4,%3 \n" + "jge 40b \n" + + "49: \n" + "add $0x3,%3 \n" + "jl 99f \n" + + // 1 pixel loop. + "91: \n" + "movd " MEMACCESS(0) ",%%xmm3 \n" + "lea " MEMLEA(0x4,0) ",%0 \n" + "movdqa %%xmm3,%%xmm0 \n" + "pxor %%xmm4,%%xmm3 \n" + "movd " MEMACCESS(1) ",%%xmm2 \n" + "pshufb %4,%%xmm3 \n" + "pand %%xmm6,%%xmm2 \n" + "paddw %%xmm7,%%xmm3 \n" + "pmullw %%xmm3,%%xmm2 \n" + "movd " MEMACCESS(1) ",%%xmm1 \n" + "lea " MEMLEA(0x4,1) ",%1 \n" + "psrlw $0x8,%%xmm1 \n" + "por %%xmm4,%%xmm0 \n" + "pmullw %%xmm3,%%xmm1 \n" + "psrlw $0x8,%%xmm2 \n" + "paddusb %%xmm2,%%xmm0 \n" + "pand %%xmm5,%%xmm1 \n" + "paddusb %%xmm1,%%xmm0 \n" + "movd %%xmm0," MEMACCESS(2) " \n" + "lea " MEMLEA(0x4,2) ",%2 \n" + "sub $0x1,%3 \n" + "jge 91b \n" + "99: \n" + : "+r"(src_argb0), // %0 + "+r"(src_argb1), // %1 + "+r"(dst_argb), // %2 + "+r"(width) // %3 + : "m"(kShuffleAlpha) // %4 + : "memory", "cc" + , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7" + ); +} +#endif // HAS_ARGBBLENDROW_SSSE3 + +#ifdef HAS_ARGBATTENUATEROW_SSE2 +// Attenuate 4 pixels at a time. +void ARGBAttenuateRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width) { + asm volatile ( + "pcmpeqb %%xmm4,%%xmm4 \n" + "pslld $0x18,%%xmm4 \n" + "pcmpeqb %%xmm5,%%xmm5 \n" + "psrld $0x8,%%xmm5 \n" + + // 4 pixel loop. + LABELALIGN + "1: \n" + "movdqu " MEMACCESS(0) ",%%xmm0 \n" + "punpcklbw %%xmm0,%%xmm0 \n" + "pshufhw $0xff,%%xmm0,%%xmm2 \n" + "pshuflw $0xff,%%xmm2,%%xmm2 \n" + "pmulhuw %%xmm2,%%xmm0 \n" + "movdqu " MEMACCESS(0) ",%%xmm1 \n" + "punpckhbw %%xmm1,%%xmm1 \n" + "pshufhw $0xff,%%xmm1,%%xmm2 \n" + "pshuflw $0xff,%%xmm2,%%xmm2 \n" + "pmulhuw %%xmm2,%%xmm1 \n" + "movdqu " MEMACCESS(0) ",%%xmm2 \n" + "lea " MEMLEA(0x10,0) ",%0 \n" + "psrlw $0x8,%%xmm0 \n" + "pand %%xmm4,%%xmm2 \n" + "psrlw $0x8,%%xmm1 \n" + "packuswb %%xmm1,%%xmm0 \n" + "pand %%xmm5,%%xmm0 \n" + "por %%xmm2,%%xmm0 \n" + "movdqu %%xmm0," MEMACCESS(1) " \n" + "lea " MEMLEA(0x10,1) ",%1 \n" + "sub $0x4,%2 \n" + "jg 1b \n" + : "+r"(src_argb), // %0 + "+r"(dst_argb), // %1 + "+r"(width) // %2 + : + : "memory", "cc" + , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5" + ); +} +#endif // HAS_ARGBATTENUATEROW_SSE2 + +#ifdef HAS_ARGBATTENUATEROW_SSSE3 +// Shuffle table duplicating alpha +static uvec8 kShuffleAlpha0 = { + 3u, 3u, 3u, 3u, 3u, 3u, 128u, 128u, 7u, 7u, 7u, 7u, 7u, 7u, 128u, 128u +}; +static uvec8 kShuffleAlpha1 = { + 11u, 11u, 11u, 11u, 11u, 11u, 128u, 128u, + 15u, 15u, 15u, 15u, 15u, 15u, 128u, 128u +}; +// Attenuate 4 pixels at a time. +void ARGBAttenuateRow_SSSE3(const uint8* src_argb, uint8* dst_argb, int width) { + asm volatile ( + "pcmpeqb %%xmm3,%%xmm3 \n" + "pslld $0x18,%%xmm3 \n" + "movdqa %3,%%xmm4 \n" + "movdqa %4,%%xmm5 \n" + + // 4 pixel loop. + LABELALIGN + "1: \n" + "movdqu " MEMACCESS(0) ",%%xmm0 \n" + "pshufb %%xmm4,%%xmm0 \n" + "movdqu " MEMACCESS(0) ",%%xmm1 \n" + "punpcklbw %%xmm1,%%xmm1 \n" + "pmulhuw %%xmm1,%%xmm0 \n" + "movdqu " MEMACCESS(0) ",%%xmm1 \n" + "pshufb %%xmm5,%%xmm1 \n" + "movdqu " MEMACCESS(0) ",%%xmm2 \n" + "punpckhbw %%xmm2,%%xmm2 \n" + "pmulhuw %%xmm2,%%xmm1 \n" + "movdqu " MEMACCESS(0) ",%%xmm2 \n" + "lea " MEMLEA(0x10,0) ",%0 \n" + "pand %%xmm3,%%xmm2 \n" + "psrlw $0x8,%%xmm0 \n" + "psrlw $0x8,%%xmm1 \n" + "packuswb %%xmm1,%%xmm0 \n" + "por %%xmm2,%%xmm0 \n" + "movdqu %%xmm0," MEMACCESS(1) " \n" + "lea " MEMLEA(0x10,1) ",%1 \n" + "sub $0x4,%2 \n" + "jg 1b \n" + : "+r"(src_argb), // %0 + "+r"(dst_argb), // %1 + "+r"(width) // %2 + : "m"(kShuffleAlpha0), // %3 + "m"(kShuffleAlpha1) // %4 + : "memory", "cc" + , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5" + ); +} +#endif // HAS_ARGBATTENUATEROW_SSSE3 + +#ifdef HAS_ARGBATTENUATEROW_AVX2 +// Shuffle table duplicating alpha. +static const uvec8 kShuffleAlpha_AVX2 = { + 6u, 7u, 6u, 7u, 6u, 7u, 128u, 128u, 14u, 15u, 14u, 15u, 14u, 15u, 128u, 128u +}; +// Attenuate 8 pixels at a time. +void ARGBAttenuateRow_AVX2(const uint8* src_argb, uint8* dst_argb, int width) { + asm volatile ( + "vbroadcastf128 %3,%%ymm4 \n" + "vpcmpeqb %%ymm5,%%ymm5,%%ymm5 \n" + "vpslld $0x18,%%ymm5,%%ymm5 \n" + "sub %0,%1 \n" + + // 8 pixel loop. + LABELALIGN + "1: \n" + "vmovdqu " MEMACCESS(0) ",%%ymm6 \n" + "vpunpcklbw %%ymm6,%%ymm6,%%ymm0 \n" + "vpunpckhbw %%ymm6,%%ymm6,%%ymm1 \n" + "vpshufb %%ymm4,%%ymm0,%%ymm2 \n" + "vpshufb %%ymm4,%%ymm1,%%ymm3 \n" + "vpmulhuw %%ymm2,%%ymm0,%%ymm0 \n" + "vpmulhuw %%ymm3,%%ymm1,%%ymm1 \n" + "vpand %%ymm5,%%ymm6,%%ymm6 \n" + "vpsrlw $0x8,%%ymm0,%%ymm0 \n" + "vpsrlw $0x8,%%ymm1,%%ymm1 \n" + "vpackuswb %%ymm1,%%ymm0,%%ymm0 \n" + "vpor %%ymm6,%%ymm0,%%ymm0 \n" + MEMOPMEM(vmovdqu,ymm0,0x00,0,1,1) // vmovdqu %%ymm0,(%0,%1) + "lea " MEMLEA(0x20,0) ",%0 \n" + "sub $0x8,%2 \n" + "jg 1b \n" + "vzeroupper \n" + : "+r"(src_argb), // %0 + "+r"(dst_argb), // %1 + "+r"(width) // %2 + : "m"(kShuffleAlpha_AVX2) // %3 + : "memory", "cc" + , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6" + ); +} +#endif // HAS_ARGBATTENUATEROW_AVX2 + +#ifdef HAS_ARGBUNATTENUATEROW_SSE2 +// Unattenuate 4 pixels at a time. +void ARGBUnattenuateRow_SSE2(const uint8* src_argb, uint8* dst_argb, + int width) { + uintptr_t alpha = 0; + asm volatile ( + // 4 pixel loop. + LABELALIGN + "1: \n" + "movdqu " MEMACCESS(0) ",%%xmm0 \n" + "movzb " MEMACCESS2(0x03,0) ",%3 \n" + "punpcklbw %%xmm0,%%xmm0 \n" + MEMOPREG(movd,0x00,4,3,4,xmm2) // movd 0x0(%4,%3,4),%%xmm2 + "movzb " MEMACCESS2(0x07,0) ",%3 \n" + MEMOPREG(movd,0x00,4,3,4,xmm3) // movd 0x0(%4,%3,4),%%xmm3 + "pshuflw $0x40,%%xmm2,%%xmm2 \n" + "pshuflw $0x40,%%xmm3,%%xmm3 \n" + "movlhps %%xmm3,%%xmm2 \n" + "pmulhuw %%xmm2,%%xmm0 \n" + "movdqu " MEMACCESS(0) ",%%xmm1 \n" + "movzb " MEMACCESS2(0x0b,0) ",%3 \n" + "punpckhbw %%xmm1,%%xmm1 \n" + MEMOPREG(movd,0x00,4,3,4,xmm2) // movd 0x0(%4,%3,4),%%xmm2 + "movzb " MEMACCESS2(0x0f,0) ",%3 \n" + MEMOPREG(movd,0x00,4,3,4,xmm3) // movd 0x0(%4,%3,4),%%xmm3 + "pshuflw $0x40,%%xmm2,%%xmm2 \n" + "pshuflw $0x40,%%xmm3,%%xmm3 \n" + "movlhps %%xmm3,%%xmm2 \n" + "pmulhuw %%xmm2,%%xmm1 \n" + "lea " MEMLEA(0x10,0) ",%0 \n" + "packuswb %%xmm1,%%xmm0 \n" + "movdqu %%xmm0," MEMACCESS(1) " \n" + "lea " MEMLEA(0x10,1) ",%1 \n" + "sub $0x4,%2 \n" + "jg 1b \n" + : "+r"(src_argb), // %0 + "+r"(dst_argb), // %1 + "+r"(width), // %2 + "+r"(alpha) // %3 + : "r"(fixed_invtbl8) // %4 + : "memory", "cc", NACL_R14 + "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5" + ); +} +#endif // HAS_ARGBUNATTENUATEROW_SSE2 + +#ifdef HAS_ARGBUNATTENUATEROW_AVX2 +// Shuffle table duplicating alpha. +static const uvec8 kUnattenShuffleAlpha_AVX2 = { + 0u, 1u, 0u, 1u, 0u, 1u, 6u, 7u, 8u, 9u, 8u, 9u, 8u, 9u, 14u, 15u +}; +// Unattenuate 8 pixels at a time. +void ARGBUnattenuateRow_AVX2(const uint8* src_argb, uint8* dst_argb, + int width) { + uintptr_t alpha = 0; + asm volatile ( + "sub %0,%1 \n" + "vbroadcastf128 %5,%%ymm5 \n" + + // 8 pixel loop. + LABELALIGN + "1: \n" + // replace VPGATHER + "movzb " MEMACCESS2(0x03,0) ",%3 \n" + MEMOPREG(vmovd,0x00,4,3,4,xmm0) // vmovd 0x0(%4,%3,4),%%xmm0 + "movzb " MEMACCESS2(0x07,0) ",%3 \n" + MEMOPREG(vmovd,0x00,4,3,4,xmm1) // vmovd 0x0(%4,%3,4),%%xmm1 + "movzb " MEMACCESS2(0x0b,0) ",%3 \n" + "vpunpckldq %%xmm1,%%xmm0,%%xmm6 \n" + MEMOPREG(vmovd,0x00,4,3,4,xmm2) // vmovd 0x0(%4,%3,4),%%xmm2 + "movzb " MEMACCESS2(0x0f,0) ",%3 \n" + MEMOPREG(vmovd,0x00,4,3,4,xmm3) // vmovd 0x0(%4,%3,4),%%xmm3 + "movzb " MEMACCESS2(0x13,0) ",%3 \n" + "vpunpckldq %%xmm3,%%xmm2,%%xmm7 \n" + MEMOPREG(vmovd,0x00,4,3,4,xmm0) // vmovd 0x0(%4,%3,4),%%xmm0 + "movzb " MEMACCESS2(0x17,0) ",%3 \n" + MEMOPREG(vmovd,0x00,4,3,4,xmm1) // vmovd 0x0(%4,%3,4),%%xmm1 + "movzb " MEMACCESS2(0x1b,0) ",%3 \n" + "vpunpckldq %%xmm1,%%xmm0,%%xmm0 \n" + MEMOPREG(vmovd,0x00,4,3,4,xmm2) // vmovd 0x0(%4,%3,4),%%xmm2 + "movzb " MEMACCESS2(0x1f,0) ",%3 \n" + MEMOPREG(vmovd,0x00,4,3,4,xmm3) // vmovd 0x0(%4,%3,4),%%xmm3 + "vpunpckldq %%xmm3,%%xmm2,%%xmm2 \n" + "vpunpcklqdq %%xmm7,%%xmm6,%%xmm3 \n" + "vpunpcklqdq %%xmm2,%%xmm0,%%xmm0 \n" + "vinserti128 $0x1,%%xmm0,%%ymm3,%%ymm3 \n" + // end of VPGATHER + + "vmovdqu " MEMACCESS(0) ",%%ymm6 \n" + "vpunpcklbw %%ymm6,%%ymm6,%%ymm0 \n" + "vpunpckhbw %%ymm6,%%ymm6,%%ymm1 \n" + "vpunpcklwd %%ymm3,%%ymm3,%%ymm2 \n" + "vpunpckhwd %%ymm3,%%ymm3,%%ymm3 \n" + "vpshufb %%ymm5,%%ymm2,%%ymm2 \n" + "vpshufb %%ymm5,%%ymm3,%%ymm3 \n" + "vpmulhuw %%ymm2,%%ymm0,%%ymm0 \n" + "vpmulhuw %%ymm3,%%ymm1,%%ymm1 \n" + "vpackuswb %%ymm1,%%ymm0,%%ymm0 \n" + MEMOPMEM(vmovdqu,ymm0,0x00,0,1,1) // vmovdqu %%ymm0,(%0,%1) + "lea " MEMLEA(0x20,0) ",%0 \n" + "sub $0x8,%2 \n" + "jg 1b \n" + "vzeroupper \n" + : "+r"(src_argb), // %0 + "+r"(dst_argb), // %1 + "+r"(width), // %2 + "+r"(alpha) // %3 + : "r"(fixed_invtbl8), // %4 + "m"(kUnattenShuffleAlpha_AVX2) // %5 + : "memory", "cc", NACL_R14 + "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7" + ); +} +#endif // HAS_ARGBUNATTENUATEROW_AVX2 + +#ifdef HAS_ARGBGRAYROW_SSSE3 +// Convert 8 ARGB pixels (64 bytes) to 8 Gray ARGB pixels +void ARGBGrayRow_SSSE3(const uint8* src_argb, uint8* dst_argb, int width) { + asm volatile ( + "movdqa %3,%%xmm4 \n" + "movdqa %4,%%xmm5 \n" + + // 8 pixel loop. + LABELALIGN + "1: \n" + "movdqu " MEMACCESS(0) ",%%xmm0 \n" + "movdqu " MEMACCESS2(0x10,0) ",%%xmm1 \n" + "pmaddubsw %%xmm4,%%xmm0 \n" + "pmaddubsw %%xmm4,%%xmm1 \n" + "phaddw %%xmm1,%%xmm0 \n" + "paddw %%xmm5,%%xmm0 \n" + "psrlw $0x7,%%xmm0 \n" + "packuswb %%xmm0,%%xmm0 \n" + "movdqu " MEMACCESS(0) ",%%xmm2 \n" + "movdqu " MEMACCESS2(0x10,0) ",%%xmm3 \n" + "lea " MEMLEA(0x20,0) ",%0 \n" + "psrld $0x18,%%xmm2 \n" + "psrld $0x18,%%xmm3 \n" + "packuswb %%xmm3,%%xmm2 \n" + "packuswb %%xmm2,%%xmm2 \n" + "movdqa %%xmm0,%%xmm3 \n" + "punpcklbw %%xmm0,%%xmm0 \n" + "punpcklbw %%xmm2,%%xmm3 \n" + "movdqa %%xmm0,%%xmm1 \n" + "punpcklwd %%xmm3,%%xmm0 \n" + "punpckhwd %%xmm3,%%xmm1 \n" + "movdqu %%xmm0," MEMACCESS(1) " \n" + "movdqu %%xmm1," MEMACCESS2(0x10,1) " \n" + "lea " MEMLEA(0x20,1) ",%1 \n" + "sub $0x8,%2 \n" + "jg 1b \n" + : "+r"(src_argb), // %0 + "+r"(dst_argb), // %1 + "+r"(width) // %2 + : "m"(kARGBToYJ), // %3 + "m"(kAddYJ64) // %4 + : "memory", "cc" + , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5" + ); +} +#endif // HAS_ARGBGRAYROW_SSSE3 + +#ifdef HAS_ARGBSEPIAROW_SSSE3 +// b = (r * 35 + g * 68 + b * 17) >> 7 +// g = (r * 45 + g * 88 + b * 22) >> 7 +// r = (r * 50 + g * 98 + b * 24) >> 7 +// Constant for ARGB color to sepia tone +static vec8 kARGBToSepiaB = { + 17, 68, 35, 0, 17, 68, 35, 0, 17, 68, 35, 0, 17, 68, 35, 0 +}; + +static vec8 kARGBToSepiaG = { + 22, 88, 45, 0, 22, 88, 45, 0, 22, 88, 45, 0, 22, 88, 45, 0 +}; + +static vec8 kARGBToSepiaR = { + 24, 98, 50, 0, 24, 98, 50, 0, 24, 98, 50, 0, 24, 98, 50, 0 +}; + +// Convert 8 ARGB pixels (32 bytes) to 8 Sepia ARGB pixels. +void ARGBSepiaRow_SSSE3(uint8* dst_argb, int width) { + asm volatile ( + "movdqa %2,%%xmm2 \n" + "movdqa %3,%%xmm3 \n" + "movdqa %4,%%xmm4 \n" + + // 8 pixel loop. + LABELALIGN + "1: \n" + "movdqu " MEMACCESS(0) ",%%xmm0 \n" + "movdqu " MEMACCESS2(0x10,0) ",%%xmm6 \n" + "pmaddubsw %%xmm2,%%xmm0 \n" + "pmaddubsw %%xmm2,%%xmm6 \n" + "phaddw %%xmm6,%%xmm0 \n" + "psrlw $0x7,%%xmm0 \n" + "packuswb %%xmm0,%%xmm0 \n" + "movdqu " MEMACCESS(0) ",%%xmm5 \n" + "movdqu " MEMACCESS2(0x10,0) ",%%xmm1 \n" + "pmaddubsw %%xmm3,%%xmm5 \n" + "pmaddubsw %%xmm3,%%xmm1 \n" + "phaddw %%xmm1,%%xmm5 \n" + "psrlw $0x7,%%xmm5 \n" + "packuswb %%xmm5,%%xmm5 \n" + "punpcklbw %%xmm5,%%xmm0 \n" + "movdqu " MEMACCESS(0) ",%%xmm5 \n" + "movdqu " MEMACCESS2(0x10,0) ",%%xmm1 \n" + "pmaddubsw %%xmm4,%%xmm5 \n" + "pmaddubsw %%xmm4,%%xmm1 \n" + "phaddw %%xmm1,%%xmm5 \n" + "psrlw $0x7,%%xmm5 \n" + "packuswb %%xmm5,%%xmm5 \n" + "movdqu " MEMACCESS(0) ",%%xmm6 \n" + "movdqu " MEMACCESS2(0x10,0) ",%%xmm1 \n" + "psrld $0x18,%%xmm6 \n" + "psrld $0x18,%%xmm1 \n" + "packuswb %%xmm1,%%xmm6 \n" + "packuswb %%xmm6,%%xmm6 \n" + "punpcklbw %%xmm6,%%xmm5 \n" + "movdqa %%xmm0,%%xmm1 \n" + "punpcklwd %%xmm5,%%xmm0 \n" + "punpckhwd %%xmm5,%%xmm1 \n" + "movdqu %%xmm0," MEMACCESS(0) " \n" + "movdqu %%xmm1," MEMACCESS2(0x10,0) " \n" + "lea " MEMLEA(0x20,0) ",%0 \n" + "sub $0x8,%1 \n" + "jg 1b \n" + : "+r"(dst_argb), // %0 + "+r"(width) // %1 + : "m"(kARGBToSepiaB), // %2 + "m"(kARGBToSepiaG), // %3 + "m"(kARGBToSepiaR) // %4 + : "memory", "cc" + , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6" + ); +} +#endif // HAS_ARGBSEPIAROW_SSSE3 + +#ifdef HAS_ARGBCOLORMATRIXROW_SSSE3 +// Tranform 8 ARGB pixels (32 bytes) with color matrix. +// Same as Sepia except matrix is provided. +void ARGBColorMatrixRow_SSSE3(const uint8* src_argb, uint8* dst_argb, + const int8* matrix_argb, int width) { + asm volatile ( + "movdqu " MEMACCESS(3) ",%%xmm5 \n" + "pshufd $0x00,%%xmm5,%%xmm2 \n" + "pshufd $0x55,%%xmm5,%%xmm3 \n" + "pshufd $0xaa,%%xmm5,%%xmm4 \n" + "pshufd $0xff,%%xmm5,%%xmm5 \n" + + // 8 pixel loop. + LABELALIGN + "1: \n" + "movdqu " MEMACCESS(0) ",%%xmm0 \n" + "movdqu " MEMACCESS2(0x10,0) ",%%xmm7 \n" + "pmaddubsw %%xmm2,%%xmm0 \n" + "pmaddubsw %%xmm2,%%xmm7 \n" + "movdqu " MEMACCESS(0) ",%%xmm6 \n" + "movdqu " MEMACCESS2(0x10,0) ",%%xmm1 \n" + "pmaddubsw %%xmm3,%%xmm6 \n" + "pmaddubsw %%xmm3,%%xmm1 \n" + "phaddsw %%xmm7,%%xmm0 \n" + "phaddsw %%xmm1,%%xmm6 \n" + "psraw $0x6,%%xmm0 \n" + "psraw $0x6,%%xmm6 \n" + "packuswb %%xmm0,%%xmm0 \n" + "packuswb %%xmm6,%%xmm6 \n" + "punpcklbw %%xmm6,%%xmm0 \n" + "movdqu " MEMACCESS(0) ",%%xmm1 \n" + "movdqu " MEMACCESS2(0x10,0) ",%%xmm7 \n" + "pmaddubsw %%xmm4,%%xmm1 \n" + "pmaddubsw %%xmm4,%%xmm7 \n" + "phaddsw %%xmm7,%%xmm1 \n" + "movdqu " MEMACCESS(0) ",%%xmm6 \n" + "movdqu " MEMACCESS2(0x10,0) ",%%xmm7 \n" + "pmaddubsw %%xmm5,%%xmm6 \n" + "pmaddubsw %%xmm5,%%xmm7 \n" + "phaddsw %%xmm7,%%xmm6 \n" + "psraw $0x6,%%xmm1 \n" + "psraw $0x6,%%xmm6 \n" + "packuswb %%xmm1,%%xmm1 \n" + "packuswb %%xmm6,%%xmm6 \n" + "punpcklbw %%xmm6,%%xmm1 \n" + "movdqa %%xmm0,%%xmm6 \n" + "punpcklwd %%xmm1,%%xmm0 \n" + "punpckhwd %%xmm1,%%xmm6 \n" + "movdqu %%xmm0," MEMACCESS(1) " \n" + "movdqu %%xmm6," MEMACCESS2(0x10,1) " \n" + "lea " MEMLEA(0x20,0) ",%0 \n" + "lea " MEMLEA(0x20,1) ",%1 \n" + "sub $0x8,%2 \n" + "jg 1b \n" + : "+r"(src_argb), // %0 + "+r"(dst_argb), // %1 + "+r"(width) // %2 + : "r"(matrix_argb) // %3 + : "memory", "cc" + , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7" + ); +} +#endif // HAS_ARGBCOLORMATRIXROW_SSSE3 + +#ifdef HAS_ARGBQUANTIZEROW_SSE2 +// Quantize 4 ARGB pixels (16 bytes). +void ARGBQuantizeRow_SSE2(uint8* dst_argb, int scale, int interval_size, + int interval_offset, int width) { + asm volatile ( + "movd %2,%%xmm2 \n" + "movd %3,%%xmm3 \n" + "movd %4,%%xmm4 \n" + "pshuflw $0x40,%%xmm2,%%xmm2 \n" + "pshufd $0x44,%%xmm2,%%xmm2 \n" + "pshuflw $0x40,%%xmm3,%%xmm3 \n" + "pshufd $0x44,%%xmm3,%%xmm3 \n" + "pshuflw $0x40,%%xmm4,%%xmm4 \n" + "pshufd $0x44,%%xmm4,%%xmm4 \n" + "pxor %%xmm5,%%xmm5 \n" + "pcmpeqb %%xmm6,%%xmm6 \n" + "pslld $0x18,%%xmm6 \n" + + // 4 pixel loop. + LABELALIGN + "1: \n" + "movdqu " MEMACCESS(0) ",%%xmm0 \n" + "punpcklbw %%xmm5,%%xmm0 \n" + "pmulhuw %%xmm2,%%xmm0 \n" + "movdqu " MEMACCESS(0) ",%%xmm1 \n" + "punpckhbw %%xmm5,%%xmm1 \n" + "pmulhuw %%xmm2,%%xmm1 \n" + "pmullw %%xmm3,%%xmm0 \n" + "movdqu " MEMACCESS(0) ",%%xmm7 \n" + "pmullw %%xmm3,%%xmm1 \n" + "pand %%xmm6,%%xmm7 \n" + "paddw %%xmm4,%%xmm0 \n" + "paddw %%xmm4,%%xmm1 \n" + "packuswb %%xmm1,%%xmm0 \n" + "por %%xmm7,%%xmm0 \n" + "movdqu %%xmm0," MEMACCESS(0) " \n" + "lea " MEMLEA(0x10,0) ",%0 \n" + "sub $0x4,%1 \n" + "jg 1b \n" + : "+r"(dst_argb), // %0 + "+r"(width) // %1 + : "r"(scale), // %2 + "r"(interval_size), // %3 + "r"(interval_offset) // %4 + : "memory", "cc" + , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7" + ); +} +#endif // HAS_ARGBQUANTIZEROW_SSE2 + +#ifdef HAS_ARGBSHADEROW_SSE2 +// Shade 4 pixels at a time by specified value. +void ARGBShadeRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width, + uint32 value) { + asm volatile ( + "movd %3,%%xmm2 \n" + "punpcklbw %%xmm2,%%xmm2 \n" + "punpcklqdq %%xmm2,%%xmm2 \n" + + // 4 pixel loop. + LABELALIGN + "1: \n" + "movdqu " MEMACCESS(0) ",%%xmm0 \n" + "lea " MEMLEA(0x10,0) ",%0 \n" + "movdqa %%xmm0,%%xmm1 \n" + "punpcklbw %%xmm0,%%xmm0 \n" + "punpckhbw %%xmm1,%%xmm1 \n" + "pmulhuw %%xmm2,%%xmm0 \n" + "pmulhuw %%xmm2,%%xmm1 \n" + "psrlw $0x8,%%xmm0 \n" + "psrlw $0x8,%%xmm1 \n" + "packuswb %%xmm1,%%xmm0 \n" + "movdqu %%xmm0," MEMACCESS(1) " \n" + "lea " MEMLEA(0x10,1) ",%1 \n" + "sub $0x4,%2 \n" + "jg 1b \n" + : "+r"(src_argb), // %0 + "+r"(dst_argb), // %1 + "+r"(width) // %2 + : "r"(value) // %3 + : "memory", "cc" + , "xmm0", "xmm1", "xmm2" + ); +} +#endif // HAS_ARGBSHADEROW_SSE2 + +#ifdef HAS_ARGBMULTIPLYROW_SSE2 +// Multiply 2 rows of ARGB pixels together, 4 pixels at a time. +void ARGBMultiplyRow_SSE2(const uint8* src_argb0, const uint8* src_argb1, + uint8* dst_argb, int width) { + asm volatile ( + "pxor %%xmm5,%%xmm5 \n" + + // 4 pixel loop. + LABELALIGN + "1: \n" + "movdqu " MEMACCESS(0) ",%%xmm0 \n" + "lea " MEMLEA(0x10,0) ",%0 \n" + "movdqu " MEMACCESS(1) ",%%xmm2 \n" + "lea " MEMLEA(0x10,1) ",%1 \n" + "movdqu %%xmm0,%%xmm1 \n" + "movdqu %%xmm2,%%xmm3 \n" + "punpcklbw %%xmm0,%%xmm0 \n" + "punpckhbw %%xmm1,%%xmm1 \n" + "punpcklbw %%xmm5,%%xmm2 \n" + "punpckhbw %%xmm5,%%xmm3 \n" + "pmulhuw %%xmm2,%%xmm0 \n" + "pmulhuw %%xmm3,%%xmm1 \n" + "packuswb %%xmm1,%%xmm0 \n" + "movdqu %%xmm0," MEMACCESS(2) " \n" + "lea " MEMLEA(0x10,2) ",%2 \n" + "sub $0x4,%3 \n" + "jg 1b \n" + : "+r"(src_argb0), // %0 + "+r"(src_argb1), // %1 + "+r"(dst_argb), // %2 + "+r"(width) // %3 + : + : "memory", "cc" + , "xmm0", "xmm1", "xmm2", "xmm3", "xmm5" + ); +} +#endif // HAS_ARGBMULTIPLYROW_SSE2 + +#ifdef HAS_ARGBMULTIPLYROW_AVX2 +// Multiply 2 rows of ARGB pixels together, 8 pixels at a time. +void ARGBMultiplyRow_AVX2(const uint8* src_argb0, const uint8* src_argb1, + uint8* dst_argb, int width) { + asm volatile ( + "vpxor %%ymm5,%%ymm5,%%ymm5 \n" + + // 4 pixel loop. + LABELALIGN + "1: \n" + "vmovdqu " MEMACCESS(0) ",%%ymm1 \n" + "lea " MEMLEA(0x20,0) ",%0 \n" + "vmovdqu " MEMACCESS(1) ",%%ymm3 \n" + "lea " MEMLEA(0x20,1) ",%1 \n" + "vpunpcklbw %%ymm1,%%ymm1,%%ymm0 \n" + "vpunpckhbw %%ymm1,%%ymm1,%%ymm1 \n" + "vpunpcklbw %%ymm5,%%ymm3,%%ymm2 \n" + "vpunpckhbw %%ymm5,%%ymm3,%%ymm3 \n" + "vpmulhuw %%ymm2,%%ymm0,%%ymm0 \n" + "vpmulhuw %%ymm3,%%ymm1,%%ymm1 \n" + "vpackuswb %%ymm1,%%ymm0,%%ymm0 \n" + "vmovdqu %%ymm0," MEMACCESS(2) " \n" + "lea " MEMLEA(0x20,2) ",%2 \n" + "sub $0x8,%3 \n" + "jg 1b \n" + "vzeroupper \n" + : "+r"(src_argb0), // %0 + "+r"(src_argb1), // %1 + "+r"(dst_argb), // %2 + "+r"(width) // %3 + : + : "memory", "cc" +#if defined(__AVX2__) + , "xmm0", "xmm1", "xmm2", "xmm3", "xmm5" +#endif + ); +} +#endif // HAS_ARGBMULTIPLYROW_AVX2 + +#ifdef HAS_ARGBADDROW_SSE2 +// Add 2 rows of ARGB pixels together, 4 pixels at a time. +void ARGBAddRow_SSE2(const uint8* src_argb0, const uint8* src_argb1, + uint8* dst_argb, int width) { + asm volatile ( + // 4 pixel loop. + LABELALIGN + "1: \n" + "movdqu " MEMACCESS(0) ",%%xmm0 \n" + "lea " MEMLEA(0x10,0) ",%0 \n" + "movdqu " MEMACCESS(1) ",%%xmm1 \n" + "lea " MEMLEA(0x10,1) ",%1 \n" + "paddusb %%xmm1,%%xmm0 \n" + "movdqu %%xmm0," MEMACCESS(2) " \n" + "lea " MEMLEA(0x10,2) ",%2 \n" + "sub $0x4,%3 \n" + "jg 1b \n" + : "+r"(src_argb0), // %0 + "+r"(src_argb1), // %1 + "+r"(dst_argb), // %2 + "+r"(width) // %3 + : + : "memory", "cc" + , "xmm0", "xmm1" + ); +} +#endif // HAS_ARGBADDROW_SSE2 + +#ifdef HAS_ARGBADDROW_AVX2 +// Add 2 rows of ARGB pixels together, 4 pixels at a time. +void ARGBAddRow_AVX2(const uint8* src_argb0, const uint8* src_argb1, + uint8* dst_argb, int width) { + asm volatile ( + // 4 pixel loop. + LABELALIGN + "1: \n" + "vmovdqu " MEMACCESS(0) ",%%ymm0 \n" + "lea " MEMLEA(0x20,0) ",%0 \n" + "vpaddusb " MEMACCESS(1) ",%%ymm0,%%ymm0 \n" + "lea " MEMLEA(0x20,1) ",%1 \n" + "vmovdqu %%ymm0," MEMACCESS(2) " \n" + "lea " MEMLEA(0x20,2) ",%2 \n" + "sub $0x8,%3 \n" + "jg 1b \n" + "vzeroupper \n" + : "+r"(src_argb0), // %0 + "+r"(src_argb1), // %1 + "+r"(dst_argb), // %2 + "+r"(width) // %3 + : + : "memory", "cc" + , "xmm0" + ); +} +#endif // HAS_ARGBADDROW_AVX2 + +#ifdef HAS_ARGBSUBTRACTROW_SSE2 +// Subtract 2 rows of ARGB pixels, 4 pixels at a time. +void ARGBSubtractRow_SSE2(const uint8* src_argb0, const uint8* src_argb1, + uint8* dst_argb, int width) { + asm volatile ( + // 4 pixel loop. + LABELALIGN + "1: \n" + "movdqu " MEMACCESS(0) ",%%xmm0 \n" + "lea " MEMLEA(0x10,0) ",%0 \n" + "movdqu " MEMACCESS(1) ",%%xmm1 \n" + "lea " MEMLEA(0x10,1) ",%1 \n" + "psubusb %%xmm1,%%xmm0 \n" + "movdqu %%xmm0," MEMACCESS(2) " \n" + "lea " MEMLEA(0x10,2) ",%2 \n" + "sub $0x4,%3 \n" + "jg 1b \n" + : "+r"(src_argb0), // %0 + "+r"(src_argb1), // %1 + "+r"(dst_argb), // %2 + "+r"(width) // %3 + : + : "memory", "cc" + , "xmm0", "xmm1" + ); +} +#endif // HAS_ARGBSUBTRACTROW_SSE2 + +#ifdef HAS_ARGBSUBTRACTROW_AVX2 +// Subtract 2 rows of ARGB pixels, 8 pixels at a time. +void ARGBSubtractRow_AVX2(const uint8* src_argb0, const uint8* src_argb1, + uint8* dst_argb, int width) { + asm volatile ( + // 4 pixel loop. + LABELALIGN + "1: \n" + "vmovdqu " MEMACCESS(0) ",%%ymm0 \n" + "lea " MEMLEA(0x20,0) ",%0 \n" + "vpsubusb " MEMACCESS(1) ",%%ymm0,%%ymm0 \n" + "lea " MEMLEA(0x20,1) ",%1 \n" + "vmovdqu %%ymm0," MEMACCESS(2) " \n" + "lea " MEMLEA(0x20,2) ",%2 \n" + "sub $0x8,%3 \n" + "jg 1b \n" + "vzeroupper \n" + : "+r"(src_argb0), // %0 + "+r"(src_argb1), // %1 + "+r"(dst_argb), // %2 + "+r"(width) // %3 + : + : "memory", "cc" + , "xmm0" + ); +} +#endif // HAS_ARGBSUBTRACTROW_AVX2 + +#ifdef HAS_SOBELXROW_SSE2 +// SobelX as a matrix is +// -1 0 1 +// -2 0 2 +// -1 0 1 +void SobelXRow_SSE2(const uint8* src_y0, const uint8* src_y1, + const uint8* src_y2, uint8* dst_sobelx, int width) { + asm volatile ( + "sub %0,%1 \n" + "sub %0,%2 \n" + "sub %0,%3 \n" + "pxor %%xmm5,%%xmm5 \n" + + // 8 pixel loop. + LABELALIGN + "1: \n" + "movq " MEMACCESS(0) ",%%xmm0 \n" + "movq " MEMACCESS2(0x2,0) ",%%xmm1 \n" + "punpcklbw %%xmm5,%%xmm0 \n" + "punpcklbw %%xmm5,%%xmm1 \n" + "psubw %%xmm1,%%xmm0 \n" + MEMOPREG(movq,0x00,0,1,1,xmm1) // movq (%0,%1,1),%%xmm1 + MEMOPREG(movq,0x02,0,1,1,xmm2) // movq 0x2(%0,%1,1),%%xmm2 + "punpcklbw %%xmm5,%%xmm1 \n" + "punpcklbw %%xmm5,%%xmm2 \n" + "psubw %%xmm2,%%xmm1 \n" + MEMOPREG(movq,0x00,0,2,1,xmm2) // movq (%0,%2,1),%%xmm2 + MEMOPREG(movq,0x02,0,2,1,xmm3) // movq 0x2(%0,%2,1),%%xmm3 + "punpcklbw %%xmm5,%%xmm2 \n" + "punpcklbw %%xmm5,%%xmm3 \n" + "psubw %%xmm3,%%xmm2 \n" + "paddw %%xmm2,%%xmm0 \n" + "paddw %%xmm1,%%xmm0 \n" + "paddw %%xmm1,%%xmm0 \n" + "pxor %%xmm1,%%xmm1 \n" + "psubw %%xmm0,%%xmm1 \n" + "pmaxsw %%xmm1,%%xmm0 \n" + "packuswb %%xmm0,%%xmm0 \n" + MEMOPMEM(movq,xmm0,0x00,0,3,1) // movq %%xmm0,(%0,%3,1) + "lea " MEMLEA(0x8,0) ",%0 \n" + "sub $0x8,%4 \n" + "jg 1b \n" + : "+r"(src_y0), // %0 + "+r"(src_y1), // %1 + "+r"(src_y2), // %2 + "+r"(dst_sobelx), // %3 + "+r"(width) // %4 + : + : "memory", "cc", NACL_R14 + "xmm0", "xmm1", "xmm2", "xmm3", "xmm5" + ); +} +#endif // HAS_SOBELXROW_SSE2 + +#ifdef HAS_SOBELYROW_SSE2 +// SobelY as a matrix is +// -1 -2 -1 +// 0 0 0 +// 1 2 1 +void SobelYRow_SSE2(const uint8* src_y0, const uint8* src_y1, + uint8* dst_sobely, int width) { + asm volatile ( + "sub %0,%1 \n" + "sub %0,%2 \n" + "pxor %%xmm5,%%xmm5 \n" + + // 8 pixel loop. + LABELALIGN + "1: \n" + "movq " MEMACCESS(0) ",%%xmm0 \n" + MEMOPREG(movq,0x00,0,1,1,xmm1) // movq (%0,%1,1),%%xmm1 + "punpcklbw %%xmm5,%%xmm0 \n" + "punpcklbw %%xmm5,%%xmm1 \n" + "psubw %%xmm1,%%xmm0 \n" + "movq " MEMACCESS2(0x1,0) ",%%xmm1 \n" + MEMOPREG(movq,0x01,0,1,1,xmm2) // movq 0x1(%0,%1,1),%%xmm2 + "punpcklbw %%xmm5,%%xmm1 \n" + "punpcklbw %%xmm5,%%xmm2 \n" + "psubw %%xmm2,%%xmm1 \n" + "movq " MEMACCESS2(0x2,0) ",%%xmm2 \n" + MEMOPREG(movq,0x02,0,1,1,xmm3) // movq 0x2(%0,%1,1),%%xmm3 + "punpcklbw %%xmm5,%%xmm2 \n" + "punpcklbw %%xmm5,%%xmm3 \n" + "psubw %%xmm3,%%xmm2 \n" + "paddw %%xmm2,%%xmm0 \n" + "paddw %%xmm1,%%xmm0 \n" + "paddw %%xmm1,%%xmm0 \n" + "pxor %%xmm1,%%xmm1 \n" + "psubw %%xmm0,%%xmm1 \n" + "pmaxsw %%xmm1,%%xmm0 \n" + "packuswb %%xmm0,%%xmm0 \n" + MEMOPMEM(movq,xmm0,0x00,0,2,1) // movq %%xmm0,(%0,%2,1) + "lea " MEMLEA(0x8,0) ",%0 \n" + "sub $0x8,%3 \n" + "jg 1b \n" + : "+r"(src_y0), // %0 + "+r"(src_y1), // %1 + "+r"(dst_sobely), // %2 + "+r"(width) // %3 + : + : "memory", "cc", NACL_R14 + "xmm0", "xmm1", "xmm2", "xmm3", "xmm5" + ); +} +#endif // HAS_SOBELYROW_SSE2 + +#ifdef HAS_SOBELROW_SSE2 +// Adds Sobel X and Sobel Y and stores Sobel into ARGB. +// A = 255 +// R = Sobel +// G = Sobel +// B = Sobel +void SobelRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely, + uint8* dst_argb, int width) { + asm volatile ( + "sub %0,%1 \n" + "pcmpeqb %%xmm5,%%xmm5 \n" + "pslld $0x18,%%xmm5 \n" + + // 8 pixel loop. + LABELALIGN + "1: \n" + "movdqu " MEMACCESS(0) ",%%xmm0 \n" + MEMOPREG(movdqu,0x00,0,1,1,xmm1) // movdqu (%0,%1,1),%%xmm1 + "lea " MEMLEA(0x10,0) ",%0 \n" + "paddusb %%xmm1,%%xmm0 \n" + "movdqa %%xmm0,%%xmm2 \n" + "punpcklbw %%xmm0,%%xmm2 \n" + "punpckhbw %%xmm0,%%xmm0 \n" + "movdqa %%xmm2,%%xmm1 \n" + "punpcklwd %%xmm2,%%xmm1 \n" + "punpckhwd %%xmm2,%%xmm2 \n" + "por %%xmm5,%%xmm1 \n" + "por %%xmm5,%%xmm2 \n" + "movdqa %%xmm0,%%xmm3 \n" + "punpcklwd %%xmm0,%%xmm3 \n" + "punpckhwd %%xmm0,%%xmm0 \n" + "por %%xmm5,%%xmm3 \n" + "por %%xmm5,%%xmm0 \n" + "movdqu %%xmm1," MEMACCESS(2) " \n" + "movdqu %%xmm2," MEMACCESS2(0x10,2) " \n" + "movdqu %%xmm3," MEMACCESS2(0x20,2) " \n" + "movdqu %%xmm0," MEMACCESS2(0x30,2) " \n" + "lea " MEMLEA(0x40,2) ",%2 \n" + "sub $0x10,%3 \n" + "jg 1b \n" + : "+r"(src_sobelx), // %0 + "+r"(src_sobely), // %1 + "+r"(dst_argb), // %2 + "+r"(width) // %3 + : + : "memory", "cc", NACL_R14 + "xmm0", "xmm1", "xmm2", "xmm3", "xmm5" + ); +} +#endif // HAS_SOBELROW_SSE2 + +#ifdef HAS_SOBELTOPLANEROW_SSE2 +// Adds Sobel X and Sobel Y and stores Sobel into a plane. +void SobelToPlaneRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely, + uint8* dst_y, int width) { + asm volatile ( + "sub %0,%1 \n" + "pcmpeqb %%xmm5,%%xmm5 \n" + "pslld $0x18,%%xmm5 \n" + + // 8 pixel loop. + LABELALIGN + "1: \n" + "movdqu " MEMACCESS(0) ",%%xmm0 \n" + MEMOPREG(movdqu,0x00,0,1,1,xmm1) // movdqu (%0,%1,1),%%xmm1 + "lea " MEMLEA(0x10,0) ",%0 \n" + "paddusb %%xmm1,%%xmm0 \n" + "movdqu %%xmm0," MEMACCESS(2) " \n" + "lea " MEMLEA(0x10,2) ",%2 \n" + "sub $0x10,%3 \n" + "jg 1b \n" + : "+r"(src_sobelx), // %0 + "+r"(src_sobely), // %1 + "+r"(dst_y), // %2 + "+r"(width) // %3 + : + : "memory", "cc", NACL_R14 + "xmm0", "xmm1" + ); +} +#endif // HAS_SOBELTOPLANEROW_SSE2 + +#ifdef HAS_SOBELXYROW_SSE2 +// Mixes Sobel X, Sobel Y and Sobel into ARGB. +// A = 255 +// R = Sobel X +// G = Sobel +// B = Sobel Y +void SobelXYRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely, + uint8* dst_argb, int width) { + asm volatile ( + "sub %0,%1 \n" + "pcmpeqb %%xmm5,%%xmm5 \n" + + // 8 pixel loop. + LABELALIGN + "1: \n" + "movdqu " MEMACCESS(0) ",%%xmm0 \n" + MEMOPREG(movdqu,0x00,0,1,1,xmm1) // movdqu (%0,%1,1),%%xmm1 + "lea " MEMLEA(0x10,0) ",%0 \n" + "movdqa %%xmm0,%%xmm2 \n" + "paddusb %%xmm1,%%xmm2 \n" + "movdqa %%xmm0,%%xmm3 \n" + "punpcklbw %%xmm5,%%xmm3 \n" + "punpckhbw %%xmm5,%%xmm0 \n" + "movdqa %%xmm1,%%xmm4 \n" + "punpcklbw %%xmm2,%%xmm4 \n" + "punpckhbw %%xmm2,%%xmm1 \n" + "movdqa %%xmm4,%%xmm6 \n" + "punpcklwd %%xmm3,%%xmm6 \n" + "punpckhwd %%xmm3,%%xmm4 \n" + "movdqa %%xmm1,%%xmm7 \n" + "punpcklwd %%xmm0,%%xmm7 \n" + "punpckhwd %%xmm0,%%xmm1 \n" + "movdqu %%xmm6," MEMACCESS(2) " \n" + "movdqu %%xmm4," MEMACCESS2(0x10,2) " \n" + "movdqu %%xmm7," MEMACCESS2(0x20,2) " \n" + "movdqu %%xmm1," MEMACCESS2(0x30,2) " \n" + "lea " MEMLEA(0x40,2) ",%2 \n" + "sub $0x10,%3 \n" + "jg 1b \n" + : "+r"(src_sobelx), // %0 + "+r"(src_sobely), // %1 + "+r"(dst_argb), // %2 + "+r"(width) // %3 + : + : "memory", "cc", NACL_R14 + "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7" + ); +} +#endif // HAS_SOBELXYROW_SSE2 + +#ifdef HAS_COMPUTECUMULATIVESUMROW_SSE2 +// Creates a table of cumulative sums where each value is a sum of all values +// above and to the left of the value, inclusive of the value. +void ComputeCumulativeSumRow_SSE2(const uint8* row, int32* cumsum, + const int32* previous_cumsum, int width) { + asm volatile ( + "pxor %%xmm0,%%xmm0 \n" + "pxor %%xmm1,%%xmm1 \n" + "sub $0x4,%3 \n" + "jl 49f \n" + "test $0xf,%1 \n" + "jne 49f \n" + + // 4 pixel loop \n" + LABELALIGN + "40: \n" + "movdqu " MEMACCESS(0) ",%%xmm2 \n" + "lea " MEMLEA(0x10,0) ",%0 \n" + "movdqa %%xmm2,%%xmm4 \n" + "punpcklbw %%xmm1,%%xmm2 \n" + "movdqa %%xmm2,%%xmm3 \n" + "punpcklwd %%xmm1,%%xmm2 \n" + "punpckhwd %%xmm1,%%xmm3 \n" + "punpckhbw %%xmm1,%%xmm4 \n" + "movdqa %%xmm4,%%xmm5 \n" + "punpcklwd %%xmm1,%%xmm4 \n" + "punpckhwd %%xmm1,%%xmm5 \n" + "paddd %%xmm2,%%xmm0 \n" + "movdqu " MEMACCESS(2) ",%%xmm2 \n" + "paddd %%xmm0,%%xmm2 \n" + "paddd %%xmm3,%%xmm0 \n" + "movdqu " MEMACCESS2(0x10,2) ",%%xmm3 \n" + "paddd %%xmm0,%%xmm3 \n" + "paddd %%xmm4,%%xmm0 \n" + "movdqu " MEMACCESS2(0x20,2) ",%%xmm4 \n" + "paddd %%xmm0,%%xmm4 \n" + "paddd %%xmm5,%%xmm0 \n" + "movdqu " MEMACCESS2(0x30,2) ",%%xmm5 \n" + "lea " MEMLEA(0x40,2) ",%2 \n" + "paddd %%xmm0,%%xmm5 \n" + "movdqu %%xmm2," MEMACCESS(1) " \n" + "movdqu %%xmm3," MEMACCESS2(0x10,1) " \n" + "movdqu %%xmm4," MEMACCESS2(0x20,1) " \n" + "movdqu %%xmm5," MEMACCESS2(0x30,1) " \n" + "lea " MEMLEA(0x40,1) ",%1 \n" + "sub $0x4,%3 \n" + "jge 40b \n" + + "49: \n" + "add $0x3,%3 \n" + "jl 19f \n" + + // 1 pixel loop \n" + LABELALIGN + "10: \n" + "movd " MEMACCESS(0) ",%%xmm2 \n" + "lea " MEMLEA(0x4,0) ",%0 \n" + "punpcklbw %%xmm1,%%xmm2 \n" + "punpcklwd %%xmm1,%%xmm2 \n" + "paddd %%xmm2,%%xmm0 \n" + "movdqu " MEMACCESS(2) ",%%xmm2 \n" + "lea " MEMLEA(0x10,2) ",%2 \n" + "paddd %%xmm0,%%xmm2 \n" + "movdqu %%xmm2," MEMACCESS(1) " \n" + "lea " MEMLEA(0x10,1) ",%1 \n" + "sub $0x1,%3 \n" + "jge 10b \n" + + "19: \n" + : "+r"(row), // %0 + "+r"(cumsum), // %1 + "+r"(previous_cumsum), // %2 + "+r"(width) // %3 + : + : "memory", "cc" + , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5" + ); +} +#endif // HAS_COMPUTECUMULATIVESUMROW_SSE2 + +#ifdef HAS_CUMULATIVESUMTOAVERAGEROW_SSE2 +void CumulativeSumToAverageRow_SSE2(const int32* topleft, const int32* botleft, + int width, int area, uint8* dst, + int count) { + asm volatile ( + "movd %5,%%xmm5 \n" + "cvtdq2ps %%xmm5,%%xmm5 \n" + "rcpss %%xmm5,%%xmm4 \n" + "pshufd $0x0,%%xmm4,%%xmm4 \n" + "sub $0x4,%3 \n" + "jl 49f \n" + "cmpl $0x80,%5 \n" + "ja 40f \n" + + "pshufd $0x0,%%xmm5,%%xmm5 \n" + "pcmpeqb %%xmm6,%%xmm6 \n" + "psrld $0x10,%%xmm6 \n" + "cvtdq2ps %%xmm6,%%xmm6 \n" + "addps %%xmm6,%%xmm5 \n" + "mulps %%xmm4,%%xmm5 \n" + "cvtps2dq %%xmm5,%%xmm5 \n" + "packssdw %%xmm5,%%xmm5 \n" + + // 4 pixel small loop \n" + LABELALIGN + "4: \n" + "movdqu " MEMACCESS(0) ",%%xmm0 \n" + "movdqu " MEMACCESS2(0x10,0) ",%%xmm1 \n" + "movdqu " MEMACCESS2(0x20,0) ",%%xmm2 \n" + "movdqu " MEMACCESS2(0x30,0) ",%%xmm3 \n" + MEMOPREG(psubd,0x00,0,4,4,xmm0) // psubd 0x00(%0,%4,4),%%xmm0 + MEMOPREG(psubd,0x10,0,4,4,xmm1) // psubd 0x10(%0,%4,4),%%xmm1 + MEMOPREG(psubd,0x20,0,4,4,xmm2) // psubd 0x20(%0,%4,4),%%xmm2 + MEMOPREG(psubd,0x30,0,4,4,xmm3) // psubd 0x30(%0,%4,4),%%xmm3 + "lea " MEMLEA(0x40,0) ",%0 \n" + "psubd " MEMACCESS(1) ",%%xmm0 \n" + "psubd " MEMACCESS2(0x10,1) ",%%xmm1 \n" + "psubd " MEMACCESS2(0x20,1) ",%%xmm2 \n" + "psubd " MEMACCESS2(0x30,1) ",%%xmm3 \n" + MEMOPREG(paddd,0x00,1,4,4,xmm0) // paddd 0x00(%1,%4,4),%%xmm0 + MEMOPREG(paddd,0x10,1,4,4,xmm1) // paddd 0x10(%1,%4,4),%%xmm1 + MEMOPREG(paddd,0x20,1,4,4,xmm2) // paddd 0x20(%1,%4,4),%%xmm2 + MEMOPREG(paddd,0x30,1,4,4,xmm3) // paddd 0x30(%1,%4,4),%%xmm3 + "lea " MEMLEA(0x40,1) ",%1 \n" + "packssdw %%xmm1,%%xmm0 \n" + "packssdw %%xmm3,%%xmm2 \n" + "pmulhuw %%xmm5,%%xmm0 \n" + "pmulhuw %%xmm5,%%xmm2 \n" + "packuswb %%xmm2,%%xmm0 \n" + "movdqu %%xmm0," MEMACCESS(2) " \n" + "lea " MEMLEA(0x10,2) ",%2 \n" + "sub $0x4,%3 \n" + "jge 4b \n" + "jmp 49f \n" + + // 4 pixel loop \n" + LABELALIGN + "40: \n" + "movdqu " MEMACCESS(0) ",%%xmm0 \n" + "movdqu " MEMACCESS2(0x10,0) ",%%xmm1 \n" + "movdqu " MEMACCESS2(0x20,0) ",%%xmm2 \n" + "movdqu " MEMACCESS2(0x30,0) ",%%xmm3 \n" + MEMOPREG(psubd,0x00,0,4,4,xmm0) // psubd 0x00(%0,%4,4),%%xmm0 + MEMOPREG(psubd,0x10,0,4,4,xmm1) // psubd 0x10(%0,%4,4),%%xmm1 + MEMOPREG(psubd,0x20,0,4,4,xmm2) // psubd 0x20(%0,%4,4),%%xmm2 + MEMOPREG(psubd,0x30,0,4,4,xmm3) // psubd 0x30(%0,%4,4),%%xmm3 + "lea " MEMLEA(0x40,0) ",%0 \n" + "psubd " MEMACCESS(1) ",%%xmm0 \n" + "psubd " MEMACCESS2(0x10,1) ",%%xmm1 \n" + "psubd " MEMACCESS2(0x20,1) ",%%xmm2 \n" + "psubd " MEMACCESS2(0x30,1) ",%%xmm3 \n" + MEMOPREG(paddd,0x00,1,4,4,xmm0) // paddd 0x00(%1,%4,4),%%xmm0 + MEMOPREG(paddd,0x10,1,4,4,xmm1) // paddd 0x10(%1,%4,4),%%xmm1 + MEMOPREG(paddd,0x20,1,4,4,xmm2) // paddd 0x20(%1,%4,4),%%xmm2 + MEMOPREG(paddd,0x30,1,4,4,xmm3) // paddd 0x30(%1,%4,4),%%xmm3 + "lea " MEMLEA(0x40,1) ",%1 \n" + "cvtdq2ps %%xmm0,%%xmm0 \n" + "cvtdq2ps %%xmm1,%%xmm1 \n" + "mulps %%xmm4,%%xmm0 \n" + "mulps %%xmm4,%%xmm1 \n" + "cvtdq2ps %%xmm2,%%xmm2 \n" + "cvtdq2ps %%xmm3,%%xmm3 \n" + "mulps %%xmm4,%%xmm2 \n" + "mulps %%xmm4,%%xmm3 \n" + "cvtps2dq %%xmm0,%%xmm0 \n" + "cvtps2dq %%xmm1,%%xmm1 \n" + "cvtps2dq %%xmm2,%%xmm2 \n" + "cvtps2dq %%xmm3,%%xmm3 \n" + "packssdw %%xmm1,%%xmm0 \n" + "packssdw %%xmm3,%%xmm2 \n" + "packuswb %%xmm2,%%xmm0 \n" + "movdqu %%xmm0," MEMACCESS(2) " \n" + "lea " MEMLEA(0x10,2) ",%2 \n" + "sub $0x4,%3 \n" + "jge 40b \n" + + "49: \n" + "add $0x3,%3 \n" + "jl 19f \n" + + // 1 pixel loop \n" + LABELALIGN + "10: \n" + "movdqu " MEMACCESS(0) ",%%xmm0 \n" + MEMOPREG(psubd,0x00,0,4,4,xmm0) // psubd 0x00(%0,%4,4),%%xmm0 + "lea " MEMLEA(0x10,0) ",%0 \n" + "psubd " MEMACCESS(1) ",%%xmm0 \n" + MEMOPREG(paddd,0x00,1,4,4,xmm0) // paddd 0x00(%1,%4,4),%%xmm0 + "lea " MEMLEA(0x10,1) ",%1 \n" + "cvtdq2ps %%xmm0,%%xmm0 \n" + "mulps %%xmm4,%%xmm0 \n" + "cvtps2dq %%xmm0,%%xmm0 \n" + "packssdw %%xmm0,%%xmm0 \n" + "packuswb %%xmm0,%%xmm0 \n" + "movd %%xmm0," MEMACCESS(2) " \n" + "lea " MEMLEA(0x4,2) ",%2 \n" + "sub $0x1,%3 \n" + "jge 10b \n" + "19: \n" + : "+r"(topleft), // %0 + "+r"(botleft), // %1 + "+r"(dst), // %2 + "+rm"(count) // %3 + : "r"((intptr_t)(width)), // %4 + "rm"(area) // %5 + : "memory", "cc", NACL_R14 + "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6" + ); +} +#endif // HAS_CUMULATIVESUMTOAVERAGEROW_SSE2 + +#ifdef HAS_ARGBAFFINEROW_SSE2 +// Copy ARGB pixels from source image with slope to a row of destination. +LIBYUV_API +void ARGBAffineRow_SSE2(const uint8* src_argb, int src_argb_stride, + uint8* dst_argb, const float* src_dudv, int width) { + intptr_t src_argb_stride_temp = src_argb_stride; + intptr_t temp = 0; + asm volatile ( + "movq " MEMACCESS(3) ",%%xmm2 \n" + "movq " MEMACCESS2(0x08,3) ",%%xmm7 \n" + "shl $0x10,%1 \n" + "add $0x4,%1 \n" + "movd %1,%%xmm5 \n" + "sub $0x4,%4 \n" + "jl 49f \n" + + "pshufd $0x44,%%xmm7,%%xmm7 \n" + "pshufd $0x0,%%xmm5,%%xmm5 \n" + "movdqa %%xmm2,%%xmm0 \n" + "addps %%xmm7,%%xmm0 \n" + "movlhps %%xmm0,%%xmm2 \n" + "movdqa %%xmm7,%%xmm4 \n" + "addps %%xmm4,%%xmm4 \n" + "movdqa %%xmm2,%%xmm3 \n" + "addps %%xmm4,%%xmm3 \n" + "addps %%xmm4,%%xmm4 \n" + + // 4 pixel loop \n" + LABELALIGN + "40: \n" + "cvttps2dq %%xmm2,%%xmm0 \n" // x, y float to int first 2 + "cvttps2dq %%xmm3,%%xmm1 \n" // x, y float to int next 2 + "packssdw %%xmm1,%%xmm0 \n" // x, y as 8 shorts + "pmaddwd %%xmm5,%%xmm0 \n" // off = x * 4 + y * stride + "movd %%xmm0,%k1 \n" + "pshufd $0x39,%%xmm0,%%xmm0 \n" + "movd %%xmm0,%k5 \n" + "pshufd $0x39,%%xmm0,%%xmm0 \n" + MEMOPREG(movd,0x00,0,1,1,xmm1) // movd (%0,%1,1),%%xmm1 + MEMOPREG(movd,0x00,0,5,1,xmm6) // movd (%0,%5,1),%%xmm6 + "punpckldq %%xmm6,%%xmm1 \n" + "addps %%xmm4,%%xmm2 \n" + "movq %%xmm1," MEMACCESS(2) " \n" + "movd %%xmm0,%k1 \n" + "pshufd $0x39,%%xmm0,%%xmm0 \n" + "movd %%xmm0,%k5 \n" + MEMOPREG(movd,0x00,0,1,1,xmm0) // movd (%0,%1,1),%%xmm0 + MEMOPREG(movd,0x00,0,5,1,xmm6) // movd (%0,%5,1),%%xmm6 + "punpckldq %%xmm6,%%xmm0 \n" + "addps %%xmm4,%%xmm3 \n" + "movq %%xmm0," MEMACCESS2(0x08,2) " \n" + "lea " MEMLEA(0x10,2) ",%2 \n" + "sub $0x4,%4 \n" + "jge 40b \n" + + "49: \n" + "add $0x3,%4 \n" + "jl 19f \n" + + // 1 pixel loop \n" + LABELALIGN + "10: \n" + "cvttps2dq %%xmm2,%%xmm0 \n" + "packssdw %%xmm0,%%xmm0 \n" + "pmaddwd %%xmm5,%%xmm0 \n" + "addps %%xmm7,%%xmm2 \n" + "movd %%xmm0,%k1 \n" + MEMOPREG(movd,0x00,0,1,1,xmm0) // movd (%0,%1,1),%%xmm0 + "movd %%xmm0," MEMACCESS(2) " \n" + "lea " MEMLEA(0x04,2) ",%2 \n" + "sub $0x1,%4 \n" + "jge 10b \n" + "19: \n" + : "+r"(src_argb), // %0 + "+r"(src_argb_stride_temp), // %1 + "+r"(dst_argb), // %2 + "+r"(src_dudv), // %3 + "+rm"(width), // %4 + "+r"(temp) // %5 + : + : "memory", "cc", NACL_R14 + "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7" + ); +} +#endif // HAS_ARGBAFFINEROW_SSE2 + +#ifdef HAS_INTERPOLATEROW_SSSE3 +// Bilinear filter 16x2 -> 16x1 +void InterpolateRow_SSSE3(uint8* dst_ptr, const uint8* src_ptr, + ptrdiff_t src_stride, int dst_width, + int source_y_fraction) { + asm volatile ( + "sub %1,%0 \n" + "shr %3 \n" + "cmp $0x0,%3 \n" + "je 100f \n" + "cmp $0x20,%3 \n" + "je 75f \n" + "cmp $0x40,%3 \n" + "je 50f \n" + "cmp $0x60,%3 \n" + "je 25f \n" + + "movd %3,%%xmm0 \n" + "neg %3 \n" + "add $0x80,%3 \n" + "movd %3,%%xmm5 \n" + "punpcklbw %%xmm0,%%xmm5 \n" + "punpcklwd %%xmm5,%%xmm5 \n" + "pshufd $0x0,%%xmm5,%%xmm5 \n" + + // General purpose row blend. + LABELALIGN + "1: \n" + "movdqu " MEMACCESS(1) ",%%xmm0 \n" + MEMOPREG(movdqu,0x00,1,4,1,xmm2) + "movdqa %%xmm0,%%xmm1 \n" + "punpcklbw %%xmm2,%%xmm0 \n" + "punpckhbw %%xmm2,%%xmm1 \n" + "pmaddubsw %%xmm5,%%xmm0 \n" + "pmaddubsw %%xmm5,%%xmm1 \n" + "psrlw $0x7,%%xmm0 \n" + "psrlw $0x7,%%xmm1 \n" + "packuswb %%xmm1,%%xmm0 \n" + MEMOPMEM(movdqu,xmm0,0x00,1,0,1) + "lea " MEMLEA(0x10,1) ",%1 \n" + "sub $0x10,%2 \n" + "jg 1b \n" + "jmp 99f \n" + + // Blend 25 / 75. + LABELALIGN + "25: \n" + "movdqu " MEMACCESS(1) ",%%xmm0 \n" + MEMOPREG(movdqu,0x00,1,4,1,xmm1) + "pavgb %%xmm1,%%xmm0 \n" + "pavgb %%xmm1,%%xmm0 \n" + MEMOPMEM(movdqu,xmm0,0x00,1,0,1) + "lea " MEMLEA(0x10,1) ",%1 \n" + "sub $0x10,%2 \n" + "jg 25b \n" + "jmp 99f \n" + + // Blend 50 / 50. + LABELALIGN + "50: \n" + "movdqu " MEMACCESS(1) ",%%xmm0 \n" + MEMOPREG(movdqu,0x00,1,4,1,xmm1) + "pavgb %%xmm1,%%xmm0 \n" + MEMOPMEM(movdqu,xmm0,0x00,1,0,1) + "lea " MEMLEA(0x10,1) ",%1 \n" + "sub $0x10,%2 \n" + "jg 50b \n" + "jmp 99f \n" + + // Blend 75 / 25. + LABELALIGN + "75: \n" + "movdqu " MEMACCESS(1) ",%%xmm1 \n" + MEMOPREG(movdqu,0x00,1,4,1,xmm0) + "pavgb %%xmm1,%%xmm0 \n" + "pavgb %%xmm1,%%xmm0 \n" + MEMOPMEM(movdqu,xmm0,0x00,1,0,1) + "lea " MEMLEA(0x10,1) ",%1 \n" + "sub $0x10,%2 \n" + "jg 75b \n" + "jmp 99f \n" + + // Blend 100 / 0 - Copy row unchanged. + LABELALIGN + "100: \n" + "movdqu " MEMACCESS(1) ",%%xmm0 \n" + MEMOPMEM(movdqu,xmm0,0x00,1,0,1) + "lea " MEMLEA(0x10,1) ",%1 \n" + "sub $0x10,%2 \n" + "jg 100b \n" + + "99: \n" + : "+r"(dst_ptr), // %0 + "+r"(src_ptr), // %1 + "+r"(dst_width), // %2 + "+r"(source_y_fraction) // %3 + : "r"((intptr_t)(src_stride)) // %4 + : "memory", "cc", NACL_R14 + "xmm0", "xmm1", "xmm2", "xmm5" + ); +} +#endif // HAS_INTERPOLATEROW_SSSE3 + +#ifdef HAS_INTERPOLATEROW_AVX2 +// Bilinear filter 32x2 -> 32x1 +void InterpolateRow_AVX2(uint8* dst_ptr, const uint8* src_ptr, + ptrdiff_t src_stride, int dst_width, + int source_y_fraction) { + asm volatile ( + "shr %3 \n" + "cmp $0x0,%3 \n" + "je 100f \n" + "sub %1,%0 \n" + "cmp $0x20,%3 \n" + "je 75f \n" + "cmp $0x40,%3 \n" + "je 50f \n" + "cmp $0x60,%3 \n" + "je 25f \n" + + "vmovd %3,%%xmm0 \n" + "neg %3 \n" + "add $0x80,%3 \n" + "vmovd %3,%%xmm5 \n" + "vpunpcklbw %%xmm0,%%xmm5,%%xmm5 \n" + "vpunpcklwd %%xmm5,%%xmm5,%%xmm5 \n" + "vpxor %%ymm0,%%ymm0,%%ymm0 \n" + "vpermd %%ymm5,%%ymm0,%%ymm5 \n" + + // General purpose row blend. + LABELALIGN + "1: \n" + "vmovdqu " MEMACCESS(1) ",%%ymm0 \n" + MEMOPREG(vmovdqu,0x00,1,4,1,ymm2) + "vpunpckhbw %%ymm2,%%ymm0,%%ymm1 \n" + "vpunpcklbw %%ymm2,%%ymm0,%%ymm0 \n" + "vpmaddubsw %%ymm5,%%ymm0,%%ymm0 \n" + "vpmaddubsw %%ymm5,%%ymm1,%%ymm1 \n" + "vpsrlw $0x7,%%ymm0,%%ymm0 \n" + "vpsrlw $0x7,%%ymm1,%%ymm1 \n" + "vpackuswb %%ymm1,%%ymm0,%%ymm0 \n" + MEMOPMEM(vmovdqu,ymm0,0x00,1,0,1) + "lea " MEMLEA(0x20,1) ",%1 \n" + "sub $0x20,%2 \n" + "jg 1b \n" + "jmp 99f \n" + + // Blend 25 / 75. + LABELALIGN + "25: \n" + "vmovdqu " MEMACCESS(1) ",%%ymm0 \n" + MEMOPREG(vmovdqu,0x00,1,4,1,ymm1) + "vpavgb %%ymm1,%%ymm0,%%ymm0 \n" + "vpavgb %%ymm1,%%ymm0,%%ymm0 \n" + MEMOPMEM(vmovdqu,ymm0,0x00,1,0,1) + "lea " MEMLEA(0x20,1) ",%1 \n" + "sub $0x20,%2 \n" + "jg 25b \n" + "jmp 99f \n" + + // Blend 50 / 50. + LABELALIGN + "50: \n" + "vmovdqu " MEMACCESS(1) ",%%ymm0 \n" + VMEMOPREG(vpavgb,0x00,1,4,1,ymm0,ymm0) // vpavgb (%1,%4,1),%%ymm0,%%ymm0 + MEMOPMEM(vmovdqu,ymm0,0x00,1,0,1) + "lea " MEMLEA(0x20,1) ",%1 \n" + "sub $0x20,%2 \n" + "jg 50b \n" + "jmp 99f \n" + + // Blend 75 / 25. + LABELALIGN + "75: \n" + "vmovdqu " MEMACCESS(1) ",%%ymm1 \n" + MEMOPREG(vmovdqu,0x00,1,4,1,ymm0) + "vpavgb %%ymm1,%%ymm0,%%ymm0 \n" + "vpavgb %%ymm1,%%ymm0,%%ymm0 \n" + MEMOPMEM(vmovdqu,ymm0,0x00,1,0,1) + "lea " MEMLEA(0x20,1) ",%1 \n" + "sub $0x20,%2 \n" + "jg 75b \n" + "jmp 99f \n" + + // Blend 100 / 0 - Copy row unchanged. + LABELALIGN + "100: \n" + "rep movsb " MEMMOVESTRING(1,0) " \n" + "jmp 999f \n" + + "99: \n" + "vzeroupper \n" + "999: \n" + : "+D"(dst_ptr), // %0 + "+S"(src_ptr), // %1 + "+c"(dst_width), // %2 + "+r"(source_y_fraction) // %3 + : "r"((intptr_t)(src_stride)) // %4 + : "memory", "cc", NACL_R14 + "xmm0", "xmm1", "xmm2", "xmm5" + ); +} +#endif // HAS_INTERPOLATEROW_AVX2 + +#ifdef HAS_INTERPOLATEROW_SSE2 +// Bilinear filter 16x2 -> 16x1 +void InterpolateRow_SSE2(uint8* dst_ptr, const uint8* src_ptr, + ptrdiff_t src_stride, int dst_width, + int source_y_fraction) { + asm volatile ( + "sub %1,%0 \n" + "shr %3 \n" + "cmp $0x0,%3 \n" + "je 100f \n" + "cmp $0x20,%3 \n" + "je 75f \n" + "cmp $0x40,%3 \n" + "je 50f \n" + "cmp $0x60,%3 \n" + "je 25f \n" + + "movd %3,%%xmm0 \n" + "neg %3 \n" + "add $0x80,%3 \n" + "movd %3,%%xmm5 \n" + "punpcklbw %%xmm0,%%xmm5 \n" + "punpcklwd %%xmm5,%%xmm5 \n" + "pshufd $0x0,%%xmm5,%%xmm5 \n" + "pxor %%xmm4,%%xmm4 \n" + + // General purpose row blend. + LABELALIGN + "1: \n" + "movdqu " MEMACCESS(1) ",%%xmm0 \n" + MEMOPREG(movdqu,0x00,1,4,1,xmm2) // movdqu (%1,%4,1),%%xmm2 + "movdqa %%xmm0,%%xmm1 \n" + "movdqa %%xmm2,%%xmm3 \n" + "punpcklbw %%xmm4,%%xmm2 \n" + "punpckhbw %%xmm4,%%xmm3 \n" + "punpcklbw %%xmm4,%%xmm0 \n" + "punpckhbw %%xmm4,%%xmm1 \n" + "psubw %%xmm0,%%xmm2 \n" + "psubw %%xmm1,%%xmm3 \n" + "paddw %%xmm2,%%xmm2 \n" + "paddw %%xmm3,%%xmm3 \n" + "pmulhw %%xmm5,%%xmm2 \n" + "pmulhw %%xmm5,%%xmm3 \n" + "paddw %%xmm2,%%xmm0 \n" + "paddw %%xmm3,%%xmm1 \n" + "packuswb %%xmm1,%%xmm0 \n" + MEMOPMEM(movdqu,xmm0,0x00,1,0,1) // movdqu %%xmm0,(%1,%0,1) + "lea " MEMLEA(0x10,1) ",%1 \n" + "sub $0x10,%2 \n" + "jg 1b \n" + "jmp 99f \n" + + // Blend 25 / 75. + LABELALIGN + "25: \n" + "movdqu " MEMACCESS(1) ",%%xmm0 \n" + MEMOPREG(movdqu,0x00,1,4,1,xmm1) // movdqu (%1,%4,1),%%xmm1 + "pavgb %%xmm1,%%xmm0 \n" + "pavgb %%xmm1,%%xmm0 \n" + MEMOPMEM(movdqu,xmm0,0x00,1,0,1) // movdqu %%xmm0,(%1,%0,1) + "lea " MEMLEA(0x10,1) ",%1 \n" + "sub $0x10,%2 \n" + "jg 25b \n" + "jmp 99f \n" + + // Blend 50 / 50. + LABELALIGN + "50: \n" + "movdqu " MEMACCESS(1) ",%%xmm0 \n" + MEMOPREG(movdqu,0x00,1,4,1,xmm1) // movdqu (%1,%4,1),%%xmm1 + "pavgb %%xmm1,%%xmm0 \n" + MEMOPMEM(movdqu,xmm0,0x00,1,0,1) // movdqu %%xmm0,(%1,%0,1) + "lea " MEMLEA(0x10,1) ",%1 \n" + "sub $0x10,%2 \n" + "jg 50b \n" + "jmp 99f \n" + + // Blend 75 / 25. + LABELALIGN + "75: \n" + "movdqu " MEMACCESS(1) ",%%xmm1 \n" + MEMOPREG(movdqu,0x00,1,4,1,xmm0) // movdqu (%1,%4,1),%%xmm0 + "pavgb %%xmm1,%%xmm0 \n" + "pavgb %%xmm1,%%xmm0 \n" + MEMOPMEM(movdqu,xmm0,0x00,1,0,1) // movdqu %%xmm0,(%1,%0,1) + "lea " MEMLEA(0x10,1) ",%1 \n" + "sub $0x10,%2 \n" + "jg 75b \n" + "jmp 99f \n" + + // Blend 100 / 0 - Copy row unchanged. + LABELALIGN + "100: \n" + "movdqu " MEMACCESS(1) ",%%xmm0 \n" + MEMOPMEM(movdqu,xmm0,0x00,1,0,1) // movdqu %%xmm0,(%1,%0,1) + "lea " MEMLEA(0x10,1) ",%1 \n" + "sub $0x10,%2 \n" + "jg 100b \n" + + "99: \n" + : "+r"(dst_ptr), // %0 + "+r"(src_ptr), // %1 + "+r"(dst_width), // %2 + "+r"(source_y_fraction) // %3 + : "r"((intptr_t)(src_stride)) // %4 + : "memory", "cc", NACL_R14 + "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5" + ); +} +#endif // HAS_INTERPOLATEROW_SSE2 + +#ifdef HAS_ARGBSHUFFLEROW_SSSE3 +// For BGRAToARGB, ABGRToARGB, RGBAToARGB, and ARGBToRGBA. +void ARGBShuffleRow_SSSE3(const uint8* src_argb, uint8* dst_argb, + const uint8* shuffler, int pix) { + asm volatile ( + "movdqu " MEMACCESS(3) ",%%xmm5 \n" + LABELALIGN + "1: \n" + "movdqu " MEMACCESS(0) ",%%xmm0 \n" + "movdqu " MEMACCESS2(0x10,0) ",%%xmm1 \n" + "lea " MEMLEA(0x20,0) ",%0 \n" + "pshufb %%xmm5,%%xmm0 \n" + "pshufb %%xmm5,%%xmm1 \n" + "movdqu %%xmm0," MEMACCESS(1) " \n" + "movdqu %%xmm1," MEMACCESS2(0x10,1) " \n" + "lea " MEMLEA(0x20,1) ",%1 \n" + "sub $0x8,%2 \n" + "jg 1b \n" + : "+r"(src_argb), // %0 + "+r"(dst_argb), // %1 + "+r"(pix) // %2 + : "r"(shuffler) // %3 + : "memory", "cc" + , "xmm0", "xmm1", "xmm5" + ); +} +#endif // HAS_ARGBSHUFFLEROW_SSSE3 + +#ifdef HAS_ARGBSHUFFLEROW_AVX2 +// For BGRAToARGB, ABGRToARGB, RGBAToARGB, and ARGBToRGBA. +void ARGBShuffleRow_AVX2(const uint8* src_argb, uint8* dst_argb, + const uint8* shuffler, int pix) { + asm volatile ( + "vbroadcastf128 " MEMACCESS(3) ",%%ymm5 \n" + LABELALIGN + "1: \n" + "vmovdqu " MEMACCESS(0) ",%%ymm0 \n" + "vmovdqu " MEMACCESS2(0x20,0) ",%%ymm1 \n" + "lea " MEMLEA(0x40,0) ",%0 \n" + "vpshufb %%ymm5,%%ymm0,%%ymm0 \n" + "vpshufb %%ymm5,%%ymm1,%%ymm1 \n" + "vmovdqu %%ymm0," MEMACCESS(1) " \n" + "vmovdqu %%ymm1," MEMACCESS2(0x20,1) " \n" + "lea " MEMLEA(0x40,1) ",%1 \n" + "sub $0x10,%2 \n" + "jg 1b \n" + "vzeroupper \n" + : "+r"(src_argb), // %0 + "+r"(dst_argb), // %1 + "+r"(pix) // %2 + : "r"(shuffler) // %3 + : "memory", "cc" + , "xmm0", "xmm1", "xmm5" + ); +} +#endif // HAS_ARGBSHUFFLEROW_AVX2 + +#ifdef HAS_ARGBSHUFFLEROW_SSE2 +// For BGRAToARGB, ABGRToARGB, RGBAToARGB, and ARGBToRGBA. +void ARGBShuffleRow_SSE2(const uint8* src_argb, uint8* dst_argb, + const uint8* shuffler, int pix) { + uintptr_t pixel_temp = 0u; + asm volatile ( + "pxor %%xmm5,%%xmm5 \n" + "mov " MEMACCESS(4) ",%k2 \n" + "cmp $0x3000102,%k2 \n" + "je 3012f \n" + "cmp $0x10203,%k2 \n" + "je 123f \n" + "cmp $0x30201,%k2 \n" + "je 321f \n" + "cmp $0x2010003,%k2 \n" + "je 2103f \n" + + LABELALIGN + "1: \n" + "movzb " MEMACCESS(4) ",%2 \n" + MEMOPARG(movzb,0x00,0,2,1,2) " \n" // movzb (%0,%2,1),%2 + "mov %b2," MEMACCESS(1) " \n" + "movzb " MEMACCESS2(0x1,4) ",%2 \n" + MEMOPARG(movzb,0x00,0,2,1,2) " \n" // movzb (%0,%2,1),%2 + "mov %b2," MEMACCESS2(0x1,1) " \n" + "movzb " MEMACCESS2(0x2,4) ",%2 \n" + MEMOPARG(movzb,0x00,0,2,1,2) " \n" // movzb (%0,%2,1),%2 + "mov %b2," MEMACCESS2(0x2,1) " \n" + "movzb " MEMACCESS2(0x3,4) ",%2 \n" + MEMOPARG(movzb,0x00,0,2,1,2) " \n" // movzb (%0,%2,1),%2 + "mov %b2," MEMACCESS2(0x3,1) " \n" + "lea " MEMLEA(0x4,0) ",%0 \n" + "lea " MEMLEA(0x4,1) ",%1 \n" + "sub $0x1,%3 \n" + "jg 1b \n" + "jmp 99f \n" + + LABELALIGN + "123: \n" + "movdqu " MEMACCESS(0) ",%%xmm0 \n" + "lea " MEMLEA(0x10,0) ",%0 \n" + "movdqa %%xmm0,%%xmm1 \n" + "punpcklbw %%xmm5,%%xmm0 \n" + "punpckhbw %%xmm5,%%xmm1 \n" + "pshufhw $0x1b,%%xmm0,%%xmm0 \n" + "pshuflw $0x1b,%%xmm0,%%xmm0 \n" + "pshufhw $0x1b,%%xmm1,%%xmm1 \n" + "pshuflw $0x1b,%%xmm1,%%xmm1 \n" + "packuswb %%xmm1,%%xmm0 \n" + "movdqu %%xmm0," MEMACCESS(1) " \n" + "lea " MEMLEA(0x10,1) ",%1 \n" + "sub $0x4,%3 \n" + "jg 123b \n" + "jmp 99f \n" + + LABELALIGN + "321: \n" + "movdqu " MEMACCESS(0) ",%%xmm0 \n" + "lea " MEMLEA(0x10,0) ",%0 \n" + "movdqa %%xmm0,%%xmm1 \n" + "punpcklbw %%xmm5,%%xmm0 \n" + "punpckhbw %%xmm5,%%xmm1 \n" + "pshufhw $0x39,%%xmm0,%%xmm0 \n" + "pshuflw $0x39,%%xmm0,%%xmm0 \n" + "pshufhw $0x39,%%xmm1,%%xmm1 \n" + "pshuflw $0x39,%%xmm1,%%xmm1 \n" + "packuswb %%xmm1,%%xmm0 \n" + "movdqu %%xmm0," MEMACCESS(1) " \n" + "lea " MEMLEA(0x10,1) ",%1 \n" + "sub $0x4,%3 \n" + "jg 321b \n" + "jmp 99f \n" + + LABELALIGN + "2103: \n" + "movdqu " MEMACCESS(0) ",%%xmm0 \n" + "lea " MEMLEA(0x10,0) ",%0 \n" + "movdqa %%xmm0,%%xmm1 \n" + "punpcklbw %%xmm5,%%xmm0 \n" + "punpckhbw %%xmm5,%%xmm1 \n" + "pshufhw $0x93,%%xmm0,%%xmm0 \n" + "pshuflw $0x93,%%xmm0,%%xmm0 \n" + "pshufhw $0x93,%%xmm1,%%xmm1 \n" + "pshuflw $0x93,%%xmm1,%%xmm1 \n" + "packuswb %%xmm1,%%xmm0 \n" + "movdqu %%xmm0," MEMACCESS(1) " \n" + "lea " MEMLEA(0x10,1) ",%1 \n" + "sub $0x4,%3 \n" + "jg 2103b \n" + "jmp 99f \n" + + LABELALIGN + "3012: \n" + "movdqu " MEMACCESS(0) ",%%xmm0 \n" + "lea " MEMLEA(0x10,0) ",%0 \n" + "movdqa %%xmm0,%%xmm1 \n" + "punpcklbw %%xmm5,%%xmm0 \n" + "punpckhbw %%xmm5,%%xmm1 \n" + "pshufhw $0xc6,%%xmm0,%%xmm0 \n" + "pshuflw $0xc6,%%xmm0,%%xmm0 \n" + "pshufhw $0xc6,%%xmm1,%%xmm1 \n" + "pshuflw $0xc6,%%xmm1,%%xmm1 \n" + "packuswb %%xmm1,%%xmm0 \n" + "movdqu %%xmm0," MEMACCESS(1) " \n" + "lea " MEMLEA(0x10,1) ",%1 \n" + "sub $0x4,%3 \n" + "jg 3012b \n" + + "99: \n" + : "+r"(src_argb), // %0 + "+r"(dst_argb), // %1 + "+d"(pixel_temp), // %2 + "+r"(pix) // %3 + : "r"(shuffler) // %4 + : "memory", "cc", NACL_R14 + "xmm0", "xmm1", "xmm5" + ); +} +#endif // HAS_ARGBSHUFFLEROW_SSE2 + +#ifdef HAS_I422TOYUY2ROW_SSE2 +void I422ToYUY2Row_SSE2(const uint8* src_y, + const uint8* src_u, + const uint8* src_v, + uint8* dst_frame, int width) { + asm volatile ( + "sub %1,%2 \n" + LABELALIGN + "1: \n" + "movq " MEMACCESS(1) ",%%xmm2 \n" + MEMOPREG(movq,0x00,1,2,1,xmm3) // movq (%1,%2,1),%%xmm3 + "lea " MEMLEA(0x8,1) ",%1 \n" + "punpcklbw %%xmm3,%%xmm2 \n" + "movdqu " MEMACCESS(0) ",%%xmm0 \n" + "lea " MEMLEA(0x10,0) ",%0 \n" + "movdqa %%xmm0,%%xmm1 \n" + "punpcklbw %%xmm2,%%xmm0 \n" + "punpckhbw %%xmm2,%%xmm1 \n" + "movdqu %%xmm0," MEMACCESS(3) " \n" + "movdqu %%xmm1," MEMACCESS2(0x10,3) " \n" + "lea " MEMLEA(0x20,3) ",%3 \n" + "sub $0x10,%4 \n" + "jg 1b \n" + : "+r"(src_y), // %0 + "+r"(src_u), // %1 + "+r"(src_v), // %2 + "+r"(dst_frame), // %3 + "+rm"(width) // %4 + : + : "memory", "cc", NACL_R14 + "xmm0", "xmm1", "xmm2", "xmm3" + ); +} +#endif // HAS_I422TOYUY2ROW_SSE2 + +#ifdef HAS_I422TOUYVYROW_SSE2 +void I422ToUYVYRow_SSE2(const uint8* src_y, + const uint8* src_u, + const uint8* src_v, + uint8* dst_frame, int width) { + asm volatile ( + "sub %1,%2 \n" + LABELALIGN + "1: \n" + "movq " MEMACCESS(1) ",%%xmm2 \n" + MEMOPREG(movq,0x00,1,2,1,xmm3) // movq (%1,%2,1),%%xmm3 + "lea " MEMLEA(0x8,1) ",%1 \n" + "punpcklbw %%xmm3,%%xmm2 \n" + "movdqu " MEMACCESS(0) ",%%xmm0 \n" + "movdqa %%xmm2,%%xmm1 \n" + "lea " MEMLEA(0x10,0) ",%0 \n" + "punpcklbw %%xmm0,%%xmm1 \n" + "punpckhbw %%xmm0,%%xmm2 \n" + "movdqu %%xmm1," MEMACCESS(3) " \n" + "movdqu %%xmm2," MEMACCESS2(0x10,3) " \n" + "lea " MEMLEA(0x20,3) ",%3 \n" + "sub $0x10,%4 \n" + "jg 1b \n" + : "+r"(src_y), // %0 + "+r"(src_u), // %1 + "+r"(src_v), // %2 + "+r"(dst_frame), // %3 + "+rm"(width) // %4 + : + : "memory", "cc", NACL_R14 + "xmm0", "xmm1", "xmm2", "xmm3" + ); +} +#endif // HAS_I422TOUYVYROW_SSE2 + +#ifdef HAS_ARGBPOLYNOMIALROW_SSE2 +void ARGBPolynomialRow_SSE2(const uint8* src_argb, + uint8* dst_argb, const float* poly, + int width) { + asm volatile ( + "pxor %%xmm3,%%xmm3 \n" + + // 2 pixel loop. + LABELALIGN + "1: \n" + "movq " MEMACCESS(0) ",%%xmm0 \n" + "lea " MEMLEA(0x8,0) ",%0 \n" + "punpcklbw %%xmm3,%%xmm0 \n" + "movdqa %%xmm0,%%xmm4 \n" + "punpcklwd %%xmm3,%%xmm0 \n" + "punpckhwd %%xmm3,%%xmm4 \n" + "cvtdq2ps %%xmm0,%%xmm0 \n" + "cvtdq2ps %%xmm4,%%xmm4 \n" + "movdqa %%xmm0,%%xmm1 \n" + "movdqa %%xmm4,%%xmm5 \n" + "mulps " MEMACCESS2(0x10,3) ",%%xmm0 \n" + "mulps " MEMACCESS2(0x10,3) ",%%xmm4 \n" + "addps " MEMACCESS(3) ",%%xmm0 \n" + "addps " MEMACCESS(3) ",%%xmm4 \n" + "movdqa %%xmm1,%%xmm2 \n" + "movdqa %%xmm5,%%xmm6 \n" + "mulps %%xmm1,%%xmm2 \n" + "mulps %%xmm5,%%xmm6 \n" + "mulps %%xmm2,%%xmm1 \n" + "mulps %%xmm6,%%xmm5 \n" + "mulps " MEMACCESS2(0x20,3) ",%%xmm2 \n" + "mulps " MEMACCESS2(0x20,3) ",%%xmm6 \n" + "mulps " MEMACCESS2(0x30,3) ",%%xmm1 \n" + "mulps " MEMACCESS2(0x30,3) ",%%xmm5 \n" + "addps %%xmm2,%%xmm0 \n" + "addps %%xmm6,%%xmm4 \n" + "addps %%xmm1,%%xmm0 \n" + "addps %%xmm5,%%xmm4 \n" + "cvttps2dq %%xmm0,%%xmm0 \n" + "cvttps2dq %%xmm4,%%xmm4 \n" + "packuswb %%xmm4,%%xmm0 \n" + "packuswb %%xmm0,%%xmm0 \n" + "movq %%xmm0," MEMACCESS(1) " \n" + "lea " MEMLEA(0x8,1) ",%1 \n" + "sub $0x2,%2 \n" + "jg 1b \n" + : "+r"(src_argb), // %0 + "+r"(dst_argb), // %1 + "+r"(width) // %2 + : "r"(poly) // %3 + : "memory", "cc" + , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6" + ); +} +#endif // HAS_ARGBPOLYNOMIALROW_SSE2 + +#ifdef HAS_ARGBPOLYNOMIALROW_AVX2 +void ARGBPolynomialRow_AVX2(const uint8* src_argb, + uint8* dst_argb, const float* poly, + int width) { + asm volatile ( + "vbroadcastf128 " MEMACCESS(3) ",%%ymm4 \n" + "vbroadcastf128 " MEMACCESS2(0x10,3) ",%%ymm5 \n" + "vbroadcastf128 " MEMACCESS2(0x20,3) ",%%ymm6 \n" + "vbroadcastf128 " MEMACCESS2(0x30,3) ",%%ymm7 \n" + + // 2 pixel loop. + LABELALIGN + "1: \n" + "vpmovzxbd " MEMACCESS(0) ",%%ymm0 \n" // 2 ARGB pixels + "lea " MEMLEA(0x8,0) ",%0 \n" + "vcvtdq2ps %%ymm0,%%ymm0 \n" // X 8 floats + "vmulps %%ymm0,%%ymm0,%%ymm2 \n" // X * X + "vmulps %%ymm7,%%ymm0,%%ymm3 \n" // C3 * X + "vfmadd132ps %%ymm5,%%ymm4,%%ymm0 \n" // result = C0 + C1 * X + "vfmadd231ps %%ymm6,%%ymm2,%%ymm0 \n" // result += C2 * X * X + "vfmadd231ps %%ymm3,%%ymm2,%%ymm0 \n" // result += C3 * X * X * X + "vcvttps2dq %%ymm0,%%ymm0 \n" + "vpackusdw %%ymm0,%%ymm0,%%ymm0 \n" + "vpermq $0xd8,%%ymm0,%%ymm0 \n" + "vpackuswb %%xmm0,%%xmm0,%%xmm0 \n" + "vmovq %%xmm0," MEMACCESS(1) " \n" + "lea " MEMLEA(0x8,1) ",%1 \n" + "sub $0x2,%2 \n" + "jg 1b \n" + "vzeroupper \n" + : "+r"(src_argb), // %0 + "+r"(dst_argb), // %1 + "+r"(width) // %2 + : "r"(poly) // %3 + : "memory", "cc", + "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7" + ); +} +#endif // HAS_ARGBPOLYNOMIALROW_AVX2 + +#ifdef HAS_ARGBCOLORTABLEROW_X86 +// Tranform ARGB pixels with color table. +void ARGBColorTableRow_X86(uint8* dst_argb, const uint8* table_argb, + int width) { + uintptr_t pixel_temp = 0u; + asm volatile ( + // 1 pixel loop. + LABELALIGN + "1: \n" + "movzb " MEMACCESS(0) ",%1 \n" + "lea " MEMLEA(0x4,0) ",%0 \n" + MEMOPARG(movzb,0x00,3,1,4,1) " \n" // movzb (%3,%1,4),%1 + "mov %b1," MEMACCESS2(-0x4,0) " \n" + "movzb " MEMACCESS2(-0x3,0) ",%1 \n" + MEMOPARG(movzb,0x01,3,1,4,1) " \n" // movzb 0x1(%3,%1,4),%1 + "mov %b1," MEMACCESS2(-0x3,0) " \n" + "movzb " MEMACCESS2(-0x2,0) ",%1 \n" + MEMOPARG(movzb,0x02,3,1,4,1) " \n" // movzb 0x2(%3,%1,4),%1 + "mov %b1," MEMACCESS2(-0x2,0) " \n" + "movzb " MEMACCESS2(-0x1,0) ",%1 \n" + MEMOPARG(movzb,0x03,3,1,4,1) " \n" // movzb 0x3(%3,%1,4),%1 + "mov %b1," MEMACCESS2(-0x1,0) " \n" + "dec %2 \n" + "jg 1b \n" + : "+r"(dst_argb), // %0 + "+d"(pixel_temp), // %1 + "+r"(width) // %2 + : "r"(table_argb) // %3 + : "memory", "cc"); +} +#endif // HAS_ARGBCOLORTABLEROW_X86 + +#ifdef HAS_RGBCOLORTABLEROW_X86 +// Tranform RGB pixels with color table. +void RGBColorTableRow_X86(uint8* dst_argb, const uint8* table_argb, int width) { + uintptr_t pixel_temp = 0u; + asm volatile ( + // 1 pixel loop. + LABELALIGN + "1: \n" + "movzb " MEMACCESS(0) ",%1 \n" + "lea " MEMLEA(0x4,0) ",%0 \n" + MEMOPARG(movzb,0x00,3,1,4,1) " \n" // movzb (%3,%1,4),%1 + "mov %b1," MEMACCESS2(-0x4,0) " \n" + "movzb " MEMACCESS2(-0x3,0) ",%1 \n" + MEMOPARG(movzb,0x01,3,1,4,1) " \n" // movzb 0x1(%3,%1,4),%1 + "mov %b1," MEMACCESS2(-0x3,0) " \n" + "movzb " MEMACCESS2(-0x2,0) ",%1 \n" + MEMOPARG(movzb,0x02,3,1,4,1) " \n" // movzb 0x2(%3,%1,4),%1 + "mov %b1," MEMACCESS2(-0x2,0) " \n" + "dec %2 \n" + "jg 1b \n" + : "+r"(dst_argb), // %0 + "+d"(pixel_temp), // %1 + "+r"(width) // %2 + : "r"(table_argb) // %3 + : "memory", "cc"); +} +#endif // HAS_RGBCOLORTABLEROW_X86 + +#ifdef HAS_ARGBLUMACOLORTABLEROW_SSSE3 +// Tranform RGB pixels with luma table. +void ARGBLumaColorTableRow_SSSE3(const uint8* src_argb, uint8* dst_argb, + int width, + const uint8* luma, uint32 lumacoeff) { + uintptr_t pixel_temp = 0u; + uintptr_t table_temp = 0u; + asm volatile ( + "movd %6,%%xmm3 \n" + "pshufd $0x0,%%xmm3,%%xmm3 \n" + "pcmpeqb %%xmm4,%%xmm4 \n" + "psllw $0x8,%%xmm4 \n" + "pxor %%xmm5,%%xmm5 \n" + + // 4 pixel loop. + LABELALIGN + "1: \n" + "movdqu " MEMACCESS(2) ",%%xmm0 \n" + "pmaddubsw %%xmm3,%%xmm0 \n" + "phaddw %%xmm0,%%xmm0 \n" + "pand %%xmm4,%%xmm0 \n" + "punpcklwd %%xmm5,%%xmm0 \n" + "movd %%xmm0,%k1 \n" // 32 bit offset + "add %5,%1 \n" + "pshufd $0x39,%%xmm0,%%xmm0 \n" + + "movzb " MEMACCESS(2) ",%0 \n" + MEMOPARG(movzb,0x00,1,0,1,0) " \n" // movzb (%1,%0,1),%0 + "mov %b0," MEMACCESS(3) " \n" + "movzb " MEMACCESS2(0x1,2) ",%0 \n" + MEMOPARG(movzb,0x00,1,0,1,0) " \n" // movzb (%1,%0,1),%0 + "mov %b0," MEMACCESS2(0x1,3) " \n" + "movzb " MEMACCESS2(0x2,2) ",%0 \n" + MEMOPARG(movzb,0x00,1,0,1,0) " \n" // movzb (%1,%0,1),%0 + "mov %b0," MEMACCESS2(0x2,3) " \n" + "movzb " MEMACCESS2(0x3,2) ",%0 \n" + "mov %b0," MEMACCESS2(0x3,3) " \n" + + "movd %%xmm0,%k1 \n" // 32 bit offset + "add %5,%1 \n" + "pshufd $0x39,%%xmm0,%%xmm0 \n" + + "movzb " MEMACCESS2(0x4,2) ",%0 \n" + MEMOPARG(movzb,0x00,1,0,1,0) " \n" // movzb (%1,%0,1),%0 + "mov %b0," MEMACCESS2(0x4,3) " \n" + "movzb " MEMACCESS2(0x5,2) ",%0 \n" + MEMOPARG(movzb,0x00,1,0,1,0) " \n" // movzb (%1,%0,1),%0 + "mov %b0," MEMACCESS2(0x5,3) " \n" + "movzb " MEMACCESS2(0x6,2) ",%0 \n" + MEMOPARG(movzb,0x00,1,0,1,0) " \n" // movzb (%1,%0,1),%0 + "mov %b0," MEMACCESS2(0x6,3) " \n" + "movzb " MEMACCESS2(0x7,2) ",%0 \n" + "mov %b0," MEMACCESS2(0x7,3) " \n" + + "movd %%xmm0,%k1 \n" // 32 bit offset + "add %5,%1 \n" + "pshufd $0x39,%%xmm0,%%xmm0 \n" + + "movzb " MEMACCESS2(0x8,2) ",%0 \n" + MEMOPARG(movzb,0x00,1,0,1,0) " \n" // movzb (%1,%0,1),%0 + "mov %b0," MEMACCESS2(0x8,3) " \n" + "movzb " MEMACCESS2(0x9,2) ",%0 \n" + MEMOPARG(movzb,0x00,1,0,1,0) " \n" // movzb (%1,%0,1),%0 + "mov %b0," MEMACCESS2(0x9,3) " \n" + "movzb " MEMACCESS2(0xa,2) ",%0 \n" + MEMOPARG(movzb,0x00,1,0,1,0) " \n" // movzb (%1,%0,1),%0 + "mov %b0," MEMACCESS2(0xa,3) " \n" + "movzb " MEMACCESS2(0xb,2) ",%0 \n" + "mov %b0," MEMACCESS2(0xb,3) " \n" + + "movd %%xmm0,%k1 \n" // 32 bit offset + "add %5,%1 \n" + + "movzb " MEMACCESS2(0xc,2) ",%0 \n" + MEMOPARG(movzb,0x00,1,0,1,0) " \n" // movzb (%1,%0,1),%0 + "mov %b0," MEMACCESS2(0xc,3) " \n" + "movzb " MEMACCESS2(0xd,2) ",%0 \n" + MEMOPARG(movzb,0x00,1,0,1,0) " \n" // movzb (%1,%0,1),%0 + "mov %b0," MEMACCESS2(0xd,3) " \n" + "movzb " MEMACCESS2(0xe,2) ",%0 \n" + MEMOPARG(movzb,0x00,1,0,1,0) " \n" // movzb (%1,%0,1),%0 + "mov %b0," MEMACCESS2(0xe,3) " \n" + "movzb " MEMACCESS2(0xf,2) ",%0 \n" + "mov %b0," MEMACCESS2(0xf,3) " \n" + "lea " MEMLEA(0x10,2) ",%2 \n" + "lea " MEMLEA(0x10,3) ",%3 \n" + "sub $0x4,%4 \n" + "jg 1b \n" + : "+d"(pixel_temp), // %0 + "+a"(table_temp), // %1 + "+r"(src_argb), // %2 + "+r"(dst_argb), // %3 + "+rm"(width) // %4 + : "r"(luma), // %5 + "rm"(lumacoeff) // %6 + : "memory", "cc", "xmm0", "xmm3", "xmm4", "xmm5" + ); +} +#endif // HAS_ARGBLUMACOLORTABLEROW_SSSE3 + +#endif // defined(__x86_64__) || defined(__i386__) + +#ifdef __cplusplus +} // extern "C" +} // namespace libyuv +#endif diff --git a/media/libaom/src/third_party/libyuv/source/row_mips.cc b/media/libaom/src/third_party/libyuv/source/row_mips.cc new file mode 100644 index 000000000..cfc9ffe03 --- /dev/null +++ b/media/libaom/src/third_party/libyuv/source/row_mips.cc @@ -0,0 +1,911 @@ +/* + * Copyright (c) 2012 The LibYuv project authors. All Rights Reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#include "libyuv/row.h" + +#ifdef __cplusplus +namespace libyuv { +extern "C" { +#endif + +// The following are available on Mips platforms: +#if !defined(LIBYUV_DISABLE_MIPS) && defined(__mips__) && \ + (_MIPS_SIM == _MIPS_SIM_ABI32) + +#ifdef HAS_COPYROW_MIPS +void CopyRow_MIPS(const uint8* src, uint8* dst, int count) { + __asm__ __volatile__ ( + ".set noreorder \n" + ".set noat \n" + "slti $at, %[count], 8 \n" + "bne $at ,$zero, $last8 \n" + "xor $t8, %[src], %[dst] \n" + "andi $t8, $t8, 0x3 \n" + + "bne $t8, $zero, unaligned \n" + "negu $a3, %[dst] \n" + // make dst/src aligned + "andi $a3, $a3, 0x3 \n" + "beq $a3, $zero, $chk16w \n" + // word-aligned now count is the remining bytes count + "subu %[count], %[count], $a3 \n" + + "lwr $t8, 0(%[src]) \n" + "addu %[src], %[src], $a3 \n" + "swr $t8, 0(%[dst]) \n" + "addu %[dst], %[dst], $a3 \n" + + // Now the dst/src are mutually word-aligned with word-aligned addresses + "$chk16w: \n" + "andi $t8, %[count], 0x3f \n" // whole 64-B chunks? + // t8 is the byte count after 64-byte chunks + "beq %[count], $t8, chk8w \n" + // There will be at most 1 32-byte chunk after it + "subu $a3, %[count], $t8 \n" // the reminder + // Here a3 counts bytes in 16w chunks + "addu $a3, %[dst], $a3 \n" + // Now a3 is the final dst after 64-byte chunks + "addu $t0, %[dst], %[count] \n" + // t0 is the "past the end" address + + // When in the loop we exercise "pref 30,x(a1)", the a1+x should not be past + // the "t0-32" address + // This means: for x=128 the last "safe" a1 address is "t0-160" + // Alternatively, for x=64 the last "safe" a1 address is "t0-96" + // we will use "pref 30,128(a1)", so "t0-160" is the limit + "subu $t9, $t0, 160 \n" + // t9 is the "last safe pref 30,128(a1)" address + "pref 0, 0(%[src]) \n" // first line of src + "pref 0, 32(%[src]) \n" // second line of src + "pref 0, 64(%[src]) \n" + "pref 30, 32(%[dst]) \n" + // In case the a1 > t9 don't use "pref 30" at all + "sgtu $v1, %[dst], $t9 \n" + "bgtz $v1, $loop16w \n" + "nop \n" + // otherwise, start with using pref30 + "pref 30, 64(%[dst]) \n" + "$loop16w: \n" + "pref 0, 96(%[src]) \n" + "lw $t0, 0(%[src]) \n" + "bgtz $v1, $skip_pref30_96 \n" // skip + "lw $t1, 4(%[src]) \n" + "pref 30, 96(%[dst]) \n" // continue + "$skip_pref30_96: \n" + "lw $t2, 8(%[src]) \n" + "lw $t3, 12(%[src]) \n" + "lw $t4, 16(%[src]) \n" + "lw $t5, 20(%[src]) \n" + "lw $t6, 24(%[src]) \n" + "lw $t7, 28(%[src]) \n" + "pref 0, 128(%[src]) \n" + // bring the next lines of src, addr 128 + "sw $t0, 0(%[dst]) \n" + "sw $t1, 4(%[dst]) \n" + "sw $t2, 8(%[dst]) \n" + "sw $t3, 12(%[dst]) \n" + "sw $t4, 16(%[dst]) \n" + "sw $t5, 20(%[dst]) \n" + "sw $t6, 24(%[dst]) \n" + "sw $t7, 28(%[dst]) \n" + "lw $t0, 32(%[src]) \n" + "bgtz $v1, $skip_pref30_128 \n" // skip pref 30,128(a1) + "lw $t1, 36(%[src]) \n" + "pref 30, 128(%[dst]) \n" // set dest, addr 128 + "$skip_pref30_128: \n" + "lw $t2, 40(%[src]) \n" + "lw $t3, 44(%[src]) \n" + "lw $t4, 48(%[src]) \n" + "lw $t5, 52(%[src]) \n" + "lw $t6, 56(%[src]) \n" + "lw $t7, 60(%[src]) \n" + "pref 0, 160(%[src]) \n" + // bring the next lines of src, addr 160 + "sw $t0, 32(%[dst]) \n" + "sw $t1, 36(%[dst]) \n" + "sw $t2, 40(%[dst]) \n" + "sw $t3, 44(%[dst]) \n" + "sw $t4, 48(%[dst]) \n" + "sw $t5, 52(%[dst]) \n" + "sw $t6, 56(%[dst]) \n" + "sw $t7, 60(%[dst]) \n" + + "addiu %[dst], %[dst], 64 \n" // adding 64 to dest + "sgtu $v1, %[dst], $t9 \n" + "bne %[dst], $a3, $loop16w \n" + " addiu %[src], %[src], 64 \n" // adding 64 to src + "move %[count], $t8 \n" + + // Here we have src and dest word-aligned but less than 64-bytes to go + + "chk8w: \n" + "pref 0, 0x0(%[src]) \n" + "andi $t8, %[count], 0x1f \n" // 32-byte chunk? + // the t8 is the reminder count past 32-bytes + "beq %[count], $t8, chk1w \n" + // count=t8,no 32-byte chunk + " nop \n" + + "lw $t0, 0(%[src]) \n" + "lw $t1, 4(%[src]) \n" + "lw $t2, 8(%[src]) \n" + "lw $t3, 12(%[src]) \n" + "lw $t4, 16(%[src]) \n" + "lw $t5, 20(%[src]) \n" + "lw $t6, 24(%[src]) \n" + "lw $t7, 28(%[src]) \n" + "addiu %[src], %[src], 32 \n" + + "sw $t0, 0(%[dst]) \n" + "sw $t1, 4(%[dst]) \n" + "sw $t2, 8(%[dst]) \n" + "sw $t3, 12(%[dst]) \n" + "sw $t4, 16(%[dst]) \n" + "sw $t5, 20(%[dst]) \n" + "sw $t6, 24(%[dst]) \n" + "sw $t7, 28(%[dst]) \n" + "addiu %[dst], %[dst], 32 \n" + + "chk1w: \n" + "andi %[count], $t8, 0x3 \n" + // now count is the reminder past 1w chunks + "beq %[count], $t8, $last8 \n" + " subu $a3, $t8, %[count] \n" + // a3 is count of bytes in 1w chunks + "addu $a3, %[dst], $a3 \n" + // now a3 is the dst address past the 1w chunks + // copying in words (4-byte chunks) + "$wordCopy_loop: \n" + "lw $t3, 0(%[src]) \n" + // the first t3 may be equal t0 ... optimize? + "addiu %[src], %[src],4 \n" + "addiu %[dst], %[dst],4 \n" + "bne %[dst], $a3,$wordCopy_loop \n" + " sw $t3, -4(%[dst]) \n" + + // For the last (<8) bytes + "$last8: \n" + "blez %[count], leave \n" + " addu $a3, %[dst], %[count] \n" // a3 -last dst address + "$last8loop: \n" + "lb $v1, 0(%[src]) \n" + "addiu %[src], %[src], 1 \n" + "addiu %[dst], %[dst], 1 \n" + "bne %[dst], $a3, $last8loop \n" + " sb $v1, -1(%[dst]) \n" + + "leave: \n" + " j $ra \n" + " nop \n" + + // + // UNALIGNED case + // + + "unaligned: \n" + // got here with a3="negu a1" + "andi $a3, $a3, 0x3 \n" // a1 is word aligned? + "beqz $a3, $ua_chk16w \n" + " subu %[count], %[count], $a3 \n" + // bytes left after initial a3 bytes + "lwr $v1, 0(%[src]) \n" + "lwl $v1, 3(%[src]) \n" + "addu %[src], %[src], $a3 \n" // a3 may be 1, 2 or 3 + "swr $v1, 0(%[dst]) \n" + "addu %[dst], %[dst], $a3 \n" + // below the dst will be word aligned (NOTE1) + "$ua_chk16w: \n" + "andi $t8, %[count], 0x3f \n" // whole 64-B chunks? + // t8 is the byte count after 64-byte chunks + "beq %[count], $t8, ua_chk8w \n" + // if a2==t8, no 64-byte chunks + // There will be at most 1 32-byte chunk after it + "subu $a3, %[count], $t8 \n" // the reminder + // Here a3 counts bytes in 16w chunks + "addu $a3, %[dst], $a3 \n" + // Now a3 is the final dst after 64-byte chunks + "addu $t0, %[dst], %[count] \n" // t0 "past the end" + "subu $t9, $t0, 160 \n" + // t9 is the "last safe pref 30,128(a1)" address + "pref 0, 0(%[src]) \n" // first line of src + "pref 0, 32(%[src]) \n" // second line addr 32 + "pref 0, 64(%[src]) \n" + "pref 30, 32(%[dst]) \n" + // safe, as we have at least 64 bytes ahead + // In case the a1 > t9 don't use "pref 30" at all + "sgtu $v1, %[dst], $t9 \n" + "bgtz $v1, $ua_loop16w \n" + // skip "pref 30,64(a1)" for too short arrays + " nop \n" + // otherwise, start with using pref30 + "pref 30, 64(%[dst]) \n" + "$ua_loop16w: \n" + "pref 0, 96(%[src]) \n" + "lwr $t0, 0(%[src]) \n" + "lwl $t0, 3(%[src]) \n" + "lwr $t1, 4(%[src]) \n" + "bgtz $v1, $ua_skip_pref30_96 \n" + " lwl $t1, 7(%[src]) \n" + "pref 30, 96(%[dst]) \n" + // continue setting up the dest, addr 96 + "$ua_skip_pref30_96: \n" + "lwr $t2, 8(%[src]) \n" + "lwl $t2, 11(%[src]) \n" + "lwr $t3, 12(%[src]) \n" + "lwl $t3, 15(%[src]) \n" + "lwr $t4, 16(%[src]) \n" + "lwl $t4, 19(%[src]) \n" + "lwr $t5, 20(%[src]) \n" + "lwl $t5, 23(%[src]) \n" + "lwr $t6, 24(%[src]) \n" + "lwl $t6, 27(%[src]) \n" + "lwr $t7, 28(%[src]) \n" + "lwl $t7, 31(%[src]) \n" + "pref 0, 128(%[src]) \n" + // bring the next lines of src, addr 128 + "sw $t0, 0(%[dst]) \n" + "sw $t1, 4(%[dst]) \n" + "sw $t2, 8(%[dst]) \n" + "sw $t3, 12(%[dst]) \n" + "sw $t4, 16(%[dst]) \n" + "sw $t5, 20(%[dst]) \n" + "sw $t6, 24(%[dst]) \n" + "sw $t7, 28(%[dst]) \n" + "lwr $t0, 32(%[src]) \n" + "lwl $t0, 35(%[src]) \n" + "lwr $t1, 36(%[src]) \n" + "bgtz $v1, ua_skip_pref30_128 \n" + " lwl $t1, 39(%[src]) \n" + "pref 30, 128(%[dst]) \n" + // continue setting up the dest, addr 128 + "ua_skip_pref30_128: \n" + + "lwr $t2, 40(%[src]) \n" + "lwl $t2, 43(%[src]) \n" + "lwr $t3, 44(%[src]) \n" + "lwl $t3, 47(%[src]) \n" + "lwr $t4, 48(%[src]) \n" + "lwl $t4, 51(%[src]) \n" + "lwr $t5, 52(%[src]) \n" + "lwl $t5, 55(%[src]) \n" + "lwr $t6, 56(%[src]) \n" + "lwl $t6, 59(%[src]) \n" + "lwr $t7, 60(%[src]) \n" + "lwl $t7, 63(%[src]) \n" + "pref 0, 160(%[src]) \n" + // bring the next lines of src, addr 160 + "sw $t0, 32(%[dst]) \n" + "sw $t1, 36(%[dst]) \n" + "sw $t2, 40(%[dst]) \n" + "sw $t3, 44(%[dst]) \n" + "sw $t4, 48(%[dst]) \n" + "sw $t5, 52(%[dst]) \n" + "sw $t6, 56(%[dst]) \n" + "sw $t7, 60(%[dst]) \n" + + "addiu %[dst],%[dst],64 \n" // adding 64 to dest + "sgtu $v1,%[dst],$t9 \n" + "bne %[dst],$a3,$ua_loop16w \n" + " addiu %[src],%[src],64 \n" // adding 64 to src + "move %[count],$t8 \n" + + // Here we have src and dest word-aligned but less than 64-bytes to go + + "ua_chk8w: \n" + "pref 0, 0x0(%[src]) \n" + "andi $t8, %[count], 0x1f \n" // 32-byte chunk? + // the t8 is the reminder count + "beq %[count], $t8, $ua_chk1w \n" + // when count==t8, no 32-byte chunk + + "lwr $t0, 0(%[src]) \n" + "lwl $t0, 3(%[src]) \n" + "lwr $t1, 4(%[src]) \n" + "lwl $t1, 7(%[src]) \n" + "lwr $t2, 8(%[src]) \n" + "lwl $t2, 11(%[src]) \n" + "lwr $t3, 12(%[src]) \n" + "lwl $t3, 15(%[src]) \n" + "lwr $t4, 16(%[src]) \n" + "lwl $t4, 19(%[src]) \n" + "lwr $t5, 20(%[src]) \n" + "lwl $t5, 23(%[src]) \n" + "lwr $t6, 24(%[src]) \n" + "lwl $t6, 27(%[src]) \n" + "lwr $t7, 28(%[src]) \n" + "lwl $t7, 31(%[src]) \n" + "addiu %[src], %[src], 32 \n" + + "sw $t0, 0(%[dst]) \n" + "sw $t1, 4(%[dst]) \n" + "sw $t2, 8(%[dst]) \n" + "sw $t3, 12(%[dst]) \n" + "sw $t4, 16(%[dst]) \n" + "sw $t5, 20(%[dst]) \n" + "sw $t6, 24(%[dst]) \n" + "sw $t7, 28(%[dst]) \n" + "addiu %[dst], %[dst], 32 \n" + + "$ua_chk1w: \n" + "andi %[count], $t8, 0x3 \n" + // now count is the reminder past 1w chunks + "beq %[count], $t8, ua_smallCopy \n" + "subu $a3, $t8, %[count] \n" + // a3 is count of bytes in 1w chunks + "addu $a3, %[dst], $a3 \n" + // now a3 is the dst address past the 1w chunks + + // copying in words (4-byte chunks) + "$ua_wordCopy_loop: \n" + "lwr $v1, 0(%[src]) \n" + "lwl $v1, 3(%[src]) \n" + "addiu %[src], %[src], 4 \n" + "addiu %[dst], %[dst], 4 \n" + // note: dst=a1 is word aligned here, see NOTE1 + "bne %[dst], $a3, $ua_wordCopy_loop \n" + " sw $v1,-4(%[dst]) \n" + + // Now less than 4 bytes (value in count) left to copy + "ua_smallCopy: \n" + "beqz %[count], leave \n" + " addu $a3, %[dst], %[count] \n" // a3 = last dst address + "$ua_smallCopy_loop: \n" + "lb $v1, 0(%[src]) \n" + "addiu %[src], %[src], 1 \n" + "addiu %[dst], %[dst], 1 \n" + "bne %[dst],$a3,$ua_smallCopy_loop \n" + " sb $v1, -1(%[dst]) \n" + + "j $ra \n" + " nop \n" + ".set at \n" + ".set reorder \n" + : [dst] "+r" (dst), [src] "+r" (src) + : [count] "r" (count) + : "t0", "t1", "t2", "t3", "t4", "t5", "t6", "t7", + "t8", "t9", "a3", "v1", "at" + ); +} +#endif // HAS_COPYROW_MIPS + +// MIPS DSPR2 functions +#if !defined(LIBYUV_DISABLE_MIPS) && defined(__mips_dsp) && \ + (__mips_dsp_rev >= 2) && \ + (_MIPS_SIM == _MIPS_SIM_ABI32) && (__mips_isa_rev < 6) + +void SplitUVRow_MIPS_DSPR2(const uint8* src_uv, uint8* dst_u, uint8* dst_v, + int width) { + __asm__ __volatile__ ( + ".set push \n" + ".set noreorder \n" + "srl $t4, %[width], 4 \n" // multiplies of 16 + "blez $t4, 2f \n" + " andi %[width], %[width], 0xf \n" // residual + + ".p2align 2 \n" + "1: \n" + "addiu $t4, $t4, -1 \n" + "lw $t0, 0(%[src_uv]) \n" // V1 | U1 | V0 | U0 + "lw $t1, 4(%[src_uv]) \n" // V3 | U3 | V2 | U2 + "lw $t2, 8(%[src_uv]) \n" // V5 | U5 | V4 | U4 + "lw $t3, 12(%[src_uv]) \n" // V7 | U7 | V6 | U6 + "lw $t5, 16(%[src_uv]) \n" // V9 | U9 | V8 | U8 + "lw $t6, 20(%[src_uv]) \n" // V11 | U11 | V10 | U10 + "lw $t7, 24(%[src_uv]) \n" // V13 | U13 | V12 | U12 + "lw $t8, 28(%[src_uv]) \n" // V15 | U15 | V14 | U14 + "addiu %[src_uv], %[src_uv], 32 \n" + "precrq.qb.ph $t9, $t1, $t0 \n" // V3 | V2 | V1 | V0 + "precr.qb.ph $t0, $t1, $t0 \n" // U3 | U2 | U1 | U0 + "precrq.qb.ph $t1, $t3, $t2 \n" // V7 | V6 | V5 | V4 + "precr.qb.ph $t2, $t3, $t2 \n" // U7 | U6 | U5 | U4 + "precrq.qb.ph $t3, $t6, $t5 \n" // V11 | V10 | V9 | V8 + "precr.qb.ph $t5, $t6, $t5 \n" // U11 | U10 | U9 | U8 + "precrq.qb.ph $t6, $t8, $t7 \n" // V15 | V14 | V13 | V12 + "precr.qb.ph $t7, $t8, $t7 \n" // U15 | U14 | U13 | U12 + "sw $t9, 0(%[dst_v]) \n" + "sw $t0, 0(%[dst_u]) \n" + "sw $t1, 4(%[dst_v]) \n" + "sw $t2, 4(%[dst_u]) \n" + "sw $t3, 8(%[dst_v]) \n" + "sw $t5, 8(%[dst_u]) \n" + "sw $t6, 12(%[dst_v]) \n" + "sw $t7, 12(%[dst_u]) \n" + "addiu %[dst_v], %[dst_v], 16 \n" + "bgtz $t4, 1b \n" + " addiu %[dst_u], %[dst_u], 16 \n" + + "beqz %[width], 3f \n" + " nop \n" + + "2: \n" + "lbu $t0, 0(%[src_uv]) \n" + "lbu $t1, 1(%[src_uv]) \n" + "addiu %[src_uv], %[src_uv], 2 \n" + "addiu %[width], %[width], -1 \n" + "sb $t0, 0(%[dst_u]) \n" + "sb $t1, 0(%[dst_v]) \n" + "addiu %[dst_u], %[dst_u], 1 \n" + "bgtz %[width], 2b \n" + " addiu %[dst_v], %[dst_v], 1 \n" + + "3: \n" + ".set pop \n" + : [src_uv] "+r" (src_uv), + [width] "+r" (width), + [dst_u] "+r" (dst_u), + [dst_v] "+r" (dst_v) + : + : "t0", "t1", "t2", "t3", + "t4", "t5", "t6", "t7", "t8", "t9" + ); +} + +void MirrorRow_MIPS_DSPR2(const uint8* src, uint8* dst, int width) { + __asm__ __volatile__ ( + ".set push \n" + ".set noreorder \n" + + "srl $t4, %[width], 4 \n" // multiplies of 16 + "andi $t5, %[width], 0xf \n" + "blez $t4, 2f \n" + " addu %[src], %[src], %[width] \n" // src += width + + ".p2align 2 \n" + "1: \n" + "lw $t0, -16(%[src]) \n" // |3|2|1|0| + "lw $t1, -12(%[src]) \n" // |7|6|5|4| + "lw $t2, -8(%[src]) \n" // |11|10|9|8| + "lw $t3, -4(%[src]) \n" // |15|14|13|12| + "wsbh $t0, $t0 \n" // |2|3|0|1| + "wsbh $t1, $t1 \n" // |6|7|4|5| + "wsbh $t2, $t2 \n" // |10|11|8|9| + "wsbh $t3, $t3 \n" // |14|15|12|13| + "rotr $t0, $t0, 16 \n" // |0|1|2|3| + "rotr $t1, $t1, 16 \n" // |4|5|6|7| + "rotr $t2, $t2, 16 \n" // |8|9|10|11| + "rotr $t3, $t3, 16 \n" // |12|13|14|15| + "addiu %[src], %[src], -16 \n" + "addiu $t4, $t4, -1 \n" + "sw $t3, 0(%[dst]) \n" // |15|14|13|12| + "sw $t2, 4(%[dst]) \n" // |11|10|9|8| + "sw $t1, 8(%[dst]) \n" // |7|6|5|4| + "sw $t0, 12(%[dst]) \n" // |3|2|1|0| + "bgtz $t4, 1b \n" + " addiu %[dst], %[dst], 16 \n" + "beqz $t5, 3f \n" + " nop \n" + + "2: \n" + "lbu $t0, -1(%[src]) \n" + "addiu $t5, $t5, -1 \n" + "addiu %[src], %[src], -1 \n" + "sb $t0, 0(%[dst]) \n" + "bgez $t5, 2b \n" + " addiu %[dst], %[dst], 1 \n" + + "3: \n" + ".set pop \n" + : [src] "+r" (src), [dst] "+r" (dst) + : [width] "r" (width) + : "t0", "t1", "t2", "t3", "t4", "t5" + ); +} + +void MirrorUVRow_MIPS_DSPR2(const uint8* src_uv, uint8* dst_u, uint8* dst_v, + int width) { + int x = 0; + int y = 0; + __asm__ __volatile__ ( + ".set push \n" + ".set noreorder \n" + + "addu $t4, %[width], %[width] \n" + "srl %[x], %[width], 4 \n" + "andi %[y], %[width], 0xf \n" + "blez %[x], 2f \n" + " addu %[src_uv], %[src_uv], $t4 \n" + + ".p2align 2 \n" + "1: \n" + "lw $t0, -32(%[src_uv]) \n" // |3|2|1|0| + "lw $t1, -28(%[src_uv]) \n" // |7|6|5|4| + "lw $t2, -24(%[src_uv]) \n" // |11|10|9|8| + "lw $t3, -20(%[src_uv]) \n" // |15|14|13|12| + "lw $t4, -16(%[src_uv]) \n" // |19|18|17|16| + "lw $t6, -12(%[src_uv]) \n" // |23|22|21|20| + "lw $t7, -8(%[src_uv]) \n" // |27|26|25|24| + "lw $t8, -4(%[src_uv]) \n" // |31|30|29|28| + + "rotr $t0, $t0, 16 \n" // |1|0|3|2| + "rotr $t1, $t1, 16 \n" // |5|4|7|6| + "rotr $t2, $t2, 16 \n" // |9|8|11|10| + "rotr $t3, $t3, 16 \n" // |13|12|15|14| + "rotr $t4, $t4, 16 \n" // |17|16|19|18| + "rotr $t6, $t6, 16 \n" // |21|20|23|22| + "rotr $t7, $t7, 16 \n" // |25|24|27|26| + "rotr $t8, $t8, 16 \n" // |29|28|31|30| + "precr.qb.ph $t9, $t0, $t1 \n" // |0|2|4|6| + "precrq.qb.ph $t5, $t0, $t1 \n" // |1|3|5|7| + "precr.qb.ph $t0, $t2, $t3 \n" // |8|10|12|14| + "precrq.qb.ph $t1, $t2, $t3 \n" // |9|11|13|15| + "precr.qb.ph $t2, $t4, $t6 \n" // |16|18|20|22| + "precrq.qb.ph $t3, $t4, $t6 \n" // |17|19|21|23| + "precr.qb.ph $t4, $t7, $t8 \n" // |24|26|28|30| + "precrq.qb.ph $t6, $t7, $t8 \n" // |25|27|29|31| + "addiu %[src_uv], %[src_uv], -32 \n" + "addiu %[x], %[x], -1 \n" + "swr $t4, 0(%[dst_u]) \n" + "swl $t4, 3(%[dst_u]) \n" // |30|28|26|24| + "swr $t6, 0(%[dst_v]) \n" + "swl $t6, 3(%[dst_v]) \n" // |31|29|27|25| + "swr $t2, 4(%[dst_u]) \n" + "swl $t2, 7(%[dst_u]) \n" // |22|20|18|16| + "swr $t3, 4(%[dst_v]) \n" + "swl $t3, 7(%[dst_v]) \n" // |23|21|19|17| + "swr $t0, 8(%[dst_u]) \n" + "swl $t0, 11(%[dst_u]) \n" // |14|12|10|8| + "swr $t1, 8(%[dst_v]) \n" + "swl $t1, 11(%[dst_v]) \n" // |15|13|11|9| + "swr $t9, 12(%[dst_u]) \n" + "swl $t9, 15(%[dst_u]) \n" // |6|4|2|0| + "swr $t5, 12(%[dst_v]) \n" + "swl $t5, 15(%[dst_v]) \n" // |7|5|3|1| + "addiu %[dst_v], %[dst_v], 16 \n" + "bgtz %[x], 1b \n" + " addiu %[dst_u], %[dst_u], 16 \n" + "beqz %[y], 3f \n" + " nop \n" + "b 2f \n" + " nop \n" + + "2: \n" + "lbu $t0, -2(%[src_uv]) \n" + "lbu $t1, -1(%[src_uv]) \n" + "addiu %[src_uv], %[src_uv], -2 \n" + "addiu %[y], %[y], -1 \n" + "sb $t0, 0(%[dst_u]) \n" + "sb $t1, 0(%[dst_v]) \n" + "addiu %[dst_u], %[dst_u], 1 \n" + "bgtz %[y], 2b \n" + " addiu %[dst_v], %[dst_v], 1 \n" + + "3: \n" + ".set pop \n" + : [src_uv] "+r" (src_uv), + [dst_u] "+r" (dst_u), + [dst_v] "+r" (dst_v), + [x] "=&r" (x), + [y] "+r" (y) + : [width] "r" (width) + : "t0", "t1", "t2", "t3", "t4", + "t5", "t7", "t8", "t9" + ); +} + +// Convert (4 Y and 2 VU) I422 and arrange RGB values into +// t5 = | 0 | B0 | 0 | b0 | +// t4 = | 0 | B1 | 0 | b1 | +// t9 = | 0 | G0 | 0 | g0 | +// t8 = | 0 | G1 | 0 | g1 | +// t2 = | 0 | R0 | 0 | r0 | +// t1 = | 0 | R1 | 0 | r1 | +#define I422ToTransientMipsRGB \ + "lw $t0, 0(%[y_buf]) \n" \ + "lhu $t1, 0(%[u_buf]) \n" \ + "lhu $t2, 0(%[v_buf]) \n" \ + "preceu.ph.qbr $t1, $t1 \n" \ + "preceu.ph.qbr $t2, $t2 \n" \ + "preceu.ph.qbra $t3, $t0 \n" \ + "preceu.ph.qbla $t0, $t0 \n" \ + "subu.ph $t1, $t1, $s5 \n" \ + "subu.ph $t2, $t2, $s5 \n" \ + "subu.ph $t3, $t3, $s4 \n" \ + "subu.ph $t0, $t0, $s4 \n" \ + "mul.ph $t3, $t3, $s0 \n" \ + "mul.ph $t0, $t0, $s0 \n" \ + "shll.ph $t4, $t1, 0x7 \n" \ + "subu.ph $t4, $t4, $t1 \n" \ + "mul.ph $t6, $t1, $s1 \n" \ + "mul.ph $t1, $t2, $s2 \n" \ + "addq_s.ph $t5, $t4, $t3 \n" \ + "addq_s.ph $t4, $t4, $t0 \n" \ + "shra.ph $t5, $t5, 6 \n" \ + "shra.ph $t4, $t4, 6 \n" \ + "addiu %[u_buf], 2 \n" \ + "addiu %[v_buf], 2 \n" \ + "addu.ph $t6, $t6, $t1 \n" \ + "mul.ph $t1, $t2, $s3 \n" \ + "addu.ph $t9, $t6, $t3 \n" \ + "addu.ph $t8, $t6, $t0 \n" \ + "shra.ph $t9, $t9, 6 \n" \ + "shra.ph $t8, $t8, 6 \n" \ + "addu.ph $t2, $t1, $t3 \n" \ + "addu.ph $t1, $t1, $t0 \n" \ + "shra.ph $t2, $t2, 6 \n" \ + "shra.ph $t1, $t1, 6 \n" \ + "subu.ph $t5, $t5, $s5 \n" \ + "subu.ph $t4, $t4, $s5 \n" \ + "subu.ph $t9, $t9, $s5 \n" \ + "subu.ph $t8, $t8, $s5 \n" \ + "subu.ph $t2, $t2, $s5 \n" \ + "subu.ph $t1, $t1, $s5 \n" \ + "shll_s.ph $t5, $t5, 8 \n" \ + "shll_s.ph $t4, $t4, 8 \n" \ + "shll_s.ph $t9, $t9, 8 \n" \ + "shll_s.ph $t8, $t8, 8 \n" \ + "shll_s.ph $t2, $t2, 8 \n" \ + "shll_s.ph $t1, $t1, 8 \n" \ + "shra.ph $t5, $t5, 8 \n" \ + "shra.ph $t4, $t4, 8 \n" \ + "shra.ph $t9, $t9, 8 \n" \ + "shra.ph $t8, $t8, 8 \n" \ + "shra.ph $t2, $t2, 8 \n" \ + "shra.ph $t1, $t1, 8 \n" \ + "addu.ph $t5, $t5, $s5 \n" \ + "addu.ph $t4, $t4, $s5 \n" \ + "addu.ph $t9, $t9, $s5 \n" \ + "addu.ph $t8, $t8, $s5 \n" \ + "addu.ph $t2, $t2, $s5 \n" \ + "addu.ph $t1, $t1, $s5 \n" + +void I422ToARGBRow_MIPS_DSPR2(const uint8* y_buf, + const uint8* u_buf, + const uint8* v_buf, + uint8* rgb_buf, + int width) { + __asm__ __volatile__ ( + ".set push \n" + ".set noreorder \n" + "beqz %[width], 2f \n" + " repl.ph $s0, 74 \n" // |YG|YG| = |74|74| + "repl.ph $s1, -25 \n" // |UG|UG| = |-25|-25| + "repl.ph $s2, -52 \n" // |VG|VG| = |-52|-52| + "repl.ph $s3, 102 \n" // |VR|VR| = |102|102| + "repl.ph $s4, 16 \n" // |0|16|0|16| + "repl.ph $s5, 128 \n" // |128|128| // clipping + "lui $s6, 0xff00 \n" + "ori $s6, 0xff00 \n" // |ff|00|ff|00|ff| + + ".p2align 2 \n" + "1: \n" + I422ToTransientMipsRGB +// Arranging into argb format + "precr.qb.ph $t4, $t8, $t4 \n" // |G1|g1|B1|b1| + "precr.qb.ph $t5, $t9, $t5 \n" // |G0|g0|B0|b0| + "addiu %[width], -4 \n" + "precrq.qb.ph $t8, $t4, $t5 \n" // |G1|B1|G0|B0| + "precr.qb.ph $t9, $t4, $t5 \n" // |g1|b1|g0|b0| + "precr.qb.ph $t2, $t1, $t2 \n" // |R1|r1|R0|r0| + + "addiu %[y_buf], 4 \n" + "preceu.ph.qbla $t1, $t2 \n" // |0 |R1|0 |R0| + "preceu.ph.qbra $t2, $t2 \n" // |0 |r1|0 |r0| + "or $t1, $t1, $s6 \n" // |ff|R1|ff|R0| + "or $t2, $t2, $s6 \n" // |ff|r1|ff|r0| + "precrq.ph.w $t0, $t2, $t9 \n" // |ff|r1|g1|b1| + "precrq.ph.w $t3, $t1, $t8 \n" // |ff|R1|G1|B1| + "sll $t9, $t9, 16 \n" + "sll $t8, $t8, 16 \n" + "packrl.ph $t2, $t2, $t9 \n" // |ff|r0|g0|b0| + "packrl.ph $t1, $t1, $t8 \n" // |ff|R0|G0|B0| +// Store results. + "sw $t2, 0(%[rgb_buf]) \n" + "sw $t0, 4(%[rgb_buf]) \n" + "sw $t1, 8(%[rgb_buf]) \n" + "sw $t3, 12(%[rgb_buf]) \n" + "bnez %[width], 1b \n" + " addiu %[rgb_buf], 16 \n" + "2: \n" + ".set pop \n" + :[y_buf] "+r" (y_buf), + [u_buf] "+r" (u_buf), + [v_buf] "+r" (v_buf), + [width] "+r" (width), + [rgb_buf] "+r" (rgb_buf) + : + : "t0", "t1", "t2", "t3", "t4", "t5", + "t6", "t7", "t8", "t9", + "s0", "s1", "s2", "s3", + "s4", "s5", "s6" + ); +} + +void I422ToABGRRow_MIPS_DSPR2(const uint8* y_buf, + const uint8* u_buf, + const uint8* v_buf, + uint8* rgb_buf, + int width) { + __asm__ __volatile__ ( + ".set push \n" + ".set noreorder \n" + "beqz %[width], 2f \n" + " repl.ph $s0, 74 \n" // |YG|YG| = |74|74| + "repl.ph $s1, -25 \n" // |UG|UG| = |-25|-25| + "repl.ph $s2, -52 \n" // |VG|VG| = |-52|-52| + "repl.ph $s3, 102 \n" // |VR|VR| = |102|102| + "repl.ph $s4, 16 \n" // |0|16|0|16| + "repl.ph $s5, 128 \n" // |128|128| + "lui $s6, 0xff00 \n" + "ori $s6, 0xff00 \n" // |ff|00|ff|00| + + ".p2align 2 \n" + "1: \n" + I422ToTransientMipsRGB +// Arranging into abgr format + "precr.qb.ph $t0, $t8, $t1 \n" // |G1|g1|R1|r1| + "precr.qb.ph $t3, $t9, $t2 \n" // |G0|g0|R0|r0| + "precrq.qb.ph $t8, $t0, $t3 \n" // |G1|R1|G0|R0| + "precr.qb.ph $t9, $t0, $t3 \n" // |g1|r1|g0|r0| + + "precr.qb.ph $t2, $t4, $t5 \n" // |B1|b1|B0|b0| + "addiu %[width], -4 \n" + "addiu %[y_buf], 4 \n" + "preceu.ph.qbla $t1, $t2 \n" // |0 |B1|0 |B0| + "preceu.ph.qbra $t2, $t2 \n" // |0 |b1|0 |b0| + "or $t1, $t1, $s6 \n" // |ff|B1|ff|B0| + "or $t2, $t2, $s6 \n" // |ff|b1|ff|b0| + "precrq.ph.w $t0, $t2, $t9 \n" // |ff|b1|g1|r1| + "precrq.ph.w $t3, $t1, $t8 \n" // |ff|B1|G1|R1| + "sll $t9, $t9, 16 \n" + "sll $t8, $t8, 16 \n" + "packrl.ph $t2, $t2, $t9 \n" // |ff|b0|g0|r0| + "packrl.ph $t1, $t1, $t8 \n" // |ff|B0|G0|R0| +// Store results. + "sw $t2, 0(%[rgb_buf]) \n" + "sw $t0, 4(%[rgb_buf]) \n" + "sw $t1, 8(%[rgb_buf]) \n" + "sw $t3, 12(%[rgb_buf]) \n" + "bnez %[width], 1b \n" + " addiu %[rgb_buf], 16 \n" + "2: \n" + ".set pop \n" + :[y_buf] "+r" (y_buf), + [u_buf] "+r" (u_buf), + [v_buf] "+r" (v_buf), + [width] "+r" (width), + [rgb_buf] "+r" (rgb_buf) + : + : "t0", "t1", "t2", "t3", "t4", "t5", + "t6", "t7", "t8", "t9", + "s0", "s1", "s2", "s3", + "s4", "s5", "s6" + ); +} + +void I422ToBGRARow_MIPS_DSPR2(const uint8* y_buf, + const uint8* u_buf, + const uint8* v_buf, + uint8* rgb_buf, + int width) { + __asm__ __volatile__ ( + ".set push \n" + ".set noreorder \n" + "beqz %[width], 2f \n" + " repl.ph $s0, 74 \n" // |YG|YG| = |74 |74 | + "repl.ph $s1, -25 \n" // |UG|UG| = |-25|-25| + "repl.ph $s2, -52 \n" // |VG|VG| = |-52|-52| + "repl.ph $s3, 102 \n" // |VR|VR| = |102|102| + "repl.ph $s4, 16 \n" // |0|16|0|16| + "repl.ph $s5, 128 \n" // |128|128| + "lui $s6, 0xff \n" + "ori $s6, 0xff \n" // |00|ff|00|ff| + + ".p2align 2 \n" + "1: \n" + I422ToTransientMipsRGB + // Arranging into bgra format + "precr.qb.ph $t4, $t4, $t8 \n" // |B1|b1|G1|g1| + "precr.qb.ph $t5, $t5, $t9 \n" // |B0|b0|G0|g0| + "precrq.qb.ph $t8, $t4, $t5 \n" // |B1|G1|B0|G0| + "precr.qb.ph $t9, $t4, $t5 \n" // |b1|g1|b0|g0| + + "precr.qb.ph $t2, $t1, $t2 \n" // |R1|r1|R0|r0| + "addiu %[width], -4 \n" + "addiu %[y_buf], 4 \n" + "preceu.ph.qbla $t1, $t2 \n" // |0 |R1|0 |R0| + "preceu.ph.qbra $t2, $t2 \n" // |0 |r1|0 |r0| + "sll $t1, $t1, 8 \n" // |R1|0 |R0|0 | + "sll $t2, $t2, 8 \n" // |r1|0 |r0|0 | + "or $t1, $t1, $s6 \n" // |R1|ff|R0|ff| + "or $t2, $t2, $s6 \n" // |r1|ff|r0|ff| + "precrq.ph.w $t0, $t9, $t2 \n" // |b1|g1|r1|ff| + "precrq.ph.w $t3, $t8, $t1 \n" // |B1|G1|R1|ff| + "sll $t1, $t1, 16 \n" + "sll $t2, $t2, 16 \n" + "packrl.ph $t2, $t9, $t2 \n" // |b0|g0|r0|ff| + "packrl.ph $t1, $t8, $t1 \n" // |B0|G0|R0|ff| +// Store results. + "sw $t2, 0(%[rgb_buf]) \n" + "sw $t0, 4(%[rgb_buf]) \n" + "sw $t1, 8(%[rgb_buf]) \n" + "sw $t3, 12(%[rgb_buf]) \n" + "bnez %[width], 1b \n" + " addiu %[rgb_buf], 16 \n" + "2: \n" + ".set pop \n" + :[y_buf] "+r" (y_buf), + [u_buf] "+r" (u_buf), + [v_buf] "+r" (v_buf), + [width] "+r" (width), + [rgb_buf] "+r" (rgb_buf) + : + : "t0", "t1", "t2", "t3", "t4", "t5", + "t6", "t7", "t8", "t9", + "s0", "s1", "s2", "s3", + "s4", "s5", "s6" + ); +} + +// Bilinear filter 8x2 -> 8x1 +void InterpolateRow_MIPS_DSPR2(uint8* dst_ptr, const uint8* src_ptr, + ptrdiff_t src_stride, int dst_width, + int source_y_fraction) { + int y0_fraction = 256 - source_y_fraction; + const uint8* src_ptr1 = src_ptr + src_stride; + + __asm__ __volatile__ ( + ".set push \n" + ".set noreorder \n" + + "replv.ph $t0, %[y0_fraction] \n" + "replv.ph $t1, %[source_y_fraction] \n" + + ".p2align 2 \n" + "1: \n" + "lw $t2, 0(%[src_ptr]) \n" + "lw $t3, 0(%[src_ptr1]) \n" + "lw $t4, 4(%[src_ptr]) \n" + "lw $t5, 4(%[src_ptr1]) \n" + "muleu_s.ph.qbl $t6, $t2, $t0 \n" + "muleu_s.ph.qbr $t7, $t2, $t0 \n" + "muleu_s.ph.qbl $t8, $t3, $t1 \n" + "muleu_s.ph.qbr $t9, $t3, $t1 \n" + "muleu_s.ph.qbl $t2, $t4, $t0 \n" + "muleu_s.ph.qbr $t3, $t4, $t0 \n" + "muleu_s.ph.qbl $t4, $t5, $t1 \n" + "muleu_s.ph.qbr $t5, $t5, $t1 \n" + "addq.ph $t6, $t6, $t8 \n" + "addq.ph $t7, $t7, $t9 \n" + "addq.ph $t2, $t2, $t4 \n" + "addq.ph $t3, $t3, $t5 \n" + "shra.ph $t6, $t6, 8 \n" + "shra.ph $t7, $t7, 8 \n" + "shra.ph $t2, $t2, 8 \n" + "shra.ph $t3, $t3, 8 \n" + "precr.qb.ph $t6, $t6, $t7 \n" + "precr.qb.ph $t2, $t2, $t3 \n" + "addiu %[src_ptr], %[src_ptr], 8 \n" + "addiu %[src_ptr1], %[src_ptr1], 8 \n" + "addiu %[dst_width], %[dst_width], -8 \n" + "sw $t6, 0(%[dst_ptr]) \n" + "sw $t2, 4(%[dst_ptr]) \n" + "bgtz %[dst_width], 1b \n" + " addiu %[dst_ptr], %[dst_ptr], 8 \n" + + ".set pop \n" + : [dst_ptr] "+r" (dst_ptr), + [src_ptr1] "+r" (src_ptr1), + [src_ptr] "+r" (src_ptr), + [dst_width] "+r" (dst_width) + : [source_y_fraction] "r" (source_y_fraction), + [y0_fraction] "r" (y0_fraction), + [src_stride] "r" (src_stride) + : "t0", "t1", "t2", "t3", "t4", "t5", + "t6", "t7", "t8", "t9" + ); +} +#endif // __mips_dsp_rev >= 2 + +#endif // defined(__mips__) + +#ifdef __cplusplus +} // extern "C" +} // namespace libyuv +#endif diff --git a/media/libaom/src/third_party/libyuv/source/row_neon.cc b/media/libaom/src/third_party/libyuv/source/row_neon.cc new file mode 100644 index 000000000..1a72eb903 --- /dev/null +++ b/media/libaom/src/third_party/libyuv/source/row_neon.cc @@ -0,0 +1,3084 @@ +/* + * Copyright 2011 The LibYuv Project Authors. All rights reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#include "libyuv/row.h" + +#ifdef __cplusplus +namespace libyuv { +extern "C" { +#endif + +// This module is for GCC Neon +#if !defined(LIBYUV_DISABLE_NEON) && defined(__ARM_NEON__) && \ + !defined(__aarch64__) + +// Read 8 Y, 4 U and 4 V from 422 +#define READYUV422 \ + MEMACCESS(0) \ + "vld1.8 {d0}, [%0]! \n" \ + MEMACCESS(1) \ + "vld1.32 {d2[0]}, [%1]! \n" \ + MEMACCESS(2) \ + "vld1.32 {d2[1]}, [%2]! \n" + +// Read 8 Y, 2 U and 2 V from 422 +#define READYUV411 \ + MEMACCESS(0) \ + "vld1.8 {d0}, [%0]! \n" \ + MEMACCESS(1) \ + "vld1.16 {d2[0]}, [%1]! \n" \ + MEMACCESS(2) \ + "vld1.16 {d2[1]}, [%2]! \n" \ + "vmov.u8 d3, d2 \n" \ + "vzip.u8 d2, d3 \n" + +// Read 8 Y, 8 U and 8 V from 444 +#define READYUV444 \ + MEMACCESS(0) \ + "vld1.8 {d0}, [%0]! \n" \ + MEMACCESS(1) \ + "vld1.8 {d2}, [%1]! \n" \ + MEMACCESS(2) \ + "vld1.8 {d3}, [%2]! \n" \ + "vpaddl.u8 q1, q1 \n" \ + "vrshrn.u16 d2, q1, #1 \n" + +// Read 8 Y, and set 4 U and 4 V to 128 +#define READYUV400 \ + MEMACCESS(0) \ + "vld1.8 {d0}, [%0]! \n" \ + "vmov.u8 d2, #128 \n" + +// Read 8 Y and 4 UV from NV12 +#define READNV12 \ + MEMACCESS(0) \ + "vld1.8 {d0}, [%0]! \n" \ + MEMACCESS(1) \ + "vld1.8 {d2}, [%1]! \n" \ + "vmov.u8 d3, d2 \n"/* split odd/even uv apart */\ + "vuzp.u8 d2, d3 \n" \ + "vtrn.u32 d2, d3 \n" + +// Read 8 Y and 4 VU from NV21 +#define READNV21 \ + MEMACCESS(0) \ + "vld1.8 {d0}, [%0]! \n" \ + MEMACCESS(1) \ + "vld1.8 {d2}, [%1]! \n" \ + "vmov.u8 d3, d2 \n"/* split odd/even uv apart */\ + "vuzp.u8 d3, d2 \n" \ + "vtrn.u32 d2, d3 \n" + +// Read 8 YUY2 +#define READYUY2 \ + MEMACCESS(0) \ + "vld2.8 {d0, d2}, [%0]! \n" \ + "vmov.u8 d3, d2 \n" \ + "vuzp.u8 d2, d3 \n" \ + "vtrn.u32 d2, d3 \n" + +// Read 8 UYVY +#define READUYVY \ + MEMACCESS(0) \ + "vld2.8 {d2, d3}, [%0]! \n" \ + "vmov.u8 d0, d3 \n" \ + "vmov.u8 d3, d2 \n" \ + "vuzp.u8 d2, d3 \n" \ + "vtrn.u32 d2, d3 \n" + +#define YUV422TORGB_SETUP_REG \ + MEMACCESS([kUVToRB]) \ + "vld1.8 {d24}, [%[kUVToRB]] \n" \ + MEMACCESS([kUVToG]) \ + "vld1.8 {d25}, [%[kUVToG]] \n" \ + MEMACCESS([kUVBiasBGR]) \ + "vld1.16 {d26[], d27[]}, [%[kUVBiasBGR]]! \n" \ + MEMACCESS([kUVBiasBGR]) \ + "vld1.16 {d8[], d9[]}, [%[kUVBiasBGR]]! \n" \ + MEMACCESS([kUVBiasBGR]) \ + "vld1.16 {d28[], d29[]}, [%[kUVBiasBGR]] \n" \ + MEMACCESS([kYToRgb]) \ + "vld1.32 {d30[], d31[]}, [%[kYToRgb]] \n" + +#define YUV422TORGB \ + "vmull.u8 q8, d2, d24 \n" /* u/v B/R component */\ + "vmull.u8 q9, d2, d25 \n" /* u/v G component */\ + "vmovl.u8 q0, d0 \n" /* Y */\ + "vmovl.s16 q10, d1 \n" \ + "vmovl.s16 q0, d0 \n" \ + "vmul.s32 q10, q10, q15 \n" \ + "vmul.s32 q0, q0, q15 \n" \ + "vqshrun.s32 d0, q0, #16 \n" \ + "vqshrun.s32 d1, q10, #16 \n" /* Y */\ + "vadd.s16 d18, d19 \n" \ + "vshll.u16 q1, d16, #16 \n" /* Replicate u * UB */\ + "vshll.u16 q10, d17, #16 \n" /* Replicate v * VR */\ + "vshll.u16 q3, d18, #16 \n" /* Replicate (v*VG + u*UG)*/\ + "vaddw.u16 q1, q1, d16 \n" \ + "vaddw.u16 q10, q10, d17 \n" \ + "vaddw.u16 q3, q3, d18 \n" \ + "vqadd.s16 q8, q0, q13 \n" /* B */ \ + "vqadd.s16 q9, q0, q14 \n" /* R */ \ + "vqadd.s16 q0, q0, q4 \n" /* G */ \ + "vqadd.s16 q8, q8, q1 \n" /* B */ \ + "vqadd.s16 q9, q9, q10 \n" /* R */ \ + "vqsub.s16 q0, q0, q3 \n" /* G */ \ + "vqshrun.s16 d20, q8, #6 \n" /* B */ \ + "vqshrun.s16 d22, q9, #6 \n" /* R */ \ + "vqshrun.s16 d21, q0, #6 \n" /* G */ + +// YUV to RGB conversion constants. +// Y contribution to R,G,B. Scale and bias. +#define YG 18997 /* round(1.164 * 64 * 256 * 256 / 257) */ +#define YGB 1160 /* 1.164 * 64 * 16 - adjusted for even error distribution */ + +// U and V contributions to R,G,B. +#define UB -128 /* -min(128, round(2.018 * 64)) */ +#define UG 25 /* -round(-0.391 * 64) */ +#define VG 52 /* -round(-0.813 * 64) */ +#define VR -102 /* -round(1.596 * 64) */ + +// Bias values to subtract 16 from Y and 128 from U and V. +#define BB (UB * 128 - YGB) +#define BG (UG * 128 + VG * 128 - YGB) +#define BR (VR * 128 - YGB) + +static uvec8 kUVToRB = { 128, 128, 128, 128, 102, 102, 102, 102, + 0, 0, 0, 0, 0, 0, 0, 0 }; +static uvec8 kUVToG = { 25, 25, 25, 25, 52, 52, 52, 52, + 0, 0, 0, 0, 0, 0, 0, 0 }; +static vec16 kUVBiasBGR = { BB, BG, BR, 0, 0, 0, 0, 0 }; +static vec32 kYToRgb = { 0x0101 * YG, 0, 0, 0 }; + +#undef YG +#undef YGB +#undef UB +#undef UG +#undef VG +#undef VR +#undef BB +#undef BG +#undef BR + +void I444ToARGBRow_NEON(const uint8* src_y, + const uint8* src_u, + const uint8* src_v, + uint8* dst_argb, + int width) { + asm volatile ( + YUV422TORGB_SETUP_REG + ".p2align 2 \n" + "1: \n" + READYUV444 + YUV422TORGB + "subs %4, %4, #8 \n" + "vmov.u8 d23, #255 \n" + MEMACCESS(3) + "vst4.8 {d20, d21, d22, d23}, [%3]! \n" + "bgt 1b \n" + : "+r"(src_y), // %0 + "+r"(src_u), // %1 + "+r"(src_v), // %2 + "+r"(dst_argb), // %3 + "+r"(width) // %4 + : [kUVToRB]"r"(&kUVToRB), // %5 + [kUVToG]"r"(&kUVToG), // %6 + [kUVBiasBGR]"r"(&kUVBiasBGR), + [kYToRgb]"r"(&kYToRgb) + : "cc", "memory", "q0", "q1", "q2", "q3", "q4", + "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15" + ); +} + +void I422ToARGBRow_NEON(const uint8* src_y, + const uint8* src_u, + const uint8* src_v, + uint8* dst_argb, + int width) { + asm volatile ( + YUV422TORGB_SETUP_REG + ".p2align 2 \n" + "1: \n" + READYUV422 + YUV422TORGB + "subs %4, %4, #8 \n" + "vmov.u8 d23, #255 \n" + MEMACCESS(3) + "vst4.8 {d20, d21, d22, d23}, [%3]! \n" + "bgt 1b \n" + : "+r"(src_y), // %0 + "+r"(src_u), // %1 + "+r"(src_v), // %2 + "+r"(dst_argb), // %3 + "+r"(width) // %4 + : [kUVToRB]"r"(&kUVToRB), // %5 + [kUVToG]"r"(&kUVToG), // %6 + [kUVBiasBGR]"r"(&kUVBiasBGR), + [kYToRgb]"r"(&kYToRgb) + : "cc", "memory", "q0", "q1", "q2", "q3", "q4", + "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15" + ); +} + +void I411ToARGBRow_NEON(const uint8* src_y, + const uint8* src_u, + const uint8* src_v, + uint8* dst_argb, + int width) { + asm volatile ( + YUV422TORGB_SETUP_REG + ".p2align 2 \n" + "1: \n" + READYUV411 + YUV422TORGB + "subs %4, %4, #8 \n" + "vmov.u8 d23, #255 \n" + MEMACCESS(3) + "vst4.8 {d20, d21, d22, d23}, [%3]! \n" + "bgt 1b \n" + : "+r"(src_y), // %0 + "+r"(src_u), // %1 + "+r"(src_v), // %2 + "+r"(dst_argb), // %3 + "+r"(width) // %4 + : [kUVToRB]"r"(&kUVToRB), // %5 + [kUVToG]"r"(&kUVToG), // %6 + [kUVBiasBGR]"r"(&kUVBiasBGR), + [kYToRgb]"r"(&kYToRgb) + : "cc", "memory", "q0", "q1", "q2", "q3", "q4", + "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15" + ); +} + +void I422ToBGRARow_NEON(const uint8* src_y, + const uint8* src_u, + const uint8* src_v, + uint8* dst_bgra, + int width) { + asm volatile ( + YUV422TORGB_SETUP_REG + ".p2align 2 \n" + "1: \n" + READYUV422 + YUV422TORGB + "subs %4, %4, #8 \n" + "vswp.u8 d20, d22 \n" + "vmov.u8 d19, #255 \n" + MEMACCESS(3) + "vst4.8 {d19, d20, d21, d22}, [%3]! \n" + "bgt 1b \n" + : "+r"(src_y), // %0 + "+r"(src_u), // %1 + "+r"(src_v), // %2 + "+r"(dst_bgra), // %3 + "+r"(width) // %4 + : [kUVToRB]"r"(&kUVToRB), // %5 + [kUVToG]"r"(&kUVToG), // %6 + [kUVBiasBGR]"r"(&kUVBiasBGR), + [kYToRgb]"r"(&kYToRgb) + : "cc", "memory", "q0", "q1", "q2", "q3", "q4", + "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15" + ); +} + +void I422ToABGRRow_NEON(const uint8* src_y, + const uint8* src_u, + const uint8* src_v, + uint8* dst_abgr, + int width) { + asm volatile ( + YUV422TORGB_SETUP_REG + ".p2align 2 \n" + "1: \n" + READYUV422 + YUV422TORGB + "subs %4, %4, #8 \n" + "vswp.u8 d20, d22 \n" + "vmov.u8 d23, #255 \n" + MEMACCESS(3) + "vst4.8 {d20, d21, d22, d23}, [%3]! \n" + "bgt 1b \n" + : "+r"(src_y), // %0 + "+r"(src_u), // %1 + "+r"(src_v), // %2 + "+r"(dst_abgr), // %3 + "+r"(width) // %4 + : [kUVToRB]"r"(&kUVToRB), // %5 + [kUVToG]"r"(&kUVToG), // %6 + [kUVBiasBGR]"r"(&kUVBiasBGR), + [kYToRgb]"r"(&kYToRgb) + : "cc", "memory", "q0", "q1", "q2", "q3", "q4", + "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15" + ); +} + +void I422ToRGBARow_NEON(const uint8* src_y, + const uint8* src_u, + const uint8* src_v, + uint8* dst_rgba, + int width) { + asm volatile ( + YUV422TORGB_SETUP_REG + ".p2align 2 \n" + "1: \n" + READYUV422 + YUV422TORGB + "subs %4, %4, #8 \n" + "vmov.u8 d19, #255 \n" + MEMACCESS(3) + "vst4.8 {d19, d20, d21, d22}, [%3]! \n" + "bgt 1b \n" + : "+r"(src_y), // %0 + "+r"(src_u), // %1 + "+r"(src_v), // %2 + "+r"(dst_rgba), // %3 + "+r"(width) // %4 + : [kUVToRB]"r"(&kUVToRB), // %5 + [kUVToG]"r"(&kUVToG), // %6 + [kUVBiasBGR]"r"(&kUVBiasBGR), + [kYToRgb]"r"(&kYToRgb) + : "cc", "memory", "q0", "q1", "q2", "q3", "q4", + "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15" + ); +} + +void I422ToRGB24Row_NEON(const uint8* src_y, + const uint8* src_u, + const uint8* src_v, + uint8* dst_rgb24, + int width) { + asm volatile ( + YUV422TORGB_SETUP_REG + ".p2align 2 \n" + "1: \n" + READYUV422 + YUV422TORGB + "subs %4, %4, #8 \n" + MEMACCESS(3) + "vst3.8 {d20, d21, d22}, [%3]! \n" + "bgt 1b \n" + : "+r"(src_y), // %0 + "+r"(src_u), // %1 + "+r"(src_v), // %2 + "+r"(dst_rgb24), // %3 + "+r"(width) // %4 + : [kUVToRB]"r"(&kUVToRB), // %5 + [kUVToG]"r"(&kUVToG), // %6 + [kUVBiasBGR]"r"(&kUVBiasBGR), + [kYToRgb]"r"(&kYToRgb) + : "cc", "memory", "q0", "q1", "q2", "q3", "q4", + "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15" + ); +} + +void I422ToRAWRow_NEON(const uint8* src_y, + const uint8* src_u, + const uint8* src_v, + uint8* dst_raw, + int width) { + asm volatile ( + YUV422TORGB_SETUP_REG + ".p2align 2 \n" + "1: \n" + READYUV422 + YUV422TORGB + "subs %4, %4, #8 \n" + "vswp.u8 d20, d22 \n" + MEMACCESS(3) + "vst3.8 {d20, d21, d22}, [%3]! \n" + "bgt 1b \n" + : "+r"(src_y), // %0 + "+r"(src_u), // %1 + "+r"(src_v), // %2 + "+r"(dst_raw), // %3 + "+r"(width) // %4 + : [kUVToRB]"r"(&kUVToRB), // %5 + [kUVToG]"r"(&kUVToG), // %6 + [kUVBiasBGR]"r"(&kUVBiasBGR), + [kYToRgb]"r"(&kYToRgb) + : "cc", "memory", "q0", "q1", "q2", "q3", "q4", + "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15" + ); +} + +#define ARGBTORGB565 \ + "vshr.u8 d20, d20, #3 \n" /* B */ \ + "vshr.u8 d21, d21, #2 \n" /* G */ \ + "vshr.u8 d22, d22, #3 \n" /* R */ \ + "vmovl.u8 q8, d20 \n" /* B */ \ + "vmovl.u8 q9, d21 \n" /* G */ \ + "vmovl.u8 q10, d22 \n" /* R */ \ + "vshl.u16 q9, q9, #5 \n" /* G */ \ + "vshl.u16 q10, q10, #11 \n" /* R */ \ + "vorr q0, q8, q9 \n" /* BG */ \ + "vorr q0, q0, q10 \n" /* BGR */ + +void I422ToRGB565Row_NEON(const uint8* src_y, + const uint8* src_u, + const uint8* src_v, + uint8* dst_rgb565, + int width) { + asm volatile ( + YUV422TORGB_SETUP_REG + ".p2align 2 \n" + "1: \n" + READYUV422 + YUV422TORGB + "subs %4, %4, #8 \n" + ARGBTORGB565 + MEMACCESS(3) + "vst1.8 {q0}, [%3]! \n" // store 8 pixels RGB565. + "bgt 1b \n" + : "+r"(src_y), // %0 + "+r"(src_u), // %1 + "+r"(src_v), // %2 + "+r"(dst_rgb565), // %3 + "+r"(width) // %4 + : [kUVToRB]"r"(&kUVToRB), // %5 + [kUVToG]"r"(&kUVToG), // %6 + [kUVBiasBGR]"r"(&kUVBiasBGR), + [kYToRgb]"r"(&kYToRgb) + : "cc", "memory", "q0", "q1", "q2", "q3", "q4", + "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15" + ); +} + +#define ARGBTOARGB1555 \ + "vshr.u8 q10, q10, #3 \n" /* B */ \ + "vshr.u8 d22, d22, #3 \n" /* R */ \ + "vshr.u8 d23, d23, #7 \n" /* A */ \ + "vmovl.u8 q8, d20 \n" /* B */ \ + "vmovl.u8 q9, d21 \n" /* G */ \ + "vmovl.u8 q10, d22 \n" /* R */ \ + "vmovl.u8 q11, d23 \n" /* A */ \ + "vshl.u16 q9, q9, #5 \n" /* G */ \ + "vshl.u16 q10, q10, #10 \n" /* R */ \ + "vshl.u16 q11, q11, #15 \n" /* A */ \ + "vorr q0, q8, q9 \n" /* BG */ \ + "vorr q1, q10, q11 \n" /* RA */ \ + "vorr q0, q0, q1 \n" /* BGRA */ + +void I422ToARGB1555Row_NEON(const uint8* src_y, + const uint8* src_u, + const uint8* src_v, + uint8* dst_argb1555, + int width) { + asm volatile ( + YUV422TORGB_SETUP_REG + ".p2align 2 \n" + "1: \n" + READYUV422 + YUV422TORGB + "subs %4, %4, #8 \n" + "vmov.u8 d23, #255 \n" + ARGBTOARGB1555 + MEMACCESS(3) + "vst1.8 {q0}, [%3]! \n" // store 8 pixels ARGB1555. + "bgt 1b \n" + : "+r"(src_y), // %0 + "+r"(src_u), // %1 + "+r"(src_v), // %2 + "+r"(dst_argb1555), // %3 + "+r"(width) // %4 + : [kUVToRB]"r"(&kUVToRB), // %5 + [kUVToG]"r"(&kUVToG), // %6 + [kUVBiasBGR]"r"(&kUVBiasBGR), + [kYToRgb]"r"(&kYToRgb) + : "cc", "memory", "q0", "q1", "q2", "q3", "q4", + "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15" + ); +} + +#define ARGBTOARGB4444 \ + "vshr.u8 d20, d20, #4 \n" /* B */ \ + "vbic.32 d21, d21, d4 \n" /* G */ \ + "vshr.u8 d22, d22, #4 \n" /* R */ \ + "vbic.32 d23, d23, d4 \n" /* A */ \ + "vorr d0, d20, d21 \n" /* BG */ \ + "vorr d1, d22, d23 \n" /* RA */ \ + "vzip.u8 d0, d1 \n" /* BGRA */ + +void I422ToARGB4444Row_NEON(const uint8* src_y, + const uint8* src_u, + const uint8* src_v, + uint8* dst_argb4444, + int width) { + asm volatile ( + YUV422TORGB_SETUP_REG + "vmov.u8 d4, #0x0f \n" // bits to clear with vbic. + ".p2align 2 \n" + "1: \n" + READYUV422 + YUV422TORGB + "subs %4, %4, #8 \n" + "vmov.u8 d23, #255 \n" + ARGBTOARGB4444 + MEMACCESS(3) + "vst1.8 {q0}, [%3]! \n" // store 8 pixels ARGB4444. + "bgt 1b \n" + : "+r"(src_y), // %0 + "+r"(src_u), // %1 + "+r"(src_v), // %2 + "+r"(dst_argb4444), // %3 + "+r"(width) // %4 + : [kUVToRB]"r"(&kUVToRB), // %5 + [kUVToG]"r"(&kUVToG), // %6 + [kUVBiasBGR]"r"(&kUVBiasBGR), + [kYToRgb]"r"(&kYToRgb) + : "cc", "memory", "q0", "q1", "q2", "q3", "q4", + "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15" + ); +} + +void I400ToARGBRow_NEON(const uint8* src_y, + uint8* dst_argb, + int width) { + asm volatile ( + YUV422TORGB_SETUP_REG + ".p2align 2 \n" + "1: \n" + READYUV400 + YUV422TORGB + "subs %2, %2, #8 \n" + "vmov.u8 d23, #255 \n" + MEMACCESS(1) + "vst4.8 {d20, d21, d22, d23}, [%1]! \n" + "bgt 1b \n" + : "+r"(src_y), // %0 + "+r"(dst_argb), // %1 + "+r"(width) // %2 + : [kUVToRB]"r"(&kUVToRB), // %3 + [kUVToG]"r"(&kUVToG), // %4 + [kUVBiasBGR]"r"(&kUVBiasBGR), + [kYToRgb]"r"(&kYToRgb) + : "cc", "memory", "q0", "q1", "q2", "q3", "q4", + "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15" + ); +} + +void J400ToARGBRow_NEON(const uint8* src_y, + uint8* dst_argb, + int width) { + asm volatile ( + "vmov.u8 d23, #255 \n" + ".p2align 2 \n" + "1: \n" + MEMACCESS(0) + "vld1.8 {d20}, [%0]! \n" + "vmov d21, d20 \n" + "vmov d22, d20 \n" + "subs %2, %2, #8 \n" + MEMACCESS(1) + "vst4.8 {d20, d21, d22, d23}, [%1]! \n" + "bgt 1b \n" + : "+r"(src_y), // %0 + "+r"(dst_argb), // %1 + "+r"(width) // %2 + : + : "cc", "memory", "d20", "d21", "d22", "d23" + ); +} + +void NV12ToARGBRow_NEON(const uint8* src_y, + const uint8* src_uv, + uint8* dst_argb, + int width) { + asm volatile ( + YUV422TORGB_SETUP_REG + ".p2align 2 \n" + "1: \n" + READNV12 + YUV422TORGB + "subs %3, %3, #8 \n" + "vmov.u8 d23, #255 \n" + MEMACCESS(2) + "vst4.8 {d20, d21, d22, d23}, [%2]! \n" + "bgt 1b \n" + : "+r"(src_y), // %0 + "+r"(src_uv), // %1 + "+r"(dst_argb), // %2 + "+r"(width) // %3 + : [kUVToRB]"r"(&kUVToRB), // %4 + [kUVToG]"r"(&kUVToG), // %5 + [kUVBiasBGR]"r"(&kUVBiasBGR), + [kYToRgb]"r"(&kYToRgb) + : "cc", "memory", "q0", "q1", "q2", "q3", "q4", + "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15" + ); +} + +void NV21ToARGBRow_NEON(const uint8* src_y, + const uint8* src_uv, + uint8* dst_argb, + int width) { + asm volatile ( + YUV422TORGB_SETUP_REG + ".p2align 2 \n" + "1: \n" + READNV21 + YUV422TORGB + "subs %3, %3, #8 \n" + "vmov.u8 d23, #255 \n" + MEMACCESS(2) + "vst4.8 {d20, d21, d22, d23}, [%2]! \n" + "bgt 1b \n" + : "+r"(src_y), // %0 + "+r"(src_uv), // %1 + "+r"(dst_argb), // %2 + "+r"(width) // %3 + : [kUVToRB]"r"(&kUVToRB), // %4 + [kUVToG]"r"(&kUVToG), // %5 + [kUVBiasBGR]"r"(&kUVBiasBGR), + [kYToRgb]"r"(&kYToRgb) + : "cc", "memory", "q0", "q1", "q2", "q3", "q4", + "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15" + ); +} + +void NV12ToRGB565Row_NEON(const uint8* src_y, + const uint8* src_uv, + uint8* dst_rgb565, + int width) { + asm volatile ( + YUV422TORGB_SETUP_REG + ".p2align 2 \n" + "1: \n" + READNV12 + YUV422TORGB + "subs %3, %3, #8 \n" + ARGBTORGB565 + MEMACCESS(2) + "vst1.8 {q0}, [%2]! \n" // store 8 pixels RGB565. + "bgt 1b \n" + : "+r"(src_y), // %0 + "+r"(src_uv), // %1 + "+r"(dst_rgb565), // %2 + "+r"(width) // %3 + : [kUVToRB]"r"(&kUVToRB), // %4 + [kUVToG]"r"(&kUVToG), // %5 + [kUVBiasBGR]"r"(&kUVBiasBGR), + [kYToRgb]"r"(&kYToRgb) + : "cc", "memory", "q0", "q1", "q2", "q3", "q4", + "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15" + ); +} + +void NV21ToRGB565Row_NEON(const uint8* src_y, + const uint8* src_uv, + uint8* dst_rgb565, + int width) { + asm volatile ( + YUV422TORGB_SETUP_REG + ".p2align 2 \n" + "1: \n" + READNV21 + YUV422TORGB + "subs %3, %3, #8 \n" + ARGBTORGB565 + MEMACCESS(2) + "vst1.8 {q0}, [%2]! \n" // store 8 pixels RGB565. + "bgt 1b \n" + : "+r"(src_y), // %0 + "+r"(src_uv), // %1 + "+r"(dst_rgb565), // %2 + "+r"(width) // %3 + : [kUVToRB]"r"(&kUVToRB), // %4 + [kUVToG]"r"(&kUVToG), // %5 + [kUVBiasBGR]"r"(&kUVBiasBGR), + [kYToRgb]"r"(&kYToRgb) + : "cc", "memory", "q0", "q1", "q2", "q3", "q4", + "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15" + ); +} + +void YUY2ToARGBRow_NEON(const uint8* src_yuy2, + uint8* dst_argb, + int width) { + asm volatile ( + YUV422TORGB_SETUP_REG + ".p2align 2 \n" + "1: \n" + READYUY2 + YUV422TORGB + "subs %2, %2, #8 \n" + "vmov.u8 d23, #255 \n" + MEMACCESS(1) + "vst4.8 {d20, d21, d22, d23}, [%1]! \n" + "bgt 1b \n" + : "+r"(src_yuy2), // %0 + "+r"(dst_argb), // %1 + "+r"(width) // %2 + : [kUVToRB]"r"(&kUVToRB), // %3 + [kUVToG]"r"(&kUVToG), // %4 + [kUVBiasBGR]"r"(&kUVBiasBGR), + [kYToRgb]"r"(&kYToRgb) + : "cc", "memory", "q0", "q1", "q2", "q3", "q4", + "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15" + ); +} + +void UYVYToARGBRow_NEON(const uint8* src_uyvy, + uint8* dst_argb, + int width) { + asm volatile ( + YUV422TORGB_SETUP_REG + ".p2align 2 \n" + "1: \n" + READUYVY + YUV422TORGB + "subs %2, %2, #8 \n" + "vmov.u8 d23, #255 \n" + MEMACCESS(1) + "vst4.8 {d20, d21, d22, d23}, [%1]! \n" + "bgt 1b \n" + : "+r"(src_uyvy), // %0 + "+r"(dst_argb), // %1 + "+r"(width) // %2 + : [kUVToRB]"r"(&kUVToRB), // %3 + [kUVToG]"r"(&kUVToG), // %4 + [kUVBiasBGR]"r"(&kUVBiasBGR), + [kYToRgb]"r"(&kYToRgb) + : "cc", "memory", "q0", "q1", "q2", "q3", "q4", + "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15" + ); +} + +// Reads 16 pairs of UV and write even values to dst_u and odd to dst_v. +void SplitUVRow_NEON(const uint8* src_uv, uint8* dst_u, uint8* dst_v, + int width) { + asm volatile ( + ".p2align 2 \n" + "1: \n" + MEMACCESS(0) + "vld2.8 {q0, q1}, [%0]! \n" // load 16 pairs of UV + "subs %3, %3, #16 \n" // 16 processed per loop + MEMACCESS(1) + "vst1.8 {q0}, [%1]! \n" // store U + MEMACCESS(2) + "vst1.8 {q1}, [%2]! \n" // store V + "bgt 1b \n" + : "+r"(src_uv), // %0 + "+r"(dst_u), // %1 + "+r"(dst_v), // %2 + "+r"(width) // %3 // Output registers + : // Input registers + : "cc", "memory", "q0", "q1" // Clobber List + ); +} + +// Reads 16 U's and V's and writes out 16 pairs of UV. +void MergeUVRow_NEON(const uint8* src_u, const uint8* src_v, uint8* dst_uv, + int width) { + asm volatile ( + ".p2align 2 \n" + "1: \n" + MEMACCESS(0) + "vld1.8 {q0}, [%0]! \n" // load U + MEMACCESS(1) + "vld1.8 {q1}, [%1]! \n" // load V + "subs %3, %3, #16 \n" // 16 processed per loop + MEMACCESS(2) + "vst2.u8 {q0, q1}, [%2]! \n" // store 16 pairs of UV + "bgt 1b \n" + : + "+r"(src_u), // %0 + "+r"(src_v), // %1 + "+r"(dst_uv), // %2 + "+r"(width) // %3 // Output registers + : // Input registers + : "cc", "memory", "q0", "q1" // Clobber List + ); +} + +// Copy multiple of 32. vld4.8 allow unaligned and is fastest on a15. +void CopyRow_NEON(const uint8* src, uint8* dst, int count) { + asm volatile ( + ".p2align 2 \n" + "1: \n" + MEMACCESS(0) + "vld1.8 {d0, d1, d2, d3}, [%0]! \n" // load 32 + "subs %2, %2, #32 \n" // 32 processed per loop + MEMACCESS(1) + "vst1.8 {d0, d1, d2, d3}, [%1]! \n" // store 32 + "bgt 1b \n" + : "+r"(src), // %0 + "+r"(dst), // %1 + "+r"(count) // %2 // Output registers + : // Input registers + : "cc", "memory", "q0", "q1" // Clobber List + ); +} + +// SetRow writes 'count' bytes using an 8 bit value repeated. +void SetRow_NEON(uint8* dst, uint8 v8, int count) { + asm volatile ( + "vdup.8 q0, %2 \n" // duplicate 16 bytes + "1: \n" + "subs %1, %1, #16 \n" // 16 bytes per loop + MEMACCESS(0) + "vst1.8 {q0}, [%0]! \n" // store + "bgt 1b \n" + : "+r"(dst), // %0 + "+r"(count) // %1 + : "r"(v8) // %2 + : "cc", "memory", "q0" + ); +} + +// ARGBSetRow writes 'count' pixels using an 32 bit value repeated. +void ARGBSetRow_NEON(uint8* dst, uint32 v32, int count) { + asm volatile ( + "vdup.u32 q0, %2 \n" // duplicate 4 ints + "1: \n" + "subs %1, %1, #4 \n" // 4 pixels per loop + MEMACCESS(0) + "vst1.8 {q0}, [%0]! \n" // store + "bgt 1b \n" + : "+r"(dst), // %0 + "+r"(count) // %1 + : "r"(v32) // %2 + : "cc", "memory", "q0" + ); +} + +void MirrorRow_NEON(const uint8* src, uint8* dst, int width) { + asm volatile ( + // Start at end of source row. + "mov r3, #-16 \n" + "add %0, %0, %2 \n" + "sub %0, #16 \n" + + ".p2align 2 \n" + "1: \n" + MEMACCESS(0) + "vld1.8 {q0}, [%0], r3 \n" // src -= 16 + "subs %2, #16 \n" // 16 pixels per loop. + "vrev64.8 q0, q0 \n" + MEMACCESS(1) + "vst1.8 {d1}, [%1]! \n" // dst += 16 + MEMACCESS(1) + "vst1.8 {d0}, [%1]! \n" + "bgt 1b \n" + : "+r"(src), // %0 + "+r"(dst), // %1 + "+r"(width) // %2 + : + : "cc", "memory", "r3", "q0" + ); +} + +void MirrorUVRow_NEON(const uint8* src_uv, uint8* dst_u, uint8* dst_v, + int width) { + asm volatile ( + // Start at end of source row. + "mov r12, #-16 \n" + "add %0, %0, %3, lsl #1 \n" + "sub %0, #16 \n" + + ".p2align 2 \n" + "1: \n" + MEMACCESS(0) + "vld2.8 {d0, d1}, [%0], r12 \n" // src -= 16 + "subs %3, #8 \n" // 8 pixels per loop. + "vrev64.8 q0, q0 \n" + MEMACCESS(1) + "vst1.8 {d0}, [%1]! \n" // dst += 8 + MEMACCESS(2) + "vst1.8 {d1}, [%2]! \n" + "bgt 1b \n" + : "+r"(src_uv), // %0 + "+r"(dst_u), // %1 + "+r"(dst_v), // %2 + "+r"(width) // %3 + : + : "cc", "memory", "r12", "q0" + ); +} + +void ARGBMirrorRow_NEON(const uint8* src, uint8* dst, int width) { + asm volatile ( + // Start at end of source row. + "mov r3, #-16 \n" + "add %0, %0, %2, lsl #2 \n" + "sub %0, #16 \n" + + ".p2align 2 \n" + "1: \n" + MEMACCESS(0) + "vld1.8 {q0}, [%0], r3 \n" // src -= 16 + "subs %2, #4 \n" // 4 pixels per loop. + "vrev64.32 q0, q0 \n" + MEMACCESS(1) + "vst1.8 {d1}, [%1]! \n" // dst += 16 + MEMACCESS(1) + "vst1.8 {d0}, [%1]! \n" + "bgt 1b \n" + : "+r"(src), // %0 + "+r"(dst), // %1 + "+r"(width) // %2 + : + : "cc", "memory", "r3", "q0" + ); +} + +void RGB24ToARGBRow_NEON(const uint8* src_rgb24, uint8* dst_argb, int pix) { + asm volatile ( + "vmov.u8 d4, #255 \n" // Alpha + ".p2align 2 \n" + "1: \n" + MEMACCESS(0) + "vld3.8 {d1, d2, d3}, [%0]! \n" // load 8 pixels of RGB24. + "subs %2, %2, #8 \n" // 8 processed per loop. + MEMACCESS(1) + "vst4.8 {d1, d2, d3, d4}, [%1]! \n" // store 8 pixels of ARGB. + "bgt 1b \n" + : "+r"(src_rgb24), // %0 + "+r"(dst_argb), // %1 + "+r"(pix) // %2 + : + : "cc", "memory", "d1", "d2", "d3", "d4" // Clobber List + ); +} + +void RAWToARGBRow_NEON(const uint8* src_raw, uint8* dst_argb, int pix) { + asm volatile ( + "vmov.u8 d4, #255 \n" // Alpha + ".p2align 2 \n" + "1: \n" + MEMACCESS(0) + "vld3.8 {d1, d2, d3}, [%0]! \n" // load 8 pixels of RAW. + "subs %2, %2, #8 \n" // 8 processed per loop. + "vswp.u8 d1, d3 \n" // swap R, B + MEMACCESS(1) + "vst4.8 {d1, d2, d3, d4}, [%1]! \n" // store 8 pixels of ARGB. + "bgt 1b \n" + : "+r"(src_raw), // %0 + "+r"(dst_argb), // %1 + "+r"(pix) // %2 + : + : "cc", "memory", "d1", "d2", "d3", "d4" // Clobber List + ); +} + +#define RGB565TOARGB \ + "vshrn.u16 d6, q0, #5 \n" /* G xxGGGGGG */ \ + "vuzp.u8 d0, d1 \n" /* d0 xxxBBBBB RRRRRxxx */ \ + "vshl.u8 d6, d6, #2 \n" /* G GGGGGG00 upper 6 */ \ + "vshr.u8 d1, d1, #3 \n" /* R 000RRRRR lower 5 */ \ + "vshl.u8 q0, q0, #3 \n" /* B,R BBBBB000 upper 5 */ \ + "vshr.u8 q2, q0, #5 \n" /* B,R 00000BBB lower 3 */ \ + "vorr.u8 d0, d0, d4 \n" /* B */ \ + "vshr.u8 d4, d6, #6 \n" /* G 000000GG lower 2 */ \ + "vorr.u8 d2, d1, d5 \n" /* R */ \ + "vorr.u8 d1, d4, d6 \n" /* G */ + +void RGB565ToARGBRow_NEON(const uint8* src_rgb565, uint8* dst_argb, int pix) { + asm volatile ( + "vmov.u8 d3, #255 \n" // Alpha + ".p2align 2 \n" + "1: \n" + MEMACCESS(0) + "vld1.8 {q0}, [%0]! \n" // load 8 RGB565 pixels. + "subs %2, %2, #8 \n" // 8 processed per loop. + RGB565TOARGB + MEMACCESS(1) + "vst4.8 {d0, d1, d2, d3}, [%1]! \n" // store 8 pixels of ARGB. + "bgt 1b \n" + : "+r"(src_rgb565), // %0 + "+r"(dst_argb), // %1 + "+r"(pix) // %2 + : + : "cc", "memory", "q0", "q1", "q2", "q3" // Clobber List + ); +} + +#define ARGB1555TOARGB \ + "vshrn.u16 d7, q0, #8 \n" /* A Arrrrrxx */ \ + "vshr.u8 d6, d7, #2 \n" /* R xxxRRRRR */ \ + "vshrn.u16 d5, q0, #5 \n" /* G xxxGGGGG */ \ + "vmovn.u16 d4, q0 \n" /* B xxxBBBBB */ \ + "vshr.u8 d7, d7, #7 \n" /* A 0000000A */ \ + "vneg.s8 d7, d7 \n" /* A AAAAAAAA upper 8 */ \ + "vshl.u8 d6, d6, #3 \n" /* R RRRRR000 upper 5 */ \ + "vshr.u8 q1, q3, #5 \n" /* R,A 00000RRR lower 3 */ \ + "vshl.u8 q0, q2, #3 \n" /* B,G BBBBB000 upper 5 */ \ + "vshr.u8 q2, q0, #5 \n" /* B,G 00000BBB lower 3 */ \ + "vorr.u8 q1, q1, q3 \n" /* R,A */ \ + "vorr.u8 q0, q0, q2 \n" /* B,G */ \ + +// RGB555TOARGB is same as ARGB1555TOARGB but ignores alpha. +#define RGB555TOARGB \ + "vshrn.u16 d6, q0, #5 \n" /* G xxxGGGGG */ \ + "vuzp.u8 d0, d1 \n" /* d0 xxxBBBBB xRRRRRxx */ \ + "vshl.u8 d6, d6, #3 \n" /* G GGGGG000 upper 5 */ \ + "vshr.u8 d1, d1, #2 \n" /* R 00xRRRRR lower 5 */ \ + "vshl.u8 q0, q0, #3 \n" /* B,R BBBBB000 upper 5 */ \ + "vshr.u8 q2, q0, #5 \n" /* B,R 00000BBB lower 3 */ \ + "vorr.u8 d0, d0, d4 \n" /* B */ \ + "vshr.u8 d4, d6, #5 \n" /* G 00000GGG lower 3 */ \ + "vorr.u8 d2, d1, d5 \n" /* R */ \ + "vorr.u8 d1, d4, d6 \n" /* G */ + +void ARGB1555ToARGBRow_NEON(const uint8* src_argb1555, uint8* dst_argb, + int pix) { + asm volatile ( + "vmov.u8 d3, #255 \n" // Alpha + ".p2align 2 \n" + "1: \n" + MEMACCESS(0) + "vld1.8 {q0}, [%0]! \n" // load 8 ARGB1555 pixels. + "subs %2, %2, #8 \n" // 8 processed per loop. + ARGB1555TOARGB + MEMACCESS(1) + "vst4.8 {d0, d1, d2, d3}, [%1]! \n" // store 8 pixels of ARGB. + "bgt 1b \n" + : "+r"(src_argb1555), // %0 + "+r"(dst_argb), // %1 + "+r"(pix) // %2 + : + : "cc", "memory", "q0", "q1", "q2", "q3" // Clobber List + ); +} + +#define ARGB4444TOARGB \ + "vuzp.u8 d0, d1 \n" /* d0 BG, d1 RA */ \ + "vshl.u8 q2, q0, #4 \n" /* B,R BBBB0000 */ \ + "vshr.u8 q1, q0, #4 \n" /* G,A 0000GGGG */ \ + "vshr.u8 q0, q2, #4 \n" /* B,R 0000BBBB */ \ + "vorr.u8 q0, q0, q2 \n" /* B,R BBBBBBBB */ \ + "vshl.u8 q2, q1, #4 \n" /* G,A GGGG0000 */ \ + "vorr.u8 q1, q1, q2 \n" /* G,A GGGGGGGG */ \ + "vswp.u8 d1, d2 \n" /* B,R,G,A -> B,G,R,A */ + +void ARGB4444ToARGBRow_NEON(const uint8* src_argb4444, uint8* dst_argb, + int pix) { + asm volatile ( + "vmov.u8 d3, #255 \n" // Alpha + ".p2align 2 \n" + "1: \n" + MEMACCESS(0) + "vld1.8 {q0}, [%0]! \n" // load 8 ARGB4444 pixels. + "subs %2, %2, #8 \n" // 8 processed per loop. + ARGB4444TOARGB + MEMACCESS(1) + "vst4.8 {d0, d1, d2, d3}, [%1]! \n" // store 8 pixels of ARGB. + "bgt 1b \n" + : "+r"(src_argb4444), // %0 + "+r"(dst_argb), // %1 + "+r"(pix) // %2 + : + : "cc", "memory", "q0", "q1", "q2" // Clobber List + ); +} + +void ARGBToRGB24Row_NEON(const uint8* src_argb, uint8* dst_rgb24, int pix) { + asm volatile ( + ".p2align 2 \n" + "1: \n" + MEMACCESS(0) + "vld4.8 {d1, d2, d3, d4}, [%0]! \n" // load 8 pixels of ARGB. + "subs %2, %2, #8 \n" // 8 processed per loop. + MEMACCESS(1) + "vst3.8 {d1, d2, d3}, [%1]! \n" // store 8 pixels of RGB24. + "bgt 1b \n" + : "+r"(src_argb), // %0 + "+r"(dst_rgb24), // %1 + "+r"(pix) // %2 + : + : "cc", "memory", "d1", "d2", "d3", "d4" // Clobber List + ); +} + +void ARGBToRAWRow_NEON(const uint8* src_argb, uint8* dst_raw, int pix) { + asm volatile ( + ".p2align 2 \n" + "1: \n" + MEMACCESS(0) + "vld4.8 {d1, d2, d3, d4}, [%0]! \n" // load 8 pixels of ARGB. + "subs %2, %2, #8 \n" // 8 processed per loop. + "vswp.u8 d1, d3 \n" // swap R, B + MEMACCESS(1) + "vst3.8 {d1, d2, d3}, [%1]! \n" // store 8 pixels of RAW. + "bgt 1b \n" + : "+r"(src_argb), // %0 + "+r"(dst_raw), // %1 + "+r"(pix) // %2 + : + : "cc", "memory", "d1", "d2", "d3", "d4" // Clobber List + ); +} + +void YUY2ToYRow_NEON(const uint8* src_yuy2, uint8* dst_y, int pix) { + asm volatile ( + ".p2align 2 \n" + "1: \n" + MEMACCESS(0) + "vld2.8 {q0, q1}, [%0]! \n" // load 16 pixels of YUY2. + "subs %2, %2, #16 \n" // 16 processed per loop. + MEMACCESS(1) + "vst1.8 {q0}, [%1]! \n" // store 16 pixels of Y. + "bgt 1b \n" + : "+r"(src_yuy2), // %0 + "+r"(dst_y), // %1 + "+r"(pix) // %2 + : + : "cc", "memory", "q0", "q1" // Clobber List + ); +} + +void UYVYToYRow_NEON(const uint8* src_uyvy, uint8* dst_y, int pix) { + asm volatile ( + ".p2align 2 \n" + "1: \n" + MEMACCESS(0) + "vld2.8 {q0, q1}, [%0]! \n" // load 16 pixels of UYVY. + "subs %2, %2, #16 \n" // 16 processed per loop. + MEMACCESS(1) + "vst1.8 {q1}, [%1]! \n" // store 16 pixels of Y. + "bgt 1b \n" + : "+r"(src_uyvy), // %0 + "+r"(dst_y), // %1 + "+r"(pix) // %2 + : + : "cc", "memory", "q0", "q1" // Clobber List + ); +} + +void YUY2ToUV422Row_NEON(const uint8* src_yuy2, uint8* dst_u, uint8* dst_v, + int pix) { + asm volatile ( + ".p2align 2 \n" + "1: \n" + MEMACCESS(0) + "vld4.8 {d0, d1, d2, d3}, [%0]! \n" // load 16 pixels of YUY2. + "subs %3, %3, #16 \n" // 16 pixels = 8 UVs. + MEMACCESS(1) + "vst1.8 {d1}, [%1]! \n" // store 8 U. + MEMACCESS(2) + "vst1.8 {d3}, [%2]! \n" // store 8 V. + "bgt 1b \n" + : "+r"(src_yuy2), // %0 + "+r"(dst_u), // %1 + "+r"(dst_v), // %2 + "+r"(pix) // %3 + : + : "cc", "memory", "d0", "d1", "d2", "d3" // Clobber List + ); +} + +void UYVYToUV422Row_NEON(const uint8* src_uyvy, uint8* dst_u, uint8* dst_v, + int pix) { + asm volatile ( + ".p2align 2 \n" + "1: \n" + MEMACCESS(0) + "vld4.8 {d0, d1, d2, d3}, [%0]! \n" // load 16 pixels of UYVY. + "subs %3, %3, #16 \n" // 16 pixels = 8 UVs. + MEMACCESS(1) + "vst1.8 {d0}, [%1]! \n" // store 8 U. + MEMACCESS(2) + "vst1.8 {d2}, [%2]! \n" // store 8 V. + "bgt 1b \n" + : "+r"(src_uyvy), // %0 + "+r"(dst_u), // %1 + "+r"(dst_v), // %2 + "+r"(pix) // %3 + : + : "cc", "memory", "d0", "d1", "d2", "d3" // Clobber List + ); +} + +void YUY2ToUVRow_NEON(const uint8* src_yuy2, int stride_yuy2, + uint8* dst_u, uint8* dst_v, int pix) { + asm volatile ( + "add %1, %0, %1 \n" // stride + src_yuy2 + ".p2align 2 \n" + "1: \n" + MEMACCESS(0) + "vld4.8 {d0, d1, d2, d3}, [%0]! \n" // load 16 pixels of YUY2. + "subs %4, %4, #16 \n" // 16 pixels = 8 UVs. + MEMACCESS(1) + "vld4.8 {d4, d5, d6, d7}, [%1]! \n" // load next row YUY2. + "vrhadd.u8 d1, d1, d5 \n" // average rows of U + "vrhadd.u8 d3, d3, d7 \n" // average rows of V + MEMACCESS(2) + "vst1.8 {d1}, [%2]! \n" // store 8 U. + MEMACCESS(3) + "vst1.8 {d3}, [%3]! \n" // store 8 V. + "bgt 1b \n" + : "+r"(src_yuy2), // %0 + "+r"(stride_yuy2), // %1 + "+r"(dst_u), // %2 + "+r"(dst_v), // %3 + "+r"(pix) // %4 + : + : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7" // Clobber List + ); +} + +void UYVYToUVRow_NEON(const uint8* src_uyvy, int stride_uyvy, + uint8* dst_u, uint8* dst_v, int pix) { + asm volatile ( + "add %1, %0, %1 \n" // stride + src_uyvy + ".p2align 2 \n" + "1: \n" + MEMACCESS(0) + "vld4.8 {d0, d1, d2, d3}, [%0]! \n" // load 16 pixels of UYVY. + "subs %4, %4, #16 \n" // 16 pixels = 8 UVs. + MEMACCESS(1) + "vld4.8 {d4, d5, d6, d7}, [%1]! \n" // load next row UYVY. + "vrhadd.u8 d0, d0, d4 \n" // average rows of U + "vrhadd.u8 d2, d2, d6 \n" // average rows of V + MEMACCESS(2) + "vst1.8 {d0}, [%2]! \n" // store 8 U. + MEMACCESS(3) + "vst1.8 {d2}, [%3]! \n" // store 8 V. + "bgt 1b \n" + : "+r"(src_uyvy), // %0 + "+r"(stride_uyvy), // %1 + "+r"(dst_u), // %2 + "+r"(dst_v), // %3 + "+r"(pix) // %4 + : + : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7" // Clobber List + ); +} + +// For BGRAToARGB, ABGRToARGB, RGBAToARGB, and ARGBToRGBA. +void ARGBShuffleRow_NEON(const uint8* src_argb, uint8* dst_argb, + const uint8* shuffler, int pix) { + asm volatile ( + MEMACCESS(3) + "vld1.8 {q2}, [%3] \n" // shuffler + "1: \n" + MEMACCESS(0) + "vld1.8 {q0}, [%0]! \n" // load 4 pixels. + "subs %2, %2, #4 \n" // 4 processed per loop + "vtbl.8 d2, {d0, d1}, d4 \n" // look up 2 first pixels + "vtbl.8 d3, {d0, d1}, d5 \n" // look up 2 next pixels + MEMACCESS(1) + "vst1.8 {q1}, [%1]! \n" // store 4. + "bgt 1b \n" + : "+r"(src_argb), // %0 + "+r"(dst_argb), // %1 + "+r"(pix) // %2 + : "r"(shuffler) // %3 + : "cc", "memory", "q0", "q1", "q2" // Clobber List + ); +} + +void I422ToYUY2Row_NEON(const uint8* src_y, + const uint8* src_u, + const uint8* src_v, + uint8* dst_yuy2, int width) { + asm volatile ( + ".p2align 2 \n" + "1: \n" + MEMACCESS(0) + "vld2.8 {d0, d2}, [%0]! \n" // load 16 Ys + MEMACCESS(1) + "vld1.8 {d1}, [%1]! \n" // load 8 Us + MEMACCESS(2) + "vld1.8 {d3}, [%2]! \n" // load 8 Vs + "subs %4, %4, #16 \n" // 16 pixels + MEMACCESS(3) + "vst4.8 {d0, d1, d2, d3}, [%3]! \n" // Store 8 YUY2/16 pixels. + "bgt 1b \n" + : "+r"(src_y), // %0 + "+r"(src_u), // %1 + "+r"(src_v), // %2 + "+r"(dst_yuy2), // %3 + "+r"(width) // %4 + : + : "cc", "memory", "d0", "d1", "d2", "d3" + ); +} + +void I422ToUYVYRow_NEON(const uint8* src_y, + const uint8* src_u, + const uint8* src_v, + uint8* dst_uyvy, int width) { + asm volatile ( + ".p2align 2 \n" + "1: \n" + MEMACCESS(0) + "vld2.8 {d1, d3}, [%0]! \n" // load 16 Ys + MEMACCESS(1) + "vld1.8 {d0}, [%1]! \n" // load 8 Us + MEMACCESS(2) + "vld1.8 {d2}, [%2]! \n" // load 8 Vs + "subs %4, %4, #16 \n" // 16 pixels + MEMACCESS(3) + "vst4.8 {d0, d1, d2, d3}, [%3]! \n" // Store 8 UYVY/16 pixels. + "bgt 1b \n" + : "+r"(src_y), // %0 + "+r"(src_u), // %1 + "+r"(src_v), // %2 + "+r"(dst_uyvy), // %3 + "+r"(width) // %4 + : + : "cc", "memory", "d0", "d1", "d2", "d3" + ); +} + +void ARGBToRGB565Row_NEON(const uint8* src_argb, uint8* dst_rgb565, int pix) { + asm volatile ( + ".p2align 2 \n" + "1: \n" + MEMACCESS(0) + "vld4.8 {d20, d21, d22, d23}, [%0]! \n" // load 8 pixels of ARGB. + "subs %2, %2, #8 \n" // 8 processed per loop. + ARGBTORGB565 + MEMACCESS(1) + "vst1.8 {q0}, [%1]! \n" // store 8 pixels RGB565. + "bgt 1b \n" + : "+r"(src_argb), // %0 + "+r"(dst_rgb565), // %1 + "+r"(pix) // %2 + : + : "cc", "memory", "q0", "q8", "q9", "q10", "q11" + ); +} + +void ARGBToRGB565DitherRow_NEON(const uint8* src_argb, uint8* dst_rgb, + const uint32 dither4, int width) { + asm volatile ( + ".p2align 2 \n" + "vdup.32 d2, %2 \n" // dither4 + "1: \n" + MEMACCESS(1) + "vld4.8 {d20, d21, d22, d23}, [%1]! \n" // load 8 pixels of ARGB. + "subs %3, %3, #8 \n" // 8 processed per loop. + "vqadd.u8 d20, d20, d2 \n" + "vqadd.u8 d21, d21, d2 \n" + "vqadd.u8 d22, d22, d2 \n" + ARGBTORGB565 + MEMACCESS(0) + "vst1.8 {q0}, [%0]! \n" // store 8 pixels RGB565. + "bgt 1b \n" + : "+r"(dst_rgb) // %0 + : "r"(src_argb), // %1 + "r"(dither4), // %2 + "r"(width) // %3 + : "cc", "memory", "q0", "q1", "q8", "q9", "q10", "q11" + ); +} + +void ARGBToARGB1555Row_NEON(const uint8* src_argb, uint8* dst_argb1555, + int pix) { + asm volatile ( + ".p2align 2 \n" + "1: \n" + MEMACCESS(0) + "vld4.8 {d20, d21, d22, d23}, [%0]! \n" // load 8 pixels of ARGB. + "subs %2, %2, #8 \n" // 8 processed per loop. + ARGBTOARGB1555 + MEMACCESS(1) + "vst1.8 {q0}, [%1]! \n" // store 8 pixels ARGB1555. + "bgt 1b \n" + : "+r"(src_argb), // %0 + "+r"(dst_argb1555), // %1 + "+r"(pix) // %2 + : + : "cc", "memory", "q0", "q8", "q9", "q10", "q11" + ); +} + +void ARGBToARGB4444Row_NEON(const uint8* src_argb, uint8* dst_argb4444, + int pix) { + asm volatile ( + "vmov.u8 d4, #0x0f \n" // bits to clear with vbic. + ".p2align 2 \n" + "1: \n" + MEMACCESS(0) + "vld4.8 {d20, d21, d22, d23}, [%0]! \n" // load 8 pixels of ARGB. + "subs %2, %2, #8 \n" // 8 processed per loop. + ARGBTOARGB4444 + MEMACCESS(1) + "vst1.8 {q0}, [%1]! \n" // store 8 pixels ARGB4444. + "bgt 1b \n" + : "+r"(src_argb), // %0 + "+r"(dst_argb4444), // %1 + "+r"(pix) // %2 + : + : "cc", "memory", "q0", "q8", "q9", "q10", "q11" + ); +} + +void ARGBToYRow_NEON(const uint8* src_argb, uint8* dst_y, int pix) { + asm volatile ( + "vmov.u8 d24, #13 \n" // B * 0.1016 coefficient + "vmov.u8 d25, #65 \n" // G * 0.5078 coefficient + "vmov.u8 d26, #33 \n" // R * 0.2578 coefficient + "vmov.u8 d27, #16 \n" // Add 16 constant + ".p2align 2 \n" + "1: \n" + MEMACCESS(0) + "vld4.8 {d0, d1, d2, d3}, [%0]! \n" // load 8 ARGB pixels. + "subs %2, %2, #8 \n" // 8 processed per loop. + "vmull.u8 q2, d0, d24 \n" // B + "vmlal.u8 q2, d1, d25 \n" // G + "vmlal.u8 q2, d2, d26 \n" // R + "vqrshrun.s16 d0, q2, #7 \n" // 16 bit to 8 bit Y + "vqadd.u8 d0, d27 \n" + MEMACCESS(1) + "vst1.8 {d0}, [%1]! \n" // store 8 pixels Y. + "bgt 1b \n" + : "+r"(src_argb), // %0 + "+r"(dst_y), // %1 + "+r"(pix) // %2 + : + : "cc", "memory", "q0", "q1", "q2", "q12", "q13" + ); +} + +void ARGBToYJRow_NEON(const uint8* src_argb, uint8* dst_y, int pix) { + asm volatile ( + "vmov.u8 d24, #15 \n" // B * 0.11400 coefficient + "vmov.u8 d25, #75 \n" // G * 0.58700 coefficient + "vmov.u8 d26, #38 \n" // R * 0.29900 coefficient + ".p2align 2 \n" + "1: \n" + MEMACCESS(0) + "vld4.8 {d0, d1, d2, d3}, [%0]! \n" // load 8 ARGB pixels. + "subs %2, %2, #8 \n" // 8 processed per loop. + "vmull.u8 q2, d0, d24 \n" // B + "vmlal.u8 q2, d1, d25 \n" // G + "vmlal.u8 q2, d2, d26 \n" // R + "vqrshrun.s16 d0, q2, #7 \n" // 15 bit to 8 bit Y + MEMACCESS(1) + "vst1.8 {d0}, [%1]! \n" // store 8 pixels Y. + "bgt 1b \n" + : "+r"(src_argb), // %0 + "+r"(dst_y), // %1 + "+r"(pix) // %2 + : + : "cc", "memory", "q0", "q1", "q2", "q12", "q13" + ); +} + +// 8x1 pixels. +void ARGBToUV444Row_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v, + int pix) { + asm volatile ( + "vmov.u8 d24, #112 \n" // UB / VR 0.875 coefficient + "vmov.u8 d25, #74 \n" // UG -0.5781 coefficient + "vmov.u8 d26, #38 \n" // UR -0.2969 coefficient + "vmov.u8 d27, #18 \n" // VB -0.1406 coefficient + "vmov.u8 d28, #94 \n" // VG -0.7344 coefficient + "vmov.u16 q15, #0x8080 \n" // 128.5 + ".p2align 2 \n" + "1: \n" + MEMACCESS(0) + "vld4.8 {d0, d1, d2, d3}, [%0]! \n" // load 8 ARGB pixels. + "subs %3, %3, #8 \n" // 8 processed per loop. + "vmull.u8 q2, d0, d24 \n" // B + "vmlsl.u8 q2, d1, d25 \n" // G + "vmlsl.u8 q2, d2, d26 \n" // R + "vadd.u16 q2, q2, q15 \n" // +128 -> unsigned + + "vmull.u8 q3, d2, d24 \n" // R + "vmlsl.u8 q3, d1, d28 \n" // G + "vmlsl.u8 q3, d0, d27 \n" // B + "vadd.u16 q3, q3, q15 \n" // +128 -> unsigned + + "vqshrn.u16 d0, q2, #8 \n" // 16 bit to 8 bit U + "vqshrn.u16 d1, q3, #8 \n" // 16 bit to 8 bit V + + MEMACCESS(1) + "vst1.8 {d0}, [%1]! \n" // store 8 pixels U. + MEMACCESS(2) + "vst1.8 {d1}, [%2]! \n" // store 8 pixels V. + "bgt 1b \n" + : "+r"(src_argb), // %0 + "+r"(dst_u), // %1 + "+r"(dst_v), // %2 + "+r"(pix) // %3 + : + : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q12", "q13", "q14", "q15" + ); +} + +// 16x1 pixels -> 8x1. pix is number of argb pixels. e.g. 16. +void ARGBToUV422Row_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v, + int pix) { + asm volatile ( + "vmov.s16 q10, #112 / 2 \n" // UB / VR 0.875 coefficient + "vmov.s16 q11, #74 / 2 \n" // UG -0.5781 coefficient + "vmov.s16 q12, #38 / 2 \n" // UR -0.2969 coefficient + "vmov.s16 q13, #18 / 2 \n" // VB -0.1406 coefficient + "vmov.s16 q14, #94 / 2 \n" // VG -0.7344 coefficient + "vmov.u16 q15, #0x8080 \n" // 128.5 + ".p2align 2 \n" + "1: \n" + MEMACCESS(0) + "vld4.8 {d0, d2, d4, d6}, [%0]! \n" // load 8 ARGB pixels. + MEMACCESS(0) + "vld4.8 {d1, d3, d5, d7}, [%0]! \n" // load next 8 ARGB pixels. + + "vpaddl.u8 q0, q0 \n" // B 16 bytes -> 8 shorts. + "vpaddl.u8 q1, q1 \n" // G 16 bytes -> 8 shorts. + "vpaddl.u8 q2, q2 \n" // R 16 bytes -> 8 shorts. + + "subs %3, %3, #16 \n" // 16 processed per loop. + "vmul.s16 q8, q0, q10 \n" // B + "vmls.s16 q8, q1, q11 \n" // G + "vmls.s16 q8, q2, q12 \n" // R + "vadd.u16 q8, q8, q15 \n" // +128 -> unsigned + + "vmul.s16 q9, q2, q10 \n" // R + "vmls.s16 q9, q1, q14 \n" // G + "vmls.s16 q9, q0, q13 \n" // B + "vadd.u16 q9, q9, q15 \n" // +128 -> unsigned + + "vqshrn.u16 d0, q8, #8 \n" // 16 bit to 8 bit U + "vqshrn.u16 d1, q9, #8 \n" // 16 bit to 8 bit V + + MEMACCESS(1) + "vst1.8 {d0}, [%1]! \n" // store 8 pixels U. + MEMACCESS(2) + "vst1.8 {d1}, [%2]! \n" // store 8 pixels V. + "bgt 1b \n" + : "+r"(src_argb), // %0 + "+r"(dst_u), // %1 + "+r"(dst_v), // %2 + "+r"(pix) // %3 + : + : "cc", "memory", "q0", "q1", "q2", "q3", + "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15" + ); +} + +// 32x1 pixels -> 8x1. pix is number of argb pixels. e.g. 32. +void ARGBToUV411Row_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v, + int pix) { + asm volatile ( + "vmov.s16 q10, #112 / 2 \n" // UB / VR 0.875 coefficient + "vmov.s16 q11, #74 / 2 \n" // UG -0.5781 coefficient + "vmov.s16 q12, #38 / 2 \n" // UR -0.2969 coefficient + "vmov.s16 q13, #18 / 2 \n" // VB -0.1406 coefficient + "vmov.s16 q14, #94 / 2 \n" // VG -0.7344 coefficient + "vmov.u16 q15, #0x8080 \n" // 128.5 + ".p2align 2 \n" + "1: \n" + MEMACCESS(0) + "vld4.8 {d0, d2, d4, d6}, [%0]! \n" // load 8 ARGB pixels. + MEMACCESS(0) + "vld4.8 {d1, d3, d5, d7}, [%0]! \n" // load next 8 ARGB pixels. + "vpaddl.u8 q0, q0 \n" // B 16 bytes -> 8 shorts. + "vpaddl.u8 q1, q1 \n" // G 16 bytes -> 8 shorts. + "vpaddl.u8 q2, q2 \n" // R 16 bytes -> 8 shorts. + MEMACCESS(0) + "vld4.8 {d8, d10, d12, d14}, [%0]! \n" // load 8 more ARGB pixels. + MEMACCESS(0) + "vld4.8 {d9, d11, d13, d15}, [%0]! \n" // load last 8 ARGB pixels. + "vpaddl.u8 q4, q4 \n" // B 16 bytes -> 8 shorts. + "vpaddl.u8 q5, q5 \n" // G 16 bytes -> 8 shorts. + "vpaddl.u8 q6, q6 \n" // R 16 bytes -> 8 shorts. + + "vpadd.u16 d0, d0, d1 \n" // B 16 shorts -> 8 shorts. + "vpadd.u16 d1, d8, d9 \n" // B + "vpadd.u16 d2, d2, d3 \n" // G 16 shorts -> 8 shorts. + "vpadd.u16 d3, d10, d11 \n" // G + "vpadd.u16 d4, d4, d5 \n" // R 16 shorts -> 8 shorts. + "vpadd.u16 d5, d12, d13 \n" // R + + "vrshr.u16 q0, q0, #1 \n" // 2x average + "vrshr.u16 q1, q1, #1 \n" + "vrshr.u16 q2, q2, #1 \n" + + "subs %3, %3, #32 \n" // 32 processed per loop. + "vmul.s16 q8, q0, q10 \n" // B + "vmls.s16 q8, q1, q11 \n" // G + "vmls.s16 q8, q2, q12 \n" // R + "vadd.u16 q8, q8, q15 \n" // +128 -> unsigned + "vmul.s16 q9, q2, q10 \n" // R + "vmls.s16 q9, q1, q14 \n" // G + "vmls.s16 q9, q0, q13 \n" // B + "vadd.u16 q9, q9, q15 \n" // +128 -> unsigned + "vqshrn.u16 d0, q8, #8 \n" // 16 bit to 8 bit U + "vqshrn.u16 d1, q9, #8 \n" // 16 bit to 8 bit V + MEMACCESS(1) + "vst1.8 {d0}, [%1]! \n" // store 8 pixels U. + MEMACCESS(2) + "vst1.8 {d1}, [%2]! \n" // store 8 pixels V. + "bgt 1b \n" + : "+r"(src_argb), // %0 + "+r"(dst_u), // %1 + "+r"(dst_v), // %2 + "+r"(pix) // %3 + : + : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", + "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15" + ); +} + +// 16x2 pixels -> 8x1. pix is number of argb pixels. e.g. 16. +#define RGBTOUV(QB, QG, QR) \ + "vmul.s16 q8, " #QB ", q10 \n" /* B */ \ + "vmls.s16 q8, " #QG ", q11 \n" /* G */ \ + "vmls.s16 q8, " #QR ", q12 \n" /* R */ \ + "vadd.u16 q8, q8, q15 \n" /* +128 -> unsigned */ \ + "vmul.s16 q9, " #QR ", q10 \n" /* R */ \ + "vmls.s16 q9, " #QG ", q14 \n" /* G */ \ + "vmls.s16 q9, " #QB ", q13 \n" /* B */ \ + "vadd.u16 q9, q9, q15 \n" /* +128 -> unsigned */ \ + "vqshrn.u16 d0, q8, #8 \n" /* 16 bit to 8 bit U */ \ + "vqshrn.u16 d1, q9, #8 \n" /* 16 bit to 8 bit V */ + +// TODO(fbarchard): Consider vhadd vertical, then vpaddl horizontal, avoid shr. +void ARGBToUVRow_NEON(const uint8* src_argb, int src_stride_argb, + uint8* dst_u, uint8* dst_v, int pix) { + asm volatile ( + "add %1, %0, %1 \n" // src_stride + src_argb + "vmov.s16 q10, #112 / 2 \n" // UB / VR 0.875 coefficient + "vmov.s16 q11, #74 / 2 \n" // UG -0.5781 coefficient + "vmov.s16 q12, #38 / 2 \n" // UR -0.2969 coefficient + "vmov.s16 q13, #18 / 2 \n" // VB -0.1406 coefficient + "vmov.s16 q14, #94 / 2 \n" // VG -0.7344 coefficient + "vmov.u16 q15, #0x8080 \n" // 128.5 + ".p2align 2 \n" + "1: \n" + MEMACCESS(0) + "vld4.8 {d0, d2, d4, d6}, [%0]! \n" // load 8 ARGB pixels. + MEMACCESS(0) + "vld4.8 {d1, d3, d5, d7}, [%0]! \n" // load next 8 ARGB pixels. + "vpaddl.u8 q0, q0 \n" // B 16 bytes -> 8 shorts. + "vpaddl.u8 q1, q1 \n" // G 16 bytes -> 8 shorts. + "vpaddl.u8 q2, q2 \n" // R 16 bytes -> 8 shorts. + MEMACCESS(1) + "vld4.8 {d8, d10, d12, d14}, [%1]! \n" // load 8 more ARGB pixels. + MEMACCESS(1) + "vld4.8 {d9, d11, d13, d15}, [%1]! \n" // load last 8 ARGB pixels. + "vpadal.u8 q0, q4 \n" // B 16 bytes -> 8 shorts. + "vpadal.u8 q1, q5 \n" // G 16 bytes -> 8 shorts. + "vpadal.u8 q2, q6 \n" // R 16 bytes -> 8 shorts. + + "vrshr.u16 q0, q0, #1 \n" // 2x average + "vrshr.u16 q1, q1, #1 \n" + "vrshr.u16 q2, q2, #1 \n" + + "subs %4, %4, #16 \n" // 32 processed per loop. + RGBTOUV(q0, q1, q2) + MEMACCESS(2) + "vst1.8 {d0}, [%2]! \n" // store 8 pixels U. + MEMACCESS(3) + "vst1.8 {d1}, [%3]! \n" // store 8 pixels V. + "bgt 1b \n" + : "+r"(src_argb), // %0 + "+r"(src_stride_argb), // %1 + "+r"(dst_u), // %2 + "+r"(dst_v), // %3 + "+r"(pix) // %4 + : + : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", + "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15" + ); +} + +// TODO(fbarchard): Subsample match C code. +void ARGBToUVJRow_NEON(const uint8* src_argb, int src_stride_argb, + uint8* dst_u, uint8* dst_v, int pix) { + asm volatile ( + "add %1, %0, %1 \n" // src_stride + src_argb + "vmov.s16 q10, #127 / 2 \n" // UB / VR 0.500 coefficient + "vmov.s16 q11, #84 / 2 \n" // UG -0.33126 coefficient + "vmov.s16 q12, #43 / 2 \n" // UR -0.16874 coefficient + "vmov.s16 q13, #20 / 2 \n" // VB -0.08131 coefficient + "vmov.s16 q14, #107 / 2 \n" // VG -0.41869 coefficient + "vmov.u16 q15, #0x8080 \n" // 128.5 + ".p2align 2 \n" + "1: \n" + MEMACCESS(0) + "vld4.8 {d0, d2, d4, d6}, [%0]! \n" // load 8 ARGB pixels. + MEMACCESS(0) + "vld4.8 {d1, d3, d5, d7}, [%0]! \n" // load next 8 ARGB pixels. + "vpaddl.u8 q0, q0 \n" // B 16 bytes -> 8 shorts. + "vpaddl.u8 q1, q1 \n" // G 16 bytes -> 8 shorts. + "vpaddl.u8 q2, q2 \n" // R 16 bytes -> 8 shorts. + MEMACCESS(1) + "vld4.8 {d8, d10, d12, d14}, [%1]! \n" // load 8 more ARGB pixels. + MEMACCESS(1) + "vld4.8 {d9, d11, d13, d15}, [%1]! \n" // load last 8 ARGB pixels. + "vpadal.u8 q0, q4 \n" // B 16 bytes -> 8 shorts. + "vpadal.u8 q1, q5 \n" // G 16 bytes -> 8 shorts. + "vpadal.u8 q2, q6 \n" // R 16 bytes -> 8 shorts. + + "vrshr.u16 q0, q0, #1 \n" // 2x average + "vrshr.u16 q1, q1, #1 \n" + "vrshr.u16 q2, q2, #1 \n" + + "subs %4, %4, #16 \n" // 32 processed per loop. + RGBTOUV(q0, q1, q2) + MEMACCESS(2) + "vst1.8 {d0}, [%2]! \n" // store 8 pixels U. + MEMACCESS(3) + "vst1.8 {d1}, [%3]! \n" // store 8 pixels V. + "bgt 1b \n" + : "+r"(src_argb), // %0 + "+r"(src_stride_argb), // %1 + "+r"(dst_u), // %2 + "+r"(dst_v), // %3 + "+r"(pix) // %4 + : + : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", + "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15" + ); +} + +void BGRAToUVRow_NEON(const uint8* src_bgra, int src_stride_bgra, + uint8* dst_u, uint8* dst_v, int pix) { + asm volatile ( + "add %1, %0, %1 \n" // src_stride + src_bgra + "vmov.s16 q10, #112 / 2 \n" // UB / VR 0.875 coefficient + "vmov.s16 q11, #74 / 2 \n" // UG -0.5781 coefficient + "vmov.s16 q12, #38 / 2 \n" // UR -0.2969 coefficient + "vmov.s16 q13, #18 / 2 \n" // VB -0.1406 coefficient + "vmov.s16 q14, #94 / 2 \n" // VG -0.7344 coefficient + "vmov.u16 q15, #0x8080 \n" // 128.5 + ".p2align 2 \n" + "1: \n" + MEMACCESS(0) + "vld4.8 {d0, d2, d4, d6}, [%0]! \n" // load 8 BGRA pixels. + MEMACCESS(0) + "vld4.8 {d1, d3, d5, d7}, [%0]! \n" // load next 8 BGRA pixels. + "vpaddl.u8 q3, q3 \n" // B 16 bytes -> 8 shorts. + "vpaddl.u8 q2, q2 \n" // G 16 bytes -> 8 shorts. + "vpaddl.u8 q1, q1 \n" // R 16 bytes -> 8 shorts. + MEMACCESS(1) + "vld4.8 {d8, d10, d12, d14}, [%1]! \n" // load 8 more BGRA pixels. + MEMACCESS(1) + "vld4.8 {d9, d11, d13, d15}, [%1]! \n" // load last 8 BGRA pixels. + "vpadal.u8 q3, q7 \n" // B 16 bytes -> 8 shorts. + "vpadal.u8 q2, q6 \n" // G 16 bytes -> 8 shorts. + "vpadal.u8 q1, q5 \n" // R 16 bytes -> 8 shorts. + + "vrshr.u16 q1, q1, #1 \n" // 2x average + "vrshr.u16 q2, q2, #1 \n" + "vrshr.u16 q3, q3, #1 \n" + + "subs %4, %4, #16 \n" // 32 processed per loop. + RGBTOUV(q3, q2, q1) + MEMACCESS(2) + "vst1.8 {d0}, [%2]! \n" // store 8 pixels U. + MEMACCESS(3) + "vst1.8 {d1}, [%3]! \n" // store 8 pixels V. + "bgt 1b \n" + : "+r"(src_bgra), // %0 + "+r"(src_stride_bgra), // %1 + "+r"(dst_u), // %2 + "+r"(dst_v), // %3 + "+r"(pix) // %4 + : + : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", + "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15" + ); +} + +void ABGRToUVRow_NEON(const uint8* src_abgr, int src_stride_abgr, + uint8* dst_u, uint8* dst_v, int pix) { + asm volatile ( + "add %1, %0, %1 \n" // src_stride + src_abgr + "vmov.s16 q10, #112 / 2 \n" // UB / VR 0.875 coefficient + "vmov.s16 q11, #74 / 2 \n" // UG -0.5781 coefficient + "vmov.s16 q12, #38 / 2 \n" // UR -0.2969 coefficient + "vmov.s16 q13, #18 / 2 \n" // VB -0.1406 coefficient + "vmov.s16 q14, #94 / 2 \n" // VG -0.7344 coefficient + "vmov.u16 q15, #0x8080 \n" // 128.5 + ".p2align 2 \n" + "1: \n" + MEMACCESS(0) + "vld4.8 {d0, d2, d4, d6}, [%0]! \n" // load 8 ABGR pixels. + MEMACCESS(0) + "vld4.8 {d1, d3, d5, d7}, [%0]! \n" // load next 8 ABGR pixels. + "vpaddl.u8 q2, q2 \n" // B 16 bytes -> 8 shorts. + "vpaddl.u8 q1, q1 \n" // G 16 bytes -> 8 shorts. + "vpaddl.u8 q0, q0 \n" // R 16 bytes -> 8 shorts. + MEMACCESS(1) + "vld4.8 {d8, d10, d12, d14}, [%1]! \n" // load 8 more ABGR pixels. + MEMACCESS(1) + "vld4.8 {d9, d11, d13, d15}, [%1]! \n" // load last 8 ABGR pixels. + "vpadal.u8 q2, q6 \n" // B 16 bytes -> 8 shorts. + "vpadal.u8 q1, q5 \n" // G 16 bytes -> 8 shorts. + "vpadal.u8 q0, q4 \n" // R 16 bytes -> 8 shorts. + + "vrshr.u16 q0, q0, #1 \n" // 2x average + "vrshr.u16 q1, q1, #1 \n" + "vrshr.u16 q2, q2, #1 \n" + + "subs %4, %4, #16 \n" // 32 processed per loop. + RGBTOUV(q2, q1, q0) + MEMACCESS(2) + "vst1.8 {d0}, [%2]! \n" // store 8 pixels U. + MEMACCESS(3) + "vst1.8 {d1}, [%3]! \n" // store 8 pixels V. + "bgt 1b \n" + : "+r"(src_abgr), // %0 + "+r"(src_stride_abgr), // %1 + "+r"(dst_u), // %2 + "+r"(dst_v), // %3 + "+r"(pix) // %4 + : + : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", + "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15" + ); +} + +void RGBAToUVRow_NEON(const uint8* src_rgba, int src_stride_rgba, + uint8* dst_u, uint8* dst_v, int pix) { + asm volatile ( + "add %1, %0, %1 \n" // src_stride + src_rgba + "vmov.s16 q10, #112 / 2 \n" // UB / VR 0.875 coefficient + "vmov.s16 q11, #74 / 2 \n" // UG -0.5781 coefficient + "vmov.s16 q12, #38 / 2 \n" // UR -0.2969 coefficient + "vmov.s16 q13, #18 / 2 \n" // VB -0.1406 coefficient + "vmov.s16 q14, #94 / 2 \n" // VG -0.7344 coefficient + "vmov.u16 q15, #0x8080 \n" // 128.5 + ".p2align 2 \n" + "1: \n" + MEMACCESS(0) + "vld4.8 {d0, d2, d4, d6}, [%0]! \n" // load 8 RGBA pixels. + MEMACCESS(0) + "vld4.8 {d1, d3, d5, d7}, [%0]! \n" // load next 8 RGBA pixels. + "vpaddl.u8 q0, q1 \n" // B 16 bytes -> 8 shorts. + "vpaddl.u8 q1, q2 \n" // G 16 bytes -> 8 shorts. + "vpaddl.u8 q2, q3 \n" // R 16 bytes -> 8 shorts. + MEMACCESS(1) + "vld4.8 {d8, d10, d12, d14}, [%1]! \n" // load 8 more RGBA pixels. + MEMACCESS(1) + "vld4.8 {d9, d11, d13, d15}, [%1]! \n" // load last 8 RGBA pixels. + "vpadal.u8 q0, q5 \n" // B 16 bytes -> 8 shorts. + "vpadal.u8 q1, q6 \n" // G 16 bytes -> 8 shorts. + "vpadal.u8 q2, q7 \n" // R 16 bytes -> 8 shorts. + + "vrshr.u16 q0, q0, #1 \n" // 2x average + "vrshr.u16 q1, q1, #1 \n" + "vrshr.u16 q2, q2, #1 \n" + + "subs %4, %4, #16 \n" // 32 processed per loop. + RGBTOUV(q0, q1, q2) + MEMACCESS(2) + "vst1.8 {d0}, [%2]! \n" // store 8 pixels U. + MEMACCESS(3) + "vst1.8 {d1}, [%3]! \n" // store 8 pixels V. + "bgt 1b \n" + : "+r"(src_rgba), // %0 + "+r"(src_stride_rgba), // %1 + "+r"(dst_u), // %2 + "+r"(dst_v), // %3 + "+r"(pix) // %4 + : + : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", + "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15" + ); +} + +void RGB24ToUVRow_NEON(const uint8* src_rgb24, int src_stride_rgb24, + uint8* dst_u, uint8* dst_v, int pix) { + asm volatile ( + "add %1, %0, %1 \n" // src_stride + src_rgb24 + "vmov.s16 q10, #112 / 2 \n" // UB / VR 0.875 coefficient + "vmov.s16 q11, #74 / 2 \n" // UG -0.5781 coefficient + "vmov.s16 q12, #38 / 2 \n" // UR -0.2969 coefficient + "vmov.s16 q13, #18 / 2 \n" // VB -0.1406 coefficient + "vmov.s16 q14, #94 / 2 \n" // VG -0.7344 coefficient + "vmov.u16 q15, #0x8080 \n" // 128.5 + ".p2align 2 \n" + "1: \n" + MEMACCESS(0) + "vld3.8 {d0, d2, d4}, [%0]! \n" // load 8 RGB24 pixels. + MEMACCESS(0) + "vld3.8 {d1, d3, d5}, [%0]! \n" // load next 8 RGB24 pixels. + "vpaddl.u8 q0, q0 \n" // B 16 bytes -> 8 shorts. + "vpaddl.u8 q1, q1 \n" // G 16 bytes -> 8 shorts. + "vpaddl.u8 q2, q2 \n" // R 16 bytes -> 8 shorts. + MEMACCESS(1) + "vld3.8 {d8, d10, d12}, [%1]! \n" // load 8 more RGB24 pixels. + MEMACCESS(1) + "vld3.8 {d9, d11, d13}, [%1]! \n" // load last 8 RGB24 pixels. + "vpadal.u8 q0, q4 \n" // B 16 bytes -> 8 shorts. + "vpadal.u8 q1, q5 \n" // G 16 bytes -> 8 shorts. + "vpadal.u8 q2, q6 \n" // R 16 bytes -> 8 shorts. + + "vrshr.u16 q0, q0, #1 \n" // 2x average + "vrshr.u16 q1, q1, #1 \n" + "vrshr.u16 q2, q2, #1 \n" + + "subs %4, %4, #16 \n" // 32 processed per loop. + RGBTOUV(q0, q1, q2) + MEMACCESS(2) + "vst1.8 {d0}, [%2]! \n" // store 8 pixels U. + MEMACCESS(3) + "vst1.8 {d1}, [%3]! \n" // store 8 pixels V. + "bgt 1b \n" + : "+r"(src_rgb24), // %0 + "+r"(src_stride_rgb24), // %1 + "+r"(dst_u), // %2 + "+r"(dst_v), // %3 + "+r"(pix) // %4 + : + : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", + "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15" + ); +} + +void RAWToUVRow_NEON(const uint8* src_raw, int src_stride_raw, + uint8* dst_u, uint8* dst_v, int pix) { + asm volatile ( + "add %1, %0, %1 \n" // src_stride + src_raw + "vmov.s16 q10, #112 / 2 \n" // UB / VR 0.875 coefficient + "vmov.s16 q11, #74 / 2 \n" // UG -0.5781 coefficient + "vmov.s16 q12, #38 / 2 \n" // UR -0.2969 coefficient + "vmov.s16 q13, #18 / 2 \n" // VB -0.1406 coefficient + "vmov.s16 q14, #94 / 2 \n" // VG -0.7344 coefficient + "vmov.u16 q15, #0x8080 \n" // 128.5 + ".p2align 2 \n" + "1: \n" + MEMACCESS(0) + "vld3.8 {d0, d2, d4}, [%0]! \n" // load 8 RAW pixels. + MEMACCESS(0) + "vld3.8 {d1, d3, d5}, [%0]! \n" // load next 8 RAW pixels. + "vpaddl.u8 q2, q2 \n" // B 16 bytes -> 8 shorts. + "vpaddl.u8 q1, q1 \n" // G 16 bytes -> 8 shorts. + "vpaddl.u8 q0, q0 \n" // R 16 bytes -> 8 shorts. + MEMACCESS(1) + "vld3.8 {d8, d10, d12}, [%1]! \n" // load 8 more RAW pixels. + MEMACCESS(1) + "vld3.8 {d9, d11, d13}, [%1]! \n" // load last 8 RAW pixels. + "vpadal.u8 q2, q6 \n" // B 16 bytes -> 8 shorts. + "vpadal.u8 q1, q5 \n" // G 16 bytes -> 8 shorts. + "vpadal.u8 q0, q4 \n" // R 16 bytes -> 8 shorts. + + "vrshr.u16 q0, q0, #1 \n" // 2x average + "vrshr.u16 q1, q1, #1 \n" + "vrshr.u16 q2, q2, #1 \n" + + "subs %4, %4, #16 \n" // 32 processed per loop. + RGBTOUV(q2, q1, q0) + MEMACCESS(2) + "vst1.8 {d0}, [%2]! \n" // store 8 pixels U. + MEMACCESS(3) + "vst1.8 {d1}, [%3]! \n" // store 8 pixels V. + "bgt 1b \n" + : "+r"(src_raw), // %0 + "+r"(src_stride_raw), // %1 + "+r"(dst_u), // %2 + "+r"(dst_v), // %3 + "+r"(pix) // %4 + : + : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", + "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15" + ); +} + +// 16x2 pixels -> 8x1. pix is number of argb pixels. e.g. 16. +void RGB565ToUVRow_NEON(const uint8* src_rgb565, int src_stride_rgb565, + uint8* dst_u, uint8* dst_v, int pix) { + asm volatile ( + "add %1, %0, %1 \n" // src_stride + src_argb + "vmov.s16 q10, #112 / 2 \n" // UB / VR 0.875 coefficient + "vmov.s16 q11, #74 / 2 \n" // UG -0.5781 coefficient + "vmov.s16 q12, #38 / 2 \n" // UR -0.2969 coefficient + "vmov.s16 q13, #18 / 2 \n" // VB -0.1406 coefficient + "vmov.s16 q14, #94 / 2 \n" // VG -0.7344 coefficient + "vmov.u16 q15, #0x8080 \n" // 128.5 + ".p2align 2 \n" + "1: \n" + MEMACCESS(0) + "vld1.8 {q0}, [%0]! \n" // load 8 RGB565 pixels. + RGB565TOARGB + "vpaddl.u8 d8, d0 \n" // B 8 bytes -> 4 shorts. + "vpaddl.u8 d10, d1 \n" // G 8 bytes -> 4 shorts. + "vpaddl.u8 d12, d2 \n" // R 8 bytes -> 4 shorts. + MEMACCESS(0) + "vld1.8 {q0}, [%0]! \n" // next 8 RGB565 pixels. + RGB565TOARGB + "vpaddl.u8 d9, d0 \n" // B 8 bytes -> 4 shorts. + "vpaddl.u8 d11, d1 \n" // G 8 bytes -> 4 shorts. + "vpaddl.u8 d13, d2 \n" // R 8 bytes -> 4 shorts. + + MEMACCESS(1) + "vld1.8 {q0}, [%1]! \n" // load 8 RGB565 pixels. + RGB565TOARGB + "vpadal.u8 d8, d0 \n" // B 8 bytes -> 4 shorts. + "vpadal.u8 d10, d1 \n" // G 8 bytes -> 4 shorts. + "vpadal.u8 d12, d2 \n" // R 8 bytes -> 4 shorts. + MEMACCESS(1) + "vld1.8 {q0}, [%1]! \n" // next 8 RGB565 pixels. + RGB565TOARGB + "vpadal.u8 d9, d0 \n" // B 8 bytes -> 4 shorts. + "vpadal.u8 d11, d1 \n" // G 8 bytes -> 4 shorts. + "vpadal.u8 d13, d2 \n" // R 8 bytes -> 4 shorts. + + "vrshr.u16 q4, q4, #1 \n" // 2x average + "vrshr.u16 q5, q5, #1 \n" + "vrshr.u16 q6, q6, #1 \n" + + "subs %4, %4, #16 \n" // 16 processed per loop. + "vmul.s16 q8, q4, q10 \n" // B + "vmls.s16 q8, q5, q11 \n" // G + "vmls.s16 q8, q6, q12 \n" // R + "vadd.u16 q8, q8, q15 \n" // +128 -> unsigned + "vmul.s16 q9, q6, q10 \n" // R + "vmls.s16 q9, q5, q14 \n" // G + "vmls.s16 q9, q4, q13 \n" // B + "vadd.u16 q9, q9, q15 \n" // +128 -> unsigned + "vqshrn.u16 d0, q8, #8 \n" // 16 bit to 8 bit U + "vqshrn.u16 d1, q9, #8 \n" // 16 bit to 8 bit V + MEMACCESS(2) + "vst1.8 {d0}, [%2]! \n" // store 8 pixels U. + MEMACCESS(3) + "vst1.8 {d1}, [%3]! \n" // store 8 pixels V. + "bgt 1b \n" + : "+r"(src_rgb565), // %0 + "+r"(src_stride_rgb565), // %1 + "+r"(dst_u), // %2 + "+r"(dst_v), // %3 + "+r"(pix) // %4 + : + : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", + "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15" + ); +} + +// 16x2 pixels -> 8x1. pix is number of argb pixels. e.g. 16. +void ARGB1555ToUVRow_NEON(const uint8* src_argb1555, int src_stride_argb1555, + uint8* dst_u, uint8* dst_v, int pix) { + asm volatile ( + "add %1, %0, %1 \n" // src_stride + src_argb + "vmov.s16 q10, #112 / 2 \n" // UB / VR 0.875 coefficient + "vmov.s16 q11, #74 / 2 \n" // UG -0.5781 coefficient + "vmov.s16 q12, #38 / 2 \n" // UR -0.2969 coefficient + "vmov.s16 q13, #18 / 2 \n" // VB -0.1406 coefficient + "vmov.s16 q14, #94 / 2 \n" // VG -0.7344 coefficient + "vmov.u16 q15, #0x8080 \n" // 128.5 + ".p2align 2 \n" + "1: \n" + MEMACCESS(0) + "vld1.8 {q0}, [%0]! \n" // load 8 ARGB1555 pixels. + RGB555TOARGB + "vpaddl.u8 d8, d0 \n" // B 8 bytes -> 4 shorts. + "vpaddl.u8 d10, d1 \n" // G 8 bytes -> 4 shorts. + "vpaddl.u8 d12, d2 \n" // R 8 bytes -> 4 shorts. + MEMACCESS(0) + "vld1.8 {q0}, [%0]! \n" // next 8 ARGB1555 pixels. + RGB555TOARGB + "vpaddl.u8 d9, d0 \n" // B 8 bytes -> 4 shorts. + "vpaddl.u8 d11, d1 \n" // G 8 bytes -> 4 shorts. + "vpaddl.u8 d13, d2 \n" // R 8 bytes -> 4 shorts. + + MEMACCESS(1) + "vld1.8 {q0}, [%1]! \n" // load 8 ARGB1555 pixels. + RGB555TOARGB + "vpadal.u8 d8, d0 \n" // B 8 bytes -> 4 shorts. + "vpadal.u8 d10, d1 \n" // G 8 bytes -> 4 shorts. + "vpadal.u8 d12, d2 \n" // R 8 bytes -> 4 shorts. + MEMACCESS(1) + "vld1.8 {q0}, [%1]! \n" // next 8 ARGB1555 pixels. + RGB555TOARGB + "vpadal.u8 d9, d0 \n" // B 8 bytes -> 4 shorts. + "vpadal.u8 d11, d1 \n" // G 8 bytes -> 4 shorts. + "vpadal.u8 d13, d2 \n" // R 8 bytes -> 4 shorts. + + "vrshr.u16 q4, q4, #1 \n" // 2x average + "vrshr.u16 q5, q5, #1 \n" + "vrshr.u16 q6, q6, #1 \n" + + "subs %4, %4, #16 \n" // 16 processed per loop. + "vmul.s16 q8, q4, q10 \n" // B + "vmls.s16 q8, q5, q11 \n" // G + "vmls.s16 q8, q6, q12 \n" // R + "vadd.u16 q8, q8, q15 \n" // +128 -> unsigned + "vmul.s16 q9, q6, q10 \n" // R + "vmls.s16 q9, q5, q14 \n" // G + "vmls.s16 q9, q4, q13 \n" // B + "vadd.u16 q9, q9, q15 \n" // +128 -> unsigned + "vqshrn.u16 d0, q8, #8 \n" // 16 bit to 8 bit U + "vqshrn.u16 d1, q9, #8 \n" // 16 bit to 8 bit V + MEMACCESS(2) + "vst1.8 {d0}, [%2]! \n" // store 8 pixels U. + MEMACCESS(3) + "vst1.8 {d1}, [%3]! \n" // store 8 pixels V. + "bgt 1b \n" + : "+r"(src_argb1555), // %0 + "+r"(src_stride_argb1555), // %1 + "+r"(dst_u), // %2 + "+r"(dst_v), // %3 + "+r"(pix) // %4 + : + : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", + "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15" + ); +} + +// 16x2 pixels -> 8x1. pix is number of argb pixels. e.g. 16. +void ARGB4444ToUVRow_NEON(const uint8* src_argb4444, int src_stride_argb4444, + uint8* dst_u, uint8* dst_v, int pix) { + asm volatile ( + "add %1, %0, %1 \n" // src_stride + src_argb + "vmov.s16 q10, #112 / 2 \n" // UB / VR 0.875 coefficient + "vmov.s16 q11, #74 / 2 \n" // UG -0.5781 coefficient + "vmov.s16 q12, #38 / 2 \n" // UR -0.2969 coefficient + "vmov.s16 q13, #18 / 2 \n" // VB -0.1406 coefficient + "vmov.s16 q14, #94 / 2 \n" // VG -0.7344 coefficient + "vmov.u16 q15, #0x8080 \n" // 128.5 + ".p2align 2 \n" + "1: \n" + MEMACCESS(0) + "vld1.8 {q0}, [%0]! \n" // load 8 ARGB4444 pixels. + ARGB4444TOARGB + "vpaddl.u8 d8, d0 \n" // B 8 bytes -> 4 shorts. + "vpaddl.u8 d10, d1 \n" // G 8 bytes -> 4 shorts. + "vpaddl.u8 d12, d2 \n" // R 8 bytes -> 4 shorts. + MEMACCESS(0) + "vld1.8 {q0}, [%0]! \n" // next 8 ARGB4444 pixels. + ARGB4444TOARGB + "vpaddl.u8 d9, d0 \n" // B 8 bytes -> 4 shorts. + "vpaddl.u8 d11, d1 \n" // G 8 bytes -> 4 shorts. + "vpaddl.u8 d13, d2 \n" // R 8 bytes -> 4 shorts. + + MEMACCESS(1) + "vld1.8 {q0}, [%1]! \n" // load 8 ARGB4444 pixels. + ARGB4444TOARGB + "vpadal.u8 d8, d0 \n" // B 8 bytes -> 4 shorts. + "vpadal.u8 d10, d1 \n" // G 8 bytes -> 4 shorts. + "vpadal.u8 d12, d2 \n" // R 8 bytes -> 4 shorts. + MEMACCESS(1) + "vld1.8 {q0}, [%1]! \n" // next 8 ARGB4444 pixels. + ARGB4444TOARGB + "vpadal.u8 d9, d0 \n" // B 8 bytes -> 4 shorts. + "vpadal.u8 d11, d1 \n" // G 8 bytes -> 4 shorts. + "vpadal.u8 d13, d2 \n" // R 8 bytes -> 4 shorts. + + "vrshr.u16 q4, q4, #1 \n" // 2x average + "vrshr.u16 q5, q5, #1 \n" + "vrshr.u16 q6, q6, #1 \n" + + "subs %4, %4, #16 \n" // 16 processed per loop. + "vmul.s16 q8, q4, q10 \n" // B + "vmls.s16 q8, q5, q11 \n" // G + "vmls.s16 q8, q6, q12 \n" // R + "vadd.u16 q8, q8, q15 \n" // +128 -> unsigned + "vmul.s16 q9, q6, q10 \n" // R + "vmls.s16 q9, q5, q14 \n" // G + "vmls.s16 q9, q4, q13 \n" // B + "vadd.u16 q9, q9, q15 \n" // +128 -> unsigned + "vqshrn.u16 d0, q8, #8 \n" // 16 bit to 8 bit U + "vqshrn.u16 d1, q9, #8 \n" // 16 bit to 8 bit V + MEMACCESS(2) + "vst1.8 {d0}, [%2]! \n" // store 8 pixels U. + MEMACCESS(3) + "vst1.8 {d1}, [%3]! \n" // store 8 pixels V. + "bgt 1b \n" + : "+r"(src_argb4444), // %0 + "+r"(src_stride_argb4444), // %1 + "+r"(dst_u), // %2 + "+r"(dst_v), // %3 + "+r"(pix) // %4 + : + : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", + "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15" + ); +} + +void RGB565ToYRow_NEON(const uint8* src_rgb565, uint8* dst_y, int pix) { + asm volatile ( + "vmov.u8 d24, #13 \n" // B * 0.1016 coefficient + "vmov.u8 d25, #65 \n" // G * 0.5078 coefficient + "vmov.u8 d26, #33 \n" // R * 0.2578 coefficient + "vmov.u8 d27, #16 \n" // Add 16 constant + ".p2align 2 \n" + "1: \n" + MEMACCESS(0) + "vld1.8 {q0}, [%0]! \n" // load 8 RGB565 pixels. + "subs %2, %2, #8 \n" // 8 processed per loop. + RGB565TOARGB + "vmull.u8 q2, d0, d24 \n" // B + "vmlal.u8 q2, d1, d25 \n" // G + "vmlal.u8 q2, d2, d26 \n" // R + "vqrshrun.s16 d0, q2, #7 \n" // 16 bit to 8 bit Y + "vqadd.u8 d0, d27 \n" + MEMACCESS(1) + "vst1.8 {d0}, [%1]! \n" // store 8 pixels Y. + "bgt 1b \n" + : "+r"(src_rgb565), // %0 + "+r"(dst_y), // %1 + "+r"(pix) // %2 + : + : "cc", "memory", "q0", "q1", "q2", "q3", "q12", "q13" + ); +} + +void ARGB1555ToYRow_NEON(const uint8* src_argb1555, uint8* dst_y, int pix) { + asm volatile ( + "vmov.u8 d24, #13 \n" // B * 0.1016 coefficient + "vmov.u8 d25, #65 \n" // G * 0.5078 coefficient + "vmov.u8 d26, #33 \n" // R * 0.2578 coefficient + "vmov.u8 d27, #16 \n" // Add 16 constant + ".p2align 2 \n" + "1: \n" + MEMACCESS(0) + "vld1.8 {q0}, [%0]! \n" // load 8 ARGB1555 pixels. + "subs %2, %2, #8 \n" // 8 processed per loop. + ARGB1555TOARGB + "vmull.u8 q2, d0, d24 \n" // B + "vmlal.u8 q2, d1, d25 \n" // G + "vmlal.u8 q2, d2, d26 \n" // R + "vqrshrun.s16 d0, q2, #7 \n" // 16 bit to 8 bit Y + "vqadd.u8 d0, d27 \n" + MEMACCESS(1) + "vst1.8 {d0}, [%1]! \n" // store 8 pixels Y. + "bgt 1b \n" + : "+r"(src_argb1555), // %0 + "+r"(dst_y), // %1 + "+r"(pix) // %2 + : + : "cc", "memory", "q0", "q1", "q2", "q3", "q12", "q13" + ); +} + +void ARGB4444ToYRow_NEON(const uint8* src_argb4444, uint8* dst_y, int pix) { + asm volatile ( + "vmov.u8 d24, #13 \n" // B * 0.1016 coefficient + "vmov.u8 d25, #65 \n" // G * 0.5078 coefficient + "vmov.u8 d26, #33 \n" // R * 0.2578 coefficient + "vmov.u8 d27, #16 \n" // Add 16 constant + ".p2align 2 \n" + "1: \n" + MEMACCESS(0) + "vld1.8 {q0}, [%0]! \n" // load 8 ARGB4444 pixels. + "subs %2, %2, #8 \n" // 8 processed per loop. + ARGB4444TOARGB + "vmull.u8 q2, d0, d24 \n" // B + "vmlal.u8 q2, d1, d25 \n" // G + "vmlal.u8 q2, d2, d26 \n" // R + "vqrshrun.s16 d0, q2, #7 \n" // 16 bit to 8 bit Y + "vqadd.u8 d0, d27 \n" + MEMACCESS(1) + "vst1.8 {d0}, [%1]! \n" // store 8 pixels Y. + "bgt 1b \n" + : "+r"(src_argb4444), // %0 + "+r"(dst_y), // %1 + "+r"(pix) // %2 + : + : "cc", "memory", "q0", "q1", "q2", "q3", "q12", "q13" + ); +} + +void BGRAToYRow_NEON(const uint8* src_bgra, uint8* dst_y, int pix) { + asm volatile ( + "vmov.u8 d4, #33 \n" // R * 0.2578 coefficient + "vmov.u8 d5, #65 \n" // G * 0.5078 coefficient + "vmov.u8 d6, #13 \n" // B * 0.1016 coefficient + "vmov.u8 d7, #16 \n" // Add 16 constant + ".p2align 2 \n" + "1: \n" + MEMACCESS(0) + "vld4.8 {d0, d1, d2, d3}, [%0]! \n" // load 8 pixels of BGRA. + "subs %2, %2, #8 \n" // 8 processed per loop. + "vmull.u8 q8, d1, d4 \n" // R + "vmlal.u8 q8, d2, d5 \n" // G + "vmlal.u8 q8, d3, d6 \n" // B + "vqrshrun.s16 d0, q8, #7 \n" // 16 bit to 8 bit Y + "vqadd.u8 d0, d7 \n" + MEMACCESS(1) + "vst1.8 {d0}, [%1]! \n" // store 8 pixels Y. + "bgt 1b \n" + : "+r"(src_bgra), // %0 + "+r"(dst_y), // %1 + "+r"(pix) // %2 + : + : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7", "q8" + ); +} + +void ABGRToYRow_NEON(const uint8* src_abgr, uint8* dst_y, int pix) { + asm volatile ( + "vmov.u8 d4, #33 \n" // R * 0.2578 coefficient + "vmov.u8 d5, #65 \n" // G * 0.5078 coefficient + "vmov.u8 d6, #13 \n" // B * 0.1016 coefficient + "vmov.u8 d7, #16 \n" // Add 16 constant + ".p2align 2 \n" + "1: \n" + MEMACCESS(0) + "vld4.8 {d0, d1, d2, d3}, [%0]! \n" // load 8 pixels of ABGR. + "subs %2, %2, #8 \n" // 8 processed per loop. + "vmull.u8 q8, d0, d4 \n" // R + "vmlal.u8 q8, d1, d5 \n" // G + "vmlal.u8 q8, d2, d6 \n" // B + "vqrshrun.s16 d0, q8, #7 \n" // 16 bit to 8 bit Y + "vqadd.u8 d0, d7 \n" + MEMACCESS(1) + "vst1.8 {d0}, [%1]! \n" // store 8 pixels Y. + "bgt 1b \n" + : "+r"(src_abgr), // %0 + "+r"(dst_y), // %1 + "+r"(pix) // %2 + : + : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7", "q8" + ); +} + +void RGBAToYRow_NEON(const uint8* src_rgba, uint8* dst_y, int pix) { + asm volatile ( + "vmov.u8 d4, #13 \n" // B * 0.1016 coefficient + "vmov.u8 d5, #65 \n" // G * 0.5078 coefficient + "vmov.u8 d6, #33 \n" // R * 0.2578 coefficient + "vmov.u8 d7, #16 \n" // Add 16 constant + ".p2align 2 \n" + "1: \n" + MEMACCESS(0) + "vld4.8 {d0, d1, d2, d3}, [%0]! \n" // load 8 pixels of RGBA. + "subs %2, %2, #8 \n" // 8 processed per loop. + "vmull.u8 q8, d1, d4 \n" // B + "vmlal.u8 q8, d2, d5 \n" // G + "vmlal.u8 q8, d3, d6 \n" // R + "vqrshrun.s16 d0, q8, #7 \n" // 16 bit to 8 bit Y + "vqadd.u8 d0, d7 \n" + MEMACCESS(1) + "vst1.8 {d0}, [%1]! \n" // store 8 pixels Y. + "bgt 1b \n" + : "+r"(src_rgba), // %0 + "+r"(dst_y), // %1 + "+r"(pix) // %2 + : + : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7", "q8" + ); +} + +void RGB24ToYRow_NEON(const uint8* src_rgb24, uint8* dst_y, int pix) { + asm volatile ( + "vmov.u8 d4, #13 \n" // B * 0.1016 coefficient + "vmov.u8 d5, #65 \n" // G * 0.5078 coefficient + "vmov.u8 d6, #33 \n" // R * 0.2578 coefficient + "vmov.u8 d7, #16 \n" // Add 16 constant + ".p2align 2 \n" + "1: \n" + MEMACCESS(0) + "vld3.8 {d0, d1, d2}, [%0]! \n" // load 8 pixels of RGB24. + "subs %2, %2, #8 \n" // 8 processed per loop. + "vmull.u8 q8, d0, d4 \n" // B + "vmlal.u8 q8, d1, d5 \n" // G + "vmlal.u8 q8, d2, d6 \n" // R + "vqrshrun.s16 d0, q8, #7 \n" // 16 bit to 8 bit Y + "vqadd.u8 d0, d7 \n" + MEMACCESS(1) + "vst1.8 {d0}, [%1]! \n" // store 8 pixels Y. + "bgt 1b \n" + : "+r"(src_rgb24), // %0 + "+r"(dst_y), // %1 + "+r"(pix) // %2 + : + : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7", "q8" + ); +} + +void RAWToYRow_NEON(const uint8* src_raw, uint8* dst_y, int pix) { + asm volatile ( + "vmov.u8 d4, #33 \n" // R * 0.2578 coefficient + "vmov.u8 d5, #65 \n" // G * 0.5078 coefficient + "vmov.u8 d6, #13 \n" // B * 0.1016 coefficient + "vmov.u8 d7, #16 \n" // Add 16 constant + ".p2align 2 \n" + "1: \n" + MEMACCESS(0) + "vld3.8 {d0, d1, d2}, [%0]! \n" // load 8 pixels of RAW. + "subs %2, %2, #8 \n" // 8 processed per loop. + "vmull.u8 q8, d0, d4 \n" // B + "vmlal.u8 q8, d1, d5 \n" // G + "vmlal.u8 q8, d2, d6 \n" // R + "vqrshrun.s16 d0, q8, #7 \n" // 16 bit to 8 bit Y + "vqadd.u8 d0, d7 \n" + MEMACCESS(1) + "vst1.8 {d0}, [%1]! \n" // store 8 pixels Y. + "bgt 1b \n" + : "+r"(src_raw), // %0 + "+r"(dst_y), // %1 + "+r"(pix) // %2 + : + : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7", "q8" + ); +} + +// Bilinear filter 16x2 -> 16x1 +void InterpolateRow_NEON(uint8* dst_ptr, + const uint8* src_ptr, ptrdiff_t src_stride, + int dst_width, int source_y_fraction) { + asm volatile ( + "cmp %4, #0 \n" + "beq 100f \n" + "add %2, %1 \n" + "cmp %4, #64 \n" + "beq 75f \n" + "cmp %4, #128 \n" + "beq 50f \n" + "cmp %4, #192 \n" + "beq 25f \n" + + "vdup.8 d5, %4 \n" + "rsb %4, #256 \n" + "vdup.8 d4, %4 \n" + // General purpose row blend. + "1: \n" + MEMACCESS(1) + "vld1.8 {q0}, [%1]! \n" + MEMACCESS(2) + "vld1.8 {q1}, [%2]! \n" + "subs %3, %3, #16 \n" + "vmull.u8 q13, d0, d4 \n" + "vmull.u8 q14, d1, d4 \n" + "vmlal.u8 q13, d2, d5 \n" + "vmlal.u8 q14, d3, d5 \n" + "vrshrn.u16 d0, q13, #8 \n" + "vrshrn.u16 d1, q14, #8 \n" + MEMACCESS(0) + "vst1.8 {q0}, [%0]! \n" + "bgt 1b \n" + "b 99f \n" + + // Blend 25 / 75. + "25: \n" + MEMACCESS(1) + "vld1.8 {q0}, [%1]! \n" + MEMACCESS(2) + "vld1.8 {q1}, [%2]! \n" + "subs %3, %3, #16 \n" + "vrhadd.u8 q0, q1 \n" + "vrhadd.u8 q0, q1 \n" + MEMACCESS(0) + "vst1.8 {q0}, [%0]! \n" + "bgt 25b \n" + "b 99f \n" + + // Blend 50 / 50. + "50: \n" + MEMACCESS(1) + "vld1.8 {q0}, [%1]! \n" + MEMACCESS(2) + "vld1.8 {q1}, [%2]! \n" + "subs %3, %3, #16 \n" + "vrhadd.u8 q0, q1 \n" + MEMACCESS(0) + "vst1.8 {q0}, [%0]! \n" + "bgt 50b \n" + "b 99f \n" + + // Blend 75 / 25. + "75: \n" + MEMACCESS(1) + "vld1.8 {q1}, [%1]! \n" + MEMACCESS(2) + "vld1.8 {q0}, [%2]! \n" + "subs %3, %3, #16 \n" + "vrhadd.u8 q0, q1 \n" + "vrhadd.u8 q0, q1 \n" + MEMACCESS(0) + "vst1.8 {q0}, [%0]! \n" + "bgt 75b \n" + "b 99f \n" + + // Blend 100 / 0 - Copy row unchanged. + "100: \n" + MEMACCESS(1) + "vld1.8 {q0}, [%1]! \n" + "subs %3, %3, #16 \n" + MEMACCESS(0) + "vst1.8 {q0}, [%0]! \n" + "bgt 100b \n" + + "99: \n" + : "+r"(dst_ptr), // %0 + "+r"(src_ptr), // %1 + "+r"(src_stride), // %2 + "+r"(dst_width), // %3 + "+r"(source_y_fraction) // %4 + : + : "cc", "memory", "q0", "q1", "d4", "d5", "q13", "q14" + ); +} + +// dr * (256 - sa) / 256 + sr = dr - dr * sa / 256 + sr +void ARGBBlendRow_NEON(const uint8* src_argb0, const uint8* src_argb1, + uint8* dst_argb, int width) { + asm volatile ( + "subs %3, #8 \n" + "blt 89f \n" + // Blend 8 pixels. + "8: \n" + MEMACCESS(0) + "vld4.8 {d0, d1, d2, d3}, [%0]! \n" // load 8 pixels of ARGB0. + MEMACCESS(1) + "vld4.8 {d4, d5, d6, d7}, [%1]! \n" // load 8 pixels of ARGB1. + "subs %3, %3, #8 \n" // 8 processed per loop. + "vmull.u8 q10, d4, d3 \n" // db * a + "vmull.u8 q11, d5, d3 \n" // dg * a + "vmull.u8 q12, d6, d3 \n" // dr * a + "vqrshrn.u16 d20, q10, #8 \n" // db >>= 8 + "vqrshrn.u16 d21, q11, #8 \n" // dg >>= 8 + "vqrshrn.u16 d22, q12, #8 \n" // dr >>= 8 + "vqsub.u8 q2, q2, q10 \n" // dbg - dbg * a / 256 + "vqsub.u8 d6, d6, d22 \n" // dr - dr * a / 256 + "vqadd.u8 q0, q0, q2 \n" // + sbg + "vqadd.u8 d2, d2, d6 \n" // + sr + "vmov.u8 d3, #255 \n" // a = 255 + MEMACCESS(2) + "vst4.8 {d0, d1, d2, d3}, [%2]! \n" // store 8 pixels of ARGB. + "bge 8b \n" + + "89: \n" + "adds %3, #8-1 \n" + "blt 99f \n" + + // Blend 1 pixels. + "1: \n" + MEMACCESS(0) + "vld4.8 {d0[0],d1[0],d2[0],d3[0]}, [%0]! \n" // load 1 pixel ARGB0. + MEMACCESS(1) + "vld4.8 {d4[0],d5[0],d6[0],d7[0]}, [%1]! \n" // load 1 pixel ARGB1. + "subs %3, %3, #1 \n" // 1 processed per loop. + "vmull.u8 q10, d4, d3 \n" // db * a + "vmull.u8 q11, d5, d3 \n" // dg * a + "vmull.u8 q12, d6, d3 \n" // dr * a + "vqrshrn.u16 d20, q10, #8 \n" // db >>= 8 + "vqrshrn.u16 d21, q11, #8 \n" // dg >>= 8 + "vqrshrn.u16 d22, q12, #8 \n" // dr >>= 8 + "vqsub.u8 q2, q2, q10 \n" // dbg - dbg * a / 256 + "vqsub.u8 d6, d6, d22 \n" // dr - dr * a / 256 + "vqadd.u8 q0, q0, q2 \n" // + sbg + "vqadd.u8 d2, d2, d6 \n" // + sr + "vmov.u8 d3, #255 \n" // a = 255 + MEMACCESS(2) + "vst4.8 {d0[0],d1[0],d2[0],d3[0]}, [%2]! \n" // store 1 pixel. + "bge 1b \n" + + "99: \n" + + : "+r"(src_argb0), // %0 + "+r"(src_argb1), // %1 + "+r"(dst_argb), // %2 + "+r"(width) // %3 + : + : "cc", "memory", "q0", "q1", "q2", "q3", "q10", "q11", "q12" + ); +} + +// Attenuate 8 pixels at a time. +void ARGBAttenuateRow_NEON(const uint8* src_argb, uint8* dst_argb, int width) { + asm volatile ( + // Attenuate 8 pixels. + "1: \n" + MEMACCESS(0) + "vld4.8 {d0, d1, d2, d3}, [%0]! \n" // load 8 pixels of ARGB. + "subs %2, %2, #8 \n" // 8 processed per loop. + "vmull.u8 q10, d0, d3 \n" // b * a + "vmull.u8 q11, d1, d3 \n" // g * a + "vmull.u8 q12, d2, d3 \n" // r * a + "vqrshrn.u16 d0, q10, #8 \n" // b >>= 8 + "vqrshrn.u16 d1, q11, #8 \n" // g >>= 8 + "vqrshrn.u16 d2, q12, #8 \n" // r >>= 8 + MEMACCESS(1) + "vst4.8 {d0, d1, d2, d3}, [%1]! \n" // store 8 pixels of ARGB. + "bgt 1b \n" + : "+r"(src_argb), // %0 + "+r"(dst_argb), // %1 + "+r"(width) // %2 + : + : "cc", "memory", "q0", "q1", "q10", "q11", "q12" + ); +} + +// Quantize 8 ARGB pixels (32 bytes). +// dst = (dst * scale >> 16) * interval_size + interval_offset; +void ARGBQuantizeRow_NEON(uint8* dst_argb, int scale, int interval_size, + int interval_offset, int width) { + asm volatile ( + "vdup.u16 q8, %2 \n" + "vshr.u16 q8, q8, #1 \n" // scale >>= 1 + "vdup.u16 q9, %3 \n" // interval multiply. + "vdup.u16 q10, %4 \n" // interval add + + // 8 pixel loop. + ".p2align 2 \n" + "1: \n" + MEMACCESS(0) + "vld4.8 {d0, d2, d4, d6}, [%0] \n" // load 8 pixels of ARGB. + "subs %1, %1, #8 \n" // 8 processed per loop. + "vmovl.u8 q0, d0 \n" // b (0 .. 255) + "vmovl.u8 q1, d2 \n" + "vmovl.u8 q2, d4 \n" + "vqdmulh.s16 q0, q0, q8 \n" // b * scale + "vqdmulh.s16 q1, q1, q8 \n" // g + "vqdmulh.s16 q2, q2, q8 \n" // r + "vmul.u16 q0, q0, q9 \n" // b * interval_size + "vmul.u16 q1, q1, q9 \n" // g + "vmul.u16 q2, q2, q9 \n" // r + "vadd.u16 q0, q0, q10 \n" // b + interval_offset + "vadd.u16 q1, q1, q10 \n" // g + "vadd.u16 q2, q2, q10 \n" // r + "vqmovn.u16 d0, q0 \n" + "vqmovn.u16 d2, q1 \n" + "vqmovn.u16 d4, q2 \n" + MEMACCESS(0) + "vst4.8 {d0, d2, d4, d6}, [%0]! \n" // store 8 pixels of ARGB. + "bgt 1b \n" + : "+r"(dst_argb), // %0 + "+r"(width) // %1 + : "r"(scale), // %2 + "r"(interval_size), // %3 + "r"(interval_offset) // %4 + : "cc", "memory", "q0", "q1", "q2", "q3", "q8", "q9", "q10" + ); +} + +// Shade 8 pixels at a time by specified value. +// NOTE vqrdmulh.s16 q10, q10, d0[0] must use a scaler register from 0 to 8. +// Rounding in vqrdmulh does +1 to high if high bit of low s16 is set. +void ARGBShadeRow_NEON(const uint8* src_argb, uint8* dst_argb, int width, + uint32 value) { + asm volatile ( + "vdup.u32 q0, %3 \n" // duplicate scale value. + "vzip.u8 d0, d1 \n" // d0 aarrggbb. + "vshr.u16 q0, q0, #1 \n" // scale / 2. + + // 8 pixel loop. + ".p2align 2 \n" + "1: \n" + MEMACCESS(0) + "vld4.8 {d20, d22, d24, d26}, [%0]! \n" // load 8 pixels of ARGB. + "subs %2, %2, #8 \n" // 8 processed per loop. + "vmovl.u8 q10, d20 \n" // b (0 .. 255) + "vmovl.u8 q11, d22 \n" + "vmovl.u8 q12, d24 \n" + "vmovl.u8 q13, d26 \n" + "vqrdmulh.s16 q10, q10, d0[0] \n" // b * scale * 2 + "vqrdmulh.s16 q11, q11, d0[1] \n" // g + "vqrdmulh.s16 q12, q12, d0[2] \n" // r + "vqrdmulh.s16 q13, q13, d0[3] \n" // a + "vqmovn.u16 d20, q10 \n" + "vqmovn.u16 d22, q11 \n" + "vqmovn.u16 d24, q12 \n" + "vqmovn.u16 d26, q13 \n" + MEMACCESS(1) + "vst4.8 {d20, d22, d24, d26}, [%1]! \n" // store 8 pixels of ARGB. + "bgt 1b \n" + : "+r"(src_argb), // %0 + "+r"(dst_argb), // %1 + "+r"(width) // %2 + : "r"(value) // %3 + : "cc", "memory", "q0", "q10", "q11", "q12", "q13" + ); +} + +// Convert 8 ARGB pixels (64 bytes) to 8 Gray ARGB pixels +// Similar to ARGBToYJ but stores ARGB. +// C code is (15 * b + 75 * g + 38 * r + 64) >> 7; +void ARGBGrayRow_NEON(const uint8* src_argb, uint8* dst_argb, int width) { + asm volatile ( + "vmov.u8 d24, #15 \n" // B * 0.11400 coefficient + "vmov.u8 d25, #75 \n" // G * 0.58700 coefficient + "vmov.u8 d26, #38 \n" // R * 0.29900 coefficient + ".p2align 2 \n" + "1: \n" + MEMACCESS(0) + "vld4.8 {d0, d1, d2, d3}, [%0]! \n" // load 8 ARGB pixels. + "subs %2, %2, #8 \n" // 8 processed per loop. + "vmull.u8 q2, d0, d24 \n" // B + "vmlal.u8 q2, d1, d25 \n" // G + "vmlal.u8 q2, d2, d26 \n" // R + "vqrshrun.s16 d0, q2, #7 \n" // 15 bit to 8 bit B + "vmov d1, d0 \n" // G + "vmov d2, d0 \n" // R + MEMACCESS(1) + "vst4.8 {d0, d1, d2, d3}, [%1]! \n" // store 8 ARGB pixels. + "bgt 1b \n" + : "+r"(src_argb), // %0 + "+r"(dst_argb), // %1 + "+r"(width) // %2 + : + : "cc", "memory", "q0", "q1", "q2", "q12", "q13" + ); +} + +// Convert 8 ARGB pixels (32 bytes) to 8 Sepia ARGB pixels. +// b = (r * 35 + g * 68 + b * 17) >> 7 +// g = (r * 45 + g * 88 + b * 22) >> 7 +// r = (r * 50 + g * 98 + b * 24) >> 7 +void ARGBSepiaRow_NEON(uint8* dst_argb, int width) { + asm volatile ( + "vmov.u8 d20, #17 \n" // BB coefficient + "vmov.u8 d21, #68 \n" // BG coefficient + "vmov.u8 d22, #35 \n" // BR coefficient + "vmov.u8 d24, #22 \n" // GB coefficient + "vmov.u8 d25, #88 \n" // GG coefficient + "vmov.u8 d26, #45 \n" // GR coefficient + "vmov.u8 d28, #24 \n" // BB coefficient + "vmov.u8 d29, #98 \n" // BG coefficient + "vmov.u8 d30, #50 \n" // BR coefficient + ".p2align 2 \n" + "1: \n" + MEMACCESS(0) + "vld4.8 {d0, d1, d2, d3}, [%0] \n" // load 8 ARGB pixels. + "subs %1, %1, #8 \n" // 8 processed per loop. + "vmull.u8 q2, d0, d20 \n" // B to Sepia B + "vmlal.u8 q2, d1, d21 \n" // G + "vmlal.u8 q2, d2, d22 \n" // R + "vmull.u8 q3, d0, d24 \n" // B to Sepia G + "vmlal.u8 q3, d1, d25 \n" // G + "vmlal.u8 q3, d2, d26 \n" // R + "vmull.u8 q8, d0, d28 \n" // B to Sepia R + "vmlal.u8 q8, d1, d29 \n" // G + "vmlal.u8 q8, d2, d30 \n" // R + "vqshrn.u16 d0, q2, #7 \n" // 16 bit to 8 bit B + "vqshrn.u16 d1, q3, #7 \n" // 16 bit to 8 bit G + "vqshrn.u16 d2, q8, #7 \n" // 16 bit to 8 bit R + MEMACCESS(0) + "vst4.8 {d0, d1, d2, d3}, [%0]! \n" // store 8 ARGB pixels. + "bgt 1b \n" + : "+r"(dst_argb), // %0 + "+r"(width) // %1 + : + : "cc", "memory", "q0", "q1", "q2", "q3", + "q10", "q11", "q12", "q13", "q14", "q15" + ); +} + +// Tranform 8 ARGB pixels (32 bytes) with color matrix. +// TODO(fbarchard): Was same as Sepia except matrix is provided. This function +// needs to saturate. Consider doing a non-saturating version. +void ARGBColorMatrixRow_NEON(const uint8* src_argb, uint8* dst_argb, + const int8* matrix_argb, int width) { + asm volatile ( + MEMACCESS(3) + "vld1.8 {q2}, [%3] \n" // load 3 ARGB vectors. + "vmovl.s8 q0, d4 \n" // B,G coefficients s16. + "vmovl.s8 q1, d5 \n" // R,A coefficients s16. + + ".p2align 2 \n" + "1: \n" + MEMACCESS(0) + "vld4.8 {d16, d18, d20, d22}, [%0]! \n" // load 8 ARGB pixels. + "subs %2, %2, #8 \n" // 8 processed per loop. + "vmovl.u8 q8, d16 \n" // b (0 .. 255) 16 bit + "vmovl.u8 q9, d18 \n" // g + "vmovl.u8 q10, d20 \n" // r + "vmovl.u8 q11, d22 \n" // a + "vmul.s16 q12, q8, d0[0] \n" // B = B * Matrix B + "vmul.s16 q13, q8, d1[0] \n" // G = B * Matrix G + "vmul.s16 q14, q8, d2[0] \n" // R = B * Matrix R + "vmul.s16 q15, q8, d3[0] \n" // A = B * Matrix A + "vmul.s16 q4, q9, d0[1] \n" // B += G * Matrix B + "vmul.s16 q5, q9, d1[1] \n" // G += G * Matrix G + "vmul.s16 q6, q9, d2[1] \n" // R += G * Matrix R + "vmul.s16 q7, q9, d3[1] \n" // A += G * Matrix A + "vqadd.s16 q12, q12, q4 \n" // Accumulate B + "vqadd.s16 q13, q13, q5 \n" // Accumulate G + "vqadd.s16 q14, q14, q6 \n" // Accumulate R + "vqadd.s16 q15, q15, q7 \n" // Accumulate A + "vmul.s16 q4, q10, d0[2] \n" // B += R * Matrix B + "vmul.s16 q5, q10, d1[2] \n" // G += R * Matrix G + "vmul.s16 q6, q10, d2[2] \n" // R += R * Matrix R + "vmul.s16 q7, q10, d3[2] \n" // A += R * Matrix A + "vqadd.s16 q12, q12, q4 \n" // Accumulate B + "vqadd.s16 q13, q13, q5 \n" // Accumulate G + "vqadd.s16 q14, q14, q6 \n" // Accumulate R + "vqadd.s16 q15, q15, q7 \n" // Accumulate A + "vmul.s16 q4, q11, d0[3] \n" // B += A * Matrix B + "vmul.s16 q5, q11, d1[3] \n" // G += A * Matrix G + "vmul.s16 q6, q11, d2[3] \n" // R += A * Matrix R + "vmul.s16 q7, q11, d3[3] \n" // A += A * Matrix A + "vqadd.s16 q12, q12, q4 \n" // Accumulate B + "vqadd.s16 q13, q13, q5 \n" // Accumulate G + "vqadd.s16 q14, q14, q6 \n" // Accumulate R + "vqadd.s16 q15, q15, q7 \n" // Accumulate A + "vqshrun.s16 d16, q12, #6 \n" // 16 bit to 8 bit B + "vqshrun.s16 d18, q13, #6 \n" // 16 bit to 8 bit G + "vqshrun.s16 d20, q14, #6 \n" // 16 bit to 8 bit R + "vqshrun.s16 d22, q15, #6 \n" // 16 bit to 8 bit A + MEMACCESS(1) + "vst4.8 {d16, d18, d20, d22}, [%1]! \n" // store 8 ARGB pixels. + "bgt 1b \n" + : "+r"(src_argb), // %0 + "+r"(dst_argb), // %1 + "+r"(width) // %2 + : "r"(matrix_argb) // %3 + : "cc", "memory", "q0", "q1", "q2", "q4", "q5", "q6", "q7", "q8", "q9", + "q10", "q11", "q12", "q13", "q14", "q15" + ); +} + +// TODO(fbarchard): fix vqshrun in ARGBMultiplyRow_NEON and reenable. +#ifdef HAS_ARGBMULTIPLYROW_NEON +// Multiply 2 rows of ARGB pixels together, 8 pixels at a time. +void ARGBMultiplyRow_NEON(const uint8* src_argb0, const uint8* src_argb1, + uint8* dst_argb, int width) { + asm volatile ( + // 8 pixel loop. + ".p2align 2 \n" + "1: \n" + MEMACCESS(0) + "vld4.8 {d0, d2, d4, d6}, [%0]! \n" // load 8 ARGB pixels. + MEMACCESS(1) + "vld4.8 {d1, d3, d5, d7}, [%1]! \n" // load 8 more ARGB pixels. + "subs %3, %3, #8 \n" // 8 processed per loop. + "vmull.u8 q0, d0, d1 \n" // multiply B + "vmull.u8 q1, d2, d3 \n" // multiply G + "vmull.u8 q2, d4, d5 \n" // multiply R + "vmull.u8 q3, d6, d7 \n" // multiply A + "vrshrn.u16 d0, q0, #8 \n" // 16 bit to 8 bit B + "vrshrn.u16 d1, q1, #8 \n" // 16 bit to 8 bit G + "vrshrn.u16 d2, q2, #8 \n" // 16 bit to 8 bit R + "vrshrn.u16 d3, q3, #8 \n" // 16 bit to 8 bit A + MEMACCESS(2) + "vst4.8 {d0, d1, d2, d3}, [%2]! \n" // store 8 ARGB pixels. + "bgt 1b \n" + + : "+r"(src_argb0), // %0 + "+r"(src_argb1), // %1 + "+r"(dst_argb), // %2 + "+r"(width) // %3 + : + : "cc", "memory", "q0", "q1", "q2", "q3" + ); +} +#endif // HAS_ARGBMULTIPLYROW_NEON + +// Add 2 rows of ARGB pixels together, 8 pixels at a time. +void ARGBAddRow_NEON(const uint8* src_argb0, const uint8* src_argb1, + uint8* dst_argb, int width) { + asm volatile ( + // 8 pixel loop. + ".p2align 2 \n" + "1: \n" + MEMACCESS(0) + "vld4.8 {d0, d1, d2, d3}, [%0]! \n" // load 8 ARGB pixels. + MEMACCESS(1) + "vld4.8 {d4, d5, d6, d7}, [%1]! \n" // load 8 more ARGB pixels. + "subs %3, %3, #8 \n" // 8 processed per loop. + "vqadd.u8 q0, q0, q2 \n" // add B, G + "vqadd.u8 q1, q1, q3 \n" // add R, A + MEMACCESS(2) + "vst4.8 {d0, d1, d2, d3}, [%2]! \n" // store 8 ARGB pixels. + "bgt 1b \n" + + : "+r"(src_argb0), // %0 + "+r"(src_argb1), // %1 + "+r"(dst_argb), // %2 + "+r"(width) // %3 + : + : "cc", "memory", "q0", "q1", "q2", "q3" + ); +} + +// Subtract 2 rows of ARGB pixels, 8 pixels at a time. +void ARGBSubtractRow_NEON(const uint8* src_argb0, const uint8* src_argb1, + uint8* dst_argb, int width) { + asm volatile ( + // 8 pixel loop. + ".p2align 2 \n" + "1: \n" + MEMACCESS(0) + "vld4.8 {d0, d1, d2, d3}, [%0]! \n" // load 8 ARGB pixels. + MEMACCESS(1) + "vld4.8 {d4, d5, d6, d7}, [%1]! \n" // load 8 more ARGB pixels. + "subs %3, %3, #8 \n" // 8 processed per loop. + "vqsub.u8 q0, q0, q2 \n" // subtract B, G + "vqsub.u8 q1, q1, q3 \n" // subtract R, A + MEMACCESS(2) + "vst4.8 {d0, d1, d2, d3}, [%2]! \n" // store 8 ARGB pixels. + "bgt 1b \n" + + : "+r"(src_argb0), // %0 + "+r"(src_argb1), // %1 + "+r"(dst_argb), // %2 + "+r"(width) // %3 + : + : "cc", "memory", "q0", "q1", "q2", "q3" + ); +} + +// Adds Sobel X and Sobel Y and stores Sobel into ARGB. +// A = 255 +// R = Sobel +// G = Sobel +// B = Sobel +void SobelRow_NEON(const uint8* src_sobelx, const uint8* src_sobely, + uint8* dst_argb, int width) { + asm volatile ( + "vmov.u8 d3, #255 \n" // alpha + // 8 pixel loop. + ".p2align 2 \n" + "1: \n" + MEMACCESS(0) + "vld1.8 {d0}, [%0]! \n" // load 8 sobelx. + MEMACCESS(1) + "vld1.8 {d1}, [%1]! \n" // load 8 sobely. + "subs %3, %3, #8 \n" // 8 processed per loop. + "vqadd.u8 d0, d0, d1 \n" // add + "vmov.u8 d1, d0 \n" + "vmov.u8 d2, d0 \n" + MEMACCESS(2) + "vst4.8 {d0, d1, d2, d3}, [%2]! \n" // store 8 ARGB pixels. + "bgt 1b \n" + : "+r"(src_sobelx), // %0 + "+r"(src_sobely), // %1 + "+r"(dst_argb), // %2 + "+r"(width) // %3 + : + : "cc", "memory", "q0", "q1" + ); +} + +// Adds Sobel X and Sobel Y and stores Sobel into plane. +void SobelToPlaneRow_NEON(const uint8* src_sobelx, const uint8* src_sobely, + uint8* dst_y, int width) { + asm volatile ( + // 16 pixel loop. + ".p2align 2 \n" + "1: \n" + MEMACCESS(0) + "vld1.8 {q0}, [%0]! \n" // load 16 sobelx. + MEMACCESS(1) + "vld1.8 {q1}, [%1]! \n" // load 16 sobely. + "subs %3, %3, #16 \n" // 16 processed per loop. + "vqadd.u8 q0, q0, q1 \n" // add + MEMACCESS(2) + "vst1.8 {q0}, [%2]! \n" // store 16 pixels. + "bgt 1b \n" + : "+r"(src_sobelx), // %0 + "+r"(src_sobely), // %1 + "+r"(dst_y), // %2 + "+r"(width) // %3 + : + : "cc", "memory", "q0", "q1" + ); +} + +// Mixes Sobel X, Sobel Y and Sobel into ARGB. +// A = 255 +// R = Sobel X +// G = Sobel +// B = Sobel Y +void SobelXYRow_NEON(const uint8* src_sobelx, const uint8* src_sobely, + uint8* dst_argb, int width) { + asm volatile ( + "vmov.u8 d3, #255 \n" // alpha + // 8 pixel loop. + ".p2align 2 \n" + "1: \n" + MEMACCESS(0) + "vld1.8 {d2}, [%0]! \n" // load 8 sobelx. + MEMACCESS(1) + "vld1.8 {d0}, [%1]! \n" // load 8 sobely. + "subs %3, %3, #8 \n" // 8 processed per loop. + "vqadd.u8 d1, d0, d2 \n" // add + MEMACCESS(2) + "vst4.8 {d0, d1, d2, d3}, [%2]! \n" // store 8 ARGB pixels. + "bgt 1b \n" + : "+r"(src_sobelx), // %0 + "+r"(src_sobely), // %1 + "+r"(dst_argb), // %2 + "+r"(width) // %3 + : + : "cc", "memory", "q0", "q1" + ); +} + +// SobelX as a matrix is +// -1 0 1 +// -2 0 2 +// -1 0 1 +void SobelXRow_NEON(const uint8* src_y0, const uint8* src_y1, + const uint8* src_y2, uint8* dst_sobelx, int width) { + asm volatile ( + ".p2align 2 \n" + "1: \n" + MEMACCESS(0) + "vld1.8 {d0}, [%0],%5 \n" // top + MEMACCESS(0) + "vld1.8 {d1}, [%0],%6 \n" + "vsubl.u8 q0, d0, d1 \n" + MEMACCESS(1) + "vld1.8 {d2}, [%1],%5 \n" // center * 2 + MEMACCESS(1) + "vld1.8 {d3}, [%1],%6 \n" + "vsubl.u8 q1, d2, d3 \n" + "vadd.s16 q0, q0, q1 \n" + "vadd.s16 q0, q0, q1 \n" + MEMACCESS(2) + "vld1.8 {d2}, [%2],%5 \n" // bottom + MEMACCESS(2) + "vld1.8 {d3}, [%2],%6 \n" + "subs %4, %4, #8 \n" // 8 pixels + "vsubl.u8 q1, d2, d3 \n" + "vadd.s16 q0, q0, q1 \n" + "vabs.s16 q0, q0 \n" + "vqmovn.u16 d0, q0 \n" + MEMACCESS(3) + "vst1.8 {d0}, [%3]! \n" // store 8 sobelx + "bgt 1b \n" + : "+r"(src_y0), // %0 + "+r"(src_y1), // %1 + "+r"(src_y2), // %2 + "+r"(dst_sobelx), // %3 + "+r"(width) // %4 + : "r"(2), // %5 + "r"(6) // %6 + : "cc", "memory", "q0", "q1" // Clobber List + ); +} + +// SobelY as a matrix is +// -1 -2 -1 +// 0 0 0 +// 1 2 1 +void SobelYRow_NEON(const uint8* src_y0, const uint8* src_y1, + uint8* dst_sobely, int width) { + asm volatile ( + ".p2align 2 \n" + "1: \n" + MEMACCESS(0) + "vld1.8 {d0}, [%0],%4 \n" // left + MEMACCESS(1) + "vld1.8 {d1}, [%1],%4 \n" + "vsubl.u8 q0, d0, d1 \n" + MEMACCESS(0) + "vld1.8 {d2}, [%0],%4 \n" // center * 2 + MEMACCESS(1) + "vld1.8 {d3}, [%1],%4 \n" + "vsubl.u8 q1, d2, d3 \n" + "vadd.s16 q0, q0, q1 \n" + "vadd.s16 q0, q0, q1 \n" + MEMACCESS(0) + "vld1.8 {d2}, [%0],%5 \n" // right + MEMACCESS(1) + "vld1.8 {d3}, [%1],%5 \n" + "subs %3, %3, #8 \n" // 8 pixels + "vsubl.u8 q1, d2, d3 \n" + "vadd.s16 q0, q0, q1 \n" + "vabs.s16 q0, q0 \n" + "vqmovn.u16 d0, q0 \n" + MEMACCESS(2) + "vst1.8 {d0}, [%2]! \n" // store 8 sobely + "bgt 1b \n" + : "+r"(src_y0), // %0 + "+r"(src_y1), // %1 + "+r"(dst_sobely), // %2 + "+r"(width) // %3 + : "r"(1), // %4 + "r"(6) // %5 + : "cc", "memory", "q0", "q1" // Clobber List + ); +} +#endif // defined(__ARM_NEON__) && !defined(__aarch64__) + +#ifdef __cplusplus +} // extern "C" +} // namespace libyuv +#endif diff --git a/media/libaom/src/third_party/libyuv/source/row_neon64.cc b/media/libaom/src/third_party/libyuv/source/row_neon64.cc new file mode 100644 index 000000000..5d015454b --- /dev/null +++ b/media/libaom/src/third_party/libyuv/source/row_neon64.cc @@ -0,0 +1,3087 @@ +/* + * Copyright 2014 The LibYuv Project Authors. All rights reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#include "libyuv/row.h" + +#ifdef __cplusplus +namespace libyuv { +extern "C" { +#endif + +// This module is for GCC Neon armv8 64 bit. +#if !defined(LIBYUV_DISABLE_NEON) && defined(__aarch64__) + +// Read 8 Y, 4 U and 4 V from 422 +#define READYUV422 \ + MEMACCESS(0) \ + "ld1 {v0.8b}, [%0], #8 \n" \ + MEMACCESS(1) \ + "ld1 {v1.s}[0], [%1], #4 \n" \ + MEMACCESS(2) \ + "ld1 {v1.s}[1], [%2], #4 \n" + +// Read 8 Y, 2 U and 2 V from 422 +#define READYUV411 \ + MEMACCESS(0) \ + "ld1 {v0.8b}, [%0], #8 \n" \ + MEMACCESS(1) \ + "ld1 {v2.h}[0], [%1], #2 \n" \ + MEMACCESS(2) \ + "ld1 {v2.h}[1], [%2], #2 \n" \ + "zip1 v1.8b, v2.8b, v2.8b \n" + +// Read 8 Y, 8 U and 8 V from 444 +#define READYUV444 \ + MEMACCESS(0) \ + "ld1 {v0.8b}, [%0], #8 \n" \ + MEMACCESS(1) \ + "ld1 {v1.d}[0], [%1], #8 \n" \ + MEMACCESS(2) \ + "ld1 {v1.d}[1], [%2], #8 \n" \ + "uaddlp v1.8h, v1.16b \n" \ + "rshrn v1.8b, v1.8h, #1 \n" + +// Read 8 Y, and set 4 U and 4 V to 128 +#define READYUV400 \ + MEMACCESS(0) \ + "ld1 {v0.8b}, [%0], #8 \n" \ + "movi v1.8b , #128 \n" + +// Read 8 Y and 4 UV from NV12 +#define READNV12 \ + MEMACCESS(0) \ + "ld1 {v0.8b}, [%0], #8 \n" \ + MEMACCESS(1) \ + "ld1 {v2.8b}, [%1], #8 \n" \ + "uzp1 v1.8b, v2.8b, v2.8b \n" \ + "uzp2 v3.8b, v2.8b, v2.8b \n" \ + "ins v1.s[1], v3.s[0] \n" + +// Read 8 Y and 4 VU from NV21 +#define READNV21 \ + MEMACCESS(0) \ + "ld1 {v0.8b}, [%0], #8 \n" \ + MEMACCESS(1) \ + "ld1 {v2.8b}, [%1], #8 \n" \ + "uzp1 v3.8b, v2.8b, v2.8b \n" \ + "uzp2 v1.8b, v2.8b, v2.8b \n" \ + "ins v1.s[1], v3.s[0] \n" + +// Read 8 YUY2 +#define READYUY2 \ + MEMACCESS(0) \ + "ld2 {v0.8b, v1.8b}, [%0], #16 \n" \ + "uzp2 v3.8b, v1.8b, v1.8b \n" \ + "uzp1 v1.8b, v1.8b, v1.8b \n" \ + "ins v1.s[1], v3.s[0] \n" + +// Read 8 UYVY +#define READUYVY \ + MEMACCESS(0) \ + "ld2 {v2.8b, v3.8b}, [%0], #16 \n" \ + "orr v0.8b, v3.8b, v3.8b \n" \ + "uzp1 v1.8b, v2.8b, v2.8b \n" \ + "uzp2 v3.8b, v2.8b, v2.8b \n" \ + "ins v1.s[1], v3.s[0] \n" + +#define YUV422TORGB_SETUP_REG \ + "ld1r {v24.8h}, [%[kUVBiasBGR]], #2 \n" \ + "ld1r {v25.8h}, [%[kUVBiasBGR]], #2 \n" \ + "ld1r {v26.8h}, [%[kUVBiasBGR]] \n" \ + "ld1r {v31.4s}, [%[kYToRgb]] \n" \ + "movi v27.8h, #128 \n" \ + "movi v28.8h, #102 \n" \ + "movi v29.8h, #25 \n" \ + "movi v30.8h, #52 \n" + +#define YUV422TORGB(vR, vG, vB) \ + "uxtl v0.8h, v0.8b \n" /* Extract Y */ \ + "shll v2.8h, v1.8b, #8 \n" /* Replicate UV */ \ + "ushll2 v3.4s, v0.8h, #0 \n" /* Y */ \ + "ushll v0.4s, v0.4h, #0 \n" \ + "mul v3.4s, v3.4s, v31.4s \n" \ + "mul v0.4s, v0.4s, v31.4s \n" \ + "sqshrun v0.4h, v0.4s, #16 \n" \ + "sqshrun2 v0.8h, v3.4s, #16 \n" /* Y */ \ + "uaddw v1.8h, v2.8h, v1.8b \n" /* Replicate UV */ \ + "mov v2.d[0], v1.d[1] \n" /* Extract V */ \ + "uxtl v2.8h, v2.8b \n" \ + "uxtl v1.8h, v1.8b \n" /* Extract U */ \ + "mul v3.8h, v1.8h, v27.8h \n" \ + "mul v5.8h, v1.8h, v29.8h \n" \ + "mul v6.8h, v2.8h, v30.8h \n" \ + "mul v7.8h, v2.8h, v28.8h \n" \ + "sqadd v6.8h, v6.8h, v5.8h \n" \ + "sqadd " #vB ".8h, v24.8h, v0.8h \n" /* B */ \ + "sqadd " #vG ".8h, v25.8h, v0.8h \n" /* G */ \ + "sqadd " #vR ".8h, v26.8h, v0.8h \n" /* R */ \ + "sqadd " #vB ".8h, " #vB ".8h, v3.8h \n" /* B */ \ + "sqsub " #vG ".8h, " #vG ".8h, v6.8h \n" /* G */ \ + "sqadd " #vR ".8h, " #vR ".8h, v7.8h \n" /* R */ \ + "sqshrun " #vB ".8b, " #vB ".8h, #6 \n" /* B */ \ + "sqshrun " #vG ".8b, " #vG ".8h, #6 \n" /* G */ \ + "sqshrun " #vR ".8b, " #vR ".8h, #6 \n" /* R */ \ + +// YUV to RGB conversion constants. +// Y contribution to R,G,B. Scale and bias. +#define YG 18997 /* round(1.164 * 64 * 256 * 256 / 257) */ +#define YGB 1160 /* 1.164 * 64 * 16 - adjusted for even error distribution */ + +// U and V contributions to R,G,B. +#define UB -128 /* -min(128, round(2.018 * 64)) */ +#define UG 25 /* -round(-0.391 * 64) */ +#define VG 52 /* -round(-0.813 * 64) */ +#define VR -102 /* -round(1.596 * 64) */ + +// Bias values to subtract 16 from Y and 128 from U and V. +#define BB (UB * 128 - YGB) +#define BG (UG * 128 + VG * 128 - YGB) +#define BR (VR * 128 - YGB) + +static vec16 kUVBiasBGR = { BB, BG, BR, 0, 0, 0, 0, 0 }; +static vec32 kYToRgb = { 0x0101 * YG, 0, 0, 0 }; + +#undef YG +#undef YGB +#undef UB +#undef UG +#undef VG +#undef VR +#undef BB +#undef BG +#undef BR + +#define RGBTOUV_SETUP_REG \ + "movi v20.8h, #56, lsl #0 \n" /* UB/VR coefficient (0.875) / 2 */ \ + "movi v21.8h, #37, lsl #0 \n" /* UG coefficient (-0.5781) / 2 */ \ + "movi v22.8h, #19, lsl #0 \n" /* UR coefficient (-0.2969) / 2 */ \ + "movi v23.8h, #9, lsl #0 \n" /* VB coefficient (-0.1406) / 2 */ \ + "movi v24.8h, #47, lsl #0 \n" /* VG coefficient (-0.7344) / 2 */ \ + "movi v25.16b, #0x80 \n" /* 128.5 (0x8080 in 16-bit) */ + + +#ifdef HAS_I444TOARGBROW_NEON +void I444ToARGBRow_NEON(const uint8* src_y, + const uint8* src_u, + const uint8* src_v, + uint8* dst_argb, + int width) { + asm volatile ( + YUV422TORGB_SETUP_REG + "1: \n" + READYUV444 + YUV422TORGB(v22, v21, v20) + "subs %w4, %w4, #8 \n" + "movi v23.8b, #255 \n" /* A */ + MEMACCESS(3) + "st4 {v20.8b,v21.8b,v22.8b,v23.8b}, [%3], #32 \n" + "b.gt 1b \n" + : "+r"(src_y), // %0 + "+r"(src_u), // %1 + "+r"(src_v), // %2 + "+r"(dst_argb), // %3 + "+r"(width) // %4 + : [kUVBiasBGR]"r"(&kUVBiasBGR), + [kYToRgb]"r"(&kYToRgb) + : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20", + "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30" + ); +} +#endif // HAS_I444TOARGBROW_NEON + +#ifdef HAS_I422TOARGBROW_NEON +void I422ToARGBRow_NEON(const uint8* src_y, + const uint8* src_u, + const uint8* src_v, + uint8* dst_argb, + int width) { + asm volatile ( + YUV422TORGB_SETUP_REG + "1: \n" + READYUV422 + YUV422TORGB(v22, v21, v20) + "subs %w4, %w4, #8 \n" + "movi v23.8b, #255 \n" /* A */ + MEMACCESS(3) + "st4 {v20.8b,v21.8b,v22.8b,v23.8b}, [%3], #32 \n" + "b.gt 1b \n" + : "+r"(src_y), // %0 + "+r"(src_u), // %1 + "+r"(src_v), // %2 + "+r"(dst_argb), // %3 + "+r"(width) // %4 + : [kUVBiasBGR]"r"(&kUVBiasBGR), + [kYToRgb]"r"(&kYToRgb) + : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20", + "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30" + ); +} +#endif // HAS_I422TOARGBROW_NEON + +#ifdef HAS_I411TOARGBROW_NEON +void I411ToARGBRow_NEON(const uint8* src_y, + const uint8* src_u, + const uint8* src_v, + uint8* dst_argb, + int width) { + asm volatile ( + YUV422TORGB_SETUP_REG + "1: \n" + READYUV411 + YUV422TORGB(v22, v21, v20) + "subs %w4, %w4, #8 \n" + "movi v23.8b, #255 \n" /* A */ + MEMACCESS(3) + "st4 {v20.8b,v21.8b,v22.8b,v23.8b}, [%3], #32 \n" + "b.gt 1b \n" + : "+r"(src_y), // %0 + "+r"(src_u), // %1 + "+r"(src_v), // %2 + "+r"(dst_argb), // %3 + "+r"(width) // %4 + : [kUVBiasBGR]"r"(&kUVBiasBGR), + [kYToRgb]"r"(&kYToRgb) + : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20", + "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30" + ); +} +#endif // HAS_I411TOARGBROW_NEON + +#ifdef HAS_I422TOBGRAROW_NEON +void I422ToBGRARow_NEON(const uint8* src_y, + const uint8* src_u, + const uint8* src_v, + uint8* dst_bgra, + int width) { + asm volatile ( + YUV422TORGB_SETUP_REG + "1: \n" + READYUV422 + YUV422TORGB(v21, v22, v23) + "subs %w4, %w4, #8 \n" + "movi v20.8b, #255 \n" /* A */ + MEMACCESS(3) + "st4 {v20.8b,v21.8b,v22.8b,v23.8b}, [%3], #32 \n" + "b.gt 1b \n" + : "+r"(src_y), // %0 + "+r"(src_u), // %1 + "+r"(src_v), // %2 + "+r"(dst_bgra), // %3 + "+r"(width) // %4 + : [kUVBiasBGR]"r"(&kUVBiasBGR), + [kYToRgb]"r"(&kYToRgb) + : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20", + "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30" + ); +} +#endif // HAS_I422TOBGRAROW_NEON + +#ifdef HAS_I422TOABGRROW_NEON +void I422ToABGRRow_NEON(const uint8* src_y, + const uint8* src_u, + const uint8* src_v, + uint8* dst_abgr, + int width) { + asm volatile ( + YUV422TORGB_SETUP_REG + "1: \n" + READYUV422 + YUV422TORGB(v20, v21, v22) + "subs %w4, %w4, #8 \n" + "movi v23.8b, #255 \n" /* A */ + MEMACCESS(3) + "st4 {v20.8b,v21.8b,v22.8b,v23.8b}, [%3], #32 \n" + "b.gt 1b \n" + : "+r"(src_y), // %0 + "+r"(src_u), // %1 + "+r"(src_v), // %2 + "+r"(dst_abgr), // %3 + "+r"(width) // %4 + : [kUVBiasBGR]"r"(&kUVBiasBGR), + [kYToRgb]"r"(&kYToRgb) + : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20", + "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30" + ); +} +#endif // HAS_I422TOABGRROW_NEON + +#ifdef HAS_I422TORGBAROW_NEON +void I422ToRGBARow_NEON(const uint8* src_y, + const uint8* src_u, + const uint8* src_v, + uint8* dst_rgba, + int width) { + asm volatile ( + YUV422TORGB_SETUP_REG + "1: \n" + READYUV422 + YUV422TORGB(v23, v22, v21) + "subs %w4, %w4, #8 \n" + "movi v20.8b, #255 \n" /* A */ + MEMACCESS(3) + "st4 {v20.8b,v21.8b,v22.8b,v23.8b}, [%3], #32 \n" + "b.gt 1b \n" + : "+r"(src_y), // %0 + "+r"(src_u), // %1 + "+r"(src_v), // %2 + "+r"(dst_rgba), // %3 + "+r"(width) // %4 + : [kUVBiasBGR]"r"(&kUVBiasBGR), + [kYToRgb]"r"(&kYToRgb) + : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20", + "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30" + ); +} +#endif // HAS_I422TORGBAROW_NEON + +#ifdef HAS_I422TORGB24ROW_NEON +void I422ToRGB24Row_NEON(const uint8* src_y, + const uint8* src_u, + const uint8* src_v, + uint8* dst_rgb24, + int width) { + asm volatile ( + YUV422TORGB_SETUP_REG + "1: \n" + READYUV422 + YUV422TORGB(v22, v21, v20) + "subs %w4, %w4, #8 \n" + MEMACCESS(3) + "st3 {v20.8b,v21.8b,v22.8b}, [%3], #24 \n" + "b.gt 1b \n" + : "+r"(src_y), // %0 + "+r"(src_u), // %1 + "+r"(src_v), // %2 + "+r"(dst_rgb24), // %3 + "+r"(width) // %4 + : [kUVBiasBGR]"r"(&kUVBiasBGR), + [kYToRgb]"r"(&kYToRgb) + : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20", + "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30" + ); +} +#endif // HAS_I422TORGB24ROW_NEON + +#ifdef HAS_I422TORAWROW_NEON +void I422ToRAWRow_NEON(const uint8* src_y, + const uint8* src_u, + const uint8* src_v, + uint8* dst_raw, + int width) { + asm volatile ( + YUV422TORGB_SETUP_REG + "1: \n" + READYUV422 + YUV422TORGB(v20, v21, v22) + "subs %w4, %w4, #8 \n" + MEMACCESS(3) + "st3 {v20.8b,v21.8b,v22.8b}, [%3], #24 \n" + "b.gt 1b \n" + : "+r"(src_y), // %0 + "+r"(src_u), // %1 + "+r"(src_v), // %2 + "+r"(dst_raw), // %3 + "+r"(width) // %4 + : [kUVBiasBGR]"r"(&kUVBiasBGR), + [kYToRgb]"r"(&kYToRgb) + : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20", + "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30" + ); +} +#endif // HAS_I422TORAWROW_NEON + +#define ARGBTORGB565 \ + "shll v0.8h, v22.8b, #8 \n" /* R */ \ + "shll v20.8h, v20.8b, #8 \n" /* B */ \ + "shll v21.8h, v21.8b, #8 \n" /* G */ \ + "sri v0.8h, v21.8h, #5 \n" /* RG */ \ + "sri v0.8h, v20.8h, #11 \n" /* RGB */ + +#ifdef HAS_I422TORGB565ROW_NEON +void I422ToRGB565Row_NEON(const uint8* src_y, + const uint8* src_u, + const uint8* src_v, + uint8* dst_rgb565, + int width) { + asm volatile ( + YUV422TORGB_SETUP_REG + "1: \n" + READYUV422 + YUV422TORGB(v22, v21, v20) + "subs %w4, %w4, #8 \n" + ARGBTORGB565 + MEMACCESS(3) + "st1 {v0.8h}, [%3], #16 \n" // store 8 pixels RGB565. + "b.gt 1b \n" + : "+r"(src_y), // %0 + "+r"(src_u), // %1 + "+r"(src_v), // %2 + "+r"(dst_rgb565), // %3 + "+r"(width) // %4 + : [kUVBiasBGR]"r"(&kUVBiasBGR), + [kYToRgb]"r"(&kYToRgb) + : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20", + "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30" + ); +} +#endif // HAS_I422TORGB565ROW_NEON + +#define ARGBTOARGB1555 \ + "shll v0.8h, v23.8b, #8 \n" /* A */ \ + "shll v22.8h, v22.8b, #8 \n" /* R */ \ + "shll v20.8h, v20.8b, #8 \n" /* B */ \ + "shll v21.8h, v21.8b, #8 \n" /* G */ \ + "sri v0.8h, v22.8h, #1 \n" /* AR */ \ + "sri v0.8h, v21.8h, #6 \n" /* ARG */ \ + "sri v0.8h, v20.8h, #11 \n" /* ARGB */ + +#ifdef HAS_I422TOARGB1555ROW_NEON +void I422ToARGB1555Row_NEON(const uint8* src_y, + const uint8* src_u, + const uint8* src_v, + uint8* dst_argb1555, + int width) { + asm volatile ( + YUV422TORGB_SETUP_REG + "1: \n" + READYUV422 + YUV422TORGB(v22, v21, v20) + "subs %w4, %w4, #8 \n" + "movi v23.8b, #255 \n" + ARGBTOARGB1555 + MEMACCESS(3) + "st1 {v0.8h}, [%3], #16 \n" // store 8 pixels RGB565. + "b.gt 1b \n" + : "+r"(src_y), // %0 + "+r"(src_u), // %1 + "+r"(src_v), // %2 + "+r"(dst_argb1555), // %3 + "+r"(width) // %4 + : [kUVBiasBGR]"r"(&kUVBiasBGR), + [kYToRgb]"r"(&kYToRgb) + : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20", + "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30" + ); +} +#endif // HAS_I422TOARGB1555ROW_NEON + +#define ARGBTOARGB4444 \ + /* Input v20.8b<=B, v21.8b<=G, v22.8b<=R, v23.8b<=A, v4.8b<=0x0f */ \ + "ushr v20.8b, v20.8b, #4 \n" /* B */ \ + "bic v21.8b, v21.8b, v4.8b \n" /* G */ \ + "ushr v22.8b, v22.8b, #4 \n" /* R */ \ + "bic v23.8b, v23.8b, v4.8b \n" /* A */ \ + "orr v0.8b, v20.8b, v21.8b \n" /* BG */ \ + "orr v1.8b, v22.8b, v23.8b \n" /* RA */ \ + "zip1 v0.16b, v0.16b, v1.16b \n" /* BGRA */ + +#ifdef HAS_I422TOARGB4444ROW_NEON +void I422ToARGB4444Row_NEON(const uint8* src_y, + const uint8* src_u, + const uint8* src_v, + uint8* dst_argb4444, + int width) { + asm volatile ( + YUV422TORGB_SETUP_REG + "movi v4.16b, #0x0f \n" // bits to clear with vbic. + "1: \n" + READYUV422 + YUV422TORGB(v22, v21, v20) + "subs %w4, %w4, #8 \n" + "movi v23.8b, #255 \n" + ARGBTOARGB4444 + MEMACCESS(3) + "st1 {v0.8h}, [%3], #16 \n" // store 8 pixels ARGB4444. + "b.gt 1b \n" + : "+r"(src_y), // %0 + "+r"(src_u), // %1 + "+r"(src_v), // %2 + "+r"(dst_argb4444), // %3 + "+r"(width) // %4 + : [kUVBiasBGR]"r"(&kUVBiasBGR), + [kYToRgb]"r"(&kYToRgb) + : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20", + "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30" + ); +} +#endif // HAS_I422TOARGB4444ROW_NEON + +#ifdef HAS_I400TOARGBROW_NEON +void I400ToARGBRow_NEON(const uint8* src_y, + uint8* dst_argb, + int width) { + int64 width64 = (int64)(width); + asm volatile ( + YUV422TORGB_SETUP_REG + "1: \n" + READYUV400 + YUV422TORGB(v22, v21, v20) + "subs %w2, %w2, #8 \n" + "movi v23.8b, #255 \n" + MEMACCESS(1) + "st4 {v20.8b,v21.8b,v22.8b,v23.8b}, [%1], #32 \n" + "b.gt 1b \n" + : "+r"(src_y), // %0 + "+r"(dst_argb), // %1 + "+r"(width64) // %2 + : [kUVBiasBGR]"r"(&kUVBiasBGR), + [kYToRgb]"r"(&kYToRgb) + : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20", + "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30" + ); +} +#endif // HAS_I400TOARGBROW_NEON + +#ifdef HAS_J400TOARGBROW_NEON +void J400ToARGBRow_NEON(const uint8* src_y, + uint8* dst_argb, + int width) { + asm volatile ( + "movi v23.8b, #255 \n" + "1: \n" + MEMACCESS(0) + "ld1 {v20.8b}, [%0], #8 \n" + "orr v21.8b, v20.8b, v20.8b \n" + "orr v22.8b, v20.8b, v20.8b \n" + "subs %w2, %w2, #8 \n" + MEMACCESS(1) + "st4 {v20.8b,v21.8b,v22.8b,v23.8b}, [%1], #32 \n" + "b.gt 1b \n" + : "+r"(src_y), // %0 + "+r"(dst_argb), // %1 + "+r"(width) // %2 + : + : "cc", "memory", "v20", "v21", "v22", "v23" + ); +} +#endif // HAS_J400TOARGBROW_NEON + +#ifdef HAS_NV12TOARGBROW_NEON +void NV12ToARGBRow_NEON(const uint8* src_y, + const uint8* src_uv, + uint8* dst_argb, + int width) { + asm volatile ( + YUV422TORGB_SETUP_REG + "1: \n" + READNV12 + YUV422TORGB(v22, v21, v20) + "subs %w3, %w3, #8 \n" + "movi v23.8b, #255 \n" + MEMACCESS(2) + "st4 {v20.8b,v21.8b,v22.8b,v23.8b}, [%2], #32 \n" + "b.gt 1b \n" + : "+r"(src_y), // %0 + "+r"(src_uv), // %1 + "+r"(dst_argb), // %2 + "+r"(width) // %3 + : [kUVBiasBGR]"r"(&kUVBiasBGR), + [kYToRgb]"r"(&kYToRgb) + : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20", + "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30" + ); +} +#endif // HAS_NV12TOARGBROW_NEON + +#ifdef HAS_NV21TOARGBROW_NEON +void NV21ToARGBRow_NEON(const uint8* src_y, + const uint8* src_uv, + uint8* dst_argb, + int width) { + asm volatile ( + YUV422TORGB_SETUP_REG + "1: \n" + READNV21 + YUV422TORGB(v22, v21, v20) + "subs %w3, %w3, #8 \n" + "movi v23.8b, #255 \n" + MEMACCESS(2) + "st4 {v20.8b,v21.8b,v22.8b,v23.8b}, [%2], #32 \n" + "b.gt 1b \n" + : "+r"(src_y), // %0 + "+r"(src_uv), // %1 + "+r"(dst_argb), // %2 + "+r"(width) // %3 + : [kUVBiasBGR]"r"(&kUVBiasBGR), + [kYToRgb]"r"(&kYToRgb) + : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20", + "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30" + ); +} +#endif // HAS_NV21TOARGBROW_NEON + +#ifdef HAS_NV12TORGB565ROW_NEON +void NV12ToRGB565Row_NEON(const uint8* src_y, + const uint8* src_uv, + uint8* dst_rgb565, + int width) { + asm volatile ( + YUV422TORGB_SETUP_REG + "1: \n" + READNV12 + YUV422TORGB(v22, v21, v20) + "subs %w3, %w3, #8 \n" + ARGBTORGB565 + MEMACCESS(2) + "st1 {v0.8h}, [%2], 16 \n" // store 8 pixels RGB565. + "b.gt 1b \n" + : "+r"(src_y), // %0 + "+r"(src_uv), // %1 + "+r"(dst_rgb565), // %2 + "+r"(width) // %3 + : [kUVBiasBGR]"r"(&kUVBiasBGR), + [kYToRgb]"r"(&kYToRgb) + : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20", + "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30" + ); +} +#endif // HAS_NV12TORGB565ROW_NEON + +#ifdef HAS_NV21TORGB565ROW_NEON +void NV21ToRGB565Row_NEON(const uint8* src_y, + const uint8* src_uv, + uint8* dst_rgb565, + int width) { + asm volatile ( + YUV422TORGB_SETUP_REG + "1: \n" + READNV21 + YUV422TORGB(v22, v21, v20) + "subs %w3, %w3, #8 \n" + ARGBTORGB565 + MEMACCESS(2) + "st1 {v0.8h}, [%2], 16 \n" // store 8 pixels RGB565. + "b.gt 1b \n" + : "+r"(src_y), // %0 + "+r"(src_uv), // %1 + "+r"(dst_rgb565), // %2 + "+r"(width) // %3 + : [kUVBiasBGR]"r"(&kUVBiasBGR), + [kYToRgb]"r"(&kYToRgb) + : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20", + "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30" + ); +} +#endif // HAS_NV21TORGB565ROW_NEON + +#ifdef HAS_YUY2TOARGBROW_NEON +void YUY2ToARGBRow_NEON(const uint8* src_yuy2, + uint8* dst_argb, + int width) { + int64 width64 = (int64)(width); + asm volatile ( + YUV422TORGB_SETUP_REG + "1: \n" + READYUY2 + YUV422TORGB(v22, v21, v20) + "subs %w2, %w2, #8 \n" + "movi v23.8b, #255 \n" + MEMACCESS(1) + "st4 {v20.8b,v21.8b,v22.8b,v23.8b}, [%1], #32 \n" + "b.gt 1b \n" + : "+r"(src_yuy2), // %0 + "+r"(dst_argb), // %1 + "+r"(width64) // %2 + : [kUVBiasBGR]"r"(&kUVBiasBGR), + [kYToRgb]"r"(&kYToRgb) + : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20", + "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30" + ); +} +#endif // HAS_YUY2TOARGBROW_NEON + +#ifdef HAS_UYVYTOARGBROW_NEON +void UYVYToARGBRow_NEON(const uint8* src_uyvy, + uint8* dst_argb, + int width) { + int64 width64 = (int64)(width); + asm volatile ( + YUV422TORGB_SETUP_REG + "1: \n" + READUYVY + YUV422TORGB(v22, v21, v20) + "subs %w2, %w2, #8 \n" + "movi v23.8b, #255 \n" + MEMACCESS(1) + "st4 {v20.8b,v21.8b,v22.8b,v23.8b}, [%1], 32 \n" + "b.gt 1b \n" + : "+r"(src_uyvy), // %0 + "+r"(dst_argb), // %1 + "+r"(width64) // %2 + : [kUVBiasBGR]"r"(&kUVBiasBGR), + [kYToRgb]"r"(&kYToRgb) + : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20", + "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30" + ); +} +#endif // HAS_UYVYTOARGBROW_NEON + +// Reads 16 pairs of UV and write even values to dst_u and odd to dst_v. +#ifdef HAS_SPLITUVROW_NEON +void SplitUVRow_NEON(const uint8* src_uv, uint8* dst_u, uint8* dst_v, + int width) { + asm volatile ( + "1: \n" + MEMACCESS(0) + "ld2 {v0.16b,v1.16b}, [%0], #32 \n" // load 16 pairs of UV + "subs %w3, %w3, #16 \n" // 16 processed per loop + MEMACCESS(1) + "st1 {v0.16b}, [%1], #16 \n" // store U + MEMACCESS(2) + "st1 {v1.16b}, [%2], #16 \n" // store V + "b.gt 1b \n" + : "+r"(src_uv), // %0 + "+r"(dst_u), // %1 + "+r"(dst_v), // %2 + "+r"(width) // %3 // Output registers + : // Input registers + : "cc", "memory", "v0", "v1" // Clobber List + ); +} +#endif // HAS_SPLITUVROW_NEON + +// Reads 16 U's and V's and writes out 16 pairs of UV. +#ifdef HAS_MERGEUVROW_NEON +void MergeUVRow_NEON(const uint8* src_u, const uint8* src_v, uint8* dst_uv, + int width) { + asm volatile ( + "1: \n" + MEMACCESS(0) + "ld1 {v0.16b}, [%0], #16 \n" // load U + MEMACCESS(1) + "ld1 {v1.16b}, [%1], #16 \n" // load V + "subs %w3, %w3, #16 \n" // 16 processed per loop + MEMACCESS(2) + "st2 {v0.16b,v1.16b}, [%2], #32 \n" // store 16 pairs of UV + "b.gt 1b \n" + : + "+r"(src_u), // %0 + "+r"(src_v), // %1 + "+r"(dst_uv), // %2 + "+r"(width) // %3 // Output registers + : // Input registers + : "cc", "memory", "v0", "v1" // Clobber List + ); +} +#endif // HAS_MERGEUVROW_NEON + +// Copy multiple of 32. vld4.8 allow unaligned and is fastest on a15. +#ifdef HAS_COPYROW_NEON +void CopyRow_NEON(const uint8* src, uint8* dst, int count) { + asm volatile ( + "1: \n" + MEMACCESS(0) + "ld1 {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n" // load 32 + "subs %w2, %w2, #32 \n" // 32 processed per loop + MEMACCESS(1) + "st1 {v0.8b,v1.8b,v2.8b,v3.8b}, [%1], #32 \n" // store 32 + "b.gt 1b \n" + : "+r"(src), // %0 + "+r"(dst), // %1 + "+r"(count) // %2 // Output registers + : // Input registers + : "cc", "memory", "v0", "v1", "v2", "v3" // Clobber List + ); +} +#endif // HAS_COPYROW_NEON + +// SetRow writes 'count' bytes using an 8 bit value repeated. +void SetRow_NEON(uint8* dst, uint8 v8, int count) { + asm volatile ( + "dup v0.16b, %w2 \n" // duplicate 16 bytes + "1: \n" + "subs %w1, %w1, #16 \n" // 16 bytes per loop + MEMACCESS(0) + "st1 {v0.16b}, [%0], #16 \n" // store + "b.gt 1b \n" + : "+r"(dst), // %0 + "+r"(count) // %1 + : "r"(v8) // %2 + : "cc", "memory", "v0" + ); +} + +void ARGBSetRow_NEON(uint8* dst, uint32 v32, int count) { + asm volatile ( + "dup v0.4s, %w2 \n" // duplicate 4 ints + "1: \n" + "subs %w1, %w1, #4 \n" // 4 ints per loop + MEMACCESS(0) + "st1 {v0.16b}, [%0], #16 \n" // store + "b.gt 1b \n" + : "+r"(dst), // %0 + "+r"(count) // %1 + : "r"(v32) // %2 + : "cc", "memory", "v0" + ); +} + +#ifdef HAS_MIRRORROW_NEON +void MirrorRow_NEON(const uint8* src, uint8* dst, int width) { + int64 width64 = (int64) width; + asm volatile ( + // Start at end of source row. + "add %0, %0, %2 \n" + "sub %0, %0, #16 \n" + + "1: \n" + MEMACCESS(0) + "ld1 {v0.16b}, [%0], %3 \n" // src -= 16 + "subs %2, %2, #16 \n" // 16 pixels per loop. + "rev64 v0.16b, v0.16b \n" + MEMACCESS(1) + "st1 {v0.D}[1], [%1], #8 \n" // dst += 16 + MEMACCESS(1) + "st1 {v0.D}[0], [%1], #8 \n" + "b.gt 1b \n" + : "+r"(src), // %0 + "+r"(dst), // %1 + "+r"(width64) // %2 + : "r"((ptrdiff_t)-16) // %3 + : "cc", "memory", "v0" + ); +} +#endif // HAS_MIRRORROW_NEON + +#ifdef HAS_MIRRORUVROW_NEON +void MirrorUVRow_NEON(const uint8* src_uv, uint8* dst_u, uint8* dst_v, + int width) { + int64 width64 = (int64) width; + asm volatile ( + // Start at end of source row. + "add %0, %0, %3, lsl #1 \n" + "sub %0, %0, #16 \n" + + "1: \n" + MEMACCESS(0) + "ld2 {v0.8b, v1.8b}, [%0], %4 \n" // src -= 16 + "subs %3, %3, #8 \n" // 8 pixels per loop. + "rev64 v0.8b, v0.8b \n" + "rev64 v1.8b, v1.8b \n" + MEMACCESS(1) + "st1 {v0.8b}, [%1], #8 \n" // dst += 8 + MEMACCESS(2) + "st1 {v1.8b}, [%2], #8 \n" + "b.gt 1b \n" + : "+r"(src_uv), // %0 + "+r"(dst_u), // %1 + "+r"(dst_v), // %2 + "+r"(width64) // %3 + : "r"((ptrdiff_t)-16) // %4 + : "cc", "memory", "v0", "v1" + ); +} +#endif // HAS_MIRRORUVROW_NEON + +#ifdef HAS_ARGBMIRRORROW_NEON +void ARGBMirrorRow_NEON(const uint8* src, uint8* dst, int width) { + int64 width64 = (int64) width; + asm volatile ( + // Start at end of source row. + "add %0, %0, %2, lsl #2 \n" + "sub %0, %0, #16 \n" + + "1: \n" + MEMACCESS(0) + "ld1 {v0.16b}, [%0], %3 \n" // src -= 16 + "subs %2, %2, #4 \n" // 4 pixels per loop. + "rev64 v0.4s, v0.4s \n" + MEMACCESS(1) + "st1 {v0.D}[1], [%1], #8 \n" // dst += 16 + MEMACCESS(1) + "st1 {v0.D}[0], [%1], #8 \n" + "b.gt 1b \n" + : "+r"(src), // %0 + "+r"(dst), // %1 + "+r"(width64) // %2 + : "r"((ptrdiff_t)-16) // %3 + : "cc", "memory", "v0" + ); +} +#endif // HAS_ARGBMIRRORROW_NEON + +#ifdef HAS_RGB24TOARGBROW_NEON +void RGB24ToARGBRow_NEON(const uint8* src_rgb24, uint8* dst_argb, int pix) { + asm volatile ( + "movi v4.8b, #255 \n" // Alpha + "1: \n" + MEMACCESS(0) + "ld3 {v1.8b,v2.8b,v3.8b}, [%0], #24 \n" // load 8 pixels of RGB24. + "subs %w2, %w2, #8 \n" // 8 processed per loop. + MEMACCESS(1) + "st4 {v1.8b,v2.8b,v3.8b,v4.8b}, [%1], #32 \n" // store 8 ARGB pixels + "b.gt 1b \n" + : "+r"(src_rgb24), // %0 + "+r"(dst_argb), // %1 + "+r"(pix) // %2 + : + : "cc", "memory", "v1", "v2", "v3", "v4" // Clobber List + ); +} +#endif // HAS_RGB24TOARGBROW_NEON + +#ifdef HAS_RAWTOARGBROW_NEON +void RAWToARGBRow_NEON(const uint8* src_raw, uint8* dst_argb, int pix) { + asm volatile ( + "movi v5.8b, #255 \n" // Alpha + "1: \n" + MEMACCESS(0) + "ld3 {v0.8b,v1.8b,v2.8b}, [%0], #24 \n" // read r g b + "subs %w2, %w2, #8 \n" // 8 processed per loop. + "orr v3.8b, v1.8b, v1.8b \n" // move g + "orr v4.8b, v0.8b, v0.8b \n" // move r + MEMACCESS(1) + "st4 {v2.8b,v3.8b,v4.8b,v5.8b}, [%1], #32 \n" // store b g r a + "b.gt 1b \n" + : "+r"(src_raw), // %0 + "+r"(dst_argb), // %1 + "+r"(pix) // %2 + : + : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5" // Clobber List + ); +} +#endif // HAS_RAWTOARGBROW_NEON + +#define RGB565TOARGB \ + "shrn v6.8b, v0.8h, #5 \n" /* G xxGGGGGG */ \ + "shl v6.8b, v6.8b, #2 \n" /* G GGGGGG00 upper 6 */ \ + "ushr v4.8b, v6.8b, #6 \n" /* G 000000GG lower 2 */ \ + "orr v1.8b, v4.8b, v6.8b \n" /* G */ \ + "xtn v2.8b, v0.8h \n" /* B xxxBBBBB */ \ + "ushr v0.8h, v0.8h, #11 \n" /* R 000RRRRR */ \ + "xtn2 v2.16b,v0.8h \n" /* R in upper part */ \ + "shl v2.16b, v2.16b, #3 \n" /* R,B BBBBB000 upper 5 */ \ + "ushr v0.16b, v2.16b, #5 \n" /* R,B 00000BBB lower 3 */ \ + "orr v0.16b, v0.16b, v2.16b \n" /* R,B */ \ + "dup v2.2D, v0.D[1] \n" /* R */ + +#ifdef HAS_RGB565TOARGBROW_NEON +void RGB565ToARGBRow_NEON(const uint8* src_rgb565, uint8* dst_argb, int pix) { + asm volatile ( + "movi v3.8b, #255 \n" // Alpha + "1: \n" + MEMACCESS(0) + "ld1 {v0.16b}, [%0], #16 \n" // load 8 RGB565 pixels. + "subs %w2, %w2, #8 \n" // 8 processed per loop. + RGB565TOARGB + MEMACCESS(1) + "st4 {v0.8b,v1.8b,v2.8b,v3.8b}, [%1], #32 \n" // store 8 ARGB pixels + "b.gt 1b \n" + : "+r"(src_rgb565), // %0 + "+r"(dst_argb), // %1 + "+r"(pix) // %2 + : + : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v6" // Clobber List + ); +} +#endif // HAS_RGB565TOARGBROW_NEON + +#define ARGB1555TOARGB \ + "ushr v2.8h, v0.8h, #10 \n" /* R xxxRRRRR */ \ + "shl v2.8h, v2.8h, #3 \n" /* R RRRRR000 upper 5 */ \ + "xtn v3.8b, v2.8h \n" /* RRRRR000 AAAAAAAA */ \ + \ + "sshr v2.8h, v0.8h, #15 \n" /* A AAAAAAAA */ \ + "xtn2 v3.16b, v2.8h \n" \ + \ + "xtn v2.8b, v0.8h \n" /* B xxxBBBBB */ \ + "shrn2 v2.16b,v0.8h, #5 \n" /* G xxxGGGGG */ \ + \ + "ushr v1.16b, v3.16b, #5 \n" /* R,A 00000RRR lower 3 */ \ + "shl v0.16b, v2.16b, #3 \n" /* B,G BBBBB000 upper 5 */ \ + "ushr v2.16b, v0.16b, #5 \n" /* B,G 00000BBB lower 3 */ \ + \ + "orr v0.16b, v0.16b, v2.16b \n" /* B,G */ \ + "orr v2.16b, v1.16b, v3.16b \n" /* R,A */ \ + "dup v1.2D, v0.D[1] \n" \ + "dup v3.2D, v2.D[1] \n" + +// RGB555TOARGB is same as ARGB1555TOARGB but ignores alpha. +#define RGB555TOARGB \ + "ushr v2.8h, v0.8h, #10 \n" /* R xxxRRRRR */ \ + "shl v2.8h, v2.8h, #3 \n" /* R RRRRR000 upper 5 */ \ + "xtn v3.8b, v2.8h \n" /* RRRRR000 */ \ + \ + "xtn v2.8b, v0.8h \n" /* B xxxBBBBB */ \ + "shrn2 v2.16b,v0.8h, #5 \n" /* G xxxGGGGG */ \ + \ + "ushr v1.16b, v3.16b, #5 \n" /* R 00000RRR lower 3 */ \ + "shl v0.16b, v2.16b, #3 \n" /* B,G BBBBB000 upper 5 */ \ + "ushr v2.16b, v0.16b, #5 \n" /* B,G 00000BBB lower 3 */ \ + \ + "orr v0.16b, v0.16b, v2.16b \n" /* B,G */ \ + "orr v2.16b, v1.16b, v3.16b \n" /* R */ \ + "dup v1.2D, v0.D[1] \n" /* G */ \ + +#ifdef HAS_ARGB1555TOARGBROW_NEON +void ARGB1555ToARGBRow_NEON(const uint8* src_argb1555, uint8* dst_argb, + int pix) { + asm volatile ( + "movi v3.8b, #255 \n" // Alpha + "1: \n" + MEMACCESS(0) + "ld1 {v0.16b}, [%0], #16 \n" // load 8 ARGB1555 pixels. + "subs %w2, %w2, #8 \n" // 8 processed per loop. + ARGB1555TOARGB + MEMACCESS(1) + "st4 {v0.8b,v1.8b,v2.8b,v3.8b}, [%1], #32 \n" // store 8 ARGB pixels + "b.gt 1b \n" + : "+r"(src_argb1555), // %0 + "+r"(dst_argb), // %1 + "+r"(pix) // %2 + : + : "cc", "memory", "v0", "v1", "v2", "v3" // Clobber List + ); +} +#endif // HAS_ARGB1555TOARGBROW_NEON + +#define ARGB4444TOARGB \ + "shrn v1.8b, v0.8h, #8 \n" /* v1(l) AR */ \ + "xtn2 v1.16b, v0.8h \n" /* v1(h) GB */ \ + "shl v2.16b, v1.16b, #4 \n" /* B,R BBBB0000 */ \ + "ushr v3.16b, v1.16b, #4 \n" /* G,A 0000GGGG */ \ + "ushr v0.16b, v2.16b, #4 \n" /* B,R 0000BBBB */ \ + "shl v1.16b, v3.16b, #4 \n" /* G,A GGGG0000 */ \ + "orr v2.16b, v0.16b, v2.16b \n" /* B,R BBBBBBBB */ \ + "orr v3.16b, v1.16b, v3.16b \n" /* G,A GGGGGGGG */ \ + "dup v0.2D, v2.D[1] \n" \ + "dup v1.2D, v3.D[1] \n" + +#ifdef HAS_ARGB4444TOARGBROW_NEON +void ARGB4444ToARGBRow_NEON(const uint8* src_argb4444, uint8* dst_argb, + int pix) { + asm volatile ( + "1: \n" + MEMACCESS(0) + "ld1 {v0.16b}, [%0], #16 \n" // load 8 ARGB4444 pixels. + "subs %w2, %w2, #8 \n" // 8 processed per loop. + ARGB4444TOARGB + MEMACCESS(1) + "st4 {v0.8b,v1.8b,v2.8b,v3.8b}, [%1], #32 \n" // store 8 ARGB pixels + "b.gt 1b \n" + : "+r"(src_argb4444), // %0 + "+r"(dst_argb), // %1 + "+r"(pix) // %2 + : + : "cc", "memory", "v0", "v1", "v2", "v3", "v4" // Clobber List + ); +} +#endif // HAS_ARGB4444TOARGBROW_NEON + +#ifdef HAS_ARGBTORGB24ROW_NEON +void ARGBToRGB24Row_NEON(const uint8* src_argb, uint8* dst_rgb24, int pix) { + asm volatile ( + "1: \n" + MEMACCESS(0) + "ld4 {v1.8b,v2.8b,v3.8b,v4.8b}, [%0], #32 \n" // load 8 ARGB pixels + "subs %w2, %w2, #8 \n" // 8 processed per loop. + MEMACCESS(1) + "st3 {v1.8b,v2.8b,v3.8b}, [%1], #24 \n" // store 8 pixels of RGB24. + "b.gt 1b \n" + : "+r"(src_argb), // %0 + "+r"(dst_rgb24), // %1 + "+r"(pix) // %2 + : + : "cc", "memory", "v1", "v2", "v3", "v4" // Clobber List + ); +} +#endif // HAS_ARGBTORGB24ROW_NEON + +#ifdef HAS_ARGBTORAWROW_NEON +void ARGBToRAWRow_NEON(const uint8* src_argb, uint8* dst_raw, int pix) { + asm volatile ( + "1: \n" + MEMACCESS(0) + "ld4 {v1.8b,v2.8b,v3.8b,v4.8b}, [%0], #32 \n" // load b g r a + "subs %w2, %w2, #8 \n" // 8 processed per loop. + "orr v4.8b, v2.8b, v2.8b \n" // mov g + "orr v5.8b, v1.8b, v1.8b \n" // mov b + MEMACCESS(1) + "st3 {v3.8b,v4.8b,v5.8b}, [%1], #24 \n" // store r g b + "b.gt 1b \n" + : "+r"(src_argb), // %0 + "+r"(dst_raw), // %1 + "+r"(pix) // %2 + : + : "cc", "memory", "v1", "v2", "v3", "v4", "v5" // Clobber List + ); +} +#endif // HAS_ARGBTORAWROW_NEON + +#ifdef HAS_YUY2TOYROW_NEON +void YUY2ToYRow_NEON(const uint8* src_yuy2, uint8* dst_y, int pix) { + asm volatile ( + "1: \n" + MEMACCESS(0) + "ld2 {v0.16b,v1.16b}, [%0], #32 \n" // load 16 pixels of YUY2. + "subs %w2, %w2, #16 \n" // 16 processed per loop. + MEMACCESS(1) + "st1 {v0.16b}, [%1], #16 \n" // store 16 pixels of Y. + "b.gt 1b \n" + : "+r"(src_yuy2), // %0 + "+r"(dst_y), // %1 + "+r"(pix) // %2 + : + : "cc", "memory", "v0", "v1" // Clobber List + ); +} +#endif // HAS_YUY2TOYROW_NEON + +#ifdef HAS_UYVYTOYROW_NEON +void UYVYToYRow_NEON(const uint8* src_uyvy, uint8* dst_y, int pix) { + asm volatile ( + "1: \n" + MEMACCESS(0) + "ld2 {v0.16b,v1.16b}, [%0], #32 \n" // load 16 pixels of UYVY. + "subs %w2, %w2, #16 \n" // 16 processed per loop. + MEMACCESS(1) + "st1 {v1.16b}, [%1], #16 \n" // store 16 pixels of Y. + "b.gt 1b \n" + : "+r"(src_uyvy), // %0 + "+r"(dst_y), // %1 + "+r"(pix) // %2 + : + : "cc", "memory", "v0", "v1" // Clobber List + ); +} +#endif // HAS_UYVYTOYROW_NEON + +#ifdef HAS_YUY2TOUV422ROW_NEON +void YUY2ToUV422Row_NEON(const uint8* src_yuy2, uint8* dst_u, uint8* dst_v, + int pix) { + asm volatile ( + "1: \n" + MEMACCESS(0) + "ld4 {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n" // load 16 YUY2 pixels + "subs %w3, %w3, #16 \n" // 16 pixels = 8 UVs. + MEMACCESS(1) + "st1 {v1.8b}, [%1], #8 \n" // store 8 U. + MEMACCESS(2) + "st1 {v3.8b}, [%2], #8 \n" // store 8 V. + "b.gt 1b \n" + : "+r"(src_yuy2), // %0 + "+r"(dst_u), // %1 + "+r"(dst_v), // %2 + "+r"(pix) // %3 + : + : "cc", "memory", "v0", "v1", "v2", "v3" // Clobber List + ); +} +#endif // HAS_YUY2TOUV422ROW_NEON + +#ifdef HAS_UYVYTOUV422ROW_NEON +void UYVYToUV422Row_NEON(const uint8* src_uyvy, uint8* dst_u, uint8* dst_v, + int pix) { + asm volatile ( + "1: \n" + MEMACCESS(0) + "ld4 {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n" // load 16 UYVY pixels + "subs %w3, %w3, #16 \n" // 16 pixels = 8 UVs. + MEMACCESS(1) + "st1 {v0.8b}, [%1], #8 \n" // store 8 U. + MEMACCESS(2) + "st1 {v2.8b}, [%2], #8 \n" // store 8 V. + "b.gt 1b \n" + : "+r"(src_uyvy), // %0 + "+r"(dst_u), // %1 + "+r"(dst_v), // %2 + "+r"(pix) // %3 + : + : "cc", "memory", "v0", "v1", "v2", "v3" // Clobber List + ); +} +#endif // HAS_UYVYTOUV422ROW_NEON + +#ifdef HAS_YUY2TOUVROW_NEON +void YUY2ToUVRow_NEON(const uint8* src_yuy2, int stride_yuy2, + uint8* dst_u, uint8* dst_v, int pix) { + const uint8* src_yuy2b = src_yuy2 + stride_yuy2; + asm volatile ( + "1: \n" + MEMACCESS(0) + "ld4 {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n" // load 16 pixels + "subs %w4, %w4, #16 \n" // 16 pixels = 8 UVs. + MEMACCESS(1) + "ld4 {v4.8b,v5.8b,v6.8b,v7.8b}, [%1], #32 \n" // load next row + "urhadd v1.8b, v1.8b, v5.8b \n" // average rows of U + "urhadd v3.8b, v3.8b, v7.8b \n" // average rows of V + MEMACCESS(2) + "st1 {v1.8b}, [%2], #8 \n" // store 8 U. + MEMACCESS(3) + "st1 {v3.8b}, [%3], #8 \n" // store 8 V. + "b.gt 1b \n" + : "+r"(src_yuy2), // %0 + "+r"(src_yuy2b), // %1 + "+r"(dst_u), // %2 + "+r"(dst_v), // %3 + "+r"(pix) // %4 + : + : "cc", "memory", "v0", "v1", "v2", "v3", "v4", + "v5", "v6", "v7" // Clobber List + ); +} +#endif // HAS_YUY2TOUVROW_NEON + +#ifdef HAS_UYVYTOUVROW_NEON +void UYVYToUVRow_NEON(const uint8* src_uyvy, int stride_uyvy, + uint8* dst_u, uint8* dst_v, int pix) { + const uint8* src_uyvyb = src_uyvy + stride_uyvy; + asm volatile ( + "1: \n" + MEMACCESS(0) + "ld4 {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n" // load 16 pixels + "subs %w4, %w4, #16 \n" // 16 pixels = 8 UVs. + MEMACCESS(1) + "ld4 {v4.8b,v5.8b,v6.8b,v7.8b}, [%1], #32 \n" // load next row + "urhadd v0.8b, v0.8b, v4.8b \n" // average rows of U + "urhadd v2.8b, v2.8b, v6.8b \n" // average rows of V + MEMACCESS(2) + "st1 {v0.8b}, [%2], #8 \n" // store 8 U. + MEMACCESS(3) + "st1 {v2.8b}, [%3], #8 \n" // store 8 V. + "b.gt 1b \n" + : "+r"(src_uyvy), // %0 + "+r"(src_uyvyb), // %1 + "+r"(dst_u), // %2 + "+r"(dst_v), // %3 + "+r"(pix) // %4 + : + : "cc", "memory", "v0", "v1", "v2", "v3", "v4", + "v5", "v6", "v7" // Clobber List + ); +} +#endif // HAS_UYVYTOUVROW_NEON + +// For BGRAToARGB, ABGRToARGB, RGBAToARGB, and ARGBToRGBA. +#ifdef HAS_ARGBSHUFFLEROW_NEON +void ARGBShuffleRow_NEON(const uint8* src_argb, uint8* dst_argb, + const uint8* shuffler, int pix) { + asm volatile ( + MEMACCESS(3) + "ld1 {v2.16b}, [%3] \n" // shuffler + "1: \n" + MEMACCESS(0) + "ld1 {v0.16b}, [%0], #16 \n" // load 4 pixels. + "subs %w2, %w2, #4 \n" // 4 processed per loop + "tbl v1.16b, {v0.16b}, v2.16b \n" // look up 4 pixels + MEMACCESS(1) + "st1 {v1.16b}, [%1], #16 \n" // store 4. + "b.gt 1b \n" + : "+r"(src_argb), // %0 + "+r"(dst_argb), // %1 + "+r"(pix) // %2 + : "r"(shuffler) // %3 + : "cc", "memory", "v0", "v1", "v2" // Clobber List + ); +} +#endif // HAS_ARGBSHUFFLEROW_NEON + +#ifdef HAS_I422TOYUY2ROW_NEON +void I422ToYUY2Row_NEON(const uint8* src_y, + const uint8* src_u, + const uint8* src_v, + uint8* dst_yuy2, int width) { + asm volatile ( + "1: \n" + MEMACCESS(0) + "ld2 {v0.8b, v1.8b}, [%0], #16 \n" // load 16 Ys + "orr v2.8b, v1.8b, v1.8b \n" + MEMACCESS(1) + "ld1 {v1.8b}, [%1], #8 \n" // load 8 Us + MEMACCESS(2) + "ld1 {v3.8b}, [%2], #8 \n" // load 8 Vs + "subs %w4, %w4, #16 \n" // 16 pixels + MEMACCESS(3) + "st4 {v0.8b,v1.8b,v2.8b,v3.8b}, [%3], #32 \n" // Store 16 pixels. + "b.gt 1b \n" + : "+r"(src_y), // %0 + "+r"(src_u), // %1 + "+r"(src_v), // %2 + "+r"(dst_yuy2), // %3 + "+r"(width) // %4 + : + : "cc", "memory", "v0", "v1", "v2", "v3" + ); +} +#endif // HAS_I422TOYUY2ROW_NEON + +#ifdef HAS_I422TOUYVYROW_NEON +void I422ToUYVYRow_NEON(const uint8* src_y, + const uint8* src_u, + const uint8* src_v, + uint8* dst_uyvy, int width) { + asm volatile ( + "1: \n" + MEMACCESS(0) + "ld2 {v1.8b,v2.8b}, [%0], #16 \n" // load 16 Ys + "orr v3.8b, v2.8b, v2.8b \n" + MEMACCESS(1) + "ld1 {v0.8b}, [%1], #8 \n" // load 8 Us + MEMACCESS(2) + "ld1 {v2.8b}, [%2], #8 \n" // load 8 Vs + "subs %w4, %w4, #16 \n" // 16 pixels + MEMACCESS(3) + "st4 {v0.8b,v1.8b,v2.8b,v3.8b}, [%3], #32 \n" // Store 16 pixels. + "b.gt 1b \n" + : "+r"(src_y), // %0 + "+r"(src_u), // %1 + "+r"(src_v), // %2 + "+r"(dst_uyvy), // %3 + "+r"(width) // %4 + : + : "cc", "memory", "v0", "v1", "v2", "v3" + ); +} +#endif // HAS_I422TOUYVYROW_NEON + +#ifdef HAS_ARGBTORGB565ROW_NEON +void ARGBToRGB565Row_NEON(const uint8* src_argb, uint8* dst_rgb565, int pix) { + asm volatile ( + "1: \n" + MEMACCESS(0) + "ld4 {v20.8b,v21.8b,v22.8b,v23.8b}, [%0], #32 \n" // load 8 pixels + "subs %w2, %w2, #8 \n" // 8 processed per loop. + ARGBTORGB565 + MEMACCESS(1) + "st1 {v0.16b}, [%1], #16 \n" // store 8 pixels RGB565. + "b.gt 1b \n" + : "+r"(src_argb), // %0 + "+r"(dst_rgb565), // %1 + "+r"(pix) // %2 + : + : "cc", "memory", "v0", "v20", "v21", "v22", "v23" + ); +} +#endif // HAS_ARGBTORGB565ROW_NEON + +#ifdef HAS_ARGBTORGB565DITHERROW_NEON +void ARGBToRGB565DitherRow_NEON(const uint8* src_argb, uint8* dst_rgb, + const uint32 dither4, int width) { + asm volatile ( + "dup v1.4s, %w2 \n" // dither4 + "1: \n" + MEMACCESS(1) + "ld4 {v20.8b,v21.8b,v22.8b,v23.8b}, [%1], #32 \n" // load 8 pixels + "subs %w3, %w3, #8 \n" // 8 processed per loop. + "uqadd v20.8b, v20.8b, v1.8b \n" + "uqadd v21.8b, v21.8b, v1.8b \n" + "uqadd v22.8b, v22.8b, v1.8b \n" + ARGBTORGB565 + MEMACCESS(0) + "st1 {v0.16b}, [%0], #16 \n" // store 8 pixels RGB565. + "b.gt 1b \n" + : "+r"(dst_rgb) // %0 + : "r"(src_argb), // %1 + "r"(dither4), // %2 + "r"(width) // %3 + : "cc", "memory", "v0", "v1", "v20", "v21", "v22", "v23" + ); +} +#endif // HAS_ARGBTORGB565ROW_NEON + +#ifdef HAS_ARGBTOARGB1555ROW_NEON +void ARGBToARGB1555Row_NEON(const uint8* src_argb, uint8* dst_argb1555, + int pix) { + asm volatile ( + "1: \n" + MEMACCESS(0) + "ld4 {v20.8b,v21.8b,v22.8b,v23.8b}, [%0], #32 \n" // load 8 pixels + "subs %w2, %w2, #8 \n" // 8 processed per loop. + ARGBTOARGB1555 + MEMACCESS(1) + "st1 {v0.16b}, [%1], #16 \n" // store 8 pixels ARGB1555. + "b.gt 1b \n" + : "+r"(src_argb), // %0 + "+r"(dst_argb1555), // %1 + "+r"(pix) // %2 + : + : "cc", "memory", "v0", "v20", "v21", "v22", "v23" + ); +} +#endif // HAS_ARGBTOARGB1555ROW_NEON + +#ifdef HAS_ARGBTOARGB4444ROW_NEON +void ARGBToARGB4444Row_NEON(const uint8* src_argb, uint8* dst_argb4444, + int pix) { + asm volatile ( + "movi v4.16b, #0x0f \n" // bits to clear with vbic. + "1: \n" + MEMACCESS(0) + "ld4 {v20.8b,v21.8b,v22.8b,v23.8b}, [%0], #32 \n" // load 8 pixels + "subs %w2, %w2, #8 \n" // 8 processed per loop. + ARGBTOARGB4444 + MEMACCESS(1) + "st1 {v0.16b}, [%1], #16 \n" // store 8 pixels ARGB4444. + "b.gt 1b \n" + : "+r"(src_argb), // %0 + "+r"(dst_argb4444), // %1 + "+r"(pix) // %2 + : + : "cc", "memory", "v0", "v1", "v4", "v20", "v21", "v22", "v23" + ); +} +#endif // HAS_ARGBTOARGB4444ROW_NEON + +#ifdef HAS_ARGBTOYROW_NEON +void ARGBToYRow_NEON(const uint8* src_argb, uint8* dst_y, int pix) { + asm volatile ( + "movi v4.8b, #13 \n" // B * 0.1016 coefficient + "movi v5.8b, #65 \n" // G * 0.5078 coefficient + "movi v6.8b, #33 \n" // R * 0.2578 coefficient + "movi v7.8b, #16 \n" // Add 16 constant + "1: \n" + MEMACCESS(0) + "ld4 {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n" // load 8 ARGB pixels. + "subs %w2, %w2, #8 \n" // 8 processed per loop. + "umull v3.8h, v0.8b, v4.8b \n" // B + "umlal v3.8h, v1.8b, v5.8b \n" // G + "umlal v3.8h, v2.8b, v6.8b \n" // R + "sqrshrun v0.8b, v3.8h, #7 \n" // 16 bit to 8 bit Y + "uqadd v0.8b, v0.8b, v7.8b \n" + MEMACCESS(1) + "st1 {v0.8b}, [%1], #8 \n" // store 8 pixels Y. + "b.gt 1b \n" + : "+r"(src_argb), // %0 + "+r"(dst_y), // %1 + "+r"(pix) // %2 + : + : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7" + ); +} +#endif // HAS_ARGBTOYROW_NEON + +#ifdef HAS_ARGBTOYJROW_NEON +void ARGBToYJRow_NEON(const uint8* src_argb, uint8* dst_y, int pix) { + asm volatile ( + "movi v4.8b, #15 \n" // B * 0.11400 coefficient + "movi v5.8b, #75 \n" // G * 0.58700 coefficient + "movi v6.8b, #38 \n" // R * 0.29900 coefficient + "1: \n" + MEMACCESS(0) + "ld4 {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n" // load 8 ARGB pixels. + "subs %w2, %w2, #8 \n" // 8 processed per loop. + "umull v3.8h, v0.8b, v4.8b \n" // B + "umlal v3.8h, v1.8b, v5.8b \n" // G + "umlal v3.8h, v2.8b, v6.8b \n" // R + "sqrshrun v0.8b, v3.8h, #7 \n" // 15 bit to 8 bit Y + MEMACCESS(1) + "st1 {v0.8b}, [%1], #8 \n" // store 8 pixels Y. + "b.gt 1b \n" + : "+r"(src_argb), // %0 + "+r"(dst_y), // %1 + "+r"(pix) // %2 + : + : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6" + ); +} +#endif // HAS_ARGBTOYJROW_NEON + +// 8x1 pixels. +#ifdef HAS_ARGBTOUV444ROW_NEON +void ARGBToUV444Row_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v, + int pix) { + asm volatile ( + "movi v24.8b, #112 \n" // UB / VR 0.875 coefficient + "movi v25.8b, #74 \n" // UG -0.5781 coefficient + "movi v26.8b, #38 \n" // UR -0.2969 coefficient + "movi v27.8b, #18 \n" // VB -0.1406 coefficient + "movi v28.8b, #94 \n" // VG -0.7344 coefficient + "movi v29.16b,#0x80 \n" // 128.5 + "1: \n" + MEMACCESS(0) + "ld4 {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n" // load 8 ARGB pixels. + "subs %w3, %w3, #8 \n" // 8 processed per loop. + "umull v4.8h, v0.8b, v24.8b \n" // B + "umlsl v4.8h, v1.8b, v25.8b \n" // G + "umlsl v4.8h, v2.8b, v26.8b \n" // R + "add v4.8h, v4.8h, v29.8h \n" // +128 -> unsigned + + "umull v3.8h, v2.8b, v24.8b \n" // R + "umlsl v3.8h, v1.8b, v28.8b \n" // G + "umlsl v3.8h, v0.8b, v27.8b \n" // B + "add v3.8h, v3.8h, v29.8h \n" // +128 -> unsigned + + "uqshrn v0.8b, v4.8h, #8 \n" // 16 bit to 8 bit U + "uqshrn v1.8b, v3.8h, #8 \n" // 16 bit to 8 bit V + + MEMACCESS(1) + "st1 {v0.8b}, [%1], #8 \n" // store 8 pixels U. + MEMACCESS(2) + "st1 {v1.8b}, [%2], #8 \n" // store 8 pixels V. + "b.gt 1b \n" + : "+r"(src_argb), // %0 + "+r"(dst_u), // %1 + "+r"(dst_v), // %2 + "+r"(pix) // %3 + : + : "cc", "memory", "v0", "v1", "v2", "v3", "v4", + "v24", "v25", "v26", "v27", "v28", "v29" + ); +} +#endif // HAS_ARGBTOUV444ROW_NEON + +// 16x1 pixels -> 8x1. pix is number of argb pixels. e.g. 16. +#ifdef HAS_ARGBTOUV422ROW_NEON +void ARGBToUV422Row_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v, + int pix) { + asm volatile ( + RGBTOUV_SETUP_REG + "1: \n" + MEMACCESS(0) + "ld4 {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64 \n" // load 16 pixels. + + "uaddlp v0.8h, v0.16b \n" // B 16 bytes -> 8 shorts. + "uaddlp v1.8h, v1.16b \n" // G 16 bytes -> 8 shorts. + "uaddlp v2.8h, v2.16b \n" // R 16 bytes -> 8 shorts. + + "subs %w3, %w3, #16 \n" // 16 processed per loop. + "mul v3.8h, v0.8h, v20.8h \n" // B + "mls v3.8h, v1.8h, v21.8h \n" // G + "mls v3.8h, v2.8h, v22.8h \n" // R + "add v3.8h, v3.8h, v25.8h \n" // +128 -> unsigned + + "mul v4.8h, v2.8h, v20.8h \n" // R + "mls v4.8h, v1.8h, v24.8h \n" // G + "mls v4.8h, v0.8h, v23.8h \n" // B + "add v4.8h, v4.8h, v25.8h \n" // +128 -> unsigned + + "uqshrn v0.8b, v3.8h, #8 \n" // 16 bit to 8 bit U + "uqshrn v1.8b, v4.8h, #8 \n" // 16 bit to 8 bit V + + MEMACCESS(1) + "st1 {v0.8b}, [%1], #8 \n" // store 8 pixels U. + MEMACCESS(2) + "st1 {v1.8b}, [%2], #8 \n" // store 8 pixels V. + "b.gt 1b \n" + : "+r"(src_argb), // %0 + "+r"(dst_u), // %1 + "+r"(dst_v), // %2 + "+r"(pix) // %3 + : + : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", + "v20", "v21", "v22", "v23", "v24", "v25" + ); +} +#endif // HAS_ARGBTOUV422ROW_NEON + +// 32x1 pixels -> 8x1. pix is number of argb pixels. e.g. 32. +#ifdef HAS_ARGBTOUV411ROW_NEON +void ARGBToUV411Row_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v, + int pix) { + asm volatile ( + RGBTOUV_SETUP_REG + "1: \n" + MEMACCESS(0) + "ld4 {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64 \n" // load 16 pixels. + "uaddlp v0.8h, v0.16b \n" // B 16 bytes -> 8 shorts. + "uaddlp v1.8h, v1.16b \n" // G 16 bytes -> 8 shorts. + "uaddlp v2.8h, v2.16b \n" // R 16 bytes -> 8 shorts. + MEMACCESS(0) + "ld4 {v4.16b,v5.16b,v6.16b,v7.16b}, [%0], #64 \n" // load next 16. + "uaddlp v4.8h, v4.16b \n" // B 16 bytes -> 8 shorts. + "uaddlp v5.8h, v5.16b \n" // G 16 bytes -> 8 shorts. + "uaddlp v6.8h, v6.16b \n" // R 16 bytes -> 8 shorts. + + "addp v0.8h, v0.8h, v4.8h \n" // B 16 shorts -> 8 shorts. + "addp v1.8h, v1.8h, v5.8h \n" // G 16 shorts -> 8 shorts. + "addp v2.8h, v2.8h, v6.8h \n" // R 16 shorts -> 8 shorts. + + "urshr v0.8h, v0.8h, #1 \n" // 2x average + "urshr v1.8h, v1.8h, #1 \n" + "urshr v2.8h, v2.8h, #1 \n" + + "subs %w3, %w3, #32 \n" // 32 processed per loop. + "mul v3.8h, v0.8h, v20.8h \n" // B + "mls v3.8h, v1.8h, v21.8h \n" // G + "mls v3.8h, v2.8h, v22.8h \n" // R + "add v3.8h, v3.8h, v25.8h \n" // +128 -> unsigned + "mul v4.8h, v2.8h, v20.8h \n" // R + "mls v4.8h, v1.8h, v24.8h \n" // G + "mls v4.8h, v0.8h, v23.8h \n" // B + "add v4.8h, v4.8h, v25.8h \n" // +128 -> unsigned + "uqshrn v0.8b, v3.8h, #8 \n" // 16 bit to 8 bit U + "uqshrn v1.8b, v4.8h, #8 \n" // 16 bit to 8 bit V + MEMACCESS(1) + "st1 {v0.8b}, [%1], #8 \n" // store 8 pixels U. + MEMACCESS(2) + "st1 {v1.8b}, [%2], #8 \n" // store 8 pixels V. + "b.gt 1b \n" + : "+r"(src_argb), // %0 + "+r"(dst_u), // %1 + "+r"(dst_v), // %2 + "+r"(pix) // %3 + : + : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", + "v20", "v21", "v22", "v23", "v24", "v25" + ); +} +#endif // HAS_ARGBTOUV411ROW_NEON + +// 16x2 pixels -> 8x1. pix is number of argb pixels. e.g. 16. +#define RGBTOUV(QB, QG, QR) \ + "mul v3.8h, " #QB ",v20.8h \n" /* B */ \ + "mul v4.8h, " #QR ",v20.8h \n" /* R */ \ + "mls v3.8h, " #QG ",v21.8h \n" /* G */ \ + "mls v4.8h, " #QG ",v24.8h \n" /* G */ \ + "mls v3.8h, " #QR ",v22.8h \n" /* R */ \ + "mls v4.8h, " #QB ",v23.8h \n" /* B */ \ + "add v3.8h, v3.8h, v25.8h \n" /* +128 -> unsigned */ \ + "add v4.8h, v4.8h, v25.8h \n" /* +128 -> unsigned */ \ + "uqshrn v0.8b, v3.8h, #8 \n" /* 16 bit to 8 bit U */ \ + "uqshrn v1.8b, v4.8h, #8 \n" /* 16 bit to 8 bit V */ + +// TODO(fbarchard): Consider vhadd vertical, then vpaddl horizontal, avoid shr. +// TODO(fbarchard): consider ptrdiff_t for all strides. + +#ifdef HAS_ARGBTOUVROW_NEON +void ARGBToUVRow_NEON(const uint8* src_argb, int src_stride_argb, + uint8* dst_u, uint8* dst_v, int pix) { + const uint8* src_argb_1 = src_argb + src_stride_argb; + asm volatile ( + RGBTOUV_SETUP_REG + "1: \n" + MEMACCESS(0) + "ld4 {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64 \n" // load 16 pixels. + "uaddlp v0.8h, v0.16b \n" // B 16 bytes -> 8 shorts. + "uaddlp v1.8h, v1.16b \n" // G 16 bytes -> 8 shorts. + "uaddlp v2.8h, v2.16b \n" // R 16 bytes -> 8 shorts. + + MEMACCESS(1) + "ld4 {v4.16b,v5.16b,v6.16b,v7.16b}, [%1], #64 \n" // load next 16 + "uadalp v0.8h, v4.16b \n" // B 16 bytes -> 8 shorts. + "uadalp v1.8h, v5.16b \n" // G 16 bytes -> 8 shorts. + "uadalp v2.8h, v6.16b \n" // R 16 bytes -> 8 shorts. + + "urshr v0.8h, v0.8h, #1 \n" // 2x average + "urshr v1.8h, v1.8h, #1 \n" + "urshr v2.8h, v2.8h, #1 \n" + + "subs %w4, %w4, #16 \n" // 32 processed per loop. + RGBTOUV(v0.8h, v1.8h, v2.8h) + MEMACCESS(2) + "st1 {v0.8b}, [%2], #8 \n" // store 8 pixels U. + MEMACCESS(3) + "st1 {v1.8b}, [%3], #8 \n" // store 8 pixels V. + "b.gt 1b \n" + : "+r"(src_argb), // %0 + "+r"(src_argb_1), // %1 + "+r"(dst_u), // %2 + "+r"(dst_v), // %3 + "+r"(pix) // %4 + : + : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", + "v20", "v21", "v22", "v23", "v24", "v25" + ); +} +#endif // HAS_ARGBTOUVROW_NEON + +// TODO(fbarchard): Subsample match C code. +#ifdef HAS_ARGBTOUVJROW_NEON +void ARGBToUVJRow_NEON(const uint8* src_argb, int src_stride_argb, + uint8* dst_u, uint8* dst_v, int pix) { + const uint8* src_argb_1 = src_argb + src_stride_argb; + asm volatile ( + "movi v20.8h, #63, lsl #0 \n" // UB/VR coeff (0.500) / 2 + "movi v21.8h, #42, lsl #0 \n" // UG coeff (-0.33126) / 2 + "movi v22.8h, #21, lsl #0 \n" // UR coeff (-0.16874) / 2 + "movi v23.8h, #10, lsl #0 \n" // VB coeff (-0.08131) / 2 + "movi v24.8h, #53, lsl #0 \n" // VG coeff (-0.41869) / 2 + "movi v25.16b, #0x80 \n" // 128.5 (0x8080 in 16-bit) + "1: \n" + MEMACCESS(0) + "ld4 {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64 \n" // load 16 pixels. + "uaddlp v0.8h, v0.16b \n" // B 16 bytes -> 8 shorts. + "uaddlp v1.8h, v1.16b \n" // G 16 bytes -> 8 shorts. + "uaddlp v2.8h, v2.16b \n" // R 16 bytes -> 8 shorts. + MEMACCESS(1) + "ld4 {v4.16b,v5.16b,v6.16b,v7.16b}, [%1], #64 \n" // load next 16 + "uadalp v0.8h, v4.16b \n" // B 16 bytes -> 8 shorts. + "uadalp v1.8h, v5.16b \n" // G 16 bytes -> 8 shorts. + "uadalp v2.8h, v6.16b \n" // R 16 bytes -> 8 shorts. + + "urshr v0.8h, v0.8h, #1 \n" // 2x average + "urshr v1.8h, v1.8h, #1 \n" + "urshr v2.8h, v2.8h, #1 \n" + + "subs %w4, %w4, #16 \n" // 32 processed per loop. + RGBTOUV(v0.8h, v1.8h, v2.8h) + MEMACCESS(2) + "st1 {v0.8b}, [%2], #8 \n" // store 8 pixels U. + MEMACCESS(3) + "st1 {v1.8b}, [%3], #8 \n" // store 8 pixels V. + "b.gt 1b \n" + : "+r"(src_argb), // %0 + "+r"(src_argb_1), // %1 + "+r"(dst_u), // %2 + "+r"(dst_v), // %3 + "+r"(pix) // %4 + : + : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", + "v20", "v21", "v22", "v23", "v24", "v25" + ); +} +#endif // HAS_ARGBTOUVJROW_NEON + +#ifdef HAS_BGRATOUVROW_NEON +void BGRAToUVRow_NEON(const uint8* src_bgra, int src_stride_bgra, + uint8* dst_u, uint8* dst_v, int pix) { + const uint8* src_bgra_1 = src_bgra + src_stride_bgra; + asm volatile ( + RGBTOUV_SETUP_REG + "1: \n" + MEMACCESS(0) + "ld4 {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64 \n" // load 16 pixels. + "uaddlp v0.8h, v3.16b \n" // B 16 bytes -> 8 shorts. + "uaddlp v3.8h, v2.16b \n" // G 16 bytes -> 8 shorts. + "uaddlp v2.8h, v1.16b \n" // R 16 bytes -> 8 shorts. + MEMACCESS(1) + "ld4 {v4.16b,v5.16b,v6.16b,v7.16b}, [%1], #64 \n" // load 16 more + "uadalp v0.8h, v7.16b \n" // B 16 bytes -> 8 shorts. + "uadalp v3.8h, v6.16b \n" // G 16 bytes -> 8 shorts. + "uadalp v2.8h, v5.16b \n" // R 16 bytes -> 8 shorts. + + "urshr v0.8h, v0.8h, #1 \n" // 2x average + "urshr v1.8h, v3.8h, #1 \n" + "urshr v2.8h, v2.8h, #1 \n" + + "subs %w4, %w4, #16 \n" // 32 processed per loop. + RGBTOUV(v0.8h, v1.8h, v2.8h) + MEMACCESS(2) + "st1 {v0.8b}, [%2], #8 \n" // store 8 pixels U. + MEMACCESS(3) + "st1 {v1.8b}, [%3], #8 \n" // store 8 pixels V. + "b.gt 1b \n" + : "+r"(src_bgra), // %0 + "+r"(src_bgra_1), // %1 + "+r"(dst_u), // %2 + "+r"(dst_v), // %3 + "+r"(pix) // %4 + : + : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", + "v20", "v21", "v22", "v23", "v24", "v25" + ); +} +#endif // HAS_BGRATOUVROW_NEON + +#ifdef HAS_ABGRTOUVROW_NEON +void ABGRToUVRow_NEON(const uint8* src_abgr, int src_stride_abgr, + uint8* dst_u, uint8* dst_v, int pix) { + const uint8* src_abgr_1 = src_abgr + src_stride_abgr; + asm volatile ( + RGBTOUV_SETUP_REG + "1: \n" + MEMACCESS(0) + "ld4 {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64 \n" // load 16 pixels. + "uaddlp v3.8h, v2.16b \n" // B 16 bytes -> 8 shorts. + "uaddlp v2.8h, v1.16b \n" // G 16 bytes -> 8 shorts. + "uaddlp v1.8h, v0.16b \n" // R 16 bytes -> 8 shorts. + MEMACCESS(1) + "ld4 {v4.16b,v5.16b,v6.16b,v7.16b}, [%1], #64 \n" // load 16 more. + "uadalp v3.8h, v6.16b \n" // B 16 bytes -> 8 shorts. + "uadalp v2.8h, v5.16b \n" // G 16 bytes -> 8 shorts. + "uadalp v1.8h, v4.16b \n" // R 16 bytes -> 8 shorts. + + "urshr v0.8h, v3.8h, #1 \n" // 2x average + "urshr v2.8h, v2.8h, #1 \n" + "urshr v1.8h, v1.8h, #1 \n" + + "subs %w4, %w4, #16 \n" // 32 processed per loop. + RGBTOUV(v0.8h, v2.8h, v1.8h) + MEMACCESS(2) + "st1 {v0.8b}, [%2], #8 \n" // store 8 pixels U. + MEMACCESS(3) + "st1 {v1.8b}, [%3], #8 \n" // store 8 pixels V. + "b.gt 1b \n" + : "+r"(src_abgr), // %0 + "+r"(src_abgr_1), // %1 + "+r"(dst_u), // %2 + "+r"(dst_v), // %3 + "+r"(pix) // %4 + : + : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", + "v20", "v21", "v22", "v23", "v24", "v25" + ); +} +#endif // HAS_ABGRTOUVROW_NEON + +#ifdef HAS_RGBATOUVROW_NEON +void RGBAToUVRow_NEON(const uint8* src_rgba, int src_stride_rgba, + uint8* dst_u, uint8* dst_v, int pix) { + const uint8* src_rgba_1 = src_rgba + src_stride_rgba; + asm volatile ( + RGBTOUV_SETUP_REG + "1: \n" + MEMACCESS(0) + "ld4 {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64 \n" // load 16 pixels. + "uaddlp v0.8h, v1.16b \n" // B 16 bytes -> 8 shorts. + "uaddlp v1.8h, v2.16b \n" // G 16 bytes -> 8 shorts. + "uaddlp v2.8h, v3.16b \n" // R 16 bytes -> 8 shorts. + MEMACCESS(1) + "ld4 {v4.16b,v5.16b,v6.16b,v7.16b}, [%1], #64 \n" // load 16 more. + "uadalp v0.8h, v5.16b \n" // B 16 bytes -> 8 shorts. + "uadalp v1.8h, v6.16b \n" // G 16 bytes -> 8 shorts. + "uadalp v2.8h, v7.16b \n" // R 16 bytes -> 8 shorts. + + "urshr v0.8h, v0.8h, #1 \n" // 2x average + "urshr v1.8h, v1.8h, #1 \n" + "urshr v2.8h, v2.8h, #1 \n" + + "subs %w4, %w4, #16 \n" // 32 processed per loop. + RGBTOUV(v0.8h, v1.8h, v2.8h) + MEMACCESS(2) + "st1 {v0.8b}, [%2], #8 \n" // store 8 pixels U. + MEMACCESS(3) + "st1 {v1.8b}, [%3], #8 \n" // store 8 pixels V. + "b.gt 1b \n" + : "+r"(src_rgba), // %0 + "+r"(src_rgba_1), // %1 + "+r"(dst_u), // %2 + "+r"(dst_v), // %3 + "+r"(pix) // %4 + : + : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", + "v20", "v21", "v22", "v23", "v24", "v25" + ); +} +#endif // HAS_RGBATOUVROW_NEON + +#ifdef HAS_RGB24TOUVROW_NEON +void RGB24ToUVRow_NEON(const uint8* src_rgb24, int src_stride_rgb24, + uint8* dst_u, uint8* dst_v, int pix) { + const uint8* src_rgb24_1 = src_rgb24 + src_stride_rgb24; + asm volatile ( + RGBTOUV_SETUP_REG + "1: \n" + MEMACCESS(0) + "ld3 {v0.16b,v1.16b,v2.16b}, [%0], #48 \n" // load 16 pixels. + "uaddlp v0.8h, v0.16b \n" // B 16 bytes -> 8 shorts. + "uaddlp v1.8h, v1.16b \n" // G 16 bytes -> 8 shorts. + "uaddlp v2.8h, v2.16b \n" // R 16 bytes -> 8 shorts. + MEMACCESS(1) + "ld3 {v4.16b,v5.16b,v6.16b}, [%1], #48 \n" // load 16 more. + "uadalp v0.8h, v4.16b \n" // B 16 bytes -> 8 shorts. + "uadalp v1.8h, v5.16b \n" // G 16 bytes -> 8 shorts. + "uadalp v2.8h, v6.16b \n" // R 16 bytes -> 8 shorts. + + "urshr v0.8h, v0.8h, #1 \n" // 2x average + "urshr v1.8h, v1.8h, #1 \n" + "urshr v2.8h, v2.8h, #1 \n" + + "subs %w4, %w4, #16 \n" // 32 processed per loop. + RGBTOUV(v0.8h, v1.8h, v2.8h) + MEMACCESS(2) + "st1 {v0.8b}, [%2], #8 \n" // store 8 pixels U. + MEMACCESS(3) + "st1 {v1.8b}, [%3], #8 \n" // store 8 pixels V. + "b.gt 1b \n" + : "+r"(src_rgb24), // %0 + "+r"(src_rgb24_1), // %1 + "+r"(dst_u), // %2 + "+r"(dst_v), // %3 + "+r"(pix) // %4 + : + : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", + "v20", "v21", "v22", "v23", "v24", "v25" + ); +} +#endif // HAS_RGB24TOUVROW_NEON + +#ifdef HAS_RAWTOUVROW_NEON +void RAWToUVRow_NEON(const uint8* src_raw, int src_stride_raw, + uint8* dst_u, uint8* dst_v, int pix) { + const uint8* src_raw_1 = src_raw + src_stride_raw; + asm volatile ( + RGBTOUV_SETUP_REG + "1: \n" + MEMACCESS(0) + "ld3 {v0.16b,v1.16b,v2.16b}, [%0], #48 \n" // load 8 RAW pixels. + "uaddlp v2.8h, v2.16b \n" // B 16 bytes -> 8 shorts. + "uaddlp v1.8h, v1.16b \n" // G 16 bytes -> 8 shorts. + "uaddlp v0.8h, v0.16b \n" // R 16 bytes -> 8 shorts. + MEMACCESS(1) + "ld3 {v4.16b,v5.16b,v6.16b}, [%1], #48 \n" // load 8 more RAW pixels + "uadalp v2.8h, v6.16b \n" // B 16 bytes -> 8 shorts. + "uadalp v1.8h, v5.16b \n" // G 16 bytes -> 8 shorts. + "uadalp v0.8h, v4.16b \n" // R 16 bytes -> 8 shorts. + + "urshr v2.8h, v2.8h, #1 \n" // 2x average + "urshr v1.8h, v1.8h, #1 \n" + "urshr v0.8h, v0.8h, #1 \n" + + "subs %w4, %w4, #16 \n" // 32 processed per loop. + RGBTOUV(v2.8h, v1.8h, v0.8h) + MEMACCESS(2) + "st1 {v0.8b}, [%2], #8 \n" // store 8 pixels U. + MEMACCESS(3) + "st1 {v1.8b}, [%3], #8 \n" // store 8 pixels V. + "b.gt 1b \n" + : "+r"(src_raw), // %0 + "+r"(src_raw_1), // %1 + "+r"(dst_u), // %2 + "+r"(dst_v), // %3 + "+r"(pix) // %4 + : + : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", + "v20", "v21", "v22", "v23", "v24", "v25" + ); +} +#endif // HAS_RAWTOUVROW_NEON + +// 16x2 pixels -> 8x1. pix is number of argb pixels. e.g. 16. +#ifdef HAS_RGB565TOUVROW_NEON +void RGB565ToUVRow_NEON(const uint8* src_rgb565, int src_stride_rgb565, + uint8* dst_u, uint8* dst_v, int pix) { + const uint8* src_rgb565_1 = src_rgb565 + src_stride_rgb565; + asm volatile ( + "movi v22.8h, #56, lsl #0 \n" // UB / VR coeff (0.875) / 2 + "movi v23.8h, #37, lsl #0 \n" // UG coeff (-0.5781) / 2 + "movi v24.8h, #19, lsl #0 \n" // UR coeff (-0.2969) / 2 + "movi v25.8h, #9 , lsl #0 \n" // VB coeff (-0.1406) / 2 + "movi v26.8h, #47, lsl #0 \n" // VG coeff (-0.7344) / 2 + "movi v27.16b, #0x80 \n" // 128.5 (0x8080 in 16-bit) + "1: \n" + MEMACCESS(0) + "ld1 {v0.16b}, [%0], #16 \n" // load 8 RGB565 pixels. + RGB565TOARGB + "uaddlp v16.4h, v0.8b \n" // B 8 bytes -> 4 shorts. + "uaddlp v18.4h, v1.8b \n" // G 8 bytes -> 4 shorts. + "uaddlp v20.4h, v2.8b \n" // R 8 bytes -> 4 shorts. + MEMACCESS(0) + "ld1 {v0.16b}, [%0], #16 \n" // next 8 RGB565 pixels. + RGB565TOARGB + "uaddlp v17.4h, v0.8b \n" // B 8 bytes -> 4 shorts. + "uaddlp v19.4h, v1.8b \n" // G 8 bytes -> 4 shorts. + "uaddlp v21.4h, v2.8b \n" // R 8 bytes -> 4 shorts. + + MEMACCESS(1) + "ld1 {v0.16b}, [%1], #16 \n" // load 8 RGB565 pixels. + RGB565TOARGB + "uadalp v16.4h, v0.8b \n" // B 8 bytes -> 4 shorts. + "uadalp v18.4h, v1.8b \n" // G 8 bytes -> 4 shorts. + "uadalp v20.4h, v2.8b \n" // R 8 bytes -> 4 shorts. + MEMACCESS(1) + "ld1 {v0.16b}, [%1], #16 \n" // next 8 RGB565 pixels. + RGB565TOARGB + "uadalp v17.4h, v0.8b \n" // B 8 bytes -> 4 shorts. + "uadalp v19.4h, v1.8b \n" // G 8 bytes -> 4 shorts. + "uadalp v21.4h, v2.8b \n" // R 8 bytes -> 4 shorts. + + "ins v16.D[1], v17.D[0] \n" + "ins v18.D[1], v19.D[0] \n" + "ins v20.D[1], v21.D[0] \n" + + "urshr v4.8h, v16.8h, #1 \n" // 2x average + "urshr v5.8h, v18.8h, #1 \n" + "urshr v6.8h, v20.8h, #1 \n" + + "subs %w4, %w4, #16 \n" // 16 processed per loop. + "mul v16.8h, v4.8h, v22.8h \n" // B + "mls v16.8h, v5.8h, v23.8h \n" // G + "mls v16.8h, v6.8h, v24.8h \n" // R + "add v16.8h, v16.8h, v27.8h \n" // +128 -> unsigned + "mul v17.8h, v6.8h, v22.8h \n" // R + "mls v17.8h, v5.8h, v26.8h \n" // G + "mls v17.8h, v4.8h, v25.8h \n" // B + "add v17.8h, v17.8h, v27.8h \n" // +128 -> unsigned + "uqshrn v0.8b, v16.8h, #8 \n" // 16 bit to 8 bit U + "uqshrn v1.8b, v17.8h, #8 \n" // 16 bit to 8 bit V + MEMACCESS(2) + "st1 {v0.8b}, [%2], #8 \n" // store 8 pixels U. + MEMACCESS(3) + "st1 {v1.8b}, [%3], #8 \n" // store 8 pixels V. + "b.gt 1b \n" + : "+r"(src_rgb565), // %0 + "+r"(src_rgb565_1), // %1 + "+r"(dst_u), // %2 + "+r"(dst_v), // %3 + "+r"(pix) // %4 + : + : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", + "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", + "v25", "v26", "v27" + ); +} +#endif // HAS_RGB565TOUVROW_NEON + +// 16x2 pixels -> 8x1. pix is number of argb pixels. e.g. 16. +#ifdef HAS_ARGB1555TOUVROW_NEON +void ARGB1555ToUVRow_NEON(const uint8* src_argb1555, int src_stride_argb1555, + uint8* dst_u, uint8* dst_v, int pix) { + const uint8* src_argb1555_1 = src_argb1555 + src_stride_argb1555; + asm volatile ( + RGBTOUV_SETUP_REG + "1: \n" + MEMACCESS(0) + "ld1 {v0.16b}, [%0], #16 \n" // load 8 ARGB1555 pixels. + RGB555TOARGB + "uaddlp v16.4h, v0.8b \n" // B 8 bytes -> 4 shorts. + "uaddlp v17.4h, v1.8b \n" // G 8 bytes -> 4 shorts. + "uaddlp v18.4h, v2.8b \n" // R 8 bytes -> 4 shorts. + MEMACCESS(0) + "ld1 {v0.16b}, [%0], #16 \n" // next 8 ARGB1555 pixels. + RGB555TOARGB + "uaddlp v26.4h, v0.8b \n" // B 8 bytes -> 4 shorts. + "uaddlp v27.4h, v1.8b \n" // G 8 bytes -> 4 shorts. + "uaddlp v28.4h, v2.8b \n" // R 8 bytes -> 4 shorts. + + MEMACCESS(1) + "ld1 {v0.16b}, [%1], #16 \n" // load 8 ARGB1555 pixels. + RGB555TOARGB + "uadalp v16.4h, v0.8b \n" // B 8 bytes -> 4 shorts. + "uadalp v17.4h, v1.8b \n" // G 8 bytes -> 4 shorts. + "uadalp v18.4h, v2.8b \n" // R 8 bytes -> 4 shorts. + MEMACCESS(1) + "ld1 {v0.16b}, [%1], #16 \n" // next 8 ARGB1555 pixels. + RGB555TOARGB + "uadalp v26.4h, v0.8b \n" // B 8 bytes -> 4 shorts. + "uadalp v27.4h, v1.8b \n" // G 8 bytes -> 4 shorts. + "uadalp v28.4h, v2.8b \n" // R 8 bytes -> 4 shorts. + + "ins v16.D[1], v26.D[0] \n" + "ins v17.D[1], v27.D[0] \n" + "ins v18.D[1], v28.D[0] \n" + + "urshr v4.8h, v16.8h, #1 \n" // 2x average + "urshr v5.8h, v17.8h, #1 \n" + "urshr v6.8h, v18.8h, #1 \n" + + "subs %w4, %w4, #16 \n" // 16 processed per loop. + "mul v2.8h, v4.8h, v20.8h \n" // B + "mls v2.8h, v5.8h, v21.8h \n" // G + "mls v2.8h, v6.8h, v22.8h \n" // R + "add v2.8h, v2.8h, v25.8h \n" // +128 -> unsigned + "mul v3.8h, v6.8h, v20.8h \n" // R + "mls v3.8h, v5.8h, v24.8h \n" // G + "mls v3.8h, v4.8h, v23.8h \n" // B + "add v3.8h, v3.8h, v25.8h \n" // +128 -> unsigned + "uqshrn v0.8b, v2.8h, #8 \n" // 16 bit to 8 bit U + "uqshrn v1.8b, v3.8h, #8 \n" // 16 bit to 8 bit V + MEMACCESS(2) + "st1 {v0.8b}, [%2], #8 \n" // store 8 pixels U. + MEMACCESS(3) + "st1 {v1.8b}, [%3], #8 \n" // store 8 pixels V. + "b.gt 1b \n" + : "+r"(src_argb1555), // %0 + "+r"(src_argb1555_1), // %1 + "+r"(dst_u), // %2 + "+r"(dst_v), // %3 + "+r"(pix) // %4 + : + : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", + "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", + "v26", "v27", "v28" + ); +} +#endif // HAS_ARGB1555TOUVROW_NEON + +// 16x2 pixels -> 8x1. pix is number of argb pixels. e.g. 16. +#ifdef HAS_ARGB4444TOUVROW_NEON +void ARGB4444ToUVRow_NEON(const uint8* src_argb4444, int src_stride_argb4444, + uint8* dst_u, uint8* dst_v, int pix) { + const uint8* src_argb4444_1 = src_argb4444 + src_stride_argb4444; + asm volatile ( + RGBTOUV_SETUP_REG + "1: \n" + MEMACCESS(0) + "ld1 {v0.16b}, [%0], #16 \n" // load 8 ARGB4444 pixels. + ARGB4444TOARGB + "uaddlp v16.4h, v0.8b \n" // B 8 bytes -> 4 shorts. + "uaddlp v17.4h, v1.8b \n" // G 8 bytes -> 4 shorts. + "uaddlp v18.4h, v2.8b \n" // R 8 bytes -> 4 shorts. + MEMACCESS(0) + "ld1 {v0.16b}, [%0], #16 \n" // next 8 ARGB4444 pixels. + ARGB4444TOARGB + "uaddlp v26.4h, v0.8b \n" // B 8 bytes -> 4 shorts. + "uaddlp v27.4h, v1.8b \n" // G 8 bytes -> 4 shorts. + "uaddlp v28.4h, v2.8b \n" // R 8 bytes -> 4 shorts. + + MEMACCESS(1) + "ld1 {v0.16b}, [%1], #16 \n" // load 8 ARGB4444 pixels. + ARGB4444TOARGB + "uadalp v16.4h, v0.8b \n" // B 8 bytes -> 4 shorts. + "uadalp v17.4h, v1.8b \n" // G 8 bytes -> 4 shorts. + "uadalp v18.4h, v2.8b \n" // R 8 bytes -> 4 shorts. + MEMACCESS(1) + "ld1 {v0.16b}, [%1], #16 \n" // next 8 ARGB4444 pixels. + ARGB4444TOARGB + "uadalp v26.4h, v0.8b \n" // B 8 bytes -> 4 shorts. + "uadalp v27.4h, v1.8b \n" // G 8 bytes -> 4 shorts. + "uadalp v28.4h, v2.8b \n" // R 8 bytes -> 4 shorts. + + "ins v16.D[1], v26.D[0] \n" + "ins v17.D[1], v27.D[0] \n" + "ins v18.D[1], v28.D[0] \n" + + "urshr v4.8h, v16.8h, #1 \n" // 2x average + "urshr v5.8h, v17.8h, #1 \n" + "urshr v6.8h, v18.8h, #1 \n" + + "subs %w4, %w4, #16 \n" // 16 processed per loop. + "mul v2.8h, v4.8h, v20.8h \n" // B + "mls v2.8h, v5.8h, v21.8h \n" // G + "mls v2.8h, v6.8h, v22.8h \n" // R + "add v2.8h, v2.8h, v25.8h \n" // +128 -> unsigned + "mul v3.8h, v6.8h, v20.8h \n" // R + "mls v3.8h, v5.8h, v24.8h \n" // G + "mls v3.8h, v4.8h, v23.8h \n" // B + "add v3.8h, v3.8h, v25.8h \n" // +128 -> unsigned + "uqshrn v0.8b, v2.8h, #8 \n" // 16 bit to 8 bit U + "uqshrn v1.8b, v3.8h, #8 \n" // 16 bit to 8 bit V + MEMACCESS(2) + "st1 {v0.8b}, [%2], #8 \n" // store 8 pixels U. + MEMACCESS(3) + "st1 {v1.8b}, [%3], #8 \n" // store 8 pixels V. + "b.gt 1b \n" + : "+r"(src_argb4444), // %0 + "+r"(src_argb4444_1), // %1 + "+r"(dst_u), // %2 + "+r"(dst_v), // %3 + "+r"(pix) // %4 + : + : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", + "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", + "v26", "v27", "v28" + + ); +} +#endif // HAS_ARGB4444TOUVROW_NEON + +#ifdef HAS_RGB565TOYROW_NEON +void RGB565ToYRow_NEON(const uint8* src_rgb565, uint8* dst_y, int pix) { + asm volatile ( + "movi v24.8b, #13 \n" // B * 0.1016 coefficient + "movi v25.8b, #65 \n" // G * 0.5078 coefficient + "movi v26.8b, #33 \n" // R * 0.2578 coefficient + "movi v27.8b, #16 \n" // Add 16 constant + "1: \n" + MEMACCESS(0) + "ld1 {v0.16b}, [%0], #16 \n" // load 8 RGB565 pixels. + "subs %w2, %w2, #8 \n" // 8 processed per loop. + RGB565TOARGB + "umull v3.8h, v0.8b, v24.8b \n" // B + "umlal v3.8h, v1.8b, v25.8b \n" // G + "umlal v3.8h, v2.8b, v26.8b \n" // R + "sqrshrun v0.8b, v3.8h, #7 \n" // 16 bit to 8 bit Y + "uqadd v0.8b, v0.8b, v27.8b \n" + MEMACCESS(1) + "st1 {v0.8b}, [%1], #8 \n" // store 8 pixels Y. + "b.gt 1b \n" + : "+r"(src_rgb565), // %0 + "+r"(dst_y), // %1 + "+r"(pix) // %2 + : + : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v6", + "v24", "v25", "v26", "v27" + ); +} +#endif // HAS_RGB565TOYROW_NEON + +#ifdef HAS_ARGB1555TOYROW_NEON +void ARGB1555ToYRow_NEON(const uint8* src_argb1555, uint8* dst_y, int pix) { + asm volatile ( + "movi v4.8b, #13 \n" // B * 0.1016 coefficient + "movi v5.8b, #65 \n" // G * 0.5078 coefficient + "movi v6.8b, #33 \n" // R * 0.2578 coefficient + "movi v7.8b, #16 \n" // Add 16 constant + "1: \n" + MEMACCESS(0) + "ld1 {v0.16b}, [%0], #16 \n" // load 8 ARGB1555 pixels. + "subs %w2, %w2, #8 \n" // 8 processed per loop. + ARGB1555TOARGB + "umull v3.8h, v0.8b, v4.8b \n" // B + "umlal v3.8h, v1.8b, v5.8b \n" // G + "umlal v3.8h, v2.8b, v6.8b \n" // R + "sqrshrun v0.8b, v3.8h, #7 \n" // 16 bit to 8 bit Y + "uqadd v0.8b, v0.8b, v7.8b \n" + MEMACCESS(1) + "st1 {v0.8b}, [%1], #8 \n" // store 8 pixels Y. + "b.gt 1b \n" + : "+r"(src_argb1555), // %0 + "+r"(dst_y), // %1 + "+r"(pix) // %2 + : + : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7" + ); +} +#endif // HAS_ARGB1555TOYROW_NEON + +#ifdef HAS_ARGB4444TOYROW_NEON +void ARGB4444ToYRow_NEON(const uint8* src_argb4444, uint8* dst_y, int pix) { + asm volatile ( + "movi v24.8b, #13 \n" // B * 0.1016 coefficient + "movi v25.8b, #65 \n" // G * 0.5078 coefficient + "movi v26.8b, #33 \n" // R * 0.2578 coefficient + "movi v27.8b, #16 \n" // Add 16 constant + "1: \n" + MEMACCESS(0) + "ld1 {v0.16b}, [%0], #16 \n" // load 8 ARGB4444 pixels. + "subs %w2, %w2, #8 \n" // 8 processed per loop. + ARGB4444TOARGB + "umull v3.8h, v0.8b, v24.8b \n" // B + "umlal v3.8h, v1.8b, v25.8b \n" // G + "umlal v3.8h, v2.8b, v26.8b \n" // R + "sqrshrun v0.8b, v3.8h, #7 \n" // 16 bit to 8 bit Y + "uqadd v0.8b, v0.8b, v27.8b \n" + MEMACCESS(1) + "st1 {v0.8b}, [%1], #8 \n" // store 8 pixels Y. + "b.gt 1b \n" + : "+r"(src_argb4444), // %0 + "+r"(dst_y), // %1 + "+r"(pix) // %2 + : + : "cc", "memory", "v0", "v1", "v2", "v3", "v24", "v25", "v26", "v27" + ); +} +#endif // HAS_ARGB4444TOYROW_NEON + +#ifdef HAS_BGRATOYROW_NEON +void BGRAToYRow_NEON(const uint8* src_bgra, uint8* dst_y, int pix) { + asm volatile ( + "movi v4.8b, #33 \n" // R * 0.2578 coefficient + "movi v5.8b, #65 \n" // G * 0.5078 coefficient + "movi v6.8b, #13 \n" // B * 0.1016 coefficient + "movi v7.8b, #16 \n" // Add 16 constant + "1: \n" + MEMACCESS(0) + "ld4 {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n" // load 8 pixels. + "subs %w2, %w2, #8 \n" // 8 processed per loop. + "umull v16.8h, v1.8b, v4.8b \n" // R + "umlal v16.8h, v2.8b, v5.8b \n" // G + "umlal v16.8h, v3.8b, v6.8b \n" // B + "sqrshrun v0.8b, v16.8h, #7 \n" // 16 bit to 8 bit Y + "uqadd v0.8b, v0.8b, v7.8b \n" + MEMACCESS(1) + "st1 {v0.8b}, [%1], #8 \n" // store 8 pixels Y. + "b.gt 1b \n" + : "+r"(src_bgra), // %0 + "+r"(dst_y), // %1 + "+r"(pix) // %2 + : + : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16" + ); +} +#endif // HAS_BGRATOYROW_NEON + +#ifdef HAS_ABGRTOYROW_NEON +void ABGRToYRow_NEON(const uint8* src_abgr, uint8* dst_y, int pix) { + asm volatile ( + "movi v4.8b, #33 \n" // R * 0.2578 coefficient + "movi v5.8b, #65 \n" // G * 0.5078 coefficient + "movi v6.8b, #13 \n" // B * 0.1016 coefficient + "movi v7.8b, #16 \n" // Add 16 constant + "1: \n" + MEMACCESS(0) + "ld4 {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n" // load 8 pixels. + "subs %w2, %w2, #8 \n" // 8 processed per loop. + "umull v16.8h, v0.8b, v4.8b \n" // R + "umlal v16.8h, v1.8b, v5.8b \n" // G + "umlal v16.8h, v2.8b, v6.8b \n" // B + "sqrshrun v0.8b, v16.8h, #7 \n" // 16 bit to 8 bit Y + "uqadd v0.8b, v0.8b, v7.8b \n" + MEMACCESS(1) + "st1 {v0.8b}, [%1], #8 \n" // store 8 pixels Y. + "b.gt 1b \n" + : "+r"(src_abgr), // %0 + "+r"(dst_y), // %1 + "+r"(pix) // %2 + : + : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16" + ); +} +#endif // HAS_ABGRTOYROW_NEON + +#ifdef HAS_RGBATOYROW_NEON +void RGBAToYRow_NEON(const uint8* src_rgba, uint8* dst_y, int pix) { + asm volatile ( + "movi v4.8b, #13 \n" // B * 0.1016 coefficient + "movi v5.8b, #65 \n" // G * 0.5078 coefficient + "movi v6.8b, #33 \n" // R * 0.2578 coefficient + "movi v7.8b, #16 \n" // Add 16 constant + "1: \n" + MEMACCESS(0) + "ld4 {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n" // load 8 pixels. + "subs %w2, %w2, #8 \n" // 8 processed per loop. + "umull v16.8h, v1.8b, v4.8b \n" // B + "umlal v16.8h, v2.8b, v5.8b \n" // G + "umlal v16.8h, v3.8b, v6.8b \n" // R + "sqrshrun v0.8b, v16.8h, #7 \n" // 16 bit to 8 bit Y + "uqadd v0.8b, v0.8b, v7.8b \n" + MEMACCESS(1) + "st1 {v0.8b}, [%1], #8 \n" // store 8 pixels Y. + "b.gt 1b \n" + : "+r"(src_rgba), // %0 + "+r"(dst_y), // %1 + "+r"(pix) // %2 + : + : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16" + ); +} +#endif // HAS_RGBATOYROW_NEON + +#ifdef HAS_RGB24TOYROW_NEON +void RGB24ToYRow_NEON(const uint8* src_rgb24, uint8* dst_y, int pix) { + asm volatile ( + "movi v4.8b, #13 \n" // B * 0.1016 coefficient + "movi v5.8b, #65 \n" // G * 0.5078 coefficient + "movi v6.8b, #33 \n" // R * 0.2578 coefficient + "movi v7.8b, #16 \n" // Add 16 constant + "1: \n" + MEMACCESS(0) + "ld3 {v0.8b,v1.8b,v2.8b}, [%0], #24 \n" // load 8 pixels. + "subs %w2, %w2, #8 \n" // 8 processed per loop. + "umull v16.8h, v0.8b, v4.8b \n" // B + "umlal v16.8h, v1.8b, v5.8b \n" // G + "umlal v16.8h, v2.8b, v6.8b \n" // R + "sqrshrun v0.8b, v16.8h, #7 \n" // 16 bit to 8 bit Y + "uqadd v0.8b, v0.8b, v7.8b \n" + MEMACCESS(1) + "st1 {v0.8b}, [%1], #8 \n" // store 8 pixels Y. + "b.gt 1b \n" + : "+r"(src_rgb24), // %0 + "+r"(dst_y), // %1 + "+r"(pix) // %2 + : + : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16" + ); +} +#endif // HAS_RGB24TOYROW_NEON + +#ifdef HAS_RAWTOYROW_NEON +void RAWToYRow_NEON(const uint8* src_raw, uint8* dst_y, int pix) { + asm volatile ( + "movi v4.8b, #33 \n" // R * 0.2578 coefficient + "movi v5.8b, #65 \n" // G * 0.5078 coefficient + "movi v6.8b, #13 \n" // B * 0.1016 coefficient + "movi v7.8b, #16 \n" // Add 16 constant + "1: \n" + MEMACCESS(0) + "ld3 {v0.8b,v1.8b,v2.8b}, [%0], #24 \n" // load 8 pixels. + "subs %w2, %w2, #8 \n" // 8 processed per loop. + "umull v16.8h, v0.8b, v4.8b \n" // B + "umlal v16.8h, v1.8b, v5.8b \n" // G + "umlal v16.8h, v2.8b, v6.8b \n" // R + "sqrshrun v0.8b, v16.8h, #7 \n" // 16 bit to 8 bit Y + "uqadd v0.8b, v0.8b, v7.8b \n" + MEMACCESS(1) + "st1 {v0.8b}, [%1], #8 \n" // store 8 pixels Y. + "b.gt 1b \n" + : "+r"(src_raw), // %0 + "+r"(dst_y), // %1 + "+r"(pix) // %2 + : + : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16" + ); +} +#endif // HAS_RAWTOYROW_NEON + +// Bilinear filter 16x2 -> 16x1 +#ifdef HAS_INTERPOLATEROW_NEON +void InterpolateRow_NEON(uint8* dst_ptr, + const uint8* src_ptr, ptrdiff_t src_stride, + int dst_width, int source_y_fraction) { + int y1_fraction = source_y_fraction; + int y0_fraction = 256 - y1_fraction; + const uint8* src_ptr1 = src_ptr + src_stride; + asm volatile ( + "cmp %w4, #0 \n" + "b.eq 100f \n" + "cmp %w4, #64 \n" + "b.eq 75f \n" + "cmp %w4, #128 \n" + "b.eq 50f \n" + "cmp %w4, #192 \n" + "b.eq 25f \n" + + "dup v5.16b, %w4 \n" + "dup v4.16b, %w5 \n" + // General purpose row blend. + "1: \n" + MEMACCESS(1) + "ld1 {v0.16b}, [%1], #16 \n" + MEMACCESS(2) + "ld1 {v1.16b}, [%2], #16 \n" + "subs %w3, %w3, #16 \n" + "umull v2.8h, v0.8b, v4.8b \n" + "umull2 v3.8h, v0.16b, v4.16b \n" + "umlal v2.8h, v1.8b, v5.8b \n" + "umlal2 v3.8h, v1.16b, v5.16b \n" + "rshrn v0.8b, v2.8h, #8 \n" + "rshrn2 v0.16b, v3.8h, #8 \n" + MEMACCESS(0) + "st1 {v0.16b}, [%0], #16 \n" + "b.gt 1b \n" + "b 99f \n" + + // Blend 25 / 75. + "25: \n" + MEMACCESS(1) + "ld1 {v0.16b}, [%1], #16 \n" + MEMACCESS(2) + "ld1 {v1.16b}, [%2], #16 \n" + "subs %w3, %w3, #16 \n" + "urhadd v0.16b, v0.16b, v1.16b \n" + "urhadd v0.16b, v0.16b, v1.16b \n" + MEMACCESS(0) + "st1 {v0.16b}, [%0], #16 \n" + "b.gt 25b \n" + "b 99f \n" + + // Blend 50 / 50. + "50: \n" + MEMACCESS(1) + "ld1 {v0.16b}, [%1], #16 \n" + MEMACCESS(2) + "ld1 {v1.16b}, [%2], #16 \n" + "subs %w3, %w3, #16 \n" + "urhadd v0.16b, v0.16b, v1.16b \n" + MEMACCESS(0) + "st1 {v0.16b}, [%0], #16 \n" + "b.gt 50b \n" + "b 99f \n" + + // Blend 75 / 25. + "75: \n" + MEMACCESS(1) + "ld1 {v1.16b}, [%1], #16 \n" + MEMACCESS(2) + "ld1 {v0.16b}, [%2], #16 \n" + "subs %w3, %w3, #16 \n" + "urhadd v0.16b, v0.16b, v1.16b \n" + "urhadd v0.16b, v0.16b, v1.16b \n" + MEMACCESS(0) + "st1 {v0.16b}, [%0], #16 \n" + "b.gt 75b \n" + "b 99f \n" + + // Blend 100 / 0 - Copy row unchanged. + "100: \n" + MEMACCESS(1) + "ld1 {v0.16b}, [%1], #16 \n" + "subs %w3, %w3, #16 \n" + MEMACCESS(0) + "st1 {v0.16b}, [%0], #16 \n" + "b.gt 100b \n" + + "99: \n" + : "+r"(dst_ptr), // %0 + "+r"(src_ptr), // %1 + "+r"(src_ptr1), // %2 + "+r"(dst_width), // %3 + "+r"(y1_fraction), // %4 + "+r"(y0_fraction) // %5 + : + : "cc", "memory", "v0", "v1", "v3", "v4", "v5" + ); +} +#endif // HAS_INTERPOLATEROW_NEON + +// dr * (256 - sa) / 256 + sr = dr - dr * sa / 256 + sr +#ifdef HAS_ARGBBLENDROW_NEON +void ARGBBlendRow_NEON(const uint8* src_argb0, const uint8* src_argb1, + uint8* dst_argb, int width) { + asm volatile ( + "subs %w3, %w3, #8 \n" + "b.lt 89f \n" + // Blend 8 pixels. + "8: \n" + MEMACCESS(0) + "ld4 {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n" // load 8 ARGB0 pixels + MEMACCESS(1) + "ld4 {v4.8b,v5.8b,v6.8b,v7.8b}, [%1], #32 \n" // load 8 ARGB1 pixels + "subs %w3, %w3, #8 \n" // 8 processed per loop. + "umull v16.8h, v4.8b, v3.8b \n" // db * a + "umull v17.8h, v5.8b, v3.8b \n" // dg * a + "umull v18.8h, v6.8b, v3.8b \n" // dr * a + "uqrshrn v16.8b, v16.8h, #8 \n" // db >>= 8 + "uqrshrn v17.8b, v17.8h, #8 \n" // dg >>= 8 + "uqrshrn v18.8b, v18.8h, #8 \n" // dr >>= 8 + "uqsub v4.8b, v4.8b, v16.8b \n" // db - (db * a / 256) + "uqsub v5.8b, v5.8b, v17.8b \n" // dg - (dg * a / 256) + "uqsub v6.8b, v6.8b, v18.8b \n" // dr - (dr * a / 256) + "uqadd v0.8b, v0.8b, v4.8b \n" // + sb + "uqadd v1.8b, v1.8b, v5.8b \n" // + sg + "uqadd v2.8b, v2.8b, v6.8b \n" // + sr + "movi v3.8b, #255 \n" // a = 255 + MEMACCESS(2) + "st4 {v0.8b,v1.8b,v2.8b,v3.8b}, [%2], #32 \n" // store 8 ARGB pixels + "b.ge 8b \n" + + "89: \n" + "adds %w3, %w3, #8-1 \n" + "b.lt 99f \n" + + // Blend 1 pixels. + "1: \n" + MEMACCESS(0) + "ld4 {v0.b,v1.b,v2.b,v3.b}[0], [%0], #4 \n" // load 1 pixel ARGB0. + MEMACCESS(1) + "ld4 {v4.b,v5.b,v6.b,v7.b}[0], [%1], #4 \n" // load 1 pixel ARGB1. + "subs %w3, %w3, #1 \n" // 1 processed per loop. + "umull v16.8h, v4.8b, v3.8b \n" // db * a + "umull v17.8h, v5.8b, v3.8b \n" // dg * a + "umull v18.8h, v6.8b, v3.8b \n" // dr * a + "uqrshrn v16.8b, v16.8h, #8 \n" // db >>= 8 + "uqrshrn v17.8b, v17.8h, #8 \n" // dg >>= 8 + "uqrshrn v18.8b, v18.8h, #8 \n" // dr >>= 8 + "uqsub v4.8b, v4.8b, v16.8b \n" // db - (db * a / 256) + "uqsub v5.8b, v5.8b, v17.8b \n" // dg - (dg * a / 256) + "uqsub v6.8b, v6.8b, v18.8b \n" // dr - (dr * a / 256) + "uqadd v0.8b, v0.8b, v4.8b \n" // + sb + "uqadd v1.8b, v1.8b, v5.8b \n" // + sg + "uqadd v2.8b, v2.8b, v6.8b \n" // + sr + "movi v3.8b, #255 \n" // a = 255 + MEMACCESS(2) + "st4 {v0.b,v1.b,v2.b,v3.b}[0], [%2], #4 \n" // store 1 pixel. + "b.ge 1b \n" + + "99: \n" + + : "+r"(src_argb0), // %0 + "+r"(src_argb1), // %1 + "+r"(dst_argb), // %2 + "+r"(width) // %3 + : + : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", + "v16", "v17", "v18" + ); +} +#endif // HAS_ARGBBLENDROW_NEON + +// Attenuate 8 pixels at a time. +#ifdef HAS_ARGBATTENUATEROW_NEON +void ARGBAttenuateRow_NEON(const uint8* src_argb, uint8* dst_argb, int width) { + asm volatile ( + // Attenuate 8 pixels. + "1: \n" + MEMACCESS(0) + "ld4 {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n" // load 8 ARGB pixels + "subs %w2, %w2, #8 \n" // 8 processed per loop. + "umull v4.8h, v0.8b, v3.8b \n" // b * a + "umull v5.8h, v1.8b, v3.8b \n" // g * a + "umull v6.8h, v2.8b, v3.8b \n" // r * a + "uqrshrn v0.8b, v4.8h, #8 \n" // b >>= 8 + "uqrshrn v1.8b, v5.8h, #8 \n" // g >>= 8 + "uqrshrn v2.8b, v6.8h, #8 \n" // r >>= 8 + MEMACCESS(1) + "st4 {v0.8b,v1.8b,v2.8b,v3.8b}, [%1], #32 \n" // store 8 ARGB pixels + "b.gt 1b \n" + : "+r"(src_argb), // %0 + "+r"(dst_argb), // %1 + "+r"(width) // %2 + : + : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6" + ); +} +#endif // HAS_ARGBATTENUATEROW_NEON + +// Quantize 8 ARGB pixels (32 bytes). +// dst = (dst * scale >> 16) * interval_size + interval_offset; +#ifdef HAS_ARGBQUANTIZEROW_NEON +void ARGBQuantizeRow_NEON(uint8* dst_argb, int scale, int interval_size, + int interval_offset, int width) { + asm volatile ( + "dup v4.8h, %w2 \n" + "ushr v4.8h, v4.8h, #1 \n" // scale >>= 1 + "dup v5.8h, %w3 \n" // interval multiply. + "dup v6.8h, %w4 \n" // interval add + + // 8 pixel loop. + "1: \n" + MEMACCESS(0) + "ld4 {v0.8b,v1.8b,v2.8b,v3.8b}, [%0] \n" // load 8 pixels of ARGB. + "subs %w1, %w1, #8 \n" // 8 processed per loop. + "uxtl v0.8h, v0.8b \n" // b (0 .. 255) + "uxtl v1.8h, v1.8b \n" + "uxtl v2.8h, v2.8b \n" + "sqdmulh v0.8h, v0.8h, v4.8h \n" // b * scale + "sqdmulh v1.8h, v1.8h, v4.8h \n" // g + "sqdmulh v2.8h, v2.8h, v4.8h \n" // r + "mul v0.8h, v0.8h, v5.8h \n" // b * interval_size + "mul v1.8h, v1.8h, v5.8h \n" // g + "mul v2.8h, v2.8h, v5.8h \n" // r + "add v0.8h, v0.8h, v6.8h \n" // b + interval_offset + "add v1.8h, v1.8h, v6.8h \n" // g + "add v2.8h, v2.8h, v6.8h \n" // r + "uqxtn v0.8b, v0.8h \n" + "uqxtn v1.8b, v1.8h \n" + "uqxtn v2.8b, v2.8h \n" + MEMACCESS(0) + "st4 {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n" // store 8 ARGB pixels + "b.gt 1b \n" + : "+r"(dst_argb), // %0 + "+r"(width) // %1 + : "r"(scale), // %2 + "r"(interval_size), // %3 + "r"(interval_offset) // %4 + : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6" + ); +} +#endif // HAS_ARGBQUANTIZEROW_NEON + +// Shade 8 pixels at a time by specified value. +// NOTE vqrdmulh.s16 q10, q10, d0[0] must use a scaler register from 0 to 8. +// Rounding in vqrdmulh does +1 to high if high bit of low s16 is set. +#ifdef HAS_ARGBSHADEROW_NEON +void ARGBShadeRow_NEON(const uint8* src_argb, uint8* dst_argb, int width, + uint32 value) { + asm volatile ( + "dup v0.4s, %w3 \n" // duplicate scale value. + "zip1 v0.8b, v0.8b, v0.8b \n" // v0.8b aarrggbb. + "ushr v0.8h, v0.8h, #1 \n" // scale / 2. + + // 8 pixel loop. + "1: \n" + MEMACCESS(0) + "ld4 {v4.8b,v5.8b,v6.8b,v7.8b}, [%0], #32 \n" // load 8 ARGB pixels. + "subs %w2, %w2, #8 \n" // 8 processed per loop. + "uxtl v4.8h, v4.8b \n" // b (0 .. 255) + "uxtl v5.8h, v5.8b \n" + "uxtl v6.8h, v6.8b \n" + "uxtl v7.8h, v7.8b \n" + "sqrdmulh v4.8h, v4.8h, v0.h[0] \n" // b * scale * 2 + "sqrdmulh v5.8h, v5.8h, v0.h[1] \n" // g + "sqrdmulh v6.8h, v6.8h, v0.h[2] \n" // r + "sqrdmulh v7.8h, v7.8h, v0.h[3] \n" // a + "uqxtn v4.8b, v4.8h \n" + "uqxtn v5.8b, v5.8h \n" + "uqxtn v6.8b, v6.8h \n" + "uqxtn v7.8b, v7.8h \n" + MEMACCESS(1) + "st4 {v4.8b,v5.8b,v6.8b,v7.8b}, [%1], #32 \n" // store 8 ARGB pixels + "b.gt 1b \n" + : "+r"(src_argb), // %0 + "+r"(dst_argb), // %1 + "+r"(width) // %2 + : "r"(value) // %3 + : "cc", "memory", "v0", "v4", "v5", "v6", "v7" + ); +} +#endif // HAS_ARGBSHADEROW_NEON + +// Convert 8 ARGB pixels (64 bytes) to 8 Gray ARGB pixels +// Similar to ARGBToYJ but stores ARGB. +// C code is (15 * b + 75 * g + 38 * r + 64) >> 7; +#ifdef HAS_ARGBGRAYROW_NEON +void ARGBGrayRow_NEON(const uint8* src_argb, uint8* dst_argb, int width) { + asm volatile ( + "movi v24.8b, #15 \n" // B * 0.11400 coefficient + "movi v25.8b, #75 \n" // G * 0.58700 coefficient + "movi v26.8b, #38 \n" // R * 0.29900 coefficient + "1: \n" + MEMACCESS(0) + "ld4 {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n" // load 8 ARGB pixels. + "subs %w2, %w2, #8 \n" // 8 processed per loop. + "umull v4.8h, v0.8b, v24.8b \n" // B + "umlal v4.8h, v1.8b, v25.8b \n" // G + "umlal v4.8h, v2.8b, v26.8b \n" // R + "sqrshrun v0.8b, v4.8h, #7 \n" // 15 bit to 8 bit B + "orr v1.8b, v0.8b, v0.8b \n" // G + "orr v2.8b, v0.8b, v0.8b \n" // R + MEMACCESS(1) + "st4 {v0.8b,v1.8b,v2.8b,v3.8b}, [%1], #32 \n" // store 8 pixels. + "b.gt 1b \n" + : "+r"(src_argb), // %0 + "+r"(dst_argb), // %1 + "+r"(width) // %2 + : + : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v24", "v25", "v26" + ); +} +#endif // HAS_ARGBGRAYROW_NEON + +// Convert 8 ARGB pixels (32 bytes) to 8 Sepia ARGB pixels. +// b = (r * 35 + g * 68 + b * 17) >> 7 +// g = (r * 45 + g * 88 + b * 22) >> 7 +// r = (r * 50 + g * 98 + b * 24) >> 7 + +#ifdef HAS_ARGBSEPIAROW_NEON +void ARGBSepiaRow_NEON(uint8* dst_argb, int width) { + asm volatile ( + "movi v20.8b, #17 \n" // BB coefficient + "movi v21.8b, #68 \n" // BG coefficient + "movi v22.8b, #35 \n" // BR coefficient + "movi v24.8b, #22 \n" // GB coefficient + "movi v25.8b, #88 \n" // GG coefficient + "movi v26.8b, #45 \n" // GR coefficient + "movi v28.8b, #24 \n" // BB coefficient + "movi v29.8b, #98 \n" // BG coefficient + "movi v30.8b, #50 \n" // BR coefficient + "1: \n" + MEMACCESS(0) + "ld4 {v0.8b,v1.8b,v2.8b,v3.8b}, [%0] \n" // load 8 ARGB pixels. + "subs %w1, %w1, #8 \n" // 8 processed per loop. + "umull v4.8h, v0.8b, v20.8b \n" // B to Sepia B + "umlal v4.8h, v1.8b, v21.8b \n" // G + "umlal v4.8h, v2.8b, v22.8b \n" // R + "umull v5.8h, v0.8b, v24.8b \n" // B to Sepia G + "umlal v5.8h, v1.8b, v25.8b \n" // G + "umlal v5.8h, v2.8b, v26.8b \n" // R + "umull v6.8h, v0.8b, v28.8b \n" // B to Sepia R + "umlal v6.8h, v1.8b, v29.8b \n" // G + "umlal v6.8h, v2.8b, v30.8b \n" // R + "uqshrn v0.8b, v4.8h, #7 \n" // 16 bit to 8 bit B + "uqshrn v1.8b, v5.8h, #7 \n" // 16 bit to 8 bit G + "uqshrn v2.8b, v6.8h, #7 \n" // 16 bit to 8 bit R + MEMACCESS(0) + "st4 {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n" // store 8 pixels. + "b.gt 1b \n" + : "+r"(dst_argb), // %0 + "+r"(width) // %1 + : + : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", + "v20", "v21", "v22", "v24", "v25", "v26", "v28", "v29", "v30" + ); +} +#endif // HAS_ARGBSEPIAROW_NEON + +// Tranform 8 ARGB pixels (32 bytes) with color matrix. +// TODO(fbarchard): Was same as Sepia except matrix is provided. This function +// needs to saturate. Consider doing a non-saturating version. +#ifdef HAS_ARGBCOLORMATRIXROW_NEON +void ARGBColorMatrixRow_NEON(const uint8* src_argb, uint8* dst_argb, + const int8* matrix_argb, int width) { + asm volatile ( + MEMACCESS(3) + "ld1 {v2.16b}, [%3] \n" // load 3 ARGB vectors. + "sxtl v0.8h, v2.8b \n" // B,G coefficients s16. + "sxtl2 v1.8h, v2.16b \n" // R,A coefficients s16. + + "1: \n" + MEMACCESS(0) + "ld4 {v16.8b,v17.8b,v18.8b,v19.8b}, [%0], #32 \n" // load 8 pixels. + "subs %w2, %w2, #8 \n" // 8 processed per loop. + "uxtl v16.8h, v16.8b \n" // b (0 .. 255) 16 bit + "uxtl v17.8h, v17.8b \n" // g + "uxtl v18.8h, v18.8b \n" // r + "uxtl v19.8h, v19.8b \n" // a + "mul v22.8h, v16.8h, v0.h[0] \n" // B = B * Matrix B + "mul v23.8h, v16.8h, v0.h[4] \n" // G = B * Matrix G + "mul v24.8h, v16.8h, v1.h[0] \n" // R = B * Matrix R + "mul v25.8h, v16.8h, v1.h[4] \n" // A = B * Matrix A + "mul v4.8h, v17.8h, v0.h[1] \n" // B += G * Matrix B + "mul v5.8h, v17.8h, v0.h[5] \n" // G += G * Matrix G + "mul v6.8h, v17.8h, v1.h[1] \n" // R += G * Matrix R + "mul v7.8h, v17.8h, v1.h[5] \n" // A += G * Matrix A + "sqadd v22.8h, v22.8h, v4.8h \n" // Accumulate B + "sqadd v23.8h, v23.8h, v5.8h \n" // Accumulate G + "sqadd v24.8h, v24.8h, v6.8h \n" // Accumulate R + "sqadd v25.8h, v25.8h, v7.8h \n" // Accumulate A + "mul v4.8h, v18.8h, v0.h[2] \n" // B += R * Matrix B + "mul v5.8h, v18.8h, v0.h[6] \n" // G += R * Matrix G + "mul v6.8h, v18.8h, v1.h[2] \n" // R += R * Matrix R + "mul v7.8h, v18.8h, v1.h[6] \n" // A += R * Matrix A + "sqadd v22.8h, v22.8h, v4.8h \n" // Accumulate B + "sqadd v23.8h, v23.8h, v5.8h \n" // Accumulate G + "sqadd v24.8h, v24.8h, v6.8h \n" // Accumulate R + "sqadd v25.8h, v25.8h, v7.8h \n" // Accumulate A + "mul v4.8h, v19.8h, v0.h[3] \n" // B += A * Matrix B + "mul v5.8h, v19.8h, v0.h[7] \n" // G += A * Matrix G + "mul v6.8h, v19.8h, v1.h[3] \n" // R += A * Matrix R + "mul v7.8h, v19.8h, v1.h[7] \n" // A += A * Matrix A + "sqadd v22.8h, v22.8h, v4.8h \n" // Accumulate B + "sqadd v23.8h, v23.8h, v5.8h \n" // Accumulate G + "sqadd v24.8h, v24.8h, v6.8h \n" // Accumulate R + "sqadd v25.8h, v25.8h, v7.8h \n" // Accumulate A + "sqshrun v16.8b, v22.8h, #6 \n" // 16 bit to 8 bit B + "sqshrun v17.8b, v23.8h, #6 \n" // 16 bit to 8 bit G + "sqshrun v18.8b, v24.8h, #6 \n" // 16 bit to 8 bit R + "sqshrun v19.8b, v25.8h, #6 \n" // 16 bit to 8 bit A + MEMACCESS(1) + "st4 {v16.8b,v17.8b,v18.8b,v19.8b}, [%1], #32 \n" // store 8 pixels. + "b.gt 1b \n" + : "+r"(src_argb), // %0 + "+r"(dst_argb), // %1 + "+r"(width) // %2 + : "r"(matrix_argb) // %3 + : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16", "v17", + "v18", "v19", "v22", "v23", "v24", "v25" + ); +} +#endif // HAS_ARGBCOLORMATRIXROW_NEON + +// TODO(fbarchard): fix vqshrun in ARGBMultiplyRow_NEON and reenable. +// Multiply 2 rows of ARGB pixels together, 8 pixels at a time. +#ifdef HAS_ARGBMULTIPLYROW_NEON +void ARGBMultiplyRow_NEON(const uint8* src_argb0, const uint8* src_argb1, + uint8* dst_argb, int width) { + asm volatile ( + // 8 pixel loop. + "1: \n" + MEMACCESS(0) + "ld4 {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n" // load 8 ARGB pixels. + MEMACCESS(1) + "ld4 {v4.8b,v5.8b,v6.8b,v7.8b}, [%1], #32 \n" // load 8 more pixels. + "subs %w3, %w3, #8 \n" // 8 processed per loop. + "umull v0.8h, v0.8b, v4.8b \n" // multiply B + "umull v1.8h, v1.8b, v5.8b \n" // multiply G + "umull v2.8h, v2.8b, v6.8b \n" // multiply R + "umull v3.8h, v3.8b, v7.8b \n" // multiply A + "rshrn v0.8b, v0.8h, #8 \n" // 16 bit to 8 bit B + "rshrn v1.8b, v1.8h, #8 \n" // 16 bit to 8 bit G + "rshrn v2.8b, v2.8h, #8 \n" // 16 bit to 8 bit R + "rshrn v3.8b, v3.8h, #8 \n" // 16 bit to 8 bit A + MEMACCESS(2) + "st4 {v0.8b,v1.8b,v2.8b,v3.8b}, [%2], #32 \n" // store 8 ARGB pixels + "b.gt 1b \n" + + : "+r"(src_argb0), // %0 + "+r"(src_argb1), // %1 + "+r"(dst_argb), // %2 + "+r"(width) // %3 + : + : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7" + ); +} +#endif // HAS_ARGBMULTIPLYROW_NEON + +// Add 2 rows of ARGB pixels together, 8 pixels at a time. +#ifdef HAS_ARGBADDROW_NEON +void ARGBAddRow_NEON(const uint8* src_argb0, const uint8* src_argb1, + uint8* dst_argb, int width) { + asm volatile ( + // 8 pixel loop. + "1: \n" + MEMACCESS(0) + "ld4 {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n" // load 8 ARGB pixels. + MEMACCESS(1) + "ld4 {v4.8b,v5.8b,v6.8b,v7.8b}, [%1], #32 \n" // load 8 more pixels. + "subs %w3, %w3, #8 \n" // 8 processed per loop. + "uqadd v0.8b, v0.8b, v4.8b \n" + "uqadd v1.8b, v1.8b, v5.8b \n" + "uqadd v2.8b, v2.8b, v6.8b \n" + "uqadd v3.8b, v3.8b, v7.8b \n" + MEMACCESS(2) + "st4 {v0.8b,v1.8b,v2.8b,v3.8b}, [%2], #32 \n" // store 8 ARGB pixels + "b.gt 1b \n" + + : "+r"(src_argb0), // %0 + "+r"(src_argb1), // %1 + "+r"(dst_argb), // %2 + "+r"(width) // %3 + : + : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7" + ); +} +#endif // HAS_ARGBADDROW_NEON + +// Subtract 2 rows of ARGB pixels, 8 pixels at a time. +#ifdef HAS_ARGBSUBTRACTROW_NEON +void ARGBSubtractRow_NEON(const uint8* src_argb0, const uint8* src_argb1, + uint8* dst_argb, int width) { + asm volatile ( + // 8 pixel loop. + "1: \n" + MEMACCESS(0) + "ld4 {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n" // load 8 ARGB pixels. + MEMACCESS(1) + "ld4 {v4.8b,v5.8b,v6.8b,v7.8b}, [%1], #32 \n" // load 8 more pixels. + "subs %w3, %w3, #8 \n" // 8 processed per loop. + "uqsub v0.8b, v0.8b, v4.8b \n" + "uqsub v1.8b, v1.8b, v5.8b \n" + "uqsub v2.8b, v2.8b, v6.8b \n" + "uqsub v3.8b, v3.8b, v7.8b \n" + MEMACCESS(2) + "st4 {v0.8b,v1.8b,v2.8b,v3.8b}, [%2], #32 \n" // store 8 ARGB pixels + "b.gt 1b \n" + + : "+r"(src_argb0), // %0 + "+r"(src_argb1), // %1 + "+r"(dst_argb), // %2 + "+r"(width) // %3 + : + : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7" + ); +} +#endif // HAS_ARGBSUBTRACTROW_NEON + +// Adds Sobel X and Sobel Y and stores Sobel into ARGB. +// A = 255 +// R = Sobel +// G = Sobel +// B = Sobel +#ifdef HAS_SOBELROW_NEON +void SobelRow_NEON(const uint8* src_sobelx, const uint8* src_sobely, + uint8* dst_argb, int width) { + asm volatile ( + "movi v3.8b, #255 \n" // alpha + // 8 pixel loop. + "1: \n" + MEMACCESS(0) + "ld1 {v0.8b}, [%0], #8 \n" // load 8 sobelx. + MEMACCESS(1) + "ld1 {v1.8b}, [%1], #8 \n" // load 8 sobely. + "subs %w3, %w3, #8 \n" // 8 processed per loop. + "uqadd v0.8b, v0.8b, v1.8b \n" // add + "orr v1.8b, v0.8b, v0.8b \n" + "orr v2.8b, v0.8b, v0.8b \n" + MEMACCESS(2) + "st4 {v0.8b,v1.8b,v2.8b,v3.8b}, [%2], #32 \n" // store 8 ARGB pixels + "b.gt 1b \n" + : "+r"(src_sobelx), // %0 + "+r"(src_sobely), // %1 + "+r"(dst_argb), // %2 + "+r"(width) // %3 + : + : "cc", "memory", "v0", "v1", "v2", "v3" + ); +} +#endif // HAS_SOBELROW_NEON + +// Adds Sobel X and Sobel Y and stores Sobel into plane. +#ifdef HAS_SOBELTOPLANEROW_NEON +void SobelToPlaneRow_NEON(const uint8* src_sobelx, const uint8* src_sobely, + uint8* dst_y, int width) { + asm volatile ( + // 16 pixel loop. + "1: \n" + MEMACCESS(0) + "ld1 {v0.16b}, [%0], #16 \n" // load 16 sobelx. + MEMACCESS(1) + "ld1 {v1.16b}, [%1], #16 \n" // load 16 sobely. + "subs %w3, %w3, #16 \n" // 16 processed per loop. + "uqadd v0.16b, v0.16b, v1.16b \n" // add + MEMACCESS(2) + "st1 {v0.16b}, [%2], #16 \n" // store 16 pixels. + "b.gt 1b \n" + : "+r"(src_sobelx), // %0 + "+r"(src_sobely), // %1 + "+r"(dst_y), // %2 + "+r"(width) // %3 + : + : "cc", "memory", "v0", "v1" + ); +} +#endif // HAS_SOBELTOPLANEROW_NEON + +// Mixes Sobel X, Sobel Y and Sobel into ARGB. +// A = 255 +// R = Sobel X +// G = Sobel +// B = Sobel Y +#ifdef HAS_SOBELXYROW_NEON +void SobelXYRow_NEON(const uint8* src_sobelx, const uint8* src_sobely, + uint8* dst_argb, int width) { + asm volatile ( + "movi v3.8b, #255 \n" // alpha + // 8 pixel loop. + "1: \n" + MEMACCESS(0) + "ld1 {v2.8b}, [%0], #8 \n" // load 8 sobelx. + MEMACCESS(1) + "ld1 {v0.8b}, [%1], #8 \n" // load 8 sobely. + "subs %w3, %w3, #8 \n" // 8 processed per loop. + "uqadd v1.8b, v0.8b, v2.8b \n" // add + MEMACCESS(2) + "st4 {v0.8b,v1.8b,v2.8b,v3.8b}, [%2], #32 \n" // store 8 ARGB pixels + "b.gt 1b \n" + : "+r"(src_sobelx), // %0 + "+r"(src_sobely), // %1 + "+r"(dst_argb), // %2 + "+r"(width) // %3 + : + : "cc", "memory", "v0", "v1", "v2", "v3" + ); +} +#endif // HAS_SOBELXYROW_NEON + +// SobelX as a matrix is +// -1 0 1 +// -2 0 2 +// -1 0 1 +#ifdef HAS_SOBELXROW_NEON +void SobelXRow_NEON(const uint8* src_y0, const uint8* src_y1, + const uint8* src_y2, uint8* dst_sobelx, int width) { + asm volatile ( + "1: \n" + MEMACCESS(0) + "ld1 {v0.8b}, [%0],%5 \n" // top + MEMACCESS(0) + "ld1 {v1.8b}, [%0],%6 \n" + "usubl v0.8h, v0.8b, v1.8b \n" + MEMACCESS(1) + "ld1 {v2.8b}, [%1],%5 \n" // center * 2 + MEMACCESS(1) + "ld1 {v3.8b}, [%1],%6 \n" + "usubl v1.8h, v2.8b, v3.8b \n" + "add v0.8h, v0.8h, v1.8h \n" + "add v0.8h, v0.8h, v1.8h \n" + MEMACCESS(2) + "ld1 {v2.8b}, [%2],%5 \n" // bottom + MEMACCESS(2) + "ld1 {v3.8b}, [%2],%6 \n" + "subs %w4, %w4, #8 \n" // 8 pixels + "usubl v1.8h, v2.8b, v3.8b \n" + "add v0.8h, v0.8h, v1.8h \n" + "abs v0.8h, v0.8h \n" + "uqxtn v0.8b, v0.8h \n" + MEMACCESS(3) + "st1 {v0.8b}, [%3], #8 \n" // store 8 sobelx + "b.gt 1b \n" + : "+r"(src_y0), // %0 + "+r"(src_y1), // %1 + "+r"(src_y2), // %2 + "+r"(dst_sobelx), // %3 + "+r"(width) // %4 + : "r"(2LL), // %5 + "r"(6LL) // %6 + : "cc", "memory", "v0", "v1", "v2", "v3" // Clobber List + ); +} +#endif // HAS_SOBELXROW_NEON + +// SobelY as a matrix is +// -1 -2 -1 +// 0 0 0 +// 1 2 1 +#ifdef HAS_SOBELYROW_NEON +void SobelYRow_NEON(const uint8* src_y0, const uint8* src_y1, + uint8* dst_sobely, int width) { + asm volatile ( + "1: \n" + MEMACCESS(0) + "ld1 {v0.8b}, [%0],%4 \n" // left + MEMACCESS(1) + "ld1 {v1.8b}, [%1],%4 \n" + "usubl v0.8h, v0.8b, v1.8b \n" + MEMACCESS(0) + "ld1 {v2.8b}, [%0],%4 \n" // center * 2 + MEMACCESS(1) + "ld1 {v3.8b}, [%1],%4 \n" + "usubl v1.8h, v2.8b, v3.8b \n" + "add v0.8h, v0.8h, v1.8h \n" + "add v0.8h, v0.8h, v1.8h \n" + MEMACCESS(0) + "ld1 {v2.8b}, [%0],%5 \n" // right + MEMACCESS(1) + "ld1 {v3.8b}, [%1],%5 \n" + "subs %w3, %w3, #8 \n" // 8 pixels + "usubl v1.8h, v2.8b, v3.8b \n" + "add v0.8h, v0.8h, v1.8h \n" + "abs v0.8h, v0.8h \n" + "uqxtn v0.8b, v0.8h \n" + MEMACCESS(2) + "st1 {v0.8b}, [%2], #8 \n" // store 8 sobely + "b.gt 1b \n" + : "+r"(src_y0), // %0 + "+r"(src_y1), // %1 + "+r"(dst_sobely), // %2 + "+r"(width) // %3 + : "r"(1LL), // %4 + "r"(6LL) // %5 + : "cc", "memory", "v0", "v1", "v2", "v3" // Clobber List + ); +} +#endif // HAS_SOBELYROW_NEON +#endif // !defined(LIBYUV_DISABLE_NEON) && defined(__aarch64__) + +#ifdef __cplusplus +} // extern "C" +} // namespace libyuv +#endif diff --git a/media/libaom/src/third_party/libyuv/source/row_win.cc b/media/libaom/src/third_party/libyuv/source/row_win.cc new file mode 100644 index 000000000..71be268b4 --- /dev/null +++ b/media/libaom/src/third_party/libyuv/source/row_win.cc @@ -0,0 +1,6331 @@ +/* + * Copyright 2011 The LibYuv Project Authors. All rights reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#include "libyuv/row.h" + +#if !defined(LIBYUV_DISABLE_X86) && defined(_M_X64) && \ + defined(_MSC_VER) && !defined(__clang__) +#include <emmintrin.h> +#include <tmmintrin.h> // For _mm_maddubs_epi16 +#endif + +#ifdef __cplusplus +namespace libyuv { +extern "C" { +#endif + +// This module is for Visual C. +#if !defined(LIBYUV_DISABLE_X86) && (defined(_M_IX86) || defined(_M_X64)) && \ + defined(_MSC_VER) && !defined(__clang__) + +struct YuvConstants { + lvec8 kUVToB; // 0 + lvec8 kUVToG; // 32 + lvec8 kUVToR; // 64 + lvec16 kUVBiasB; // 96 + lvec16 kUVBiasG; // 128 + lvec16 kUVBiasR; // 160 + lvec16 kYToRgb; // 192 +}; + +// BT.601 YUV to RGB reference +// R = (Y - 16) * 1.164 - V * -1.596 +// G = (Y - 16) * 1.164 - U * 0.391 - V * 0.813 +// B = (Y - 16) * 1.164 - U * -2.018 + +// Y contribution to R,G,B. Scale and bias. +// TODO(fbarchard): Consider moving constants into a common header. +#define YG 18997 /* round(1.164 * 64 * 256 * 256 / 257) */ +#define YGB -1160 /* 1.164 * 64 * -16 + 64 / 2 */ + +// U and V contributions to R,G,B. +#define UB -128 /* max(-128, round(-2.018 * 64)) */ +#define UG 25 /* round(0.391 * 64) */ +#define VG 52 /* round(0.813 * 64) */ +#define VR -102 /* round(-1.596 * 64) */ + +// Bias values to subtract 16 from Y and 128 from U and V. +#define BB (UB * 128 + YGB) +#define BG (UG * 128 + VG * 128 + YGB) +#define BR (VR * 128 + YGB) + +// BT601 constants for YUV to RGB. +static YuvConstants SIMD_ALIGNED(kYuvConstants) = { + { UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, + UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0 }, + { UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, + UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG }, + { 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, + 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR }, + { BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB }, + { BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG }, + { BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR }, + { YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG } +}; + +// BT601 constants for NV21 where chroma plane is VU instead of UV. +static YuvConstants SIMD_ALIGNED(kYvuConstants) = { + { 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, + 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB }, + { VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, + VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG }, + { VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, + VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0 }, + { BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB }, + { BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG }, + { BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR }, + { YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG } +}; + +#undef YG +#undef YGB +#undef UB +#undef UG +#undef VG +#undef VR +#undef BB +#undef BG +#undef BR + +// JPEG YUV to RGB reference +// * R = Y - V * -1.40200 +// * G = Y - U * 0.34414 - V * 0.71414 +// * B = Y - U * -1.77200 + +// Y contribution to R,G,B. Scale and bias. +// TODO(fbarchard): Consider moving constants into a common header. +#define YGJ 16320 /* round(1.000 * 64 * 256 * 256 / 257) */ +#define YGBJ 32 /* 64 / 2 */ + +// U and V contributions to R,G,B. +#define UBJ -113 /* round(-1.77200 * 64) */ +#define UGJ 22 /* round(0.34414 * 64) */ +#define VGJ 46 /* round(0.71414 * 64) */ +#define VRJ -90 /* round(-1.40200 * 64) */ + +// Bias values to subtract 16 from Y and 128 from U and V. +#define BBJ (UBJ * 128 + YGBJ) +#define BGJ (UGJ * 128 + VGJ * 128 + YGBJ) +#define BRJ (VRJ * 128 + YGBJ) + +// JPEG constants for YUV to RGB. +static YuvConstants SIMD_ALIGNED(kYuvJConstants) = { + { UBJ, 0, UBJ, 0, UBJ, 0, UBJ, 0, UBJ, 0, UBJ, 0, UBJ, 0, UBJ, 0, + UBJ, 0, UBJ, 0, UBJ, 0, UBJ, 0, UBJ, 0, UBJ, 0, UBJ, 0, UBJ, 0 }, + { UGJ, VGJ, UGJ, VGJ, UGJ, VGJ, UGJ, VGJ, + UGJ, VGJ, UGJ, VGJ, UGJ, VGJ, UGJ, VGJ, + UGJ, VGJ, UGJ, VGJ, UGJ, VGJ, UGJ, VGJ, + UGJ, VGJ, UGJ, VGJ, UGJ, VGJ, UGJ, VGJ }, + { 0, VRJ, 0, VRJ, 0, VRJ, 0, VRJ, 0, VRJ, 0, VRJ, 0, VRJ, 0, VRJ, + 0, VRJ, 0, VRJ, 0, VRJ, 0, VRJ, 0, VRJ, 0, VRJ, 0, VRJ, 0, VRJ }, + { BBJ, BBJ, BBJ, BBJ, BBJ, BBJ, BBJ, BBJ, + BBJ, BBJ, BBJ, BBJ, BBJ, BBJ, BBJ, BBJ }, + { BGJ, BGJ, BGJ, BGJ, BGJ, BGJ, BGJ, BGJ, + BGJ, BGJ, BGJ, BGJ, BGJ, BGJ, BGJ, BGJ }, + { BRJ, BRJ, BRJ, BRJ, BRJ, BRJ, BRJ, BRJ, + BRJ, BRJ, BRJ, BRJ, BRJ, BRJ, BRJ, BRJ }, + { YGJ, YGJ, YGJ, YGJ, YGJ, YGJ, YGJ, YGJ, + YGJ, YGJ, YGJ, YGJ, YGJ, YGJ, YGJ, YGJ } +}; + +#undef YGJ +#undef YGBJ +#undef UBJ +#undef UGJ +#undef VGJ +#undef VRJ +#undef BBJ +#undef BGJ +#undef BRJ + +// 64 bit +#if defined(_M_X64) +#if defined(HAS_I422TOARGBROW_SSSE3) +void I422ToARGBRow_SSSE3(const uint8* y_buf, + const uint8* u_buf, + const uint8* v_buf, + uint8* dst_argb, + int width) { + __m128i xmm0, xmm1, xmm2, xmm3; + const __m128i xmm5 = _mm_set1_epi8(-1); + const ptrdiff_t offset = (uint8*)v_buf - (uint8*)u_buf; + + while (width > 0) { + xmm0 = _mm_cvtsi32_si128(*(uint32*)u_buf); + xmm1 = _mm_cvtsi32_si128(*(uint32*)(u_buf + offset)); + xmm0 = _mm_unpacklo_epi8(xmm0, xmm1); + xmm0 = _mm_unpacklo_epi16(xmm0, xmm0); + xmm1 = _mm_loadu_si128(&xmm0); + xmm2 = _mm_loadu_si128(&xmm0); + xmm0 = _mm_maddubs_epi16(xmm0, *(__m128i*)kYuvConstants.kUVToB); + xmm1 = _mm_maddubs_epi16(xmm1, *(__m128i*)kYuvConstants.kUVToG); + xmm2 = _mm_maddubs_epi16(xmm2, *(__m128i*)kYuvConstants.kUVToR); + xmm0 = _mm_sub_epi16(*(__m128i*)kYuvConstants.kUVBiasB, xmm0); + xmm1 = _mm_sub_epi16(*(__m128i*)kYuvConstants.kUVBiasG, xmm1); + xmm2 = _mm_sub_epi16(*(__m128i*)kYuvConstants.kUVBiasR, xmm2); + xmm3 = _mm_loadl_epi64((__m128i*)y_buf); + xmm3 = _mm_unpacklo_epi8(xmm3, xmm3); + xmm3 = _mm_mulhi_epu16(xmm3, *(__m128i*)kYuvConstants.kYToRgb); + xmm0 = _mm_adds_epi16(xmm0, xmm3); + xmm1 = _mm_adds_epi16(xmm1, xmm3); + xmm2 = _mm_adds_epi16(xmm2, xmm3); + xmm0 = _mm_srai_epi16(xmm0, 6); + xmm1 = _mm_srai_epi16(xmm1, 6); + xmm2 = _mm_srai_epi16(xmm2, 6); + xmm0 = _mm_packus_epi16(xmm0, xmm0); + xmm1 = _mm_packus_epi16(xmm1, xmm1); + xmm2 = _mm_packus_epi16(xmm2, xmm2); + xmm0 = _mm_unpacklo_epi8(xmm0, xmm1); + xmm2 = _mm_unpacklo_epi8(xmm2, xmm5); + xmm1 = _mm_loadu_si128(&xmm0); + xmm0 = _mm_unpacklo_epi16(xmm0, xmm2); + xmm1 = _mm_unpackhi_epi16(xmm1, xmm2); + + _mm_storeu_si128((__m128i *)dst_argb, xmm0); + _mm_storeu_si128((__m128i *)(dst_argb + 16), xmm1); + + y_buf += 8; + u_buf += 4; + dst_argb += 32; + width -= 8; + } +} +#endif +// 32 bit +#else // defined(_M_X64) +#ifdef HAS_ARGBTOYROW_SSSE3 + +// Constants for ARGB. +static const vec8 kARGBToY = { + 13, 65, 33, 0, 13, 65, 33, 0, 13, 65, 33, 0, 13, 65, 33, 0 +}; + +// JPeg full range. +static const vec8 kARGBToYJ = { + 15, 75, 38, 0, 15, 75, 38, 0, 15, 75, 38, 0, 15, 75, 38, 0 +}; + +static const vec8 kARGBToU = { + 112, -74, -38, 0, 112, -74, -38, 0, 112, -74, -38, 0, 112, -74, -38, 0 +}; + +static const vec8 kARGBToUJ = { + 127, -84, -43, 0, 127, -84, -43, 0, 127, -84, -43, 0, 127, -84, -43, 0 +}; + +static const vec8 kARGBToV = { + -18, -94, 112, 0, -18, -94, 112, 0, -18, -94, 112, 0, -18, -94, 112, 0, +}; + +static const vec8 kARGBToVJ = { + -20, -107, 127, 0, -20, -107, 127, 0, -20, -107, 127, 0, -20, -107, 127, 0 +}; + +// vpshufb for vphaddw + vpackuswb packed to shorts. +static const lvec8 kShufARGBToUV_AVX = { + 0, 1, 8, 9, 2, 3, 10, 11, 4, 5, 12, 13, 6, 7, 14, 15, + 0, 1, 8, 9, 2, 3, 10, 11, 4, 5, 12, 13, 6, 7, 14, 15 +}; + +// Constants for BGRA. +static const vec8 kBGRAToY = { + 0, 33, 65, 13, 0, 33, 65, 13, 0, 33, 65, 13, 0, 33, 65, 13 +}; + +static const vec8 kBGRAToU = { + 0, -38, -74, 112, 0, -38, -74, 112, 0, -38, -74, 112, 0, -38, -74, 112 +}; + +static const vec8 kBGRAToV = { + 0, 112, -94, -18, 0, 112, -94, -18, 0, 112, -94, -18, 0, 112, -94, -18 +}; + +// Constants for ABGR. +static const vec8 kABGRToY = { + 33, 65, 13, 0, 33, 65, 13, 0, 33, 65, 13, 0, 33, 65, 13, 0 +}; + +static const vec8 kABGRToU = { + -38, -74, 112, 0, -38, -74, 112, 0, -38, -74, 112, 0, -38, -74, 112, 0 +}; + +static const vec8 kABGRToV = { + 112, -94, -18, 0, 112, -94, -18, 0, 112, -94, -18, 0, 112, -94, -18, 0 +}; + +// Constants for RGBA. +static const vec8 kRGBAToY = { + 0, 13, 65, 33, 0, 13, 65, 33, 0, 13, 65, 33, 0, 13, 65, 33 +}; + +static const vec8 kRGBAToU = { + 0, 112, -74, -38, 0, 112, -74, -38, 0, 112, -74, -38, 0, 112, -74, -38 +}; + +static const vec8 kRGBAToV = { + 0, -18, -94, 112, 0, -18, -94, 112, 0, -18, -94, 112, 0, -18, -94, 112 +}; + +static const uvec8 kAddY16 = { + 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u +}; + +// 7 bit fixed point 0.5. +static const vec16 kAddYJ64 = { + 64, 64, 64, 64, 64, 64, 64, 64 +}; + +static const uvec8 kAddUV128 = { + 128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u, + 128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u +}; + +static const uvec16 kAddUVJ128 = { + 0x8080u, 0x8080u, 0x8080u, 0x8080u, 0x8080u, 0x8080u, 0x8080u, 0x8080u +}; + +// Shuffle table for converting RGB24 to ARGB. +static const uvec8 kShuffleMaskRGB24ToARGB = { + 0u, 1u, 2u, 12u, 3u, 4u, 5u, 13u, 6u, 7u, 8u, 14u, 9u, 10u, 11u, 15u +}; + +// Shuffle table for converting RAW to ARGB. +static const uvec8 kShuffleMaskRAWToARGB = { + 2u, 1u, 0u, 12u, 5u, 4u, 3u, 13u, 8u, 7u, 6u, 14u, 11u, 10u, 9u, 15u +}; + +// Shuffle table for converting ARGB to RGB24. +static const uvec8 kShuffleMaskARGBToRGB24 = { + 0u, 1u, 2u, 4u, 5u, 6u, 8u, 9u, 10u, 12u, 13u, 14u, 128u, 128u, 128u, 128u +}; + +// Shuffle table for converting ARGB to RAW. +static const uvec8 kShuffleMaskARGBToRAW = { + 2u, 1u, 0u, 6u, 5u, 4u, 10u, 9u, 8u, 14u, 13u, 12u, 128u, 128u, 128u, 128u +}; + +// Shuffle table for converting ARGBToRGB24 for I422ToRGB24. First 8 + next 4 +static const uvec8 kShuffleMaskARGBToRGB24_0 = { + 0u, 1u, 2u, 4u, 5u, 6u, 8u, 9u, 128u, 128u, 128u, 128u, 10u, 12u, 13u, 14u +}; + +// Shuffle table for converting ARGB to RAW. +static const uvec8 kShuffleMaskARGBToRAW_0 = { + 2u, 1u, 0u, 6u, 5u, 4u, 10u, 9u, 128u, 128u, 128u, 128u, 8u, 14u, 13u, 12u +}; + +// Duplicates gray value 3 times and fills in alpha opaque. +__declspec(naked) +void J400ToARGBRow_SSE2(const uint8* src_y, uint8* dst_argb, int pix) { + __asm { + mov eax, [esp + 4] // src_y + mov edx, [esp + 8] // dst_argb + mov ecx, [esp + 12] // pix + pcmpeqb xmm5, xmm5 // generate mask 0xff000000 + pslld xmm5, 24 + + convertloop: + movq xmm0, qword ptr [eax] + lea eax, [eax + 8] + punpcklbw xmm0, xmm0 + movdqa xmm1, xmm0 + punpcklwd xmm0, xmm0 + punpckhwd xmm1, xmm1 + por xmm0, xmm5 + por xmm1, xmm5 + movdqu [edx], xmm0 + movdqu [edx + 16], xmm1 + lea edx, [edx + 32] + sub ecx, 8 + jg convertloop + ret + } +} + +#ifdef HAS_J400TOARGBROW_AVX2 +// Duplicates gray value 3 times and fills in alpha opaque. +__declspec(naked) +void J400ToARGBRow_AVX2(const uint8* src_y, uint8* dst_argb, int pix) { + __asm { + mov eax, [esp + 4] // src_y + mov edx, [esp + 8] // dst_argb + mov ecx, [esp + 12] // pix + vpcmpeqb ymm5, ymm5, ymm5 // generate mask 0xff000000 + vpslld ymm5, ymm5, 24 + + convertloop: + vmovdqu xmm0, [eax] + lea eax, [eax + 16] + vpermq ymm0, ymm0, 0xd8 + vpunpcklbw ymm0, ymm0, ymm0 + vpermq ymm0, ymm0, 0xd8 + vpunpckhwd ymm1, ymm0, ymm0 + vpunpcklwd ymm0, ymm0, ymm0 + vpor ymm0, ymm0, ymm5 + vpor ymm1, ymm1, ymm5 + vmovdqu [edx], ymm0 + vmovdqu [edx + 32], ymm1 + lea edx, [edx + 64] + sub ecx, 16 + jg convertloop + vzeroupper + ret + } +} +#endif // HAS_J400TOARGBROW_AVX2 + +__declspec(naked) +void RGB24ToARGBRow_SSSE3(const uint8* src_rgb24, uint8* dst_argb, int pix) { + __asm { + mov eax, [esp + 4] // src_rgb24 + mov edx, [esp + 8] // dst_argb + mov ecx, [esp + 12] // pix + pcmpeqb xmm5, xmm5 // generate mask 0xff000000 + pslld xmm5, 24 + movdqa xmm4, kShuffleMaskRGB24ToARGB + + convertloop: + movdqu xmm0, [eax] + movdqu xmm1, [eax + 16] + movdqu xmm3, [eax + 32] + lea eax, [eax + 48] + movdqa xmm2, xmm3 + palignr xmm2, xmm1, 8 // xmm2 = { xmm3[0:3] xmm1[8:15]} + pshufb xmm2, xmm4 + por xmm2, xmm5 + palignr xmm1, xmm0, 12 // xmm1 = { xmm3[0:7] xmm0[12:15]} + pshufb xmm0, xmm4 + movdqu [edx + 32], xmm2 + por xmm0, xmm5 + pshufb xmm1, xmm4 + movdqu [edx], xmm0 + por xmm1, xmm5 + palignr xmm3, xmm3, 4 // xmm3 = { xmm3[4:15]} + pshufb xmm3, xmm4 + movdqu [edx + 16], xmm1 + por xmm3, xmm5 + movdqu [edx + 48], xmm3 + lea edx, [edx + 64] + sub ecx, 16 + jg convertloop + ret + } +} + +__declspec(naked) +void RAWToARGBRow_SSSE3(const uint8* src_raw, uint8* dst_argb, + int pix) { + __asm { + mov eax, [esp + 4] // src_raw + mov edx, [esp + 8] // dst_argb + mov ecx, [esp + 12] // pix + pcmpeqb xmm5, xmm5 // generate mask 0xff000000 + pslld xmm5, 24 + movdqa xmm4, kShuffleMaskRAWToARGB + + convertloop: + movdqu xmm0, [eax] + movdqu xmm1, [eax + 16] + movdqu xmm3, [eax + 32] + lea eax, [eax + 48] + movdqa xmm2, xmm3 + palignr xmm2, xmm1, 8 // xmm2 = { xmm3[0:3] xmm1[8:15]} + pshufb xmm2, xmm4 + por xmm2, xmm5 + palignr xmm1, xmm0, 12 // xmm1 = { xmm3[0:7] xmm0[12:15]} + pshufb xmm0, xmm4 + movdqu [edx + 32], xmm2 + por xmm0, xmm5 + pshufb xmm1, xmm4 + movdqu [edx], xmm0 + por xmm1, xmm5 + palignr xmm3, xmm3, 4 // xmm3 = { xmm3[4:15]} + pshufb xmm3, xmm4 + movdqu [edx + 16], xmm1 + por xmm3, xmm5 + movdqu [edx + 48], xmm3 + lea edx, [edx + 64] + sub ecx, 16 + jg convertloop + ret + } +} + +// pmul method to replicate bits. +// Math to replicate bits: +// (v << 8) | (v << 3) +// v * 256 + v * 8 +// v * (256 + 8) +// G shift of 5 is incorporated, so shift is 5 + 8 and 5 + 3 +// 20 instructions. +__declspec(naked) +void RGB565ToARGBRow_SSE2(const uint8* src_rgb565, uint8* dst_argb, + int pix) { + __asm { + mov eax, 0x01080108 // generate multiplier to repeat 5 bits + movd xmm5, eax + pshufd xmm5, xmm5, 0 + mov eax, 0x20802080 // multiplier shift by 5 and then repeat 6 bits + movd xmm6, eax + pshufd xmm6, xmm6, 0 + pcmpeqb xmm3, xmm3 // generate mask 0xf800f800 for Red + psllw xmm3, 11 + pcmpeqb xmm4, xmm4 // generate mask 0x07e007e0 for Green + psllw xmm4, 10 + psrlw xmm4, 5 + pcmpeqb xmm7, xmm7 // generate mask 0xff00ff00 for Alpha + psllw xmm7, 8 + + mov eax, [esp + 4] // src_rgb565 + mov edx, [esp + 8] // dst_argb + mov ecx, [esp + 12] // pix + sub edx, eax + sub edx, eax + + convertloop: + movdqu xmm0, [eax] // fetch 8 pixels of bgr565 + movdqa xmm1, xmm0 + movdqa xmm2, xmm0 + pand xmm1, xmm3 // R in upper 5 bits + psllw xmm2, 11 // B in upper 5 bits + pmulhuw xmm1, xmm5 // * (256 + 8) + pmulhuw xmm2, xmm5 // * (256 + 8) + psllw xmm1, 8 + por xmm1, xmm2 // RB + pand xmm0, xmm4 // G in middle 6 bits + pmulhuw xmm0, xmm6 // << 5 * (256 + 4) + por xmm0, xmm7 // AG + movdqa xmm2, xmm1 + punpcklbw xmm1, xmm0 + punpckhbw xmm2, xmm0 + movdqu [eax * 2 + edx], xmm1 // store 4 pixels of ARGB + movdqu [eax * 2 + edx + 16], xmm2 // store next 4 pixels of ARGB + lea eax, [eax + 16] + sub ecx, 8 + jg convertloop + ret + } +} + +#ifdef HAS_RGB565TOARGBROW_AVX2 +// pmul method to replicate bits. +// Math to replicate bits: +// (v << 8) | (v << 3) +// v * 256 + v * 8 +// v * (256 + 8) +// G shift of 5 is incorporated, so shift is 5 + 8 and 5 + 3 +__declspec(naked) +void RGB565ToARGBRow_AVX2(const uint8* src_rgb565, uint8* dst_argb, + int pix) { + __asm { + mov eax, 0x01080108 // generate multiplier to repeat 5 bits + vmovd xmm5, eax + vbroadcastss ymm5, xmm5 + mov eax, 0x20802080 // multiplier shift by 5 and then repeat 6 bits + movd xmm6, eax + vbroadcastss ymm6, xmm6 + vpcmpeqb ymm3, ymm3, ymm3 // generate mask 0xf800f800 for Red + vpsllw ymm3, ymm3, 11 + vpcmpeqb ymm4, ymm4, ymm4 // generate mask 0x07e007e0 for Green + vpsllw ymm4, ymm4, 10 + vpsrlw ymm4, ymm4, 5 + vpcmpeqb ymm7, ymm7, ymm7 // generate mask 0xff00ff00 for Alpha + vpsllw ymm7, ymm7, 8 + + mov eax, [esp + 4] // src_rgb565 + mov edx, [esp + 8] // dst_argb + mov ecx, [esp + 12] // pix + sub edx, eax + sub edx, eax + + convertloop: + vmovdqu ymm0, [eax] // fetch 16 pixels of bgr565 + vpand ymm1, ymm0, ymm3 // R in upper 5 bits + vpsllw ymm2, ymm0, 11 // B in upper 5 bits + vpmulhuw ymm1, ymm1, ymm5 // * (256 + 8) + vpmulhuw ymm2, ymm2, ymm5 // * (256 + 8) + vpsllw ymm1, ymm1, 8 + vpor ymm1, ymm1, ymm2 // RB + vpand ymm0, ymm0, ymm4 // G in middle 6 bits + vpmulhuw ymm0, ymm0, ymm6 // << 5 * (256 + 4) + vpor ymm0, ymm0, ymm7 // AG + vpermq ymm0, ymm0, 0xd8 // mutate for unpack + vpermq ymm1, ymm1, 0xd8 + vpunpckhbw ymm2, ymm1, ymm0 + vpunpcklbw ymm1, ymm1, ymm0 + vmovdqu [eax * 2 + edx], ymm1 // store 4 pixels of ARGB + vmovdqu [eax * 2 + edx + 32], ymm2 // store next 4 pixels of ARGB + lea eax, [eax + 32] + sub ecx, 16 + jg convertloop + vzeroupper + ret + } +} +#endif // HAS_RGB565TOARGBROW_AVX2 + +#ifdef HAS_ARGB1555TOARGBROW_AVX2 +__declspec(naked) +void ARGB1555ToARGBRow_AVX2(const uint8* src_argb1555, uint8* dst_argb, + int pix) { + __asm { + mov eax, 0x01080108 // generate multiplier to repeat 5 bits + vmovd xmm5, eax + vbroadcastss ymm5, xmm5 + mov eax, 0x42004200 // multiplier shift by 6 and then repeat 5 bits + movd xmm6, eax + vbroadcastss ymm6, xmm6 + vpcmpeqb ymm3, ymm3, ymm3 // generate mask 0xf800f800 for Red + vpsllw ymm3, ymm3, 11 + vpsrlw ymm4, ymm3, 6 // generate mask 0x03e003e0 for Green + vpcmpeqb ymm7, ymm7, ymm7 // generate mask 0xff00ff00 for Alpha + vpsllw ymm7, ymm7, 8 + + mov eax, [esp + 4] // src_argb1555 + mov edx, [esp + 8] // dst_argb + mov ecx, [esp + 12] // pix + sub edx, eax + sub edx, eax + + convertloop: + vmovdqu ymm0, [eax] // fetch 16 pixels of 1555 + vpsllw ymm1, ymm0, 1 // R in upper 5 bits + vpsllw ymm2, ymm0, 11 // B in upper 5 bits + vpand ymm1, ymm1, ymm3 + vpmulhuw ymm2, ymm2, ymm5 // * (256 + 8) + vpmulhuw ymm1, ymm1, ymm5 // * (256 + 8) + vpsllw ymm1, ymm1, 8 + vpor ymm1, ymm1, ymm2 // RB + vpsraw ymm2, ymm0, 8 // A + vpand ymm0, ymm0, ymm4 // G in middle 5 bits + vpmulhuw ymm0, ymm0, ymm6 // << 6 * (256 + 8) + vpand ymm2, ymm2, ymm7 + vpor ymm0, ymm0, ymm2 // AG + vpermq ymm0, ymm0, 0xd8 // mutate for unpack + vpermq ymm1, ymm1, 0xd8 + vpunpckhbw ymm2, ymm1, ymm0 + vpunpcklbw ymm1, ymm1, ymm0 + vmovdqu [eax * 2 + edx], ymm1 // store 8 pixels of ARGB + vmovdqu [eax * 2 + edx + 32], ymm2 // store next 8 pixels of ARGB + lea eax, [eax + 32] + sub ecx, 16 + jg convertloop + vzeroupper + ret + } +} +#endif // HAS_ARGB1555TOARGBROW_AVX2 + +#ifdef HAS_ARGB4444TOARGBROW_AVX2 +__declspec(naked) +void ARGB4444ToARGBRow_AVX2(const uint8* src_argb4444, uint8* dst_argb, + int pix) { + __asm { + mov eax, 0x0f0f0f0f // generate mask 0x0f0f0f0f + vmovd xmm4, eax + vbroadcastss ymm4, xmm4 + vpslld ymm5, ymm4, 4 // 0xf0f0f0f0 for high nibbles + mov eax, [esp + 4] // src_argb4444 + mov edx, [esp + 8] // dst_argb + mov ecx, [esp + 12] // pix + sub edx, eax + sub edx, eax + + convertloop: + vmovdqu ymm0, [eax] // fetch 16 pixels of bgra4444 + vpand ymm2, ymm0, ymm5 // mask high nibbles + vpand ymm0, ymm0, ymm4 // mask low nibbles + vpsrlw ymm3, ymm2, 4 + vpsllw ymm1, ymm0, 4 + vpor ymm2, ymm2, ymm3 + vpor ymm0, ymm0, ymm1 + vpermq ymm0, ymm0, 0xd8 // mutate for unpack + vpermq ymm2, ymm2, 0xd8 + vpunpckhbw ymm1, ymm0, ymm2 + vpunpcklbw ymm0, ymm0, ymm2 + vmovdqu [eax * 2 + edx], ymm0 // store 8 pixels of ARGB + vmovdqu [eax * 2 + edx + 32], ymm1 // store next 8 pixels of ARGB + lea eax, [eax + 32] + sub ecx, 16 + jg convertloop + vzeroupper + ret + } +} +#endif // HAS_ARGB4444TOARGBROW_AVX2 + +// 24 instructions +__declspec(naked) +void ARGB1555ToARGBRow_SSE2(const uint8* src_argb1555, uint8* dst_argb, + int pix) { + __asm { + mov eax, 0x01080108 // generate multiplier to repeat 5 bits + movd xmm5, eax + pshufd xmm5, xmm5, 0 + mov eax, 0x42004200 // multiplier shift by 6 and then repeat 5 bits + movd xmm6, eax + pshufd xmm6, xmm6, 0 + pcmpeqb xmm3, xmm3 // generate mask 0xf800f800 for Red + psllw xmm3, 11 + movdqa xmm4, xmm3 // generate mask 0x03e003e0 for Green + psrlw xmm4, 6 + pcmpeqb xmm7, xmm7 // generate mask 0xff00ff00 for Alpha + psllw xmm7, 8 + + mov eax, [esp + 4] // src_argb1555 + mov edx, [esp + 8] // dst_argb + mov ecx, [esp + 12] // pix + sub edx, eax + sub edx, eax + + convertloop: + movdqu xmm0, [eax] // fetch 8 pixels of 1555 + movdqa xmm1, xmm0 + movdqa xmm2, xmm0 + psllw xmm1, 1 // R in upper 5 bits + psllw xmm2, 11 // B in upper 5 bits + pand xmm1, xmm3 + pmulhuw xmm2, xmm5 // * (256 + 8) + pmulhuw xmm1, xmm5 // * (256 + 8) + psllw xmm1, 8 + por xmm1, xmm2 // RB + movdqa xmm2, xmm0 + pand xmm0, xmm4 // G in middle 5 bits + psraw xmm2, 8 // A + pmulhuw xmm0, xmm6 // << 6 * (256 + 8) + pand xmm2, xmm7 + por xmm0, xmm2 // AG + movdqa xmm2, xmm1 + punpcklbw xmm1, xmm0 + punpckhbw xmm2, xmm0 + movdqu [eax * 2 + edx], xmm1 // store 4 pixels of ARGB + movdqu [eax * 2 + edx + 16], xmm2 // store next 4 pixels of ARGB + lea eax, [eax + 16] + sub ecx, 8 + jg convertloop + ret + } +} + +// 18 instructions. +__declspec(naked) +void ARGB4444ToARGBRow_SSE2(const uint8* src_argb4444, uint8* dst_argb, + int pix) { + __asm { + mov eax, 0x0f0f0f0f // generate mask 0x0f0f0f0f + movd xmm4, eax + pshufd xmm4, xmm4, 0 + movdqa xmm5, xmm4 // 0xf0f0f0f0 for high nibbles + pslld xmm5, 4 + mov eax, [esp + 4] // src_argb4444 + mov edx, [esp + 8] // dst_argb + mov ecx, [esp + 12] // pix + sub edx, eax + sub edx, eax + + convertloop: + movdqu xmm0, [eax] // fetch 8 pixels of bgra4444 + movdqa xmm2, xmm0 + pand xmm0, xmm4 // mask low nibbles + pand xmm2, xmm5 // mask high nibbles + movdqa xmm1, xmm0 + movdqa xmm3, xmm2 + psllw xmm1, 4 + psrlw xmm3, 4 + por xmm0, xmm1 + por xmm2, xmm3 + movdqa xmm1, xmm0 + punpcklbw xmm0, xmm2 + punpckhbw xmm1, xmm2 + movdqu [eax * 2 + edx], xmm0 // store 4 pixels of ARGB + movdqu [eax * 2 + edx + 16], xmm1 // store next 4 pixels of ARGB + lea eax, [eax + 16] + sub ecx, 8 + jg convertloop + ret + } +} + +__declspec(naked) +void ARGBToRGB24Row_SSSE3(const uint8* src_argb, uint8* dst_rgb, int pix) { + __asm { + mov eax, [esp + 4] // src_argb + mov edx, [esp + 8] // dst_rgb + mov ecx, [esp + 12] // pix + movdqa xmm6, kShuffleMaskARGBToRGB24 + + convertloop: + movdqu xmm0, [eax] // fetch 16 pixels of argb + movdqu xmm1, [eax + 16] + movdqu xmm2, [eax + 32] + movdqu xmm3, [eax + 48] + lea eax, [eax + 64] + pshufb xmm0, xmm6 // pack 16 bytes of ARGB to 12 bytes of RGB + pshufb xmm1, xmm6 + pshufb xmm2, xmm6 + pshufb xmm3, xmm6 + movdqa xmm4, xmm1 // 4 bytes from 1 for 0 + psrldq xmm1, 4 // 8 bytes from 1 + pslldq xmm4, 12 // 4 bytes from 1 for 0 + movdqa xmm5, xmm2 // 8 bytes from 2 for 1 + por xmm0, xmm4 // 4 bytes from 1 for 0 + pslldq xmm5, 8 // 8 bytes from 2 for 1 + movdqu [edx], xmm0 // store 0 + por xmm1, xmm5 // 8 bytes from 2 for 1 + psrldq xmm2, 8 // 4 bytes from 2 + pslldq xmm3, 4 // 12 bytes from 3 for 2 + por xmm2, xmm3 // 12 bytes from 3 for 2 + movdqu [edx + 16], xmm1 // store 1 + movdqu [edx + 32], xmm2 // store 2 + lea edx, [edx + 48] + sub ecx, 16 + jg convertloop + ret + } +} + +__declspec(naked) +void ARGBToRAWRow_SSSE3(const uint8* src_argb, uint8* dst_rgb, int pix) { + __asm { + mov eax, [esp + 4] // src_argb + mov edx, [esp + 8] // dst_rgb + mov ecx, [esp + 12] // pix + movdqa xmm6, kShuffleMaskARGBToRAW + + convertloop: + movdqu xmm0, [eax] // fetch 16 pixels of argb + movdqu xmm1, [eax + 16] + movdqu xmm2, [eax + 32] + movdqu xmm3, [eax + 48] + lea eax, [eax + 64] + pshufb xmm0, xmm6 // pack 16 bytes of ARGB to 12 bytes of RGB + pshufb xmm1, xmm6 + pshufb xmm2, xmm6 + pshufb xmm3, xmm6 + movdqa xmm4, xmm1 // 4 bytes from 1 for 0 + psrldq xmm1, 4 // 8 bytes from 1 + pslldq xmm4, 12 // 4 bytes from 1 for 0 + movdqa xmm5, xmm2 // 8 bytes from 2 for 1 + por xmm0, xmm4 // 4 bytes from 1 for 0 + pslldq xmm5, 8 // 8 bytes from 2 for 1 + movdqu [edx], xmm0 // store 0 + por xmm1, xmm5 // 8 bytes from 2 for 1 + psrldq xmm2, 8 // 4 bytes from 2 + pslldq xmm3, 4 // 12 bytes from 3 for 2 + por xmm2, xmm3 // 12 bytes from 3 for 2 + movdqu [edx + 16], xmm1 // store 1 + movdqu [edx + 32], xmm2 // store 2 + lea edx, [edx + 48] + sub ecx, 16 + jg convertloop + ret + } +} + +// 4 pixels +__declspec(naked) +void ARGBToRGB565Row_SSE2(const uint8* src_argb, uint8* dst_rgb, int pix) { + __asm { + mov eax, [esp + 4] // src_argb + mov edx, [esp + 8] // dst_rgb + mov ecx, [esp + 12] // pix + pcmpeqb xmm3, xmm3 // generate mask 0x0000001f + psrld xmm3, 27 + pcmpeqb xmm4, xmm4 // generate mask 0x000007e0 + psrld xmm4, 26 + pslld xmm4, 5 + pcmpeqb xmm5, xmm5 // generate mask 0xfffff800 + pslld xmm5, 11 + + convertloop: + movdqu xmm0, [eax] // fetch 4 pixels of argb + movdqa xmm1, xmm0 // B + movdqa xmm2, xmm0 // G + pslld xmm0, 8 // R + psrld xmm1, 3 // B + psrld xmm2, 5 // G + psrad xmm0, 16 // R + pand xmm1, xmm3 // B + pand xmm2, xmm4 // G + pand xmm0, xmm5 // R + por xmm1, xmm2 // BG + por xmm0, xmm1 // BGR + packssdw xmm0, xmm0 + lea eax, [eax + 16] + movq qword ptr [edx], xmm0 // store 4 pixels of RGB565 + lea edx, [edx + 8] + sub ecx, 4 + jg convertloop + ret + } +} + +// 8 pixels +__declspec(naked) +void ARGBToRGB565DitherRow_SSE2(const uint8* src_argb, uint8* dst_rgb, + const uint32 dither4, int pix) { + __asm { + + mov eax, [esp + 4] // src_argb + mov edx, [esp + 8] // dst_rgb + movd xmm6, [esp + 12] // dither4 + mov ecx, [esp + 16] // pix + punpcklbw xmm6, xmm6 // make dither 16 bytes + movdqa xmm7, xmm6 + punpcklwd xmm6, xmm6 + punpckhwd xmm7, xmm7 + pcmpeqb xmm3, xmm3 // generate mask 0x0000001f + psrld xmm3, 27 + pcmpeqb xmm4, xmm4 // generate mask 0x000007e0 + psrld xmm4, 26 + pslld xmm4, 5 + pcmpeqb xmm5, xmm5 // generate mask 0xfffff800 + pslld xmm5, 11 + + convertloop: + movdqu xmm0, [eax] // fetch 4 pixels of argb + paddusb xmm0, xmm6 // add dither + movdqa xmm1, xmm0 // B + movdqa xmm2, xmm0 // G + pslld xmm0, 8 // R + psrld xmm1, 3 // B + psrld xmm2, 5 // G + psrad xmm0, 16 // R + pand xmm1, xmm3 // B + pand xmm2, xmm4 // G + pand xmm0, xmm5 // R + por xmm1, xmm2 // BG + por xmm0, xmm1 // BGR + packssdw xmm0, xmm0 + lea eax, [eax + 16] + movq qword ptr [edx], xmm0 // store 4 pixels of RGB565 + lea edx, [edx + 8] + sub ecx, 4 + jg convertloop + ret + } +} + +#ifdef HAS_ARGBTORGB565DITHERROW_AVX2 +__declspec(naked) +void ARGBToRGB565DitherRow_AVX2(const uint8* src_argb, uint8* dst_rgb, + const uint32 dither4, int pix) { + __asm { + mov eax, [esp + 4] // src_argb + mov edx, [esp + 8] // dst_rgb + vbroadcastss xmm6, [esp + 12] // dither4 + mov ecx, [esp + 16] // pix + vpunpcklbw xmm6, xmm6, xmm6 // make dither 32 bytes + vpermq ymm6, ymm6, 0xd8 + vpunpcklwd ymm6, ymm6, ymm6 + vpcmpeqb ymm3, ymm3, ymm3 // generate mask 0x0000001f + vpsrld ymm3, ymm3, 27 + vpcmpeqb ymm4, ymm4, ymm4 // generate mask 0x000007e0 + vpsrld ymm4, ymm4, 26 + vpslld ymm4, ymm4, 5 + vpslld ymm5, ymm3, 11 // generate mask 0x0000f800 + + convertloop: + vmovdqu ymm0, [eax] // fetch 8 pixels of argb + vpaddusb ymm0, ymm0, ymm6 // add dither + vpsrld ymm2, ymm0, 5 // G + vpsrld ymm1, ymm0, 3 // B + vpsrld ymm0, ymm0, 8 // R + vpand ymm2, ymm2, ymm4 // G + vpand ymm1, ymm1, ymm3 // B + vpand ymm0, ymm0, ymm5 // R + vpor ymm1, ymm1, ymm2 // BG + vpor ymm0, ymm0, ymm1 // BGR + vpackusdw ymm0, ymm0, ymm0 + vpermq ymm0, ymm0, 0xd8 + lea eax, [eax + 32] + vmovdqu [edx], xmm0 // store 8 pixels of RGB565 + lea edx, [edx + 16] + sub ecx, 8 + jg convertloop + vzeroupper + ret + } +} +#endif // HAS_ARGBTORGB565DITHERROW_AVX2 + +// TODO(fbarchard): Improve sign extension/packing. +__declspec(naked) +void ARGBToARGB1555Row_SSE2(const uint8* src_argb, uint8* dst_rgb, int pix) { + __asm { + mov eax, [esp + 4] // src_argb + mov edx, [esp + 8] // dst_rgb + mov ecx, [esp + 12] // pix + pcmpeqb xmm4, xmm4 // generate mask 0x0000001f + psrld xmm4, 27 + movdqa xmm5, xmm4 // generate mask 0x000003e0 + pslld xmm5, 5 + movdqa xmm6, xmm4 // generate mask 0x00007c00 + pslld xmm6, 10 + pcmpeqb xmm7, xmm7 // generate mask 0xffff8000 + pslld xmm7, 15 + + convertloop: + movdqu xmm0, [eax] // fetch 4 pixels of argb + movdqa xmm1, xmm0 // B + movdqa xmm2, xmm0 // G + movdqa xmm3, xmm0 // R + psrad xmm0, 16 // A + psrld xmm1, 3 // B + psrld xmm2, 6 // G + psrld xmm3, 9 // R + pand xmm0, xmm7 // A + pand xmm1, xmm4 // B + pand xmm2, xmm5 // G + pand xmm3, xmm6 // R + por xmm0, xmm1 // BA + por xmm2, xmm3 // GR + por xmm0, xmm2 // BGRA + packssdw xmm0, xmm0 + lea eax, [eax + 16] + movq qword ptr [edx], xmm0 // store 4 pixels of ARGB1555 + lea edx, [edx + 8] + sub ecx, 4 + jg convertloop + ret + } +} + +__declspec(naked) +void ARGBToARGB4444Row_SSE2(const uint8* src_argb, uint8* dst_rgb, int pix) { + __asm { + mov eax, [esp + 4] // src_argb + mov edx, [esp + 8] // dst_rgb + mov ecx, [esp + 12] // pix + pcmpeqb xmm4, xmm4 // generate mask 0xf000f000 + psllw xmm4, 12 + movdqa xmm3, xmm4 // generate mask 0x00f000f0 + psrlw xmm3, 8 + + convertloop: + movdqu xmm0, [eax] // fetch 4 pixels of argb + movdqa xmm1, xmm0 + pand xmm0, xmm3 // low nibble + pand xmm1, xmm4 // high nibble + psrld xmm0, 4 + psrld xmm1, 8 + por xmm0, xmm1 + packuswb xmm0, xmm0 + lea eax, [eax + 16] + movq qword ptr [edx], xmm0 // store 4 pixels of ARGB4444 + lea edx, [edx + 8] + sub ecx, 4 + jg convertloop + ret + } +} + +#ifdef HAS_ARGBTORGB565ROW_AVX2 +__declspec(naked) +void ARGBToRGB565Row_AVX2(const uint8* src_argb, uint8* dst_rgb, int pix) { + __asm { + mov eax, [esp + 4] // src_argb + mov edx, [esp + 8] // dst_rgb + mov ecx, [esp + 12] // pix + vpcmpeqb ymm3, ymm3, ymm3 // generate mask 0x0000001f + vpsrld ymm3, ymm3, 27 + vpcmpeqb ymm4, ymm4, ymm4 // generate mask 0x000007e0 + vpsrld ymm4, ymm4, 26 + vpslld ymm4, ymm4, 5 + vpslld ymm5, ymm3, 11 // generate mask 0x0000f800 + + convertloop: + vmovdqu ymm0, [eax] // fetch 8 pixels of argb + vpsrld ymm2, ymm0, 5 // G + vpsrld ymm1, ymm0, 3 // B + vpsrld ymm0, ymm0, 8 // R + vpand ymm2, ymm2, ymm4 // G + vpand ymm1, ymm1, ymm3 // B + vpand ymm0, ymm0, ymm5 // R + vpor ymm1, ymm1, ymm2 // BG + vpor ymm0, ymm0, ymm1 // BGR + vpackusdw ymm0, ymm0, ymm0 + vpermq ymm0, ymm0, 0xd8 + lea eax, [eax + 32] + vmovdqu [edx], xmm0 // store 8 pixels of RGB565 + lea edx, [edx + 16] + sub ecx, 8 + jg convertloop + vzeroupper + ret + } +} +#endif // HAS_ARGBTORGB565ROW_AVX2 + +#ifdef HAS_ARGBTOARGB1555ROW_AVX2 +__declspec(naked) +void ARGBToARGB1555Row_AVX2(const uint8* src_argb, uint8* dst_rgb, int pix) { + __asm { + mov eax, [esp + 4] // src_argb + mov edx, [esp + 8] // dst_rgb + mov ecx, [esp + 12] // pix + vpcmpeqb ymm4, ymm4, ymm4 + vpsrld ymm4, ymm4, 27 // generate mask 0x0000001f + vpslld ymm5, ymm4, 5 // generate mask 0x000003e0 + vpslld ymm6, ymm4, 10 // generate mask 0x00007c00 + vpcmpeqb ymm7, ymm7, ymm7 // generate mask 0xffff8000 + vpslld ymm7, ymm7, 15 + + convertloop: + vmovdqu ymm0, [eax] // fetch 8 pixels of argb + vpsrld ymm3, ymm0, 9 // R + vpsrld ymm2, ymm0, 6 // G + vpsrld ymm1, ymm0, 3 // B + vpsrad ymm0, ymm0, 16 // A + vpand ymm3, ymm3, ymm6 // R + vpand ymm2, ymm2, ymm5 // G + vpand ymm1, ymm1, ymm4 // B + vpand ymm0, ymm0, ymm7 // A + vpor ymm0, ymm0, ymm1 // BA + vpor ymm2, ymm2, ymm3 // GR + vpor ymm0, ymm0, ymm2 // BGRA + vpackssdw ymm0, ymm0, ymm0 + vpermq ymm0, ymm0, 0xd8 + lea eax, [eax + 32] + vmovdqu [edx], xmm0 // store 8 pixels of ARGB1555 + lea edx, [edx + 16] + sub ecx, 8 + jg convertloop + vzeroupper + ret + } +} +#endif // HAS_ARGBTOARGB1555ROW_AVX2 + +#ifdef HAS_ARGBTOARGB4444ROW_AVX2 +__declspec(naked) +void ARGBToARGB4444Row_AVX2(const uint8* src_argb, uint8* dst_rgb, int pix) { + __asm { + mov eax, [esp + 4] // src_argb + mov edx, [esp + 8] // dst_rgb + mov ecx, [esp + 12] // pix + vpcmpeqb ymm4, ymm4, ymm4 // generate mask 0xf000f000 + vpsllw ymm4, ymm4, 12 + vpsrlw ymm3, ymm4, 8 // generate mask 0x00f000f0 + + convertloop: + vmovdqu ymm0, [eax] // fetch 8 pixels of argb + vpand ymm1, ymm0, ymm4 // high nibble + vpand ymm0, ymm0, ymm3 // low nibble + vpsrld ymm1, ymm1, 8 + vpsrld ymm0, ymm0, 4 + vpor ymm0, ymm0, ymm1 + vpackuswb ymm0, ymm0, ymm0 + vpermq ymm0, ymm0, 0xd8 + lea eax, [eax + 32] + vmovdqu [edx], xmm0 // store 8 pixels of ARGB4444 + lea edx, [edx + 16] + sub ecx, 8 + jg convertloop + vzeroupper + ret + } +} +#endif // HAS_ARGBTOARGB4444ROW_AVX2 + +// Convert 16 ARGB pixels (64 bytes) to 16 Y values. +__declspec(naked) +void ARGBToYRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) { + __asm { + mov eax, [esp + 4] /* src_argb */ + mov edx, [esp + 8] /* dst_y */ + mov ecx, [esp + 12] /* pix */ + movdqa xmm4, kARGBToY + movdqa xmm5, kAddY16 + + convertloop: + movdqu xmm0, [eax] + movdqu xmm1, [eax + 16] + movdqu xmm2, [eax + 32] + movdqu xmm3, [eax + 48] + pmaddubsw xmm0, xmm4 + pmaddubsw xmm1, xmm4 + pmaddubsw xmm2, xmm4 + pmaddubsw xmm3, xmm4 + lea eax, [eax + 64] + phaddw xmm0, xmm1 + phaddw xmm2, xmm3 + psrlw xmm0, 7 + psrlw xmm2, 7 + packuswb xmm0, xmm2 + paddb xmm0, xmm5 + movdqu [edx], xmm0 + lea edx, [edx + 16] + sub ecx, 16 + jg convertloop + ret + } +} + +// Convert 16 ARGB pixels (64 bytes) to 16 YJ values. +// Same as ARGBToYRow but different coefficients, no add 16, but do rounding. +__declspec(naked) +void ARGBToYJRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) { + __asm { + mov eax, [esp + 4] /* src_argb */ + mov edx, [esp + 8] /* dst_y */ + mov ecx, [esp + 12] /* pix */ + movdqa xmm4, kARGBToYJ + movdqa xmm5, kAddYJ64 + + convertloop: + movdqu xmm0, [eax] + movdqu xmm1, [eax + 16] + movdqu xmm2, [eax + 32] + movdqu xmm3, [eax + 48] + pmaddubsw xmm0, xmm4 + pmaddubsw xmm1, xmm4 + pmaddubsw xmm2, xmm4 + pmaddubsw xmm3, xmm4 + lea eax, [eax + 64] + phaddw xmm0, xmm1 + phaddw xmm2, xmm3 + paddw xmm0, xmm5 // Add .5 for rounding. + paddw xmm2, xmm5 + psrlw xmm0, 7 + psrlw xmm2, 7 + packuswb xmm0, xmm2 + movdqu [edx], xmm0 + lea edx, [edx + 16] + sub ecx, 16 + jg convertloop + ret + } +} + +#ifdef HAS_ARGBTOYROW_AVX2 +// vpermd for vphaddw + vpackuswb vpermd. +static const lvec32 kPermdARGBToY_AVX = { + 0, 4, 1, 5, 2, 6, 3, 7 +}; + +// Convert 32 ARGB pixels (128 bytes) to 32 Y values. +__declspec(naked) +void ARGBToYRow_AVX2(const uint8* src_argb, uint8* dst_y, int pix) { + __asm { + mov eax, [esp + 4] /* src_argb */ + mov edx, [esp + 8] /* dst_y */ + mov ecx, [esp + 12] /* pix */ + vbroadcastf128 ymm4, kARGBToY + vbroadcastf128 ymm5, kAddY16 + vmovdqu ymm6, kPermdARGBToY_AVX + + convertloop: + vmovdqu ymm0, [eax] + vmovdqu ymm1, [eax + 32] + vmovdqu ymm2, [eax + 64] + vmovdqu ymm3, [eax + 96] + vpmaddubsw ymm0, ymm0, ymm4 + vpmaddubsw ymm1, ymm1, ymm4 + vpmaddubsw ymm2, ymm2, ymm4 + vpmaddubsw ymm3, ymm3, ymm4 + lea eax, [eax + 128] + vphaddw ymm0, ymm0, ymm1 // mutates. + vphaddw ymm2, ymm2, ymm3 + vpsrlw ymm0, ymm0, 7 + vpsrlw ymm2, ymm2, 7 + vpackuswb ymm0, ymm0, ymm2 // mutates. + vpermd ymm0, ymm6, ymm0 // For vphaddw + vpackuswb mutation. + vpaddb ymm0, ymm0, ymm5 // add 16 for Y + vmovdqu [edx], ymm0 + lea edx, [edx + 32] + sub ecx, 32 + jg convertloop + vzeroupper + ret + } +} +#endif // HAS_ARGBTOYROW_AVX2 + +#ifdef HAS_ARGBTOYJROW_AVX2 +// Convert 32 ARGB pixels (128 bytes) to 32 Y values. +__declspec(naked) +void ARGBToYJRow_AVX2(const uint8* src_argb, uint8* dst_y, int pix) { + __asm { + mov eax, [esp + 4] /* src_argb */ + mov edx, [esp + 8] /* dst_y */ + mov ecx, [esp + 12] /* pix */ + vbroadcastf128 ymm4, kARGBToYJ + vbroadcastf128 ymm5, kAddYJ64 + vmovdqu ymm6, kPermdARGBToY_AVX + + convertloop: + vmovdqu ymm0, [eax] + vmovdqu ymm1, [eax + 32] + vmovdqu ymm2, [eax + 64] + vmovdqu ymm3, [eax + 96] + vpmaddubsw ymm0, ymm0, ymm4 + vpmaddubsw ymm1, ymm1, ymm4 + vpmaddubsw ymm2, ymm2, ymm4 + vpmaddubsw ymm3, ymm3, ymm4 + lea eax, [eax + 128] + vphaddw ymm0, ymm0, ymm1 // mutates. + vphaddw ymm2, ymm2, ymm3 + vpaddw ymm0, ymm0, ymm5 // Add .5 for rounding. + vpaddw ymm2, ymm2, ymm5 + vpsrlw ymm0, ymm0, 7 + vpsrlw ymm2, ymm2, 7 + vpackuswb ymm0, ymm0, ymm2 // mutates. + vpermd ymm0, ymm6, ymm0 // For vphaddw + vpackuswb mutation. + vmovdqu [edx], ymm0 + lea edx, [edx + 32] + sub ecx, 32 + jg convertloop + + vzeroupper + ret + } +} +#endif // HAS_ARGBTOYJROW_AVX2 + +__declspec(naked) +void BGRAToYRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) { + __asm { + mov eax, [esp + 4] /* src_argb */ + mov edx, [esp + 8] /* dst_y */ + mov ecx, [esp + 12] /* pix */ + movdqa xmm4, kBGRAToY + movdqa xmm5, kAddY16 + + convertloop: + movdqu xmm0, [eax] + movdqu xmm1, [eax + 16] + movdqu xmm2, [eax + 32] + movdqu xmm3, [eax + 48] + pmaddubsw xmm0, xmm4 + pmaddubsw xmm1, xmm4 + pmaddubsw xmm2, xmm4 + pmaddubsw xmm3, xmm4 + lea eax, [eax + 64] + phaddw xmm0, xmm1 + phaddw xmm2, xmm3 + psrlw xmm0, 7 + psrlw xmm2, 7 + packuswb xmm0, xmm2 + paddb xmm0, xmm5 + movdqu [edx], xmm0 + lea edx, [edx + 16] + sub ecx, 16 + jg convertloop + ret + } +} + +__declspec(naked) +void ABGRToYRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) { + __asm { + mov eax, [esp + 4] /* src_argb */ + mov edx, [esp + 8] /* dst_y */ + mov ecx, [esp + 12] /* pix */ + movdqa xmm4, kABGRToY + movdqa xmm5, kAddY16 + + convertloop: + movdqu xmm0, [eax] + movdqu xmm1, [eax + 16] + movdqu xmm2, [eax + 32] + movdqu xmm3, [eax + 48] + pmaddubsw xmm0, xmm4 + pmaddubsw xmm1, xmm4 + pmaddubsw xmm2, xmm4 + pmaddubsw xmm3, xmm4 + lea eax, [eax + 64] + phaddw xmm0, xmm1 + phaddw xmm2, xmm3 + psrlw xmm0, 7 + psrlw xmm2, 7 + packuswb xmm0, xmm2 + paddb xmm0, xmm5 + movdqu [edx], xmm0 + lea edx, [edx + 16] + sub ecx, 16 + jg convertloop + ret + } +} + +__declspec(naked) +void RGBAToYRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) { + __asm { + mov eax, [esp + 4] /* src_argb */ + mov edx, [esp + 8] /* dst_y */ + mov ecx, [esp + 12] /* pix */ + movdqa xmm4, kRGBAToY + movdqa xmm5, kAddY16 + + convertloop: + movdqu xmm0, [eax] + movdqu xmm1, [eax + 16] + movdqu xmm2, [eax + 32] + movdqu xmm3, [eax + 48] + pmaddubsw xmm0, xmm4 + pmaddubsw xmm1, xmm4 + pmaddubsw xmm2, xmm4 + pmaddubsw xmm3, xmm4 + lea eax, [eax + 64] + phaddw xmm0, xmm1 + phaddw xmm2, xmm3 + psrlw xmm0, 7 + psrlw xmm2, 7 + packuswb xmm0, xmm2 + paddb xmm0, xmm5 + movdqu [edx], xmm0 + lea edx, [edx + 16] + sub ecx, 16 + jg convertloop + ret + } +} + +__declspec(naked) +void ARGBToUVRow_SSSE3(const uint8* src_argb0, int src_stride_argb, + uint8* dst_u, uint8* dst_v, int width) { + __asm { + push esi + push edi + mov eax, [esp + 8 + 4] // src_argb + mov esi, [esp + 8 + 8] // src_stride_argb + mov edx, [esp + 8 + 12] // dst_u + mov edi, [esp + 8 + 16] // dst_v + mov ecx, [esp + 8 + 20] // pix + movdqa xmm5, kAddUV128 + movdqa xmm6, kARGBToV + movdqa xmm7, kARGBToU + sub edi, edx // stride from u to v + + convertloop: + /* step 1 - subsample 16x2 argb pixels to 8x1 */ + movdqu xmm0, [eax] + movdqu xmm4, [eax + esi] + pavgb xmm0, xmm4 + movdqu xmm1, [eax + 16] + movdqu xmm4, [eax + esi + 16] + pavgb xmm1, xmm4 + movdqu xmm2, [eax + 32] + movdqu xmm4, [eax + esi + 32] + pavgb xmm2, xmm4 + movdqu xmm3, [eax + 48] + movdqu xmm4, [eax + esi + 48] + pavgb xmm3, xmm4 + + lea eax, [eax + 64] + movdqa xmm4, xmm0 + shufps xmm0, xmm1, 0x88 + shufps xmm4, xmm1, 0xdd + pavgb xmm0, xmm4 + movdqa xmm4, xmm2 + shufps xmm2, xmm3, 0x88 + shufps xmm4, xmm3, 0xdd + pavgb xmm2, xmm4 + + // step 2 - convert to U and V + // from here down is very similar to Y code except + // instead of 16 different pixels, its 8 pixels of U and 8 of V + movdqa xmm1, xmm0 + movdqa xmm3, xmm2 + pmaddubsw xmm0, xmm7 // U + pmaddubsw xmm2, xmm7 + pmaddubsw xmm1, xmm6 // V + pmaddubsw xmm3, xmm6 + phaddw xmm0, xmm2 + phaddw xmm1, xmm3 + psraw xmm0, 8 + psraw xmm1, 8 + packsswb xmm0, xmm1 + paddb xmm0, xmm5 // -> unsigned + + // step 3 - store 8 U and 8 V values + movlps qword ptr [edx], xmm0 // U + movhps qword ptr [edx + edi], xmm0 // V + lea edx, [edx + 8] + sub ecx, 16 + jg convertloop + + pop edi + pop esi + ret + } +} + +__declspec(naked) +void ARGBToUVJRow_SSSE3(const uint8* src_argb0, int src_stride_argb, + uint8* dst_u, uint8* dst_v, int width) { + __asm { + push esi + push edi + mov eax, [esp + 8 + 4] // src_argb + mov esi, [esp + 8 + 8] // src_stride_argb + mov edx, [esp + 8 + 12] // dst_u + mov edi, [esp + 8 + 16] // dst_v + mov ecx, [esp + 8 + 20] // pix + movdqa xmm5, kAddUVJ128 + movdqa xmm6, kARGBToVJ + movdqa xmm7, kARGBToUJ + sub edi, edx // stride from u to v + + convertloop: + /* step 1 - subsample 16x2 argb pixels to 8x1 */ + movdqu xmm0, [eax] + movdqu xmm4, [eax + esi] + pavgb xmm0, xmm4 + movdqu xmm1, [eax + 16] + movdqu xmm4, [eax + esi + 16] + pavgb xmm1, xmm4 + movdqu xmm2, [eax + 32] + movdqu xmm4, [eax + esi + 32] + pavgb xmm2, xmm4 + movdqu xmm3, [eax + 48] + movdqu xmm4, [eax + esi + 48] + pavgb xmm3, xmm4 + + lea eax, [eax + 64] + movdqa xmm4, xmm0 + shufps xmm0, xmm1, 0x88 + shufps xmm4, xmm1, 0xdd + pavgb xmm0, xmm4 + movdqa xmm4, xmm2 + shufps xmm2, xmm3, 0x88 + shufps xmm4, xmm3, 0xdd + pavgb xmm2, xmm4 + + // step 2 - convert to U and V + // from here down is very similar to Y code except + // instead of 16 different pixels, its 8 pixels of U and 8 of V + movdqa xmm1, xmm0 + movdqa xmm3, xmm2 + pmaddubsw xmm0, xmm7 // U + pmaddubsw xmm2, xmm7 + pmaddubsw xmm1, xmm6 // V + pmaddubsw xmm3, xmm6 + phaddw xmm0, xmm2 + phaddw xmm1, xmm3 + paddw xmm0, xmm5 // +.5 rounding -> unsigned + paddw xmm1, xmm5 + psraw xmm0, 8 + psraw xmm1, 8 + packsswb xmm0, xmm1 + + // step 3 - store 8 U and 8 V values + movlps qword ptr [edx], xmm0 // U + movhps qword ptr [edx + edi], xmm0 // V + lea edx, [edx + 8] + sub ecx, 16 + jg convertloop + + pop edi + pop esi + ret + } +} + +#ifdef HAS_ARGBTOUVROW_AVX2 +__declspec(naked) +void ARGBToUVRow_AVX2(const uint8* src_argb0, int src_stride_argb, + uint8* dst_u, uint8* dst_v, int width) { + __asm { + push esi + push edi + mov eax, [esp + 8 + 4] // src_argb + mov esi, [esp + 8 + 8] // src_stride_argb + mov edx, [esp + 8 + 12] // dst_u + mov edi, [esp + 8 + 16] // dst_v + mov ecx, [esp + 8 + 20] // pix + vbroadcastf128 ymm5, kAddUV128 + vbroadcastf128 ymm6, kARGBToV + vbroadcastf128 ymm7, kARGBToU + sub edi, edx // stride from u to v + + convertloop: + /* step 1 - subsample 32x2 argb pixels to 16x1 */ + vmovdqu ymm0, [eax] + vmovdqu ymm1, [eax + 32] + vmovdqu ymm2, [eax + 64] + vmovdqu ymm3, [eax + 96] + vpavgb ymm0, ymm0, [eax + esi] + vpavgb ymm1, ymm1, [eax + esi + 32] + vpavgb ymm2, ymm2, [eax + esi + 64] + vpavgb ymm3, ymm3, [eax + esi + 96] + lea eax, [eax + 128] + vshufps ymm4, ymm0, ymm1, 0x88 + vshufps ymm0, ymm0, ymm1, 0xdd + vpavgb ymm0, ymm0, ymm4 // mutated by vshufps + vshufps ymm4, ymm2, ymm3, 0x88 + vshufps ymm2, ymm2, ymm3, 0xdd + vpavgb ymm2, ymm2, ymm4 // mutated by vshufps + + // step 2 - convert to U and V + // from here down is very similar to Y code except + // instead of 32 different pixels, its 16 pixels of U and 16 of V + vpmaddubsw ymm1, ymm0, ymm7 // U + vpmaddubsw ymm3, ymm2, ymm7 + vpmaddubsw ymm0, ymm0, ymm6 // V + vpmaddubsw ymm2, ymm2, ymm6 + vphaddw ymm1, ymm1, ymm3 // mutates + vphaddw ymm0, ymm0, ymm2 + vpsraw ymm1, ymm1, 8 + vpsraw ymm0, ymm0, 8 + vpacksswb ymm0, ymm1, ymm0 // mutates + vpermq ymm0, ymm0, 0xd8 // For vpacksswb + vpshufb ymm0, ymm0, kShufARGBToUV_AVX // For vshufps + vphaddw + vpaddb ymm0, ymm0, ymm5 // -> unsigned + + // step 3 - store 16 U and 16 V values + vextractf128 [edx], ymm0, 0 // U + vextractf128 [edx + edi], ymm0, 1 // V + lea edx, [edx + 16] + sub ecx, 32 + jg convertloop + + pop edi + pop esi + vzeroupper + ret + } +} +#endif // HAS_ARGBTOUVROW_AVX2 + +__declspec(naked) +void ARGBToUV444Row_SSSE3(const uint8* src_argb0, + uint8* dst_u, uint8* dst_v, int width) { + __asm { + push edi + mov eax, [esp + 4 + 4] // src_argb + mov edx, [esp + 4 + 8] // dst_u + mov edi, [esp + 4 + 12] // dst_v + mov ecx, [esp + 4 + 16] // pix + movdqa xmm5, kAddUV128 + movdqa xmm6, kARGBToV + movdqa xmm7, kARGBToU + sub edi, edx // stride from u to v + + convertloop: + /* convert to U and V */ + movdqu xmm0, [eax] // U + movdqu xmm1, [eax + 16] + movdqu xmm2, [eax + 32] + movdqu xmm3, [eax + 48] + pmaddubsw xmm0, xmm7 + pmaddubsw xmm1, xmm7 + pmaddubsw xmm2, xmm7 + pmaddubsw xmm3, xmm7 + phaddw xmm0, xmm1 + phaddw xmm2, xmm3 + psraw xmm0, 8 + psraw xmm2, 8 + packsswb xmm0, xmm2 + paddb xmm0, xmm5 + movdqu [edx], xmm0 + + movdqu xmm0, [eax] // V + movdqu xmm1, [eax + 16] + movdqu xmm2, [eax + 32] + movdqu xmm3, [eax + 48] + pmaddubsw xmm0, xmm6 + pmaddubsw xmm1, xmm6 + pmaddubsw xmm2, xmm6 + pmaddubsw xmm3, xmm6 + phaddw xmm0, xmm1 + phaddw xmm2, xmm3 + psraw xmm0, 8 + psraw xmm2, 8 + packsswb xmm0, xmm2 + paddb xmm0, xmm5 + lea eax, [eax + 64] + movdqu [edx + edi], xmm0 + lea edx, [edx + 16] + sub ecx, 16 + jg convertloop + + pop edi + ret + } +} + +__declspec(naked) +void ARGBToUV422Row_SSSE3(const uint8* src_argb0, + uint8* dst_u, uint8* dst_v, int width) { + __asm { + push edi + mov eax, [esp + 4 + 4] // src_argb + mov edx, [esp + 4 + 8] // dst_u + mov edi, [esp + 4 + 12] // dst_v + mov ecx, [esp + 4 + 16] // pix + movdqa xmm5, kAddUV128 + movdqa xmm6, kARGBToV + movdqa xmm7, kARGBToU + sub edi, edx // stride from u to v + + convertloop: + /* step 1 - subsample 16x2 argb pixels to 8x1 */ + movdqu xmm0, [eax] + movdqu xmm1, [eax + 16] + movdqu xmm2, [eax + 32] + movdqu xmm3, [eax + 48] + lea eax, [eax + 64] + movdqa xmm4, xmm0 + shufps xmm0, xmm1, 0x88 + shufps xmm4, xmm1, 0xdd + pavgb xmm0, xmm4 + movdqa xmm4, xmm2 + shufps xmm2, xmm3, 0x88 + shufps xmm4, xmm3, 0xdd + pavgb xmm2, xmm4 + + // step 2 - convert to U and V + // from here down is very similar to Y code except + // instead of 16 different pixels, its 8 pixels of U and 8 of V + movdqa xmm1, xmm0 + movdqa xmm3, xmm2 + pmaddubsw xmm0, xmm7 // U + pmaddubsw xmm2, xmm7 + pmaddubsw xmm1, xmm6 // V + pmaddubsw xmm3, xmm6 + phaddw xmm0, xmm2 + phaddw xmm1, xmm3 + psraw xmm0, 8 + psraw xmm1, 8 + packsswb xmm0, xmm1 + paddb xmm0, xmm5 // -> unsigned + + // step 3 - store 8 U and 8 V values + movlps qword ptr [edx], xmm0 // U + movhps qword ptr [edx + edi], xmm0 // V + lea edx, [edx + 8] + sub ecx, 16 + jg convertloop + + pop edi + ret + } +} + +__declspec(naked) +void BGRAToUVRow_SSSE3(const uint8* src_argb0, int src_stride_argb, + uint8* dst_u, uint8* dst_v, int width) { + __asm { + push esi + push edi + mov eax, [esp + 8 + 4] // src_argb + mov esi, [esp + 8 + 8] // src_stride_argb + mov edx, [esp + 8 + 12] // dst_u + mov edi, [esp + 8 + 16] // dst_v + mov ecx, [esp + 8 + 20] // pix + movdqa xmm5, kAddUV128 + movdqa xmm6, kBGRAToV + movdqa xmm7, kBGRAToU + sub edi, edx // stride from u to v + + convertloop: + /* step 1 - subsample 16x2 argb pixels to 8x1 */ + movdqu xmm0, [eax] + movdqu xmm4, [eax + esi] + pavgb xmm0, xmm4 + movdqu xmm1, [eax + 16] + movdqu xmm4, [eax + esi + 16] + pavgb xmm1, xmm4 + movdqu xmm2, [eax + 32] + movdqu xmm4, [eax + esi + 32] + pavgb xmm2, xmm4 + movdqu xmm3, [eax + 48] + movdqu xmm4, [eax + esi + 48] + pavgb xmm3, xmm4 + + lea eax, [eax + 64] + movdqa xmm4, xmm0 + shufps xmm0, xmm1, 0x88 + shufps xmm4, xmm1, 0xdd + pavgb xmm0, xmm4 + movdqa xmm4, xmm2 + shufps xmm2, xmm3, 0x88 + shufps xmm4, xmm3, 0xdd + pavgb xmm2, xmm4 + + // step 2 - convert to U and V + // from here down is very similar to Y code except + // instead of 16 different pixels, its 8 pixels of U and 8 of V + movdqa xmm1, xmm0 + movdqa xmm3, xmm2 + pmaddubsw xmm0, xmm7 // U + pmaddubsw xmm2, xmm7 + pmaddubsw xmm1, xmm6 // V + pmaddubsw xmm3, xmm6 + phaddw xmm0, xmm2 + phaddw xmm1, xmm3 + psraw xmm0, 8 + psraw xmm1, 8 + packsswb xmm0, xmm1 + paddb xmm0, xmm5 // -> unsigned + + // step 3 - store 8 U and 8 V values + movlps qword ptr [edx], xmm0 // U + movhps qword ptr [edx + edi], xmm0 // V + lea edx, [edx + 8] + sub ecx, 16 + jg convertloop + + pop edi + pop esi + ret + } +} + +__declspec(naked) +void ABGRToUVRow_SSSE3(const uint8* src_argb0, int src_stride_argb, + uint8* dst_u, uint8* dst_v, int width) { + __asm { + push esi + push edi + mov eax, [esp + 8 + 4] // src_argb + mov esi, [esp + 8 + 8] // src_stride_argb + mov edx, [esp + 8 + 12] // dst_u + mov edi, [esp + 8 + 16] // dst_v + mov ecx, [esp + 8 + 20] // pix + movdqa xmm5, kAddUV128 + movdqa xmm6, kABGRToV + movdqa xmm7, kABGRToU + sub edi, edx // stride from u to v + + convertloop: + /* step 1 - subsample 16x2 argb pixels to 8x1 */ + movdqu xmm0, [eax] + movdqu xmm4, [eax + esi] + pavgb xmm0, xmm4 + movdqu xmm1, [eax + 16] + movdqu xmm4, [eax + esi + 16] + pavgb xmm1, xmm4 + movdqu xmm2, [eax + 32] + movdqu xmm4, [eax + esi + 32] + pavgb xmm2, xmm4 + movdqu xmm3, [eax + 48] + movdqu xmm4, [eax + esi + 48] + pavgb xmm3, xmm4 + + lea eax, [eax + 64] + movdqa xmm4, xmm0 + shufps xmm0, xmm1, 0x88 + shufps xmm4, xmm1, 0xdd + pavgb xmm0, xmm4 + movdqa xmm4, xmm2 + shufps xmm2, xmm3, 0x88 + shufps xmm4, xmm3, 0xdd + pavgb xmm2, xmm4 + + // step 2 - convert to U and V + // from here down is very similar to Y code except + // instead of 16 different pixels, its 8 pixels of U and 8 of V + movdqa xmm1, xmm0 + movdqa xmm3, xmm2 + pmaddubsw xmm0, xmm7 // U + pmaddubsw xmm2, xmm7 + pmaddubsw xmm1, xmm6 // V + pmaddubsw xmm3, xmm6 + phaddw xmm0, xmm2 + phaddw xmm1, xmm3 + psraw xmm0, 8 + psraw xmm1, 8 + packsswb xmm0, xmm1 + paddb xmm0, xmm5 // -> unsigned + + // step 3 - store 8 U and 8 V values + movlps qword ptr [edx], xmm0 // U + movhps qword ptr [edx + edi], xmm0 // V + lea edx, [edx + 8] + sub ecx, 16 + jg convertloop + + pop edi + pop esi + ret + } +} + +__declspec(naked) +void RGBAToUVRow_SSSE3(const uint8* src_argb0, int src_stride_argb, + uint8* dst_u, uint8* dst_v, int width) { + __asm { + push esi + push edi + mov eax, [esp + 8 + 4] // src_argb + mov esi, [esp + 8 + 8] // src_stride_argb + mov edx, [esp + 8 + 12] // dst_u + mov edi, [esp + 8 + 16] // dst_v + mov ecx, [esp + 8 + 20] // pix + movdqa xmm5, kAddUV128 + movdqa xmm6, kRGBAToV + movdqa xmm7, kRGBAToU + sub edi, edx // stride from u to v + + convertloop: + /* step 1 - subsample 16x2 argb pixels to 8x1 */ + movdqu xmm0, [eax] + movdqu xmm4, [eax + esi] + pavgb xmm0, xmm4 + movdqu xmm1, [eax + 16] + movdqu xmm4, [eax + esi + 16] + pavgb xmm1, xmm4 + movdqu xmm2, [eax + 32] + movdqu xmm4, [eax + esi + 32] + pavgb xmm2, xmm4 + movdqu xmm3, [eax + 48] + movdqu xmm4, [eax + esi + 48] + pavgb xmm3, xmm4 + + lea eax, [eax + 64] + movdqa xmm4, xmm0 + shufps xmm0, xmm1, 0x88 + shufps xmm4, xmm1, 0xdd + pavgb xmm0, xmm4 + movdqa xmm4, xmm2 + shufps xmm2, xmm3, 0x88 + shufps xmm4, xmm3, 0xdd + pavgb xmm2, xmm4 + + // step 2 - convert to U and V + // from here down is very similar to Y code except + // instead of 16 different pixels, its 8 pixels of U and 8 of V + movdqa xmm1, xmm0 + movdqa xmm3, xmm2 + pmaddubsw xmm0, xmm7 // U + pmaddubsw xmm2, xmm7 + pmaddubsw xmm1, xmm6 // V + pmaddubsw xmm3, xmm6 + phaddw xmm0, xmm2 + phaddw xmm1, xmm3 + psraw xmm0, 8 + psraw xmm1, 8 + packsswb xmm0, xmm1 + paddb xmm0, xmm5 // -> unsigned + + // step 3 - store 8 U and 8 V values + movlps qword ptr [edx], xmm0 // U + movhps qword ptr [edx + edi], xmm0 // V + lea edx, [edx + 8] + sub ecx, 16 + jg convertloop + + pop edi + pop esi + ret + } +} +#endif // HAS_ARGBTOYROW_SSSE3 + +// Read 16 UV from 444 +#define READYUV444_AVX2 __asm { \ + __asm vmovdqu xmm0, [esi] /* U */ /* NOLINT */ \ + __asm vmovdqu xmm1, [esi + edi] /* V */ /* NOLINT */ \ + __asm lea esi, [esi + 16] \ + __asm vpermq ymm0, ymm0, 0xd8 \ + __asm vpermq ymm1, ymm1, 0xd8 \ + __asm vpunpcklbw ymm0, ymm0, ymm1 /* UV */ \ + } + +// Read 8 UV from 422, upsample to 16 UV. +#define READYUV422_AVX2 __asm { \ + __asm vmovq xmm0, qword ptr [esi] /* U */ /* NOLINT */ \ + __asm vmovq xmm1, qword ptr [esi + edi] /* V */ /* NOLINT */ \ + __asm lea esi, [esi + 8] \ + __asm vpunpcklbw ymm0, ymm0, ymm1 /* UV */ \ + __asm vpermq ymm0, ymm0, 0xd8 \ + __asm vpunpcklwd ymm0, ymm0, ymm0 /* UVUV (upsample) */ \ + } + +// Read 4 UV from 411, upsample to 16 UV. +#define READYUV411_AVX2 __asm { \ + __asm vmovd xmm0, dword ptr [esi] /* U */ /* NOLINT */ \ + __asm vmovd xmm1, dword ptr [esi + edi] /* V */ /* NOLINT */ \ + __asm lea esi, [esi + 4] \ + __asm vpunpcklbw ymm0, ymm0, ymm1 /* UV */ \ + __asm vpunpcklwd ymm0, ymm0, ymm0 /* UVUV (upsample) */ \ + __asm vpermq ymm0, ymm0, 0xd8 \ + __asm vpunpckldq ymm0, ymm0, ymm0 /* UVUVUVUV (upsample) */ \ + } + +// Read 8 UV from NV12, upsample to 16 UV. +#define READNV12_AVX2 __asm { \ + __asm vmovdqu xmm0, [esi] /* UV */ \ + __asm lea esi, [esi + 16] \ + __asm vpermq ymm0, ymm0, 0xd8 \ + __asm vpunpcklwd ymm0, ymm0, ymm0 /* UVUV (upsample) */ \ + } + +// Convert 16 pixels: 16 UV and 16 Y. +#define YUVTORGB_AVX2(YuvConstants) __asm { \ + /* Step 1: Find 8 UV contributions to 16 R,G,B values */ \ + __asm vpmaddubsw ymm2, ymm0, YuvConstants.kUVToR /* scale R UV */ \ + __asm vpmaddubsw ymm1, ymm0, YuvConstants.kUVToG /* scale G UV */ \ + __asm vpmaddubsw ymm0, ymm0, YuvConstants.kUVToB /* scale B UV */ \ + __asm vmovdqu ymm3, YuvConstants.kUVBiasR \ + __asm vpsubw ymm2, ymm3, ymm2 \ + __asm vmovdqu ymm3, YuvConstants.kUVBiasG \ + __asm vpsubw ymm1, ymm3, ymm1 \ + __asm vmovdqu ymm3, YuvConstants.kUVBiasB \ + __asm vpsubw ymm0, ymm3, ymm0 \ + /* Step 2: Find Y contribution to 16 R,G,B values */ \ + __asm vmovdqu xmm3, [eax] /* NOLINT */ \ + __asm lea eax, [eax + 16] \ + __asm vpermq ymm3, ymm3, 0xd8 \ + __asm vpunpcklbw ymm3, ymm3, ymm3 \ + __asm vpmulhuw ymm3, ymm3, YuvConstants.kYToRgb \ + __asm vpaddsw ymm0, ymm0, ymm3 /* B += Y */ \ + __asm vpaddsw ymm1, ymm1, ymm3 /* G += Y */ \ + __asm vpaddsw ymm2, ymm2, ymm3 /* R += Y */ \ + __asm vpsraw ymm0, ymm0, 6 \ + __asm vpsraw ymm1, ymm1, 6 \ + __asm vpsraw ymm2, ymm2, 6 \ + __asm vpackuswb ymm0, ymm0, ymm0 /* B */ \ + __asm vpackuswb ymm1, ymm1, ymm1 /* G */ \ + __asm vpackuswb ymm2, ymm2, ymm2 /* R */ \ + } + +// Store 16 ARGB values. +#define STOREARGB_AVX2 __asm { \ + /* Step 3: Weave into ARGB */ \ + __asm vpunpcklbw ymm0, ymm0, ymm1 /* BG */ \ + __asm vpermq ymm0, ymm0, 0xd8 \ + __asm vpunpcklbw ymm2, ymm2, ymm5 /* RA */ \ + __asm vpermq ymm2, ymm2, 0xd8 \ + __asm vpunpcklwd ymm1, ymm0, ymm2 /* BGRA first 8 pixels */ \ + __asm vpunpckhwd ymm0, ymm0, ymm2 /* BGRA next 8 pixels */ \ + __asm vmovdqu 0[edx], ymm1 \ + __asm vmovdqu 32[edx], ymm0 \ + __asm lea edx, [edx + 64] \ + } + +#ifdef HAS_I422TOARGBROW_AVX2 +// 16 pixels +// 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 ARGB (64 bytes). +__declspec(naked) +void I422ToARGBRow_AVX2(const uint8* y_buf, + const uint8* u_buf, + const uint8* v_buf, + uint8* dst_argb, + int width) { + __asm { + push esi + push edi + mov eax, [esp + 8 + 4] // Y + mov esi, [esp + 8 + 8] // U + mov edi, [esp + 8 + 12] // V + mov edx, [esp + 8 + 16] // argb + mov ecx, [esp + 8 + 20] // width + sub edi, esi + vpcmpeqb ymm5, ymm5, ymm5 // generate 0xffffffffffffffff for alpha + + convertloop: + READYUV422_AVX2 + YUVTORGB_AVX2(kYuvConstants) + STOREARGB_AVX2 + + sub ecx, 16 + jg convertloop + + pop edi + pop esi + vzeroupper + ret + } +} +#endif // HAS_I422TOARGBROW_AVX2 + +#ifdef HAS_J422TOARGBROW_AVX2 +// 16 pixels +// 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 ARGB (64 bytes). +__declspec(naked) +void J422ToARGBRow_AVX2(const uint8* y_buf, + const uint8* u_buf, + const uint8* v_buf, + uint8* dst_argb, + int width) { + __asm { + push esi + push edi + mov eax, [esp + 8 + 4] // Y + mov esi, [esp + 8 + 8] // U + mov edi, [esp + 8 + 12] // V + mov edx, [esp + 8 + 16] // argb + mov ecx, [esp + 8 + 20] // width + sub edi, esi + vpcmpeqb ymm5, ymm5, ymm5 // generate 0xffffffffffffffff for alpha + + convertloop: + READYUV422_AVX2 + YUVTORGB_AVX2(kYuvJConstants) + STOREARGB_AVX2 + + sub ecx, 16 + jg convertloop + + pop edi + pop esi + vzeroupper + ret + } +} +#endif // HAS_J422TOARGBROW_AVX2 + +#ifdef HAS_I444TOARGBROW_AVX2 +// 16 pixels +// 16 UV values with 16 Y producing 16 ARGB (64 bytes). +__declspec(naked) +void I444ToARGBRow_AVX2(const uint8* y_buf, + const uint8* u_buf, + const uint8* v_buf, + uint8* dst_argb, + int width) { + __asm { + push esi + push edi + mov eax, [esp + 8 + 4] // Y + mov esi, [esp + 8 + 8] // U + mov edi, [esp + 8 + 12] // V + mov edx, [esp + 8 + 16] // argb + mov ecx, [esp + 8 + 20] // width + sub edi, esi + vpcmpeqb ymm5, ymm5, ymm5 // generate 0xffffffffffffffff for alpha + + convertloop: + READYUV444_AVX2 + YUVTORGB_AVX2(kYuvConstants) + STOREARGB_AVX2 + + sub ecx, 16 + jg convertloop + + pop edi + pop esi + vzeroupper + ret + } +} +#endif // HAS_I444TOARGBROW_AVX2 + +#ifdef HAS_I411TOARGBROW_AVX2 +// 16 pixels +// 4 UV values upsampled to 16 UV, mixed with 16 Y producing 16 ARGB (64 bytes). +__declspec(naked) +void I411ToARGBRow_AVX2(const uint8* y_buf, + const uint8* u_buf, + const uint8* v_buf, + uint8* dst_argb, + int width) { + __asm { + push esi + push edi + mov eax, [esp + 8 + 4] // Y + mov esi, [esp + 8 + 8] // U + mov edi, [esp + 8 + 12] // V + mov edx, [esp + 8 + 16] // argb + mov ecx, [esp + 8 + 20] // width + sub edi, esi + vpcmpeqb ymm5, ymm5, ymm5 // generate 0xffffffffffffffff for alpha + + convertloop: + READYUV411_AVX2 + YUVTORGB_AVX2(kYuvConstants) + STOREARGB_AVX2 + + sub ecx, 16 + jg convertloop + + pop edi + pop esi + vzeroupper + ret + } +} +#endif // HAS_I411TOARGBROW_AVX2 + +#ifdef HAS_NV12TOARGBROW_AVX2 +// 16 pixels. +// 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 ARGB (64 bytes). +__declspec(naked) +void NV12ToARGBRow_AVX2(const uint8* y_buf, + const uint8* uv_buf, + uint8* dst_argb, + int width) { + __asm { + push esi + mov eax, [esp + 4 + 4] // Y + mov esi, [esp + 4 + 8] // UV + mov edx, [esp + 4 + 12] // argb + mov ecx, [esp + 4 + 16] // width + vpcmpeqb ymm5, ymm5, ymm5 // generate 0xffffffffffffffff for alpha + + convertloop: + READNV12_AVX2 + YUVTORGB_AVX2(kYuvConstants) + STOREARGB_AVX2 + + sub ecx, 16 + jg convertloop + + pop esi + vzeroupper + ret + } +} +#endif // HAS_NV12TOARGBROW_AVX2 + +#ifdef HAS_NV21TOARGBROW_AVX2 +// 16 pixels. +// 8 VU values upsampled to 16 VU, mixed with 16 Y producing 16 ARGB (64 bytes). +__declspec(naked) +void NV21ToARGBRow_AVX2(const uint8* y_buf, + const uint8* uv_buf, + uint8* dst_argb, + int width) { + __asm { + push esi + mov eax, [esp + 4 + 4] // Y + mov esi, [esp + 4 + 8] // UV + mov edx, [esp + 4 + 12] // argb + mov ecx, [esp + 4 + 16] // width + vpcmpeqb ymm5, ymm5, ymm5 // generate 0xffffffffffffffff for alpha + + convertloop: + READNV12_AVX2 + YUVTORGB_AVX2(kYvuConstants) + STOREARGB_AVX2 + + sub ecx, 16 + jg convertloop + + pop esi + vzeroupper + ret + } +} +#endif // HAS_NV21TOARGBROW_AVX2 + +#ifdef HAS_I422TOBGRAROW_AVX2 +// 16 pixels +// 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 BGRA (64 bytes). +// TODO(fbarchard): Use macros to reduce duplicate code. See SSSE3. +__declspec(naked) +void I422ToBGRARow_AVX2(const uint8* y_buf, + const uint8* u_buf, + const uint8* v_buf, + uint8* dst_argb, + int width) { + __asm { + push esi + push edi + mov eax, [esp + 8 + 4] // Y + mov esi, [esp + 8 + 8] // U + mov edi, [esp + 8 + 12] // V + mov edx, [esp + 8 + 16] // argb + mov ecx, [esp + 8 + 20] // width + sub edi, esi + vpcmpeqb ymm5, ymm5, ymm5 // generate 0xffffffffffffffff for alpha + + convertloop: + READYUV422_AVX2 + YUVTORGB_AVX2(kYuvConstants) + + // Step 3: Weave into BGRA + vpunpcklbw ymm1, ymm1, ymm0 // GB + vpermq ymm1, ymm1, 0xd8 + vpunpcklbw ymm2, ymm5, ymm2 // AR + vpermq ymm2, ymm2, 0xd8 + vpunpcklwd ymm0, ymm2, ymm1 // ARGB first 8 pixels + vpunpckhwd ymm2, ymm2, ymm1 // ARGB next 8 pixels + vmovdqu [edx], ymm0 + vmovdqu [edx + 32], ymm2 + lea edx, [edx + 64] + sub ecx, 16 + jg convertloop + + pop edi + pop esi + vzeroupper + ret + } +} +#endif // HAS_I422TOBGRAROW_AVX2 + +#ifdef HAS_I422TORGBAROW_AVX2 +// 16 pixels +// 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 RGBA (64 bytes). +// TODO(fbarchard): Use macros to reduce duplicate code. See SSSE3. +__declspec(naked) +void I422ToRGBARow_AVX2(const uint8* y_buf, + const uint8* u_buf, + const uint8* v_buf, + uint8* dst_argb, + int width) { + __asm { + push esi + push edi + mov eax, [esp + 8 + 4] // Y + mov esi, [esp + 8 + 8] // U + mov edi, [esp + 8 + 12] // V + mov edx, [esp + 8 + 16] // argb + mov ecx, [esp + 8 + 20] // width + sub edi, esi + vpcmpeqb ymm5, ymm5, ymm5 // generate 0xffffffffffffffff for alpha + + convertloop: + READYUV422_AVX2 + YUVTORGB_AVX2(kYuvConstants) + + // Step 3: Weave into RGBA + vpunpcklbw ymm1, ymm1, ymm2 // GR + vpermq ymm1, ymm1, 0xd8 + vpunpcklbw ymm2, ymm5, ymm0 // AB + vpermq ymm2, ymm2, 0xd8 + vpunpcklwd ymm0, ymm2, ymm1 // ABGR first 8 pixels + vpunpckhwd ymm1, ymm2, ymm1 // ABGR next 8 pixels + vmovdqu [edx], ymm0 + vmovdqu [edx + 32], ymm1 + lea edx, [edx + 64] + sub ecx, 16 + jg convertloop + + pop edi + pop esi + vzeroupper + ret + } +} +#endif // HAS_I422TORGBAROW_AVX2 + +#ifdef HAS_I422TOABGRROW_AVX2 +// 16 pixels +// 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 ABGR (64 bytes). +// TODO(fbarchard): Use macros to reduce duplicate code. See SSSE3. +__declspec(naked) +void I422ToABGRRow_AVX2(const uint8* y_buf, + const uint8* u_buf, + const uint8* v_buf, + uint8* dst_argb, + int width) { + __asm { + push esi + push edi + mov eax, [esp + 8 + 4] // Y + mov esi, [esp + 8 + 8] // U + mov edi, [esp + 8 + 12] // V + mov edx, [esp + 8 + 16] // argb + mov ecx, [esp + 8 + 20] // width + sub edi, esi + vpcmpeqb ymm5, ymm5, ymm5 // generate 0xffffffffffffffff for alpha + + convertloop: + READYUV422_AVX2 + YUVTORGB_AVX2(kYuvConstants) + + // Step 3: Weave into ABGR + vpunpcklbw ymm1, ymm2, ymm1 // RG + vpermq ymm1, ymm1, 0xd8 + vpunpcklbw ymm2, ymm0, ymm5 // BA + vpermq ymm2, ymm2, 0xd8 + vpunpcklwd ymm0, ymm1, ymm2 // RGBA first 8 pixels + vpunpckhwd ymm1, ymm1, ymm2 // RGBA next 8 pixels + vmovdqu [edx], ymm0 + vmovdqu [edx + 32], ymm1 + lea edx, [edx + 64] + sub ecx, 16 + jg convertloop + + pop edi + pop esi + vzeroupper + ret + } +} +#endif // HAS_I422TOABGRROW_AVX2 + +#if defined(HAS_I422TOARGBROW_SSSE3) +// TODO(fbarchard): Read that does half size on Y and treats 420 as 444. + +// Read 8 UV from 444. +#define READYUV444 __asm { \ + __asm movq xmm0, qword ptr [esi] /* U */ /* NOLINT */ \ + __asm movq xmm1, qword ptr [esi + edi] /* V */ /* NOLINT */ \ + __asm lea esi, [esi + 8] \ + __asm punpcklbw xmm0, xmm1 /* UV */ \ + } + +// Read 4 UV from 422, upsample to 8 UV. +#define READYUV422 __asm { \ + __asm movd xmm0, [esi] /* U */ \ + __asm movd xmm1, [esi + edi] /* V */ \ + __asm lea esi, [esi + 4] \ + __asm punpcklbw xmm0, xmm1 /* UV */ \ + __asm punpcklwd xmm0, xmm0 /* UVUV (upsample) */ \ + } + +// Read 2 UV from 411, upsample to 8 UV. +#define READYUV411 __asm { \ + __asm movzx ebx, word ptr [esi] /* U */ /* NOLINT */ \ + __asm movd xmm0, ebx \ + __asm movzx ebx, word ptr [esi + edi] /* V */ /* NOLINT */ \ + __asm movd xmm1, ebx \ + __asm lea esi, [esi + 2] \ + __asm punpcklbw xmm0, xmm1 /* UV */ \ + __asm punpcklwd xmm0, xmm0 /* UVUV (upsample) */ \ + __asm punpckldq xmm0, xmm0 /* UVUVUVUV (upsample) */ \ + } + +// Read 4 UV from NV12, upsample to 8 UV. +#define READNV12 __asm { \ + __asm movq xmm0, qword ptr [esi] /* UV */ /* NOLINT */ \ + __asm lea esi, [esi + 8] \ + __asm punpcklwd xmm0, xmm0 /* UVUV (upsample) */ \ + } + +// Convert 8 pixels: 8 UV and 8 Y. +#define YUVTORGB(YuvConstants) __asm { \ + /* Step 1: Find 4 UV contributions to 8 R,G,B values */ \ + __asm movdqa xmm1, xmm0 \ + __asm movdqa xmm2, xmm0 \ + __asm movdqa xmm3, xmm0 \ + __asm movdqa xmm0, YuvConstants.kUVBiasB /* unbias back to signed */ \ + __asm pmaddubsw xmm1, YuvConstants.kUVToB /* scale B UV */ \ + __asm psubw xmm0, xmm1 \ + __asm movdqa xmm1, YuvConstants.kUVBiasG \ + __asm pmaddubsw xmm2, YuvConstants.kUVToG /* scale G UV */ \ + __asm psubw xmm1, xmm2 \ + __asm movdqa xmm2, YuvConstants.kUVBiasR \ + __asm pmaddubsw xmm3, YuvConstants.kUVToR /* scale R UV */ \ + __asm psubw xmm2, xmm3 \ + /* Step 2: Find Y contribution to 8 R,G,B values */ \ + __asm movq xmm3, qword ptr [eax] /* NOLINT */ \ + __asm lea eax, [eax + 8] \ + __asm punpcklbw xmm3, xmm3 \ + __asm pmulhuw xmm3, YuvConstants.kYToRgb \ + __asm paddsw xmm0, xmm3 /* B += Y */ \ + __asm paddsw xmm1, xmm3 /* G += Y */ \ + __asm paddsw xmm2, xmm3 /* R += Y */ \ + __asm psraw xmm0, 6 \ + __asm psraw xmm1, 6 \ + __asm psraw xmm2, 6 \ + __asm packuswb xmm0, xmm0 /* B */ \ + __asm packuswb xmm1, xmm1 /* G */ \ + __asm packuswb xmm2, xmm2 /* R */ \ + } + +// Store 8 ARGB values. +#define STOREARGB __asm { \ + /* Step 3: Weave into ARGB */ \ + __asm punpcklbw xmm0, xmm1 /* BG */ \ + __asm punpcklbw xmm2, xmm5 /* RA */ \ + __asm movdqa xmm1, xmm0 \ + __asm punpcklwd xmm0, xmm2 /* BGRA first 4 pixels */ \ + __asm punpckhwd xmm1, xmm2 /* BGRA next 4 pixels */ \ + __asm movdqu 0[edx], xmm0 \ + __asm movdqu 16[edx], xmm1 \ + __asm lea edx, [edx + 32] \ + } + +// Store 8 BGRA values. +#define STOREBGRA __asm { \ + /* Step 3: Weave into BGRA */ \ + __asm pcmpeqb xmm5, xmm5 /* generate 0xffffffff for alpha */ \ + __asm punpcklbw xmm1, xmm0 /* GB */ \ + __asm punpcklbw xmm5, xmm2 /* AR */ \ + __asm movdqa xmm0, xmm5 \ + __asm punpcklwd xmm5, xmm1 /* BGRA first 4 pixels */ \ + __asm punpckhwd xmm0, xmm1 /* BGRA next 4 pixels */ \ + __asm movdqu 0[edx], xmm5 \ + __asm movdqu 16[edx], xmm0 \ + __asm lea edx, [edx + 32] \ + } + +// Store 8 ABGR values. +#define STOREABGR __asm { \ + /* Step 3: Weave into ABGR */ \ + __asm punpcklbw xmm2, xmm1 /* RG */ \ + __asm punpcklbw xmm0, xmm5 /* BA */ \ + __asm movdqa xmm1, xmm2 \ + __asm punpcklwd xmm2, xmm0 /* RGBA first 4 pixels */ \ + __asm punpckhwd xmm1, xmm0 /* RGBA next 4 pixels */ \ + __asm movdqu 0[edx], xmm2 \ + __asm movdqu 16[edx], xmm1 \ + __asm lea edx, [edx + 32] \ + } + +// Store 8 RGBA values. +#define STORERGBA __asm { \ + /* Step 3: Weave into RGBA */ \ + __asm pcmpeqb xmm5, xmm5 /* generate 0xffffffff for alpha */ \ + __asm punpcklbw xmm1, xmm2 /* GR */ \ + __asm punpcklbw xmm5, xmm0 /* AB */ \ + __asm movdqa xmm0, xmm5 \ + __asm punpcklwd xmm5, xmm1 /* RGBA first 4 pixels */ \ + __asm punpckhwd xmm0, xmm1 /* RGBA next 4 pixels */ \ + __asm movdqu 0[edx], xmm5 \ + __asm movdqu 16[edx], xmm0 \ + __asm lea edx, [edx + 32] \ + } + +// Store 8 RGB24 values. +#define STORERGB24 __asm { \ + /* Step 3: Weave into RRGB */ \ + __asm punpcklbw xmm0, xmm1 /* BG */ \ + __asm punpcklbw xmm2, xmm2 /* RR */ \ + __asm movdqa xmm1, xmm0 \ + __asm punpcklwd xmm0, xmm2 /* BGRR first 4 pixels */ \ + __asm punpckhwd xmm1, xmm2 /* BGRR next 4 pixels */ \ + /* Step 4: RRGB -> RGB24 */ \ + __asm pshufb xmm0, xmm5 /* Pack first 8 and last 4 bytes. */ \ + __asm pshufb xmm1, xmm6 /* Pack first 12 bytes. */ \ + __asm palignr xmm1, xmm0, 12 /* last 4 bytes of xmm0 + 12 xmm1 */ \ + __asm movq qword ptr 0[edx], xmm0 /* First 8 bytes */ \ + __asm movdqu 8[edx], xmm1 /* Last 16 bytes */ \ + __asm lea edx, [edx + 24] \ + } + +// Store 8 RAW values. +#define STORERAW __asm { \ + /* Step 3: Weave into RRGB */ \ + __asm punpcklbw xmm0, xmm1 /* BG */ \ + __asm punpcklbw xmm2, xmm2 /* RR */ \ + __asm movdqa xmm1, xmm0 \ + __asm punpcklwd xmm0, xmm2 /* BGRR first 4 pixels */ \ + __asm punpckhwd xmm1, xmm2 /* BGRR next 4 pixels */ \ + /* Step 4: RRGB -> RAW */ \ + __asm pshufb xmm0, xmm5 /* Pack first 8 and last 4 bytes. */ \ + __asm pshufb xmm1, xmm6 /* Pack first 12 bytes. */ \ + __asm palignr xmm1, xmm0, 12 /* last 4 bytes of xmm0 + 12 xmm1 */ \ + __asm movq qword ptr 0[edx], xmm0 /* First 8 bytes */ \ + __asm movdqu 8[edx], xmm1 /* Last 16 bytes */ \ + __asm lea edx, [edx + 24] \ + } + +// Store 8 RGB565 values. +#define STORERGB565 __asm { \ + /* Step 3: Weave into RRGB */ \ + __asm punpcklbw xmm0, xmm1 /* BG */ \ + __asm punpcklbw xmm2, xmm2 /* RR */ \ + __asm movdqa xmm1, xmm0 \ + __asm punpcklwd xmm0, xmm2 /* BGRR first 4 pixels */ \ + __asm punpckhwd xmm1, xmm2 /* BGRR next 4 pixels */ \ + /* Step 4: RRGB -> RGB565 */ \ + __asm movdqa xmm3, xmm0 /* B first 4 pixels of argb */ \ + __asm movdqa xmm2, xmm0 /* G */ \ + __asm pslld xmm0, 8 /* R */ \ + __asm psrld xmm3, 3 /* B */ \ + __asm psrld xmm2, 5 /* G */ \ + __asm psrad xmm0, 16 /* R */ \ + __asm pand xmm3, xmm5 /* B */ \ + __asm pand xmm2, xmm6 /* G */ \ + __asm pand xmm0, xmm7 /* R */ \ + __asm por xmm3, xmm2 /* BG */ \ + __asm por xmm0, xmm3 /* BGR */ \ + __asm movdqa xmm3, xmm1 /* B next 4 pixels of argb */ \ + __asm movdqa xmm2, xmm1 /* G */ \ + __asm pslld xmm1, 8 /* R */ \ + __asm psrld xmm3, 3 /* B */ \ + __asm psrld xmm2, 5 /* G */ \ + __asm psrad xmm1, 16 /* R */ \ + __asm pand xmm3, xmm5 /* B */ \ + __asm pand xmm2, xmm6 /* G */ \ + __asm pand xmm1, xmm7 /* R */ \ + __asm por xmm3, xmm2 /* BG */ \ + __asm por xmm1, xmm3 /* BGR */ \ + __asm packssdw xmm0, xmm1 \ + __asm movdqu 0[edx], xmm0 /* store 8 pixels of RGB565 */ \ + __asm lea edx, [edx + 16] \ + } + +// 8 pixels. +// 8 UV values, mixed with 8 Y producing 8 ARGB (32 bytes). +__declspec(naked) +void I444ToARGBRow_SSSE3(const uint8* y_buf, + const uint8* u_buf, + const uint8* v_buf, + uint8* dst_argb, + int width) { + __asm { + push esi + push edi + mov eax, [esp + 8 + 4] // Y + mov esi, [esp + 8 + 8] // U + mov edi, [esp + 8 + 12] // V + mov edx, [esp + 8 + 16] // argb + mov ecx, [esp + 8 + 20] // width + sub edi, esi + pcmpeqb xmm5, xmm5 // generate 0xffffffff for alpha + + convertloop: + READYUV444 + YUVTORGB(kYuvConstants) + STOREARGB + + sub ecx, 8 + jg convertloop + + pop edi + pop esi + ret + } +} + +// 8 pixels. +// 4 UV values upsampled to 8 UV, mixed with 8 Y producing 8 RGB24 (24 bytes). +__declspec(naked) +void I422ToRGB24Row_SSSE3(const uint8* y_buf, + const uint8* u_buf, + const uint8* v_buf, + uint8* dst_rgb24, + int width) { + __asm { + push esi + push edi + mov eax, [esp + 8 + 4] // Y + mov esi, [esp + 8 + 8] // U + mov edi, [esp + 8 + 12] // V + mov edx, [esp + 8 + 16] // rgb24 + mov ecx, [esp + 8 + 20] // width + sub edi, esi + movdqa xmm5, kShuffleMaskARGBToRGB24_0 + movdqa xmm6, kShuffleMaskARGBToRGB24 + + convertloop: + READYUV422 + YUVTORGB(kYuvConstants) + STORERGB24 + + sub ecx, 8 + jg convertloop + + pop edi + pop esi + ret + } +} + +// 8 pixels. +// 4 UV values upsampled to 8 UV, mixed with 8 Y producing 8 RAW (24 bytes). +__declspec(naked) +void I422ToRAWRow_SSSE3(const uint8* y_buf, + const uint8* u_buf, + const uint8* v_buf, + uint8* dst_raw, + int width) { + __asm { + push esi + push edi + mov eax, [esp + 8 + 4] // Y + mov esi, [esp + 8 + 8] // U + mov edi, [esp + 8 + 12] // V + mov edx, [esp + 8 + 16] // raw + mov ecx, [esp + 8 + 20] // width + sub edi, esi + movdqa xmm5, kShuffleMaskARGBToRAW_0 + movdqa xmm6, kShuffleMaskARGBToRAW + + convertloop: + READYUV422 + YUVTORGB(kYuvConstants) + STORERAW + + sub ecx, 8 + jg convertloop + + pop edi + pop esi + ret + } +} + +// 8 pixels +// 4 UV values upsampled to 8 UV, mixed with 8 Y producing 8 RGB565 (16 bytes). +__declspec(naked) +void I422ToRGB565Row_SSSE3(const uint8* y_buf, + const uint8* u_buf, + const uint8* v_buf, + uint8* rgb565_buf, + int width) { + __asm { + push esi + push edi + mov eax, [esp + 8 + 4] // Y + mov esi, [esp + 8 + 8] // U + mov edi, [esp + 8 + 12] // V + mov edx, [esp + 8 + 16] // rgb565 + mov ecx, [esp + 8 + 20] // width + sub edi, esi + pcmpeqb xmm5, xmm5 // generate mask 0x0000001f + psrld xmm5, 27 + pcmpeqb xmm6, xmm6 // generate mask 0x000007e0 + psrld xmm6, 26 + pslld xmm6, 5 + pcmpeqb xmm7, xmm7 // generate mask 0xfffff800 + pslld xmm7, 11 + + convertloop: + READYUV422 + YUVTORGB(kYuvConstants) + STORERGB565 + + sub ecx, 8 + jg convertloop + + pop edi + pop esi + ret + } +} + +// 8 pixels. +// 4 UV values upsampled to 8 UV, mixed with 8 Y producing 8 ARGB (32 bytes). +__declspec(naked) +void I422ToARGBRow_SSSE3(const uint8* y_buf, + const uint8* u_buf, + const uint8* v_buf, + uint8* dst_argb, + int width) { + __asm { + push esi + push edi + mov eax, [esp + 8 + 4] // Y + mov esi, [esp + 8 + 8] // U + mov edi, [esp + 8 + 12] // V + mov edx, [esp + 8 + 16] // argb + mov ecx, [esp + 8 + 20] // width + sub edi, esi + pcmpeqb xmm5, xmm5 // generate 0xffffffff for alpha + + convertloop: + READYUV422 + YUVTORGB(kYuvConstants) + STOREARGB + + sub ecx, 8 + jg convertloop + + pop edi + pop esi + ret + } +} + +// 8 pixels. +// JPeg color space version of I422ToARGB +// 4 UV values upsampled to 8 UV, mixed with 8 Y producing 8 ARGB (32 bytes). +__declspec(naked) +void J422ToARGBRow_SSSE3(const uint8* y_buf, + const uint8* u_buf, + const uint8* v_buf, + uint8* dst_argb, + int width) { + __asm { + push esi + push edi + mov eax, [esp + 8 + 4] // Y + mov esi, [esp + 8 + 8] // U + mov edi, [esp + 8 + 12] // V + mov edx, [esp + 8 + 16] // argb + mov ecx, [esp + 8 + 20] // width + sub edi, esi + pcmpeqb xmm5, xmm5 // generate 0xffffffff for alpha + + convertloop: + READYUV422 + YUVTORGB(kYuvJConstants) + STOREARGB + + sub ecx, 8 + jg convertloop + + pop edi + pop esi + ret + } +} + +// 8 pixels. +// 2 UV values upsampled to 8 UV, mixed with 8 Y producing 8 ARGB (32 bytes). +// Similar to I420 but duplicate UV once more. +__declspec(naked) +void I411ToARGBRow_SSSE3(const uint8* y_buf, + const uint8* u_buf, + const uint8* v_buf, + uint8* dst_argb, + int width) { + __asm { + push ebx + push esi + push edi + mov eax, [esp + 12 + 4] // Y + mov esi, [esp + 12 + 8] // U + mov edi, [esp + 12 + 12] // V + mov edx, [esp + 12 + 16] // argb + mov ecx, [esp + 12 + 20] // width + sub edi, esi + pcmpeqb xmm5, xmm5 // generate 0xffffffff for alpha + + convertloop: + READYUV411 // modifies EBX + YUVTORGB(kYuvConstants) + STOREARGB + + sub ecx, 8 + jg convertloop + + pop edi + pop esi + pop ebx + ret + } +} + +// 8 pixels. +// 4 UV values upsampled to 8 UV, mixed with 8 Y producing 8 ARGB (32 bytes). +__declspec(naked) +void NV12ToARGBRow_SSSE3(const uint8* y_buf, + const uint8* uv_buf, + uint8* dst_argb, + int width) { + __asm { + push esi + mov eax, [esp + 4 + 4] // Y + mov esi, [esp + 4 + 8] // UV + mov edx, [esp + 4 + 12] // argb + mov ecx, [esp + 4 + 16] // width + pcmpeqb xmm5, xmm5 // generate 0xffffffff for alpha + + convertloop: + READNV12 + YUVTORGB(kYuvConstants) + STOREARGB + + sub ecx, 8 + jg convertloop + + pop esi + ret + } +} + +// 8 pixels. +// 4 VU values upsampled to 8 VU, mixed with 8 Y producing 8 ARGB (32 bytes). +__declspec(naked) +void NV21ToARGBRow_SSSE3(const uint8* y_buf, + const uint8* uv_buf, + uint8* dst_argb, + int width) { + __asm { + push esi + mov eax, [esp + 4 + 4] // Y + mov esi, [esp + 4 + 8] // UV + mov edx, [esp + 4 + 12] // argb + mov ecx, [esp + 4 + 16] // width + pcmpeqb xmm5, xmm5 // generate 0xffffffff for alpha + + convertloop: + READNV12 + YUVTORGB(kYvuConstants) + STOREARGB + + sub ecx, 8 + jg convertloop + + pop esi + ret + } +} + +__declspec(naked) +void I422ToBGRARow_SSSE3(const uint8* y_buf, + const uint8* u_buf, + const uint8* v_buf, + uint8* dst_bgra, + int width) { + __asm { + push esi + push edi + mov eax, [esp + 8 + 4] // Y + mov esi, [esp + 8 + 8] // U + mov edi, [esp + 8 + 12] // V + mov edx, [esp + 8 + 16] // bgra + mov ecx, [esp + 8 + 20] // width + sub edi, esi + + convertloop: + READYUV422 + YUVTORGB(kYuvConstants) + STOREBGRA + + sub ecx, 8 + jg convertloop + + pop edi + pop esi + ret + } +} + +__declspec(naked) +void I422ToABGRRow_SSSE3(const uint8* y_buf, + const uint8* u_buf, + const uint8* v_buf, + uint8* dst_abgr, + int width) { + __asm { + push esi + push edi + mov eax, [esp + 8 + 4] // Y + mov esi, [esp + 8 + 8] // U + mov edi, [esp + 8 + 12] // V + mov edx, [esp + 8 + 16] // abgr + mov ecx, [esp + 8 + 20] // width + sub edi, esi + pcmpeqb xmm5, xmm5 // generate 0xffffffff for alpha + + convertloop: + READYUV422 + YUVTORGB(kYuvConstants) + STOREABGR + + sub ecx, 8 + jg convertloop + + pop edi + pop esi + ret + } +} + +__declspec(naked) +void I422ToRGBARow_SSSE3(const uint8* y_buf, + const uint8* u_buf, + const uint8* v_buf, + uint8* dst_rgba, + int width) { + __asm { + push esi + push edi + mov eax, [esp + 8 + 4] // Y + mov esi, [esp + 8 + 8] // U + mov edi, [esp + 8 + 12] // V + mov edx, [esp + 8 + 16] // rgba + mov ecx, [esp + 8 + 20] // width + sub edi, esi + + convertloop: + READYUV422 + YUVTORGB(kYuvConstants) + STORERGBA + + sub ecx, 8 + jg convertloop + + pop edi + pop esi + ret + } +} + +#endif // HAS_I422TOARGBROW_SSSE3 + +#ifdef HAS_I400TOARGBROW_SSE2 +// 8 pixels of Y converted to 8 pixels of ARGB (32 bytes). +__declspec(naked) +void I400ToARGBRow_SSE2(const uint8* y_buf, + uint8* rgb_buf, + int width) { + __asm { + mov eax, 0x4a354a35 // 4a35 = 18997 = round(1.164 * 64 * 256) + movd xmm2, eax + pshufd xmm2, xmm2,0 + mov eax, 0x04880488 // 0488 = 1160 = round(1.164 * 64 * 16) + movd xmm3, eax + pshufd xmm3, xmm3, 0 + pcmpeqb xmm4, xmm4 // generate mask 0xff000000 + pslld xmm4, 24 + + mov eax, [esp + 4] // Y + mov edx, [esp + 8] // rgb + mov ecx, [esp + 12] // width + + convertloop: + // Step 1: Scale Y contribution to 8 G values. G = (y - 16) * 1.164 + movq xmm0, qword ptr [eax] + lea eax, [eax + 8] + punpcklbw xmm0, xmm0 // Y.Y + pmulhuw xmm0, xmm2 + psubusw xmm0, xmm3 + psrlw xmm0, 6 + packuswb xmm0, xmm0 // G + + // Step 2: Weave into ARGB + punpcklbw xmm0, xmm0 // GG + movdqa xmm1, xmm0 + punpcklwd xmm0, xmm0 // BGRA first 4 pixels + punpckhwd xmm1, xmm1 // BGRA next 4 pixels + por xmm0, xmm4 + por xmm1, xmm4 + movdqu [edx], xmm0 + movdqu [edx + 16], xmm1 + lea edx, [edx + 32] + sub ecx, 8 + jg convertloop + ret + } +} +#endif // HAS_I400TOARGBROW_SSE2 + +#ifdef HAS_I400TOARGBROW_AVX2 +// 16 pixels of Y converted to 16 pixels of ARGB (64 bytes). +// note: vpunpcklbw mutates and vpackuswb unmutates. +__declspec(naked) +void I400ToARGBRow_AVX2(const uint8* y_buf, + uint8* rgb_buf, + int width) { + __asm { + mov eax, 0x4a354a35 // 4a35 = 18997 = round(1.164 * 64 * 256) + vmovd xmm2, eax + vbroadcastss ymm2, xmm2 + mov eax, 0x04880488 // 0488 = 1160 = round(1.164 * 64 * 16) + vmovd xmm3, eax + vbroadcastss ymm3, xmm3 + vpcmpeqb ymm4, ymm4, ymm4 // generate mask 0xff000000 + vpslld ymm4, ymm4, 24 + + mov eax, [esp + 4] // Y + mov edx, [esp + 8] // rgb + mov ecx, [esp + 12] // width + + convertloop: + // Step 1: Scale Y contriportbution to 16 G values. G = (y - 16) * 1.164 + vmovdqu xmm0, [eax] + lea eax, [eax + 16] + vpermq ymm0, ymm0, 0xd8 // vpunpcklbw mutates + vpunpcklbw ymm0, ymm0, ymm0 // Y.Y + vpmulhuw ymm0, ymm0, ymm2 + vpsubusw ymm0, ymm0, ymm3 + vpsrlw ymm0, ymm0, 6 + vpackuswb ymm0, ymm0, ymm0 // G. still mutated: 3120 + + // TODO(fbarchard): Weave alpha with unpack. + // Step 2: Weave into ARGB + vpunpcklbw ymm1, ymm0, ymm0 // GG - mutates + vpermq ymm1, ymm1, 0xd8 + vpunpcklwd ymm0, ymm1, ymm1 // GGGG first 8 pixels + vpunpckhwd ymm1, ymm1, ymm1 // GGGG next 8 pixels + vpor ymm0, ymm0, ymm4 + vpor ymm1, ymm1, ymm4 + vmovdqu [edx], ymm0 + vmovdqu [edx + 32], ymm1 + lea edx, [edx + 64] + sub ecx, 16 + jg convertloop + vzeroupper + ret + } +} +#endif // HAS_I400TOARGBROW_AVX2 + +#ifdef HAS_MIRRORROW_SSSE3 +// Shuffle table for reversing the bytes. +static const uvec8 kShuffleMirror = { + 15u, 14u, 13u, 12u, 11u, 10u, 9u, 8u, 7u, 6u, 5u, 4u, 3u, 2u, 1u, 0u +}; + +// TODO(fbarchard): Replace lea with -16 offset. +__declspec(naked) +void MirrorRow_SSSE3(const uint8* src, uint8* dst, int width) { + __asm { + mov eax, [esp + 4] // src + mov edx, [esp + 8] // dst + mov ecx, [esp + 12] // width + movdqa xmm5, kShuffleMirror + + convertloop: + movdqu xmm0, [eax - 16 + ecx] + pshufb xmm0, xmm5 + movdqu [edx], xmm0 + lea edx, [edx + 16] + sub ecx, 16 + jg convertloop + ret + } +} +#endif // HAS_MIRRORROW_SSSE3 + +#ifdef HAS_MIRRORROW_AVX2 +__declspec(naked) +void MirrorRow_AVX2(const uint8* src, uint8* dst, int width) { + __asm { + mov eax, [esp + 4] // src + mov edx, [esp + 8] // dst + mov ecx, [esp + 12] // width + vbroadcastf128 ymm5, kShuffleMirror + + convertloop: + vmovdqu ymm0, [eax - 32 + ecx] + vpshufb ymm0, ymm0, ymm5 + vpermq ymm0, ymm0, 0x4e // swap high and low halfs + vmovdqu [edx], ymm0 + lea edx, [edx + 32] + sub ecx, 32 + jg convertloop + vzeroupper + ret + } +} +#endif // HAS_MIRRORROW_AVX2 + +#ifdef HAS_MIRRORROW_SSE2 +__declspec(naked) +void MirrorRow_SSE2(const uint8* src, uint8* dst, int width) { + __asm { + mov eax, [esp + 4] // src + mov edx, [esp + 8] // dst + mov ecx, [esp + 12] // width + + convertloop: + movdqu xmm0, [eax - 16 + ecx] + movdqa xmm1, xmm0 // swap bytes + psllw xmm0, 8 + psrlw xmm1, 8 + por xmm0, xmm1 + pshuflw xmm0, xmm0, 0x1b // swap words + pshufhw xmm0, xmm0, 0x1b + pshufd xmm0, xmm0, 0x4e // swap qwords + movdqu [edx], xmm0 + lea edx, [edx + 16] + sub ecx, 16 + jg convertloop + ret + } +} +#endif // HAS_MIRRORROW_SSE2 + +#ifdef HAS_MIRRORROW_UV_SSSE3 +// Shuffle table for reversing the bytes of UV channels. +static const uvec8 kShuffleMirrorUV = { + 14u, 12u, 10u, 8u, 6u, 4u, 2u, 0u, 15u, 13u, 11u, 9u, 7u, 5u, 3u, 1u +}; + +__declspec(naked) +void MirrorUVRow_SSSE3(const uint8* src, uint8* dst_u, uint8* dst_v, + int width) { + __asm { + push edi + mov eax, [esp + 4 + 4] // src + mov edx, [esp + 4 + 8] // dst_u + mov edi, [esp + 4 + 12] // dst_v + mov ecx, [esp + 4 + 16] // width + movdqa xmm1, kShuffleMirrorUV + lea eax, [eax + ecx * 2 - 16] + sub edi, edx + + convertloop: + movdqu xmm0, [eax] + lea eax, [eax - 16] + pshufb xmm0, xmm1 + movlpd qword ptr [edx], xmm0 + movhpd qword ptr [edx + edi], xmm0 + lea edx, [edx + 8] + sub ecx, 8 + jg convertloop + + pop edi + ret + } +} +#endif // HAS_MIRRORROW_UV_SSSE3 + +#ifdef HAS_ARGBMIRRORROW_SSE2 +__declspec(naked) +void ARGBMirrorRow_SSE2(const uint8* src, uint8* dst, int width) { + __asm { + mov eax, [esp + 4] // src + mov edx, [esp + 8] // dst + mov ecx, [esp + 12] // width + lea eax, [eax - 16 + ecx * 4] // last 4 pixels. + + convertloop: + movdqu xmm0, [eax] + lea eax, [eax - 16] + pshufd xmm0, xmm0, 0x1b + movdqu [edx], xmm0 + lea edx, [edx + 16] + sub ecx, 4 + jg convertloop + ret + } +} +#endif // HAS_ARGBMIRRORROW_SSE2 + +#ifdef HAS_ARGBMIRRORROW_AVX2 +// Shuffle table for reversing the bytes. +static const ulvec32 kARGBShuffleMirror_AVX2 = { + 7u, 6u, 5u, 4u, 3u, 2u, 1u, 0u +}; + +__declspec(naked) +void ARGBMirrorRow_AVX2(const uint8* src, uint8* dst, int width) { + __asm { + mov eax, [esp + 4] // src + mov edx, [esp + 8] // dst + mov ecx, [esp + 12] // width + vmovdqu ymm5, kARGBShuffleMirror_AVX2 + + convertloop: + vpermd ymm0, ymm5, [eax - 32 + ecx * 4] // permute dword order + vmovdqu [edx], ymm0 + lea edx, [edx + 32] + sub ecx, 8 + jg convertloop + vzeroupper + ret + } +} +#endif // HAS_ARGBMIRRORROW_AVX2 + +#ifdef HAS_SPLITUVROW_SSE2 +__declspec(naked) +void SplitUVRow_SSE2(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int pix) { + __asm { + push edi + mov eax, [esp + 4 + 4] // src_uv + mov edx, [esp + 4 + 8] // dst_u + mov edi, [esp + 4 + 12] // dst_v + mov ecx, [esp + 4 + 16] // pix + pcmpeqb xmm5, xmm5 // generate mask 0x00ff00ff + psrlw xmm5, 8 + sub edi, edx + + convertloop: + movdqu xmm0, [eax] + movdqu xmm1, [eax + 16] + lea eax, [eax + 32] + movdqa xmm2, xmm0 + movdqa xmm3, xmm1 + pand xmm0, xmm5 // even bytes + pand xmm1, xmm5 + packuswb xmm0, xmm1 + psrlw xmm2, 8 // odd bytes + psrlw xmm3, 8 + packuswb xmm2, xmm3 + movdqu [edx], xmm0 + movdqu [edx + edi], xmm2 + lea edx, [edx + 16] + sub ecx, 16 + jg convertloop + + pop edi + ret + } +} + +#endif // HAS_SPLITUVROW_SSE2 + +#ifdef HAS_SPLITUVROW_AVX2 +__declspec(naked) +void SplitUVRow_AVX2(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int pix) { + __asm { + push edi + mov eax, [esp + 4 + 4] // src_uv + mov edx, [esp + 4 + 8] // dst_u + mov edi, [esp + 4 + 12] // dst_v + mov ecx, [esp + 4 + 16] // pix + vpcmpeqb ymm5, ymm5, ymm5 // generate mask 0x00ff00ff + vpsrlw ymm5, ymm5, 8 + sub edi, edx + + convertloop: + vmovdqu ymm0, [eax] + vmovdqu ymm1, [eax + 32] + lea eax, [eax + 64] + vpsrlw ymm2, ymm0, 8 // odd bytes + vpsrlw ymm3, ymm1, 8 + vpand ymm0, ymm0, ymm5 // even bytes + vpand ymm1, ymm1, ymm5 + vpackuswb ymm0, ymm0, ymm1 + vpackuswb ymm2, ymm2, ymm3 + vpermq ymm0, ymm0, 0xd8 + vpermq ymm2, ymm2, 0xd8 + vmovdqu [edx], ymm0 + vmovdqu [edx + edi], ymm2 + lea edx, [edx + 32] + sub ecx, 32 + jg convertloop + + pop edi + vzeroupper + ret + } +} +#endif // HAS_SPLITUVROW_AVX2 + +#ifdef HAS_MERGEUVROW_SSE2 +__declspec(naked) +void MergeUVRow_SSE2(const uint8* src_u, const uint8* src_v, uint8* dst_uv, + int width) { + __asm { + push edi + mov eax, [esp + 4 + 4] // src_u + mov edx, [esp + 4 + 8] // src_v + mov edi, [esp + 4 + 12] // dst_uv + mov ecx, [esp + 4 + 16] // width + sub edx, eax + + convertloop: + movdqu xmm0, [eax] // read 16 U's + movdqu xmm1, [eax + edx] // and 16 V's + lea eax, [eax + 16] + movdqa xmm2, xmm0 + punpcklbw xmm0, xmm1 // first 8 UV pairs + punpckhbw xmm2, xmm1 // next 8 UV pairs + movdqu [edi], xmm0 + movdqu [edi + 16], xmm2 + lea edi, [edi + 32] + sub ecx, 16 + jg convertloop + + pop edi + ret + } +} +#endif // HAS_MERGEUVROW_SSE2 + +#ifdef HAS_MERGEUVROW_AVX2 +__declspec(naked) +void MergeUVRow_AVX2(const uint8* src_u, const uint8* src_v, uint8* dst_uv, + int width) { + __asm { + push edi + mov eax, [esp + 4 + 4] // src_u + mov edx, [esp + 4 + 8] // src_v + mov edi, [esp + 4 + 12] // dst_uv + mov ecx, [esp + 4 + 16] // width + sub edx, eax + + convertloop: + vmovdqu ymm0, [eax] // read 32 U's + vmovdqu ymm1, [eax + edx] // and 32 V's + lea eax, [eax + 32] + vpunpcklbw ymm2, ymm0, ymm1 // low 16 UV pairs. mutated qqword 0,2 + vpunpckhbw ymm0, ymm0, ymm1 // high 16 UV pairs. mutated qqword 1,3 + vextractf128 [edi], ymm2, 0 // bytes 0..15 + vextractf128 [edi + 16], ymm0, 0 // bytes 16..31 + vextractf128 [edi + 32], ymm2, 1 // bytes 32..47 + vextractf128 [edi + 48], ymm0, 1 // bytes 47..63 + lea edi, [edi + 64] + sub ecx, 32 + jg convertloop + + pop edi + vzeroupper + ret + } +} +#endif // HAS_MERGEUVROW_AVX2 + +#ifdef HAS_COPYROW_SSE2 +// CopyRow copys 'count' bytes using a 16 byte load/store, 32 bytes at time. +__declspec(naked) +void CopyRow_SSE2(const uint8* src, uint8* dst, int count) { + __asm { + mov eax, [esp + 4] // src + mov edx, [esp + 8] // dst + mov ecx, [esp + 12] // count + + convertloop: + movdqu xmm0, [eax] + movdqu xmm1, [eax + 16] + lea eax, [eax + 32] + movdqu [edx], xmm0 + movdqu [edx + 16], xmm1 + lea edx, [edx + 32] + sub ecx, 32 + jg convertloop + ret + } +} +#endif // HAS_COPYROW_SSE2 + +#ifdef HAS_COPYROW_AVX +// CopyRow copys 'count' bytes using a 32 byte load/store, 64 bytes at time. +__declspec(naked) +void CopyRow_AVX(const uint8* src, uint8* dst, int count) { + __asm { + mov eax, [esp + 4] // src + mov edx, [esp + 8] // dst + mov ecx, [esp + 12] // count + + convertloop: + vmovdqu ymm0, [eax] + vmovdqu ymm1, [eax + 32] + lea eax, [eax + 64] + vmovdqu [edx], ymm0 + vmovdqu [edx + 32], ymm1 + lea edx, [edx + 64] + sub ecx, 64 + jg convertloop + + vzeroupper + ret + } +} +#endif // HAS_COPYROW_AVX + +// Multiple of 1. +__declspec(naked) +void CopyRow_ERMS(const uint8* src, uint8* dst, int count) { + __asm { + mov eax, esi + mov edx, edi + mov esi, [esp + 4] // src + mov edi, [esp + 8] // dst + mov ecx, [esp + 12] // count + rep movsb + mov edi, edx + mov esi, eax + ret + } +} + +#ifdef HAS_ARGBCOPYALPHAROW_SSE2 +// width in pixels +__declspec(naked) +void ARGBCopyAlphaRow_SSE2(const uint8* src, uint8* dst, int width) { + __asm { + mov eax, [esp + 4] // src + mov edx, [esp + 8] // dst + mov ecx, [esp + 12] // count + pcmpeqb xmm0, xmm0 // generate mask 0xff000000 + pslld xmm0, 24 + pcmpeqb xmm1, xmm1 // generate mask 0x00ffffff + psrld xmm1, 8 + + convertloop: + movdqu xmm2, [eax] + movdqu xmm3, [eax + 16] + lea eax, [eax + 32] + movdqu xmm4, [edx] + movdqu xmm5, [edx + 16] + pand xmm2, xmm0 + pand xmm3, xmm0 + pand xmm4, xmm1 + pand xmm5, xmm1 + por xmm2, xmm4 + por xmm3, xmm5 + movdqu [edx], xmm2 + movdqu [edx + 16], xmm3 + lea edx, [edx + 32] + sub ecx, 8 + jg convertloop + + ret + } +} +#endif // HAS_ARGBCOPYALPHAROW_SSE2 + +#ifdef HAS_ARGBCOPYALPHAROW_AVX2 +// width in pixels +__declspec(naked) +void ARGBCopyAlphaRow_AVX2(const uint8* src, uint8* dst, int width) { + __asm { + mov eax, [esp + 4] // src + mov edx, [esp + 8] // dst + mov ecx, [esp + 12] // count + vpcmpeqb ymm0, ymm0, ymm0 + vpsrld ymm0, ymm0, 8 // generate mask 0x00ffffff + + convertloop: + vmovdqu ymm1, [eax] + vmovdqu ymm2, [eax + 32] + lea eax, [eax + 64] + vpblendvb ymm1, ymm1, [edx], ymm0 + vpblendvb ymm2, ymm2, [edx + 32], ymm0 + vmovdqu [edx], ymm1 + vmovdqu [edx + 32], ymm2 + lea edx, [edx + 64] + sub ecx, 16 + jg convertloop + + vzeroupper + ret + } +} +#endif // HAS_ARGBCOPYALPHAROW_AVX2 + +#ifdef HAS_ARGBCOPYYTOALPHAROW_SSE2 +// width in pixels +__declspec(naked) +void ARGBCopyYToAlphaRow_SSE2(const uint8* src, uint8* dst, int width) { + __asm { + mov eax, [esp + 4] // src + mov edx, [esp + 8] // dst + mov ecx, [esp + 12] // count + pcmpeqb xmm0, xmm0 // generate mask 0xff000000 + pslld xmm0, 24 + pcmpeqb xmm1, xmm1 // generate mask 0x00ffffff + psrld xmm1, 8 + + convertloop: + movq xmm2, qword ptr [eax] // 8 Y's + lea eax, [eax + 8] + punpcklbw xmm2, xmm2 + punpckhwd xmm3, xmm2 + punpcklwd xmm2, xmm2 + movdqu xmm4, [edx] + movdqu xmm5, [edx + 16] + pand xmm2, xmm0 + pand xmm3, xmm0 + pand xmm4, xmm1 + pand xmm5, xmm1 + por xmm2, xmm4 + por xmm3, xmm5 + movdqu [edx], xmm2 + movdqu [edx + 16], xmm3 + lea edx, [edx + 32] + sub ecx, 8 + jg convertloop + + ret + } +} +#endif // HAS_ARGBCOPYYTOALPHAROW_SSE2 + +#ifdef HAS_ARGBCOPYYTOALPHAROW_AVX2 +// width in pixels +__declspec(naked) +void ARGBCopyYToAlphaRow_AVX2(const uint8* src, uint8* dst, int width) { + __asm { + mov eax, [esp + 4] // src + mov edx, [esp + 8] // dst + mov ecx, [esp + 12] // count + vpcmpeqb ymm0, ymm0, ymm0 + vpsrld ymm0, ymm0, 8 // generate mask 0x00ffffff + + convertloop: + vpmovzxbd ymm1, qword ptr [eax] + vpmovzxbd ymm2, qword ptr [eax + 8] + lea eax, [eax + 16] + vpslld ymm1, ymm1, 24 + vpslld ymm2, ymm2, 24 + vpblendvb ymm1, ymm1, [edx], ymm0 + vpblendvb ymm2, ymm2, [edx + 32], ymm0 + vmovdqu [edx], ymm1 + vmovdqu [edx + 32], ymm2 + lea edx, [edx + 64] + sub ecx, 16 + jg convertloop + + vzeroupper + ret + } +} +#endif // HAS_ARGBCOPYYTOALPHAROW_AVX2 + +#ifdef HAS_SETROW_X86 +// Write 'count' bytes using an 8 bit value repeated. +// Count should be multiple of 4. +__declspec(naked) +void SetRow_X86(uint8* dst, uint8 v8, int count) { + __asm { + movzx eax, byte ptr [esp + 8] // v8 + mov edx, 0x01010101 // Duplicate byte to all bytes. + mul edx // overwrites edx with upper part of result. + mov edx, edi + mov edi, [esp + 4] // dst + mov ecx, [esp + 12] // count + shr ecx, 2 + rep stosd + mov edi, edx + ret + } +} + +// Write 'count' bytes using an 8 bit value repeated. +__declspec(naked) +void SetRow_ERMS(uint8* dst, uint8 v8, int count) { + __asm { + mov edx, edi + mov edi, [esp + 4] // dst + mov eax, [esp + 8] // v8 + mov ecx, [esp + 12] // count + rep stosb + mov edi, edx + ret + } +} + +// Write 'count' 32 bit values. +__declspec(naked) +void ARGBSetRow_X86(uint8* dst_argb, uint32 v32, int count) { + __asm { + mov edx, edi + mov edi, [esp + 4] // dst + mov eax, [esp + 8] // v32 + mov ecx, [esp + 12] // count + rep stosd + mov edi, edx + ret + } +} +#endif // HAS_SETROW_X86 + +#ifdef HAS_YUY2TOYROW_AVX2 +__declspec(naked) +void YUY2ToYRow_AVX2(const uint8* src_yuy2, + uint8* dst_y, int pix) { + __asm { + mov eax, [esp + 4] // src_yuy2 + mov edx, [esp + 8] // dst_y + mov ecx, [esp + 12] // pix + vpcmpeqb ymm5, ymm5, ymm5 // generate mask 0x00ff00ff + vpsrlw ymm5, ymm5, 8 + + convertloop: + vmovdqu ymm0, [eax] + vmovdqu ymm1, [eax + 32] + lea eax, [eax + 64] + vpand ymm0, ymm0, ymm5 // even bytes are Y + vpand ymm1, ymm1, ymm5 + vpackuswb ymm0, ymm0, ymm1 // mutates. + vpermq ymm0, ymm0, 0xd8 + vmovdqu [edx], ymm0 + lea edx, [edx + 32] + sub ecx, 32 + jg convertloop + vzeroupper + ret + } +} + +__declspec(naked) +void YUY2ToUVRow_AVX2(const uint8* src_yuy2, int stride_yuy2, + uint8* dst_u, uint8* dst_v, int pix) { + __asm { + push esi + push edi + mov eax, [esp + 8 + 4] // src_yuy2 + mov esi, [esp + 8 + 8] // stride_yuy2 + mov edx, [esp + 8 + 12] // dst_u + mov edi, [esp + 8 + 16] // dst_v + mov ecx, [esp + 8 + 20] // pix + vpcmpeqb ymm5, ymm5, ymm5 // generate mask 0x00ff00ff + vpsrlw ymm5, ymm5, 8 + sub edi, edx + + convertloop: + vmovdqu ymm0, [eax] + vmovdqu ymm1, [eax + 32] + vpavgb ymm0, ymm0, [eax + esi] + vpavgb ymm1, ymm1, [eax + esi + 32] + lea eax, [eax + 64] + vpsrlw ymm0, ymm0, 8 // YUYV -> UVUV + vpsrlw ymm1, ymm1, 8 + vpackuswb ymm0, ymm0, ymm1 // mutates. + vpermq ymm0, ymm0, 0xd8 + vpand ymm1, ymm0, ymm5 // U + vpsrlw ymm0, ymm0, 8 // V + vpackuswb ymm1, ymm1, ymm1 // mutates. + vpackuswb ymm0, ymm0, ymm0 // mutates. + vpermq ymm1, ymm1, 0xd8 + vpermq ymm0, ymm0, 0xd8 + vextractf128 [edx], ymm1, 0 // U + vextractf128 [edx + edi], ymm0, 0 // V + lea edx, [edx + 16] + sub ecx, 32 + jg convertloop + + pop edi + pop esi + vzeroupper + ret + } +} + +__declspec(naked) +void YUY2ToUV422Row_AVX2(const uint8* src_yuy2, + uint8* dst_u, uint8* dst_v, int pix) { + __asm { + push edi + mov eax, [esp + 4 + 4] // src_yuy2 + mov edx, [esp + 4 + 8] // dst_u + mov edi, [esp + 4 + 12] // dst_v + mov ecx, [esp + 4 + 16] // pix + vpcmpeqb ymm5, ymm5, ymm5 // generate mask 0x00ff00ff + vpsrlw ymm5, ymm5, 8 + sub edi, edx + + convertloop: + vmovdqu ymm0, [eax] + vmovdqu ymm1, [eax + 32] + lea eax, [eax + 64] + vpsrlw ymm0, ymm0, 8 // YUYV -> UVUV + vpsrlw ymm1, ymm1, 8 + vpackuswb ymm0, ymm0, ymm1 // mutates. + vpermq ymm0, ymm0, 0xd8 + vpand ymm1, ymm0, ymm5 // U + vpsrlw ymm0, ymm0, 8 // V + vpackuswb ymm1, ymm1, ymm1 // mutates. + vpackuswb ymm0, ymm0, ymm0 // mutates. + vpermq ymm1, ymm1, 0xd8 + vpermq ymm0, ymm0, 0xd8 + vextractf128 [edx], ymm1, 0 // U + vextractf128 [edx + edi], ymm0, 0 // V + lea edx, [edx + 16] + sub ecx, 32 + jg convertloop + + pop edi + vzeroupper + ret + } +} + +__declspec(naked) +void UYVYToYRow_AVX2(const uint8* src_uyvy, + uint8* dst_y, int pix) { + __asm { + mov eax, [esp + 4] // src_uyvy + mov edx, [esp + 8] // dst_y + mov ecx, [esp + 12] // pix + + convertloop: + vmovdqu ymm0, [eax] + vmovdqu ymm1, [eax + 32] + lea eax, [eax + 64] + vpsrlw ymm0, ymm0, 8 // odd bytes are Y + vpsrlw ymm1, ymm1, 8 + vpackuswb ymm0, ymm0, ymm1 // mutates. + vpermq ymm0, ymm0, 0xd8 + vmovdqu [edx], ymm0 + lea edx, [edx + 32] + sub ecx, 32 + jg convertloop + vzeroupper + ret + } +} + +__declspec(naked) +void UYVYToUVRow_AVX2(const uint8* src_uyvy, int stride_uyvy, + uint8* dst_u, uint8* dst_v, int pix) { + __asm { + push esi + push edi + mov eax, [esp + 8 + 4] // src_yuy2 + mov esi, [esp + 8 + 8] // stride_yuy2 + mov edx, [esp + 8 + 12] // dst_u + mov edi, [esp + 8 + 16] // dst_v + mov ecx, [esp + 8 + 20] // pix + vpcmpeqb ymm5, ymm5, ymm5 // generate mask 0x00ff00ff + vpsrlw ymm5, ymm5, 8 + sub edi, edx + + convertloop: + vmovdqu ymm0, [eax] + vmovdqu ymm1, [eax + 32] + vpavgb ymm0, ymm0, [eax + esi] + vpavgb ymm1, ymm1, [eax + esi + 32] + lea eax, [eax + 64] + vpand ymm0, ymm0, ymm5 // UYVY -> UVUV + vpand ymm1, ymm1, ymm5 + vpackuswb ymm0, ymm0, ymm1 // mutates. + vpermq ymm0, ymm0, 0xd8 + vpand ymm1, ymm0, ymm5 // U + vpsrlw ymm0, ymm0, 8 // V + vpackuswb ymm1, ymm1, ymm1 // mutates. + vpackuswb ymm0, ymm0, ymm0 // mutates. + vpermq ymm1, ymm1, 0xd8 + vpermq ymm0, ymm0, 0xd8 + vextractf128 [edx], ymm1, 0 // U + vextractf128 [edx + edi], ymm0, 0 // V + lea edx, [edx + 16] + sub ecx, 32 + jg convertloop + + pop edi + pop esi + vzeroupper + ret + } +} + +__declspec(naked) +void UYVYToUV422Row_AVX2(const uint8* src_uyvy, + uint8* dst_u, uint8* dst_v, int pix) { + __asm { + push edi + mov eax, [esp + 4 + 4] // src_yuy2 + mov edx, [esp + 4 + 8] // dst_u + mov edi, [esp + 4 + 12] // dst_v + mov ecx, [esp + 4 + 16] // pix + vpcmpeqb ymm5, ymm5, ymm5 // generate mask 0x00ff00ff + vpsrlw ymm5, ymm5, 8 + sub edi, edx + + convertloop: + vmovdqu ymm0, [eax] + vmovdqu ymm1, [eax + 32] + lea eax, [eax + 64] + vpand ymm0, ymm0, ymm5 // UYVY -> UVUV + vpand ymm1, ymm1, ymm5 + vpackuswb ymm0, ymm0, ymm1 // mutates. + vpermq ymm0, ymm0, 0xd8 + vpand ymm1, ymm0, ymm5 // U + vpsrlw ymm0, ymm0, 8 // V + vpackuswb ymm1, ymm1, ymm1 // mutates. + vpackuswb ymm0, ymm0, ymm0 // mutates. + vpermq ymm1, ymm1, 0xd8 + vpermq ymm0, ymm0, 0xd8 + vextractf128 [edx], ymm1, 0 // U + vextractf128 [edx + edi], ymm0, 0 // V + lea edx, [edx + 16] + sub ecx, 32 + jg convertloop + + pop edi + vzeroupper + ret + } +} +#endif // HAS_YUY2TOYROW_AVX2 + +#ifdef HAS_YUY2TOYROW_SSE2 +__declspec(naked) +void YUY2ToYRow_SSE2(const uint8* src_yuy2, + uint8* dst_y, int pix) { + __asm { + mov eax, [esp + 4] // src_yuy2 + mov edx, [esp + 8] // dst_y + mov ecx, [esp + 12] // pix + pcmpeqb xmm5, xmm5 // generate mask 0x00ff00ff + psrlw xmm5, 8 + + convertloop: + movdqu xmm0, [eax] + movdqu xmm1, [eax + 16] + lea eax, [eax + 32] + pand xmm0, xmm5 // even bytes are Y + pand xmm1, xmm5 + packuswb xmm0, xmm1 + movdqu [edx], xmm0 + lea edx, [edx + 16] + sub ecx, 16 + jg convertloop + ret + } +} + +__declspec(naked) +void YUY2ToUVRow_SSE2(const uint8* src_yuy2, int stride_yuy2, + uint8* dst_u, uint8* dst_v, int pix) { + __asm { + push esi + push edi + mov eax, [esp + 8 + 4] // src_yuy2 + mov esi, [esp + 8 + 8] // stride_yuy2 + mov edx, [esp + 8 + 12] // dst_u + mov edi, [esp + 8 + 16] // dst_v + mov ecx, [esp + 8 + 20] // pix + pcmpeqb xmm5, xmm5 // generate mask 0x00ff00ff + psrlw xmm5, 8 + sub edi, edx + + convertloop: + movdqu xmm0, [eax] + movdqu xmm1, [eax + 16] + movdqu xmm2, [eax + esi] + movdqu xmm3, [eax + esi + 16] + lea eax, [eax + 32] + pavgb xmm0, xmm2 + pavgb xmm1, xmm3 + psrlw xmm0, 8 // YUYV -> UVUV + psrlw xmm1, 8 + packuswb xmm0, xmm1 + movdqa xmm1, xmm0 + pand xmm0, xmm5 // U + packuswb xmm0, xmm0 + psrlw xmm1, 8 // V + packuswb xmm1, xmm1 + movq qword ptr [edx], xmm0 + movq qword ptr [edx + edi], xmm1 + lea edx, [edx + 8] + sub ecx, 16 + jg convertloop + + pop edi + pop esi + ret + } +} + +__declspec(naked) +void YUY2ToUV422Row_SSE2(const uint8* src_yuy2, + uint8* dst_u, uint8* dst_v, int pix) { + __asm { + push edi + mov eax, [esp + 4 + 4] // src_yuy2 + mov edx, [esp + 4 + 8] // dst_u + mov edi, [esp + 4 + 12] // dst_v + mov ecx, [esp + 4 + 16] // pix + pcmpeqb xmm5, xmm5 // generate mask 0x00ff00ff + psrlw xmm5, 8 + sub edi, edx + + convertloop: + movdqu xmm0, [eax] + movdqu xmm1, [eax + 16] + lea eax, [eax + 32] + psrlw xmm0, 8 // YUYV -> UVUV + psrlw xmm1, 8 + packuswb xmm0, xmm1 + movdqa xmm1, xmm0 + pand xmm0, xmm5 // U + packuswb xmm0, xmm0 + psrlw xmm1, 8 // V + packuswb xmm1, xmm1 + movq qword ptr [edx], xmm0 + movq qword ptr [edx + edi], xmm1 + lea edx, [edx + 8] + sub ecx, 16 + jg convertloop + + pop edi + ret + } +} + +__declspec(naked) +void UYVYToYRow_SSE2(const uint8* src_uyvy, + uint8* dst_y, int pix) { + __asm { + mov eax, [esp + 4] // src_uyvy + mov edx, [esp + 8] // dst_y + mov ecx, [esp + 12] // pix + + convertloop: + movdqu xmm0, [eax] + movdqu xmm1, [eax + 16] + lea eax, [eax + 32] + psrlw xmm0, 8 // odd bytes are Y + psrlw xmm1, 8 + packuswb xmm0, xmm1 + movdqu [edx], xmm0 + lea edx, [edx + 16] + sub ecx, 16 + jg convertloop + ret + } +} + +__declspec(naked) +void UYVYToUVRow_SSE2(const uint8* src_uyvy, int stride_uyvy, + uint8* dst_u, uint8* dst_v, int pix) { + __asm { + push esi + push edi + mov eax, [esp + 8 + 4] // src_yuy2 + mov esi, [esp + 8 + 8] // stride_yuy2 + mov edx, [esp + 8 + 12] // dst_u + mov edi, [esp + 8 + 16] // dst_v + mov ecx, [esp + 8 + 20] // pix + pcmpeqb xmm5, xmm5 // generate mask 0x00ff00ff + psrlw xmm5, 8 + sub edi, edx + + convertloop: + movdqu xmm0, [eax] + movdqu xmm1, [eax + 16] + movdqu xmm2, [eax + esi] + movdqu xmm3, [eax + esi + 16] + lea eax, [eax + 32] + pavgb xmm0, xmm2 + pavgb xmm1, xmm3 + pand xmm0, xmm5 // UYVY -> UVUV + pand xmm1, xmm5 + packuswb xmm0, xmm1 + movdqa xmm1, xmm0 + pand xmm0, xmm5 // U + packuswb xmm0, xmm0 + psrlw xmm1, 8 // V + packuswb xmm1, xmm1 + movq qword ptr [edx], xmm0 + movq qword ptr [edx + edi], xmm1 + lea edx, [edx + 8] + sub ecx, 16 + jg convertloop + + pop edi + pop esi + ret + } +} + +__declspec(naked) +void UYVYToUV422Row_SSE2(const uint8* src_uyvy, + uint8* dst_u, uint8* dst_v, int pix) { + __asm { + push edi + mov eax, [esp + 4 + 4] // src_yuy2 + mov edx, [esp + 4 + 8] // dst_u + mov edi, [esp + 4 + 12] // dst_v + mov ecx, [esp + 4 + 16] // pix + pcmpeqb xmm5, xmm5 // generate mask 0x00ff00ff + psrlw xmm5, 8 + sub edi, edx + + convertloop: + movdqu xmm0, [eax] + movdqu xmm1, [eax + 16] + lea eax, [eax + 32] + pand xmm0, xmm5 // UYVY -> UVUV + pand xmm1, xmm5 + packuswb xmm0, xmm1 + movdqa xmm1, xmm0 + pand xmm0, xmm5 // U + packuswb xmm0, xmm0 + psrlw xmm1, 8 // V + packuswb xmm1, xmm1 + movq qword ptr [edx], xmm0 + movq qword ptr [edx + edi], xmm1 + lea edx, [edx + 8] + sub ecx, 16 + jg convertloop + + pop edi + ret + } +} +#endif // HAS_YUY2TOYROW_SSE2 + +#ifdef HAS_ARGBBLENDROW_SSE2 +// Blend 8 pixels at a time. +__declspec(naked) +void ARGBBlendRow_SSE2(const uint8* src_argb0, const uint8* src_argb1, + uint8* dst_argb, int width) { + __asm { + push esi + mov eax, [esp + 4 + 4] // src_argb0 + mov esi, [esp + 4 + 8] // src_argb1 + mov edx, [esp + 4 + 12] // dst_argb + mov ecx, [esp + 4 + 16] // width + pcmpeqb xmm7, xmm7 // generate constant 1 + psrlw xmm7, 15 + pcmpeqb xmm6, xmm6 // generate mask 0x00ff00ff + psrlw xmm6, 8 + pcmpeqb xmm5, xmm5 // generate mask 0xff00ff00 + psllw xmm5, 8 + pcmpeqb xmm4, xmm4 // generate mask 0xff000000 + pslld xmm4, 24 + sub ecx, 4 + jl convertloop4b // less than 4 pixels? + + // 4 pixel loop. + convertloop4: + movdqu xmm3, [eax] // src argb + lea eax, [eax + 16] + movdqa xmm0, xmm3 // src argb + pxor xmm3, xmm4 // ~alpha + movdqu xmm2, [esi] // _r_b + psrlw xmm3, 8 // alpha + pshufhw xmm3, xmm3, 0F5h // 8 alpha words + pshuflw xmm3, xmm3, 0F5h + pand xmm2, xmm6 // _r_b + paddw xmm3, xmm7 // 256 - alpha + pmullw xmm2, xmm3 // _r_b * alpha + movdqu xmm1, [esi] // _a_g + lea esi, [esi + 16] + psrlw xmm1, 8 // _a_g + por xmm0, xmm4 // set alpha to 255 + pmullw xmm1, xmm3 // _a_g * alpha + psrlw xmm2, 8 // _r_b convert to 8 bits again + paddusb xmm0, xmm2 // + src argb + pand xmm1, xmm5 // a_g_ convert to 8 bits again + paddusb xmm0, xmm1 // + src argb + movdqu [edx], xmm0 + lea edx, [edx + 16] + sub ecx, 4 + jge convertloop4 + + convertloop4b: + add ecx, 4 - 1 + jl convertloop1b + + // 1 pixel loop. + convertloop1: + movd xmm3, [eax] // src argb + lea eax, [eax + 4] + movdqa xmm0, xmm3 // src argb + pxor xmm3, xmm4 // ~alpha + movd xmm2, [esi] // _r_b + psrlw xmm3, 8 // alpha + pshufhw xmm3, xmm3, 0F5h // 8 alpha words + pshuflw xmm3, xmm3, 0F5h + pand xmm2, xmm6 // _r_b + paddw xmm3, xmm7 // 256 - alpha + pmullw xmm2, xmm3 // _r_b * alpha + movd xmm1, [esi] // _a_g + lea esi, [esi + 4] + psrlw xmm1, 8 // _a_g + por xmm0, xmm4 // set alpha to 255 + pmullw xmm1, xmm3 // _a_g * alpha + psrlw xmm2, 8 // _r_b convert to 8 bits again + paddusb xmm0, xmm2 // + src argb + pand xmm1, xmm5 // a_g_ convert to 8 bits again + paddusb xmm0, xmm1 // + src argb + movd [edx], xmm0 + lea edx, [edx + 4] + sub ecx, 1 + jge convertloop1 + + convertloop1b: + pop esi + ret + } +} +#endif // HAS_ARGBBLENDROW_SSE2 + +#ifdef HAS_ARGBBLENDROW_SSSE3 +// Shuffle table for isolating alpha. +static const uvec8 kShuffleAlpha = { + 3u, 0x80, 3u, 0x80, 7u, 0x80, 7u, 0x80, + 11u, 0x80, 11u, 0x80, 15u, 0x80, 15u, 0x80 +}; +// Same as SSE2, but replaces: +// psrlw xmm3, 8 // alpha +// pshufhw xmm3, xmm3, 0F5h // 8 alpha words +// pshuflw xmm3, xmm3, 0F5h +// with.. +// pshufb xmm3, kShuffleAlpha // alpha +// Blend 8 pixels at a time. + +__declspec(naked) +void ARGBBlendRow_SSSE3(const uint8* src_argb0, const uint8* src_argb1, + uint8* dst_argb, int width) { + __asm { + push esi + mov eax, [esp + 4 + 4] // src_argb0 + mov esi, [esp + 4 + 8] // src_argb1 + mov edx, [esp + 4 + 12] // dst_argb + mov ecx, [esp + 4 + 16] // width + pcmpeqb xmm7, xmm7 // generate constant 0x0001 + psrlw xmm7, 15 + pcmpeqb xmm6, xmm6 // generate mask 0x00ff00ff + psrlw xmm6, 8 + pcmpeqb xmm5, xmm5 // generate mask 0xff00ff00 + psllw xmm5, 8 + pcmpeqb xmm4, xmm4 // generate mask 0xff000000 + pslld xmm4, 24 + sub ecx, 4 + jl convertloop4b // less than 4 pixels? + + // 4 pixel loop. + convertloop4: + movdqu xmm3, [eax] // src argb + lea eax, [eax + 16] + movdqa xmm0, xmm3 // src argb + pxor xmm3, xmm4 // ~alpha + movdqu xmm2, [esi] // _r_b + pshufb xmm3, kShuffleAlpha // alpha + pand xmm2, xmm6 // _r_b + paddw xmm3, xmm7 // 256 - alpha + pmullw xmm2, xmm3 // _r_b * alpha + movdqu xmm1, [esi] // _a_g + lea esi, [esi + 16] + psrlw xmm1, 8 // _a_g + por xmm0, xmm4 // set alpha to 255 + pmullw xmm1, xmm3 // _a_g * alpha + psrlw xmm2, 8 // _r_b convert to 8 bits again + paddusb xmm0, xmm2 // + src argb + pand xmm1, xmm5 // a_g_ convert to 8 bits again + paddusb xmm0, xmm1 // + src argb + movdqu [edx], xmm0 + lea edx, [edx + 16] + sub ecx, 4 + jge convertloop4 + + convertloop4b: + add ecx, 4 - 1 + jl convertloop1b + + // 1 pixel loop. + convertloop1: + movd xmm3, [eax] // src argb + lea eax, [eax + 4] + movdqa xmm0, xmm3 // src argb + pxor xmm3, xmm4 // ~alpha + movd xmm2, [esi] // _r_b + pshufb xmm3, kShuffleAlpha // alpha + pand xmm2, xmm6 // _r_b + paddw xmm3, xmm7 // 256 - alpha + pmullw xmm2, xmm3 // _r_b * alpha + movd xmm1, [esi] // _a_g + lea esi, [esi + 4] + psrlw xmm1, 8 // _a_g + por xmm0, xmm4 // set alpha to 255 + pmullw xmm1, xmm3 // _a_g * alpha + psrlw xmm2, 8 // _r_b convert to 8 bits again + paddusb xmm0, xmm2 // + src argb + pand xmm1, xmm5 // a_g_ convert to 8 bits again + paddusb xmm0, xmm1 // + src argb + movd [edx], xmm0 + lea edx, [edx + 4] + sub ecx, 1 + jge convertloop1 + + convertloop1b: + pop esi + ret + } +} +#endif // HAS_ARGBBLENDROW_SSSE3 + +#ifdef HAS_ARGBATTENUATEROW_SSE2 +// Attenuate 4 pixels at a time. +__declspec(naked) +void ARGBAttenuateRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width) { + __asm { + mov eax, [esp + 4] // src_argb0 + mov edx, [esp + 8] // dst_argb + mov ecx, [esp + 12] // width + pcmpeqb xmm4, xmm4 // generate mask 0xff000000 + pslld xmm4, 24 + pcmpeqb xmm5, xmm5 // generate mask 0x00ffffff + psrld xmm5, 8 + + convertloop: + movdqu xmm0, [eax] // read 4 pixels + punpcklbw xmm0, xmm0 // first 2 + pshufhw xmm2, xmm0, 0FFh // 8 alpha words + pshuflw xmm2, xmm2, 0FFh + pmulhuw xmm0, xmm2 // rgb * a + movdqu xmm1, [eax] // read 4 pixels + punpckhbw xmm1, xmm1 // next 2 pixels + pshufhw xmm2, xmm1, 0FFh // 8 alpha words + pshuflw xmm2, xmm2, 0FFh + pmulhuw xmm1, xmm2 // rgb * a + movdqu xmm2, [eax] // alphas + lea eax, [eax + 16] + psrlw xmm0, 8 + pand xmm2, xmm4 + psrlw xmm1, 8 + packuswb xmm0, xmm1 + pand xmm0, xmm5 // keep original alphas + por xmm0, xmm2 + movdqu [edx], xmm0 + lea edx, [edx + 16] + sub ecx, 4 + jg convertloop + + ret + } +} +#endif // HAS_ARGBATTENUATEROW_SSE2 + +#ifdef HAS_ARGBATTENUATEROW_SSSE3 +// Shuffle table duplicating alpha. +static const uvec8 kShuffleAlpha0 = { + 3u, 3u, 3u, 3u, 3u, 3u, 128u, 128u, 7u, 7u, 7u, 7u, 7u, 7u, 128u, 128u, +}; +static const uvec8 kShuffleAlpha1 = { + 11u, 11u, 11u, 11u, 11u, 11u, 128u, 128u, + 15u, 15u, 15u, 15u, 15u, 15u, 128u, 128u, +}; +__declspec(naked) +void ARGBAttenuateRow_SSSE3(const uint8* src_argb, uint8* dst_argb, int width) { + __asm { + mov eax, [esp + 4] // src_argb0 + mov edx, [esp + 8] // dst_argb + mov ecx, [esp + 12] // width + pcmpeqb xmm3, xmm3 // generate mask 0xff000000 + pslld xmm3, 24 + movdqa xmm4, kShuffleAlpha0 + movdqa xmm5, kShuffleAlpha1 + + convertloop: + movdqu xmm0, [eax] // read 4 pixels + pshufb xmm0, xmm4 // isolate first 2 alphas + movdqu xmm1, [eax] // read 4 pixels + punpcklbw xmm1, xmm1 // first 2 pixel rgbs + pmulhuw xmm0, xmm1 // rgb * a + movdqu xmm1, [eax] // read 4 pixels + pshufb xmm1, xmm5 // isolate next 2 alphas + movdqu xmm2, [eax] // read 4 pixels + punpckhbw xmm2, xmm2 // next 2 pixel rgbs + pmulhuw xmm1, xmm2 // rgb * a + movdqu xmm2, [eax] // mask original alpha + lea eax, [eax + 16] + pand xmm2, xmm3 + psrlw xmm0, 8 + psrlw xmm1, 8 + packuswb xmm0, xmm1 + por xmm0, xmm2 // copy original alpha + movdqu [edx], xmm0 + lea edx, [edx + 16] + sub ecx, 4 + jg convertloop + + ret + } +} +#endif // HAS_ARGBATTENUATEROW_SSSE3 + +#ifdef HAS_ARGBATTENUATEROW_AVX2 +// Shuffle table duplicating alpha. +static const uvec8 kShuffleAlpha_AVX2 = { + 6u, 7u, 6u, 7u, 6u, 7u, 128u, 128u, 14u, 15u, 14u, 15u, 14u, 15u, 128u, 128u +}; +__declspec(naked) +void ARGBAttenuateRow_AVX2(const uint8* src_argb, uint8* dst_argb, int width) { + __asm { + mov eax, [esp + 4] // src_argb0 + mov edx, [esp + 8] // dst_argb + mov ecx, [esp + 12] // width + sub edx, eax + vbroadcastf128 ymm4,kShuffleAlpha_AVX2 + vpcmpeqb ymm5, ymm5, ymm5 // generate mask 0xff000000 + vpslld ymm5, ymm5, 24 + + convertloop: + vmovdqu ymm6, [eax] // read 8 pixels. + vpunpcklbw ymm0, ymm6, ymm6 // low 4 pixels. mutated. + vpunpckhbw ymm1, ymm6, ymm6 // high 4 pixels. mutated. + vpshufb ymm2, ymm0, ymm4 // low 4 alphas + vpshufb ymm3, ymm1, ymm4 // high 4 alphas + vpmulhuw ymm0, ymm0, ymm2 // rgb * a + vpmulhuw ymm1, ymm1, ymm3 // rgb * a + vpand ymm6, ymm6, ymm5 // isolate alpha + vpsrlw ymm0, ymm0, 8 + vpsrlw ymm1, ymm1, 8 + vpackuswb ymm0, ymm0, ymm1 // unmutated. + vpor ymm0, ymm0, ymm6 // copy original alpha + vmovdqu [eax + edx], ymm0 + lea eax, [eax + 32] + sub ecx, 8 + jg convertloop + + vzeroupper + ret + } +} +#endif // HAS_ARGBATTENUATEROW_AVX2 + +#ifdef HAS_ARGBUNATTENUATEROW_SSE2 +// Unattenuate 4 pixels at a time. +__declspec(naked) +void ARGBUnattenuateRow_SSE2(const uint8* src_argb, uint8* dst_argb, + int width) { + __asm { + push esi + push edi + mov eax, [esp + 8 + 4] // src_argb0 + mov edx, [esp + 8 + 8] // dst_argb + mov ecx, [esp + 8 + 12] // width + + convertloop: + movdqu xmm0, [eax] // read 4 pixels + movzx esi, byte ptr [eax + 3] // first alpha + movzx edi, byte ptr [eax + 7] // second alpha + punpcklbw xmm0, xmm0 // first 2 + movd xmm2, dword ptr fixed_invtbl8[esi * 4] + movd xmm3, dword ptr fixed_invtbl8[edi * 4] + pshuflw xmm2, xmm2, 040h // first 4 inv_alpha words. 1, a, a, a + pshuflw xmm3, xmm3, 040h // next 4 inv_alpha words + movlhps xmm2, xmm3 + pmulhuw xmm0, xmm2 // rgb * a + + movdqu xmm1, [eax] // read 4 pixels + movzx esi, byte ptr [eax + 11] // third alpha + movzx edi, byte ptr [eax + 15] // forth alpha + punpckhbw xmm1, xmm1 // next 2 + movd xmm2, dword ptr fixed_invtbl8[esi * 4] + movd xmm3, dword ptr fixed_invtbl8[edi * 4] + pshuflw xmm2, xmm2, 040h // first 4 inv_alpha words + pshuflw xmm3, xmm3, 040h // next 4 inv_alpha words + movlhps xmm2, xmm3 + pmulhuw xmm1, xmm2 // rgb * a + lea eax, [eax + 16] + + packuswb xmm0, xmm1 + movdqu [edx], xmm0 + lea edx, [edx + 16] + sub ecx, 4 + jg convertloop + pop edi + pop esi + ret + } +} +#endif // HAS_ARGBUNATTENUATEROW_SSE2 + +#ifdef HAS_ARGBUNATTENUATEROW_AVX2 +// Shuffle table duplicating alpha. +static const uvec8 kUnattenShuffleAlpha_AVX2 = { + 0u, 1u, 0u, 1u, 0u, 1u, 6u, 7u, 8u, 9u, 8u, 9u, 8u, 9u, 14u, 15u +}; +// TODO(fbarchard): Enable USE_GATHER for future hardware if faster. +// USE_GATHER is not on by default, due to being a slow instruction. +#ifdef USE_GATHER +__declspec(naked) +void ARGBUnattenuateRow_AVX2(const uint8* src_argb, uint8* dst_argb, + int width) { + __asm { + mov eax, [esp + 4] // src_argb0 + mov edx, [esp + 8] // dst_argb + mov ecx, [esp + 12] // width + sub edx, eax + vbroadcastf128 ymm4, kUnattenShuffleAlpha_AVX2 + + convertloop: + vmovdqu ymm6, [eax] // read 8 pixels. + vpcmpeqb ymm5, ymm5, ymm5 // generate mask 0xffffffff for gather. + vpsrld ymm2, ymm6, 24 // alpha in low 8 bits. + vpunpcklbw ymm0, ymm6, ymm6 // low 4 pixels. mutated. + vpunpckhbw ymm1, ymm6, ymm6 // high 4 pixels. mutated. + vpgatherdd ymm3, [ymm2 * 4 + fixed_invtbl8], ymm5 // ymm5 cleared. 1, a + vpunpcklwd ymm2, ymm3, ymm3 // low 4 inverted alphas. mutated. 1, 1, a, a + vpunpckhwd ymm3, ymm3, ymm3 // high 4 inverted alphas. mutated. + vpshufb ymm2, ymm2, ymm4 // replicate low 4 alphas. 1, a, a, a + vpshufb ymm3, ymm3, ymm4 // replicate high 4 alphas + vpmulhuw ymm0, ymm0, ymm2 // rgb * ia + vpmulhuw ymm1, ymm1, ymm3 // rgb * ia + vpackuswb ymm0, ymm0, ymm1 // unmutated. + vmovdqu [eax + edx], ymm0 + lea eax, [eax + 32] + sub ecx, 8 + jg convertloop + + vzeroupper + ret + } +} +#else // USE_GATHER +__declspec(naked) +void ARGBUnattenuateRow_AVX2(const uint8* src_argb, uint8* dst_argb, + int width) { + __asm { + + mov eax, [esp + 4] // src_argb0 + mov edx, [esp + 8] // dst_argb + mov ecx, [esp + 12] // width + sub edx, eax + vbroadcastf128 ymm5, kUnattenShuffleAlpha_AVX2 + + push esi + push edi + + convertloop: + // replace VPGATHER + movzx esi, byte ptr [eax + 3] // alpha0 + movzx edi, byte ptr [eax + 7] // alpha1 + vmovd xmm0, dword ptr fixed_invtbl8[esi * 4] // [1,a0] + vmovd xmm1, dword ptr fixed_invtbl8[edi * 4] // [1,a1] + movzx esi, byte ptr [eax + 11] // alpha2 + movzx edi, byte ptr [eax + 15] // alpha3 + vpunpckldq xmm6, xmm0, xmm1 // [1,a1,1,a0] + vmovd xmm2, dword ptr fixed_invtbl8[esi * 4] // [1,a2] + vmovd xmm3, dword ptr fixed_invtbl8[edi * 4] // [1,a3] + movzx esi, byte ptr [eax + 19] // alpha4 + movzx edi, byte ptr [eax + 23] // alpha5 + vpunpckldq xmm7, xmm2, xmm3 // [1,a3,1,a2] + vmovd xmm0, dword ptr fixed_invtbl8[esi * 4] // [1,a4] + vmovd xmm1, dword ptr fixed_invtbl8[edi * 4] // [1,a5] + movzx esi, byte ptr [eax + 27] // alpha6 + movzx edi, byte ptr [eax + 31] // alpha7 + vpunpckldq xmm0, xmm0, xmm1 // [1,a5,1,a4] + vmovd xmm2, dword ptr fixed_invtbl8[esi * 4] // [1,a6] + vmovd xmm3, dword ptr fixed_invtbl8[edi * 4] // [1,a7] + vpunpckldq xmm2, xmm2, xmm3 // [1,a7,1,a6] + vpunpcklqdq xmm3, xmm6, xmm7 // [1,a3,1,a2,1,a1,1,a0] + vpunpcklqdq xmm0, xmm0, xmm2 // [1,a7,1,a6,1,a5,1,a4] + vinserti128 ymm3, ymm3, xmm0, 1 // [1,a7,1,a6,1,a5,1,a4,1,a3,1,a2,1,a1,1,a0] + // end of VPGATHER + + vmovdqu ymm6, [eax] // read 8 pixels. + vpunpcklbw ymm0, ymm6, ymm6 // low 4 pixels. mutated. + vpunpckhbw ymm1, ymm6, ymm6 // high 4 pixels. mutated. + vpunpcklwd ymm2, ymm3, ymm3 // low 4 inverted alphas. mutated. 1, 1, a, a + vpunpckhwd ymm3, ymm3, ymm3 // high 4 inverted alphas. mutated. + vpshufb ymm2, ymm2, ymm5 // replicate low 4 alphas. 1, a, a, a + vpshufb ymm3, ymm3, ymm5 // replicate high 4 alphas + vpmulhuw ymm0, ymm0, ymm2 // rgb * ia + vpmulhuw ymm1, ymm1, ymm3 // rgb * ia + vpackuswb ymm0, ymm0, ymm1 // unmutated. + vmovdqu [eax + edx], ymm0 + lea eax, [eax + 32] + sub ecx, 8 + jg convertloop + + pop edi + pop esi + vzeroupper + ret + } +} +#endif // USE_GATHER +#endif // HAS_ARGBATTENUATEROW_AVX2 + +#ifdef HAS_ARGBGRAYROW_SSSE3 +// Convert 8 ARGB pixels (64 bytes) to 8 Gray ARGB pixels. +__declspec(naked) +void ARGBGrayRow_SSSE3(const uint8* src_argb, uint8* dst_argb, int width) { + __asm { + mov eax, [esp + 4] /* src_argb */ + mov edx, [esp + 8] /* dst_argb */ + mov ecx, [esp + 12] /* width */ + movdqa xmm4, kARGBToYJ + movdqa xmm5, kAddYJ64 + + convertloop: + movdqu xmm0, [eax] // G + movdqu xmm1, [eax + 16] + pmaddubsw xmm0, xmm4 + pmaddubsw xmm1, xmm4 + phaddw xmm0, xmm1 + paddw xmm0, xmm5 // Add .5 for rounding. + psrlw xmm0, 7 + packuswb xmm0, xmm0 // 8 G bytes + movdqu xmm2, [eax] // A + movdqu xmm3, [eax + 16] + lea eax, [eax + 32] + psrld xmm2, 24 + psrld xmm3, 24 + packuswb xmm2, xmm3 + packuswb xmm2, xmm2 // 8 A bytes + movdqa xmm3, xmm0 // Weave into GG, GA, then GGGA + punpcklbw xmm0, xmm0 // 8 GG words + punpcklbw xmm3, xmm2 // 8 GA words + movdqa xmm1, xmm0 + punpcklwd xmm0, xmm3 // GGGA first 4 + punpckhwd xmm1, xmm3 // GGGA next 4 + movdqu [edx], xmm0 + movdqu [edx + 16], xmm1 + lea edx, [edx + 32] + sub ecx, 8 + jg convertloop + ret + } +} +#endif // HAS_ARGBGRAYROW_SSSE3 + +#ifdef HAS_ARGBSEPIAROW_SSSE3 +// b = (r * 35 + g * 68 + b * 17) >> 7 +// g = (r * 45 + g * 88 + b * 22) >> 7 +// r = (r * 50 + g * 98 + b * 24) >> 7 +// Constant for ARGB color to sepia tone. +static const vec8 kARGBToSepiaB = { + 17, 68, 35, 0, 17, 68, 35, 0, 17, 68, 35, 0, 17, 68, 35, 0 +}; + +static const vec8 kARGBToSepiaG = { + 22, 88, 45, 0, 22, 88, 45, 0, 22, 88, 45, 0, 22, 88, 45, 0 +}; + +static const vec8 kARGBToSepiaR = { + 24, 98, 50, 0, 24, 98, 50, 0, 24, 98, 50, 0, 24, 98, 50, 0 +}; + +// Convert 8 ARGB pixels (32 bytes) to 8 Sepia ARGB pixels. +__declspec(naked) +void ARGBSepiaRow_SSSE3(uint8* dst_argb, int width) { + __asm { + mov eax, [esp + 4] /* dst_argb */ + mov ecx, [esp + 8] /* width */ + movdqa xmm2, kARGBToSepiaB + movdqa xmm3, kARGBToSepiaG + movdqa xmm4, kARGBToSepiaR + + convertloop: + movdqu xmm0, [eax] // B + movdqu xmm6, [eax + 16] + pmaddubsw xmm0, xmm2 + pmaddubsw xmm6, xmm2 + phaddw xmm0, xmm6 + psrlw xmm0, 7 + packuswb xmm0, xmm0 // 8 B values + movdqu xmm5, [eax] // G + movdqu xmm1, [eax + 16] + pmaddubsw xmm5, xmm3 + pmaddubsw xmm1, xmm3 + phaddw xmm5, xmm1 + psrlw xmm5, 7 + packuswb xmm5, xmm5 // 8 G values + punpcklbw xmm0, xmm5 // 8 BG values + movdqu xmm5, [eax] // R + movdqu xmm1, [eax + 16] + pmaddubsw xmm5, xmm4 + pmaddubsw xmm1, xmm4 + phaddw xmm5, xmm1 + psrlw xmm5, 7 + packuswb xmm5, xmm5 // 8 R values + movdqu xmm6, [eax] // A + movdqu xmm1, [eax + 16] + psrld xmm6, 24 + psrld xmm1, 24 + packuswb xmm6, xmm1 + packuswb xmm6, xmm6 // 8 A values + punpcklbw xmm5, xmm6 // 8 RA values + movdqa xmm1, xmm0 // Weave BG, RA together + punpcklwd xmm0, xmm5 // BGRA first 4 + punpckhwd xmm1, xmm5 // BGRA next 4 + movdqu [eax], xmm0 + movdqu [eax + 16], xmm1 + lea eax, [eax + 32] + sub ecx, 8 + jg convertloop + ret + } +} +#endif // HAS_ARGBSEPIAROW_SSSE3 + +#ifdef HAS_ARGBCOLORMATRIXROW_SSSE3 +// Tranform 8 ARGB pixels (32 bytes) with color matrix. +// Same as Sepia except matrix is provided. +// TODO(fbarchard): packuswbs only use half of the reg. To make RGBA, combine R +// and B into a high and low, then G/A, unpackl/hbw and then unpckl/hwd. +__declspec(naked) +void ARGBColorMatrixRow_SSSE3(const uint8* src_argb, uint8* dst_argb, + const int8* matrix_argb, int width) { + __asm { + mov eax, [esp + 4] /* src_argb */ + mov edx, [esp + 8] /* dst_argb */ + mov ecx, [esp + 12] /* matrix_argb */ + movdqu xmm5, [ecx] + pshufd xmm2, xmm5, 0x00 + pshufd xmm3, xmm5, 0x55 + pshufd xmm4, xmm5, 0xaa + pshufd xmm5, xmm5, 0xff + mov ecx, [esp + 16] /* width */ + + convertloop: + movdqu xmm0, [eax] // B + movdqu xmm7, [eax + 16] + pmaddubsw xmm0, xmm2 + pmaddubsw xmm7, xmm2 + movdqu xmm6, [eax] // G + movdqu xmm1, [eax + 16] + pmaddubsw xmm6, xmm3 + pmaddubsw xmm1, xmm3 + phaddsw xmm0, xmm7 // B + phaddsw xmm6, xmm1 // G + psraw xmm0, 6 // B + psraw xmm6, 6 // G + packuswb xmm0, xmm0 // 8 B values + packuswb xmm6, xmm6 // 8 G values + punpcklbw xmm0, xmm6 // 8 BG values + movdqu xmm1, [eax] // R + movdqu xmm7, [eax + 16] + pmaddubsw xmm1, xmm4 + pmaddubsw xmm7, xmm4 + phaddsw xmm1, xmm7 // R + movdqu xmm6, [eax] // A + movdqu xmm7, [eax + 16] + pmaddubsw xmm6, xmm5 + pmaddubsw xmm7, xmm5 + phaddsw xmm6, xmm7 // A + psraw xmm1, 6 // R + psraw xmm6, 6 // A + packuswb xmm1, xmm1 // 8 R values + packuswb xmm6, xmm6 // 8 A values + punpcklbw xmm1, xmm6 // 8 RA values + movdqa xmm6, xmm0 // Weave BG, RA together + punpcklwd xmm0, xmm1 // BGRA first 4 + punpckhwd xmm6, xmm1 // BGRA next 4 + movdqu [edx], xmm0 + movdqu [edx + 16], xmm6 + lea eax, [eax + 32] + lea edx, [edx + 32] + sub ecx, 8 + jg convertloop + ret + } +} +#endif // HAS_ARGBCOLORMATRIXROW_SSSE3 + +#ifdef HAS_ARGBQUANTIZEROW_SSE2 +// Quantize 4 ARGB pixels (16 bytes). +__declspec(naked) +void ARGBQuantizeRow_SSE2(uint8* dst_argb, int scale, int interval_size, + int interval_offset, int width) { + __asm { + mov eax, [esp + 4] /* dst_argb */ + movd xmm2, [esp + 8] /* scale */ + movd xmm3, [esp + 12] /* interval_size */ + movd xmm4, [esp + 16] /* interval_offset */ + mov ecx, [esp + 20] /* width */ + pshuflw xmm2, xmm2, 040h + pshufd xmm2, xmm2, 044h + pshuflw xmm3, xmm3, 040h + pshufd xmm3, xmm3, 044h + pshuflw xmm4, xmm4, 040h + pshufd xmm4, xmm4, 044h + pxor xmm5, xmm5 // constant 0 + pcmpeqb xmm6, xmm6 // generate mask 0xff000000 + pslld xmm6, 24 + + convertloop: + movdqu xmm0, [eax] // read 4 pixels + punpcklbw xmm0, xmm5 // first 2 pixels + pmulhuw xmm0, xmm2 // pixel * scale >> 16 + movdqu xmm1, [eax] // read 4 pixels + punpckhbw xmm1, xmm5 // next 2 pixels + pmulhuw xmm1, xmm2 + pmullw xmm0, xmm3 // * interval_size + movdqu xmm7, [eax] // read 4 pixels + pmullw xmm1, xmm3 + pand xmm7, xmm6 // mask alpha + paddw xmm0, xmm4 // + interval_size / 2 + paddw xmm1, xmm4 + packuswb xmm0, xmm1 + por xmm0, xmm7 + movdqu [eax], xmm0 + lea eax, [eax + 16] + sub ecx, 4 + jg convertloop + ret + } +} +#endif // HAS_ARGBQUANTIZEROW_SSE2 + +#ifdef HAS_ARGBSHADEROW_SSE2 +// Shade 4 pixels at a time by specified value. +__declspec(naked) +void ARGBShadeRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width, + uint32 value) { + __asm { + mov eax, [esp + 4] // src_argb + mov edx, [esp + 8] // dst_argb + mov ecx, [esp + 12] // width + movd xmm2, [esp + 16] // value + punpcklbw xmm2, xmm2 + punpcklqdq xmm2, xmm2 + + convertloop: + movdqu xmm0, [eax] // read 4 pixels + lea eax, [eax + 16] + movdqa xmm1, xmm0 + punpcklbw xmm0, xmm0 // first 2 + punpckhbw xmm1, xmm1 // next 2 + pmulhuw xmm0, xmm2 // argb * value + pmulhuw xmm1, xmm2 // argb * value + psrlw xmm0, 8 + psrlw xmm1, 8 + packuswb xmm0, xmm1 + movdqu [edx], xmm0 + lea edx, [edx + 16] + sub ecx, 4 + jg convertloop + + ret + } +} +#endif // HAS_ARGBSHADEROW_SSE2 + +#ifdef HAS_ARGBMULTIPLYROW_SSE2 +// Multiply 2 rows of ARGB pixels together, 4 pixels at a time. +__declspec(naked) +void ARGBMultiplyRow_SSE2(const uint8* src_argb0, const uint8* src_argb1, + uint8* dst_argb, int width) { + __asm { + push esi + mov eax, [esp + 4 + 4] // src_argb0 + mov esi, [esp + 4 + 8] // src_argb1 + mov edx, [esp + 4 + 12] // dst_argb + mov ecx, [esp + 4 + 16] // width + pxor xmm5, xmm5 // constant 0 + + convertloop: + movdqu xmm0, [eax] // read 4 pixels from src_argb0 + movdqu xmm2, [esi] // read 4 pixels from src_argb1 + movdqu xmm1, xmm0 + movdqu xmm3, xmm2 + punpcklbw xmm0, xmm0 // first 2 + punpckhbw xmm1, xmm1 // next 2 + punpcklbw xmm2, xmm5 // first 2 + punpckhbw xmm3, xmm5 // next 2 + pmulhuw xmm0, xmm2 // src_argb0 * src_argb1 first 2 + pmulhuw xmm1, xmm3 // src_argb0 * src_argb1 next 2 + lea eax, [eax + 16] + lea esi, [esi + 16] + packuswb xmm0, xmm1 + movdqu [edx], xmm0 + lea edx, [edx + 16] + sub ecx, 4 + jg convertloop + + pop esi + ret + } +} +#endif // HAS_ARGBMULTIPLYROW_SSE2 + +#ifdef HAS_ARGBADDROW_SSE2 +// Add 2 rows of ARGB pixels together, 4 pixels at a time. +// TODO(fbarchard): Port this to posix, neon and other math functions. +__declspec(naked) +void ARGBAddRow_SSE2(const uint8* src_argb0, const uint8* src_argb1, + uint8* dst_argb, int width) { + __asm { + push esi + mov eax, [esp + 4 + 4] // src_argb0 + mov esi, [esp + 4 + 8] // src_argb1 + mov edx, [esp + 4 + 12] // dst_argb + mov ecx, [esp + 4 + 16] // width + + sub ecx, 4 + jl convertloop49 + + convertloop4: + movdqu xmm0, [eax] // read 4 pixels from src_argb0 + lea eax, [eax + 16] + movdqu xmm1, [esi] // read 4 pixels from src_argb1 + lea esi, [esi + 16] + paddusb xmm0, xmm1 // src_argb0 + src_argb1 + movdqu [edx], xmm0 + lea edx, [edx + 16] + sub ecx, 4 + jge convertloop4 + + convertloop49: + add ecx, 4 - 1 + jl convertloop19 + + convertloop1: + movd xmm0, [eax] // read 1 pixels from src_argb0 + lea eax, [eax + 4] + movd xmm1, [esi] // read 1 pixels from src_argb1 + lea esi, [esi + 4] + paddusb xmm0, xmm1 // src_argb0 + src_argb1 + movd [edx], xmm0 + lea edx, [edx + 4] + sub ecx, 1 + jge convertloop1 + + convertloop19: + pop esi + ret + } +} +#endif // HAS_ARGBADDROW_SSE2 + +#ifdef HAS_ARGBSUBTRACTROW_SSE2 +// Subtract 2 rows of ARGB pixels together, 4 pixels at a time. +__declspec(naked) +void ARGBSubtractRow_SSE2(const uint8* src_argb0, const uint8* src_argb1, + uint8* dst_argb, int width) { + __asm { + push esi + mov eax, [esp + 4 + 4] // src_argb0 + mov esi, [esp + 4 + 8] // src_argb1 + mov edx, [esp + 4 + 12] // dst_argb + mov ecx, [esp + 4 + 16] // width + + convertloop: + movdqu xmm0, [eax] // read 4 pixels from src_argb0 + lea eax, [eax + 16] + movdqu xmm1, [esi] // read 4 pixels from src_argb1 + lea esi, [esi + 16] + psubusb xmm0, xmm1 // src_argb0 - src_argb1 + movdqu [edx], xmm0 + lea edx, [edx + 16] + sub ecx, 4 + jg convertloop + + pop esi + ret + } +} +#endif // HAS_ARGBSUBTRACTROW_SSE2 + +#ifdef HAS_ARGBMULTIPLYROW_AVX2 +// Multiply 2 rows of ARGB pixels together, 8 pixels at a time. +__declspec(naked) +void ARGBMultiplyRow_AVX2(const uint8* src_argb0, const uint8* src_argb1, + uint8* dst_argb, int width) { + __asm { + push esi + mov eax, [esp + 4 + 4] // src_argb0 + mov esi, [esp + 4 + 8] // src_argb1 + mov edx, [esp + 4 + 12] // dst_argb + mov ecx, [esp + 4 + 16] // width + vpxor ymm5, ymm5, ymm5 // constant 0 + + convertloop: + vmovdqu ymm1, [eax] // read 8 pixels from src_argb0 + lea eax, [eax + 32] + vmovdqu ymm3, [esi] // read 8 pixels from src_argb1 + lea esi, [esi + 32] + vpunpcklbw ymm0, ymm1, ymm1 // low 4 + vpunpckhbw ymm1, ymm1, ymm1 // high 4 + vpunpcklbw ymm2, ymm3, ymm5 // low 4 + vpunpckhbw ymm3, ymm3, ymm5 // high 4 + vpmulhuw ymm0, ymm0, ymm2 // src_argb0 * src_argb1 low 4 + vpmulhuw ymm1, ymm1, ymm3 // src_argb0 * src_argb1 high 4 + vpackuswb ymm0, ymm0, ymm1 + vmovdqu [edx], ymm0 + lea edx, [edx + 32] + sub ecx, 8 + jg convertloop + + pop esi + vzeroupper + ret + } +} +#endif // HAS_ARGBMULTIPLYROW_AVX2 + +#ifdef HAS_ARGBADDROW_AVX2 +// Add 2 rows of ARGB pixels together, 8 pixels at a time. +__declspec(naked) +void ARGBAddRow_AVX2(const uint8* src_argb0, const uint8* src_argb1, + uint8* dst_argb, int width) { + __asm { + push esi + mov eax, [esp + 4 + 4] // src_argb0 + mov esi, [esp + 4 + 8] // src_argb1 + mov edx, [esp + 4 + 12] // dst_argb + mov ecx, [esp + 4 + 16] // width + + convertloop: + vmovdqu ymm0, [eax] // read 8 pixels from src_argb0 + lea eax, [eax + 32] + vpaddusb ymm0, ymm0, [esi] // add 8 pixels from src_argb1 + lea esi, [esi + 32] + vmovdqu [edx], ymm0 + lea edx, [edx + 32] + sub ecx, 8 + jg convertloop + + pop esi + vzeroupper + ret + } +} +#endif // HAS_ARGBADDROW_AVX2 + +#ifdef HAS_ARGBSUBTRACTROW_AVX2 +// Subtract 2 rows of ARGB pixels together, 8 pixels at a time. +__declspec(naked) +void ARGBSubtractRow_AVX2(const uint8* src_argb0, const uint8* src_argb1, + uint8* dst_argb, int width) { + __asm { + push esi + mov eax, [esp + 4 + 4] // src_argb0 + mov esi, [esp + 4 + 8] // src_argb1 + mov edx, [esp + 4 + 12] // dst_argb + mov ecx, [esp + 4 + 16] // width + + convertloop: + vmovdqu ymm0, [eax] // read 8 pixels from src_argb0 + lea eax, [eax + 32] + vpsubusb ymm0, ymm0, [esi] // src_argb0 - src_argb1 + lea esi, [esi + 32] + vmovdqu [edx], ymm0 + lea edx, [edx + 32] + sub ecx, 8 + jg convertloop + + pop esi + vzeroupper + ret + } +} +#endif // HAS_ARGBSUBTRACTROW_AVX2 + +#ifdef HAS_SOBELXROW_SSE2 +// SobelX as a matrix is +// -1 0 1 +// -2 0 2 +// -1 0 1 +__declspec(naked) +void SobelXRow_SSE2(const uint8* src_y0, const uint8* src_y1, + const uint8* src_y2, uint8* dst_sobelx, int width) { + __asm { + push esi + push edi + mov eax, [esp + 8 + 4] // src_y0 + mov esi, [esp + 8 + 8] // src_y1 + mov edi, [esp + 8 + 12] // src_y2 + mov edx, [esp + 8 + 16] // dst_sobelx + mov ecx, [esp + 8 + 20] // width + sub esi, eax + sub edi, eax + sub edx, eax + pxor xmm5, xmm5 // constant 0 + + convertloop: + movq xmm0, qword ptr [eax] // read 8 pixels from src_y0[0] + movq xmm1, qword ptr [eax + 2] // read 8 pixels from src_y0[2] + punpcklbw xmm0, xmm5 + punpcklbw xmm1, xmm5 + psubw xmm0, xmm1 + movq xmm1, qword ptr [eax + esi] // read 8 pixels from src_y1[0] + movq xmm2, qword ptr [eax + esi + 2] // read 8 pixels from src_y1[2] + punpcklbw xmm1, xmm5 + punpcklbw xmm2, xmm5 + psubw xmm1, xmm2 + movq xmm2, qword ptr [eax + edi] // read 8 pixels from src_y2[0] + movq xmm3, qword ptr [eax + edi + 2] // read 8 pixels from src_y2[2] + punpcklbw xmm2, xmm5 + punpcklbw xmm3, xmm5 + psubw xmm2, xmm3 + paddw xmm0, xmm2 + paddw xmm0, xmm1 + paddw xmm0, xmm1 + pxor xmm1, xmm1 // abs = max(xmm0, -xmm0). SSSE3 could use pabsw + psubw xmm1, xmm0 + pmaxsw xmm0, xmm1 + packuswb xmm0, xmm0 + movq qword ptr [eax + edx], xmm0 + lea eax, [eax + 8] + sub ecx, 8 + jg convertloop + + pop edi + pop esi + ret + } +} +#endif // HAS_SOBELXROW_SSE2 + +#ifdef HAS_SOBELYROW_SSE2 +// SobelY as a matrix is +// -1 -2 -1 +// 0 0 0 +// 1 2 1 +__declspec(naked) +void SobelYRow_SSE2(const uint8* src_y0, const uint8* src_y1, + uint8* dst_sobely, int width) { + __asm { + push esi + mov eax, [esp + 4 + 4] // src_y0 + mov esi, [esp + 4 + 8] // src_y1 + mov edx, [esp + 4 + 12] // dst_sobely + mov ecx, [esp + 4 + 16] // width + sub esi, eax + sub edx, eax + pxor xmm5, xmm5 // constant 0 + + convertloop: + movq xmm0, qword ptr [eax] // read 8 pixels from src_y0[0] + movq xmm1, qword ptr [eax + esi] // read 8 pixels from src_y1[0] + punpcklbw xmm0, xmm5 + punpcklbw xmm1, xmm5 + psubw xmm0, xmm1 + movq xmm1, qword ptr [eax + 1] // read 8 pixels from src_y0[1] + movq xmm2, qword ptr [eax + esi + 1] // read 8 pixels from src_y1[1] + punpcklbw xmm1, xmm5 + punpcklbw xmm2, xmm5 + psubw xmm1, xmm2 + movq xmm2, qword ptr [eax + 2] // read 8 pixels from src_y0[2] + movq xmm3, qword ptr [eax + esi + 2] // read 8 pixels from src_y1[2] + punpcklbw xmm2, xmm5 + punpcklbw xmm3, xmm5 + psubw xmm2, xmm3 + paddw xmm0, xmm2 + paddw xmm0, xmm1 + paddw xmm0, xmm1 + pxor xmm1, xmm1 // abs = max(xmm0, -xmm0). SSSE3 could use pabsw + psubw xmm1, xmm0 + pmaxsw xmm0, xmm1 + packuswb xmm0, xmm0 + movq qword ptr [eax + edx], xmm0 + lea eax, [eax + 8] + sub ecx, 8 + jg convertloop + + pop esi + ret + } +} +#endif // HAS_SOBELYROW_SSE2 + +#ifdef HAS_SOBELROW_SSE2 +// Adds Sobel X and Sobel Y and stores Sobel into ARGB. +// A = 255 +// R = Sobel +// G = Sobel +// B = Sobel +__declspec(naked) +void SobelRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely, + uint8* dst_argb, int width) { + __asm { + push esi + mov eax, [esp + 4 + 4] // src_sobelx + mov esi, [esp + 4 + 8] // src_sobely + mov edx, [esp + 4 + 12] // dst_argb + mov ecx, [esp + 4 + 16] // width + sub esi, eax + pcmpeqb xmm5, xmm5 // alpha 255 + pslld xmm5, 24 // 0xff000000 + + convertloop: + movdqu xmm0, [eax] // read 16 pixels src_sobelx + movdqu xmm1, [eax + esi] // read 16 pixels src_sobely + lea eax, [eax + 16] + paddusb xmm0, xmm1 // sobel = sobelx + sobely + movdqa xmm2, xmm0 // GG + punpcklbw xmm2, xmm0 // First 8 + punpckhbw xmm0, xmm0 // Next 8 + movdqa xmm1, xmm2 // GGGG + punpcklwd xmm1, xmm2 // First 4 + punpckhwd xmm2, xmm2 // Next 4 + por xmm1, xmm5 // GGGA + por xmm2, xmm5 + movdqa xmm3, xmm0 // GGGG + punpcklwd xmm3, xmm0 // Next 4 + punpckhwd xmm0, xmm0 // Last 4 + por xmm3, xmm5 // GGGA + por xmm0, xmm5 + movdqu [edx], xmm1 + movdqu [edx + 16], xmm2 + movdqu [edx + 32], xmm3 + movdqu [edx + 48], xmm0 + lea edx, [edx + 64] + sub ecx, 16 + jg convertloop + + pop esi + ret + } +} +#endif // HAS_SOBELROW_SSE2 + +#ifdef HAS_SOBELTOPLANEROW_SSE2 +// Adds Sobel X and Sobel Y and stores Sobel into a plane. +__declspec(naked) +void SobelToPlaneRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely, + uint8* dst_y, int width) { + __asm { + push esi + mov eax, [esp + 4 + 4] // src_sobelx + mov esi, [esp + 4 + 8] // src_sobely + mov edx, [esp + 4 + 12] // dst_argb + mov ecx, [esp + 4 + 16] // width + sub esi, eax + + convertloop: + movdqu xmm0, [eax] // read 16 pixels src_sobelx + movdqu xmm1, [eax + esi] // read 16 pixels src_sobely + lea eax, [eax + 16] + paddusb xmm0, xmm1 // sobel = sobelx + sobely + movdqu [edx], xmm0 + lea edx, [edx + 16] + sub ecx, 16 + jg convertloop + + pop esi + ret + } +} +#endif // HAS_SOBELTOPLANEROW_SSE2 + +#ifdef HAS_SOBELXYROW_SSE2 +// Mixes Sobel X, Sobel Y and Sobel into ARGB. +// A = 255 +// R = Sobel X +// G = Sobel +// B = Sobel Y +__declspec(naked) +void SobelXYRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely, + uint8* dst_argb, int width) { + __asm { + push esi + mov eax, [esp + 4 + 4] // src_sobelx + mov esi, [esp + 4 + 8] // src_sobely + mov edx, [esp + 4 + 12] // dst_argb + mov ecx, [esp + 4 + 16] // width + sub esi, eax + pcmpeqb xmm5, xmm5 // alpha 255 + + convertloop: + movdqu xmm0, [eax] // read 16 pixels src_sobelx + movdqu xmm1, [eax + esi] // read 16 pixels src_sobely + lea eax, [eax + 16] + movdqa xmm2, xmm0 + paddusb xmm2, xmm1 // sobel = sobelx + sobely + movdqa xmm3, xmm0 // XA + punpcklbw xmm3, xmm5 + punpckhbw xmm0, xmm5 + movdqa xmm4, xmm1 // YS + punpcklbw xmm4, xmm2 + punpckhbw xmm1, xmm2 + movdqa xmm6, xmm4 // YSXA + punpcklwd xmm6, xmm3 // First 4 + punpckhwd xmm4, xmm3 // Next 4 + movdqa xmm7, xmm1 // YSXA + punpcklwd xmm7, xmm0 // Next 4 + punpckhwd xmm1, xmm0 // Last 4 + movdqu [edx], xmm6 + movdqu [edx + 16], xmm4 + movdqu [edx + 32], xmm7 + movdqu [edx + 48], xmm1 + lea edx, [edx + 64] + sub ecx, 16 + jg convertloop + + pop esi + ret + } +} +#endif // HAS_SOBELXYROW_SSE2 + +#ifdef HAS_CUMULATIVESUMTOAVERAGEROW_SSE2 +// Consider float CumulativeSum. +// Consider calling CumulativeSum one row at time as needed. +// Consider circular CumulativeSum buffer of radius * 2 + 1 height. +// Convert cumulative sum for an area to an average for 1 pixel. +// topleft is pointer to top left of CumulativeSum buffer for area. +// botleft is pointer to bottom left of CumulativeSum buffer. +// width is offset from left to right of area in CumulativeSum buffer measured +// in number of ints. +// area is the number of pixels in the area being averaged. +// dst points to pixel to store result to. +// count is number of averaged pixels to produce. +// Does 4 pixels at a time. +void CumulativeSumToAverageRow_SSE2(const int32* topleft, const int32* botleft, + int width, int area, uint8* dst, + int count) { + __asm { + mov eax, topleft // eax topleft + mov esi, botleft // esi botleft + mov edx, width + movd xmm5, area + mov edi, dst + mov ecx, count + cvtdq2ps xmm5, xmm5 + rcpss xmm4, xmm5 // 1.0f / area + pshufd xmm4, xmm4, 0 + sub ecx, 4 + jl l4b + + cmp area, 128 // 128 pixels will not overflow 15 bits. + ja l4 + + pshufd xmm5, xmm5, 0 // area + pcmpeqb xmm6, xmm6 // constant of 65536.0 - 1 = 65535.0 + psrld xmm6, 16 + cvtdq2ps xmm6, xmm6 + addps xmm5, xmm6 // (65536.0 + area - 1) + mulps xmm5, xmm4 // (65536.0 + area - 1) * 1 / area + cvtps2dq xmm5, xmm5 // 0.16 fixed point + packssdw xmm5, xmm5 // 16 bit shorts + + // 4 pixel loop small blocks. + s4: + // top left + movdqu xmm0, [eax] + movdqu xmm1, [eax + 16] + movdqu xmm2, [eax + 32] + movdqu xmm3, [eax + 48] + + // - top right + psubd xmm0, [eax + edx * 4] + psubd xmm1, [eax + edx * 4 + 16] + psubd xmm2, [eax + edx * 4 + 32] + psubd xmm3, [eax + edx * 4 + 48] + lea eax, [eax + 64] + + // - bottom left + psubd xmm0, [esi] + psubd xmm1, [esi + 16] + psubd xmm2, [esi + 32] + psubd xmm3, [esi + 48] + + // + bottom right + paddd xmm0, [esi + edx * 4] + paddd xmm1, [esi + edx * 4 + 16] + paddd xmm2, [esi + edx * 4 + 32] + paddd xmm3, [esi + edx * 4 + 48] + lea esi, [esi + 64] + + packssdw xmm0, xmm1 // pack 4 pixels into 2 registers + packssdw xmm2, xmm3 + + pmulhuw xmm0, xmm5 + pmulhuw xmm2, xmm5 + + packuswb xmm0, xmm2 + movdqu [edi], xmm0 + lea edi, [edi + 16] + sub ecx, 4 + jge s4 + + jmp l4b + + // 4 pixel loop + l4: + // top left + movdqu xmm0, [eax] + movdqu xmm1, [eax + 16] + movdqu xmm2, [eax + 32] + movdqu xmm3, [eax + 48] + + // - top right + psubd xmm0, [eax + edx * 4] + psubd xmm1, [eax + edx * 4 + 16] + psubd xmm2, [eax + edx * 4 + 32] + psubd xmm3, [eax + edx * 4 + 48] + lea eax, [eax + 64] + + // - bottom left + psubd xmm0, [esi] + psubd xmm1, [esi + 16] + psubd xmm2, [esi + 32] + psubd xmm3, [esi + 48] + + // + bottom right + paddd xmm0, [esi + edx * 4] + paddd xmm1, [esi + edx * 4 + 16] + paddd xmm2, [esi + edx * 4 + 32] + paddd xmm3, [esi + edx * 4 + 48] + lea esi, [esi + 64] + + cvtdq2ps xmm0, xmm0 // Average = Sum * 1 / Area + cvtdq2ps xmm1, xmm1 + mulps xmm0, xmm4 + mulps xmm1, xmm4 + cvtdq2ps xmm2, xmm2 + cvtdq2ps xmm3, xmm3 + mulps xmm2, xmm4 + mulps xmm3, xmm4 + cvtps2dq xmm0, xmm0 + cvtps2dq xmm1, xmm1 + cvtps2dq xmm2, xmm2 + cvtps2dq xmm3, xmm3 + packssdw xmm0, xmm1 + packssdw xmm2, xmm3 + packuswb xmm0, xmm2 + movdqu [edi], xmm0 + lea edi, [edi + 16] + sub ecx, 4 + jge l4 + + l4b: + add ecx, 4 - 1 + jl l1b + + // 1 pixel loop + l1: + movdqu xmm0, [eax] + psubd xmm0, [eax + edx * 4] + lea eax, [eax + 16] + psubd xmm0, [esi] + paddd xmm0, [esi + edx * 4] + lea esi, [esi + 16] + cvtdq2ps xmm0, xmm0 + mulps xmm0, xmm4 + cvtps2dq xmm0, xmm0 + packssdw xmm0, xmm0 + packuswb xmm0, xmm0 + movd dword ptr [edi], xmm0 + lea edi, [edi + 4] + sub ecx, 1 + jge l1 + l1b: + } +} +#endif // HAS_CUMULATIVESUMTOAVERAGEROW_SSE2 + +#ifdef HAS_COMPUTECUMULATIVESUMROW_SSE2 +// Creates a table of cumulative sums where each value is a sum of all values +// above and to the left of the value. +void ComputeCumulativeSumRow_SSE2(const uint8* row, int32* cumsum, + const int32* previous_cumsum, int width) { + __asm { + mov eax, row + mov edx, cumsum + mov esi, previous_cumsum + mov ecx, width + pxor xmm0, xmm0 + pxor xmm1, xmm1 + + sub ecx, 4 + jl l4b + test edx, 15 + jne l4b + + // 4 pixel loop + l4: + movdqu xmm2, [eax] // 4 argb pixels 16 bytes. + lea eax, [eax + 16] + movdqa xmm4, xmm2 + + punpcklbw xmm2, xmm1 + movdqa xmm3, xmm2 + punpcklwd xmm2, xmm1 + punpckhwd xmm3, xmm1 + + punpckhbw xmm4, xmm1 + movdqa xmm5, xmm4 + punpcklwd xmm4, xmm1 + punpckhwd xmm5, xmm1 + + paddd xmm0, xmm2 + movdqu xmm2, [esi] // previous row above. + paddd xmm2, xmm0 + + paddd xmm0, xmm3 + movdqu xmm3, [esi + 16] + paddd xmm3, xmm0 + + paddd xmm0, xmm4 + movdqu xmm4, [esi + 32] + paddd xmm4, xmm0 + + paddd xmm0, xmm5 + movdqu xmm5, [esi + 48] + lea esi, [esi + 64] + paddd xmm5, xmm0 + + movdqu [edx], xmm2 + movdqu [edx + 16], xmm3 + movdqu [edx + 32], xmm4 + movdqu [edx + 48], xmm5 + + lea edx, [edx + 64] + sub ecx, 4 + jge l4 + + l4b: + add ecx, 4 - 1 + jl l1b + + // 1 pixel loop + l1: + movd xmm2, dword ptr [eax] // 1 argb pixel 4 bytes. + lea eax, [eax + 4] + punpcklbw xmm2, xmm1 + punpcklwd xmm2, xmm1 + paddd xmm0, xmm2 + movdqu xmm2, [esi] + lea esi, [esi + 16] + paddd xmm2, xmm0 + movdqu [edx], xmm2 + lea edx, [edx + 16] + sub ecx, 1 + jge l1 + + l1b: + } +} +#endif // HAS_COMPUTECUMULATIVESUMROW_SSE2 + +#ifdef HAS_ARGBAFFINEROW_SSE2 +// Copy ARGB pixels from source image with slope to a row of destination. +__declspec(naked) +LIBYUV_API +void ARGBAffineRow_SSE2(const uint8* src_argb, int src_argb_stride, + uint8* dst_argb, const float* uv_dudv, int width) { + __asm { + push esi + push edi + mov eax, [esp + 12] // src_argb + mov esi, [esp + 16] // stride + mov edx, [esp + 20] // dst_argb + mov ecx, [esp + 24] // pointer to uv_dudv + movq xmm2, qword ptr [ecx] // uv + movq xmm7, qword ptr [ecx + 8] // dudv + mov ecx, [esp + 28] // width + shl esi, 16 // 4, stride + add esi, 4 + movd xmm5, esi + sub ecx, 4 + jl l4b + + // setup for 4 pixel loop + pshufd xmm7, xmm7, 0x44 // dup dudv + pshufd xmm5, xmm5, 0 // dup 4, stride + movdqa xmm0, xmm2 // x0, y0, x1, y1 + addps xmm0, xmm7 + movlhps xmm2, xmm0 + movdqa xmm4, xmm7 + addps xmm4, xmm4 // dudv *= 2 + movdqa xmm3, xmm2 // x2, y2, x3, y3 + addps xmm3, xmm4 + addps xmm4, xmm4 // dudv *= 4 + + // 4 pixel loop + l4: + cvttps2dq xmm0, xmm2 // x, y float to int first 2 + cvttps2dq xmm1, xmm3 // x, y float to int next 2 + packssdw xmm0, xmm1 // x, y as 8 shorts + pmaddwd xmm0, xmm5 // offsets = x * 4 + y * stride. + movd esi, xmm0 + pshufd xmm0, xmm0, 0x39 // shift right + movd edi, xmm0 + pshufd xmm0, xmm0, 0x39 // shift right + movd xmm1, [eax + esi] // read pixel 0 + movd xmm6, [eax + edi] // read pixel 1 + punpckldq xmm1, xmm6 // combine pixel 0 and 1 + addps xmm2, xmm4 // x, y += dx, dy first 2 + movq qword ptr [edx], xmm1 + movd esi, xmm0 + pshufd xmm0, xmm0, 0x39 // shift right + movd edi, xmm0 + movd xmm6, [eax + esi] // read pixel 2 + movd xmm0, [eax + edi] // read pixel 3 + punpckldq xmm6, xmm0 // combine pixel 2 and 3 + addps xmm3, xmm4 // x, y += dx, dy next 2 + movq qword ptr 8[edx], xmm6 + lea edx, [edx + 16] + sub ecx, 4 + jge l4 + + l4b: + add ecx, 4 - 1 + jl l1b + + // 1 pixel loop + l1: + cvttps2dq xmm0, xmm2 // x, y float to int + packssdw xmm0, xmm0 // x, y as shorts + pmaddwd xmm0, xmm5 // offset = x * 4 + y * stride + addps xmm2, xmm7 // x, y += dx, dy + movd esi, xmm0 + movd xmm0, [eax + esi] // copy a pixel + movd [edx], xmm0 + lea edx, [edx + 4] + sub ecx, 1 + jge l1 + l1b: + pop edi + pop esi + ret + } +} +#endif // HAS_ARGBAFFINEROW_SSE2 + +#ifdef HAS_INTERPOLATEROW_AVX2 +// Bilinear filter 32x2 -> 32x1 +__declspec(naked) +void InterpolateRow_AVX2(uint8* dst_ptr, const uint8* src_ptr, + ptrdiff_t src_stride, int dst_width, + int source_y_fraction) { + __asm { + push esi + push edi + mov edi, [esp + 8 + 4] // dst_ptr + mov esi, [esp + 8 + 8] // src_ptr + mov edx, [esp + 8 + 12] // src_stride + mov ecx, [esp + 8 + 16] // dst_width + mov eax, [esp + 8 + 20] // source_y_fraction (0..255) + shr eax, 1 + // Dispatch to specialized filters if applicable. + cmp eax, 0 + je xloop100 // 0 / 128. Blend 100 / 0. + sub edi, esi + cmp eax, 32 + je xloop75 // 32 / 128 is 0.25. Blend 75 / 25. + cmp eax, 64 + je xloop50 // 64 / 128 is 0.50. Blend 50 / 50. + cmp eax, 96 + je xloop25 // 96 / 128 is 0.75. Blend 25 / 75. + + vmovd xmm0, eax // high fraction 0..127 + neg eax + add eax, 128 + vmovd xmm5, eax // low fraction 128..1 + vpunpcklbw xmm5, xmm5, xmm0 + vpunpcklwd xmm5, xmm5, xmm5 + vpxor ymm0, ymm0, ymm0 + vpermd ymm5, ymm0, ymm5 + + xloop: + vmovdqu ymm0, [esi] + vmovdqu ymm2, [esi + edx] + vpunpckhbw ymm1, ymm0, ymm2 // mutates + vpunpcklbw ymm0, ymm0, ymm2 // mutates + vpmaddubsw ymm0, ymm0, ymm5 + vpmaddubsw ymm1, ymm1, ymm5 + vpsrlw ymm0, ymm0, 7 + vpsrlw ymm1, ymm1, 7 + vpackuswb ymm0, ymm0, ymm1 // unmutates + vmovdqu [esi + edi], ymm0 + lea esi, [esi + 32] + sub ecx, 32 + jg xloop + jmp xloop99 + + // Blend 25 / 75. + xloop25: + vmovdqu ymm0, [esi] + vmovdqu ymm1, [esi + edx] + vpavgb ymm0, ymm0, ymm1 + vpavgb ymm0, ymm0, ymm1 + vmovdqu [esi + edi], ymm0 + lea esi, [esi + 32] + sub ecx, 32 + jg xloop25 + jmp xloop99 + + // Blend 50 / 50. + xloop50: + vmovdqu ymm0, [esi] + vpavgb ymm0, ymm0, [esi + edx] + vmovdqu [esi + edi], ymm0 + lea esi, [esi + 32] + sub ecx, 32 + jg xloop50 + jmp xloop99 + + // Blend 75 / 25. + xloop75: + vmovdqu ymm1, [esi] + vmovdqu ymm0, [esi + edx] + vpavgb ymm0, ymm0, ymm1 + vpavgb ymm0, ymm0, ymm1 + vmovdqu [esi + edi], ymm0 + lea esi, [esi + 32] + sub ecx, 32 + jg xloop75 + jmp xloop99 + + // Blend 100 / 0 - Copy row unchanged. + xloop100: + rep movsb + + xloop99: + pop edi + pop esi + vzeroupper + ret + } +} +#endif // HAS_INTERPOLATEROW_AVX2 + +// Bilinear filter 16x2 -> 16x1 +__declspec(naked) +void InterpolateRow_SSSE3(uint8* dst_ptr, const uint8* src_ptr, + ptrdiff_t src_stride, int dst_width, + int source_y_fraction) { + __asm { + push esi + push edi + mov edi, [esp + 8 + 4] // dst_ptr + mov esi, [esp + 8 + 8] // src_ptr + mov edx, [esp + 8 + 12] // src_stride + mov ecx, [esp + 8 + 16] // dst_width + mov eax, [esp + 8 + 20] // source_y_fraction (0..255) + sub edi, esi + shr eax, 1 + // Dispatch to specialized filters if applicable. + cmp eax, 0 + je xloop100 // 0 / 128. Blend 100 / 0. + cmp eax, 32 + je xloop75 // 32 / 128 is 0.25. Blend 75 / 25. + cmp eax, 64 + je xloop50 // 64 / 128 is 0.50. Blend 50 / 50. + cmp eax, 96 + je xloop25 // 96 / 128 is 0.75. Blend 25 / 75. + + movd xmm0, eax // high fraction 0..127 + neg eax + add eax, 128 + movd xmm5, eax // low fraction 128..1 + punpcklbw xmm5, xmm0 + punpcklwd xmm5, xmm5 + pshufd xmm5, xmm5, 0 + + xloop: + movdqu xmm0, [esi] + movdqu xmm2, [esi + edx] + movdqu xmm1, xmm0 + punpcklbw xmm0, xmm2 + punpckhbw xmm1, xmm2 + pmaddubsw xmm0, xmm5 + pmaddubsw xmm1, xmm5 + psrlw xmm0, 7 + psrlw xmm1, 7 + packuswb xmm0, xmm1 + movdqu [esi + edi], xmm0 + lea esi, [esi + 16] + sub ecx, 16 + jg xloop + jmp xloop99 + + // Blend 25 / 75. + xloop25: + movdqu xmm0, [esi] + movdqu xmm1, [esi + edx] + pavgb xmm0, xmm1 + pavgb xmm0, xmm1 + movdqu [esi + edi], xmm0 + lea esi, [esi + 16] + sub ecx, 16 + jg xloop25 + jmp xloop99 + + // Blend 50 / 50. + xloop50: + movdqu xmm0, [esi] + movdqu xmm1, [esi + edx] + pavgb xmm0, xmm1 + movdqu [esi + edi], xmm0 + lea esi, [esi + 16] + sub ecx, 16 + jg xloop50 + jmp xloop99 + + // Blend 75 / 25. + xloop75: + movdqu xmm1, [esi] + movdqu xmm0, [esi + edx] + pavgb xmm0, xmm1 + pavgb xmm0, xmm1 + movdqu [esi + edi], xmm0 + lea esi, [esi + 16] + sub ecx, 16 + jg xloop75 + jmp xloop99 + + // Blend 100 / 0 - Copy row unchanged. + xloop100: + movdqu xmm0, [esi] + movdqu [esi + edi], xmm0 + lea esi, [esi + 16] + sub ecx, 16 + jg xloop100 + + xloop99: + pop edi + pop esi + ret + } +} + +#ifdef HAS_INTERPOLATEROW_SSE2 +// Bilinear filter 16x2 -> 16x1 +__declspec(naked) +void InterpolateRow_SSE2(uint8* dst_ptr, const uint8* src_ptr, + ptrdiff_t src_stride, int dst_width, + int source_y_fraction) { + __asm { + push esi + push edi + mov edi, [esp + 8 + 4] // dst_ptr + mov esi, [esp + 8 + 8] // src_ptr + mov edx, [esp + 8 + 12] // src_stride + mov ecx, [esp + 8 + 16] // dst_width + mov eax, [esp + 8 + 20] // source_y_fraction (0..255) + sub edi, esi + // Dispatch to specialized filters if applicable. + cmp eax, 0 + je xloop100 // 0 / 256. Blend 100 / 0. + cmp eax, 64 + je xloop75 // 64 / 256 is 0.25. Blend 75 / 25. + cmp eax, 128 + je xloop50 // 128 / 256 is 0.50. Blend 50 / 50. + cmp eax, 192 + je xloop25 // 192 / 256 is 0.75. Blend 25 / 75. + + movd xmm5, eax // xmm5 = y fraction + punpcklbw xmm5, xmm5 + psrlw xmm5, 1 + punpcklwd xmm5, xmm5 + punpckldq xmm5, xmm5 + punpcklqdq xmm5, xmm5 + pxor xmm4, xmm4 + + xloop: + movdqu xmm0, [esi] // row0 + movdqu xmm2, [esi + edx] // row1 + movdqu xmm1, xmm0 + movdqu xmm3, xmm2 + punpcklbw xmm2, xmm4 + punpckhbw xmm3, xmm4 + punpcklbw xmm0, xmm4 + punpckhbw xmm1, xmm4 + psubw xmm2, xmm0 // row1 - row0 + psubw xmm3, xmm1 + paddw xmm2, xmm2 // 9 bits * 15 bits = 8.16 + paddw xmm3, xmm3 + pmulhw xmm2, xmm5 // scale diff + pmulhw xmm3, xmm5 + paddw xmm0, xmm2 // sum rows + paddw xmm1, xmm3 + packuswb xmm0, xmm1 + movdqu [esi + edi], xmm0 + lea esi, [esi + 16] + sub ecx, 16 + jg xloop + jmp xloop99 + + // Blend 25 / 75. + xloop25: + movdqu xmm0, [esi] + movdqu xmm1, [esi + edx] + pavgb xmm0, xmm1 + pavgb xmm0, xmm1 + movdqu [esi + edi], xmm0 + lea esi, [esi + 16] + sub ecx, 16 + jg xloop25 + jmp xloop99 + + // Blend 50 / 50. + xloop50: + movdqu xmm0, [esi] + movdqu xmm1, [esi + edx] + pavgb xmm0, xmm1 + movdqu [esi + edi], xmm0 + lea esi, [esi + 16] + sub ecx, 16 + jg xloop50 + jmp xloop99 + + // Blend 75 / 25. + xloop75: + movdqu xmm1, [esi] + movdqu xmm0, [esi + edx] + pavgb xmm0, xmm1 + pavgb xmm0, xmm1 + movdqu [esi + edi], xmm0 + lea esi, [esi + 16] + sub ecx, 16 + jg xloop75 + jmp xloop99 + + // Blend 100 / 0 - Copy row unchanged. + xloop100: + movdqu xmm0, [esi] + movdqu [esi + edi], xmm0 + lea esi, [esi + 16] + sub ecx, 16 + jg xloop100 + + xloop99: + pop edi + pop esi + ret + } +} +#endif // HAS_INTERPOLATEROW_SSE2 + +// For BGRAToARGB, ABGRToARGB, RGBAToARGB, and ARGBToRGBA. +__declspec(naked) +void ARGBShuffleRow_SSSE3(const uint8* src_argb, uint8* dst_argb, + const uint8* shuffler, int pix) { + __asm { + mov eax, [esp + 4] // src_argb + mov edx, [esp + 8] // dst_argb + mov ecx, [esp + 12] // shuffler + movdqu xmm5, [ecx] + mov ecx, [esp + 16] // pix + + wloop: + movdqu xmm0, [eax] + movdqu xmm1, [eax + 16] + lea eax, [eax + 32] + pshufb xmm0, xmm5 + pshufb xmm1, xmm5 + movdqu [edx], xmm0 + movdqu [edx + 16], xmm1 + lea edx, [edx + 32] + sub ecx, 8 + jg wloop + ret + } +} + +#ifdef HAS_ARGBSHUFFLEROW_AVX2 +__declspec(naked) +void ARGBShuffleRow_AVX2(const uint8* src_argb, uint8* dst_argb, + const uint8* shuffler, int pix) { + __asm { + mov eax, [esp + 4] // src_argb + mov edx, [esp + 8] // dst_argb + mov ecx, [esp + 12] // shuffler + vbroadcastf128 ymm5, [ecx] // same shuffle in high as low. + mov ecx, [esp + 16] // pix + + wloop: + vmovdqu ymm0, [eax] + vmovdqu ymm1, [eax + 32] + lea eax, [eax + 64] + vpshufb ymm0, ymm0, ymm5 + vpshufb ymm1, ymm1, ymm5 + vmovdqu [edx], ymm0 + vmovdqu [edx + 32], ymm1 + lea edx, [edx + 64] + sub ecx, 16 + jg wloop + + vzeroupper + ret + } +} +#endif // HAS_ARGBSHUFFLEROW_AVX2 + +__declspec(naked) +void ARGBShuffleRow_SSE2(const uint8* src_argb, uint8* dst_argb, + const uint8* shuffler, int pix) { + __asm { + push ebx + push esi + mov eax, [esp + 8 + 4] // src_argb + mov edx, [esp + 8 + 8] // dst_argb + mov esi, [esp + 8 + 12] // shuffler + mov ecx, [esp + 8 + 16] // pix + pxor xmm5, xmm5 + + mov ebx, [esi] // shuffler + cmp ebx, 0x03000102 + je shuf_3012 + cmp ebx, 0x00010203 + je shuf_0123 + cmp ebx, 0x00030201 + je shuf_0321 + cmp ebx, 0x02010003 + je shuf_2103 + + // TODO(fbarchard): Use one source pointer and 3 offsets. + shuf_any1: + movzx ebx, byte ptr [esi] + movzx ebx, byte ptr [eax + ebx] + mov [edx], bl + movzx ebx, byte ptr [esi + 1] + movzx ebx, byte ptr [eax + ebx] + mov [edx + 1], bl + movzx ebx, byte ptr [esi + 2] + movzx ebx, byte ptr [eax + ebx] + mov [edx + 2], bl + movzx ebx, byte ptr [esi + 3] + movzx ebx, byte ptr [eax + ebx] + mov [edx + 3], bl + lea eax, [eax + 4] + lea edx, [edx + 4] + sub ecx, 1 + jg shuf_any1 + jmp shuf99 + + shuf_0123: + movdqu xmm0, [eax] + lea eax, [eax + 16] + movdqa xmm1, xmm0 + punpcklbw xmm0, xmm5 + punpckhbw xmm1, xmm5 + pshufhw xmm0, xmm0, 01Bh // 1B = 00011011 = 0x0123 = BGRAToARGB + pshuflw xmm0, xmm0, 01Bh + pshufhw xmm1, xmm1, 01Bh + pshuflw xmm1, xmm1, 01Bh + packuswb xmm0, xmm1 + movdqu [edx], xmm0 + lea edx, [edx + 16] + sub ecx, 4 + jg shuf_0123 + jmp shuf99 + + shuf_0321: + movdqu xmm0, [eax] + lea eax, [eax + 16] + movdqa xmm1, xmm0 + punpcklbw xmm0, xmm5 + punpckhbw xmm1, xmm5 + pshufhw xmm0, xmm0, 039h // 39 = 00111001 = 0x0321 = RGBAToARGB + pshuflw xmm0, xmm0, 039h + pshufhw xmm1, xmm1, 039h + pshuflw xmm1, xmm1, 039h + packuswb xmm0, xmm1 + movdqu [edx], xmm0 + lea edx, [edx + 16] + sub ecx, 4 + jg shuf_0321 + jmp shuf99 + + shuf_2103: + movdqu xmm0, [eax] + lea eax, [eax + 16] + movdqa xmm1, xmm0 + punpcklbw xmm0, xmm5 + punpckhbw xmm1, xmm5 + pshufhw xmm0, xmm0, 093h // 93 = 10010011 = 0x2103 = ARGBToRGBA + pshuflw xmm0, xmm0, 093h + pshufhw xmm1, xmm1, 093h + pshuflw xmm1, xmm1, 093h + packuswb xmm0, xmm1 + movdqu [edx], xmm0 + lea edx, [edx + 16] + sub ecx, 4 + jg shuf_2103 + jmp shuf99 + + shuf_3012: + movdqu xmm0, [eax] + lea eax, [eax + 16] + movdqa xmm1, xmm0 + punpcklbw xmm0, xmm5 + punpckhbw xmm1, xmm5 + pshufhw xmm0, xmm0, 0C6h // C6 = 11000110 = 0x3012 = ABGRToARGB + pshuflw xmm0, xmm0, 0C6h + pshufhw xmm1, xmm1, 0C6h + pshuflw xmm1, xmm1, 0C6h + packuswb xmm0, xmm1 + movdqu [edx], xmm0 + lea edx, [edx + 16] + sub ecx, 4 + jg shuf_3012 + + shuf99: + pop esi + pop ebx + ret + } +} + +// YUY2 - Macro-pixel = 2 image pixels +// Y0U0Y1V0....Y2U2Y3V2...Y4U4Y5V4.... + +// UYVY - Macro-pixel = 2 image pixels +// U0Y0V0Y1 + +__declspec(naked) +void I422ToYUY2Row_SSE2(const uint8* src_y, + const uint8* src_u, + const uint8* src_v, + uint8* dst_frame, int width) { + __asm { + push esi + push edi + mov eax, [esp + 8 + 4] // src_y + mov esi, [esp + 8 + 8] // src_u + mov edx, [esp + 8 + 12] // src_v + mov edi, [esp + 8 + 16] // dst_frame + mov ecx, [esp + 8 + 20] // width + sub edx, esi + + convertloop: + movq xmm2, qword ptr [esi] // U + movq xmm3, qword ptr [esi + edx] // V + lea esi, [esi + 8] + punpcklbw xmm2, xmm3 // UV + movdqu xmm0, [eax] // Y + lea eax, [eax + 16] + movdqa xmm1, xmm0 + punpcklbw xmm0, xmm2 // YUYV + punpckhbw xmm1, xmm2 + movdqu [edi], xmm0 + movdqu [edi + 16], xmm1 + lea edi, [edi + 32] + sub ecx, 16 + jg convertloop + + pop edi + pop esi + ret + } +} + +__declspec(naked) +void I422ToUYVYRow_SSE2(const uint8* src_y, + const uint8* src_u, + const uint8* src_v, + uint8* dst_frame, int width) { + __asm { + push esi + push edi + mov eax, [esp + 8 + 4] // src_y + mov esi, [esp + 8 + 8] // src_u + mov edx, [esp + 8 + 12] // src_v + mov edi, [esp + 8 + 16] // dst_frame + mov ecx, [esp + 8 + 20] // width + sub edx, esi + + convertloop: + movq xmm2, qword ptr [esi] // U + movq xmm3, qword ptr [esi + edx] // V + lea esi, [esi + 8] + punpcklbw xmm2, xmm3 // UV + movdqu xmm0, [eax] // Y + movdqa xmm1, xmm2 + lea eax, [eax + 16] + punpcklbw xmm1, xmm0 // UYVY + punpckhbw xmm2, xmm0 + movdqu [edi], xmm1 + movdqu [edi + 16], xmm2 + lea edi, [edi + 32] + sub ecx, 16 + jg convertloop + + pop edi + pop esi + ret + } +} + +#ifdef HAS_ARGBPOLYNOMIALROW_SSE2 +__declspec(naked) +void ARGBPolynomialRow_SSE2(const uint8* src_argb, + uint8* dst_argb, const float* poly, + int width) { + __asm { + push esi + mov eax, [esp + 4 + 4] /* src_argb */ + mov edx, [esp + 4 + 8] /* dst_argb */ + mov esi, [esp + 4 + 12] /* poly */ + mov ecx, [esp + 4 + 16] /* width */ + pxor xmm3, xmm3 // 0 constant for zero extending bytes to ints. + + // 2 pixel loop. + convertloop: +// pmovzxbd xmm0, dword ptr [eax] // BGRA pixel +// pmovzxbd xmm4, dword ptr [eax + 4] // BGRA pixel + movq xmm0, qword ptr [eax] // BGRABGRA + lea eax, [eax + 8] + punpcklbw xmm0, xmm3 + movdqa xmm4, xmm0 + punpcklwd xmm0, xmm3 // pixel 0 + punpckhwd xmm4, xmm3 // pixel 1 + cvtdq2ps xmm0, xmm0 // 4 floats + cvtdq2ps xmm4, xmm4 + movdqa xmm1, xmm0 // X + movdqa xmm5, xmm4 + mulps xmm0, [esi + 16] // C1 * X + mulps xmm4, [esi + 16] + addps xmm0, [esi] // result = C0 + C1 * X + addps xmm4, [esi] + movdqa xmm2, xmm1 + movdqa xmm6, xmm5 + mulps xmm2, xmm1 // X * X + mulps xmm6, xmm5 + mulps xmm1, xmm2 // X * X * X + mulps xmm5, xmm6 + mulps xmm2, [esi + 32] // C2 * X * X + mulps xmm6, [esi + 32] + mulps xmm1, [esi + 48] // C3 * X * X * X + mulps xmm5, [esi + 48] + addps xmm0, xmm2 // result += C2 * X * X + addps xmm4, xmm6 + addps xmm0, xmm1 // result += C3 * X * X * X + addps xmm4, xmm5 + cvttps2dq xmm0, xmm0 + cvttps2dq xmm4, xmm4 + packuswb xmm0, xmm4 + packuswb xmm0, xmm0 + movq qword ptr [edx], xmm0 + lea edx, [edx + 8] + sub ecx, 2 + jg convertloop + pop esi + ret + } +} +#endif // HAS_ARGBPOLYNOMIALROW_SSE2 + +#ifdef HAS_ARGBPOLYNOMIALROW_AVX2 +__declspec(naked) +void ARGBPolynomialRow_AVX2(const uint8* src_argb, + uint8* dst_argb, const float* poly, + int width) { + __asm { + mov eax, [esp + 4] /* src_argb */ + mov edx, [esp + 8] /* dst_argb */ + mov ecx, [esp + 12] /* poly */ + vbroadcastf128 ymm4, [ecx] // C0 + vbroadcastf128 ymm5, [ecx + 16] // C1 + vbroadcastf128 ymm6, [ecx + 32] // C2 + vbroadcastf128 ymm7, [ecx + 48] // C3 + mov ecx, [esp + 16] /* width */ + + // 2 pixel loop. + convertloop: + vpmovzxbd ymm0, qword ptr [eax] // 2 BGRA pixels + lea eax, [eax + 8] + vcvtdq2ps ymm0, ymm0 // X 8 floats + vmulps ymm2, ymm0, ymm0 // X * X + vmulps ymm3, ymm0, ymm7 // C3 * X + vfmadd132ps ymm0, ymm4, ymm5 // result = C0 + C1 * X + vfmadd231ps ymm0, ymm2, ymm6 // result += C2 * X * X + vfmadd231ps ymm0, ymm2, ymm3 // result += C3 * X * X * X + vcvttps2dq ymm0, ymm0 + vpackusdw ymm0, ymm0, ymm0 // b0g0r0a0_00000000_b0g0r0a0_00000000 + vpermq ymm0, ymm0, 0xd8 // b0g0r0a0_b0g0r0a0_00000000_00000000 + vpackuswb xmm0, xmm0, xmm0 // bgrabgra_00000000_00000000_00000000 + vmovq qword ptr [edx], xmm0 + lea edx, [edx + 8] + sub ecx, 2 + jg convertloop + vzeroupper + ret + } +} +#endif // HAS_ARGBPOLYNOMIALROW_AVX2 + +#ifdef HAS_ARGBCOLORTABLEROW_X86 +// Tranform ARGB pixels with color table. +__declspec(naked) +void ARGBColorTableRow_X86(uint8* dst_argb, const uint8* table_argb, + int width) { + __asm { + push esi + mov eax, [esp + 4 + 4] /* dst_argb */ + mov esi, [esp + 4 + 8] /* table_argb */ + mov ecx, [esp + 4 + 12] /* width */ + + // 1 pixel loop. + convertloop: + movzx edx, byte ptr [eax] + lea eax, [eax + 4] + movzx edx, byte ptr [esi + edx * 4] + mov byte ptr [eax - 4], dl + movzx edx, byte ptr [eax - 4 + 1] + movzx edx, byte ptr [esi + edx * 4 + 1] + mov byte ptr [eax - 4 + 1], dl + movzx edx, byte ptr [eax - 4 + 2] + movzx edx, byte ptr [esi + edx * 4 + 2] + mov byte ptr [eax - 4 + 2], dl + movzx edx, byte ptr [eax - 4 + 3] + movzx edx, byte ptr [esi + edx * 4 + 3] + mov byte ptr [eax - 4 + 3], dl + dec ecx + jg convertloop + pop esi + ret + } +} +#endif // HAS_ARGBCOLORTABLEROW_X86 + +#ifdef HAS_RGBCOLORTABLEROW_X86 +// Tranform RGB pixels with color table. +__declspec(naked) +void RGBColorTableRow_X86(uint8* dst_argb, const uint8* table_argb, int width) { + __asm { + push esi + mov eax, [esp + 4 + 4] /* dst_argb */ + mov esi, [esp + 4 + 8] /* table_argb */ + mov ecx, [esp + 4 + 12] /* width */ + + // 1 pixel loop. + convertloop: + movzx edx, byte ptr [eax] + lea eax, [eax + 4] + movzx edx, byte ptr [esi + edx * 4] + mov byte ptr [eax - 4], dl + movzx edx, byte ptr [eax - 4 + 1] + movzx edx, byte ptr [esi + edx * 4 + 1] + mov byte ptr [eax - 4 + 1], dl + movzx edx, byte ptr [eax - 4 + 2] + movzx edx, byte ptr [esi + edx * 4 + 2] + mov byte ptr [eax - 4 + 2], dl + dec ecx + jg convertloop + + pop esi + ret + } +} +#endif // HAS_RGBCOLORTABLEROW_X86 + +#ifdef HAS_ARGBLUMACOLORTABLEROW_SSSE3 +// Tranform RGB pixels with luma table. +__declspec(naked) +void ARGBLumaColorTableRow_SSSE3(const uint8* src_argb, uint8* dst_argb, + int width, + const uint8* luma, uint32 lumacoeff) { + __asm { + push esi + push edi + mov eax, [esp + 8 + 4] /* src_argb */ + mov edi, [esp + 8 + 8] /* dst_argb */ + mov ecx, [esp + 8 + 12] /* width */ + movd xmm2, dword ptr [esp + 8 + 16] // luma table + movd xmm3, dword ptr [esp + 8 + 20] // lumacoeff + pshufd xmm2, xmm2, 0 + pshufd xmm3, xmm3, 0 + pcmpeqb xmm4, xmm4 // generate mask 0xff00ff00 + psllw xmm4, 8 + pxor xmm5, xmm5 + + // 4 pixel loop. + convertloop: + movdqu xmm0, qword ptr [eax] // generate luma ptr + pmaddubsw xmm0, xmm3 + phaddw xmm0, xmm0 + pand xmm0, xmm4 // mask out low bits + punpcklwd xmm0, xmm5 + paddd xmm0, xmm2 // add table base + movd esi, xmm0 + pshufd xmm0, xmm0, 0x39 // 00111001 to rotate right 32 + + movzx edx, byte ptr [eax] + movzx edx, byte ptr [esi + edx] + mov byte ptr [edi], dl + movzx edx, byte ptr [eax + 1] + movzx edx, byte ptr [esi + edx] + mov byte ptr [edi + 1], dl + movzx edx, byte ptr [eax + 2] + movzx edx, byte ptr [esi + edx] + mov byte ptr [edi + 2], dl + movzx edx, byte ptr [eax + 3] // copy alpha. + mov byte ptr [edi + 3], dl + + movd esi, xmm0 + pshufd xmm0, xmm0, 0x39 // 00111001 to rotate right 32 + + movzx edx, byte ptr [eax + 4] + movzx edx, byte ptr [esi + edx] + mov byte ptr [edi + 4], dl + movzx edx, byte ptr [eax + 5] + movzx edx, byte ptr [esi + edx] + mov byte ptr [edi + 5], dl + movzx edx, byte ptr [eax + 6] + movzx edx, byte ptr [esi + edx] + mov byte ptr [edi + 6], dl + movzx edx, byte ptr [eax + 7] // copy alpha. + mov byte ptr [edi + 7], dl + + movd esi, xmm0 + pshufd xmm0, xmm0, 0x39 // 00111001 to rotate right 32 + + movzx edx, byte ptr [eax + 8] + movzx edx, byte ptr [esi + edx] + mov byte ptr [edi + 8], dl + movzx edx, byte ptr [eax + 9] + movzx edx, byte ptr [esi + edx] + mov byte ptr [edi + 9], dl + movzx edx, byte ptr [eax + 10] + movzx edx, byte ptr [esi + edx] + mov byte ptr [edi + 10], dl + movzx edx, byte ptr [eax + 11] // copy alpha. + mov byte ptr [edi + 11], dl + + movd esi, xmm0 + + movzx edx, byte ptr [eax + 12] + movzx edx, byte ptr [esi + edx] + mov byte ptr [edi + 12], dl + movzx edx, byte ptr [eax + 13] + movzx edx, byte ptr [esi + edx] + mov byte ptr [edi + 13], dl + movzx edx, byte ptr [eax + 14] + movzx edx, byte ptr [esi + edx] + mov byte ptr [edi + 14], dl + movzx edx, byte ptr [eax + 15] // copy alpha. + mov byte ptr [edi + 15], dl + + lea eax, [eax + 16] + lea edi, [edi + 16] + sub ecx, 4 + jg convertloop + + pop edi + pop esi + ret + } +} +#endif // HAS_ARGBLUMACOLORTABLEROW_SSSE3 + +#endif // defined(_M_X64) +#endif // !defined(LIBYUV_DISABLE_X86) && (defined(_M_IX86) || defined(_M_X64)) + +#ifdef __cplusplus +} // extern "C" +} // namespace libyuv +#endif diff --git a/media/libaom/src/third_party/libyuv/source/row_x86.asm b/media/libaom/src/third_party/libyuv/source/row_x86.asm new file mode 100644 index 000000000..0cb326f8e --- /dev/null +++ b/media/libaom/src/third_party/libyuv/source/row_x86.asm @@ -0,0 +1,146 @@ +; +; Copyright 2012 The LibYuv Project Authors. All rights reserved. +; +; Use of this source code is governed by a BSD-style license +; that can be found in the LICENSE file in the root of the source +; tree. An additional intellectual property rights grant can be found +; in the file PATENTS. All contributing project authors may +; be found in the AUTHORS file in the root of the source tree. +; + +%ifdef __YASM_VERSION_ID__ +%if __YASM_VERSION_ID__ < 01020000h +%error AVX2 is supported only by yasm 1.2.0 or later. +%endif +%endif +%include "x86inc.asm" + +SECTION .text + +; cglobal numeric constants are parameters, gpr regs, mm regs + +; void YUY2ToYRow_SSE2(const uint8* src_yuy2, uint8* dst_y, int pix) + +%macro YUY2TOYROW 2-3 +cglobal %1ToYRow%3, 3, 3, 3, src_yuy2, dst_y, pix +%ifidn %1,YUY2 + pcmpeqb m2, m2, m2 ; generate mask 0x00ff00ff + psrlw m2, m2, 8 +%endif + + ALIGN 4 +.convertloop: + mov%2 m0, [src_yuy2q] + mov%2 m1, [src_yuy2q + mmsize] + lea src_yuy2q, [src_yuy2q + mmsize * 2] +%ifidn %1,YUY2 + pand m0, m0, m2 ; YUY2 even bytes are Y + pand m1, m1, m2 +%else + psrlw m0, m0, 8 ; UYVY odd bytes are Y + psrlw m1, m1, 8 +%endif + packuswb m0, m0, m1 +%if cpuflag(AVX2) + vpermq m0, m0, 0xd8 +%endif + sub pixd, mmsize + mov%2 [dst_yq], m0 + lea dst_yq, [dst_yq + mmsize] + jg .convertloop + REP_RET +%endmacro + +; TODO(fbarchard): Remove MMX. Add SSSE3 pshufb version. +INIT_MMX MMX +YUY2TOYROW YUY2,a, +YUY2TOYROW YUY2,u,_Unaligned +YUY2TOYROW UYVY,a, +YUY2TOYROW UYVY,u,_Unaligned +INIT_XMM SSE2 +YUY2TOYROW YUY2,a, +YUY2TOYROW YUY2,u,_Unaligned +YUY2TOYROW UYVY,a, +YUY2TOYROW UYVY,u,_Unaligned +INIT_YMM AVX2 +YUY2TOYROW YUY2,a, +YUY2TOYROW UYVY,a, + +; void SplitUVRow_SSE2(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int pix) + +%macro SplitUVRow 1-2 +cglobal SplitUVRow%2, 4, 4, 5, src_uv, dst_u, dst_v, pix + pcmpeqb m4, m4, m4 ; generate mask 0x00ff00ff + psrlw m4, m4, 8 + sub dst_vq, dst_uq + + ALIGN 4 +.convertloop: + mov%1 m0, [src_uvq] + mov%1 m1, [src_uvq + mmsize] + lea src_uvq, [src_uvq + mmsize * 2] + psrlw m2, m0, 8 ; odd bytes + psrlw m3, m1, 8 + pand m0, m0, m4 ; even bytes + pand m1, m1, m4 + packuswb m0, m0, m1 + packuswb m2, m2, m3 +%if cpuflag(AVX2) + vpermq m0, m0, 0xd8 + vpermq m2, m2, 0xd8 +%endif + mov%1 [dst_uq], m0 + mov%1 [dst_uq + dst_vq], m2 + lea dst_uq, [dst_uq + mmsize] + sub pixd, mmsize + jg .convertloop + REP_RET +%endmacro + +INIT_MMX MMX +SplitUVRow a, +SplitUVRow u,_Unaligned +INIT_XMM SSE2 +SplitUVRow a, +SplitUVRow u,_Unaligned +INIT_YMM AVX2 +SplitUVRow a, + +; void MergeUVRow_SSE2(const uint8* src_u, const uint8* src_v, uint8* dst_uv, +; int width); + +%macro MergeUVRow_ 1-2 +cglobal MergeUVRow_%2, 4, 4, 3, src_u, src_v, dst_uv, pix + sub src_vq, src_uq + + ALIGN 4 +.convertloop: + mov%1 m0, [src_uq] + mov%1 m1, [src_vq] + lea src_uq, [src_uq + mmsize] + punpcklbw m2, m0, m1 // first 8 UV pairs + punpckhbw m0, m0, m1 // next 8 UV pairs +%if cpuflag(AVX2) + vperm2i128 m1, m2, m0, 0x20 // low 128 of ymm2 and low 128 of ymm0 + vperm2i128 m2, m2, m0, 0x31 // high 128 of ymm2 and high 128 of ymm0 + mov%1 [dst_uvq], m1 + mov%1 [dst_uvq + mmsize], m2 +%else + mov%1 [dst_uvq], m2 + mov%1 [dst_uvq + mmsize], m0 +%endif + lea dst_uvq, [dst_uvq + mmsize * 2] + sub pixd, mmsize + jg .convertloop + REP_RET +%endmacro + +INIT_MMX MMX +MergeUVRow_ a, +MergeUVRow_ u,_Unaligned +INIT_XMM SSE2 +MergeUVRow_ a, +MergeUVRow_ u,_Unaligned +INIT_YMM AVX2 +MergeUVRow_ a, + diff --git a/media/libaom/src/third_party/libyuv/source/scale.cc b/media/libaom/src/third_party/libyuv/source/scale.cc new file mode 100644 index 000000000..0a01304c4 --- /dev/null +++ b/media/libaom/src/third_party/libyuv/source/scale.cc @@ -0,0 +1,1689 @@ +/* + * Copyright 2011 The LibYuv Project Authors. All rights reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#include "libyuv/scale.h" + +#include <assert.h> +#include <string.h> + +#include "libyuv/cpu_id.h" +#include "libyuv/planar_functions.h" // For CopyPlane +#include "libyuv/row.h" +#include "libyuv/scale_row.h" + +#ifdef __cplusplus +namespace libyuv { +extern "C" { +#endif + +static __inline int Abs(int v) { + return v >= 0 ? v : -v; +} + +#define SUBSAMPLE(v, a, s) (v < 0) ? (-((-v + a) >> s)) : ((v + a) >> s) + +// Scale plane, 1/2 +// This is an optimized version for scaling down a plane to 1/2 of +// its original size. + +static void ScalePlaneDown2(int src_width, int src_height, + int dst_width, int dst_height, + int src_stride, int dst_stride, + const uint8* src_ptr, uint8* dst_ptr, + enum FilterMode filtering) { + int y; + void (*ScaleRowDown2)(const uint8* src_ptr, ptrdiff_t src_stride, + uint8* dst_ptr, int dst_width) = + filtering == kFilterNone ? ScaleRowDown2_C : + (filtering == kFilterLinear ? ScaleRowDown2Linear_C : ScaleRowDown2Box_C); + int row_stride = src_stride << 1; + if (!filtering) { + src_ptr += src_stride; // Point to odd rows. + src_stride = 0; + } + +#if defined(HAS_SCALEROWDOWN2_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + ScaleRowDown2 = filtering == kFilterNone ? ScaleRowDown2_Any_NEON : + (filtering == kFilterLinear ? ScaleRowDown2Linear_Any_NEON : + ScaleRowDown2Box_Any_NEON); + if (IS_ALIGNED(dst_width, 16)) { + ScaleRowDown2 = filtering == kFilterNone ? ScaleRowDown2_NEON : + (filtering == kFilterLinear ? ScaleRowDown2Linear_NEON : + ScaleRowDown2Box_NEON); + } + } +#endif +#if defined(HAS_SCALEROWDOWN2_SSE2) + if (TestCpuFlag(kCpuHasSSE2)) { + ScaleRowDown2 = filtering == kFilterNone ? ScaleRowDown2_Any_SSE2 : + (filtering == kFilterLinear ? ScaleRowDown2Linear_Any_SSE2 : + ScaleRowDown2Box_Any_SSE2); + if (IS_ALIGNED(dst_width, 16)) { + ScaleRowDown2 = filtering == kFilterNone ? ScaleRowDown2_SSE2 : + (filtering == kFilterLinear ? ScaleRowDown2Linear_SSE2 : + ScaleRowDown2Box_SSE2); + } + } +#endif +#if defined(HAS_SCALEROWDOWN2_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + ScaleRowDown2 = filtering == kFilterNone ? ScaleRowDown2_Any_AVX2 : + (filtering == kFilterLinear ? ScaleRowDown2Linear_Any_AVX2 : + ScaleRowDown2Box_Any_AVX2); + if (IS_ALIGNED(dst_width, 32)) { + ScaleRowDown2 = filtering == kFilterNone ? ScaleRowDown2_AVX2 : + (filtering == kFilterLinear ? ScaleRowDown2Linear_AVX2 : + ScaleRowDown2Box_AVX2); + } + } +#endif +#if defined(HAS_SCALEROWDOWN2_MIPS_DSPR2) + if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(src_ptr, 4) && + IS_ALIGNED(src_stride, 4) && IS_ALIGNED(row_stride, 4) && + IS_ALIGNED(dst_ptr, 4) && IS_ALIGNED(dst_stride, 4)) { + ScaleRowDown2 = filtering ? + ScaleRowDown2Box_MIPS_DSPR2 : ScaleRowDown2_MIPS_DSPR2; + } +#endif + + if (filtering == kFilterLinear) { + src_stride = 0; + } + // TODO(fbarchard): Loop through source height to allow odd height. + for (y = 0; y < dst_height; ++y) { + ScaleRowDown2(src_ptr, src_stride, dst_ptr, dst_width); + src_ptr += row_stride; + dst_ptr += dst_stride; + } +} + +static void ScalePlaneDown2_16(int src_width, int src_height, + int dst_width, int dst_height, + int src_stride, int dst_stride, + const uint16* src_ptr, uint16* dst_ptr, + enum FilterMode filtering) { + int y; + void (*ScaleRowDown2)(const uint16* src_ptr, ptrdiff_t src_stride, + uint16* dst_ptr, int dst_width) = + filtering == kFilterNone ? ScaleRowDown2_16_C : + (filtering == kFilterLinear ? ScaleRowDown2Linear_16_C : + ScaleRowDown2Box_16_C); + int row_stride = src_stride << 1; + if (!filtering) { + src_ptr += src_stride; // Point to odd rows. + src_stride = 0; + } + +#if defined(HAS_SCALEROWDOWN2_16_NEON) + if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(dst_width, 16)) { + ScaleRowDown2 = filtering ? ScaleRowDown2Box_16_NEON : + ScaleRowDown2_16_NEON; + } +#endif +#if defined(HAS_SCALEROWDOWN2_16_SSE2) + if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 16)) { + ScaleRowDown2 = filtering == kFilterNone ? ScaleRowDown2_16_SSE2 : + (filtering == kFilterLinear ? ScaleRowDown2Linear_16_SSE2 : + ScaleRowDown2Box_16_SSE2); + } +#endif +#if defined(HAS_SCALEROWDOWN2_16_MIPS_DSPR2) + if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(src_ptr, 4) && + IS_ALIGNED(src_stride, 4) && IS_ALIGNED(row_stride, 4) && + IS_ALIGNED(dst_ptr, 4) && IS_ALIGNED(dst_stride, 4)) { + ScaleRowDown2 = filtering ? + ScaleRowDown2Box_16_MIPS_DSPR2 : ScaleRowDown2_16_MIPS_DSPR2; + } +#endif + + if (filtering == kFilterLinear) { + src_stride = 0; + } + // TODO(fbarchard): Loop through source height to allow odd height. + for (y = 0; y < dst_height; ++y) { + ScaleRowDown2(src_ptr, src_stride, dst_ptr, dst_width); + src_ptr += row_stride; + dst_ptr += dst_stride; + } +} + +// Scale plane, 1/4 +// This is an optimized version for scaling down a plane to 1/4 of +// its original size. + +static void ScalePlaneDown4(int src_width, int src_height, + int dst_width, int dst_height, + int src_stride, int dst_stride, + const uint8* src_ptr, uint8* dst_ptr, + enum FilterMode filtering) { + int y; + void (*ScaleRowDown4)(const uint8* src_ptr, ptrdiff_t src_stride, + uint8* dst_ptr, int dst_width) = + filtering ? ScaleRowDown4Box_C : ScaleRowDown4_C; + int row_stride = src_stride << 2; + if (!filtering) { + src_ptr += src_stride * 2; // Point to row 2. + src_stride = 0; + } +#if defined(HAS_SCALEROWDOWN4_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + ScaleRowDown4 = filtering ? + ScaleRowDown4Box_Any_NEON : ScaleRowDown4_Any_NEON; + if (IS_ALIGNED(dst_width, 8)) { + ScaleRowDown4 = filtering ? ScaleRowDown4Box_NEON : ScaleRowDown4_NEON; + } + } +#endif +#if defined(HAS_SCALEROWDOWN4_SSE2) + if (TestCpuFlag(kCpuHasSSE2)) { + ScaleRowDown4 = filtering ? + ScaleRowDown4Box_Any_SSE2 : ScaleRowDown4_Any_SSE2; + if (IS_ALIGNED(dst_width, 8)) { + ScaleRowDown4 = filtering ? ScaleRowDown4Box_SSE2 : ScaleRowDown4_SSE2; + } + } +#endif +#if defined(HAS_SCALEROWDOWN4_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + ScaleRowDown4 = filtering ? + ScaleRowDown4Box_Any_AVX2 : ScaleRowDown4_Any_AVX2; + if (IS_ALIGNED(dst_width, 16)) { + ScaleRowDown4 = filtering ? ScaleRowDown4Box_AVX2 : ScaleRowDown4_AVX2; + } + } +#endif +#if defined(HAS_SCALEROWDOWN4_MIPS_DSPR2) + if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(row_stride, 4) && + IS_ALIGNED(src_ptr, 4) && IS_ALIGNED(src_stride, 4) && + IS_ALIGNED(dst_ptr, 4) && IS_ALIGNED(dst_stride, 4)) { + ScaleRowDown4 = filtering ? + ScaleRowDown4Box_MIPS_DSPR2 : ScaleRowDown4_MIPS_DSPR2; + } +#endif + + if (filtering == kFilterLinear) { + src_stride = 0; + } + for (y = 0; y < dst_height; ++y) { + ScaleRowDown4(src_ptr, src_stride, dst_ptr, dst_width); + src_ptr += row_stride; + dst_ptr += dst_stride; + } +} + +static void ScalePlaneDown4_16(int src_width, int src_height, + int dst_width, int dst_height, + int src_stride, int dst_stride, + const uint16* src_ptr, uint16* dst_ptr, + enum FilterMode filtering) { + int y; + void (*ScaleRowDown4)(const uint16* src_ptr, ptrdiff_t src_stride, + uint16* dst_ptr, int dst_width) = + filtering ? ScaleRowDown4Box_16_C : ScaleRowDown4_16_C; + int row_stride = src_stride << 2; + if (!filtering) { + src_ptr += src_stride * 2; // Point to row 2. + src_stride = 0; + } +#if defined(HAS_SCALEROWDOWN4_16_NEON) + if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(dst_width, 8)) { + ScaleRowDown4 = filtering ? ScaleRowDown4Box_16_NEON : + ScaleRowDown4_16_NEON; + } +#endif +#if defined(HAS_SCALEROWDOWN4_16_SSE2) + if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 8)) { + ScaleRowDown4 = filtering ? ScaleRowDown4Box_16_SSE2 : + ScaleRowDown4_16_SSE2; + } +#endif +#if defined(HAS_SCALEROWDOWN4_16_MIPS_DSPR2) + if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(row_stride, 4) && + IS_ALIGNED(src_ptr, 4) && IS_ALIGNED(src_stride, 4) && + IS_ALIGNED(dst_ptr, 4) && IS_ALIGNED(dst_stride, 4)) { + ScaleRowDown4 = filtering ? + ScaleRowDown4Box_16_MIPS_DSPR2 : ScaleRowDown4_16_MIPS_DSPR2; + } +#endif + + if (filtering == kFilterLinear) { + src_stride = 0; + } + for (y = 0; y < dst_height; ++y) { + ScaleRowDown4(src_ptr, src_stride, dst_ptr, dst_width); + src_ptr += row_stride; + dst_ptr += dst_stride; + } +} + +// Scale plane down, 3/4 + +static void ScalePlaneDown34(int src_width, int src_height, + int dst_width, int dst_height, + int src_stride, int dst_stride, + const uint8* src_ptr, uint8* dst_ptr, + enum FilterMode filtering) { + int y; + void (*ScaleRowDown34_0)(const uint8* src_ptr, ptrdiff_t src_stride, + uint8* dst_ptr, int dst_width); + void (*ScaleRowDown34_1)(const uint8* src_ptr, ptrdiff_t src_stride, + uint8* dst_ptr, int dst_width); + const int filter_stride = (filtering == kFilterLinear) ? 0 : src_stride; + assert(dst_width % 3 == 0); + if (!filtering) { + ScaleRowDown34_0 = ScaleRowDown34_C; + ScaleRowDown34_1 = ScaleRowDown34_C; + } else { + ScaleRowDown34_0 = ScaleRowDown34_0_Box_C; + ScaleRowDown34_1 = ScaleRowDown34_1_Box_C; + } +#if defined(HAS_SCALEROWDOWN34_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + if (!filtering) { + ScaleRowDown34_0 = ScaleRowDown34_Any_NEON; + ScaleRowDown34_1 = ScaleRowDown34_Any_NEON; + } else { + ScaleRowDown34_0 = ScaleRowDown34_0_Box_Any_NEON; + ScaleRowDown34_1 = ScaleRowDown34_1_Box_Any_NEON; + } + if (dst_width % 24 == 0) { + if (!filtering) { + ScaleRowDown34_0 = ScaleRowDown34_NEON; + ScaleRowDown34_1 = ScaleRowDown34_NEON; + } else { + ScaleRowDown34_0 = ScaleRowDown34_0_Box_NEON; + ScaleRowDown34_1 = ScaleRowDown34_1_Box_NEON; + } + } + } +#endif +#if defined(HAS_SCALEROWDOWN34_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + if (!filtering) { + ScaleRowDown34_0 = ScaleRowDown34_Any_SSSE3; + ScaleRowDown34_1 = ScaleRowDown34_Any_SSSE3; + } else { + ScaleRowDown34_0 = ScaleRowDown34_0_Box_Any_SSSE3; + ScaleRowDown34_1 = ScaleRowDown34_1_Box_Any_SSSE3; + } + if (dst_width % 24 == 0) { + if (!filtering) { + ScaleRowDown34_0 = ScaleRowDown34_SSSE3; + ScaleRowDown34_1 = ScaleRowDown34_SSSE3; + } else { + ScaleRowDown34_0 = ScaleRowDown34_0_Box_SSSE3; + ScaleRowDown34_1 = ScaleRowDown34_1_Box_SSSE3; + } + } + } +#endif +#if defined(HAS_SCALEROWDOWN34_MIPS_DSPR2) + if (TestCpuFlag(kCpuHasMIPS_DSPR2) && (dst_width % 24 == 0) && + IS_ALIGNED(src_ptr, 4) && IS_ALIGNED(src_stride, 4) && + IS_ALIGNED(dst_ptr, 4) && IS_ALIGNED(dst_stride, 4)) { + if (!filtering) { + ScaleRowDown34_0 = ScaleRowDown34_MIPS_DSPR2; + ScaleRowDown34_1 = ScaleRowDown34_MIPS_DSPR2; + } else { + ScaleRowDown34_0 = ScaleRowDown34_0_Box_MIPS_DSPR2; + ScaleRowDown34_1 = ScaleRowDown34_1_Box_MIPS_DSPR2; + } + } +#endif + + for (y = 0; y < dst_height - 2; y += 3) { + ScaleRowDown34_0(src_ptr, filter_stride, dst_ptr, dst_width); + src_ptr += src_stride; + dst_ptr += dst_stride; + ScaleRowDown34_1(src_ptr, filter_stride, dst_ptr, dst_width); + src_ptr += src_stride; + dst_ptr += dst_stride; + ScaleRowDown34_0(src_ptr + src_stride, -filter_stride, + dst_ptr, dst_width); + src_ptr += src_stride * 2; + dst_ptr += dst_stride; + } + + // Remainder 1 or 2 rows with last row vertically unfiltered + if ((dst_height % 3) == 2) { + ScaleRowDown34_0(src_ptr, filter_stride, dst_ptr, dst_width); + src_ptr += src_stride; + dst_ptr += dst_stride; + ScaleRowDown34_1(src_ptr, 0, dst_ptr, dst_width); + } else if ((dst_height % 3) == 1) { + ScaleRowDown34_0(src_ptr, 0, dst_ptr, dst_width); + } +} + +static void ScalePlaneDown34_16(int src_width, int src_height, + int dst_width, int dst_height, + int src_stride, int dst_stride, + const uint16* src_ptr, uint16* dst_ptr, + enum FilterMode filtering) { + int y; + void (*ScaleRowDown34_0)(const uint16* src_ptr, ptrdiff_t src_stride, + uint16* dst_ptr, int dst_width); + void (*ScaleRowDown34_1)(const uint16* src_ptr, ptrdiff_t src_stride, + uint16* dst_ptr, int dst_width); + const int filter_stride = (filtering == kFilterLinear) ? 0 : src_stride; + assert(dst_width % 3 == 0); + if (!filtering) { + ScaleRowDown34_0 = ScaleRowDown34_16_C; + ScaleRowDown34_1 = ScaleRowDown34_16_C; + } else { + ScaleRowDown34_0 = ScaleRowDown34_0_Box_16_C; + ScaleRowDown34_1 = ScaleRowDown34_1_Box_16_C; + } +#if defined(HAS_SCALEROWDOWN34_16_NEON) + if (TestCpuFlag(kCpuHasNEON) && (dst_width % 24 == 0)) { + if (!filtering) { + ScaleRowDown34_0 = ScaleRowDown34_16_NEON; + ScaleRowDown34_1 = ScaleRowDown34_16_NEON; + } else { + ScaleRowDown34_0 = ScaleRowDown34_0_Box_16_NEON; + ScaleRowDown34_1 = ScaleRowDown34_1_Box_16_NEON; + } + } +#endif +#if defined(HAS_SCALEROWDOWN34_16_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3) && (dst_width % 24 == 0)) { + if (!filtering) { + ScaleRowDown34_0 = ScaleRowDown34_16_SSSE3; + ScaleRowDown34_1 = ScaleRowDown34_16_SSSE3; + } else { + ScaleRowDown34_0 = ScaleRowDown34_0_Box_16_SSSE3; + ScaleRowDown34_1 = ScaleRowDown34_1_Box_16_SSSE3; + } + } +#endif +#if defined(HAS_SCALEROWDOWN34_16_MIPS_DSPR2) + if (TestCpuFlag(kCpuHasMIPS_DSPR2) && (dst_width % 24 == 0) && + IS_ALIGNED(src_ptr, 4) && IS_ALIGNED(src_stride, 4) && + IS_ALIGNED(dst_ptr, 4) && IS_ALIGNED(dst_stride, 4)) { + if (!filtering) { + ScaleRowDown34_0 = ScaleRowDown34_16_MIPS_DSPR2; + ScaleRowDown34_1 = ScaleRowDown34_16_MIPS_DSPR2; + } else { + ScaleRowDown34_0 = ScaleRowDown34_0_Box_16_MIPS_DSPR2; + ScaleRowDown34_1 = ScaleRowDown34_1_Box_16_MIPS_DSPR2; + } + } +#endif + + for (y = 0; y < dst_height - 2; y += 3) { + ScaleRowDown34_0(src_ptr, filter_stride, dst_ptr, dst_width); + src_ptr += src_stride; + dst_ptr += dst_stride; + ScaleRowDown34_1(src_ptr, filter_stride, dst_ptr, dst_width); + src_ptr += src_stride; + dst_ptr += dst_stride; + ScaleRowDown34_0(src_ptr + src_stride, -filter_stride, + dst_ptr, dst_width); + src_ptr += src_stride * 2; + dst_ptr += dst_stride; + } + + // Remainder 1 or 2 rows with last row vertically unfiltered + if ((dst_height % 3) == 2) { + ScaleRowDown34_0(src_ptr, filter_stride, dst_ptr, dst_width); + src_ptr += src_stride; + dst_ptr += dst_stride; + ScaleRowDown34_1(src_ptr, 0, dst_ptr, dst_width); + } else if ((dst_height % 3) == 1) { + ScaleRowDown34_0(src_ptr, 0, dst_ptr, dst_width); + } +} + + +// Scale plane, 3/8 +// This is an optimized version for scaling down a plane to 3/8 +// of its original size. +// +// Uses box filter arranges like this +// aaabbbcc -> abc +// aaabbbcc def +// aaabbbcc ghi +// dddeeeff +// dddeeeff +// dddeeeff +// ggghhhii +// ggghhhii +// Boxes are 3x3, 2x3, 3x2 and 2x2 + +static void ScalePlaneDown38(int src_width, int src_height, + int dst_width, int dst_height, + int src_stride, int dst_stride, + const uint8* src_ptr, uint8* dst_ptr, + enum FilterMode filtering) { + int y; + void (*ScaleRowDown38_3)(const uint8* src_ptr, ptrdiff_t src_stride, + uint8* dst_ptr, int dst_width); + void (*ScaleRowDown38_2)(const uint8* src_ptr, ptrdiff_t src_stride, + uint8* dst_ptr, int dst_width); + const int filter_stride = (filtering == kFilterLinear) ? 0 : src_stride; + assert(dst_width % 3 == 0); + if (!filtering) { + ScaleRowDown38_3 = ScaleRowDown38_C; + ScaleRowDown38_2 = ScaleRowDown38_C; + } else { + ScaleRowDown38_3 = ScaleRowDown38_3_Box_C; + ScaleRowDown38_2 = ScaleRowDown38_2_Box_C; + } + +#if defined(HAS_SCALEROWDOWN38_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + if (!filtering) { + ScaleRowDown38_3 = ScaleRowDown38_Any_NEON; + ScaleRowDown38_2 = ScaleRowDown38_Any_NEON; + } else { + ScaleRowDown38_3 = ScaleRowDown38_3_Box_Any_NEON; + ScaleRowDown38_2 = ScaleRowDown38_2_Box_Any_NEON; + } + if (dst_width % 12 == 0) { + if (!filtering) { + ScaleRowDown38_3 = ScaleRowDown38_NEON; + ScaleRowDown38_2 = ScaleRowDown38_NEON; + } else { + ScaleRowDown38_3 = ScaleRowDown38_3_Box_NEON; + ScaleRowDown38_2 = ScaleRowDown38_2_Box_NEON; + } + } + } +#endif +#if defined(HAS_SCALEROWDOWN38_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + if (!filtering) { + ScaleRowDown38_3 = ScaleRowDown38_Any_SSSE3; + ScaleRowDown38_2 = ScaleRowDown38_Any_SSSE3; + } else { + ScaleRowDown38_3 = ScaleRowDown38_3_Box_Any_SSSE3; + ScaleRowDown38_2 = ScaleRowDown38_2_Box_Any_SSSE3; + } + if (dst_width % 12 == 0 && !filtering) { + ScaleRowDown38_3 = ScaleRowDown38_SSSE3; + ScaleRowDown38_2 = ScaleRowDown38_SSSE3; + } + if (dst_width % 6 == 0 && filtering) { + ScaleRowDown38_3 = ScaleRowDown38_3_Box_SSSE3; + ScaleRowDown38_2 = ScaleRowDown38_2_Box_SSSE3; + } + } +#endif +#if defined(HAS_SCALEROWDOWN38_MIPS_DSPR2) + if (TestCpuFlag(kCpuHasMIPS_DSPR2) && (dst_width % 12 == 0) && + IS_ALIGNED(src_ptr, 4) && IS_ALIGNED(src_stride, 4) && + IS_ALIGNED(dst_ptr, 4) && IS_ALIGNED(dst_stride, 4)) { + if (!filtering) { + ScaleRowDown38_3 = ScaleRowDown38_MIPS_DSPR2; + ScaleRowDown38_2 = ScaleRowDown38_MIPS_DSPR2; + } else { + ScaleRowDown38_3 = ScaleRowDown38_3_Box_MIPS_DSPR2; + ScaleRowDown38_2 = ScaleRowDown38_2_Box_MIPS_DSPR2; + } + } +#endif + + for (y = 0; y < dst_height - 2; y += 3) { + ScaleRowDown38_3(src_ptr, filter_stride, dst_ptr, dst_width); + src_ptr += src_stride * 3; + dst_ptr += dst_stride; + ScaleRowDown38_3(src_ptr, filter_stride, dst_ptr, dst_width); + src_ptr += src_stride * 3; + dst_ptr += dst_stride; + ScaleRowDown38_2(src_ptr, filter_stride, dst_ptr, dst_width); + src_ptr += src_stride * 2; + dst_ptr += dst_stride; + } + + // Remainder 1 or 2 rows with last row vertically unfiltered + if ((dst_height % 3) == 2) { + ScaleRowDown38_3(src_ptr, filter_stride, dst_ptr, dst_width); + src_ptr += src_stride * 3; + dst_ptr += dst_stride; + ScaleRowDown38_3(src_ptr, 0, dst_ptr, dst_width); + } else if ((dst_height % 3) == 1) { + ScaleRowDown38_3(src_ptr, 0, dst_ptr, dst_width); + } +} + +static void ScalePlaneDown38_16(int src_width, int src_height, + int dst_width, int dst_height, + int src_stride, int dst_stride, + const uint16* src_ptr, uint16* dst_ptr, + enum FilterMode filtering) { + int y; + void (*ScaleRowDown38_3)(const uint16* src_ptr, ptrdiff_t src_stride, + uint16* dst_ptr, int dst_width); + void (*ScaleRowDown38_2)(const uint16* src_ptr, ptrdiff_t src_stride, + uint16* dst_ptr, int dst_width); + const int filter_stride = (filtering == kFilterLinear) ? 0 : src_stride; + assert(dst_width % 3 == 0); + if (!filtering) { + ScaleRowDown38_3 = ScaleRowDown38_16_C; + ScaleRowDown38_2 = ScaleRowDown38_16_C; + } else { + ScaleRowDown38_3 = ScaleRowDown38_3_Box_16_C; + ScaleRowDown38_2 = ScaleRowDown38_2_Box_16_C; + } +#if defined(HAS_SCALEROWDOWN38_16_NEON) + if (TestCpuFlag(kCpuHasNEON) && (dst_width % 12 == 0)) { + if (!filtering) { + ScaleRowDown38_3 = ScaleRowDown38_16_NEON; + ScaleRowDown38_2 = ScaleRowDown38_16_NEON; + } else { + ScaleRowDown38_3 = ScaleRowDown38_3_Box_16_NEON; + ScaleRowDown38_2 = ScaleRowDown38_2_Box_16_NEON; + } + } +#endif +#if defined(HAS_SCALEROWDOWN38_16_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3) && (dst_width % 24 == 0)) { + if (!filtering) { + ScaleRowDown38_3 = ScaleRowDown38_16_SSSE3; + ScaleRowDown38_2 = ScaleRowDown38_16_SSSE3; + } else { + ScaleRowDown38_3 = ScaleRowDown38_3_Box_16_SSSE3; + ScaleRowDown38_2 = ScaleRowDown38_2_Box_16_SSSE3; + } + } +#endif +#if defined(HAS_SCALEROWDOWN38_16_MIPS_DSPR2) + if (TestCpuFlag(kCpuHasMIPS_DSPR2) && (dst_width % 12 == 0) && + IS_ALIGNED(src_ptr, 4) && IS_ALIGNED(src_stride, 4) && + IS_ALIGNED(dst_ptr, 4) && IS_ALIGNED(dst_stride, 4)) { + if (!filtering) { + ScaleRowDown38_3 = ScaleRowDown38_16_MIPS_DSPR2; + ScaleRowDown38_2 = ScaleRowDown38_16_MIPS_DSPR2; + } else { + ScaleRowDown38_3 = ScaleRowDown38_3_Box_16_MIPS_DSPR2; + ScaleRowDown38_2 = ScaleRowDown38_2_Box_16_MIPS_DSPR2; + } + } +#endif + + for (y = 0; y < dst_height - 2; y += 3) { + ScaleRowDown38_3(src_ptr, filter_stride, dst_ptr, dst_width); + src_ptr += src_stride * 3; + dst_ptr += dst_stride; + ScaleRowDown38_3(src_ptr, filter_stride, dst_ptr, dst_width); + src_ptr += src_stride * 3; + dst_ptr += dst_stride; + ScaleRowDown38_2(src_ptr, filter_stride, dst_ptr, dst_width); + src_ptr += src_stride * 2; + dst_ptr += dst_stride; + } + + // Remainder 1 or 2 rows with last row vertically unfiltered + if ((dst_height % 3) == 2) { + ScaleRowDown38_3(src_ptr, filter_stride, dst_ptr, dst_width); + src_ptr += src_stride * 3; + dst_ptr += dst_stride; + ScaleRowDown38_3(src_ptr, 0, dst_ptr, dst_width); + } else if ((dst_height % 3) == 1) { + ScaleRowDown38_3(src_ptr, 0, dst_ptr, dst_width); + } +} + +#define MIN1(x) ((x) < 1 ? 1 : (x)) + +static __inline uint32 SumPixels(int iboxwidth, const uint16* src_ptr) { + uint32 sum = 0u; + int x; + assert(iboxwidth > 0); + for (x = 0; x < iboxwidth; ++x) { + sum += src_ptr[x]; + } + return sum; +} + +static __inline uint32 SumPixels_16(int iboxwidth, const uint32* src_ptr) { + uint32 sum = 0u; + int x; + assert(iboxwidth > 0); + for (x = 0; x < iboxwidth; ++x) { + sum += src_ptr[x]; + } + return sum; +} + +static void ScaleAddCols2_C(int dst_width, int boxheight, int x, int dx, + const uint16* src_ptr, uint8* dst_ptr) { + int i; + int scaletbl[2]; + int minboxwidth = dx >> 16; + int* scaleptr = scaletbl - minboxwidth; + int boxwidth; + scaletbl[0] = 65536 / (MIN1(minboxwidth) * boxheight); + scaletbl[1] = 65536 / (MIN1(minboxwidth + 1) * boxheight); + for (i = 0; i < dst_width; ++i) { + int ix = x >> 16; + x += dx; + boxwidth = MIN1((x >> 16) - ix); + *dst_ptr++ = SumPixels(boxwidth, src_ptr + ix) * scaleptr[boxwidth] >> 16; + } +} + +static void ScaleAddCols2_16_C(int dst_width, int boxheight, int x, int dx, + const uint32* src_ptr, uint16* dst_ptr) { + int i; + int scaletbl[2]; + int minboxwidth = dx >> 16; + int* scaleptr = scaletbl - minboxwidth; + int boxwidth; + scaletbl[0] = 65536 / (MIN1(minboxwidth) * boxheight); + scaletbl[1] = 65536 / (MIN1(minboxwidth + 1) * boxheight); + for (i = 0; i < dst_width; ++i) { + int ix = x >> 16; + x += dx; + boxwidth = MIN1((x >> 16) - ix); + *dst_ptr++ = + SumPixels_16(boxwidth, src_ptr + ix) * scaleptr[boxwidth] >> 16; + } +} + +static void ScaleAddCols0_C(int dst_width, int boxheight, int x, int, + const uint16* src_ptr, uint8* dst_ptr) { + int scaleval = 65536 / boxheight; + int i; + src_ptr += (x >> 16); + for (i = 0; i < dst_width; ++i) { + *dst_ptr++ = src_ptr[i] * scaleval >> 16; + } +} + +static void ScaleAddCols1_C(int dst_width, int boxheight, int x, int dx, + const uint16* src_ptr, uint8* dst_ptr) { + int boxwidth = MIN1(dx >> 16); + int scaleval = 65536 / (boxwidth * boxheight); + int i; + x >>= 16; + for (i = 0; i < dst_width; ++i) { + *dst_ptr++ = SumPixels(boxwidth, src_ptr + x) * scaleval >> 16; + x += boxwidth; + } +} + +static void ScaleAddCols1_16_C(int dst_width, int boxheight, int x, int dx, + const uint32* src_ptr, uint16* dst_ptr) { + int boxwidth = MIN1(dx >> 16); + int scaleval = 65536 / (boxwidth * boxheight); + int i; + for (i = 0; i < dst_width; ++i) { + *dst_ptr++ = SumPixels_16(boxwidth, src_ptr + x) * scaleval >> 16; + x += boxwidth; + } +} + +// Scale plane down to any dimensions, with interpolation. +// (boxfilter). +// +// Same method as SimpleScale, which is fixed point, outputting +// one pixel of destination using fixed point (16.16) to step +// through source, sampling a box of pixel with simple +// averaging. +static void ScalePlaneBox(int src_width, int src_height, + int dst_width, int dst_height, + int src_stride, int dst_stride, + const uint8* src_ptr, uint8* dst_ptr) { + int j, k; + // Initial source x/y coordinate and step values as 16.16 fixed point. + int x = 0; + int y = 0; + int dx = 0; + int dy = 0; + const int max_y = (src_height << 16); + ScaleSlope(src_width, src_height, dst_width, dst_height, kFilterBox, + &x, &y, &dx, &dy); + src_width = Abs(src_width); + { + // Allocate a row buffer of uint16. + align_buffer_64(row16, src_width * 2); + void (*ScaleAddCols)(int dst_width, int boxheight, int x, int dx, + const uint16* src_ptr, uint8* dst_ptr) = + (dx & 0xffff) ? ScaleAddCols2_C: + ((dx != 0x10000) ? ScaleAddCols1_C : ScaleAddCols0_C); + void (*ScaleAddRow)(const uint8* src_ptr, uint16* dst_ptr, int src_width) = + ScaleAddRow_C; +#if defined(HAS_SCALEADDROW_SSE2) + if (TestCpuFlag(kCpuHasSSE2)) { + ScaleAddRow = ScaleAddRow_Any_SSE2; + if (IS_ALIGNED(src_width, 16)) { + ScaleAddRow = ScaleAddRow_SSE2; + } + } +#endif +#if defined(HAS_SCALEADDROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + ScaleAddRow = ScaleAddRow_Any_AVX2; + if (IS_ALIGNED(src_width, 32)) { + ScaleAddRow = ScaleAddRow_AVX2; + } + } +#endif +#if defined(HAS_SCALEADDROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + ScaleAddRow = ScaleAddRow_Any_NEON; + if (IS_ALIGNED(src_width, 16)) { + ScaleAddRow = ScaleAddRow_NEON; + } + } +#endif + + for (j = 0; j < dst_height; ++j) { + int boxheight; + int iy = y >> 16; + const uint8* src = src_ptr + iy * src_stride; + y += dy; + if (y > max_y) { + y = max_y; + } + boxheight = MIN1((y >> 16) - iy); + memset(row16, 0, src_width * 2); + for (k = 0; k < boxheight; ++k) { + ScaleAddRow(src, (uint16 *)(row16), src_width); + src += src_stride; + } + ScaleAddCols(dst_width, boxheight, x, dx, (uint16*)(row16), dst_ptr); + dst_ptr += dst_stride; + } + free_aligned_buffer_64(row16); + } +} + +static void ScalePlaneBox_16(int src_width, int src_height, + int dst_width, int dst_height, + int src_stride, int dst_stride, + const uint16* src_ptr, uint16* dst_ptr) { + int j, k; + // Initial source x/y coordinate and step values as 16.16 fixed point. + int x = 0; + int y = 0; + int dx = 0; + int dy = 0; + const int max_y = (src_height << 16); + ScaleSlope(src_width, src_height, dst_width, dst_height, kFilterBox, + &x, &y, &dx, &dy); + src_width = Abs(src_width); + { + // Allocate a row buffer of uint32. + align_buffer_64(row32, src_width * 4); + void (*ScaleAddCols)(int dst_width, int boxheight, int x, int dx, + const uint32* src_ptr, uint16* dst_ptr) = + (dx & 0xffff) ? ScaleAddCols2_16_C: ScaleAddCols1_16_C; + void (*ScaleAddRow)(const uint16* src_ptr, uint32* dst_ptr, int src_width) = + ScaleAddRow_16_C; + +#if defined(HAS_SCALEADDROW_16_SSE2) + if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(src_width, 16)) { + ScaleAddRow = ScaleAddRow_16_SSE2; + } +#endif + + for (j = 0; j < dst_height; ++j) { + int boxheight; + int iy = y >> 16; + const uint16* src = src_ptr + iy * src_stride; + y += dy; + if (y > max_y) { + y = max_y; + } + boxheight = MIN1((y >> 16) - iy); + memset(row32, 0, src_width * 4); + for (k = 0; k < boxheight; ++k) { + ScaleAddRow(src, (uint32 *)(row32), src_width); + src += src_stride; + } + ScaleAddCols(dst_width, boxheight, x, dx, (uint32*)(row32), dst_ptr); + dst_ptr += dst_stride; + } + free_aligned_buffer_64(row32); + } +} + +// Scale plane down with bilinear interpolation. +void ScalePlaneBilinearDown(int src_width, int src_height, + int dst_width, int dst_height, + int src_stride, int dst_stride, + const uint8* src_ptr, uint8* dst_ptr, + enum FilterMode filtering) { + // Initial source x/y coordinate and step values as 16.16 fixed point. + int x = 0; + int y = 0; + int dx = 0; + int dy = 0; + // TODO(fbarchard): Consider not allocating row buffer for kFilterLinear. + // Allocate a row buffer. + align_buffer_64(row, src_width); + + const int max_y = (src_height - 1) << 16; + int j; + void (*ScaleFilterCols)(uint8* dst_ptr, const uint8* src_ptr, + int dst_width, int x, int dx) = + (src_width >= 32768) ? ScaleFilterCols64_C : ScaleFilterCols_C; + void (*InterpolateRow)(uint8* dst_ptr, const uint8* src_ptr, + ptrdiff_t src_stride, int dst_width, int source_y_fraction) = + InterpolateRow_C; + ScaleSlope(src_width, src_height, dst_width, dst_height, filtering, + &x, &y, &dx, &dy); + src_width = Abs(src_width); + +#if defined(HAS_INTERPOLATEROW_SSE2) + if (TestCpuFlag(kCpuHasSSE2)) { + InterpolateRow = InterpolateRow_Any_SSE2; + if (IS_ALIGNED(src_width, 16)) { + InterpolateRow = InterpolateRow_SSE2; + } + } +#endif +#if defined(HAS_INTERPOLATEROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + InterpolateRow = InterpolateRow_Any_SSSE3; + if (IS_ALIGNED(src_width, 16)) { + InterpolateRow = InterpolateRow_SSSE3; + } + } +#endif +#if defined(HAS_INTERPOLATEROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + InterpolateRow = InterpolateRow_Any_AVX2; + if (IS_ALIGNED(src_width, 32)) { + InterpolateRow = InterpolateRow_AVX2; + } + } +#endif +#if defined(HAS_INTERPOLATEROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + InterpolateRow = InterpolateRow_Any_NEON; + if (IS_ALIGNED(src_width, 16)) { + InterpolateRow = InterpolateRow_NEON; + } + } +#endif +#if defined(HAS_INTERPOLATEROW_MIPS_DSPR2) + if (TestCpuFlag(kCpuHasMIPS_DSPR2)) { + InterpolateRow = InterpolateRow_Any_MIPS_DSPR2; + if (IS_ALIGNED(src_width, 4)) { + InterpolateRow = InterpolateRow_MIPS_DSPR2; + } + } +#endif + + +#if defined(HAS_SCALEFILTERCOLS_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3) && src_width < 32768) { + ScaleFilterCols = ScaleFilterCols_SSSE3; + } +#endif +#if defined(HAS_SCALEFILTERCOLS_NEON) + if (TestCpuFlag(kCpuHasNEON) && src_width < 32768) { + ScaleFilterCols = ScaleFilterCols_Any_NEON; + if (IS_ALIGNED(dst_width, 8)) { + ScaleFilterCols = ScaleFilterCols_NEON; + } + } +#endif + if (y > max_y) { + y = max_y; + } + + for (j = 0; j < dst_height; ++j) { + int yi = y >> 16; + const uint8* src = src_ptr + yi * src_stride; + if (filtering == kFilterLinear) { + ScaleFilterCols(dst_ptr, src, dst_width, x, dx); + } else { + int yf = (y >> 8) & 255; + InterpolateRow(row, src, src_stride, src_width, yf); + ScaleFilterCols(dst_ptr, row, dst_width, x, dx); + } + dst_ptr += dst_stride; + y += dy; + if (y > max_y) { + y = max_y; + } + } + free_aligned_buffer_64(row); +} + +void ScalePlaneBilinearDown_16(int src_width, int src_height, + int dst_width, int dst_height, + int src_stride, int dst_stride, + const uint16* src_ptr, uint16* dst_ptr, + enum FilterMode filtering) { + // Initial source x/y coordinate and step values as 16.16 fixed point. + int x = 0; + int y = 0; + int dx = 0; + int dy = 0; + // TODO(fbarchard): Consider not allocating row buffer for kFilterLinear. + // Allocate a row buffer. + align_buffer_64(row, src_width * 2); + + const int max_y = (src_height - 1) << 16; + int j; + void (*ScaleFilterCols)(uint16* dst_ptr, const uint16* src_ptr, + int dst_width, int x, int dx) = + (src_width >= 32768) ? ScaleFilterCols64_16_C : ScaleFilterCols_16_C; + void (*InterpolateRow)(uint16* dst_ptr, const uint16* src_ptr, + ptrdiff_t src_stride, int dst_width, int source_y_fraction) = + InterpolateRow_16_C; + ScaleSlope(src_width, src_height, dst_width, dst_height, filtering, + &x, &y, &dx, &dy); + src_width = Abs(src_width); + +#if defined(HAS_INTERPOLATEROW_16_SSE2) + if (TestCpuFlag(kCpuHasSSE2)) { + InterpolateRow = InterpolateRow_Any_16_SSE2; + if (IS_ALIGNED(src_width, 16)) { + InterpolateRow = InterpolateRow_16_SSE2; + } + } +#endif +#if defined(HAS_INTERPOLATEROW_16_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + InterpolateRow = InterpolateRow_Any_16_SSSE3; + if (IS_ALIGNED(src_width, 16)) { + InterpolateRow = InterpolateRow_16_SSSE3; + } + } +#endif +#if defined(HAS_INTERPOLATEROW_16_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + InterpolateRow = InterpolateRow_Any_16_AVX2; + if (IS_ALIGNED(src_width, 32)) { + InterpolateRow = InterpolateRow_16_AVX2; + } + } +#endif +#if defined(HAS_INTERPOLATEROW_16_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + InterpolateRow = InterpolateRow_Any_16_NEON; + if (IS_ALIGNED(src_width, 16)) { + InterpolateRow = InterpolateRow_16_NEON; + } + } +#endif +#if defined(HAS_INTERPOLATEROW_16_MIPS_DSPR2) + if (TestCpuFlag(kCpuHasMIPS_DSPR2)) { + InterpolateRow = InterpolateRow_Any_16_MIPS_DSPR2; + if (IS_ALIGNED(src_width, 4)) { + InterpolateRow = InterpolateRow_16_MIPS_DSPR2; + } + } +#endif + + +#if defined(HAS_SCALEFILTERCOLS_16_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3) && src_width < 32768) { + ScaleFilterCols = ScaleFilterCols_16_SSSE3; + } +#endif + if (y > max_y) { + y = max_y; + } + + for (j = 0; j < dst_height; ++j) { + int yi = y >> 16; + const uint16* src = src_ptr + yi * src_stride; + if (filtering == kFilterLinear) { + ScaleFilterCols(dst_ptr, src, dst_width, x, dx); + } else { + int yf = (y >> 8) & 255; + InterpolateRow((uint16*)row, src, src_stride, src_width, yf); + ScaleFilterCols(dst_ptr, (uint16*)row, dst_width, x, dx); + } + dst_ptr += dst_stride; + y += dy; + if (y > max_y) { + y = max_y; + } + } + free_aligned_buffer_64(row); +} + +// Scale up down with bilinear interpolation. +void ScalePlaneBilinearUp(int src_width, int src_height, + int dst_width, int dst_height, + int src_stride, int dst_stride, + const uint8* src_ptr, uint8* dst_ptr, + enum FilterMode filtering) { + int j; + // Initial source x/y coordinate and step values as 16.16 fixed point. + int x = 0; + int y = 0; + int dx = 0; + int dy = 0; + const int max_y = (src_height - 1) << 16; + void (*InterpolateRow)(uint8* dst_ptr, const uint8* src_ptr, + ptrdiff_t src_stride, int dst_width, int source_y_fraction) = + InterpolateRow_C; + void (*ScaleFilterCols)(uint8* dst_ptr, const uint8* src_ptr, + int dst_width, int x, int dx) = + filtering ? ScaleFilterCols_C : ScaleCols_C; + ScaleSlope(src_width, src_height, dst_width, dst_height, filtering, + &x, &y, &dx, &dy); + src_width = Abs(src_width); + +#if defined(HAS_INTERPOLATEROW_SSE2) + if (TestCpuFlag(kCpuHasSSE2)) { + InterpolateRow = InterpolateRow_Any_SSE2; + if (IS_ALIGNED(dst_width, 16)) { + InterpolateRow = InterpolateRow_SSE2; + } + } +#endif +#if defined(HAS_INTERPOLATEROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + InterpolateRow = InterpolateRow_Any_SSSE3; + if (IS_ALIGNED(dst_width, 16)) { + InterpolateRow = InterpolateRow_SSSE3; + } + } +#endif +#if defined(HAS_INTERPOLATEROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + InterpolateRow = InterpolateRow_Any_AVX2; + if (IS_ALIGNED(dst_width, 32)) { + InterpolateRow = InterpolateRow_AVX2; + } + } +#endif +#if defined(HAS_INTERPOLATEROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + InterpolateRow = InterpolateRow_Any_NEON; + if (IS_ALIGNED(dst_width, 16)) { + InterpolateRow = InterpolateRow_NEON; + } + } +#endif +#if defined(HAS_INTERPOLATEROW_MIPS_DSPR2) + if (TestCpuFlag(kCpuHasMIPS_DSPR2)) { + InterpolateRow = InterpolateRow_Any_MIPS_DSPR2; + if (IS_ALIGNED(dst_width, 4)) { + InterpolateRow = InterpolateRow_MIPS_DSPR2; + } + } +#endif + + if (filtering && src_width >= 32768) { + ScaleFilterCols = ScaleFilterCols64_C; + } +#if defined(HAS_SCALEFILTERCOLS_SSSE3) + if (filtering && TestCpuFlag(kCpuHasSSSE3) && src_width < 32768) { + ScaleFilterCols = ScaleFilterCols_SSSE3; + } +#endif +#if defined(HAS_SCALEFILTERCOLS_NEON) + if (filtering && TestCpuFlag(kCpuHasNEON) && src_width < 32768) { + ScaleFilterCols = ScaleFilterCols_Any_NEON; + if (IS_ALIGNED(dst_width, 8)) { + ScaleFilterCols = ScaleFilterCols_NEON; + } + } +#endif + if (!filtering && src_width * 2 == dst_width && x < 0x8000) { + ScaleFilterCols = ScaleColsUp2_C; +#if defined(HAS_SCALECOLS_SSE2) + if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 8)) { + ScaleFilterCols = ScaleColsUp2_SSE2; + } +#endif + } + + if (y > max_y) { + y = max_y; + } + { + int yi = y >> 16; + const uint8* src = src_ptr + yi * src_stride; + + // Allocate 2 row buffers. + const int kRowSize = (dst_width + 31) & ~31; + align_buffer_64(row, kRowSize * 2); + + uint8* rowptr = row; + int rowstride = kRowSize; + int lasty = yi; + + ScaleFilterCols(rowptr, src, dst_width, x, dx); + if (src_height > 1) { + src += src_stride; + } + ScaleFilterCols(rowptr + rowstride, src, dst_width, x, dx); + src += src_stride; + + for (j = 0; j < dst_height; ++j) { + yi = y >> 16; + if (yi != lasty) { + if (y > max_y) { + y = max_y; + yi = y >> 16; + src = src_ptr + yi * src_stride; + } + if (yi != lasty) { + ScaleFilterCols(rowptr, src, dst_width, x, dx); + rowptr += rowstride; + rowstride = -rowstride; + lasty = yi; + src += src_stride; + } + } + if (filtering == kFilterLinear) { + InterpolateRow(dst_ptr, rowptr, 0, dst_width, 0); + } else { + int yf = (y >> 8) & 255; + InterpolateRow(dst_ptr, rowptr, rowstride, dst_width, yf); + } + dst_ptr += dst_stride; + y += dy; + } + free_aligned_buffer_64(row); + } +} + +void ScalePlaneBilinearUp_16(int src_width, int src_height, + int dst_width, int dst_height, + int src_stride, int dst_stride, + const uint16* src_ptr, uint16* dst_ptr, + enum FilterMode filtering) { + int j; + // Initial source x/y coordinate and step values as 16.16 fixed point. + int x = 0; + int y = 0; + int dx = 0; + int dy = 0; + const int max_y = (src_height - 1) << 16; + void (*InterpolateRow)(uint16* dst_ptr, const uint16* src_ptr, + ptrdiff_t src_stride, int dst_width, int source_y_fraction) = + InterpolateRow_16_C; + void (*ScaleFilterCols)(uint16* dst_ptr, const uint16* src_ptr, + int dst_width, int x, int dx) = + filtering ? ScaleFilterCols_16_C : ScaleCols_16_C; + ScaleSlope(src_width, src_height, dst_width, dst_height, filtering, + &x, &y, &dx, &dy); + src_width = Abs(src_width); + +#if defined(HAS_INTERPOLATEROW_16_SSE2) + if (TestCpuFlag(kCpuHasSSE2)) { + InterpolateRow = InterpolateRow_Any_16_SSE2; + if (IS_ALIGNED(dst_width, 16)) { + InterpolateRow = InterpolateRow_16_SSE2; + } + } +#endif +#if defined(HAS_INTERPOLATEROW_16_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + InterpolateRow = InterpolateRow_Any_16_SSSE3; + if (IS_ALIGNED(dst_width, 16)) { + InterpolateRow = InterpolateRow_16_SSSE3; + } + } +#endif +#if defined(HAS_INTERPOLATEROW_16_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + InterpolateRow = InterpolateRow_Any_16_AVX2; + if (IS_ALIGNED(dst_width, 32)) { + InterpolateRow = InterpolateRow_16_AVX2; + } + } +#endif +#if defined(HAS_INTERPOLATEROW_16_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + InterpolateRow = InterpolateRow_Any_16_NEON; + if (IS_ALIGNED(dst_width, 16)) { + InterpolateRow = InterpolateRow_16_NEON; + } + } +#endif +#if defined(HAS_INTERPOLATEROW_16_MIPS_DSPR2) + if (TestCpuFlag(kCpuHasMIPS_DSPR2)) { + InterpolateRow = InterpolateRow_Any_16_MIPS_DSPR2; + if (IS_ALIGNED(dst_width, 4)) { + InterpolateRow = InterpolateRow_16_MIPS_DSPR2; + } + } +#endif + + if (filtering && src_width >= 32768) { + ScaleFilterCols = ScaleFilterCols64_16_C; + } +#if defined(HAS_SCALEFILTERCOLS_16_SSSE3) + if (filtering && TestCpuFlag(kCpuHasSSSE3) && src_width < 32768) { + ScaleFilterCols = ScaleFilterCols_16_SSSE3; + } +#endif + if (!filtering && src_width * 2 == dst_width && x < 0x8000) { + ScaleFilterCols = ScaleColsUp2_16_C; +#if defined(HAS_SCALECOLS_16_SSE2) + if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 8)) { + ScaleFilterCols = ScaleColsUp2_16_SSE2; + } +#endif + } + + if (y > max_y) { + y = max_y; + } + { + int yi = y >> 16; + const uint16* src = src_ptr + yi * src_stride; + + // Allocate 2 row buffers. + const int kRowSize = (dst_width + 31) & ~31; + align_buffer_64(row, kRowSize * 4); + + uint16* rowptr = (uint16*)row; + int rowstride = kRowSize; + int lasty = yi; + + ScaleFilterCols(rowptr, src, dst_width, x, dx); + if (src_height > 1) { + src += src_stride; + } + ScaleFilterCols(rowptr + rowstride, src, dst_width, x, dx); + src += src_stride; + + for (j = 0; j < dst_height; ++j) { + yi = y >> 16; + if (yi != lasty) { + if (y > max_y) { + y = max_y; + yi = y >> 16; + src = src_ptr + yi * src_stride; + } + if (yi != lasty) { + ScaleFilterCols(rowptr, src, dst_width, x, dx); + rowptr += rowstride; + rowstride = -rowstride; + lasty = yi; + src += src_stride; + } + } + if (filtering == kFilterLinear) { + InterpolateRow(dst_ptr, rowptr, 0, dst_width, 0); + } else { + int yf = (y >> 8) & 255; + InterpolateRow(dst_ptr, rowptr, rowstride, dst_width, yf); + } + dst_ptr += dst_stride; + y += dy; + } + free_aligned_buffer_64(row); + } +} + +// Scale Plane to/from any dimensions, without interpolation. +// Fixed point math is used for performance: The upper 16 bits +// of x and dx is the integer part of the source position and +// the lower 16 bits are the fixed decimal part. + +static void ScalePlaneSimple(int src_width, int src_height, + int dst_width, int dst_height, + int src_stride, int dst_stride, + const uint8* src_ptr, uint8* dst_ptr) { + int i; + void (*ScaleCols)(uint8* dst_ptr, const uint8* src_ptr, + int dst_width, int x, int dx) = ScaleCols_C; + // Initial source x/y coordinate and step values as 16.16 fixed point. + int x = 0; + int y = 0; + int dx = 0; + int dy = 0; + ScaleSlope(src_width, src_height, dst_width, dst_height, kFilterNone, + &x, &y, &dx, &dy); + src_width = Abs(src_width); + + if (src_width * 2 == dst_width && x < 0x8000) { + ScaleCols = ScaleColsUp2_C; +#if defined(HAS_SCALECOLS_SSE2) + if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 8)) { + ScaleCols = ScaleColsUp2_SSE2; + } +#endif + } + + for (i = 0; i < dst_height; ++i) { + ScaleCols(dst_ptr, src_ptr + (y >> 16) * src_stride, dst_width, x, dx); + dst_ptr += dst_stride; + y += dy; + } +} + +static void ScalePlaneSimple_16(int src_width, int src_height, + int dst_width, int dst_height, + int src_stride, int dst_stride, + const uint16* src_ptr, uint16* dst_ptr) { + int i; + void (*ScaleCols)(uint16* dst_ptr, const uint16* src_ptr, + int dst_width, int x, int dx) = ScaleCols_16_C; + // Initial source x/y coordinate and step values as 16.16 fixed point. + int x = 0; + int y = 0; + int dx = 0; + int dy = 0; + ScaleSlope(src_width, src_height, dst_width, dst_height, kFilterNone, + &x, &y, &dx, &dy); + src_width = Abs(src_width); + + if (src_width * 2 == dst_width && x < 0x8000) { + ScaleCols = ScaleColsUp2_16_C; +#if defined(HAS_SCALECOLS_16_SSE2) + if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 8)) { + ScaleCols = ScaleColsUp2_16_SSE2; + } +#endif + } + + for (i = 0; i < dst_height; ++i) { + ScaleCols(dst_ptr, src_ptr + (y >> 16) * src_stride, + dst_width, x, dx); + dst_ptr += dst_stride; + y += dy; + } +} + +// Scale a plane. +// This function dispatches to a specialized scaler based on scale factor. + +LIBYUV_API +void ScalePlane(const uint8* src, int src_stride, + int src_width, int src_height, + uint8* dst, int dst_stride, + int dst_width, int dst_height, + enum FilterMode filtering) { + // Simplify filtering when possible. + filtering = ScaleFilterReduce(src_width, src_height, + dst_width, dst_height, filtering); + + // Negative height means invert the image. + if (src_height < 0) { + src_height = -src_height; + src = src + (src_height - 1) * src_stride; + src_stride = -src_stride; + } + + // Use specialized scales to improve performance for common resolutions. + // For example, all the 1/2 scalings will use ScalePlaneDown2() + if (dst_width == src_width && dst_height == src_height) { + // Straight copy. + CopyPlane(src, src_stride, dst, dst_stride, dst_width, dst_height); + return; + } + if (dst_width == src_width && filtering != kFilterBox) { + int dy = FixedDiv(src_height, dst_height); + // Arbitrary scale vertically, but unscaled horizontally. + ScalePlaneVertical(src_height, + dst_width, dst_height, + src_stride, dst_stride, src, dst, + 0, 0, dy, 1, filtering); + return; + } + if (dst_width <= Abs(src_width) && dst_height <= src_height) { + // Scale down. + if (4 * dst_width == 3 * src_width && + 4 * dst_height == 3 * src_height) { + // optimized, 3/4 + ScalePlaneDown34(src_width, src_height, dst_width, dst_height, + src_stride, dst_stride, src, dst, filtering); + return; + } + if (2 * dst_width == src_width && 2 * dst_height == src_height) { + // optimized, 1/2 + ScalePlaneDown2(src_width, src_height, dst_width, dst_height, + src_stride, dst_stride, src, dst, filtering); + return; + } + // 3/8 rounded up for odd sized chroma height. + if (8 * dst_width == 3 * src_width && + dst_height == ((src_height * 3 + 7) / 8)) { + // optimized, 3/8 + ScalePlaneDown38(src_width, src_height, dst_width, dst_height, + src_stride, dst_stride, src, dst, filtering); + return; + } + if (4 * dst_width == src_width && 4 * dst_height == src_height && + (filtering == kFilterBox || filtering == kFilterNone)) { + // optimized, 1/4 + ScalePlaneDown4(src_width, src_height, dst_width, dst_height, + src_stride, dst_stride, src, dst, filtering); + return; + } + } + if (filtering == kFilterBox && dst_height * 2 < src_height) { + ScalePlaneBox(src_width, src_height, dst_width, dst_height, + src_stride, dst_stride, src, dst); + return; + } + if (filtering && dst_height > src_height) { + ScalePlaneBilinearUp(src_width, src_height, dst_width, dst_height, + src_stride, dst_stride, src, dst, filtering); + return; + } + if (filtering) { + ScalePlaneBilinearDown(src_width, src_height, dst_width, dst_height, + src_stride, dst_stride, src, dst, filtering); + return; + } + ScalePlaneSimple(src_width, src_height, dst_width, dst_height, + src_stride, dst_stride, src, dst); +} + +LIBYUV_API +void ScalePlane_16(const uint16* src, int src_stride, + int src_width, int src_height, + uint16* dst, int dst_stride, + int dst_width, int dst_height, + enum FilterMode filtering) { + // Simplify filtering when possible. + filtering = ScaleFilterReduce(src_width, src_height, + dst_width, dst_height, filtering); + + // Negative height means invert the image. + if (src_height < 0) { + src_height = -src_height; + src = src + (src_height - 1) * src_stride; + src_stride = -src_stride; + } + + // Use specialized scales to improve performance for common resolutions. + // For example, all the 1/2 scalings will use ScalePlaneDown2() + if (dst_width == src_width && dst_height == src_height) { + // Straight copy. + CopyPlane_16(src, src_stride, dst, dst_stride, dst_width, dst_height); + return; + } + if (dst_width == src_width) { + int dy = FixedDiv(src_height, dst_height); + // Arbitrary scale vertically, but unscaled vertically. + ScalePlaneVertical_16(src_height, + dst_width, dst_height, + src_stride, dst_stride, src, dst, + 0, 0, dy, 1, filtering); + return; + } + if (dst_width <= Abs(src_width) && dst_height <= src_height) { + // Scale down. + if (4 * dst_width == 3 * src_width && + 4 * dst_height == 3 * src_height) { + // optimized, 3/4 + ScalePlaneDown34_16(src_width, src_height, dst_width, dst_height, + src_stride, dst_stride, src, dst, filtering); + return; + } + if (2 * dst_width == src_width && 2 * dst_height == src_height) { + // optimized, 1/2 + ScalePlaneDown2_16(src_width, src_height, dst_width, dst_height, + src_stride, dst_stride, src, dst, filtering); + return; + } + // 3/8 rounded up for odd sized chroma height. + if (8 * dst_width == 3 * src_width && + dst_height == ((src_height * 3 + 7) / 8)) { + // optimized, 3/8 + ScalePlaneDown38_16(src_width, src_height, dst_width, dst_height, + src_stride, dst_stride, src, dst, filtering); + return; + } + if (4 * dst_width == src_width && 4 * dst_height == src_height && + filtering != kFilterBilinear) { + // optimized, 1/4 + ScalePlaneDown4_16(src_width, src_height, dst_width, dst_height, + src_stride, dst_stride, src, dst, filtering); + return; + } + } + if (filtering == kFilterBox && dst_height * 2 < src_height) { + ScalePlaneBox_16(src_width, src_height, dst_width, dst_height, + src_stride, dst_stride, src, dst); + return; + } + if (filtering && dst_height > src_height) { + ScalePlaneBilinearUp_16(src_width, src_height, dst_width, dst_height, + src_stride, dst_stride, src, dst, filtering); + return; + } + if (filtering) { + ScalePlaneBilinearDown_16(src_width, src_height, dst_width, dst_height, + src_stride, dst_stride, src, dst, filtering); + return; + } + ScalePlaneSimple_16(src_width, src_height, dst_width, dst_height, + src_stride, dst_stride, src, dst); +} + +// Scale an I420 image. +// This function in turn calls a scaling function for each plane. + +LIBYUV_API +int I420Scale(const uint8* src_y, int src_stride_y, + const uint8* src_u, int src_stride_u, + const uint8* src_v, int src_stride_v, + int src_width, int src_height, + uint8* dst_y, int dst_stride_y, + uint8* dst_u, int dst_stride_u, + uint8* dst_v, int dst_stride_v, + int dst_width, int dst_height, + enum FilterMode filtering) { + int src_halfwidth = SUBSAMPLE(src_width, 1, 1); + int src_halfheight = SUBSAMPLE(src_height, 1, 1); + int dst_halfwidth = SUBSAMPLE(dst_width, 1, 1); + int dst_halfheight = SUBSAMPLE(dst_height, 1, 1); + if (!src_y || !src_u || !src_v || src_width == 0 || src_height == 0 || + src_width > 32768 || src_height > 32768 || + !dst_y || !dst_u || !dst_v || dst_width <= 0 || dst_height <= 0) { + return -1; + } + + ScalePlane(src_y, src_stride_y, src_width, src_height, + dst_y, dst_stride_y, dst_width, dst_height, + filtering); + ScalePlane(src_u, src_stride_u, src_halfwidth, src_halfheight, + dst_u, dst_stride_u, dst_halfwidth, dst_halfheight, + filtering); + ScalePlane(src_v, src_stride_v, src_halfwidth, src_halfheight, + dst_v, dst_stride_v, dst_halfwidth, dst_halfheight, + filtering); + return 0; +} + +LIBYUV_API +int I420Scale_16(const uint16* src_y, int src_stride_y, + const uint16* src_u, int src_stride_u, + const uint16* src_v, int src_stride_v, + int src_width, int src_height, + uint16* dst_y, int dst_stride_y, + uint16* dst_u, int dst_stride_u, + uint16* dst_v, int dst_stride_v, + int dst_width, int dst_height, + enum FilterMode filtering) { + int src_halfwidth = SUBSAMPLE(src_width, 1, 1); + int src_halfheight = SUBSAMPLE(src_height, 1, 1); + int dst_halfwidth = SUBSAMPLE(dst_width, 1, 1); + int dst_halfheight = SUBSAMPLE(dst_height, 1, 1); + if (!src_y || !src_u || !src_v || src_width == 0 || src_height == 0 || + src_width > 32768 || src_height > 32768 || + !dst_y || !dst_u || !dst_v || dst_width <= 0 || dst_height <= 0) { + return -1; + } + + ScalePlane_16(src_y, src_stride_y, src_width, src_height, + dst_y, dst_stride_y, dst_width, dst_height, + filtering); + ScalePlane_16(src_u, src_stride_u, src_halfwidth, src_halfheight, + dst_u, dst_stride_u, dst_halfwidth, dst_halfheight, + filtering); + ScalePlane_16(src_v, src_stride_v, src_halfwidth, src_halfheight, + dst_v, dst_stride_v, dst_halfwidth, dst_halfheight, + filtering); + return 0; +} + +// Deprecated api +LIBYUV_API +int Scale(const uint8* src_y, const uint8* src_u, const uint8* src_v, + int src_stride_y, int src_stride_u, int src_stride_v, + int src_width, int src_height, + uint8* dst_y, uint8* dst_u, uint8* dst_v, + int dst_stride_y, int dst_stride_u, int dst_stride_v, + int dst_width, int dst_height, + LIBYUV_BOOL interpolate) { + return I420Scale(src_y, src_stride_y, + src_u, src_stride_u, + src_v, src_stride_v, + src_width, src_height, + dst_y, dst_stride_y, + dst_u, dst_stride_u, + dst_v, dst_stride_v, + dst_width, dst_height, + interpolate ? kFilterBox : kFilterNone); +} + +// Deprecated api +LIBYUV_API +int ScaleOffset(const uint8* src, int src_width, int src_height, + uint8* dst, int dst_width, int dst_height, int dst_yoffset, + LIBYUV_BOOL interpolate) { + // Chroma requires offset to multiple of 2. + int dst_yoffset_even = dst_yoffset & ~1; + int src_halfwidth = SUBSAMPLE(src_width, 1, 1); + int src_halfheight = SUBSAMPLE(src_height, 1, 1); + int dst_halfwidth = SUBSAMPLE(dst_width, 1, 1); + int dst_halfheight = SUBSAMPLE(dst_height, 1, 1); + int aheight = dst_height - dst_yoffset_even * 2; // actual output height + const uint8* src_y = src; + const uint8* src_u = src + src_width * src_height; + const uint8* src_v = src + src_width * src_height + + src_halfwidth * src_halfheight; + uint8* dst_y = dst + dst_yoffset_even * dst_width; + uint8* dst_u = dst + dst_width * dst_height + + (dst_yoffset_even >> 1) * dst_halfwidth; + uint8* dst_v = dst + dst_width * dst_height + dst_halfwidth * dst_halfheight + + (dst_yoffset_even >> 1) * dst_halfwidth; + if (!src || src_width <= 0 || src_height <= 0 || + !dst || dst_width <= 0 || dst_height <= 0 || dst_yoffset_even < 0 || + dst_yoffset_even >= dst_height) { + return -1; + } + return I420Scale(src_y, src_width, + src_u, src_halfwidth, + src_v, src_halfwidth, + src_width, src_height, + dst_y, dst_width, + dst_u, dst_halfwidth, + dst_v, dst_halfwidth, + dst_width, aheight, + interpolate ? kFilterBox : kFilterNone); +} + +#ifdef __cplusplus +} // extern "C" +} // namespace libyuv +#endif diff --git a/media/libaom/src/third_party/libyuv/source/scale_any.cc b/media/libaom/src/third_party/libyuv/source/scale_any.cc new file mode 100644 index 000000000..2f6a2c8ba --- /dev/null +++ b/media/libaom/src/third_party/libyuv/source/scale_any.cc @@ -0,0 +1,200 @@ +/* + * Copyright 2015 The LibYuv Project Authors. All rights reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#include "libyuv/scale.h" +#include "libyuv/scale_row.h" + +#include "libyuv/basic_types.h" + +#ifdef __cplusplus +namespace libyuv { +extern "C" { +#endif + +// Definition for ScaleFilterCols, ScaleARGBCols and ScaleARGBFilterCols +#define CANY(NAMEANY, TERP_SIMD, TERP_C, BPP, MASK) \ + void NAMEANY(uint8* dst_ptr, const uint8* src_ptr, \ + int dst_width, int x, int dx) { \ + int n = dst_width & ~MASK; \ + if (n > 0) { \ + TERP_SIMD(dst_ptr, src_ptr, n, x, dx); \ + } \ + TERP_C(dst_ptr + n * BPP, src_ptr, \ + dst_width & MASK, x + n * dx, dx); \ + } + +#ifdef HAS_SCALEFILTERCOLS_NEON +CANY(ScaleFilterCols_Any_NEON, ScaleFilterCols_NEON, ScaleFilterCols_C, 1, 7) +#endif +#ifdef HAS_SCALEARGBCOLS_NEON +CANY(ScaleARGBCols_Any_NEON, ScaleARGBCols_NEON, ScaleARGBCols_C, 4, 7) +#endif +#ifdef HAS_SCALEARGBFILTERCOLS_NEON +CANY(ScaleARGBFilterCols_Any_NEON, ScaleARGBFilterCols_NEON, + ScaleARGBFilterCols_C, 4, 3) +#endif +#undef CANY + +// Fixed scale down. +#define SDANY(NAMEANY, SCALEROWDOWN_SIMD, SCALEROWDOWN_C, FACTOR, BPP, MASK) \ + void NAMEANY(const uint8* src_ptr, ptrdiff_t src_stride, \ + uint8* dst_ptr, int dst_width) { \ + int r = (int)((unsigned int)dst_width % (MASK + 1)); \ + int n = dst_width - r; \ + if (n > 0) { \ + SCALEROWDOWN_SIMD(src_ptr, src_stride, dst_ptr, n); \ + } \ + SCALEROWDOWN_C(src_ptr + (n * FACTOR) * BPP, src_stride, \ + dst_ptr + n * BPP, r); \ + } + +#ifdef HAS_SCALEROWDOWN2_SSE2 +SDANY(ScaleRowDown2_Any_SSE2, ScaleRowDown2_SSE2, ScaleRowDown2_C, 2, 1, 15) +SDANY(ScaleRowDown2Linear_Any_SSE2, ScaleRowDown2Linear_SSE2, + ScaleRowDown2Linear_C, 2, 1, 15) +SDANY(ScaleRowDown2Box_Any_SSE2, ScaleRowDown2Box_SSE2, ScaleRowDown2Box_C, + 2, 1, 15) +#endif +#ifdef HAS_SCALEROWDOWN2_AVX2 +SDANY(ScaleRowDown2_Any_AVX2, ScaleRowDown2_AVX2, ScaleRowDown2_C, 2, 1, 31) +SDANY(ScaleRowDown2Linear_Any_AVX2, ScaleRowDown2Linear_AVX2, + ScaleRowDown2Linear_C, 2, 1, 31) +SDANY(ScaleRowDown2Box_Any_AVX2, ScaleRowDown2Box_AVX2, ScaleRowDown2Box_C, + 2, 1, 31) +#endif +#ifdef HAS_SCALEROWDOWN2_NEON +SDANY(ScaleRowDown2_Any_NEON, ScaleRowDown2_NEON, ScaleRowDown2_C, 2, 1, 15) +SDANY(ScaleRowDown2Linear_Any_NEON, ScaleRowDown2Linear_NEON, + ScaleRowDown2Linear_C, 2, 1, 15) +SDANY(ScaleRowDown2Box_Any_NEON, ScaleRowDown2Box_NEON, + ScaleRowDown2Box_C, 2, 1, 15) +#endif +#ifdef HAS_SCALEROWDOWN4_SSE2 +SDANY(ScaleRowDown4_Any_SSE2, ScaleRowDown4_SSE2, ScaleRowDown4_C, 4, 1, 7) +SDANY(ScaleRowDown4Box_Any_SSE2, ScaleRowDown4Box_SSE2, ScaleRowDown4Box_C, + 4, 1, 7) +#endif +#ifdef HAS_SCALEROWDOWN4_AVX2 +SDANY(ScaleRowDown4_Any_AVX2, ScaleRowDown4_AVX2, ScaleRowDown4_C, 4, 1, 15) +SDANY(ScaleRowDown4Box_Any_AVX2, ScaleRowDown4Box_AVX2, ScaleRowDown4Box_C, + 4, 1, 15) +#endif +#ifdef HAS_SCALEROWDOWN4_NEON +SDANY(ScaleRowDown4_Any_NEON, ScaleRowDown4_NEON, ScaleRowDown4_C, 4, 1, 7) +SDANY(ScaleRowDown4Box_Any_NEON, ScaleRowDown4Box_NEON, ScaleRowDown4Box_C, + 4, 1, 7) +#endif +#ifdef HAS_SCALEROWDOWN34_SSSE3 +SDANY(ScaleRowDown34_Any_SSSE3, ScaleRowDown34_SSSE3, + ScaleRowDown34_C, 4 / 3, 1, 23) +SDANY(ScaleRowDown34_0_Box_Any_SSSE3, ScaleRowDown34_0_Box_SSSE3, + ScaleRowDown34_0_Box_C, 4 / 3, 1, 23) +SDANY(ScaleRowDown34_1_Box_Any_SSSE3, ScaleRowDown34_1_Box_SSSE3, + ScaleRowDown34_1_Box_C, 4 / 3, 1, 23) +#endif +#ifdef HAS_SCALEROWDOWN34_NEON +SDANY(ScaleRowDown34_Any_NEON, ScaleRowDown34_NEON, + ScaleRowDown34_C, 4 / 3, 1, 23) +SDANY(ScaleRowDown34_0_Box_Any_NEON, ScaleRowDown34_0_Box_NEON, + ScaleRowDown34_0_Box_C, 4 / 3, 1, 23) +SDANY(ScaleRowDown34_1_Box_Any_NEON, ScaleRowDown34_1_Box_NEON, + ScaleRowDown34_1_Box_C, 4 / 3, 1, 23) +#endif +#ifdef HAS_SCALEROWDOWN38_SSSE3 +SDANY(ScaleRowDown38_Any_SSSE3, ScaleRowDown38_SSSE3, + ScaleRowDown38_C, 8 / 3, 1, 11) +SDANY(ScaleRowDown38_3_Box_Any_SSSE3, ScaleRowDown38_3_Box_SSSE3, + ScaleRowDown38_3_Box_C, 8 / 3, 1, 5) +SDANY(ScaleRowDown38_2_Box_Any_SSSE3, ScaleRowDown38_2_Box_SSSE3, + ScaleRowDown38_2_Box_C, 8 / 3, 1, 5) +#endif +#ifdef HAS_SCALEROWDOWN38_NEON +SDANY(ScaleRowDown38_Any_NEON, ScaleRowDown38_NEON, + ScaleRowDown38_C, 8 / 3, 1, 11) +SDANY(ScaleRowDown38_3_Box_Any_NEON, ScaleRowDown38_3_Box_NEON, + ScaleRowDown38_3_Box_C, 8 / 3, 1, 11) +SDANY(ScaleRowDown38_2_Box_Any_NEON, ScaleRowDown38_2_Box_NEON, + ScaleRowDown38_2_Box_C, 8 / 3, 1, 11) +#endif + +#ifdef HAS_SCALEARGBROWDOWN2_SSE2 +SDANY(ScaleARGBRowDown2_Any_SSE2, ScaleARGBRowDown2_SSE2, + ScaleARGBRowDown2_C, 2, 4, 3) +SDANY(ScaleARGBRowDown2Linear_Any_SSE2, ScaleARGBRowDown2Linear_SSE2, + ScaleARGBRowDown2Linear_C, 2, 4, 3) +SDANY(ScaleARGBRowDown2Box_Any_SSE2, ScaleARGBRowDown2Box_SSE2, + ScaleARGBRowDown2Box_C, 2, 4, 3) +#endif +#ifdef HAS_SCALEARGBROWDOWN2_NEON +SDANY(ScaleARGBRowDown2_Any_NEON, ScaleARGBRowDown2_NEON, + ScaleARGBRowDown2_C, 2, 4, 7) +SDANY(ScaleARGBRowDown2Linear_Any_NEON, ScaleARGBRowDown2Linear_NEON, + ScaleARGBRowDown2Linear_C, 2, 4, 7) +SDANY(ScaleARGBRowDown2Box_Any_NEON, ScaleARGBRowDown2Box_NEON, + ScaleARGBRowDown2Box_C, 2, 4, 7) +#endif +#undef SDANY + +// Scale down by even scale factor. +#define SDAANY(NAMEANY, SCALEROWDOWN_SIMD, SCALEROWDOWN_C, BPP, MASK) \ + void NAMEANY(const uint8* src_ptr, ptrdiff_t src_stride, int src_stepx, \ + uint8* dst_ptr, int dst_width) { \ + int r = (int)((unsigned int)dst_width % (MASK + 1)); \ + int n = dst_width - r; \ + if (n > 0) { \ + SCALEROWDOWN_SIMD(src_ptr, src_stride, src_stepx, dst_ptr, n); \ + } \ + SCALEROWDOWN_C(src_ptr + (n * src_stepx) * BPP, src_stride, \ + src_stepx, dst_ptr + n * BPP, r); \ + } + +#ifdef HAS_SCALEARGBROWDOWNEVEN_SSE2 +SDAANY(ScaleARGBRowDownEven_Any_SSE2, ScaleARGBRowDownEven_SSE2, + ScaleARGBRowDownEven_C, 4, 3) +SDAANY(ScaleARGBRowDownEvenBox_Any_SSE2, ScaleARGBRowDownEvenBox_SSE2, + ScaleARGBRowDownEvenBox_C, 4, 3) +#endif +#ifdef HAS_SCALEARGBROWDOWNEVEN_NEON +SDAANY(ScaleARGBRowDownEven_Any_NEON, ScaleARGBRowDownEven_NEON, + ScaleARGBRowDownEven_C, 4, 3) +SDAANY(ScaleARGBRowDownEvenBox_Any_NEON, ScaleARGBRowDownEvenBox_NEON, + ScaleARGBRowDownEvenBox_C, 4, 3) +#endif + +// Add rows box filter scale down. +#define SAANY(NAMEANY, SCALEADDROW_SIMD, SCALEADDROW_C, MASK) \ + void NAMEANY(const uint8* src_ptr, uint16* dst_ptr, int src_width) { \ + int n = src_width & ~MASK; \ + if (n > 0) { \ + SCALEADDROW_SIMD(src_ptr, dst_ptr, n); \ + } \ + SCALEADDROW_C(src_ptr + n, dst_ptr + n, src_width & MASK); \ + } + +#ifdef HAS_SCALEADDROW_SSE2 +SAANY(ScaleAddRow_Any_SSE2, ScaleAddRow_SSE2, ScaleAddRow_C, 15) +#endif +#ifdef HAS_SCALEADDROW_AVX2 +SAANY(ScaleAddRow_Any_AVX2, ScaleAddRow_AVX2, ScaleAddRow_C, 31) +#endif +#ifdef HAS_SCALEADDROW_NEON +SAANY(ScaleAddRow_Any_NEON, ScaleAddRow_NEON, ScaleAddRow_C, 15) +#endif +#undef SAANY + +#ifdef __cplusplus +} // extern "C" +} // namespace libyuv +#endif + + + + + diff --git a/media/libaom/src/third_party/libyuv/source/scale_argb.cc b/media/libaom/src/third_party/libyuv/source/scale_argb.cc new file mode 100644 index 000000000..40a2d1ab2 --- /dev/null +++ b/media/libaom/src/third_party/libyuv/source/scale_argb.cc @@ -0,0 +1,853 @@ +/* + * Copyright 2011 The LibYuv Project Authors. All rights reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#include "libyuv/scale.h" + +#include <assert.h> +#include <string.h> + +#include "libyuv/cpu_id.h" +#include "libyuv/planar_functions.h" // For CopyARGB +#include "libyuv/row.h" +#include "libyuv/scale_row.h" + +#ifdef __cplusplus +namespace libyuv { +extern "C" { +#endif + +static __inline int Abs(int v) { + return v >= 0 ? v : -v; +} + +// ScaleARGB ARGB, 1/2 +// This is an optimized version for scaling down a ARGB to 1/2 of +// its original size. +static void ScaleARGBDown2(int src_width, int src_height, + int dst_width, int dst_height, + int src_stride, int dst_stride, + const uint8* src_argb, uint8* dst_argb, + int x, int dx, int y, int dy, + enum FilterMode filtering) { + int j; + int row_stride = src_stride * (dy >> 16); + void (*ScaleARGBRowDown2)(const uint8* src_argb, ptrdiff_t src_stride, + uint8* dst_argb, int dst_width) = + filtering == kFilterNone ? ScaleARGBRowDown2_C : + (filtering == kFilterLinear ? ScaleARGBRowDown2Linear_C : + ScaleARGBRowDown2Box_C); + assert(dx == 65536 * 2); // Test scale factor of 2. + assert((dy & 0x1ffff) == 0); // Test vertical scale is multiple of 2. + // Advance to odd row, even column. + if (filtering == kFilterBilinear) { + src_argb += (y >> 16) * src_stride + (x >> 16) * 4; + } else { + src_argb += (y >> 16) * src_stride + ((x >> 16) - 1) * 4; + } + +#if defined(HAS_SCALEARGBROWDOWN2_SSE2) + if (TestCpuFlag(kCpuHasSSE2)) { + ScaleARGBRowDown2 = filtering == kFilterNone ? ScaleARGBRowDown2_Any_SSE2 : + (filtering == kFilterLinear ? ScaleARGBRowDown2Linear_Any_SSE2 : + ScaleARGBRowDown2Box_Any_SSE2); + if (IS_ALIGNED(dst_width, 4)) { + ScaleARGBRowDown2 = filtering == kFilterNone ? ScaleARGBRowDown2_SSE2 : + (filtering == kFilterLinear ? ScaleARGBRowDown2Linear_SSE2 : + ScaleARGBRowDown2Box_SSE2); + } + } +#endif +#if defined(HAS_SCALEARGBROWDOWN2_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + ScaleARGBRowDown2 = filtering == kFilterNone ? ScaleARGBRowDown2_Any_NEON : + (filtering == kFilterLinear ? ScaleARGBRowDown2Linear_Any_NEON : + ScaleARGBRowDown2Box_Any_NEON); + if (IS_ALIGNED(dst_width, 8)) { + ScaleARGBRowDown2 = filtering == kFilterNone ? ScaleARGBRowDown2_NEON : + (filtering == kFilterLinear ? ScaleARGBRowDown2Linear_NEON : + ScaleARGBRowDown2Box_NEON); + } + } +#endif + + if (filtering == kFilterLinear) { + src_stride = 0; + } + for (j = 0; j < dst_height; ++j) { + ScaleARGBRowDown2(src_argb, src_stride, dst_argb, dst_width); + src_argb += row_stride; + dst_argb += dst_stride; + } +} + +// ScaleARGB ARGB, 1/4 +// This is an optimized version for scaling down a ARGB to 1/4 of +// its original size. +static void ScaleARGBDown4Box(int src_width, int src_height, + int dst_width, int dst_height, + int src_stride, int dst_stride, + const uint8* src_argb, uint8* dst_argb, + int x, int dx, int y, int dy) { + int j; + // Allocate 2 rows of ARGB. + const int kRowSize = (dst_width * 2 * 4 + 31) & ~31; + align_buffer_64(row, kRowSize * 2); + int row_stride = src_stride * (dy >> 16); + void (*ScaleARGBRowDown2)(const uint8* src_argb, ptrdiff_t src_stride, + uint8* dst_argb, int dst_width) = ScaleARGBRowDown2Box_C; + // Advance to odd row, even column. + src_argb += (y >> 16) * src_stride + (x >> 16) * 4; + assert(dx == 65536 * 4); // Test scale factor of 4. + assert((dy & 0x3ffff) == 0); // Test vertical scale is multiple of 4. +#if defined(HAS_SCALEARGBROWDOWN2_SSE2) + if (TestCpuFlag(kCpuHasSSE2)) { + ScaleARGBRowDown2 = ScaleARGBRowDown2Box_Any_SSE2; + if (IS_ALIGNED(dst_width, 4)) { + ScaleARGBRowDown2 = ScaleARGBRowDown2Box_SSE2; + } + } +#endif +#if defined(HAS_SCALEARGBROWDOWN2_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + ScaleARGBRowDown2 = ScaleARGBRowDown2Box_Any_NEON; + if (IS_ALIGNED(dst_width, 8)) { + ScaleARGBRowDown2 = ScaleARGBRowDown2Box_NEON; + } + } +#endif + + for (j = 0; j < dst_height; ++j) { + ScaleARGBRowDown2(src_argb, src_stride, row, dst_width * 2); + ScaleARGBRowDown2(src_argb + src_stride * 2, src_stride, + row + kRowSize, dst_width * 2); + ScaleARGBRowDown2(row, kRowSize, dst_argb, dst_width); + src_argb += row_stride; + dst_argb += dst_stride; + } + free_aligned_buffer_64(row); +} + +// ScaleARGB ARGB Even +// This is an optimized version for scaling down a ARGB to even +// multiple of its original size. +static void ScaleARGBDownEven(int src_width, int src_height, + int dst_width, int dst_height, + int src_stride, int dst_stride, + const uint8* src_argb, uint8* dst_argb, + int x, int dx, int y, int dy, + enum FilterMode filtering) { + int j; + int col_step = dx >> 16; + int row_stride = (dy >> 16) * src_stride; + void (*ScaleARGBRowDownEven)(const uint8* src_argb, ptrdiff_t src_stride, + int src_step, uint8* dst_argb, int dst_width) = + filtering ? ScaleARGBRowDownEvenBox_C : ScaleARGBRowDownEven_C; + assert(IS_ALIGNED(src_width, 2)); + assert(IS_ALIGNED(src_height, 2)); + src_argb += (y >> 16) * src_stride + (x >> 16) * 4; +#if defined(HAS_SCALEARGBROWDOWNEVEN_SSE2) + if (TestCpuFlag(kCpuHasSSE2)) { + ScaleARGBRowDownEven = filtering ? ScaleARGBRowDownEvenBox_Any_SSE2 : + ScaleARGBRowDownEven_Any_SSE2; + if (IS_ALIGNED(dst_width, 4)) { + ScaleARGBRowDownEven = filtering ? ScaleARGBRowDownEvenBox_SSE2 : + ScaleARGBRowDownEven_SSE2; + } + } +#endif +#if defined(HAS_SCALEARGBROWDOWNEVEN_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + ScaleARGBRowDownEven = filtering ? ScaleARGBRowDownEvenBox_Any_NEON : + ScaleARGBRowDownEven_Any_NEON; + if (IS_ALIGNED(dst_width, 4)) { + ScaleARGBRowDownEven = filtering ? ScaleARGBRowDownEvenBox_NEON : + ScaleARGBRowDownEven_NEON; + } + } +#endif + + if (filtering == kFilterLinear) { + src_stride = 0; + } + for (j = 0; j < dst_height; ++j) { + ScaleARGBRowDownEven(src_argb, src_stride, col_step, dst_argb, dst_width); + src_argb += row_stride; + dst_argb += dst_stride; + } +} + +// Scale ARGB down with bilinear interpolation. +static void ScaleARGBBilinearDown(int src_width, int src_height, + int dst_width, int dst_height, + int src_stride, int dst_stride, + const uint8* src_argb, uint8* dst_argb, + int x, int dx, int y, int dy, + enum FilterMode filtering) { + int j; + void (*InterpolateRow)(uint8* dst_argb, const uint8* src_argb, + ptrdiff_t src_stride, int dst_width, int source_y_fraction) = + InterpolateRow_C; + void (*ScaleARGBFilterCols)(uint8* dst_argb, const uint8* src_argb, + int dst_width, int x, int dx) = + (src_width >= 32768) ? ScaleARGBFilterCols64_C : ScaleARGBFilterCols_C; + int64 xlast = x + (int64)(dst_width - 1) * dx; + int64 xl = (dx >= 0) ? x : xlast; + int64 xr = (dx >= 0) ? xlast : x; + int clip_src_width; + xl = (xl >> 16) & ~3; // Left edge aligned. + xr = (xr >> 16) + 1; // Right most pixel used. Bilinear uses 2 pixels. + xr = (xr + 1 + 3) & ~3; // 1 beyond 4 pixel aligned right most pixel. + if (xr > src_width) { + xr = src_width; + } + clip_src_width = (int)(xr - xl) * 4; // Width aligned to 4. + src_argb += xl * 4; + x -= (int)(xl << 16); +#if defined(HAS_INTERPOLATEROW_SSE2) + if (TestCpuFlag(kCpuHasSSE2)) { + InterpolateRow = InterpolateRow_Any_SSE2; + if (IS_ALIGNED(clip_src_width, 16)) { + InterpolateRow = InterpolateRow_SSE2; + } + } +#endif +#if defined(HAS_INTERPOLATEROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + InterpolateRow = InterpolateRow_Any_SSSE3; + if (IS_ALIGNED(clip_src_width, 16)) { + InterpolateRow = InterpolateRow_SSSE3; + } + } +#endif +#if defined(HAS_INTERPOLATEROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + InterpolateRow = InterpolateRow_Any_AVX2; + if (IS_ALIGNED(clip_src_width, 32)) { + InterpolateRow = InterpolateRow_AVX2; + } + } +#endif +#if defined(HAS_INTERPOLATEROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + InterpolateRow = InterpolateRow_Any_NEON; + if (IS_ALIGNED(clip_src_width, 16)) { + InterpolateRow = InterpolateRow_NEON; + } + } +#endif +#if defined(HAS_INTERPOLATEROW_MIPS_DSPR2) + if (TestCpuFlag(kCpuHasMIPS_DSPR2) && + IS_ALIGNED(src_argb, 4) && IS_ALIGNED(src_stride, 4)) { + InterpolateRow = InterpolateRow_Any_MIPS_DSPR2; + if (IS_ALIGNED(clip_src_width, 4)) { + InterpolateRow = InterpolateRow_MIPS_DSPR2; + } + } +#endif +#if defined(HAS_SCALEARGBFILTERCOLS_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3) && src_width < 32768) { + ScaleARGBFilterCols = ScaleARGBFilterCols_SSSE3; + } +#endif +#if defined(HAS_SCALEARGBFILTERCOLS_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + ScaleARGBFilterCols = ScaleARGBFilterCols_Any_NEON; + if (IS_ALIGNED(dst_width, 4)) { + ScaleARGBFilterCols = ScaleARGBFilterCols_NEON; + } + } +#endif + // TODO(fbarchard): Consider not allocating row buffer for kFilterLinear. + // Allocate a row of ARGB. + { + align_buffer_64(row, clip_src_width * 4); + + const int max_y = (src_height - 1) << 16; + if (y > max_y) { + y = max_y; + } + for (j = 0; j < dst_height; ++j) { + int yi = y >> 16; + const uint8* src = src_argb + yi * src_stride; + if (filtering == kFilterLinear) { + ScaleARGBFilterCols(dst_argb, src, dst_width, x, dx); + } else { + int yf = (y >> 8) & 255; + InterpolateRow(row, src, src_stride, clip_src_width, yf); + ScaleARGBFilterCols(dst_argb, row, dst_width, x, dx); + } + dst_argb += dst_stride; + y += dy; + if (y > max_y) { + y = max_y; + } + } + free_aligned_buffer_64(row); + } +} + +// Scale ARGB up with bilinear interpolation. +static void ScaleARGBBilinearUp(int src_width, int src_height, + int dst_width, int dst_height, + int src_stride, int dst_stride, + const uint8* src_argb, uint8* dst_argb, + int x, int dx, int y, int dy, + enum FilterMode filtering) { + int j; + void (*InterpolateRow)(uint8* dst_argb, const uint8* src_argb, + ptrdiff_t src_stride, int dst_width, int source_y_fraction) = + InterpolateRow_C; + void (*ScaleARGBFilterCols)(uint8* dst_argb, const uint8* src_argb, + int dst_width, int x, int dx) = + filtering ? ScaleARGBFilterCols_C : ScaleARGBCols_C; + const int max_y = (src_height - 1) << 16; +#if defined(HAS_INTERPOLATEROW_SSE2) + if (TestCpuFlag(kCpuHasSSE2)) { + InterpolateRow = InterpolateRow_Any_SSE2; + if (IS_ALIGNED(dst_width, 4)) { + InterpolateRow = InterpolateRow_SSE2; + } + } +#endif +#if defined(HAS_INTERPOLATEROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + InterpolateRow = InterpolateRow_Any_SSSE3; + if (IS_ALIGNED(dst_width, 4)) { + InterpolateRow = InterpolateRow_SSSE3; + } + } +#endif +#if defined(HAS_INTERPOLATEROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + InterpolateRow = InterpolateRow_Any_AVX2; + if (IS_ALIGNED(dst_width, 8)) { + InterpolateRow = InterpolateRow_AVX2; + } + } +#endif +#if defined(HAS_INTERPOLATEROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + InterpolateRow = InterpolateRow_Any_NEON; + if (IS_ALIGNED(dst_width, 4)) { + InterpolateRow = InterpolateRow_NEON; + } + } +#endif +#if defined(HAS_INTERPOLATEROW_MIPS_DSPR2) + if (TestCpuFlag(kCpuHasMIPS_DSPR2) && + IS_ALIGNED(dst_argb, 4) && IS_ALIGNED(dst_stride, 4)) { + InterpolateRow = InterpolateRow_MIPS_DSPR2; + } +#endif + if (src_width >= 32768) { + ScaleARGBFilterCols = filtering ? + ScaleARGBFilterCols64_C : ScaleARGBCols64_C; + } +#if defined(HAS_SCALEARGBFILTERCOLS_SSSE3) + if (filtering && TestCpuFlag(kCpuHasSSSE3) && src_width < 32768) { + ScaleARGBFilterCols = ScaleARGBFilterCols_SSSE3; + } +#endif +#if defined(HAS_SCALEARGBFILTERCOLS_NEON) + if (filtering && TestCpuFlag(kCpuHasNEON)) { + ScaleARGBFilterCols = ScaleARGBFilterCols_Any_NEON; + if (IS_ALIGNED(dst_width, 4)) { + ScaleARGBFilterCols = ScaleARGBFilterCols_NEON; + } + } +#endif +#if defined(HAS_SCALEARGBCOLS_SSE2) + if (!filtering && TestCpuFlag(kCpuHasSSE2) && src_width < 32768) { + ScaleARGBFilterCols = ScaleARGBCols_SSE2; + } +#endif +#if defined(HAS_SCALEARGBCOLS_NEON) + if (!filtering && TestCpuFlag(kCpuHasNEON)) { + ScaleARGBFilterCols = ScaleARGBCols_Any_NEON; + if (IS_ALIGNED(dst_width, 8)) { + ScaleARGBFilterCols = ScaleARGBCols_NEON; + } + } +#endif + if (!filtering && src_width * 2 == dst_width && x < 0x8000) { + ScaleARGBFilterCols = ScaleARGBColsUp2_C; +#if defined(HAS_SCALEARGBCOLSUP2_SSE2) + if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 8)) { + ScaleARGBFilterCols = ScaleARGBColsUp2_SSE2; + } +#endif + } + + if (y > max_y) { + y = max_y; + } + + { + int yi = y >> 16; + const uint8* src = src_argb + yi * src_stride; + + // Allocate 2 rows of ARGB. + const int kRowSize = (dst_width * 4 + 31) & ~31; + align_buffer_64(row, kRowSize * 2); + + uint8* rowptr = row; + int rowstride = kRowSize; + int lasty = yi; + + ScaleARGBFilterCols(rowptr, src, dst_width, x, dx); + if (src_height > 1) { + src += src_stride; + } + ScaleARGBFilterCols(rowptr + rowstride, src, dst_width, x, dx); + src += src_stride; + + for (j = 0; j < dst_height; ++j) { + yi = y >> 16; + if (yi != lasty) { + if (y > max_y) { + y = max_y; + yi = y >> 16; + src = src_argb + yi * src_stride; + } + if (yi != lasty) { + ScaleARGBFilterCols(rowptr, src, dst_width, x, dx); + rowptr += rowstride; + rowstride = -rowstride; + lasty = yi; + src += src_stride; + } + } + if (filtering == kFilterLinear) { + InterpolateRow(dst_argb, rowptr, 0, dst_width * 4, 0); + } else { + int yf = (y >> 8) & 255; + InterpolateRow(dst_argb, rowptr, rowstride, dst_width * 4, yf); + } + dst_argb += dst_stride; + y += dy; + } + free_aligned_buffer_64(row); + } +} + +#ifdef YUVSCALEUP +// Scale YUV to ARGB up with bilinear interpolation. +static void ScaleYUVToARGBBilinearUp(int src_width, int src_height, + int dst_width, int dst_height, + int src_stride_y, + int src_stride_u, + int src_stride_v, + int dst_stride_argb, + const uint8* src_y, + const uint8* src_u, + const uint8* src_v, + uint8* dst_argb, + int x, int dx, int y, int dy, + enum FilterMode filtering) { + int j; + void (*I422ToARGBRow)(const uint8* y_buf, + const uint8* u_buf, + const uint8* v_buf, + uint8* rgb_buf, + int width) = I422ToARGBRow_C; +#if defined(HAS_I422TOARGBROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + I422ToARGBRow = I422ToARGBRow_Any_SSSE3; + if (IS_ALIGNED(src_width, 8)) { + I422ToARGBRow = I422ToARGBRow_SSSE3; + } + } +#endif +#if defined(HAS_I422TOARGBROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + I422ToARGBRow = I422ToARGBRow_Any_AVX2; + if (IS_ALIGNED(src_width, 16)) { + I422ToARGBRow = I422ToARGBRow_AVX2; + } + } +#endif +#if defined(HAS_I422TOARGBROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + I422ToARGBRow = I422ToARGBRow_Any_NEON; + if (IS_ALIGNED(src_width, 8)) { + I422ToARGBRow = I422ToARGBRow_NEON; + } + } +#endif +#if defined(HAS_I422TOARGBROW_MIPS_DSPR2) + if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(src_width, 4) && + IS_ALIGNED(src_y, 4) && IS_ALIGNED(src_stride_y, 4) && + IS_ALIGNED(src_u, 2) && IS_ALIGNED(src_stride_u, 2) && + IS_ALIGNED(src_v, 2) && IS_ALIGNED(src_stride_v, 2) && + IS_ALIGNED(dst_argb, 4) && IS_ALIGNED(dst_stride_argb, 4)) { + I422ToARGBRow = I422ToARGBRow_MIPS_DSPR2; + } +#endif + + void (*InterpolateRow)(uint8* dst_argb, const uint8* src_argb, + ptrdiff_t src_stride, int dst_width, int source_y_fraction) = + InterpolateRow_C; +#if defined(HAS_INTERPOLATEROW_SSE2) + if (TestCpuFlag(kCpuHasSSE2)) { + InterpolateRow = InterpolateRow_Any_SSE2; + if (IS_ALIGNED(dst_width, 4)) { + InterpolateRow = InterpolateRow_SSE2; + } + } +#endif +#if defined(HAS_INTERPOLATEROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + InterpolateRow = InterpolateRow_Any_SSSE3; + if (IS_ALIGNED(dst_width, 4)) { + InterpolateRow = InterpolateRow_SSSE3; + } + } +#endif +#if defined(HAS_INTERPOLATEROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + InterpolateRow = InterpolateRow_Any_AVX2; + if (IS_ALIGNED(dst_width, 8)) { + InterpolateRow = InterpolateRow_AVX2; + } + } +#endif +#if defined(HAS_INTERPOLATEROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + InterpolateRow = InterpolateRow_Any_NEON; + if (IS_ALIGNED(dst_width, 4)) { + InterpolateRow = InterpolateRow_NEON; + } + } +#endif +#if defined(HAS_INTERPOLATEROW_MIPS_DSPR2) + if (TestCpuFlag(kCpuHasMIPS_DSPR2) && + IS_ALIGNED(dst_argb, 4) && IS_ALIGNED(dst_stride_argb, 4)) { + InterpolateRow = InterpolateRow_MIPS_DSPR2; + } +#endif + + void (*ScaleARGBFilterCols)(uint8* dst_argb, const uint8* src_argb, + int dst_width, int x, int dx) = + filtering ? ScaleARGBFilterCols_C : ScaleARGBCols_C; + if (src_width >= 32768) { + ScaleARGBFilterCols = filtering ? + ScaleARGBFilterCols64_C : ScaleARGBCols64_C; + } +#if defined(HAS_SCALEARGBFILTERCOLS_SSSE3) + if (filtering && TestCpuFlag(kCpuHasSSSE3) && src_width < 32768) { + ScaleARGBFilterCols = ScaleARGBFilterCols_SSSE3; + } +#endif +#if defined(HAS_SCALEARGBFILTERCOLS_NEON) + if (filtering && TestCpuFlag(kCpuHasNEON)) { + ScaleARGBFilterCols = ScaleARGBFilterCols_Any_NEON; + if (IS_ALIGNED(dst_width, 4)) { + ScaleARGBFilterCols = ScaleARGBFilterCols_NEON; + } + } +#endif +#if defined(HAS_SCALEARGBCOLS_SSE2) + if (!filtering && TestCpuFlag(kCpuHasSSE2) && src_width < 32768) { + ScaleARGBFilterCols = ScaleARGBCols_SSE2; + } +#endif +#if defined(HAS_SCALEARGBCOLS_NEON) + if (!filtering && TestCpuFlag(kCpuHasNEON)) { + ScaleARGBFilterCols = ScaleARGBCols_Any_NEON; + if (IS_ALIGNED(dst_width, 8)) { + ScaleARGBFilterCols = ScaleARGBCols_NEON; + } + } +#endif + if (!filtering && src_width * 2 == dst_width && x < 0x8000) { + ScaleARGBFilterCols = ScaleARGBColsUp2_C; +#if defined(HAS_SCALEARGBCOLSUP2_SSE2) + if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 8)) { + ScaleARGBFilterCols = ScaleARGBColsUp2_SSE2; + } +#endif + } + + const int max_y = (src_height - 1) << 16; + if (y > max_y) { + y = max_y; + } + const int kYShift = 1; // Shift Y by 1 to convert Y plane to UV coordinate. + int yi = y >> 16; + int uv_yi = yi >> kYShift; + const uint8* src_row_y = src_y + yi * src_stride_y; + const uint8* src_row_u = src_u + uv_yi * src_stride_u; + const uint8* src_row_v = src_v + uv_yi * src_stride_v; + + // Allocate 2 rows of ARGB. + const int kRowSize = (dst_width * 4 + 31) & ~31; + align_buffer_64(row, kRowSize * 2); + + // Allocate 1 row of ARGB for source conversion. + align_buffer_64(argb_row, src_width * 4); + + uint8* rowptr = row; + int rowstride = kRowSize; + int lasty = yi; + + // TODO(fbarchard): Convert first 2 rows of YUV to ARGB. + ScaleARGBFilterCols(rowptr, src_row_y, dst_width, x, dx); + if (src_height > 1) { + src_row_y += src_stride_y; + if (yi & 1) { + src_row_u += src_stride_u; + src_row_v += src_stride_v; + } + } + ScaleARGBFilterCols(rowptr + rowstride, src_row_y, dst_width, x, dx); + if (src_height > 2) { + src_row_y += src_stride_y; + if (!(yi & 1)) { + src_row_u += src_stride_u; + src_row_v += src_stride_v; + } + } + + for (j = 0; j < dst_height; ++j) { + yi = y >> 16; + if (yi != lasty) { + if (y > max_y) { + y = max_y; + yi = y >> 16; + uv_yi = yi >> kYShift; + src_row_y = src_y + yi * src_stride_y; + src_row_u = src_u + uv_yi * src_stride_u; + src_row_v = src_v + uv_yi * src_stride_v; + } + if (yi != lasty) { + // TODO(fbarchard): Convert the clipped region of row. + I422ToARGBRow(src_row_y, src_row_u, src_row_v, argb_row, src_width); + ScaleARGBFilterCols(rowptr, argb_row, dst_width, x, dx); + rowptr += rowstride; + rowstride = -rowstride; + lasty = yi; + src_row_y += src_stride_y; + if (yi & 1) { + src_row_u += src_stride_u; + src_row_v += src_stride_v; + } + } + } + if (filtering == kFilterLinear) { + InterpolateRow(dst_argb, rowptr, 0, dst_width * 4, 0); + } else { + int yf = (y >> 8) & 255; + InterpolateRow(dst_argb, rowptr, rowstride, dst_width * 4, yf); + } + dst_argb += dst_stride_argb; + y += dy; + } + free_aligned_buffer_64(row); + free_aligned_buffer_64(row_argb); +} +#endif + +// Scale ARGB to/from any dimensions, without interpolation. +// Fixed point math is used for performance: The upper 16 bits +// of x and dx is the integer part of the source position and +// the lower 16 bits are the fixed decimal part. + +static void ScaleARGBSimple(int src_width, int src_height, + int dst_width, int dst_height, + int src_stride, int dst_stride, + const uint8* src_argb, uint8* dst_argb, + int x, int dx, int y, int dy) { + int j; + void (*ScaleARGBCols)(uint8* dst_argb, const uint8* src_argb, + int dst_width, int x, int dx) = + (src_width >= 32768) ? ScaleARGBCols64_C : ScaleARGBCols_C; +#if defined(HAS_SCALEARGBCOLS_SSE2) + if (TestCpuFlag(kCpuHasSSE2) && src_width < 32768) { + ScaleARGBCols = ScaleARGBCols_SSE2; + } +#endif +#if defined(HAS_SCALEARGBCOLS_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + ScaleARGBCols = ScaleARGBCols_Any_NEON; + if (IS_ALIGNED(dst_width, 8)) { + ScaleARGBCols = ScaleARGBCols_NEON; + } + } +#endif + if (src_width * 2 == dst_width && x < 0x8000) { + ScaleARGBCols = ScaleARGBColsUp2_C; +#if defined(HAS_SCALEARGBCOLSUP2_SSE2) + if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 8)) { + ScaleARGBCols = ScaleARGBColsUp2_SSE2; + } +#endif + } + + for (j = 0; j < dst_height; ++j) { + ScaleARGBCols(dst_argb, src_argb + (y >> 16) * src_stride, + dst_width, x, dx); + dst_argb += dst_stride; + y += dy; + } +} + +// ScaleARGB a ARGB. +// This function in turn calls a scaling function +// suitable for handling the desired resolutions. +static void ScaleARGB(const uint8* src, int src_stride, + int src_width, int src_height, + uint8* dst, int dst_stride, + int dst_width, int dst_height, + int clip_x, int clip_y, int clip_width, int clip_height, + enum FilterMode filtering) { + // Initial source x/y coordinate and step values as 16.16 fixed point. + int x = 0; + int y = 0; + int dx = 0; + int dy = 0; + // ARGB does not support box filter yet, but allow the user to pass it. + // Simplify filtering when possible. + filtering = ScaleFilterReduce(src_width, src_height, + dst_width, dst_height, + filtering); + + // Negative src_height means invert the image. + if (src_height < 0) { + src_height = -src_height; + src = src + (src_height - 1) * src_stride; + src_stride = -src_stride; + } + ScaleSlope(src_width, src_height, dst_width, dst_height, filtering, + &x, &y, &dx, &dy); + src_width = Abs(src_width); + if (clip_x) { + int64 clipf = (int64)(clip_x) * dx; + x += (clipf & 0xffff); + src += (clipf >> 16) * 4; + dst += clip_x * 4; + } + if (clip_y) { + int64 clipf = (int64)(clip_y) * dy; + y += (clipf & 0xffff); + src += (clipf >> 16) * src_stride; + dst += clip_y * dst_stride; + } + + // Special case for integer step values. + if (((dx | dy) & 0xffff) == 0) { + if (!dx || !dy) { // 1 pixel wide and/or tall. + filtering = kFilterNone; + } else { + // Optimized even scale down. ie 2, 4, 6, 8, 10x. + if (!(dx & 0x10000) && !(dy & 0x10000)) { + if (dx == 0x20000) { + // Optimized 1/2 downsample. + ScaleARGBDown2(src_width, src_height, + clip_width, clip_height, + src_stride, dst_stride, src, dst, + x, dx, y, dy, filtering); + return; + } + if (dx == 0x40000 && filtering == kFilterBox) { + // Optimized 1/4 box downsample. + ScaleARGBDown4Box(src_width, src_height, + clip_width, clip_height, + src_stride, dst_stride, src, dst, + x, dx, y, dy); + return; + } + ScaleARGBDownEven(src_width, src_height, + clip_width, clip_height, + src_stride, dst_stride, src, dst, + x, dx, y, dy, filtering); + return; + } + // Optimized odd scale down. ie 3, 5, 7, 9x. + if ((dx & 0x10000) && (dy & 0x10000)) { + filtering = kFilterNone; + if (dx == 0x10000 && dy == 0x10000) { + // Straight copy. + ARGBCopy(src + (y >> 16) * src_stride + (x >> 16) * 4, src_stride, + dst, dst_stride, clip_width, clip_height); + return; + } + } + } + } + if (dx == 0x10000 && (x & 0xffff) == 0) { + // Arbitrary scale vertically, but unscaled vertically. + ScalePlaneVertical(src_height, + clip_width, clip_height, + src_stride, dst_stride, src, dst, + x, y, dy, 4, filtering); + return; + } + if (filtering && dy < 65536) { + ScaleARGBBilinearUp(src_width, src_height, + clip_width, clip_height, + src_stride, dst_stride, src, dst, + x, dx, y, dy, filtering); + return; + } + if (filtering) { + ScaleARGBBilinearDown(src_width, src_height, + clip_width, clip_height, + src_stride, dst_stride, src, dst, + x, dx, y, dy, filtering); + return; + } + ScaleARGBSimple(src_width, src_height, clip_width, clip_height, + src_stride, dst_stride, src, dst, + x, dx, y, dy); +} + +LIBYUV_API +int ARGBScaleClip(const uint8* src_argb, int src_stride_argb, + int src_width, int src_height, + uint8* dst_argb, int dst_stride_argb, + int dst_width, int dst_height, + int clip_x, int clip_y, int clip_width, int clip_height, + enum FilterMode filtering) { + if (!src_argb || src_width == 0 || src_height == 0 || + !dst_argb || dst_width <= 0 || dst_height <= 0 || + clip_x < 0 || clip_y < 0 || + clip_width > 32768 || clip_height > 32768 || + (clip_x + clip_width) > dst_width || + (clip_y + clip_height) > dst_height) { + return -1; + } + ScaleARGB(src_argb, src_stride_argb, src_width, src_height, + dst_argb, dst_stride_argb, dst_width, dst_height, + clip_x, clip_y, clip_width, clip_height, filtering); + return 0; +} + +// Scale an ARGB image. +LIBYUV_API +int ARGBScale(const uint8* src_argb, int src_stride_argb, + int src_width, int src_height, + uint8* dst_argb, int dst_stride_argb, + int dst_width, int dst_height, + enum FilterMode filtering) { + if (!src_argb || src_width == 0 || src_height == 0 || + src_width > 32768 || src_height > 32768 || + !dst_argb || dst_width <= 0 || dst_height <= 0) { + return -1; + } + ScaleARGB(src_argb, src_stride_argb, src_width, src_height, + dst_argb, dst_stride_argb, dst_width, dst_height, + 0, 0, dst_width, dst_height, filtering); + return 0; +} + +#ifdef __cplusplus +} // extern "C" +} // namespace libyuv +#endif diff --git a/media/libaom/src/third_party/libyuv/source/scale_common.cc b/media/libaom/src/third_party/libyuv/source/scale_common.cc new file mode 100644 index 000000000..1711f3d54 --- /dev/null +++ b/media/libaom/src/third_party/libyuv/source/scale_common.cc @@ -0,0 +1,1137 @@ +/* + * Copyright 2013 The LibYuv Project Authors. All rights reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#include "libyuv/scale.h" + +#include <assert.h> +#include <string.h> + +#include "libyuv/cpu_id.h" +#include "libyuv/planar_functions.h" // For CopyARGB +#include "libyuv/row.h" +#include "libyuv/scale_row.h" + +#ifdef __cplusplus +namespace libyuv { +extern "C" { +#endif + +static __inline int Abs(int v) { + return v >= 0 ? v : -v; +} + +// CPU agnostic row functions +void ScaleRowDown2_C(const uint8* src_ptr, ptrdiff_t src_stride, + uint8* dst, int dst_width) { + int x; + for (x = 0; x < dst_width - 1; x += 2) { + dst[0] = src_ptr[1]; + dst[1] = src_ptr[3]; + dst += 2; + src_ptr += 4; + } + if (dst_width & 1) { + dst[0] = src_ptr[1]; + } +} + +void ScaleRowDown2_16_C(const uint16* src_ptr, ptrdiff_t src_stride, + uint16* dst, int dst_width) { + int x; + for (x = 0; x < dst_width - 1; x += 2) { + dst[0] = src_ptr[1]; + dst[1] = src_ptr[3]; + dst += 2; + src_ptr += 4; + } + if (dst_width & 1) { + dst[0] = src_ptr[1]; + } +} + +void ScaleRowDown2Linear_C(const uint8* src_ptr, ptrdiff_t src_stride, + uint8* dst, int dst_width) { + const uint8* s = src_ptr; + int x; + for (x = 0; x < dst_width - 1; x += 2) { + dst[0] = (s[0] + s[1] + 1) >> 1; + dst[1] = (s[2] + s[3] + 1) >> 1; + dst += 2; + s += 4; + } + if (dst_width & 1) { + dst[0] = (s[0] + s[1] + 1) >> 1; + } +} + +void ScaleRowDown2Linear_16_C(const uint16* src_ptr, ptrdiff_t src_stride, + uint16* dst, int dst_width) { + const uint16* s = src_ptr; + int x; + for (x = 0; x < dst_width - 1; x += 2) { + dst[0] = (s[0] + s[1] + 1) >> 1; + dst[1] = (s[2] + s[3] + 1) >> 1; + dst += 2; + s += 4; + } + if (dst_width & 1) { + dst[0] = (s[0] + s[1] + 1) >> 1; + } +} + +void ScaleRowDown2Box_C(const uint8* src_ptr, ptrdiff_t src_stride, + uint8* dst, int dst_width) { + const uint8* s = src_ptr; + const uint8* t = src_ptr + src_stride; + int x; + for (x = 0; x < dst_width - 1; x += 2) { + dst[0] = (s[0] + s[1] + t[0] + t[1] + 2) >> 2; + dst[1] = (s[2] + s[3] + t[2] + t[3] + 2) >> 2; + dst += 2; + s += 4; + t += 4; + } + if (dst_width & 1) { + dst[0] = (s[0] + s[1] + t[0] + t[1] + 2) >> 2; + } +} + +void ScaleRowDown2Box_16_C(const uint16* src_ptr, ptrdiff_t src_stride, + uint16* dst, int dst_width) { + const uint16* s = src_ptr; + const uint16* t = src_ptr + src_stride; + int x; + for (x = 0; x < dst_width - 1; x += 2) { + dst[0] = (s[0] + s[1] + t[0] + t[1] + 2) >> 2; + dst[1] = (s[2] + s[3] + t[2] + t[3] + 2) >> 2; + dst += 2; + s += 4; + t += 4; + } + if (dst_width & 1) { + dst[0] = (s[0] + s[1] + t[0] + t[1] + 2) >> 2; + } +} + +void ScaleRowDown4_C(const uint8* src_ptr, ptrdiff_t src_stride, + uint8* dst, int dst_width) { + int x; + for (x = 0; x < dst_width - 1; x += 2) { + dst[0] = src_ptr[2]; + dst[1] = src_ptr[6]; + dst += 2; + src_ptr += 8; + } + if (dst_width & 1) { + dst[0] = src_ptr[2]; + } +} + +void ScaleRowDown4_16_C(const uint16* src_ptr, ptrdiff_t src_stride, + uint16* dst, int dst_width) { + int x; + for (x = 0; x < dst_width - 1; x += 2) { + dst[0] = src_ptr[2]; + dst[1] = src_ptr[6]; + dst += 2; + src_ptr += 8; + } + if (dst_width & 1) { + dst[0] = src_ptr[2]; + } +} + +void ScaleRowDown4Box_C(const uint8* src_ptr, ptrdiff_t src_stride, + uint8* dst, int dst_width) { + intptr_t stride = src_stride; + int x; + for (x = 0; x < dst_width - 1; x += 2) { + dst[0] = (src_ptr[0] + src_ptr[1] + src_ptr[2] + src_ptr[3] + + src_ptr[stride + 0] + src_ptr[stride + 1] + + src_ptr[stride + 2] + src_ptr[stride + 3] + + src_ptr[stride * 2 + 0] + src_ptr[stride * 2 + 1] + + src_ptr[stride * 2 + 2] + src_ptr[stride * 2 + 3] + + src_ptr[stride * 3 + 0] + src_ptr[stride * 3 + 1] + + src_ptr[stride * 3 + 2] + src_ptr[stride * 3 + 3] + + 8) >> 4; + dst[1] = (src_ptr[4] + src_ptr[5] + src_ptr[6] + src_ptr[7] + + src_ptr[stride + 4] + src_ptr[stride + 5] + + src_ptr[stride + 6] + src_ptr[stride + 7] + + src_ptr[stride * 2 + 4] + src_ptr[stride * 2 + 5] + + src_ptr[stride * 2 + 6] + src_ptr[stride * 2 + 7] + + src_ptr[stride * 3 + 4] + src_ptr[stride * 3 + 5] + + src_ptr[stride * 3 + 6] + src_ptr[stride * 3 + 7] + + 8) >> 4; + dst += 2; + src_ptr += 8; + } + if (dst_width & 1) { + dst[0] = (src_ptr[0] + src_ptr[1] + src_ptr[2] + src_ptr[3] + + src_ptr[stride + 0] + src_ptr[stride + 1] + + src_ptr[stride + 2] + src_ptr[stride + 3] + + src_ptr[stride * 2 + 0] + src_ptr[stride * 2 + 1] + + src_ptr[stride * 2 + 2] + src_ptr[stride * 2 + 3] + + src_ptr[stride * 3 + 0] + src_ptr[stride * 3 + 1] + + src_ptr[stride * 3 + 2] + src_ptr[stride * 3 + 3] + + 8) >> 4; + } +} + +void ScaleRowDown4Box_16_C(const uint16* src_ptr, ptrdiff_t src_stride, + uint16* dst, int dst_width) { + intptr_t stride = src_stride; + int x; + for (x = 0; x < dst_width - 1; x += 2) { + dst[0] = (src_ptr[0] + src_ptr[1] + src_ptr[2] + src_ptr[3] + + src_ptr[stride + 0] + src_ptr[stride + 1] + + src_ptr[stride + 2] + src_ptr[stride + 3] + + src_ptr[stride * 2 + 0] + src_ptr[stride * 2 + 1] + + src_ptr[stride * 2 + 2] + src_ptr[stride * 2 + 3] + + src_ptr[stride * 3 + 0] + src_ptr[stride * 3 + 1] + + src_ptr[stride * 3 + 2] + src_ptr[stride * 3 + 3] + + 8) >> 4; + dst[1] = (src_ptr[4] + src_ptr[5] + src_ptr[6] + src_ptr[7] + + src_ptr[stride + 4] + src_ptr[stride + 5] + + src_ptr[stride + 6] + src_ptr[stride + 7] + + src_ptr[stride * 2 + 4] + src_ptr[stride * 2 + 5] + + src_ptr[stride * 2 + 6] + src_ptr[stride * 2 + 7] + + src_ptr[stride * 3 + 4] + src_ptr[stride * 3 + 5] + + src_ptr[stride * 3 + 6] + src_ptr[stride * 3 + 7] + + 8) >> 4; + dst += 2; + src_ptr += 8; + } + if (dst_width & 1) { + dst[0] = (src_ptr[0] + src_ptr[1] + src_ptr[2] + src_ptr[3] + + src_ptr[stride + 0] + src_ptr[stride + 1] + + src_ptr[stride + 2] + src_ptr[stride + 3] + + src_ptr[stride * 2 + 0] + src_ptr[stride * 2 + 1] + + src_ptr[stride * 2 + 2] + src_ptr[stride * 2 + 3] + + src_ptr[stride * 3 + 0] + src_ptr[stride * 3 + 1] + + src_ptr[stride * 3 + 2] + src_ptr[stride * 3 + 3] + + 8) >> 4; + } +} + +void ScaleRowDown34_C(const uint8* src_ptr, ptrdiff_t src_stride, + uint8* dst, int dst_width) { + int x; + assert((dst_width % 3 == 0) && (dst_width > 0)); + for (x = 0; x < dst_width; x += 3) { + dst[0] = src_ptr[0]; + dst[1] = src_ptr[1]; + dst[2] = src_ptr[3]; + dst += 3; + src_ptr += 4; + } +} + +void ScaleRowDown34_16_C(const uint16* src_ptr, ptrdiff_t src_stride, + uint16* dst, int dst_width) { + int x; + assert((dst_width % 3 == 0) && (dst_width > 0)); + for (x = 0; x < dst_width; x += 3) { + dst[0] = src_ptr[0]; + dst[1] = src_ptr[1]; + dst[2] = src_ptr[3]; + dst += 3; + src_ptr += 4; + } +} + +// Filter rows 0 and 1 together, 3 : 1 +void ScaleRowDown34_0_Box_C(const uint8* src_ptr, ptrdiff_t src_stride, + uint8* d, int dst_width) { + const uint8* s = src_ptr; + const uint8* t = src_ptr + src_stride; + int x; + assert((dst_width % 3 == 0) && (dst_width > 0)); + for (x = 0; x < dst_width; x += 3) { + uint8 a0 = (s[0] * 3 + s[1] * 1 + 2) >> 2; + uint8 a1 = (s[1] * 1 + s[2] * 1 + 1) >> 1; + uint8 a2 = (s[2] * 1 + s[3] * 3 + 2) >> 2; + uint8 b0 = (t[0] * 3 + t[1] * 1 + 2) >> 2; + uint8 b1 = (t[1] * 1 + t[2] * 1 + 1) >> 1; + uint8 b2 = (t[2] * 1 + t[3] * 3 + 2) >> 2; + d[0] = (a0 * 3 + b0 + 2) >> 2; + d[1] = (a1 * 3 + b1 + 2) >> 2; + d[2] = (a2 * 3 + b2 + 2) >> 2; + d += 3; + s += 4; + t += 4; + } +} + +void ScaleRowDown34_0_Box_16_C(const uint16* src_ptr, ptrdiff_t src_stride, + uint16* d, int dst_width) { + const uint16* s = src_ptr; + const uint16* t = src_ptr + src_stride; + int x; + assert((dst_width % 3 == 0) && (dst_width > 0)); + for (x = 0; x < dst_width; x += 3) { + uint16 a0 = (s[0] * 3 + s[1] * 1 + 2) >> 2; + uint16 a1 = (s[1] * 1 + s[2] * 1 + 1) >> 1; + uint16 a2 = (s[2] * 1 + s[3] * 3 + 2) >> 2; + uint16 b0 = (t[0] * 3 + t[1] * 1 + 2) >> 2; + uint16 b1 = (t[1] * 1 + t[2] * 1 + 1) >> 1; + uint16 b2 = (t[2] * 1 + t[3] * 3 + 2) >> 2; + d[0] = (a0 * 3 + b0 + 2) >> 2; + d[1] = (a1 * 3 + b1 + 2) >> 2; + d[2] = (a2 * 3 + b2 + 2) >> 2; + d += 3; + s += 4; + t += 4; + } +} + +// Filter rows 1 and 2 together, 1 : 1 +void ScaleRowDown34_1_Box_C(const uint8* src_ptr, ptrdiff_t src_stride, + uint8* d, int dst_width) { + const uint8* s = src_ptr; + const uint8* t = src_ptr + src_stride; + int x; + assert((dst_width % 3 == 0) && (dst_width > 0)); + for (x = 0; x < dst_width; x += 3) { + uint8 a0 = (s[0] * 3 + s[1] * 1 + 2) >> 2; + uint8 a1 = (s[1] * 1 + s[2] * 1 + 1) >> 1; + uint8 a2 = (s[2] * 1 + s[3] * 3 + 2) >> 2; + uint8 b0 = (t[0] * 3 + t[1] * 1 + 2) >> 2; + uint8 b1 = (t[1] * 1 + t[2] * 1 + 1) >> 1; + uint8 b2 = (t[2] * 1 + t[3] * 3 + 2) >> 2; + d[0] = (a0 + b0 + 1) >> 1; + d[1] = (a1 + b1 + 1) >> 1; + d[2] = (a2 + b2 + 1) >> 1; + d += 3; + s += 4; + t += 4; + } +} + +void ScaleRowDown34_1_Box_16_C(const uint16* src_ptr, ptrdiff_t src_stride, + uint16* d, int dst_width) { + const uint16* s = src_ptr; + const uint16* t = src_ptr + src_stride; + int x; + assert((dst_width % 3 == 0) && (dst_width > 0)); + for (x = 0; x < dst_width; x += 3) { + uint16 a0 = (s[0] * 3 + s[1] * 1 + 2) >> 2; + uint16 a1 = (s[1] * 1 + s[2] * 1 + 1) >> 1; + uint16 a2 = (s[2] * 1 + s[3] * 3 + 2) >> 2; + uint16 b0 = (t[0] * 3 + t[1] * 1 + 2) >> 2; + uint16 b1 = (t[1] * 1 + t[2] * 1 + 1) >> 1; + uint16 b2 = (t[2] * 1 + t[3] * 3 + 2) >> 2; + d[0] = (a0 + b0 + 1) >> 1; + d[1] = (a1 + b1 + 1) >> 1; + d[2] = (a2 + b2 + 1) >> 1; + d += 3; + s += 4; + t += 4; + } +} + +// Scales a single row of pixels using point sampling. +void ScaleCols_C(uint8* dst_ptr, const uint8* src_ptr, + int dst_width, int x, int dx) { + int j; + for (j = 0; j < dst_width - 1; j += 2) { + dst_ptr[0] = src_ptr[x >> 16]; + x += dx; + dst_ptr[1] = src_ptr[x >> 16]; + x += dx; + dst_ptr += 2; + } + if (dst_width & 1) { + dst_ptr[0] = src_ptr[x >> 16]; + } +} + +void ScaleCols_16_C(uint16* dst_ptr, const uint16* src_ptr, + int dst_width, int x, int dx) { + int j; + for (j = 0; j < dst_width - 1; j += 2) { + dst_ptr[0] = src_ptr[x >> 16]; + x += dx; + dst_ptr[1] = src_ptr[x >> 16]; + x += dx; + dst_ptr += 2; + } + if (dst_width & 1) { + dst_ptr[0] = src_ptr[x >> 16]; + } +} + +// Scales a single row of pixels up by 2x using point sampling. +void ScaleColsUp2_C(uint8* dst_ptr, const uint8* src_ptr, + int dst_width, int x, int dx) { + int j; + for (j = 0; j < dst_width - 1; j += 2) { + dst_ptr[1] = dst_ptr[0] = src_ptr[0]; + src_ptr += 1; + dst_ptr += 2; + } + if (dst_width & 1) { + dst_ptr[0] = src_ptr[0]; + } +} + +void ScaleColsUp2_16_C(uint16* dst_ptr, const uint16* src_ptr, + int dst_width, int x, int dx) { + int j; + for (j = 0; j < dst_width - 1; j += 2) { + dst_ptr[1] = dst_ptr[0] = src_ptr[0]; + src_ptr += 1; + dst_ptr += 2; + } + if (dst_width & 1) { + dst_ptr[0] = src_ptr[0]; + } +} + +// (1-f)a + fb can be replaced with a + f(b-a) +#define BLENDER(a, b, f) (uint8)((int)(a) + \ + ((int)(f) * ((int)(b) - (int)(a)) >> 16)) + +void ScaleFilterCols_C(uint8* dst_ptr, const uint8* src_ptr, + int dst_width, int x, int dx) { + int j; + for (j = 0; j < dst_width - 1; j += 2) { + int xi = x >> 16; + int a = src_ptr[xi]; + int b = src_ptr[xi + 1]; + dst_ptr[0] = BLENDER(a, b, x & 0xffff); + x += dx; + xi = x >> 16; + a = src_ptr[xi]; + b = src_ptr[xi + 1]; + dst_ptr[1] = BLENDER(a, b, x & 0xffff); + x += dx; + dst_ptr += 2; + } + if (dst_width & 1) { + int xi = x >> 16; + int a = src_ptr[xi]; + int b = src_ptr[xi + 1]; + dst_ptr[0] = BLENDER(a, b, x & 0xffff); + } +} + +void ScaleFilterCols64_C(uint8* dst_ptr, const uint8* src_ptr, + int dst_width, int x32, int dx) { + int64 x = (int64)(x32); + int j; + for (j = 0; j < dst_width - 1; j += 2) { + int64 xi = x >> 16; + int a = src_ptr[xi]; + int b = src_ptr[xi + 1]; + dst_ptr[0] = BLENDER(a, b, x & 0xffff); + x += dx; + xi = x >> 16; + a = src_ptr[xi]; + b = src_ptr[xi + 1]; + dst_ptr[1] = BLENDER(a, b, x & 0xffff); + x += dx; + dst_ptr += 2; + } + if (dst_width & 1) { + int64 xi = x >> 16; + int a = src_ptr[xi]; + int b = src_ptr[xi + 1]; + dst_ptr[0] = BLENDER(a, b, x & 0xffff); + } +} +#undef BLENDER + +#define BLENDER(a, b, f) (uint16)((int)(a) + \ + ((int)(f) * ((int)(b) - (int)(a)) >> 16)) + +void ScaleFilterCols_16_C(uint16* dst_ptr, const uint16* src_ptr, + int dst_width, int x, int dx) { + int j; + for (j = 0; j < dst_width - 1; j += 2) { + int xi = x >> 16; + int a = src_ptr[xi]; + int b = src_ptr[xi + 1]; + dst_ptr[0] = BLENDER(a, b, x & 0xffff); + x += dx; + xi = x >> 16; + a = src_ptr[xi]; + b = src_ptr[xi + 1]; + dst_ptr[1] = BLENDER(a, b, x & 0xffff); + x += dx; + dst_ptr += 2; + } + if (dst_width & 1) { + int xi = x >> 16; + int a = src_ptr[xi]; + int b = src_ptr[xi + 1]; + dst_ptr[0] = BLENDER(a, b, x & 0xffff); + } +} + +void ScaleFilterCols64_16_C(uint16* dst_ptr, const uint16* src_ptr, + int dst_width, int x32, int dx) { + int64 x = (int64)(x32); + int j; + for (j = 0; j < dst_width - 1; j += 2) { + int64 xi = x >> 16; + int a = src_ptr[xi]; + int b = src_ptr[xi + 1]; + dst_ptr[0] = BLENDER(a, b, x & 0xffff); + x += dx; + xi = x >> 16; + a = src_ptr[xi]; + b = src_ptr[xi + 1]; + dst_ptr[1] = BLENDER(a, b, x & 0xffff); + x += dx; + dst_ptr += 2; + } + if (dst_width & 1) { + int64 xi = x >> 16; + int a = src_ptr[xi]; + int b = src_ptr[xi + 1]; + dst_ptr[0] = BLENDER(a, b, x & 0xffff); + } +} +#undef BLENDER + +void ScaleRowDown38_C(const uint8* src_ptr, ptrdiff_t src_stride, + uint8* dst, int dst_width) { + int x; + assert(dst_width % 3 == 0); + for (x = 0; x < dst_width; x += 3) { + dst[0] = src_ptr[0]; + dst[1] = src_ptr[3]; + dst[2] = src_ptr[6]; + dst += 3; + src_ptr += 8; + } +} + +void ScaleRowDown38_16_C(const uint16* src_ptr, ptrdiff_t src_stride, + uint16* dst, int dst_width) { + int x; + assert(dst_width % 3 == 0); + for (x = 0; x < dst_width; x += 3) { + dst[0] = src_ptr[0]; + dst[1] = src_ptr[3]; + dst[2] = src_ptr[6]; + dst += 3; + src_ptr += 8; + } +} + +// 8x3 -> 3x1 +void ScaleRowDown38_3_Box_C(const uint8* src_ptr, + ptrdiff_t src_stride, + uint8* dst_ptr, int dst_width) { + intptr_t stride = src_stride; + int i; + assert((dst_width % 3 == 0) && (dst_width > 0)); + for (i = 0; i < dst_width; i += 3) { + dst_ptr[0] = (src_ptr[0] + src_ptr[1] + src_ptr[2] + + src_ptr[stride + 0] + src_ptr[stride + 1] + + src_ptr[stride + 2] + src_ptr[stride * 2 + 0] + + src_ptr[stride * 2 + 1] + src_ptr[stride * 2 + 2]) * + (65536 / 9) >> 16; + dst_ptr[1] = (src_ptr[3] + src_ptr[4] + src_ptr[5] + + src_ptr[stride + 3] + src_ptr[stride + 4] + + src_ptr[stride + 5] + src_ptr[stride * 2 + 3] + + src_ptr[stride * 2 + 4] + src_ptr[stride * 2 + 5]) * + (65536 / 9) >> 16; + dst_ptr[2] = (src_ptr[6] + src_ptr[7] + + src_ptr[stride + 6] + src_ptr[stride + 7] + + src_ptr[stride * 2 + 6] + src_ptr[stride * 2 + 7]) * + (65536 / 6) >> 16; + src_ptr += 8; + dst_ptr += 3; + } +} + +void ScaleRowDown38_3_Box_16_C(const uint16* src_ptr, + ptrdiff_t src_stride, + uint16* dst_ptr, int dst_width) { + intptr_t stride = src_stride; + int i; + assert((dst_width % 3 == 0) && (dst_width > 0)); + for (i = 0; i < dst_width; i += 3) { + dst_ptr[0] = (src_ptr[0] + src_ptr[1] + src_ptr[2] + + src_ptr[stride + 0] + src_ptr[stride + 1] + + src_ptr[stride + 2] + src_ptr[stride * 2 + 0] + + src_ptr[stride * 2 + 1] + src_ptr[stride * 2 + 2]) * + (65536 / 9) >> 16; + dst_ptr[1] = (src_ptr[3] + src_ptr[4] + src_ptr[5] + + src_ptr[stride + 3] + src_ptr[stride + 4] + + src_ptr[stride + 5] + src_ptr[stride * 2 + 3] + + src_ptr[stride * 2 + 4] + src_ptr[stride * 2 + 5]) * + (65536 / 9) >> 16; + dst_ptr[2] = (src_ptr[6] + src_ptr[7] + + src_ptr[stride + 6] + src_ptr[stride + 7] + + src_ptr[stride * 2 + 6] + src_ptr[stride * 2 + 7]) * + (65536 / 6) >> 16; + src_ptr += 8; + dst_ptr += 3; + } +} + +// 8x2 -> 3x1 +void ScaleRowDown38_2_Box_C(const uint8* src_ptr, ptrdiff_t src_stride, + uint8* dst_ptr, int dst_width) { + intptr_t stride = src_stride; + int i; + assert((dst_width % 3 == 0) && (dst_width > 0)); + for (i = 0; i < dst_width; i += 3) { + dst_ptr[0] = (src_ptr[0] + src_ptr[1] + src_ptr[2] + + src_ptr[stride + 0] + src_ptr[stride + 1] + + src_ptr[stride + 2]) * (65536 / 6) >> 16; + dst_ptr[1] = (src_ptr[3] + src_ptr[4] + src_ptr[5] + + src_ptr[stride + 3] + src_ptr[stride + 4] + + src_ptr[stride + 5]) * (65536 / 6) >> 16; + dst_ptr[2] = (src_ptr[6] + src_ptr[7] + + src_ptr[stride + 6] + src_ptr[stride + 7]) * + (65536 / 4) >> 16; + src_ptr += 8; + dst_ptr += 3; + } +} + +void ScaleRowDown38_2_Box_16_C(const uint16* src_ptr, ptrdiff_t src_stride, + uint16* dst_ptr, int dst_width) { + intptr_t stride = src_stride; + int i; + assert((dst_width % 3 == 0) && (dst_width > 0)); + for (i = 0; i < dst_width; i += 3) { + dst_ptr[0] = (src_ptr[0] + src_ptr[1] + src_ptr[2] + + src_ptr[stride + 0] + src_ptr[stride + 1] + + src_ptr[stride + 2]) * (65536 / 6) >> 16; + dst_ptr[1] = (src_ptr[3] + src_ptr[4] + src_ptr[5] + + src_ptr[stride + 3] + src_ptr[stride + 4] + + src_ptr[stride + 5]) * (65536 / 6) >> 16; + dst_ptr[2] = (src_ptr[6] + src_ptr[7] + + src_ptr[stride + 6] + src_ptr[stride + 7]) * + (65536 / 4) >> 16; + src_ptr += 8; + dst_ptr += 3; + } +} + +void ScaleAddRow_C(const uint8* src_ptr, uint16* dst_ptr, int src_width) { + int x; + assert(src_width > 0); + for (x = 0; x < src_width - 1; x += 2) { + dst_ptr[0] += src_ptr[0]; + dst_ptr[1] += src_ptr[1]; + src_ptr += 2; + dst_ptr += 2; + } + if (src_width & 1) { + dst_ptr[0] += src_ptr[0]; + } +} + +void ScaleAddRow_16_C(const uint16* src_ptr, uint32* dst_ptr, int src_width) { + int x; + assert(src_width > 0); + for (x = 0; x < src_width - 1; x += 2) { + dst_ptr[0] += src_ptr[0]; + dst_ptr[1] += src_ptr[1]; + src_ptr += 2; + dst_ptr += 2; + } + if (src_width & 1) { + dst_ptr[0] += src_ptr[0]; + } +} + +void ScaleARGBRowDown2_C(const uint8* src_argb, + ptrdiff_t src_stride, + uint8* dst_argb, int dst_width) { + const uint32* src = (const uint32*)(src_argb); + uint32* dst = (uint32*)(dst_argb); + + int x; + for (x = 0; x < dst_width - 1; x += 2) { + dst[0] = src[1]; + dst[1] = src[3]; + src += 4; + dst += 2; + } + if (dst_width & 1) { + dst[0] = src[1]; + } +} + +void ScaleARGBRowDown2Linear_C(const uint8* src_argb, + ptrdiff_t src_stride, + uint8* dst_argb, int dst_width) { + int x; + for (x = 0; x < dst_width; ++x) { + dst_argb[0] = (src_argb[0] + src_argb[4] + 1) >> 1; + dst_argb[1] = (src_argb[1] + src_argb[5] + 1) >> 1; + dst_argb[2] = (src_argb[2] + src_argb[6] + 1) >> 1; + dst_argb[3] = (src_argb[3] + src_argb[7] + 1) >> 1; + src_argb += 8; + dst_argb += 4; + } +} + +void ScaleARGBRowDown2Box_C(const uint8* src_argb, ptrdiff_t src_stride, + uint8* dst_argb, int dst_width) { + int x; + for (x = 0; x < dst_width; ++x) { + dst_argb[0] = (src_argb[0] + src_argb[4] + + src_argb[src_stride] + src_argb[src_stride + 4] + 2) >> 2; + dst_argb[1] = (src_argb[1] + src_argb[5] + + src_argb[src_stride + 1] + src_argb[src_stride + 5] + 2) >> 2; + dst_argb[2] = (src_argb[2] + src_argb[6] + + src_argb[src_stride + 2] + src_argb[src_stride + 6] + 2) >> 2; + dst_argb[3] = (src_argb[3] + src_argb[7] + + src_argb[src_stride + 3] + src_argb[src_stride + 7] + 2) >> 2; + src_argb += 8; + dst_argb += 4; + } +} + +void ScaleARGBRowDownEven_C(const uint8* src_argb, ptrdiff_t src_stride, + int src_stepx, + uint8* dst_argb, int dst_width) { + const uint32* src = (const uint32*)(src_argb); + uint32* dst = (uint32*)(dst_argb); + + int x; + for (x = 0; x < dst_width - 1; x += 2) { + dst[0] = src[0]; + dst[1] = src[src_stepx]; + src += src_stepx * 2; + dst += 2; + } + if (dst_width & 1) { + dst[0] = src[0]; + } +} + +void ScaleARGBRowDownEvenBox_C(const uint8* src_argb, + ptrdiff_t src_stride, + int src_stepx, + uint8* dst_argb, int dst_width) { + int x; + for (x = 0; x < dst_width; ++x) { + dst_argb[0] = (src_argb[0] + src_argb[4] + + src_argb[src_stride] + src_argb[src_stride + 4] + 2) >> 2; + dst_argb[1] = (src_argb[1] + src_argb[5] + + src_argb[src_stride + 1] + src_argb[src_stride + 5] + 2) >> 2; + dst_argb[2] = (src_argb[2] + src_argb[6] + + src_argb[src_stride + 2] + src_argb[src_stride + 6] + 2) >> 2; + dst_argb[3] = (src_argb[3] + src_argb[7] + + src_argb[src_stride + 3] + src_argb[src_stride + 7] + 2) >> 2; + src_argb += src_stepx * 4; + dst_argb += 4; + } +} + +// Scales a single row of pixels using point sampling. +void ScaleARGBCols_C(uint8* dst_argb, const uint8* src_argb, + int dst_width, int x, int dx) { + const uint32* src = (const uint32*)(src_argb); + uint32* dst = (uint32*)(dst_argb); + int j; + for (j = 0; j < dst_width - 1; j += 2) { + dst[0] = src[x >> 16]; + x += dx; + dst[1] = src[x >> 16]; + x += dx; + dst += 2; + } + if (dst_width & 1) { + dst[0] = src[x >> 16]; + } +} + +void ScaleARGBCols64_C(uint8* dst_argb, const uint8* src_argb, + int dst_width, int x32, int dx) { + int64 x = (int64)(x32); + const uint32* src = (const uint32*)(src_argb); + uint32* dst = (uint32*)(dst_argb); + int j; + for (j = 0; j < dst_width - 1; j += 2) { + dst[0] = src[x >> 16]; + x += dx; + dst[1] = src[x >> 16]; + x += dx; + dst += 2; + } + if (dst_width & 1) { + dst[0] = src[x >> 16]; + } +} + +// Scales a single row of pixels up by 2x using point sampling. +void ScaleARGBColsUp2_C(uint8* dst_argb, const uint8* src_argb, + int dst_width, int x, int dx) { + const uint32* src = (const uint32*)(src_argb); + uint32* dst = (uint32*)(dst_argb); + int j; + for (j = 0; j < dst_width - 1; j += 2) { + dst[1] = dst[0] = src[0]; + src += 1; + dst += 2; + } + if (dst_width & 1) { + dst[0] = src[0]; + } +} + +// Mimics SSSE3 blender +#define BLENDER1(a, b, f) ((a) * (0x7f ^ f) + (b) * f) >> 7 +#define BLENDERC(a, b, f, s) (uint32)( \ + BLENDER1(((a) >> s) & 255, ((b) >> s) & 255, f) << s) +#define BLENDER(a, b, f) \ + BLENDERC(a, b, f, 24) | BLENDERC(a, b, f, 16) | \ + BLENDERC(a, b, f, 8) | BLENDERC(a, b, f, 0) + +void ScaleARGBFilterCols_C(uint8* dst_argb, const uint8* src_argb, + int dst_width, int x, int dx) { + const uint32* src = (const uint32*)(src_argb); + uint32* dst = (uint32*)(dst_argb); + int j; + for (j = 0; j < dst_width - 1; j += 2) { + int xi = x >> 16; + int xf = (x >> 9) & 0x7f; + uint32 a = src[xi]; + uint32 b = src[xi + 1]; + dst[0] = BLENDER(a, b, xf); + x += dx; + xi = x >> 16; + xf = (x >> 9) & 0x7f; + a = src[xi]; + b = src[xi + 1]; + dst[1] = BLENDER(a, b, xf); + x += dx; + dst += 2; + } + if (dst_width & 1) { + int xi = x >> 16; + int xf = (x >> 9) & 0x7f; + uint32 a = src[xi]; + uint32 b = src[xi + 1]; + dst[0] = BLENDER(a, b, xf); + } +} + +void ScaleARGBFilterCols64_C(uint8* dst_argb, const uint8* src_argb, + int dst_width, int x32, int dx) { + int64 x = (int64)(x32); + const uint32* src = (const uint32*)(src_argb); + uint32* dst = (uint32*)(dst_argb); + int j; + for (j = 0; j < dst_width - 1; j += 2) { + int64 xi = x >> 16; + int xf = (x >> 9) & 0x7f; + uint32 a = src[xi]; + uint32 b = src[xi + 1]; + dst[0] = BLENDER(a, b, xf); + x += dx; + xi = x >> 16; + xf = (x >> 9) & 0x7f; + a = src[xi]; + b = src[xi + 1]; + dst[1] = BLENDER(a, b, xf); + x += dx; + dst += 2; + } + if (dst_width & 1) { + int64 xi = x >> 16; + int xf = (x >> 9) & 0x7f; + uint32 a = src[xi]; + uint32 b = src[xi + 1]; + dst[0] = BLENDER(a, b, xf); + } +} +#undef BLENDER1 +#undef BLENDERC +#undef BLENDER + +// Scale plane vertically with bilinear interpolation. +void ScalePlaneVertical(int src_height, + int dst_width, int dst_height, + int src_stride, int dst_stride, + const uint8* src_argb, uint8* dst_argb, + int x, int y, int dy, + int bpp, enum FilterMode filtering) { + // TODO(fbarchard): Allow higher bpp. + int dst_width_bytes = dst_width * bpp; + void (*InterpolateRow)(uint8* dst_argb, const uint8* src_argb, + ptrdiff_t src_stride, int dst_width, int source_y_fraction) = + InterpolateRow_C; + const int max_y = (src_height > 1) ? ((src_height - 1) << 16) - 1 : 0; + int j; + assert(bpp >= 1 && bpp <= 4); + assert(src_height != 0); + assert(dst_width > 0); + assert(dst_height > 0); + src_argb += (x >> 16) * bpp; +#if defined(HAS_INTERPOLATEROW_SSE2) + if (TestCpuFlag(kCpuHasSSE2)) { + InterpolateRow = InterpolateRow_Any_SSE2; + if (IS_ALIGNED(dst_width_bytes, 16)) { + InterpolateRow = InterpolateRow_SSE2; + } + } +#endif +#if defined(HAS_INTERPOLATEROW_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + InterpolateRow = InterpolateRow_Any_SSSE3; + if (IS_ALIGNED(dst_width_bytes, 16)) { + InterpolateRow = InterpolateRow_SSSE3; + } + } +#endif +#if defined(HAS_INTERPOLATEROW_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + InterpolateRow = InterpolateRow_Any_AVX2; + if (IS_ALIGNED(dst_width_bytes, 32)) { + InterpolateRow = InterpolateRow_AVX2; + } + } +#endif +#if defined(HAS_INTERPOLATEROW_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + InterpolateRow = InterpolateRow_Any_NEON; + if (IS_ALIGNED(dst_width_bytes, 16)) { + InterpolateRow = InterpolateRow_NEON; + } + } +#endif +#if defined(HAS_INTERPOLATEROW_MIPS_DSPR2) + if (TestCpuFlag(kCpuHasMIPS_DSPR2) && + IS_ALIGNED(src_argb, 4) && IS_ALIGNED(src_stride, 4) && + IS_ALIGNED(dst_argb, 4) && IS_ALIGNED(dst_stride, 4)) { + InterpolateRow = InterpolateRow_Any_MIPS_DSPR2; + if (IS_ALIGNED(dst_width_bytes, 4)) { + InterpolateRow = InterpolateRow_MIPS_DSPR2; + } + } +#endif + for (j = 0; j < dst_height; ++j) { + int yi; + int yf; + if (y > max_y) { + y = max_y; + } + yi = y >> 16; + yf = filtering ? ((y >> 8) & 255) : 0; + InterpolateRow(dst_argb, src_argb + yi * src_stride, + src_stride, dst_width_bytes, yf); + dst_argb += dst_stride; + y += dy; + } +} +void ScalePlaneVertical_16(int src_height, + int dst_width, int dst_height, + int src_stride, int dst_stride, + const uint16* src_argb, uint16* dst_argb, + int x, int y, int dy, + int wpp, enum FilterMode filtering) { + // TODO(fbarchard): Allow higher wpp. + int dst_width_words = dst_width * wpp; + void (*InterpolateRow)(uint16* dst_argb, const uint16* src_argb, + ptrdiff_t src_stride, int dst_width, int source_y_fraction) = + InterpolateRow_16_C; + const int max_y = (src_height > 1) ? ((src_height - 1) << 16) - 1 : 0; + int j; + assert(wpp >= 1 && wpp <= 2); + assert(src_height != 0); + assert(dst_width > 0); + assert(dst_height > 0); + src_argb += (x >> 16) * wpp; +#if defined(HAS_INTERPOLATEROW_16_SSE2) + if (TestCpuFlag(kCpuHasSSE2)) { + InterpolateRow = InterpolateRow_Any_16_SSE2; + if (IS_ALIGNED(dst_width_bytes, 16)) { + InterpolateRow = InterpolateRow_16_SSE2; + } + } +#endif +#if defined(HAS_INTERPOLATEROW_16_SSSE3) + if (TestCpuFlag(kCpuHasSSSE3)) { + InterpolateRow = InterpolateRow_Any_16_SSSE3; + if (IS_ALIGNED(dst_width_bytes, 16)) { + InterpolateRow = InterpolateRow_16_SSSE3; + } + } +#endif +#if defined(HAS_INTERPOLATEROW_16_AVX2) + if (TestCpuFlag(kCpuHasAVX2)) { + InterpolateRow = InterpolateRow_Any_16_AVX2; + if (IS_ALIGNED(dst_width_bytes, 32)) { + InterpolateRow = InterpolateRow_16_AVX2; + } + } +#endif +#if defined(HAS_INTERPOLATEROW_16_NEON) + if (TestCpuFlag(kCpuHasNEON)) { + InterpolateRow = InterpolateRow_Any_16_NEON; + if (IS_ALIGNED(dst_width_bytes, 16)) { + InterpolateRow = InterpolateRow_16_NEON; + } + } +#endif +#if defined(HAS_INTERPOLATEROW_16_MIPS_DSPR2) + if (TestCpuFlag(kCpuHasMIPS_DSPR2) && + IS_ALIGNED(src_argb, 4) && IS_ALIGNED(src_stride, 4) && + IS_ALIGNED(dst_argb, 4) && IS_ALIGNED(dst_stride, 4)) { + InterpolateRow = InterpolateRow_Any_16_MIPS_DSPR2; + if (IS_ALIGNED(dst_width_bytes, 4)) { + InterpolateRow = InterpolateRow_16_MIPS_DSPR2; + } + } +#endif + for (j = 0; j < dst_height; ++j) { + int yi; + int yf; + if (y > max_y) { + y = max_y; + } + yi = y >> 16; + yf = filtering ? ((y >> 8) & 255) : 0; + InterpolateRow(dst_argb, src_argb + yi * src_stride, + src_stride, dst_width_words, yf); + dst_argb += dst_stride; + y += dy; + } +} + +// Simplify the filtering based on scale factors. +enum FilterMode ScaleFilterReduce(int src_width, int src_height, + int dst_width, int dst_height, + enum FilterMode filtering) { + if (src_width < 0) { + src_width = -src_width; + } + if (src_height < 0) { + src_height = -src_height; + } + if (filtering == kFilterBox) { + // If scaling both axis to 0.5 or larger, switch from Box to Bilinear. + if (dst_width * 2 >= src_width && dst_height * 2 >= src_height) { + filtering = kFilterBilinear; + } + } + if (filtering == kFilterBilinear) { + if (src_height == 1) { + filtering = kFilterLinear; + } + // TODO(fbarchard): Detect any odd scale factor and reduce to Linear. + if (dst_height == src_height || dst_height * 3 == src_height) { + filtering = kFilterLinear; + } + // TODO(fbarchard): Remove 1 pixel wide filter restriction, which is to + // avoid reading 2 pixels horizontally that causes memory exception. + if (src_width == 1) { + filtering = kFilterNone; + } + } + if (filtering == kFilterLinear) { + if (src_width == 1) { + filtering = kFilterNone; + } + // TODO(fbarchard): Detect any odd scale factor and reduce to None. + if (dst_width == src_width || dst_width * 3 == src_width) { + filtering = kFilterNone; + } + } + return filtering; +} + +// Divide num by div and return as 16.16 fixed point result. +int FixedDiv_C(int num, int div) { + return (int)(((int64)(num) << 16) / div); +} + +// Divide num by div and return as 16.16 fixed point result. +int FixedDiv1_C(int num, int div) { + return (int)((((int64)(num) << 16) - 0x00010001) / + (div - 1)); +} + +#define CENTERSTART(dx, s) (dx < 0) ? -((-dx >> 1) + s) : ((dx >> 1) + s) + +// Compute slope values for stepping. +void ScaleSlope(int src_width, int src_height, + int dst_width, int dst_height, + enum FilterMode filtering, + int* x, int* y, int* dx, int* dy) { + assert(x != NULL); + assert(y != NULL); + assert(dx != NULL); + assert(dy != NULL); + assert(src_width != 0); + assert(src_height != 0); + assert(dst_width > 0); + assert(dst_height > 0); + // Check for 1 pixel and avoid FixedDiv overflow. + if (dst_width == 1 && src_width >= 32768) { + dst_width = src_width; + } + if (dst_height == 1 && src_height >= 32768) { + dst_height = src_height; + } + if (filtering == kFilterBox) { + // Scale step for point sampling duplicates all pixels equally. + *dx = FixedDiv(Abs(src_width), dst_width); + *dy = FixedDiv(src_height, dst_height); + *x = 0; + *y = 0; + } else if (filtering == kFilterBilinear) { + // Scale step for bilinear sampling renders last pixel once for upsample. + if (dst_width <= Abs(src_width)) { + *dx = FixedDiv(Abs(src_width), dst_width); + *x = CENTERSTART(*dx, -32768); // Subtract 0.5 (32768) to center filter. + } else if (dst_width > 1) { + *dx = FixedDiv1(Abs(src_width), dst_width); + *x = 0; + } + if (dst_height <= src_height) { + *dy = FixedDiv(src_height, dst_height); + *y = CENTERSTART(*dy, -32768); // Subtract 0.5 (32768) to center filter. + } else if (dst_height > 1) { + *dy = FixedDiv1(src_height, dst_height); + *y = 0; + } + } else if (filtering == kFilterLinear) { + // Scale step for bilinear sampling renders last pixel once for upsample. + if (dst_width <= Abs(src_width)) { + *dx = FixedDiv(Abs(src_width), dst_width); + *x = CENTERSTART(*dx, -32768); // Subtract 0.5 (32768) to center filter. + } else if (dst_width > 1) { + *dx = FixedDiv1(Abs(src_width), dst_width); + *x = 0; + } + *dy = FixedDiv(src_height, dst_height); + *y = *dy >> 1; + } else { + // Scale step for point sampling duplicates all pixels equally. + *dx = FixedDiv(Abs(src_width), dst_width); + *dy = FixedDiv(src_height, dst_height); + *x = CENTERSTART(*dx, 0); + *y = CENTERSTART(*dy, 0); + } + // Negative src_width means horizontally mirror. + if (src_width < 0) { + *x += (dst_width - 1) * *dx; + *dx = -*dx; + // src_width = -src_width; // Caller must do this. + } +} +#undef CENTERSTART + +#ifdef __cplusplus +} // extern "C" +} // namespace libyuv +#endif diff --git a/media/libaom/src/third_party/libyuv/source/scale_gcc.cc b/media/libaom/src/third_party/libyuv/source/scale_gcc.cc new file mode 100644 index 000000000..8a6ac5459 --- /dev/null +++ b/media/libaom/src/third_party/libyuv/source/scale_gcc.cc @@ -0,0 +1,1089 @@ +/* + * Copyright 2013 The LibYuv Project Authors. All rights reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#include "libyuv/row.h" + +#ifdef __cplusplus +namespace libyuv { +extern "C" { +#endif + +// This module is for GCC x86 and x64. +#if !defined(LIBYUV_DISABLE_X86) && (defined(__x86_64__) || defined(__i386__)) + +// Offsets for source bytes 0 to 9 +static uvec8 kShuf0 = + { 0, 1, 3, 4, 5, 7, 8, 9, 128, 128, 128, 128, 128, 128, 128, 128 }; + +// Offsets for source bytes 11 to 20 with 8 subtracted = 3 to 12. +static uvec8 kShuf1 = + { 3, 4, 5, 7, 8, 9, 11, 12, 128, 128, 128, 128, 128, 128, 128, 128 }; + +// Offsets for source bytes 21 to 31 with 16 subtracted = 5 to 31. +static uvec8 kShuf2 = + { 5, 7, 8, 9, 11, 12, 13, 15, 128, 128, 128, 128, 128, 128, 128, 128 }; + +// Offsets for source bytes 0 to 10 +static uvec8 kShuf01 = + { 0, 1, 1, 2, 2, 3, 4, 5, 5, 6, 6, 7, 8, 9, 9, 10 }; + +// Offsets for source bytes 10 to 21 with 8 subtracted = 3 to 13. +static uvec8 kShuf11 = + { 2, 3, 4, 5, 5, 6, 6, 7, 8, 9, 9, 10, 10, 11, 12, 13 }; + +// Offsets for source bytes 21 to 31 with 16 subtracted = 5 to 31. +static uvec8 kShuf21 = + { 5, 6, 6, 7, 8, 9, 9, 10, 10, 11, 12, 13, 13, 14, 14, 15 }; + +// Coefficients for source bytes 0 to 10 +static uvec8 kMadd01 = + { 3, 1, 2, 2, 1, 3, 3, 1, 2, 2, 1, 3, 3, 1, 2, 2 }; + +// Coefficients for source bytes 10 to 21 +static uvec8 kMadd11 = + { 1, 3, 3, 1, 2, 2, 1, 3, 3, 1, 2, 2, 1, 3, 3, 1 }; + +// Coefficients for source bytes 21 to 31 +static uvec8 kMadd21 = + { 2, 2, 1, 3, 3, 1, 2, 2, 1, 3, 3, 1, 2, 2, 1, 3 }; + +// Coefficients for source bytes 21 to 31 +static vec16 kRound34 = + { 2, 2, 2, 2, 2, 2, 2, 2 }; + +static uvec8 kShuf38a = + { 0, 3, 6, 8, 11, 14, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128 }; + +static uvec8 kShuf38b = + { 128, 128, 128, 128, 128, 128, 0, 3, 6, 8, 11, 14, 128, 128, 128, 128 }; + +// Arrange words 0,3,6 into 0,1,2 +static uvec8 kShufAc = + { 0, 1, 6, 7, 12, 13, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128 }; + +// Arrange words 0,3,6 into 3,4,5 +static uvec8 kShufAc3 = + { 128, 128, 128, 128, 128, 128, 0, 1, 6, 7, 12, 13, 128, 128, 128, 128 }; + +// Scaling values for boxes of 3x3 and 2x3 +static uvec16 kScaleAc33 = + { 65536 / 9, 65536 / 9, 65536 / 6, 65536 / 9, 65536 / 9, 65536 / 6, 0, 0 }; + +// Arrange first value for pixels 0,1,2,3,4,5 +static uvec8 kShufAb0 = + { 0, 128, 3, 128, 6, 128, 8, 128, 11, 128, 14, 128, 128, 128, 128, 128 }; + +// Arrange second value for pixels 0,1,2,3,4,5 +static uvec8 kShufAb1 = + { 1, 128, 4, 128, 7, 128, 9, 128, 12, 128, 15, 128, 128, 128, 128, 128 }; + +// Arrange third value for pixels 0,1,2,3,4,5 +static uvec8 kShufAb2 = + { 2, 128, 5, 128, 128, 128, 10, 128, 13, 128, 128, 128, 128, 128, 128, 128 }; + +// Scaling values for boxes of 3x2 and 2x2 +static uvec16 kScaleAb2 = + { 65536 / 3, 65536 / 3, 65536 / 2, 65536 / 3, 65536 / 3, 65536 / 2, 0, 0 }; + +// GCC versions of row functions are verbatim conversions from Visual C. +// Generated using gcc disassembly on Visual C object file: +// objdump -D yuvscaler.obj >yuvscaler.txt + +void ScaleRowDown2_SSE2(const uint8* src_ptr, ptrdiff_t src_stride, + uint8* dst_ptr, int dst_width) { + asm volatile ( + LABELALIGN + "1: \n" + "movdqu " MEMACCESS(0) ",%%xmm0 \n" + "movdqu " MEMACCESS2(0x10,0) ",%%xmm1 \n" + "lea " MEMLEA(0x20,0) ",%0 \n" + "psrlw $0x8,%%xmm0 \n" + "psrlw $0x8,%%xmm1 \n" + "packuswb %%xmm1,%%xmm0 \n" + "movdqu %%xmm0," MEMACCESS(1) " \n" + "lea " MEMLEA(0x10,1) ",%1 \n" + "sub $0x10,%2 \n" + "jg 1b \n" + : "+r"(src_ptr), // %0 + "+r"(dst_ptr), // %1 + "+r"(dst_width) // %2 + :: "memory", "cc", "xmm0", "xmm1" + ); +} + +void ScaleRowDown2Linear_SSE2(const uint8* src_ptr, ptrdiff_t src_stride, + uint8* dst_ptr, int dst_width) { + asm volatile ( + "pcmpeqb %%xmm5,%%xmm5 \n" + "psrlw $0x8,%%xmm5 \n" + + LABELALIGN + "1: \n" + "movdqu " MEMACCESS(0) ",%%xmm0 \n" + "movdqu " MEMACCESS2(0x10, 0) ",%%xmm1 \n" + "lea " MEMLEA(0x20,0) ",%0 \n" + "movdqa %%xmm0,%%xmm2 \n" + "psrlw $0x8,%%xmm0 \n" + "movdqa %%xmm1,%%xmm3 \n" + "psrlw $0x8,%%xmm1 \n" + "pand %%xmm5,%%xmm2 \n" + "pand %%xmm5,%%xmm3 \n" + "pavgw %%xmm2,%%xmm0 \n" + "pavgw %%xmm3,%%xmm1 \n" + "packuswb %%xmm1,%%xmm0 \n" + "movdqu %%xmm0," MEMACCESS(1) " \n" + "lea " MEMLEA(0x10,1) ",%1 \n" + "sub $0x10,%2 \n" + "jg 1b \n" + : "+r"(src_ptr), // %0 + "+r"(dst_ptr), // %1 + "+r"(dst_width) // %2 + :: "memory", "cc", "xmm0", "xmm1", "xmm5" + ); +} + +void ScaleRowDown2Box_SSE2(const uint8* src_ptr, ptrdiff_t src_stride, + uint8* dst_ptr, int dst_width) { + asm volatile ( + "pcmpeqb %%xmm5,%%xmm5 \n" + "psrlw $0x8,%%xmm5 \n" + + LABELALIGN + "1: \n" + "movdqu " MEMACCESS(0) ",%%xmm0 \n" + "movdqu " MEMACCESS2(0x10,0) ",%%xmm1 \n" + MEMOPREG(movdqu,0x00,0,3,1,xmm2) // movdqu (%0,%3,1),%%xmm2 + MEMOPREG(movdqu,0x10,0,3,1,xmm3) // movdqu 0x10(%0,%3,1),%%xmm3 + "lea " MEMLEA(0x20,0) ",%0 \n" + "pavgb %%xmm2,%%xmm0 \n" + "pavgb %%xmm3,%%xmm1 \n" + "movdqa %%xmm0,%%xmm2 \n" + "psrlw $0x8,%%xmm0 \n" + "movdqa %%xmm1,%%xmm3 \n" + "psrlw $0x8,%%xmm1 \n" + "pand %%xmm5,%%xmm2 \n" + "pand %%xmm5,%%xmm3 \n" + "pavgw %%xmm2,%%xmm0 \n" + "pavgw %%xmm3,%%xmm1 \n" + "packuswb %%xmm1,%%xmm0 \n" + "movdqu %%xmm0," MEMACCESS(1) " \n" + "lea " MEMLEA(0x10,1) ",%1 \n" + "sub $0x10,%2 \n" + "jg 1b \n" + : "+r"(src_ptr), // %0 + "+r"(dst_ptr), // %1 + "+r"(dst_width) // %2 + : "r"((intptr_t)(src_stride)) // %3 + : "memory", "cc", NACL_R14 + "xmm0", "xmm1", "xmm2", "xmm3", "xmm5" + ); +} + +void ScaleRowDown4_SSE2(const uint8* src_ptr, ptrdiff_t src_stride, + uint8* dst_ptr, int dst_width) { + asm volatile ( + "pcmpeqb %%xmm5,%%xmm5 \n" + "psrld $0x18,%%xmm5 \n" + "pslld $0x10,%%xmm5 \n" + + LABELALIGN + "1: \n" + "movdqu " MEMACCESS(0) ",%%xmm0 \n" + "movdqu " MEMACCESS2(0x10,0) ",%%xmm1 \n" + "lea " MEMLEA(0x20,0) ",%0 \n" + "pand %%xmm5,%%xmm0 \n" + "pand %%xmm5,%%xmm1 \n" + "packuswb %%xmm1,%%xmm0 \n" + "psrlw $0x8,%%xmm0 \n" + "packuswb %%xmm0,%%xmm0 \n" + "movq %%xmm0," MEMACCESS(1) " \n" + "lea " MEMLEA(0x8,1) ",%1 \n" + "sub $0x8,%2 \n" + "jg 1b \n" + : "+r"(src_ptr), // %0 + "+r"(dst_ptr), // %1 + "+r"(dst_width) // %2 + :: "memory", "cc", "xmm0", "xmm1", "xmm5" + ); +} + +void ScaleRowDown4Box_SSE2(const uint8* src_ptr, ptrdiff_t src_stride, + uint8* dst_ptr, int dst_width) { + intptr_t stridex3 = 0; + asm volatile ( + "pcmpeqb %%xmm7,%%xmm7 \n" + "psrlw $0x8,%%xmm7 \n" + "lea " MEMLEA4(0x00,4,4,2) ",%3 \n" + + LABELALIGN + "1: \n" + "movdqu " MEMACCESS(0) ",%%xmm0 \n" + "movdqu " MEMACCESS2(0x10,0) ",%%xmm1 \n" + MEMOPREG(movdqu,0x00,0,4,1,xmm2) // movdqu (%0,%4,1),%%xmm2 + MEMOPREG(movdqu,0x10,0,4,1,xmm3) // movdqu 0x10(%0,%4,1),%%xmm3 + "pavgb %%xmm2,%%xmm0 \n" + "pavgb %%xmm3,%%xmm1 \n" + MEMOPREG(movdqu,0x00,0,4,2,xmm2) // movdqu (%0,%4,2),%%xmm2 + MEMOPREG(movdqu,0x10,0,4,2,xmm3) // movdqu 0x10(%0,%4,2),%%xmm3 + MEMOPREG(movdqu,0x00,0,3,1,xmm4) // movdqu (%0,%3,1),%%xmm4 + MEMOPREG(movdqu,0x10,0,3,1,xmm5) // movdqu 0x10(%0,%3,1),%%xmm5 + "lea " MEMLEA(0x20,0) ",%0 \n" + "pavgb %%xmm4,%%xmm2 \n" + "pavgb %%xmm2,%%xmm0 \n" + "pavgb %%xmm5,%%xmm3 \n" + "pavgb %%xmm3,%%xmm1 \n" + "movdqa %%xmm0,%%xmm2 \n" + "psrlw $0x8,%%xmm0 \n" + "movdqa %%xmm1,%%xmm3 \n" + "psrlw $0x8,%%xmm1 \n" + "pand %%xmm7,%%xmm2 \n" + "pand %%xmm7,%%xmm3 \n" + "pavgw %%xmm2,%%xmm0 \n" + "pavgw %%xmm3,%%xmm1 \n" + "packuswb %%xmm1,%%xmm0 \n" + "movdqa %%xmm0,%%xmm2 \n" + "psrlw $0x8,%%xmm0 \n" + "pand %%xmm7,%%xmm2 \n" + "pavgw %%xmm2,%%xmm0 \n" + "packuswb %%xmm0,%%xmm0 \n" + "movq %%xmm0," MEMACCESS(1) " \n" + "lea " MEMLEA(0x8,1) ",%1 \n" + "sub $0x8,%2 \n" + "jg 1b \n" + : "+r"(src_ptr), // %0 + "+r"(dst_ptr), // %1 + "+r"(dst_width), // %2 + "+r"(stridex3) // %3 + : "r"((intptr_t)(src_stride)) // %4 + : "memory", "cc", NACL_R14 + "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm7" + ); +} + +void ScaleRowDown34_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride, + uint8* dst_ptr, int dst_width) { + asm volatile ( + "movdqa %0,%%xmm3 \n" + "movdqa %1,%%xmm4 \n" + "movdqa %2,%%xmm5 \n" + : + : "m"(kShuf0), // %0 + "m"(kShuf1), // %1 + "m"(kShuf2) // %2 + ); + asm volatile ( + LABELALIGN + "1: \n" + "movdqu " MEMACCESS(0) ",%%xmm0 \n" + "movdqu " MEMACCESS2(0x10,0) ",%%xmm2 \n" + "lea " MEMLEA(0x20,0) ",%0 \n" + "movdqa %%xmm2,%%xmm1 \n" + "palignr $0x8,%%xmm0,%%xmm1 \n" + "pshufb %%xmm3,%%xmm0 \n" + "pshufb %%xmm4,%%xmm1 \n" + "pshufb %%xmm5,%%xmm2 \n" + "movq %%xmm0," MEMACCESS(1) " \n" + "movq %%xmm1," MEMACCESS2(0x8,1) " \n" + "movq %%xmm2," MEMACCESS2(0x10,1) " \n" + "lea " MEMLEA(0x18,1) ",%1 \n" + "sub $0x18,%2 \n" + "jg 1b \n" + : "+r"(src_ptr), // %0 + "+r"(dst_ptr), // %1 + "+r"(dst_width) // %2 + :: "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5" + ); +} + +void ScaleRowDown34_1_Box_SSSE3(const uint8* src_ptr, + ptrdiff_t src_stride, + uint8* dst_ptr, int dst_width) { + asm volatile ( + "movdqa %0,%%xmm2 \n" // kShuf01 + "movdqa %1,%%xmm3 \n" // kShuf11 + "movdqa %2,%%xmm4 \n" // kShuf21 + : + : "m"(kShuf01), // %0 + "m"(kShuf11), // %1 + "m"(kShuf21) // %2 + ); + asm volatile ( + "movdqa %0,%%xmm5 \n" // kMadd01 + "movdqa %1,%%xmm0 \n" // kMadd11 + "movdqa %2,%%xmm1 \n" // kRound34 + : + : "m"(kMadd01), // %0 + "m"(kMadd11), // %1 + "m"(kRound34) // %2 + ); + asm volatile ( + LABELALIGN + "1: \n" + "movdqu " MEMACCESS(0) ",%%xmm6 \n" + MEMOPREG(movdqu,0x00,0,3,1,xmm7) // movdqu (%0,%3),%%xmm7 + "pavgb %%xmm7,%%xmm6 \n" + "pshufb %%xmm2,%%xmm6 \n" + "pmaddubsw %%xmm5,%%xmm6 \n" + "paddsw %%xmm1,%%xmm6 \n" + "psrlw $0x2,%%xmm6 \n" + "packuswb %%xmm6,%%xmm6 \n" + "movq %%xmm6," MEMACCESS(1) " \n" + "movdqu " MEMACCESS2(0x8,0) ",%%xmm6 \n" + MEMOPREG(movdqu,0x8,0,3,1,xmm7) // movdqu 0x8(%0,%3),%%xmm7 + "pavgb %%xmm7,%%xmm6 \n" + "pshufb %%xmm3,%%xmm6 \n" + "pmaddubsw %%xmm0,%%xmm6 \n" + "paddsw %%xmm1,%%xmm6 \n" + "psrlw $0x2,%%xmm6 \n" + "packuswb %%xmm6,%%xmm6 \n" + "movq %%xmm6," MEMACCESS2(0x8,1) " \n" + "movdqu " MEMACCESS2(0x10,0) ",%%xmm6 \n" + MEMOPREG(movdqu,0x10,0,3,1,xmm7) // movdqu 0x10(%0,%3),%%xmm7 + "lea " MEMLEA(0x20,0) ",%0 \n" + "pavgb %%xmm7,%%xmm6 \n" + "pshufb %%xmm4,%%xmm6 \n" + "pmaddubsw %4,%%xmm6 \n" + "paddsw %%xmm1,%%xmm6 \n" + "psrlw $0x2,%%xmm6 \n" + "packuswb %%xmm6,%%xmm6 \n" + "movq %%xmm6," MEMACCESS2(0x10,1) " \n" + "lea " MEMLEA(0x18,1) ",%1 \n" + "sub $0x18,%2 \n" + "jg 1b \n" + : "+r"(src_ptr), // %0 + "+r"(dst_ptr), // %1 + "+r"(dst_width) // %2 + : "r"((intptr_t)(src_stride)), // %3 + "m"(kMadd21) // %4 + : "memory", "cc", NACL_R14 + "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7" + ); +} + +void ScaleRowDown34_0_Box_SSSE3(const uint8* src_ptr, + ptrdiff_t src_stride, + uint8* dst_ptr, int dst_width) { + asm volatile ( + "movdqa %0,%%xmm2 \n" // kShuf01 + "movdqa %1,%%xmm3 \n" // kShuf11 + "movdqa %2,%%xmm4 \n" // kShuf21 + : + : "m"(kShuf01), // %0 + "m"(kShuf11), // %1 + "m"(kShuf21) // %2 + ); + asm volatile ( + "movdqa %0,%%xmm5 \n" // kMadd01 + "movdqa %1,%%xmm0 \n" // kMadd11 + "movdqa %2,%%xmm1 \n" // kRound34 + : + : "m"(kMadd01), // %0 + "m"(kMadd11), // %1 + "m"(kRound34) // %2 + ); + + asm volatile ( + LABELALIGN + "1: \n" + "movdqu " MEMACCESS(0) ",%%xmm6 \n" + MEMOPREG(movdqu,0x00,0,3,1,xmm7) // movdqu (%0,%3,1),%%xmm7 + "pavgb %%xmm6,%%xmm7 \n" + "pavgb %%xmm7,%%xmm6 \n" + "pshufb %%xmm2,%%xmm6 \n" + "pmaddubsw %%xmm5,%%xmm6 \n" + "paddsw %%xmm1,%%xmm6 \n" + "psrlw $0x2,%%xmm6 \n" + "packuswb %%xmm6,%%xmm6 \n" + "movq %%xmm6," MEMACCESS(1) " \n" + "movdqu " MEMACCESS2(0x8,0) ",%%xmm6 \n" + MEMOPREG(movdqu,0x8,0,3,1,xmm7) // movdqu 0x8(%0,%3,1),%%xmm7 + "pavgb %%xmm6,%%xmm7 \n" + "pavgb %%xmm7,%%xmm6 \n" + "pshufb %%xmm3,%%xmm6 \n" + "pmaddubsw %%xmm0,%%xmm6 \n" + "paddsw %%xmm1,%%xmm6 \n" + "psrlw $0x2,%%xmm6 \n" + "packuswb %%xmm6,%%xmm6 \n" + "movq %%xmm6," MEMACCESS2(0x8,1) " \n" + "movdqu " MEMACCESS2(0x10,0) ",%%xmm6 \n" + MEMOPREG(movdqu,0x10,0,3,1,xmm7) // movdqu 0x10(%0,%3,1),%%xmm7 + "lea " MEMLEA(0x20,0) ",%0 \n" + "pavgb %%xmm6,%%xmm7 \n" + "pavgb %%xmm7,%%xmm6 \n" + "pshufb %%xmm4,%%xmm6 \n" + "pmaddubsw %4,%%xmm6 \n" + "paddsw %%xmm1,%%xmm6 \n" + "psrlw $0x2,%%xmm6 \n" + "packuswb %%xmm6,%%xmm6 \n" + "movq %%xmm6," MEMACCESS2(0x10,1) " \n" + "lea " MEMLEA(0x18,1) ",%1 \n" + "sub $0x18,%2 \n" + "jg 1b \n" + : "+r"(src_ptr), // %0 + "+r"(dst_ptr), // %1 + "+r"(dst_width) // %2 + : "r"((intptr_t)(src_stride)), // %3 + "m"(kMadd21) // %4 + : "memory", "cc", NACL_R14 + "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7" + ); +} + +void ScaleRowDown38_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride, + uint8* dst_ptr, int dst_width) { + asm volatile ( + "movdqa %3,%%xmm4 \n" + "movdqa %4,%%xmm5 \n" + + LABELALIGN + "1: \n" + "movdqu " MEMACCESS(0) ",%%xmm0 \n" + "movdqu " MEMACCESS2(0x10,0) ",%%xmm1 \n" + "lea " MEMLEA(0x20,0) ",%0 \n" + "pshufb %%xmm4,%%xmm0 \n" + "pshufb %%xmm5,%%xmm1 \n" + "paddusb %%xmm1,%%xmm0 \n" + "movq %%xmm0," MEMACCESS(1) " \n" + "movhlps %%xmm0,%%xmm1 \n" + "movd %%xmm1," MEMACCESS2(0x8,1) " \n" + "lea " MEMLEA(0xc,1) ",%1 \n" + "sub $0xc,%2 \n" + "jg 1b \n" + : "+r"(src_ptr), // %0 + "+r"(dst_ptr), // %1 + "+r"(dst_width) // %2 + : "m"(kShuf38a), // %3 + "m"(kShuf38b) // %4 + : "memory", "cc", "xmm0", "xmm1", "xmm4", "xmm5" + ); +} + +void ScaleRowDown38_2_Box_SSSE3(const uint8* src_ptr, + ptrdiff_t src_stride, + uint8* dst_ptr, int dst_width) { + asm volatile ( + "movdqa %0,%%xmm2 \n" + "movdqa %1,%%xmm3 \n" + "movdqa %2,%%xmm4 \n" + "movdqa %3,%%xmm5 \n" + : + : "m"(kShufAb0), // %0 + "m"(kShufAb1), // %1 + "m"(kShufAb2), // %2 + "m"(kScaleAb2) // %3 + ); + asm volatile ( + LABELALIGN + "1: \n" + "movdqu " MEMACCESS(0) ",%%xmm0 \n" + MEMOPREG(movdqu,0x00,0,3,1,xmm1) // movdqu (%0,%3,1),%%xmm1 + "lea " MEMLEA(0x10,0) ",%0 \n" + "pavgb %%xmm1,%%xmm0 \n" + "movdqa %%xmm0,%%xmm1 \n" + "pshufb %%xmm2,%%xmm1 \n" + "movdqa %%xmm0,%%xmm6 \n" + "pshufb %%xmm3,%%xmm6 \n" + "paddusw %%xmm6,%%xmm1 \n" + "pshufb %%xmm4,%%xmm0 \n" + "paddusw %%xmm0,%%xmm1 \n" + "pmulhuw %%xmm5,%%xmm1 \n" + "packuswb %%xmm1,%%xmm1 \n" + "movd %%xmm1," MEMACCESS(1) " \n" + "psrlq $0x10,%%xmm1 \n" + "movd %%xmm1," MEMACCESS2(0x2,1) " \n" + "lea " MEMLEA(0x6,1) ",%1 \n" + "sub $0x6,%2 \n" + "jg 1b \n" + : "+r"(src_ptr), // %0 + "+r"(dst_ptr), // %1 + "+r"(dst_width) // %2 + : "r"((intptr_t)(src_stride)) // %3 + : "memory", "cc", NACL_R14 + "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6" + ); +} + +void ScaleRowDown38_3_Box_SSSE3(const uint8* src_ptr, + ptrdiff_t src_stride, + uint8* dst_ptr, int dst_width) { + asm volatile ( + "movdqa %0,%%xmm2 \n" + "movdqa %1,%%xmm3 \n" + "movdqa %2,%%xmm4 \n" + "pxor %%xmm5,%%xmm5 \n" + : + : "m"(kShufAc), // %0 + "m"(kShufAc3), // %1 + "m"(kScaleAc33) // %2 + ); + asm volatile ( + LABELALIGN + "1: \n" + "movdqu " MEMACCESS(0) ",%%xmm0 \n" + MEMOPREG(movdqu,0x00,0,3,1,xmm6) // movdqu (%0,%3,1),%%xmm6 + "movhlps %%xmm0,%%xmm1 \n" + "movhlps %%xmm6,%%xmm7 \n" + "punpcklbw %%xmm5,%%xmm0 \n" + "punpcklbw %%xmm5,%%xmm1 \n" + "punpcklbw %%xmm5,%%xmm6 \n" + "punpcklbw %%xmm5,%%xmm7 \n" + "paddusw %%xmm6,%%xmm0 \n" + "paddusw %%xmm7,%%xmm1 \n" + MEMOPREG(movdqu,0x00,0,3,2,xmm6) // movdqu (%0,%3,2),%%xmm6 + "lea " MEMLEA(0x10,0) ",%0 \n" + "movhlps %%xmm6,%%xmm7 \n" + "punpcklbw %%xmm5,%%xmm6 \n" + "punpcklbw %%xmm5,%%xmm7 \n" + "paddusw %%xmm6,%%xmm0 \n" + "paddusw %%xmm7,%%xmm1 \n" + "movdqa %%xmm0,%%xmm6 \n" + "psrldq $0x2,%%xmm0 \n" + "paddusw %%xmm0,%%xmm6 \n" + "psrldq $0x2,%%xmm0 \n" + "paddusw %%xmm0,%%xmm6 \n" + "pshufb %%xmm2,%%xmm6 \n" + "movdqa %%xmm1,%%xmm7 \n" + "psrldq $0x2,%%xmm1 \n" + "paddusw %%xmm1,%%xmm7 \n" + "psrldq $0x2,%%xmm1 \n" + "paddusw %%xmm1,%%xmm7 \n" + "pshufb %%xmm3,%%xmm7 \n" + "paddusw %%xmm7,%%xmm6 \n" + "pmulhuw %%xmm4,%%xmm6 \n" + "packuswb %%xmm6,%%xmm6 \n" + "movd %%xmm6," MEMACCESS(1) " \n" + "psrlq $0x10,%%xmm6 \n" + "movd %%xmm6," MEMACCESS2(0x2,1) " \n" + "lea " MEMLEA(0x6,1) ",%1 \n" + "sub $0x6,%2 \n" + "jg 1b \n" + : "+r"(src_ptr), // %0 + "+r"(dst_ptr), // %1 + "+r"(dst_width) // %2 + : "r"((intptr_t)(src_stride)) // %3 + : "memory", "cc", NACL_R14 + "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7" + ); +} + +// Reads 16xN bytes and produces 16 shorts at a time. +void ScaleAddRows_SSE2(const uint8* src_ptr, ptrdiff_t src_stride, + uint16* dst_ptr, int src_width, int src_height) { + int tmp_height = 0; + intptr_t tmp_src = 0; + asm volatile ( + "mov %0,%3 \n" // row pointer + "mov %5,%2 \n" // height + "pxor %%xmm0,%%xmm0 \n" // clear accumulators + "pxor %%xmm1,%%xmm1 \n" + "pxor %%xmm4,%%xmm4 \n" + + LABELALIGN + "1: \n" + "movdqu " MEMACCESS(3) ",%%xmm2 \n" + "add %6,%3 \n" + "movdqa %%xmm2,%%xmm3 \n" + "punpcklbw %%xmm4,%%xmm2 \n" + "punpckhbw %%xmm4,%%xmm3 \n" + "paddusw %%xmm2,%%xmm0 \n" + "paddusw %%xmm3,%%xmm1 \n" + "sub $0x1,%2 \n" + "jg 1b \n" + + "movdqu %%xmm0," MEMACCESS(1) " \n" + "movdqu %%xmm1," MEMACCESS2(0x10,1) " \n" + "lea " MEMLEA(0x20,1) ",%1 \n" + "lea " MEMLEA(0x10,0) ",%0 \n" // src_ptr += 16 + "mov %0,%3 \n" // row pointer + "mov %5,%2 \n" // height + "pxor %%xmm0,%%xmm0 \n" // clear accumulators + "pxor %%xmm1,%%xmm1 \n" + "sub $0x10,%4 \n" + "jg 1b \n" + : "+r"(src_ptr), // %0 + "+r"(dst_ptr), // %1 + "+r"(tmp_height), // %2 + "+r"(tmp_src), // %3 + "+r"(src_width), // %4 + "+rm"(src_height) // %5 + : "rm"((intptr_t)(src_stride)) // %6 + : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4" + ); +} + +// Bilinear column filtering. SSSE3 version. +void ScaleFilterCols_SSSE3(uint8* dst_ptr, const uint8* src_ptr, + int dst_width, int x, int dx) { + intptr_t x0 = 0, x1 = 0, temp_pixel = 0; + asm volatile ( + "movd %6,%%xmm2 \n" + "movd %7,%%xmm3 \n" + "movl $0x04040000,%k2 \n" + "movd %k2,%%xmm5 \n" + "pcmpeqb %%xmm6,%%xmm6 \n" + "psrlw $0x9,%%xmm6 \n" + "pextrw $0x1,%%xmm2,%k3 \n" + "subl $0x2,%5 \n" + "jl 29f \n" + "movdqa %%xmm2,%%xmm0 \n" + "paddd %%xmm3,%%xmm0 \n" + "punpckldq %%xmm0,%%xmm2 \n" + "punpckldq %%xmm3,%%xmm3 \n" + "paddd %%xmm3,%%xmm3 \n" + "pextrw $0x3,%%xmm2,%k4 \n" + + LABELALIGN + "2: \n" + "movdqa %%xmm2,%%xmm1 \n" + "paddd %%xmm3,%%xmm2 \n" + MEMOPARG(movzwl,0x00,1,3,1,k2) // movzwl (%1,%3,1),%k2 + "movd %k2,%%xmm0 \n" + "psrlw $0x9,%%xmm1 \n" + MEMOPARG(movzwl,0x00,1,4,1,k2) // movzwl (%1,%4,1),%k2 + "movd %k2,%%xmm4 \n" + "pshufb %%xmm5,%%xmm1 \n" + "punpcklwd %%xmm4,%%xmm0 \n" + "pxor %%xmm6,%%xmm1 \n" + "pmaddubsw %%xmm1,%%xmm0 \n" + "pextrw $0x1,%%xmm2,%k3 \n" + "pextrw $0x3,%%xmm2,%k4 \n" + "psrlw $0x7,%%xmm0 \n" + "packuswb %%xmm0,%%xmm0 \n" + "movd %%xmm0,%k2 \n" + "mov %w2," MEMACCESS(0) " \n" + "lea " MEMLEA(0x2,0) ",%0 \n" + "sub $0x2,%5 \n" + "jge 2b \n" + + LABELALIGN + "29: \n" + "addl $0x1,%5 \n" + "jl 99f \n" + MEMOPARG(movzwl,0x00,1,3,1,k2) // movzwl (%1,%3,1),%k2 + "movd %k2,%%xmm0 \n" + "psrlw $0x9,%%xmm2 \n" + "pshufb %%xmm5,%%xmm2 \n" + "pxor %%xmm6,%%xmm2 \n" + "pmaddubsw %%xmm2,%%xmm0 \n" + "psrlw $0x7,%%xmm0 \n" + "packuswb %%xmm0,%%xmm0 \n" + "movd %%xmm0,%k2 \n" + "mov %b2," MEMACCESS(0) " \n" + "99: \n" + : "+r"(dst_ptr), // %0 + "+r"(src_ptr), // %1 + "+a"(temp_pixel), // %2 + "+r"(x0), // %3 + "+r"(x1), // %4 + "+rm"(dst_width) // %5 + : "rm"(x), // %6 + "rm"(dx) // %7 + : "memory", "cc", NACL_R14 + "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6" + ); +} + +// Reads 4 pixels, duplicates them and writes 8 pixels. +// Alignment requirement: src_argb 16 byte aligned, dst_argb 16 byte aligned. +void ScaleColsUp2_SSE2(uint8* dst_ptr, const uint8* src_ptr, + int dst_width, int x, int dx) { + asm volatile ( + LABELALIGN + "1: \n" + "movdqu " MEMACCESS(1) ",%%xmm0 \n" + "lea " MEMLEA(0x10,1) ",%1 \n" + "movdqa %%xmm0,%%xmm1 \n" + "punpcklbw %%xmm0,%%xmm0 \n" + "punpckhbw %%xmm1,%%xmm1 \n" + "movdqu %%xmm0," MEMACCESS(0) " \n" + "movdqu %%xmm1," MEMACCESS2(0x10,0) " \n" + "lea " MEMLEA(0x20,0) ",%0 \n" + "sub $0x20,%2 \n" + "jg 1b \n" + + : "+r"(dst_ptr), // %0 + "+r"(src_ptr), // %1 + "+r"(dst_width) // %2 + :: "memory", "cc", "xmm0", "xmm1" + ); +} + +void ScaleARGBRowDown2_SSE2(const uint8* src_argb, + ptrdiff_t src_stride, + uint8* dst_argb, int dst_width) { + asm volatile ( + LABELALIGN + "1: \n" + "movdqu " MEMACCESS(0) ",%%xmm0 \n" + "movdqu " MEMACCESS2(0x10,0) ",%%xmm1 \n" + "lea " MEMLEA(0x20,0) ",%0 \n" + "shufps $0xdd,%%xmm1,%%xmm0 \n" + "movdqu %%xmm0," MEMACCESS(1) " \n" + "lea " MEMLEA(0x10,1) ",%1 \n" + "sub $0x4,%2 \n" + "jg 1b \n" + : "+r"(src_argb), // %0 + "+r"(dst_argb), // %1 + "+r"(dst_width) // %2 + :: "memory", "cc", "xmm0", "xmm1" + ); +} + +void ScaleARGBRowDown2Linear_SSE2(const uint8* src_argb, + ptrdiff_t src_stride, + uint8* dst_argb, int dst_width) { + asm volatile ( + LABELALIGN + "1: \n" + "movdqu " MEMACCESS(0) ",%%xmm0 \n" + "movdqu " MEMACCESS2(0x10,0) ",%%xmm1 \n" + "lea " MEMLEA(0x20,0) ",%0 \n" + "movdqa %%xmm0,%%xmm2 \n" + "shufps $0x88,%%xmm1,%%xmm0 \n" + "shufps $0xdd,%%xmm1,%%xmm2 \n" + "pavgb %%xmm2,%%xmm0 \n" + "movdqu %%xmm0," MEMACCESS(1) " \n" + "lea " MEMLEA(0x10,1) ",%1 \n" + "sub $0x4,%2 \n" + "jg 1b \n" + : "+r"(src_argb), // %0 + "+r"(dst_argb), // %1 + "+r"(dst_width) // %2 + :: "memory", "cc", "xmm0", "xmm1" + ); +} + +void ScaleARGBRowDown2Box_SSE2(const uint8* src_argb, + ptrdiff_t src_stride, + uint8* dst_argb, int dst_width) { + asm volatile ( + LABELALIGN + "1: \n" + "movdqu " MEMACCESS(0) ",%%xmm0 \n" + "movdqu " MEMACCESS2(0x10,0) ",%%xmm1 \n" + MEMOPREG(movdqu,0x00,0,3,1,xmm2) // movdqu (%0,%3,1),%%xmm2 + MEMOPREG(movdqu,0x10,0,3,1,xmm3) // movdqu 0x10(%0,%3,1),%%xmm3 + "lea " MEMLEA(0x20,0) ",%0 \n" + "pavgb %%xmm2,%%xmm0 \n" + "pavgb %%xmm3,%%xmm1 \n" + "movdqa %%xmm0,%%xmm2 \n" + "shufps $0x88,%%xmm1,%%xmm0 \n" + "shufps $0xdd,%%xmm1,%%xmm2 \n" + "pavgb %%xmm2,%%xmm0 \n" + "movdqu %%xmm0," MEMACCESS(1) " \n" + "lea " MEMLEA(0x10,1) ",%1 \n" + "sub $0x4,%2 \n" + "jg 1b \n" + : "+r"(src_argb), // %0 + "+r"(dst_argb), // %1 + "+r"(dst_width) // %2 + : "r"((intptr_t)(src_stride)) // %3 + : "memory", "cc", NACL_R14 + "xmm0", "xmm1", "xmm2", "xmm3" + ); +} + +// Reads 4 pixels at a time. +// Alignment requirement: dst_argb 16 byte aligned. +void ScaleARGBRowDownEven_SSE2(const uint8* src_argb, ptrdiff_t src_stride, + int src_stepx, uint8* dst_argb, int dst_width) { + intptr_t src_stepx_x4 = (intptr_t)(src_stepx); + intptr_t src_stepx_x12 = 0; + asm volatile ( + "lea " MEMLEA3(0x00,1,4) ",%1 \n" + "lea " MEMLEA4(0x00,1,1,2) ",%4 \n" + LABELALIGN + "1: \n" + "movd " MEMACCESS(0) ",%%xmm0 \n" + MEMOPREG(movd,0x00,0,1,1,xmm1) // movd (%0,%1,1),%%xmm1 + "punpckldq %%xmm1,%%xmm0 \n" + MEMOPREG(movd,0x00,0,1,2,xmm2) // movd (%0,%1,2),%%xmm2 + MEMOPREG(movd,0x00,0,4,1,xmm3) // movd (%0,%4,1),%%xmm3 + "lea " MEMLEA4(0x00,0,1,4) ",%0 \n" + "punpckldq %%xmm3,%%xmm2 \n" + "punpcklqdq %%xmm2,%%xmm0 \n" + "movdqu %%xmm0," MEMACCESS(2) " \n" + "lea " MEMLEA(0x10,2) ",%2 \n" + "sub $0x4,%3 \n" + "jg 1b \n" + : "+r"(src_argb), // %0 + "+r"(src_stepx_x4), // %1 + "+r"(dst_argb), // %2 + "+r"(dst_width), // %3 + "+r"(src_stepx_x12) // %4 + :: "memory", "cc", NACL_R14 + "xmm0", "xmm1", "xmm2", "xmm3" + ); +} + +// Blends four 2x2 to 4x1. +// Alignment requirement: dst_argb 16 byte aligned. +void ScaleARGBRowDownEvenBox_SSE2(const uint8* src_argb, + ptrdiff_t src_stride, int src_stepx, + uint8* dst_argb, int dst_width) { + intptr_t src_stepx_x4 = (intptr_t)(src_stepx); + intptr_t src_stepx_x12 = 0; + intptr_t row1 = (intptr_t)(src_stride); + asm volatile ( + "lea " MEMLEA3(0x00,1,4) ",%1 \n" + "lea " MEMLEA4(0x00,1,1,2) ",%4 \n" + "lea " MEMLEA4(0x00,0,5,1) ",%5 \n" + + LABELALIGN + "1: \n" + "movq " MEMACCESS(0) ",%%xmm0 \n" + MEMOPREG(movhps,0x00,0,1,1,xmm0) // movhps (%0,%1,1),%%xmm0 + MEMOPREG(movq,0x00,0,1,2,xmm1) // movq (%0,%1,2),%%xmm1 + MEMOPREG(movhps,0x00,0,4,1,xmm1) // movhps (%0,%4,1),%%xmm1 + "lea " MEMLEA4(0x00,0,1,4) ",%0 \n" + "movq " MEMACCESS(5) ",%%xmm2 \n" + MEMOPREG(movhps,0x00,5,1,1,xmm2) // movhps (%5,%1,1),%%xmm2 + MEMOPREG(movq,0x00,5,1,2,xmm3) // movq (%5,%1,2),%%xmm3 + MEMOPREG(movhps,0x00,5,4,1,xmm3) // movhps (%5,%4,1),%%xmm3 + "lea " MEMLEA4(0x00,5,1,4) ",%5 \n" + "pavgb %%xmm2,%%xmm0 \n" + "pavgb %%xmm3,%%xmm1 \n" + "movdqa %%xmm0,%%xmm2 \n" + "shufps $0x88,%%xmm1,%%xmm0 \n" + "shufps $0xdd,%%xmm1,%%xmm2 \n" + "pavgb %%xmm2,%%xmm0 \n" + "movdqu %%xmm0," MEMACCESS(2) " \n" + "lea " MEMLEA(0x10,2) ",%2 \n" + "sub $0x4,%3 \n" + "jg 1b \n" + : "+r"(src_argb), // %0 + "+r"(src_stepx_x4), // %1 + "+r"(dst_argb), // %2 + "+rm"(dst_width), // %3 + "+r"(src_stepx_x12), // %4 + "+r"(row1) // %5 + :: "memory", "cc", NACL_R14 + "xmm0", "xmm1", "xmm2", "xmm3" + ); +} + +void ScaleARGBCols_SSE2(uint8* dst_argb, const uint8* src_argb, + int dst_width, int x, int dx) { + intptr_t x0 = 0, x1 = 0; + asm volatile ( + "movd %5,%%xmm2 \n" + "movd %6,%%xmm3 \n" + "pshufd $0x0,%%xmm2,%%xmm2 \n" + "pshufd $0x11,%%xmm3,%%xmm0 \n" + "paddd %%xmm0,%%xmm2 \n" + "paddd %%xmm3,%%xmm3 \n" + "pshufd $0x5,%%xmm3,%%xmm0 \n" + "paddd %%xmm0,%%xmm2 \n" + "paddd %%xmm3,%%xmm3 \n" + "pshufd $0x0,%%xmm3,%%xmm3 \n" + "pextrw $0x1,%%xmm2,%k0 \n" + "pextrw $0x3,%%xmm2,%k1 \n" + "cmp $0x0,%4 \n" + "jl 99f \n" + "sub $0x4,%4 \n" + "jl 49f \n" + + LABELALIGN + "40: \n" + MEMOPREG(movd,0x00,3,0,4,xmm0) // movd (%3,%0,4),%%xmm0 + MEMOPREG(movd,0x00,3,1,4,xmm1) // movd (%3,%1,4),%%xmm1 + "pextrw $0x5,%%xmm2,%k0 \n" + "pextrw $0x7,%%xmm2,%k1 \n" + "paddd %%xmm3,%%xmm2 \n" + "punpckldq %%xmm1,%%xmm0 \n" + MEMOPREG(movd,0x00,3,0,4,xmm1) // movd (%3,%0,4),%%xmm1 + MEMOPREG(movd,0x00,3,1,4,xmm4) // movd (%3,%1,4),%%xmm4 + "pextrw $0x1,%%xmm2,%k0 \n" + "pextrw $0x3,%%xmm2,%k1 \n" + "punpckldq %%xmm4,%%xmm1 \n" + "punpcklqdq %%xmm1,%%xmm0 \n" + "movdqu %%xmm0," MEMACCESS(2) " \n" + "lea " MEMLEA(0x10,2) ",%2 \n" + "sub $0x4,%4 \n" + "jge 40b \n" + + "49: \n" + "test $0x2,%4 \n" + "je 29f \n" + MEMOPREG(movd,0x00,3,0,4,xmm0) // movd (%3,%0,4),%%xmm0 + MEMOPREG(movd,0x00,3,1,4,xmm1) // movd (%3,%1,4),%%xmm1 + "pextrw $0x5,%%xmm2,%k0 \n" + "punpckldq %%xmm1,%%xmm0 \n" + "movq %%xmm0," MEMACCESS(2) " \n" + "lea " MEMLEA(0x8,2) ",%2 \n" + "29: \n" + "test $0x1,%4 \n" + "je 99f \n" + MEMOPREG(movd,0x00,3,0,4,xmm0) // movd (%3,%0,4),%%xmm0 + "movd %%xmm0," MEMACCESS(2) " \n" + "99: \n" + : "+a"(x0), // %0 + "+d"(x1), // %1 + "+r"(dst_argb), // %2 + "+r"(src_argb), // %3 + "+r"(dst_width) // %4 + : "rm"(x), // %5 + "rm"(dx) // %6 + : "memory", "cc", NACL_R14 + "xmm0", "xmm1", "xmm2", "xmm3", "xmm4" + ); +} + +// Reads 4 pixels, duplicates them and writes 8 pixels. +// Alignment requirement: src_argb 16 byte aligned, dst_argb 16 byte aligned. +void ScaleARGBColsUp2_SSE2(uint8* dst_argb, const uint8* src_argb, + int dst_width, int x, int dx) { + asm volatile ( + LABELALIGN + "1: \n" + "movdqu " MEMACCESS(1) ",%%xmm0 \n" + "lea " MEMLEA(0x10,1) ",%1 \n" + "movdqa %%xmm0,%%xmm1 \n" + "punpckldq %%xmm0,%%xmm0 \n" + "punpckhdq %%xmm1,%%xmm1 \n" + "movdqu %%xmm0," MEMACCESS(0) " \n" + "movdqu %%xmm1," MEMACCESS2(0x10,0) " \n" + "lea " MEMLEA(0x20,0) ",%0 \n" + "sub $0x8,%2 \n" + "jg 1b \n" + + : "+r"(dst_argb), // %0 + "+r"(src_argb), // %1 + "+r"(dst_width) // %2 + :: "memory", "cc", NACL_R14 + "xmm0", "xmm1" + ); +} + +// Shuffle table for arranging 2 pixels into pairs for pmaddubsw +static uvec8 kShuffleColARGB = { + 0u, 4u, 1u, 5u, 2u, 6u, 3u, 7u, // bbggrraa 1st pixel + 8u, 12u, 9u, 13u, 10u, 14u, 11u, 15u // bbggrraa 2nd pixel +}; + +// Shuffle table for duplicating 2 fractions into 8 bytes each +static uvec8 kShuffleFractions = { + 0u, 0u, 0u, 0u, 0u, 0u, 0u, 0u, 4u, 4u, 4u, 4u, 4u, 4u, 4u, 4u, +}; + +// Bilinear row filtering combines 4x2 -> 4x1. SSSE3 version +void ScaleARGBFilterCols_SSSE3(uint8* dst_argb, const uint8* src_argb, + int dst_width, int x, int dx) { + intptr_t x0 = 0, x1 = 0; + asm volatile ( + "movdqa %0,%%xmm4 \n" + "movdqa %1,%%xmm5 \n" + : + : "m"(kShuffleColARGB), // %0 + "m"(kShuffleFractions) // %1 + ); + + asm volatile ( + "movd %5,%%xmm2 \n" + "movd %6,%%xmm3 \n" + "pcmpeqb %%xmm6,%%xmm6 \n" + "psrlw $0x9,%%xmm6 \n" + "pextrw $0x1,%%xmm2,%k3 \n" + "sub $0x2,%2 \n" + "jl 29f \n" + "movdqa %%xmm2,%%xmm0 \n" + "paddd %%xmm3,%%xmm0 \n" + "punpckldq %%xmm0,%%xmm2 \n" + "punpckldq %%xmm3,%%xmm3 \n" + "paddd %%xmm3,%%xmm3 \n" + "pextrw $0x3,%%xmm2,%k4 \n" + + LABELALIGN + "2: \n" + "movdqa %%xmm2,%%xmm1 \n" + "paddd %%xmm3,%%xmm2 \n" + MEMOPREG(movq,0x00,1,3,4,xmm0) // movq (%1,%3,4),%%xmm0 + "psrlw $0x9,%%xmm1 \n" + MEMOPREG(movhps,0x00,1,4,4,xmm0) // movhps (%1,%4,4),%%xmm0 + "pshufb %%xmm5,%%xmm1 \n" + "pshufb %%xmm4,%%xmm0 \n" + "pxor %%xmm6,%%xmm1 \n" + "pmaddubsw %%xmm1,%%xmm0 \n" + "psrlw $0x7,%%xmm0 \n" + "pextrw $0x1,%%xmm2,%k3 \n" + "pextrw $0x3,%%xmm2,%k4 \n" + "packuswb %%xmm0,%%xmm0 \n" + "movq %%xmm0," MEMACCESS(0) " \n" + "lea " MEMLEA(0x8,0) ",%0 \n" + "sub $0x2,%2 \n" + "jge 2b \n" + + LABELALIGN + "29: \n" + "add $0x1,%2 \n" + "jl 99f \n" + "psrlw $0x9,%%xmm2 \n" + MEMOPREG(movq,0x00,1,3,4,xmm0) // movq (%1,%3,4),%%xmm0 + "pshufb %%xmm5,%%xmm2 \n" + "pshufb %%xmm4,%%xmm0 \n" + "pxor %%xmm6,%%xmm2 \n" + "pmaddubsw %%xmm2,%%xmm0 \n" + "psrlw $0x7,%%xmm0 \n" + "packuswb %%xmm0,%%xmm0 \n" + "movd %%xmm0," MEMACCESS(0) " \n" + + LABELALIGN + "99: \n" + : "+r"(dst_argb), // %0 + "+r"(src_argb), // %1 + "+rm"(dst_width), // %2 + "+r"(x0), // %3 + "+r"(x1) // %4 + : "rm"(x), // %5 + "rm"(dx) // %6 + : "memory", "cc", NACL_R14 + "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6" + ); +} + +// Divide num by div and return as 16.16 fixed point result. +int FixedDiv_X86(int num, int div) { + asm volatile ( + "cdq \n" + "shld $0x10,%%eax,%%edx \n" + "shl $0x10,%%eax \n" + "idiv %1 \n" + "mov %0, %%eax \n" + : "+a"(num) // %0 + : "c"(div) // %1 + : "memory", "cc", "edx" + ); + return num; +} + +// Divide num - 1 by div - 1 and return as 16.16 fixed point result. +int FixedDiv1_X86(int num, int div) { + asm volatile ( + "cdq \n" + "shld $0x10,%%eax,%%edx \n" + "shl $0x10,%%eax \n" + "sub $0x10001,%%eax \n" + "sbb $0x0,%%edx \n" + "sub $0x1,%1 \n" + "idiv %1 \n" + "mov %0, %%eax \n" + : "+a"(num) // %0 + : "c"(div) // %1 + : "memory", "cc", "edx" + ); + return num; +} + +#endif // defined(__x86_64__) || defined(__i386__) + +#ifdef __cplusplus +} // extern "C" +} // namespace libyuv +#endif diff --git a/media/libaom/src/third_party/libyuv/source/scale_mips.cc b/media/libaom/src/third_party/libyuv/source/scale_mips.cc new file mode 100644 index 000000000..3eb4f27c4 --- /dev/null +++ b/media/libaom/src/third_party/libyuv/source/scale_mips.cc @@ -0,0 +1,654 @@ +/* + * Copyright 2012 The LibYuv Project Authors. All rights reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#include "libyuv/basic_types.h" +#include "libyuv/row.h" + +#ifdef __cplusplus +namespace libyuv { +extern "C" { +#endif + +// This module is for GCC MIPS DSPR2 +#if !defined(LIBYUV_DISABLE_MIPS) && \ + defined(__mips_dsp) && (__mips_dsp_rev >= 2) && \ + (_MIPS_SIM == _MIPS_SIM_ABI32) + +void ScaleRowDown2_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride, + uint8* dst, int dst_width) { + __asm__ __volatile__( + ".set push \n" + ".set noreorder \n" + + "srl $t9, %[dst_width], 4 \n" // iterations -> by 16 + "beqz $t9, 2f \n" + " nop \n" + + ".p2align 2 \n" + "1: \n" + "lw $t0, 0(%[src_ptr]) \n" // |3|2|1|0| + "lw $t1, 4(%[src_ptr]) \n" // |7|6|5|4| + "lw $t2, 8(%[src_ptr]) \n" // |11|10|9|8| + "lw $t3, 12(%[src_ptr]) \n" // |15|14|13|12| + "lw $t4, 16(%[src_ptr]) \n" // |19|18|17|16| + "lw $t5, 20(%[src_ptr]) \n" // |23|22|21|20| + "lw $t6, 24(%[src_ptr]) \n" // |27|26|25|24| + "lw $t7, 28(%[src_ptr]) \n" // |31|30|29|28| + // TODO(fbarchard): Use odd pixels instead of even. + "precr.qb.ph $t8, $t1, $t0 \n" // |6|4|2|0| + "precr.qb.ph $t0, $t3, $t2 \n" // |14|12|10|8| + "precr.qb.ph $t1, $t5, $t4 \n" // |22|20|18|16| + "precr.qb.ph $t2, $t7, $t6 \n" // |30|28|26|24| + "addiu %[src_ptr], %[src_ptr], 32 \n" + "addiu $t9, $t9, -1 \n" + "sw $t8, 0(%[dst]) \n" + "sw $t0, 4(%[dst]) \n" + "sw $t1, 8(%[dst]) \n" + "sw $t2, 12(%[dst]) \n" + "bgtz $t9, 1b \n" + " addiu %[dst], %[dst], 16 \n" + + "2: \n" + "andi $t9, %[dst_width], 0xf \n" // residue + "beqz $t9, 3f \n" + " nop \n" + + "21: \n" + "lbu $t0, 0(%[src_ptr]) \n" + "addiu %[src_ptr], %[src_ptr], 2 \n" + "addiu $t9, $t9, -1 \n" + "sb $t0, 0(%[dst]) \n" + "bgtz $t9, 21b \n" + " addiu %[dst], %[dst], 1 \n" + + "3: \n" + ".set pop \n" + : [src_ptr] "+r" (src_ptr), + [dst] "+r" (dst) + : [dst_width] "r" (dst_width) + : "t0", "t1", "t2", "t3", "t4", "t5", + "t6", "t7", "t8", "t9" + ); +} + +void ScaleRowDown2Box_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride, + uint8* dst, int dst_width) { + const uint8* t = src_ptr + src_stride; + + __asm__ __volatile__ ( + ".set push \n" + ".set noreorder \n" + + "srl $t9, %[dst_width], 3 \n" // iterations -> step 8 + "bltz $t9, 2f \n" + " nop \n" + + ".p2align 2 \n" + "1: \n" + "lw $t0, 0(%[src_ptr]) \n" // |3|2|1|0| + "lw $t1, 4(%[src_ptr]) \n" // |7|6|5|4| + "lw $t2, 8(%[src_ptr]) \n" // |11|10|9|8| + "lw $t3, 12(%[src_ptr]) \n" // |15|14|13|12| + "lw $t4, 0(%[t]) \n" // |19|18|17|16| + "lw $t5, 4(%[t]) \n" // |23|22|21|20| + "lw $t6, 8(%[t]) \n" // |27|26|25|24| + "lw $t7, 12(%[t]) \n" // |31|30|29|28| + "addiu $t9, $t9, -1 \n" + "srl $t8, $t0, 16 \n" // |X|X|3|2| + "ins $t0, $t4, 16, 16 \n" // |17|16|1|0| + "ins $t4, $t8, 0, 16 \n" // |19|18|3|2| + "raddu.w.qb $t0, $t0 \n" // |17+16+1+0| + "raddu.w.qb $t4, $t4 \n" // |19+18+3+2| + "shra_r.w $t0, $t0, 2 \n" // |t0+2|>>2 + "shra_r.w $t4, $t4, 2 \n" // |t4+2|>>2 + "srl $t8, $t1, 16 \n" // |X|X|7|6| + "ins $t1, $t5, 16, 16 \n" // |21|20|5|4| + "ins $t5, $t8, 0, 16 \n" // |22|23|7|6| + "raddu.w.qb $t1, $t1 \n" // |21+20+5+4| + "raddu.w.qb $t5, $t5 \n" // |23+22+7+6| + "shra_r.w $t1, $t1, 2 \n" // |t1+2|>>2 + "shra_r.w $t5, $t5, 2 \n" // |t5+2|>>2 + "srl $t8, $t2, 16 \n" // |X|X|11|10| + "ins $t2, $t6, 16, 16 \n" // |25|24|9|8| + "ins $t6, $t8, 0, 16 \n" // |27|26|11|10| + "raddu.w.qb $t2, $t2 \n" // |25+24+9+8| + "raddu.w.qb $t6, $t6 \n" // |27+26+11+10| + "shra_r.w $t2, $t2, 2 \n" // |t2+2|>>2 + "shra_r.w $t6, $t6, 2 \n" // |t5+2|>>2 + "srl $t8, $t3, 16 \n" // |X|X|15|14| + "ins $t3, $t7, 16, 16 \n" // |29|28|13|12| + "ins $t7, $t8, 0, 16 \n" // |31|30|15|14| + "raddu.w.qb $t3, $t3 \n" // |29+28+13+12| + "raddu.w.qb $t7, $t7 \n" // |31+30+15+14| + "shra_r.w $t3, $t3, 2 \n" // |t3+2|>>2 + "shra_r.w $t7, $t7, 2 \n" // |t7+2|>>2 + "addiu %[src_ptr], %[src_ptr], 16 \n" + "addiu %[t], %[t], 16 \n" + "sb $t0, 0(%[dst]) \n" + "sb $t4, 1(%[dst]) \n" + "sb $t1, 2(%[dst]) \n" + "sb $t5, 3(%[dst]) \n" + "sb $t2, 4(%[dst]) \n" + "sb $t6, 5(%[dst]) \n" + "sb $t3, 6(%[dst]) \n" + "sb $t7, 7(%[dst]) \n" + "bgtz $t9, 1b \n" + " addiu %[dst], %[dst], 8 \n" + + "2: \n" + "andi $t9, %[dst_width], 0x7 \n" // x = residue + "beqz $t9, 3f \n" + " nop \n" + + "21: \n" + "lwr $t1, 0(%[src_ptr]) \n" + "lwl $t1, 3(%[src_ptr]) \n" + "lwr $t2, 0(%[t]) \n" + "lwl $t2, 3(%[t]) \n" + "srl $t8, $t1, 16 \n" + "ins $t1, $t2, 16, 16 \n" + "ins $t2, $t8, 0, 16 \n" + "raddu.w.qb $t1, $t1 \n" + "raddu.w.qb $t2, $t2 \n" + "shra_r.w $t1, $t1, 2 \n" + "shra_r.w $t2, $t2, 2 \n" + "sb $t1, 0(%[dst]) \n" + "sb $t2, 1(%[dst]) \n" + "addiu %[src_ptr], %[src_ptr], 4 \n" + "addiu $t9, $t9, -2 \n" + "addiu %[t], %[t], 4 \n" + "bgtz $t9, 21b \n" + " addiu %[dst], %[dst], 2 \n" + + "3: \n" + ".set pop \n" + + : [src_ptr] "+r" (src_ptr), + [dst] "+r" (dst), [t] "+r" (t) + : [dst_width] "r" (dst_width) + : "t0", "t1", "t2", "t3", "t4", "t5", + "t6", "t7", "t8", "t9" + ); +} + +void ScaleRowDown4_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride, + uint8* dst, int dst_width) { + __asm__ __volatile__ ( + ".set push \n" + ".set noreorder \n" + + "srl $t9, %[dst_width], 3 \n" + "beqz $t9, 2f \n" + " nop \n" + + ".p2align 2 \n" + "1: \n" + "lw $t1, 0(%[src_ptr]) \n" // |3|2|1|0| + "lw $t2, 4(%[src_ptr]) \n" // |7|6|5|4| + "lw $t3, 8(%[src_ptr]) \n" // |11|10|9|8| + "lw $t4, 12(%[src_ptr]) \n" // |15|14|13|12| + "lw $t5, 16(%[src_ptr]) \n" // |19|18|17|16| + "lw $t6, 20(%[src_ptr]) \n" // |23|22|21|20| + "lw $t7, 24(%[src_ptr]) \n" // |27|26|25|24| + "lw $t8, 28(%[src_ptr]) \n" // |31|30|29|28| + "precr.qb.ph $t1, $t2, $t1 \n" // |6|4|2|0| + "precr.qb.ph $t2, $t4, $t3 \n" // |14|12|10|8| + "precr.qb.ph $t5, $t6, $t5 \n" // |22|20|18|16| + "precr.qb.ph $t6, $t8, $t7 \n" // |30|28|26|24| + "precr.qb.ph $t1, $t2, $t1 \n" // |12|8|4|0| + "precr.qb.ph $t5, $t6, $t5 \n" // |28|24|20|16| + "addiu %[src_ptr], %[src_ptr], 32 \n" + "addiu $t9, $t9, -1 \n" + "sw $t1, 0(%[dst]) \n" + "sw $t5, 4(%[dst]) \n" + "bgtz $t9, 1b \n" + " addiu %[dst], %[dst], 8 \n" + + "2: \n" + "andi $t9, %[dst_width], 7 \n" // residue + "beqz $t9, 3f \n" + " nop \n" + + "21: \n" + "lbu $t1, 0(%[src_ptr]) \n" + "addiu %[src_ptr], %[src_ptr], 4 \n" + "addiu $t9, $t9, -1 \n" + "sb $t1, 0(%[dst]) \n" + "bgtz $t9, 21b \n" + " addiu %[dst], %[dst], 1 \n" + + "3: \n" + ".set pop \n" + : [src_ptr] "+r" (src_ptr), + [dst] "+r" (dst) + : [dst_width] "r" (dst_width) + : "t1", "t2", "t3", "t4", "t5", + "t6", "t7", "t8", "t9" + ); +} + +void ScaleRowDown4Box_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride, + uint8* dst, int dst_width) { + intptr_t stride = src_stride; + const uint8* s1 = src_ptr + stride; + const uint8* s2 = s1 + stride; + const uint8* s3 = s2 + stride; + + __asm__ __volatile__ ( + ".set push \n" + ".set noreorder \n" + + "srl $t9, %[dst_width], 1 \n" + "andi $t8, %[dst_width], 1 \n" + + ".p2align 2 \n" + "1: \n" + "lw $t0, 0(%[src_ptr]) \n" // |3|2|1|0| + "lw $t1, 0(%[s1]) \n" // |7|6|5|4| + "lw $t2, 0(%[s2]) \n" // |11|10|9|8| + "lw $t3, 0(%[s3]) \n" // |15|14|13|12| + "lw $t4, 4(%[src_ptr]) \n" // |19|18|17|16| + "lw $t5, 4(%[s1]) \n" // |23|22|21|20| + "lw $t6, 4(%[s2]) \n" // |27|26|25|24| + "lw $t7, 4(%[s3]) \n" // |31|30|29|28| + "raddu.w.qb $t0, $t0 \n" // |3 + 2 + 1 + 0| + "raddu.w.qb $t1, $t1 \n" // |7 + 6 + 5 + 4| + "raddu.w.qb $t2, $t2 \n" // |11 + 10 + 9 + 8| + "raddu.w.qb $t3, $t3 \n" // |15 + 14 + 13 + 12| + "raddu.w.qb $t4, $t4 \n" // |19 + 18 + 17 + 16| + "raddu.w.qb $t5, $t5 \n" // |23 + 22 + 21 + 20| + "raddu.w.qb $t6, $t6 \n" // |27 + 26 + 25 + 24| + "raddu.w.qb $t7, $t7 \n" // |31 + 30 + 29 + 28| + "add $t0, $t0, $t1 \n" + "add $t1, $t2, $t3 \n" + "add $t0, $t0, $t1 \n" + "add $t4, $t4, $t5 \n" + "add $t6, $t6, $t7 \n" + "add $t4, $t4, $t6 \n" + "shra_r.w $t0, $t0, 4 \n" + "shra_r.w $t4, $t4, 4 \n" + "sb $t0, 0(%[dst]) \n" + "sb $t4, 1(%[dst]) \n" + "addiu %[src_ptr], %[src_ptr], 8 \n" + "addiu %[s1], %[s1], 8 \n" + "addiu %[s2], %[s2], 8 \n" + "addiu %[s3], %[s3], 8 \n" + "addiu $t9, $t9, -1 \n" + "bgtz $t9, 1b \n" + " addiu %[dst], %[dst], 2 \n" + "beqz $t8, 2f \n" + " nop \n" + + "lw $t0, 0(%[src_ptr]) \n" // |3|2|1|0| + "lw $t1, 0(%[s1]) \n" // |7|6|5|4| + "lw $t2, 0(%[s2]) \n" // |11|10|9|8| + "lw $t3, 0(%[s3]) \n" // |15|14|13|12| + "raddu.w.qb $t0, $t0 \n" // |3 + 2 + 1 + 0| + "raddu.w.qb $t1, $t1 \n" // |7 + 6 + 5 + 4| + "raddu.w.qb $t2, $t2 \n" // |11 + 10 + 9 + 8| + "raddu.w.qb $t3, $t3 \n" // |15 + 14 + 13 + 12| + "add $t0, $t0, $t1 \n" + "add $t1, $t2, $t3 \n" + "add $t0, $t0, $t1 \n" + "shra_r.w $t0, $t0, 4 \n" + "sb $t0, 0(%[dst]) \n" + + "2: \n" + ".set pop \n" + + : [src_ptr] "+r" (src_ptr), + [dst] "+r" (dst), + [s1] "+r" (s1), + [s2] "+r" (s2), + [s3] "+r" (s3) + : [dst_width] "r" (dst_width) + : "t0", "t1", "t2", "t3", "t4", "t5", + "t6","t7", "t8", "t9" + ); +} + +void ScaleRowDown34_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride, + uint8* dst, int dst_width) { + __asm__ __volatile__ ( + ".set push \n" + ".set noreorder \n" + ".p2align 2 \n" + "1: \n" + "lw $t1, 0(%[src_ptr]) \n" // |3|2|1|0| + "lw $t2, 4(%[src_ptr]) \n" // |7|6|5|4| + "lw $t3, 8(%[src_ptr]) \n" // |11|10|9|8| + "lw $t4, 12(%[src_ptr]) \n" // |15|14|13|12| + "lw $t5, 16(%[src_ptr]) \n" // |19|18|17|16| + "lw $t6, 20(%[src_ptr]) \n" // |23|22|21|20| + "lw $t7, 24(%[src_ptr]) \n" // |27|26|25|24| + "lw $t8, 28(%[src_ptr]) \n" // |31|30|29|28| + "precrq.qb.ph $t0, $t2, $t4 \n" // |7|5|15|13| + "precrq.qb.ph $t9, $t6, $t8 \n" // |23|21|31|30| + "addiu %[dst_width], %[dst_width], -24 \n" + "ins $t1, $t1, 8, 16 \n" // |3|1|0|X| + "ins $t4, $t0, 8, 16 \n" // |X|15|13|12| + "ins $t5, $t5, 8, 16 \n" // |19|17|16|X| + "ins $t8, $t9, 8, 16 \n" // |X|31|29|28| + "addiu %[src_ptr], %[src_ptr], 32 \n" + "packrl.ph $t0, $t3, $t0 \n" // |9|8|7|5| + "packrl.ph $t9, $t7, $t9 \n" // |25|24|23|21| + "prepend $t1, $t2, 8 \n" // |4|3|1|0| + "prepend $t3, $t4, 24 \n" // |15|13|12|11| + "prepend $t5, $t6, 8 \n" // |20|19|17|16| + "prepend $t7, $t8, 24 \n" // |31|29|28|27| + "sw $t1, 0(%[dst]) \n" + "sw $t0, 4(%[dst]) \n" + "sw $t3, 8(%[dst]) \n" + "sw $t5, 12(%[dst]) \n" + "sw $t9, 16(%[dst]) \n" + "sw $t7, 20(%[dst]) \n" + "bnez %[dst_width], 1b \n" + " addiu %[dst], %[dst], 24 \n" + ".set pop \n" + : [src_ptr] "+r" (src_ptr), + [dst] "+r" (dst), + [dst_width] "+r" (dst_width) + : + : "t0", "t1", "t2", "t3", "t4", "t5", + "t6","t7", "t8", "t9" + ); +} + +void ScaleRowDown34_0_Box_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride, + uint8* d, int dst_width) { + __asm__ __volatile__ ( + ".set push \n" + ".set noreorder \n" + "repl.ph $t3, 3 \n" // 0x00030003 + + ".p2align 2 \n" + "1: \n" + "lw $t0, 0(%[src_ptr]) \n" // |S3|S2|S1|S0| + "lwx $t1, %[src_stride](%[src_ptr]) \n" // |T3|T2|T1|T0| + "rotr $t2, $t0, 8 \n" // |S0|S3|S2|S1| + "rotr $t6, $t1, 8 \n" // |T0|T3|T2|T1| + "muleu_s.ph.qbl $t4, $t2, $t3 \n" // |S0*3|S3*3| + "muleu_s.ph.qbl $t5, $t6, $t3 \n" // |T0*3|T3*3| + "andi $t0, $t2, 0xFFFF \n" // |0|0|S2|S1| + "andi $t1, $t6, 0xFFFF \n" // |0|0|T2|T1| + "raddu.w.qb $t0, $t0 \n" + "raddu.w.qb $t1, $t1 \n" + "shra_r.w $t0, $t0, 1 \n" + "shra_r.w $t1, $t1, 1 \n" + "preceu.ph.qbr $t2, $t2 \n" // |0|S2|0|S1| + "preceu.ph.qbr $t6, $t6 \n" // |0|T2|0|T1| + "rotr $t2, $t2, 16 \n" // |0|S1|0|S2| + "rotr $t6, $t6, 16 \n" // |0|T1|0|T2| + "addu.ph $t2, $t2, $t4 \n" + "addu.ph $t6, $t6, $t5 \n" + "sll $t5, $t0, 1 \n" + "add $t0, $t5, $t0 \n" + "shra_r.ph $t2, $t2, 2 \n" + "shra_r.ph $t6, $t6, 2 \n" + "shll.ph $t4, $t2, 1 \n" + "addq.ph $t4, $t4, $t2 \n" + "addu $t0, $t0, $t1 \n" + "addiu %[src_ptr], %[src_ptr], 4 \n" + "shra_r.w $t0, $t0, 2 \n" + "addu.ph $t6, $t6, $t4 \n" + "shra_r.ph $t6, $t6, 2 \n" + "srl $t1, $t6, 16 \n" + "addiu %[dst_width], %[dst_width], -3 \n" + "sb $t1, 0(%[d]) \n" + "sb $t0, 1(%[d]) \n" + "sb $t6, 2(%[d]) \n" + "bgtz %[dst_width], 1b \n" + " addiu %[d], %[d], 3 \n" + "3: \n" + ".set pop \n" + : [src_ptr] "+r" (src_ptr), + [src_stride] "+r" (src_stride), + [d] "+r" (d), + [dst_width] "+r" (dst_width) + : + : "t0", "t1", "t2", "t3", + "t4", "t5", "t6" + ); +} + +void ScaleRowDown34_1_Box_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride, + uint8* d, int dst_width) { + __asm__ __volatile__ ( + ".set push \n" + ".set noreorder \n" + "repl.ph $t2, 3 \n" // 0x00030003 + + ".p2align 2 \n" + "1: \n" + "lw $t0, 0(%[src_ptr]) \n" // |S3|S2|S1|S0| + "lwx $t1, %[src_stride](%[src_ptr]) \n" // |T3|T2|T1|T0| + "rotr $t4, $t0, 8 \n" // |S0|S3|S2|S1| + "rotr $t6, $t1, 8 \n" // |T0|T3|T2|T1| + "muleu_s.ph.qbl $t3, $t4, $t2 \n" // |S0*3|S3*3| + "muleu_s.ph.qbl $t5, $t6, $t2 \n" // |T0*3|T3*3| + "andi $t0, $t4, 0xFFFF \n" // |0|0|S2|S1| + "andi $t1, $t6, 0xFFFF \n" // |0|0|T2|T1| + "raddu.w.qb $t0, $t0 \n" + "raddu.w.qb $t1, $t1 \n" + "shra_r.w $t0, $t0, 1 \n" + "shra_r.w $t1, $t1, 1 \n" + "preceu.ph.qbr $t4, $t4 \n" // |0|S2|0|S1| + "preceu.ph.qbr $t6, $t6 \n" // |0|T2|0|T1| + "rotr $t4, $t4, 16 \n" // |0|S1|0|S2| + "rotr $t6, $t6, 16 \n" // |0|T1|0|T2| + "addu.ph $t4, $t4, $t3 \n" + "addu.ph $t6, $t6, $t5 \n" + "shra_r.ph $t6, $t6, 2 \n" + "shra_r.ph $t4, $t4, 2 \n" + "addu.ph $t6, $t6, $t4 \n" + "addiu %[src_ptr], %[src_ptr], 4 \n" + "shra_r.ph $t6, $t6, 1 \n" + "addu $t0, $t0, $t1 \n" + "addiu %[dst_width], %[dst_width], -3 \n" + "shra_r.w $t0, $t0, 1 \n" + "srl $t1, $t6, 16 \n" + "sb $t1, 0(%[d]) \n" + "sb $t0, 1(%[d]) \n" + "sb $t6, 2(%[d]) \n" + "bgtz %[dst_width], 1b \n" + " addiu %[d], %[d], 3 \n" + "3: \n" + ".set pop \n" + : [src_ptr] "+r" (src_ptr), + [src_stride] "+r" (src_stride), + [d] "+r" (d), + [dst_width] "+r" (dst_width) + : + : "t0", "t1", "t2", "t3", + "t4", "t5", "t6" + ); +} + +void ScaleRowDown38_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride, + uint8* dst, int dst_width) { + __asm__ __volatile__ ( + ".set push \n" + ".set noreorder \n" + + ".p2align 2 \n" + "1: \n" + "lw $t0, 0(%[src_ptr]) \n" // |3|2|1|0| + "lw $t1, 4(%[src_ptr]) \n" // |7|6|5|4| + "lw $t2, 8(%[src_ptr]) \n" // |11|10|9|8| + "lw $t3, 12(%[src_ptr]) \n" // |15|14|13|12| + "lw $t4, 16(%[src_ptr]) \n" // |19|18|17|16| + "lw $t5, 20(%[src_ptr]) \n" // |23|22|21|20| + "lw $t6, 24(%[src_ptr]) \n" // |27|26|25|24| + "lw $t7, 28(%[src_ptr]) \n" // |31|30|29|28| + "wsbh $t0, $t0 \n" // |2|3|0|1| + "wsbh $t6, $t6 \n" // |26|27|24|25| + "srl $t0, $t0, 8 \n" // |X|2|3|0| + "srl $t3, $t3, 16 \n" // |X|X|15|14| + "srl $t5, $t5, 16 \n" // |X|X|23|22| + "srl $t7, $t7, 16 \n" // |X|X|31|30| + "ins $t1, $t2, 24, 8 \n" // |8|6|5|4| + "ins $t6, $t5, 0, 8 \n" // |26|27|24|22| + "ins $t1, $t0, 0, 16 \n" // |8|6|3|0| + "ins $t6, $t7, 24, 8 \n" // |30|27|24|22| + "prepend $t2, $t3, 24 \n" // |X|15|14|11| + "ins $t4, $t4, 16, 8 \n" // |19|16|17|X| + "ins $t4, $t2, 0, 16 \n" // |19|16|14|11| + "addiu %[src_ptr], %[src_ptr], 32 \n" + "addiu %[dst_width], %[dst_width], -12 \n" + "addiu $t8,%[dst_width], -12 \n" + "sw $t1, 0(%[dst]) \n" + "sw $t4, 4(%[dst]) \n" + "sw $t6, 8(%[dst]) \n" + "bgez $t8, 1b \n" + " addiu %[dst], %[dst], 12 \n" + ".set pop \n" + : [src_ptr] "+r" (src_ptr), + [dst] "+r" (dst), + [dst_width] "+r" (dst_width) + : + : "t0", "t1", "t2", "t3", "t4", + "t5", "t6", "t7", "t8" + ); +} + +void ScaleRowDown38_2_Box_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride, + uint8* dst_ptr, int dst_width) { + intptr_t stride = src_stride; + const uint8* t = src_ptr + stride; + const int c = 0x2AAA; + + __asm__ __volatile__ ( + ".set push \n" + ".set noreorder \n" + + ".p2align 2 \n" + "1: \n" + "lw $t0, 0(%[src_ptr]) \n" // |S3|S2|S1|S0| + "lw $t1, 4(%[src_ptr]) \n" // |S7|S6|S5|S4| + "lw $t2, 0(%[t]) \n" // |T3|T2|T1|T0| + "lw $t3, 4(%[t]) \n" // |T7|T6|T5|T4| + "rotr $t1, $t1, 16 \n" // |S5|S4|S7|S6| + "packrl.ph $t4, $t1, $t3 \n" // |S7|S6|T7|T6| + "packrl.ph $t5, $t3, $t1 \n" // |T5|T4|S5|S4| + "raddu.w.qb $t4, $t4 \n" // S7+S6+T7+T6 + "raddu.w.qb $t5, $t5 \n" // T5+T4+S5+S4 + "precrq.qb.ph $t6, $t0, $t2 \n" // |S3|S1|T3|T1| + "precrq.qb.ph $t6, $t6, $t6 \n" // |S3|T3|S3|T3| + "srl $t4, $t4, 2 \n" // t4 / 4 + "srl $t6, $t6, 16 \n" // |0|0|S3|T3| + "raddu.w.qb $t6, $t6 \n" // 0+0+S3+T3 + "addu $t6, $t5, $t6 \n" + "mul $t6, $t6, %[c] \n" // t6 * 0x2AAA + "sll $t0, $t0, 8 \n" // |S2|S1|S0|0| + "sll $t2, $t2, 8 \n" // |T2|T1|T0|0| + "raddu.w.qb $t0, $t0 \n" // S2+S1+S0+0 + "raddu.w.qb $t2, $t2 \n" // T2+T1+T0+0 + "addu $t0, $t0, $t2 \n" + "mul $t0, $t0, %[c] \n" // t0 * 0x2AAA + "addiu %[src_ptr], %[src_ptr], 8 \n" + "addiu %[t], %[t], 8 \n" + "addiu %[dst_width], %[dst_width], -3 \n" + "addiu %[dst_ptr], %[dst_ptr], 3 \n" + "srl $t6, $t6, 16 \n" + "srl $t0, $t0, 16 \n" + "sb $t4, -1(%[dst_ptr]) \n" + "sb $t6, -2(%[dst_ptr]) \n" + "bgtz %[dst_width], 1b \n" + " sb $t0, -3(%[dst_ptr]) \n" + ".set pop \n" + : [src_ptr] "+r" (src_ptr), + [dst_ptr] "+r" (dst_ptr), + [t] "+r" (t), + [dst_width] "+r" (dst_width) + : [c] "r" (c) + : "t0", "t1", "t2", "t3", "t4", "t5", "t6" + ); +} + +void ScaleRowDown38_3_Box_MIPS_DSPR2(const uint8* src_ptr, + ptrdiff_t src_stride, + uint8* dst_ptr, int dst_width) { + intptr_t stride = src_stride; + const uint8* s1 = src_ptr + stride; + stride += stride; + const uint8* s2 = src_ptr + stride; + const int c1 = 0x1C71; + const int c2 = 0x2AAA; + + __asm__ __volatile__ ( + ".set push \n" + ".set noreorder \n" + + ".p2align 2 \n" + "1: \n" + "lw $t0, 0(%[src_ptr]) \n" // |S3|S2|S1|S0| + "lw $t1, 4(%[src_ptr]) \n" // |S7|S6|S5|S4| + "lw $t2, 0(%[s1]) \n" // |T3|T2|T1|T0| + "lw $t3, 4(%[s1]) \n" // |T7|T6|T5|T4| + "lw $t4, 0(%[s2]) \n" // |R3|R2|R1|R0| + "lw $t5, 4(%[s2]) \n" // |R7|R6|R5|R4| + "rotr $t1, $t1, 16 \n" // |S5|S4|S7|S6| + "packrl.ph $t6, $t1, $t3 \n" // |S7|S6|T7|T6| + "raddu.w.qb $t6, $t6 \n" // S7+S6+T7+T6 + "packrl.ph $t7, $t3, $t1 \n" // |T5|T4|S5|S4| + "raddu.w.qb $t7, $t7 \n" // T5+T4+S5+S4 + "sll $t8, $t5, 16 \n" // |R5|R4|0|0| + "raddu.w.qb $t8, $t8 \n" // R5+R4 + "addu $t7, $t7, $t8 \n" + "srl $t8, $t5, 16 \n" // |0|0|R7|R6| + "raddu.w.qb $t8, $t8 \n" // R7 + R6 + "addu $t6, $t6, $t8 \n" + "mul $t6, $t6, %[c2] \n" // t6 * 0x2AAA + "precrq.qb.ph $t8, $t0, $t2 \n" // |S3|S1|T3|T1| + "precrq.qb.ph $t8, $t8, $t4 \n" // |S3|T3|R3|R1| + "srl $t8, $t8, 8 \n" // |0|S3|T3|R3| + "raddu.w.qb $t8, $t8 \n" // S3 + T3 + R3 + "addu $t7, $t7, $t8 \n" + "mul $t7, $t7, %[c1] \n" // t7 * 0x1C71 + "sll $t0, $t0, 8 \n" // |S2|S1|S0|0| + "sll $t2, $t2, 8 \n" // |T2|T1|T0|0| + "sll $t4, $t4, 8 \n" // |R2|R1|R0|0| + "raddu.w.qb $t0, $t0 \n" + "raddu.w.qb $t2, $t2 \n" + "raddu.w.qb $t4, $t4 \n" + "addu $t0, $t0, $t2 \n" + "addu $t0, $t0, $t4 \n" + "mul $t0, $t0, %[c1] \n" // t0 * 0x1C71 + "addiu %[src_ptr], %[src_ptr], 8 \n" + "addiu %[s1], %[s1], 8 \n" + "addiu %[s2], %[s2], 8 \n" + "addiu %[dst_width], %[dst_width], -3 \n" + "addiu %[dst_ptr], %[dst_ptr], 3 \n" + "srl $t6, $t6, 16 \n" + "srl $t7, $t7, 16 \n" + "srl $t0, $t0, 16 \n" + "sb $t6, -1(%[dst_ptr]) \n" + "sb $t7, -2(%[dst_ptr]) \n" + "bgtz %[dst_width], 1b \n" + " sb $t0, -3(%[dst_ptr]) \n" + ".set pop \n" + : [src_ptr] "+r" (src_ptr), + [dst_ptr] "+r" (dst_ptr), + [s1] "+r" (s1), + [s2] "+r" (s2), + [dst_width] "+r" (dst_width) + : [c1] "r" (c1), [c2] "r" (c2) + : "t0", "t1", "t2", "t3", "t4", + "t5", "t6", "t7", "t8" + ); +} + +#endif // defined(__mips_dsp) && (__mips_dsp_rev >= 2) + +#ifdef __cplusplus +} // extern "C" +} // namespace libyuv +#endif + diff --git a/media/libaom/src/third_party/libyuv/source/scale_neon.cc b/media/libaom/src/third_party/libyuv/source/scale_neon.cc new file mode 100644 index 000000000..7825878e9 --- /dev/null +++ b/media/libaom/src/third_party/libyuv/source/scale_neon.cc @@ -0,0 +1,1037 @@ +/* + * Copyright 2011 The LibYuv Project Authors. All rights reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#include "libyuv/row.h" + +#ifdef __cplusplus +namespace libyuv { +extern "C" { +#endif + +// This module is for GCC Neon. +#if !defined(LIBYUV_DISABLE_NEON) && defined(__ARM_NEON__) && \ + !defined(__aarch64__) + +// NEON downscalers with interpolation. +// Provided by Fritz Koenig + +// Read 32x1 throw away even pixels, and write 16x1. +void ScaleRowDown2_NEON(const uint8* src_ptr, ptrdiff_t src_stride, + uint8* dst, int dst_width) { + asm volatile ( + ".p2align 2 \n" + "1: \n" + // load even pixels into q0, odd into q1 + MEMACCESS(0) + "vld2.8 {q0, q1}, [%0]! \n" + "subs %2, %2, #16 \n" // 16 processed per loop + MEMACCESS(1) + "vst1.8 {q1}, [%1]! \n" // store odd pixels + "bgt 1b \n" + : "+r"(src_ptr), // %0 + "+r"(dst), // %1 + "+r"(dst_width) // %2 + : + : "q0", "q1" // Clobber List + ); +} + +// Read 32x1 average down and write 16x1. +void ScaleRowDown2Linear_NEON(const uint8* src_ptr, ptrdiff_t src_stride, + uint8* dst, int dst_width) { + asm volatile ( + ".p2align 2 \n" + "1: \n" + MEMACCESS(0) + "vld1.8 {q0, q1}, [%0]! \n" // load pixels and post inc + "subs %2, %2, #16 \n" // 16 processed per loop + "vpaddl.u8 q0, q0 \n" // add adjacent + "vpaddl.u8 q1, q1 \n" + "vrshrn.u16 d0, q0, #1 \n" // downshift, round and pack + "vrshrn.u16 d1, q1, #1 \n" + MEMACCESS(1) + "vst1.8 {q0}, [%1]! \n" + "bgt 1b \n" + : "+r"(src_ptr), // %0 + "+r"(dst), // %1 + "+r"(dst_width) // %2 + : + : "q0", "q1" // Clobber List + ); +} + +// Read 32x2 average down and write 16x1. +void ScaleRowDown2Box_NEON(const uint8* src_ptr, ptrdiff_t src_stride, + uint8* dst, int dst_width) { + asm volatile ( + // change the stride to row 2 pointer + "add %1, %0 \n" + ".p2align 2 \n" + "1: \n" + MEMACCESS(0) + "vld1.8 {q0, q1}, [%0]! \n" // load row 1 and post inc + MEMACCESS(1) + "vld1.8 {q2, q3}, [%1]! \n" // load row 2 and post inc + "subs %3, %3, #16 \n" // 16 processed per loop + "vpaddl.u8 q0, q0 \n" // row 1 add adjacent + "vpaddl.u8 q1, q1 \n" + "vpadal.u8 q0, q2 \n" // row 2 add adjacent + row1 + "vpadal.u8 q1, q3 \n" + "vrshrn.u16 d0, q0, #2 \n" // downshift, round and pack + "vrshrn.u16 d1, q1, #2 \n" + MEMACCESS(2) + "vst1.8 {q0}, [%2]! \n" + "bgt 1b \n" + : "+r"(src_ptr), // %0 + "+r"(src_stride), // %1 + "+r"(dst), // %2 + "+r"(dst_width) // %3 + : + : "q0", "q1", "q2", "q3" // Clobber List + ); +} + +void ScaleRowDown4_NEON(const uint8* src_ptr, ptrdiff_t src_stride, + uint8* dst_ptr, int dst_width) { + asm volatile ( + ".p2align 2 \n" + "1: \n" + MEMACCESS(0) + "vld4.8 {d0, d1, d2, d3}, [%0]! \n" // src line 0 + "subs %2, %2, #8 \n" // 8 processed per loop + MEMACCESS(1) + "vst1.8 {d2}, [%1]! \n" + "bgt 1b \n" + : "+r"(src_ptr), // %0 + "+r"(dst_ptr), // %1 + "+r"(dst_width) // %2 + : + : "q0", "q1", "memory", "cc" + ); +} + +void ScaleRowDown4Box_NEON(const uint8* src_ptr, ptrdiff_t src_stride, + uint8* dst_ptr, int dst_width) { + const uint8* src_ptr1 = src_ptr + src_stride; + const uint8* src_ptr2 = src_ptr + src_stride * 2; + const uint8* src_ptr3 = src_ptr + src_stride * 3; +asm volatile ( + ".p2align 2 \n" + "1: \n" + MEMACCESS(0) + "vld1.8 {q0}, [%0]! \n" // load up 16x4 + MEMACCESS(3) + "vld1.8 {q1}, [%3]! \n" + MEMACCESS(4) + "vld1.8 {q2}, [%4]! \n" + MEMACCESS(5) + "vld1.8 {q3}, [%5]! \n" + "subs %2, %2, #4 \n" + "vpaddl.u8 q0, q0 \n" + "vpadal.u8 q0, q1 \n" + "vpadal.u8 q0, q2 \n" + "vpadal.u8 q0, q3 \n" + "vpaddl.u16 q0, q0 \n" + "vrshrn.u32 d0, q0, #4 \n" // divide by 16 w/rounding + "vmovn.u16 d0, q0 \n" + MEMACCESS(1) + "vst1.32 {d0[0]}, [%1]! \n" + "bgt 1b \n" + : "+r"(src_ptr), // %0 + "+r"(dst_ptr), // %1 + "+r"(dst_width), // %2 + "+r"(src_ptr1), // %3 + "+r"(src_ptr2), // %4 + "+r"(src_ptr3) // %5 + : + : "q0", "q1", "q2", "q3", "memory", "cc" + ); +} + +// Down scale from 4 to 3 pixels. Use the neon multilane read/write +// to load up the every 4th pixel into a 4 different registers. +// Point samples 32 pixels to 24 pixels. +void ScaleRowDown34_NEON(const uint8* src_ptr, + ptrdiff_t src_stride, + uint8* dst_ptr, int dst_width) { + asm volatile ( + ".p2align 2 \n" + "1: \n" + MEMACCESS(0) + "vld4.8 {d0, d1, d2, d3}, [%0]! \n" // src line 0 + "subs %2, %2, #24 \n" + "vmov d2, d3 \n" // order d0, d1, d2 + MEMACCESS(1) + "vst3.8 {d0, d1, d2}, [%1]! \n" + "bgt 1b \n" + : "+r"(src_ptr), // %0 + "+r"(dst_ptr), // %1 + "+r"(dst_width) // %2 + : + : "d0", "d1", "d2", "d3", "memory", "cc" + ); +} + +void ScaleRowDown34_0_Box_NEON(const uint8* src_ptr, + ptrdiff_t src_stride, + uint8* dst_ptr, int dst_width) { + asm volatile ( + "vmov.u8 d24, #3 \n" + "add %3, %0 \n" + ".p2align 2 \n" + "1: \n" + MEMACCESS(0) + "vld4.8 {d0, d1, d2, d3}, [%0]! \n" // src line 0 + MEMACCESS(3) + "vld4.8 {d4, d5, d6, d7}, [%3]! \n" // src line 1 + "subs %2, %2, #24 \n" + + // filter src line 0 with src line 1 + // expand chars to shorts to allow for room + // when adding lines together + "vmovl.u8 q8, d4 \n" + "vmovl.u8 q9, d5 \n" + "vmovl.u8 q10, d6 \n" + "vmovl.u8 q11, d7 \n" + + // 3 * line_0 + line_1 + "vmlal.u8 q8, d0, d24 \n" + "vmlal.u8 q9, d1, d24 \n" + "vmlal.u8 q10, d2, d24 \n" + "vmlal.u8 q11, d3, d24 \n" + + // (3 * line_0 + line_1) >> 2 + "vqrshrn.u16 d0, q8, #2 \n" + "vqrshrn.u16 d1, q9, #2 \n" + "vqrshrn.u16 d2, q10, #2 \n" + "vqrshrn.u16 d3, q11, #2 \n" + + // a0 = (src[0] * 3 + s[1] * 1) >> 2 + "vmovl.u8 q8, d1 \n" + "vmlal.u8 q8, d0, d24 \n" + "vqrshrn.u16 d0, q8, #2 \n" + + // a1 = (src[1] * 1 + s[2] * 1) >> 1 + "vrhadd.u8 d1, d1, d2 \n" + + // a2 = (src[2] * 1 + s[3] * 3) >> 2 + "vmovl.u8 q8, d2 \n" + "vmlal.u8 q8, d3, d24 \n" + "vqrshrn.u16 d2, q8, #2 \n" + + MEMACCESS(1) + "vst3.8 {d0, d1, d2}, [%1]! \n" + + "bgt 1b \n" + : "+r"(src_ptr), // %0 + "+r"(dst_ptr), // %1 + "+r"(dst_width), // %2 + "+r"(src_stride) // %3 + : + : "q0", "q1", "q2", "q3", "q8", "q9", "q10", "q11", "d24", "memory", "cc" + ); +} + +void ScaleRowDown34_1_Box_NEON(const uint8* src_ptr, + ptrdiff_t src_stride, + uint8* dst_ptr, int dst_width) { + asm volatile ( + "vmov.u8 d24, #3 \n" + "add %3, %0 \n" + ".p2align 2 \n" + "1: \n" + MEMACCESS(0) + "vld4.8 {d0, d1, d2, d3}, [%0]! \n" // src line 0 + MEMACCESS(3) + "vld4.8 {d4, d5, d6, d7}, [%3]! \n" // src line 1 + "subs %2, %2, #24 \n" + // average src line 0 with src line 1 + "vrhadd.u8 q0, q0, q2 \n" + "vrhadd.u8 q1, q1, q3 \n" + + // a0 = (src[0] * 3 + s[1] * 1) >> 2 + "vmovl.u8 q3, d1 \n" + "vmlal.u8 q3, d0, d24 \n" + "vqrshrn.u16 d0, q3, #2 \n" + + // a1 = (src[1] * 1 + s[2] * 1) >> 1 + "vrhadd.u8 d1, d1, d2 \n" + + // a2 = (src[2] * 1 + s[3] * 3) >> 2 + "vmovl.u8 q3, d2 \n" + "vmlal.u8 q3, d3, d24 \n" + "vqrshrn.u16 d2, q3, #2 \n" + + MEMACCESS(1) + "vst3.8 {d0, d1, d2}, [%1]! \n" + "bgt 1b \n" + : "+r"(src_ptr), // %0 + "+r"(dst_ptr), // %1 + "+r"(dst_width), // %2 + "+r"(src_stride) // %3 + : + : "r4", "q0", "q1", "q2", "q3", "d24", "memory", "cc" + ); +} + +#define HAS_SCALEROWDOWN38_NEON +static uvec8 kShuf38 = + { 0, 3, 6, 8, 11, 14, 16, 19, 22, 24, 27, 30, 0, 0, 0, 0 }; +static uvec8 kShuf38_2 = + { 0, 8, 16, 2, 10, 17, 4, 12, 18, 6, 14, 19, 0, 0, 0, 0 }; +static vec16 kMult38_Div6 = + { 65536 / 12, 65536 / 12, 65536 / 12, 65536 / 12, + 65536 / 12, 65536 / 12, 65536 / 12, 65536 / 12 }; +static vec16 kMult38_Div9 = + { 65536 / 18, 65536 / 18, 65536 / 18, 65536 / 18, + 65536 / 18, 65536 / 18, 65536 / 18, 65536 / 18 }; + +// 32 -> 12 +void ScaleRowDown38_NEON(const uint8* src_ptr, + ptrdiff_t src_stride, + uint8* dst_ptr, int dst_width) { + asm volatile ( + MEMACCESS(3) + "vld1.8 {q3}, [%3] \n" + ".p2align 2 \n" + "1: \n" + MEMACCESS(0) + "vld1.8 {d0, d1, d2, d3}, [%0]! \n" + "subs %2, %2, #12 \n" + "vtbl.u8 d4, {d0, d1, d2, d3}, d6 \n" + "vtbl.u8 d5, {d0, d1, d2, d3}, d7 \n" + MEMACCESS(1) + "vst1.8 {d4}, [%1]! \n" + MEMACCESS(1) + "vst1.32 {d5[0]}, [%1]! \n" + "bgt 1b \n" + : "+r"(src_ptr), // %0 + "+r"(dst_ptr), // %1 + "+r"(dst_width) // %2 + : "r"(&kShuf38) // %3 + : "d0", "d1", "d2", "d3", "d4", "d5", "memory", "cc" + ); +} + +// 32x3 -> 12x1 +void OMITFP ScaleRowDown38_3_Box_NEON(const uint8* src_ptr, + ptrdiff_t src_stride, + uint8* dst_ptr, int dst_width) { + const uint8* src_ptr1 = src_ptr + src_stride * 2; + + asm volatile ( + MEMACCESS(5) + "vld1.16 {q13}, [%5] \n" + MEMACCESS(6) + "vld1.8 {q14}, [%6] \n" + MEMACCESS(7) + "vld1.8 {q15}, [%7] \n" + "add %3, %0 \n" + ".p2align 2 \n" + "1: \n" + + // d0 = 00 40 01 41 02 42 03 43 + // d1 = 10 50 11 51 12 52 13 53 + // d2 = 20 60 21 61 22 62 23 63 + // d3 = 30 70 31 71 32 72 33 73 + MEMACCESS(0) + "vld4.8 {d0, d1, d2, d3}, [%0]! \n" + MEMACCESS(3) + "vld4.8 {d4, d5, d6, d7}, [%3]! \n" + MEMACCESS(4) + "vld4.8 {d16, d17, d18, d19}, [%4]! \n" + "subs %2, %2, #12 \n" + + // Shuffle the input data around to get align the data + // so adjacent data can be added. 0,1 - 2,3 - 4,5 - 6,7 + // d0 = 00 10 01 11 02 12 03 13 + // d1 = 40 50 41 51 42 52 43 53 + "vtrn.u8 d0, d1 \n" + "vtrn.u8 d4, d5 \n" + "vtrn.u8 d16, d17 \n" + + // d2 = 20 30 21 31 22 32 23 33 + // d3 = 60 70 61 71 62 72 63 73 + "vtrn.u8 d2, d3 \n" + "vtrn.u8 d6, d7 \n" + "vtrn.u8 d18, d19 \n" + + // d0 = 00+10 01+11 02+12 03+13 + // d2 = 40+50 41+51 42+52 43+53 + "vpaddl.u8 q0, q0 \n" + "vpaddl.u8 q2, q2 \n" + "vpaddl.u8 q8, q8 \n" + + // d3 = 60+70 61+71 62+72 63+73 + "vpaddl.u8 d3, d3 \n" + "vpaddl.u8 d7, d7 \n" + "vpaddl.u8 d19, d19 \n" + + // combine source lines + "vadd.u16 q0, q2 \n" + "vadd.u16 q0, q8 \n" + "vadd.u16 d4, d3, d7 \n" + "vadd.u16 d4, d19 \n" + + // dst_ptr[3] = (s[6 + st * 0] + s[7 + st * 0] + // + s[6 + st * 1] + s[7 + st * 1] + // + s[6 + st * 2] + s[7 + st * 2]) / 6 + "vqrdmulh.s16 q2, q2, q13 \n" + "vmovn.u16 d4, q2 \n" + + // Shuffle 2,3 reg around so that 2 can be added to the + // 0,1 reg and 3 can be added to the 4,5 reg. This + // requires expanding from u8 to u16 as the 0,1 and 4,5 + // registers are already expanded. Then do transposes + // to get aligned. + // q2 = xx 20 xx 30 xx 21 xx 31 xx 22 xx 32 xx 23 xx 33 + "vmovl.u8 q1, d2 \n" + "vmovl.u8 q3, d6 \n" + "vmovl.u8 q9, d18 \n" + + // combine source lines + "vadd.u16 q1, q3 \n" + "vadd.u16 q1, q9 \n" + + // d4 = xx 20 xx 30 xx 22 xx 32 + // d5 = xx 21 xx 31 xx 23 xx 33 + "vtrn.u32 d2, d3 \n" + + // d4 = xx 20 xx 21 xx 22 xx 23 + // d5 = xx 30 xx 31 xx 32 xx 33 + "vtrn.u16 d2, d3 \n" + + // 0+1+2, 3+4+5 + "vadd.u16 q0, q1 \n" + + // Need to divide, but can't downshift as the the value + // isn't a power of 2. So multiply by 65536 / n + // and take the upper 16 bits. + "vqrdmulh.s16 q0, q0, q15 \n" + + // Align for table lookup, vtbl requires registers to + // be adjacent + "vmov.u8 d2, d4 \n" + + "vtbl.u8 d3, {d0, d1, d2}, d28 \n" + "vtbl.u8 d4, {d0, d1, d2}, d29 \n" + + MEMACCESS(1) + "vst1.8 {d3}, [%1]! \n" + MEMACCESS(1) + "vst1.32 {d4[0]}, [%1]! \n" + "bgt 1b \n" + : "+r"(src_ptr), // %0 + "+r"(dst_ptr), // %1 + "+r"(dst_width), // %2 + "+r"(src_stride), // %3 + "+r"(src_ptr1) // %4 + : "r"(&kMult38_Div6), // %5 + "r"(&kShuf38_2), // %6 + "r"(&kMult38_Div9) // %7 + : "q0", "q1", "q2", "q3", "q8", "q9", "q13", "q14", "q15", "memory", "cc" + ); +} + +// 32x2 -> 12x1 +void ScaleRowDown38_2_Box_NEON(const uint8* src_ptr, + ptrdiff_t src_stride, + uint8* dst_ptr, int dst_width) { + asm volatile ( + MEMACCESS(4) + "vld1.16 {q13}, [%4] \n" + MEMACCESS(5) + "vld1.8 {q14}, [%5] \n" + "add %3, %0 \n" + ".p2align 2 \n" + "1: \n" + + // d0 = 00 40 01 41 02 42 03 43 + // d1 = 10 50 11 51 12 52 13 53 + // d2 = 20 60 21 61 22 62 23 63 + // d3 = 30 70 31 71 32 72 33 73 + MEMACCESS(0) + "vld4.8 {d0, d1, d2, d3}, [%0]! \n" + MEMACCESS(3) + "vld4.8 {d4, d5, d6, d7}, [%3]! \n" + "subs %2, %2, #12 \n" + + // Shuffle the input data around to get align the data + // so adjacent data can be added. 0,1 - 2,3 - 4,5 - 6,7 + // d0 = 00 10 01 11 02 12 03 13 + // d1 = 40 50 41 51 42 52 43 53 + "vtrn.u8 d0, d1 \n" + "vtrn.u8 d4, d5 \n" + + // d2 = 20 30 21 31 22 32 23 33 + // d3 = 60 70 61 71 62 72 63 73 + "vtrn.u8 d2, d3 \n" + "vtrn.u8 d6, d7 \n" + + // d0 = 00+10 01+11 02+12 03+13 + // d2 = 40+50 41+51 42+52 43+53 + "vpaddl.u8 q0, q0 \n" + "vpaddl.u8 q2, q2 \n" + + // d3 = 60+70 61+71 62+72 63+73 + "vpaddl.u8 d3, d3 \n" + "vpaddl.u8 d7, d7 \n" + + // combine source lines + "vadd.u16 q0, q2 \n" + "vadd.u16 d4, d3, d7 \n" + + // dst_ptr[3] = (s[6] + s[7] + s[6+st] + s[7+st]) / 4 + "vqrshrn.u16 d4, q2, #2 \n" + + // Shuffle 2,3 reg around so that 2 can be added to the + // 0,1 reg and 3 can be added to the 4,5 reg. This + // requires expanding from u8 to u16 as the 0,1 and 4,5 + // registers are already expanded. Then do transposes + // to get aligned. + // q2 = xx 20 xx 30 xx 21 xx 31 xx 22 xx 32 xx 23 xx 33 + "vmovl.u8 q1, d2 \n" + "vmovl.u8 q3, d6 \n" + + // combine source lines + "vadd.u16 q1, q3 \n" + + // d4 = xx 20 xx 30 xx 22 xx 32 + // d5 = xx 21 xx 31 xx 23 xx 33 + "vtrn.u32 d2, d3 \n" + + // d4 = xx 20 xx 21 xx 22 xx 23 + // d5 = xx 30 xx 31 xx 32 xx 33 + "vtrn.u16 d2, d3 \n" + + // 0+1+2, 3+4+5 + "vadd.u16 q0, q1 \n" + + // Need to divide, but can't downshift as the the value + // isn't a power of 2. So multiply by 65536 / n + // and take the upper 16 bits. + "vqrdmulh.s16 q0, q0, q13 \n" + + // Align for table lookup, vtbl requires registers to + // be adjacent + "vmov.u8 d2, d4 \n" + + "vtbl.u8 d3, {d0, d1, d2}, d28 \n" + "vtbl.u8 d4, {d0, d1, d2}, d29 \n" + + MEMACCESS(1) + "vst1.8 {d3}, [%1]! \n" + MEMACCESS(1) + "vst1.32 {d4[0]}, [%1]! \n" + "bgt 1b \n" + : "+r"(src_ptr), // %0 + "+r"(dst_ptr), // %1 + "+r"(dst_width), // %2 + "+r"(src_stride) // %3 + : "r"(&kMult38_Div6), // %4 + "r"(&kShuf38_2) // %5 + : "q0", "q1", "q2", "q3", "q13", "q14", "memory", "cc" + ); +} + +void ScaleAddRows_NEON(const uint8* src_ptr, ptrdiff_t src_stride, + uint16* dst_ptr, int src_width, int src_height) { + const uint8* src_tmp = NULL; + asm volatile ( + ".p2align 2 \n" + "1: \n" + "mov %0, %1 \n" + "mov r12, %5 \n" + "veor q2, q2, q2 \n" + "veor q3, q3, q3 \n" + "2: \n" + // load 16 pixels into q0 + MEMACCESS(0) + "vld1.8 {q0}, [%0], %3 \n" + "vaddw.u8 q3, q3, d1 \n" + "vaddw.u8 q2, q2, d0 \n" + "subs r12, r12, #1 \n" + "bgt 2b \n" + MEMACCESS(2) + "vst1.16 {q2, q3}, [%2]! \n" // store pixels + "add %1, %1, #16 \n" + "subs %4, %4, #16 \n" // 16 processed per loop + "bgt 1b \n" + : "+r"(src_tmp), // %0 + "+r"(src_ptr), // %1 + "+r"(dst_ptr), // %2 + "+r"(src_stride), // %3 + "+r"(src_width), // %4 + "+r"(src_height) // %5 + : + : "memory", "cc", "r12", "q0", "q1", "q2", "q3" // Clobber List + ); +} + +// TODO(Yang Zhang): Investigate less load instructions for +// the x/dx stepping +#define LOAD2_DATA8_LANE(n) \ + "lsr %5, %3, #16 \n" \ + "add %6, %1, %5 \n" \ + "add %3, %3, %4 \n" \ + MEMACCESS(6) \ + "vld2.8 {d6["#n"], d7["#n"]}, [%6] \n" + +void ScaleFilterCols_NEON(uint8* dst_ptr, const uint8* src_ptr, + int dst_width, int x, int dx) { + int dx_offset[4] = {0, 1, 2, 3}; + int* tmp = dx_offset; + const uint8* src_tmp = src_ptr; + asm volatile ( + ".p2align 2 \n" + "vdup.32 q0, %3 \n" // x + "vdup.32 q1, %4 \n" // dx + "vld1.32 {q2}, [%5] \n" // 0 1 2 3 + "vshl.i32 q3, q1, #2 \n" // 4 * dx + "vmul.s32 q1, q1, q2 \n" + // x , x + 1 * dx, x + 2 * dx, x + 3 * dx + "vadd.s32 q1, q1, q0 \n" + // x + 4 * dx, x + 5 * dx, x + 6 * dx, x + 7 * dx + "vadd.s32 q2, q1, q3 \n" + "vshl.i32 q0, q3, #1 \n" // 8 * dx + "1: \n" + LOAD2_DATA8_LANE(0) + LOAD2_DATA8_LANE(1) + LOAD2_DATA8_LANE(2) + LOAD2_DATA8_LANE(3) + LOAD2_DATA8_LANE(4) + LOAD2_DATA8_LANE(5) + LOAD2_DATA8_LANE(6) + LOAD2_DATA8_LANE(7) + "vmov q10, q1 \n" + "vmov q11, q2 \n" + "vuzp.16 q10, q11 \n" + "vmovl.u8 q8, d6 \n" + "vmovl.u8 q9, d7 \n" + "vsubl.s16 q11, d18, d16 \n" + "vsubl.s16 q12, d19, d17 \n" + "vmovl.u16 q13, d20 \n" + "vmovl.u16 q10, d21 \n" + "vmul.s32 q11, q11, q13 \n" + "vmul.s32 q12, q12, q10 \n" + "vshrn.s32 d18, q11, #16 \n" + "vshrn.s32 d19, q12, #16 \n" + "vadd.s16 q8, q8, q9 \n" + "vmovn.s16 d6, q8 \n" + + MEMACCESS(0) + "vst1.8 {d6}, [%0]! \n" // store pixels + "vadd.s32 q1, q1, q0 \n" + "vadd.s32 q2, q2, q0 \n" + "subs %2, %2, #8 \n" // 8 processed per loop + "bgt 1b \n" + : "+r"(dst_ptr), // %0 + "+r"(src_ptr), // %1 + "+r"(dst_width), // %2 + "+r"(x), // %3 + "+r"(dx), // %4 + "+r"(tmp), // %5 + "+r"(src_tmp) // %6 + : + : "memory", "cc", "q0", "q1", "q2", "q3", + "q8", "q9", "q10", "q11", "q12", "q13" + ); +} + +#undef LOAD2_DATA8_LANE + +// 16x2 -> 16x1 +void ScaleFilterRows_NEON(uint8* dst_ptr, + const uint8* src_ptr, ptrdiff_t src_stride, + int dst_width, int source_y_fraction) { + asm volatile ( + "cmp %4, #0 \n" + "beq 100f \n" + "add %2, %1 \n" + "cmp %4, #64 \n" + "beq 75f \n" + "cmp %4, #128 \n" + "beq 50f \n" + "cmp %4, #192 \n" + "beq 25f \n" + + "vdup.8 d5, %4 \n" + "rsb %4, #256 \n" + "vdup.8 d4, %4 \n" + // General purpose row blend. + "1: \n" + MEMACCESS(1) + "vld1.8 {q0}, [%1]! \n" + MEMACCESS(2) + "vld1.8 {q1}, [%2]! \n" + "subs %3, %3, #16 \n" + "vmull.u8 q13, d0, d4 \n" + "vmull.u8 q14, d1, d4 \n" + "vmlal.u8 q13, d2, d5 \n" + "vmlal.u8 q14, d3, d5 \n" + "vrshrn.u16 d0, q13, #8 \n" + "vrshrn.u16 d1, q14, #8 \n" + MEMACCESS(0) + "vst1.8 {q0}, [%0]! \n" + "bgt 1b \n" + "b 99f \n" + + // Blend 25 / 75. + "25: \n" + MEMACCESS(1) + "vld1.8 {q0}, [%1]! \n" + MEMACCESS(2) + "vld1.8 {q1}, [%2]! \n" + "subs %3, %3, #16 \n" + "vrhadd.u8 q0, q1 \n" + "vrhadd.u8 q0, q1 \n" + MEMACCESS(0) + "vst1.8 {q0}, [%0]! \n" + "bgt 25b \n" + "b 99f \n" + + // Blend 50 / 50. + "50: \n" + MEMACCESS(1) + "vld1.8 {q0}, [%1]! \n" + MEMACCESS(2) + "vld1.8 {q1}, [%2]! \n" + "subs %3, %3, #16 \n" + "vrhadd.u8 q0, q1 \n" + MEMACCESS(0) + "vst1.8 {q0}, [%0]! \n" + "bgt 50b \n" + "b 99f \n" + + // Blend 75 / 25. + "75: \n" + MEMACCESS(1) + "vld1.8 {q1}, [%1]! \n" + MEMACCESS(2) + "vld1.8 {q0}, [%2]! \n" + "subs %3, %3, #16 \n" + "vrhadd.u8 q0, q1 \n" + "vrhadd.u8 q0, q1 \n" + MEMACCESS(0) + "vst1.8 {q0}, [%0]! \n" + "bgt 75b \n" + "b 99f \n" + + // Blend 100 / 0 - Copy row unchanged. + "100: \n" + MEMACCESS(1) + "vld1.8 {q0}, [%1]! \n" + "subs %3, %3, #16 \n" + MEMACCESS(0) + "vst1.8 {q0}, [%0]! \n" + "bgt 100b \n" + + "99: \n" + MEMACCESS(0) + "vst1.8 {d1[7]}, [%0] \n" + : "+r"(dst_ptr), // %0 + "+r"(src_ptr), // %1 + "+r"(src_stride), // %2 + "+r"(dst_width), // %3 + "+r"(source_y_fraction) // %4 + : + : "q0", "q1", "d4", "d5", "q13", "q14", "memory", "cc" + ); +} + +void ScaleARGBRowDown2_NEON(const uint8* src_ptr, ptrdiff_t src_stride, + uint8* dst, int dst_width) { + asm volatile ( + ".p2align 2 \n" + "1: \n" + // load even pixels into q0, odd into q1 + MEMACCESS(0) + "vld2.32 {q0, q1}, [%0]! \n" + MEMACCESS(0) + "vld2.32 {q2, q3}, [%0]! \n" + "subs %2, %2, #8 \n" // 8 processed per loop + MEMACCESS(1) + "vst1.8 {q1}, [%1]! \n" // store odd pixels + MEMACCESS(1) + "vst1.8 {q3}, [%1]! \n" + "bgt 1b \n" + : "+r"(src_ptr), // %0 + "+r"(dst), // %1 + "+r"(dst_width) // %2 + : + : "memory", "cc", "q0", "q1", "q2", "q3" // Clobber List + ); +} + +void ScaleARGBRowDown2Linear_NEON(const uint8* src_argb, ptrdiff_t src_stride, + uint8* dst_argb, int dst_width) { + asm volatile ( + ".p2align 2 \n" + "1: \n" + MEMACCESS(0) + "vld4.8 {d0, d2, d4, d6}, [%0]! \n" // load 8 ARGB pixels. + MEMACCESS(0) + "vld4.8 {d1, d3, d5, d7}, [%0]! \n" // load next 8 ARGB pixels. + "subs %2, %2, #8 \n" // 8 processed per loop + "vpaddl.u8 q0, q0 \n" // B 16 bytes -> 8 shorts. + "vpaddl.u8 q1, q1 \n" // G 16 bytes -> 8 shorts. + "vpaddl.u8 q2, q2 \n" // R 16 bytes -> 8 shorts. + "vpaddl.u8 q3, q3 \n" // A 16 bytes -> 8 shorts. + "vrshrn.u16 d0, q0, #1 \n" // downshift, round and pack + "vrshrn.u16 d1, q1, #1 \n" + "vrshrn.u16 d2, q2, #1 \n" + "vrshrn.u16 d3, q3, #1 \n" + MEMACCESS(1) + "vst4.8 {d0, d1, d2, d3}, [%1]! \n" + "bgt 1b \n" + : "+r"(src_argb), // %0 + "+r"(dst_argb), // %1 + "+r"(dst_width) // %2 + : + : "memory", "cc", "q0", "q1", "q2", "q3" // Clobber List + ); +} + +void ScaleARGBRowDown2Box_NEON(const uint8* src_ptr, ptrdiff_t src_stride, + uint8* dst, int dst_width) { + asm volatile ( + // change the stride to row 2 pointer + "add %1, %1, %0 \n" + ".p2align 2 \n" + "1: \n" + MEMACCESS(0) + "vld4.8 {d0, d2, d4, d6}, [%0]! \n" // load 8 ARGB pixels. + MEMACCESS(0) + "vld4.8 {d1, d3, d5, d7}, [%0]! \n" // load next 8 ARGB pixels. + "subs %3, %3, #8 \n" // 8 processed per loop. + "vpaddl.u8 q0, q0 \n" // B 16 bytes -> 8 shorts. + "vpaddl.u8 q1, q1 \n" // G 16 bytes -> 8 shorts. + "vpaddl.u8 q2, q2 \n" // R 16 bytes -> 8 shorts. + "vpaddl.u8 q3, q3 \n" // A 16 bytes -> 8 shorts. + MEMACCESS(1) + "vld4.8 {d16, d18, d20, d22}, [%1]! \n" // load 8 more ARGB pixels. + MEMACCESS(1) + "vld4.8 {d17, d19, d21, d23}, [%1]! \n" // load last 8 ARGB pixels. + "vpadal.u8 q0, q8 \n" // B 16 bytes -> 8 shorts. + "vpadal.u8 q1, q9 \n" // G 16 bytes -> 8 shorts. + "vpadal.u8 q2, q10 \n" // R 16 bytes -> 8 shorts. + "vpadal.u8 q3, q11 \n" // A 16 bytes -> 8 shorts. + "vrshrn.u16 d0, q0, #2 \n" // downshift, round and pack + "vrshrn.u16 d1, q1, #2 \n" + "vrshrn.u16 d2, q2, #2 \n" + "vrshrn.u16 d3, q3, #2 \n" + MEMACCESS(2) + "vst4.8 {d0, d1, d2, d3}, [%2]! \n" + "bgt 1b \n" + : "+r"(src_ptr), // %0 + "+r"(src_stride), // %1 + "+r"(dst), // %2 + "+r"(dst_width) // %3 + : + : "memory", "cc", "q0", "q1", "q2", "q3", "q8", "q9", "q10", "q11" + ); +} + +// Reads 4 pixels at a time. +// Alignment requirement: src_argb 4 byte aligned. +void ScaleARGBRowDownEven_NEON(const uint8* src_argb, ptrdiff_t src_stride, + int src_stepx, uint8* dst_argb, int dst_width) { + asm volatile ( + "mov r12, %3, lsl #2 \n" + ".p2align 2 \n" + "1: \n" + MEMACCESS(0) + "vld1.32 {d0[0]}, [%0], r12 \n" + MEMACCESS(0) + "vld1.32 {d0[1]}, [%0], r12 \n" + MEMACCESS(0) + "vld1.32 {d1[0]}, [%0], r12 \n" + MEMACCESS(0) + "vld1.32 {d1[1]}, [%0], r12 \n" + "subs %2, %2, #4 \n" // 4 pixels per loop. + MEMACCESS(1) + "vst1.8 {q0}, [%1]! \n" + "bgt 1b \n" + : "+r"(src_argb), // %0 + "+r"(dst_argb), // %1 + "+r"(dst_width) // %2 + : "r"(src_stepx) // %3 + : "memory", "cc", "r12", "q0" + ); +} + +// Reads 4 pixels at a time. +// Alignment requirement: src_argb 4 byte aligned. +void ScaleARGBRowDownEvenBox_NEON(const uint8* src_argb, ptrdiff_t src_stride, + int src_stepx, + uint8* dst_argb, int dst_width) { + asm volatile ( + "mov r12, %4, lsl #2 \n" + "add %1, %1, %0 \n" + ".p2align 2 \n" + "1: \n" + MEMACCESS(0) + "vld1.8 {d0}, [%0], r12 \n" // Read 4 2x2 blocks -> 2x1 + MEMACCESS(1) + "vld1.8 {d1}, [%1], r12 \n" + MEMACCESS(0) + "vld1.8 {d2}, [%0], r12 \n" + MEMACCESS(1) + "vld1.8 {d3}, [%1], r12 \n" + MEMACCESS(0) + "vld1.8 {d4}, [%0], r12 \n" + MEMACCESS(1) + "vld1.8 {d5}, [%1], r12 \n" + MEMACCESS(0) + "vld1.8 {d6}, [%0], r12 \n" + MEMACCESS(1) + "vld1.8 {d7}, [%1], r12 \n" + "vaddl.u8 q0, d0, d1 \n" + "vaddl.u8 q1, d2, d3 \n" + "vaddl.u8 q2, d4, d5 \n" + "vaddl.u8 q3, d6, d7 \n" + "vswp.8 d1, d2 \n" // ab_cd -> ac_bd + "vswp.8 d5, d6 \n" // ef_gh -> eg_fh + "vadd.u16 q0, q0, q1 \n" // (a+b)_(c+d) + "vadd.u16 q2, q2, q3 \n" // (e+f)_(g+h) + "vrshrn.u16 d0, q0, #2 \n" // first 2 pixels. + "vrshrn.u16 d1, q2, #2 \n" // next 2 pixels. + "subs %3, %3, #4 \n" // 4 pixels per loop. + MEMACCESS(2) + "vst1.8 {q0}, [%2]! \n" + "bgt 1b \n" + : "+r"(src_argb), // %0 + "+r"(src_stride), // %1 + "+r"(dst_argb), // %2 + "+r"(dst_width) // %3 + : "r"(src_stepx) // %4 + : "memory", "cc", "r12", "q0", "q1", "q2", "q3" + ); +} + +// TODO(Yang Zhang): Investigate less load instructions for +// the x/dx stepping +#define LOAD1_DATA32_LANE(dn, n) \ + "lsr %5, %3, #16 \n" \ + "add %6, %1, %5, lsl #2 \n" \ + "add %3, %3, %4 \n" \ + MEMACCESS(6) \ + "vld1.32 {"#dn"["#n"]}, [%6] \n" + +void ScaleARGBCols_NEON(uint8* dst_argb, const uint8* src_argb, + int dst_width, int x, int dx) { + int tmp = 0; + const uint8* src_tmp = src_argb; + asm volatile ( + ".p2align 2 \n" + "1: \n" + LOAD1_DATA32_LANE(d0, 0) + LOAD1_DATA32_LANE(d0, 1) + LOAD1_DATA32_LANE(d1, 0) + LOAD1_DATA32_LANE(d1, 1) + LOAD1_DATA32_LANE(d2, 0) + LOAD1_DATA32_LANE(d2, 1) + LOAD1_DATA32_LANE(d3, 0) + LOAD1_DATA32_LANE(d3, 1) + + MEMACCESS(0) + "vst1.32 {q0, q1}, [%0]! \n" // store pixels + "subs %2, %2, #8 \n" // 8 processed per loop + "bgt 1b \n" + : "+r"(dst_argb), // %0 + "+r"(src_argb), // %1 + "+r"(dst_width), // %2 + "+r"(x), // %3 + "+r"(dx), // %4 + "+r"(tmp), // %5 + "+r"(src_tmp) // %6 + : + : "memory", "cc", "q0", "q1" + ); +} + +#undef LOAD1_DATA32_LANE + +// TODO(Yang Zhang): Investigate less load instructions for +// the x/dx stepping +#define LOAD2_DATA32_LANE(dn1, dn2, n) \ + "lsr %5, %3, #16 \n" \ + "add %6, %1, %5, lsl #2 \n" \ + "add %3, %3, %4 \n" \ + MEMACCESS(6) \ + "vld2.32 {"#dn1"["#n"], "#dn2"["#n"]}, [%6] \n" + +void ScaleARGBFilterCols_NEON(uint8* dst_argb, const uint8* src_argb, + int dst_width, int x, int dx) { + int dx_offset[4] = {0, 1, 2, 3}; + int* tmp = dx_offset; + const uint8* src_tmp = src_argb; + asm volatile ( + ".p2align 2 \n" + "vdup.32 q0, %3 \n" // x + "vdup.32 q1, %4 \n" // dx + "vld1.32 {q2}, [%5] \n" // 0 1 2 3 + "vshl.i32 q9, q1, #2 \n" // 4 * dx + "vmul.s32 q1, q1, q2 \n" + "vmov.i8 q3, #0x7f \n" // 0x7F + "vmov.i16 q15, #0x7f \n" // 0x7F + // x , x + 1 * dx, x + 2 * dx, x + 3 * dx + "vadd.s32 q8, q1, q0 \n" + "1: \n" + // d0, d1: a + // d2, d3: b + LOAD2_DATA32_LANE(d0, d2, 0) + LOAD2_DATA32_LANE(d0, d2, 1) + LOAD2_DATA32_LANE(d1, d3, 0) + LOAD2_DATA32_LANE(d1, d3, 1) + "vshrn.i32 d22, q8, #9 \n" + "vand.16 d22, d22, d30 \n" + "vdup.8 d24, d22[0] \n" + "vdup.8 d25, d22[2] \n" + "vdup.8 d26, d22[4] \n" + "vdup.8 d27, d22[6] \n" + "vext.8 d4, d24, d25, #4 \n" + "vext.8 d5, d26, d27, #4 \n" // f + "veor.8 q10, q2, q3 \n" // 0x7f ^ f + "vmull.u8 q11, d0, d20 \n" + "vmull.u8 q12, d1, d21 \n" + "vmull.u8 q13, d2, d4 \n" + "vmull.u8 q14, d3, d5 \n" + "vadd.i16 q11, q11, q13 \n" + "vadd.i16 q12, q12, q14 \n" + "vshrn.i16 d0, q11, #7 \n" + "vshrn.i16 d1, q12, #7 \n" + + MEMACCESS(0) + "vst1.32 {d0, d1}, [%0]! \n" // store pixels + "vadd.s32 q8, q8, q9 \n" + "subs %2, %2, #4 \n" // 4 processed per loop + "bgt 1b \n" + : "+r"(dst_argb), // %0 + "+r"(src_argb), // %1 + "+r"(dst_width), // %2 + "+r"(x), // %3 + "+r"(dx), // %4 + "+r"(tmp), // %5 + "+r"(src_tmp) // %6 + : + : "memory", "cc", "q0", "q1", "q2", "q3", "q8", "q9", + "q10", "q11", "q12", "q13", "q14", "q15" + ); +} + +#undef LOAD2_DATA32_LANE + +#endif // defined(__ARM_NEON__) && !defined(__aarch64__) + +#ifdef __cplusplus +} // extern "C" +} // namespace libyuv +#endif diff --git a/media/libaom/src/third_party/libyuv/source/scale_neon64.cc b/media/libaom/src/third_party/libyuv/source/scale_neon64.cc new file mode 100644 index 000000000..1d5519357 --- /dev/null +++ b/media/libaom/src/third_party/libyuv/source/scale_neon64.cc @@ -0,0 +1,1042 @@ +/* + * Copyright 2014 The LibYuv Project Authors. All rights reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#include "libyuv/scale.h" +#include "libyuv/row.h" +#include "libyuv/scale_row.h" + +#ifdef __cplusplus +namespace libyuv { +extern "C" { +#endif + +// This module is for GCC Neon armv8 64 bit. +#if !defined(LIBYUV_DISABLE_NEON) && defined(__aarch64__) + +// Read 32x1 throw away even pixels, and write 16x1. +void ScaleRowDown2_NEON(const uint8* src_ptr, ptrdiff_t src_stride, + uint8* dst, int dst_width) { + asm volatile ( + "1: \n" + // load even pixels into v0, odd into v1 + MEMACCESS(0) + "ld2 {v0.16b,v1.16b}, [%0], #32 \n" + "subs %w2, %w2, #16 \n" // 16 processed per loop + MEMACCESS(1) + "st1 {v1.16b}, [%1], #16 \n" // store odd pixels + "b.gt 1b \n" + : "+r"(src_ptr), // %0 + "+r"(dst), // %1 + "+r"(dst_width) // %2 + : + : "v0", "v1" // Clobber List + ); +} + +// Read 32x1 average down and write 16x1. +void ScaleRowDown2Linear_NEON(const uint8* src_ptr, ptrdiff_t src_stride, + uint8* dst, int dst_width) { + asm volatile ( + "1: \n" + MEMACCESS(0) + "ld1 {v0.16b,v1.16b}, [%0], #32 \n" // load pixels and post inc + "subs %w2, %w2, #16 \n" // 16 processed per loop + "uaddlp v0.8h, v0.16b \n" // add adjacent + "uaddlp v1.8h, v1.16b \n" + "rshrn v0.8b, v0.8h, #1 \n" // downshift, round and pack + "rshrn2 v0.16b, v1.8h, #1 \n" + MEMACCESS(1) + "st1 {v0.16b}, [%1], #16 \n" + "b.gt 1b \n" + : "+r"(src_ptr), // %0 + "+r"(dst), // %1 + "+r"(dst_width) // %2 + : + : "v0", "v1" // Clobber List + ); +} + +// Read 32x2 average down and write 16x1. +void ScaleRowDown2Box_NEON(const uint8* src_ptr, ptrdiff_t src_stride, + uint8* dst, int dst_width) { + asm volatile ( + // change the stride to row 2 pointer + "add %1, %1, %0 \n" + "1: \n" + MEMACCESS(0) + "ld1 {v0.16b,v1.16b}, [%0], #32 \n" // load row 1 and post inc + MEMACCESS(1) + "ld1 {v2.16b, v3.16b}, [%1], #32 \n" // load row 2 and post inc + "subs %w3, %w3, #16 \n" // 16 processed per loop + "uaddlp v0.8h, v0.16b \n" // row 1 add adjacent + "uaddlp v1.8h, v1.16b \n" + "uadalp v0.8h, v2.16b \n" // row 2 add adjacent + row1 + "uadalp v1.8h, v3.16b \n" + "rshrn v0.8b, v0.8h, #2 \n" // downshift, round and pack + "rshrn2 v0.16b, v1.8h, #2 \n" + MEMACCESS(2) + "st1 {v0.16b}, [%2], #16 \n" + "b.gt 1b \n" + : "+r"(src_ptr), // %0 + "+r"(src_stride), // %1 + "+r"(dst), // %2 + "+r"(dst_width) // %3 + : + : "v0", "v1", "v2", "v3" // Clobber List + ); +} + +void ScaleRowDown4_NEON(const uint8* src_ptr, ptrdiff_t src_stride, + uint8* dst_ptr, int dst_width) { + asm volatile ( + "1: \n" + MEMACCESS(0) + "ld4 {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n" // src line 0 + "subs %w2, %w2, #8 \n" // 8 processed per loop + MEMACCESS(1) + "st1 {v2.8b}, [%1], #8 \n" + "b.gt 1b \n" + : "+r"(src_ptr), // %0 + "+r"(dst_ptr), // %1 + "+r"(dst_width) // %2 + : + : "v0", "v1", "v2", "v3", "memory", "cc" + ); +} + +void ScaleRowDown4Box_NEON(const uint8* src_ptr, ptrdiff_t src_stride, + uint8* dst_ptr, int dst_width) { + const uint8* src_ptr1 = src_ptr + src_stride; + const uint8* src_ptr2 = src_ptr + src_stride * 2; + const uint8* src_ptr3 = src_ptr + src_stride * 3; +asm volatile ( + "1: \n" + MEMACCESS(0) + "ld1 {v0.16b}, [%0], #16 \n" // load up 16x4 + MEMACCESS(3) + "ld1 {v1.16b}, [%2], #16 \n" + MEMACCESS(4) + "ld1 {v2.16b}, [%3], #16 \n" + MEMACCESS(5) + "ld1 {v3.16b}, [%4], #16 \n" + "subs %w5, %w5, #4 \n" + "uaddlp v0.8h, v0.16b \n" + "uadalp v0.8h, v1.16b \n" + "uadalp v0.8h, v2.16b \n" + "uadalp v0.8h, v3.16b \n" + "addp v0.8h, v0.8h, v0.8h \n" + "rshrn v0.8b, v0.8h, #4 \n" // divide by 16 w/rounding + MEMACCESS(1) + "st1 {v0.s}[0], [%1], #4 \n" + "b.gt 1b \n" + : "+r"(src_ptr), // %0 + "+r"(dst_ptr), // %1 + "+r"(src_ptr1), // %2 + "+r"(src_ptr2), // %3 + "+r"(src_ptr3), // %4 + "+r"(dst_width) // %5 + : + : "v0", "v1", "v2", "v3", "memory", "cc" + ); +} + +// Down scale from 4 to 3 pixels. Use the neon multilane read/write +// to load up the every 4th pixel into a 4 different registers. +// Point samples 32 pixels to 24 pixels. +void ScaleRowDown34_NEON(const uint8* src_ptr, + ptrdiff_t src_stride, + uint8* dst_ptr, int dst_width) { + asm volatile ( + "1: \n" + MEMACCESS(0) + "ld4 {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n" // src line 0 + "subs %w2, %w2, #24 \n" + "orr v2.16b, v3.16b, v3.16b \n" // order v0, v1, v2 + MEMACCESS(1) + "st3 {v0.8b,v1.8b,v2.8b}, [%1], #24 \n" + "b.gt 1b \n" + : "+r"(src_ptr), // %0 + "+r"(dst_ptr), // %1 + "+r"(dst_width) // %2 + : + : "v0", "v1", "v2", "v3", "memory", "cc" + ); +} + +void ScaleRowDown34_0_Box_NEON(const uint8* src_ptr, + ptrdiff_t src_stride, + uint8* dst_ptr, int dst_width) { + asm volatile ( + "movi v20.8b, #3 \n" + "add %3, %3, %0 \n" + "1: \n" + MEMACCESS(0) + "ld4 {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n" // src line 0 + MEMACCESS(3) + "ld4 {v4.8b,v5.8b,v6.8b,v7.8b}, [%3], #32 \n" // src line 1 + "subs %w2, %w2, #24 \n" + + // filter src line 0 with src line 1 + // expand chars to shorts to allow for room + // when adding lines together + "ushll v16.8h, v4.8b, #0 \n" + "ushll v17.8h, v5.8b, #0 \n" + "ushll v18.8h, v6.8b, #0 \n" + "ushll v19.8h, v7.8b, #0 \n" + + // 3 * line_0 + line_1 + "umlal v16.8h, v0.8b, v20.8b \n" + "umlal v17.8h, v1.8b, v20.8b \n" + "umlal v18.8h, v2.8b, v20.8b \n" + "umlal v19.8h, v3.8b, v20.8b \n" + + // (3 * line_0 + line_1) >> 2 + "uqrshrn v0.8b, v16.8h, #2 \n" + "uqrshrn v1.8b, v17.8h, #2 \n" + "uqrshrn v2.8b, v18.8h, #2 \n" + "uqrshrn v3.8b, v19.8h, #2 \n" + + // a0 = (src[0] * 3 + s[1] * 1) >> 2 + "ushll v16.8h, v1.8b, #0 \n" + "umlal v16.8h, v0.8b, v20.8b \n" + "uqrshrn v0.8b, v16.8h, #2 \n" + + // a1 = (src[1] * 1 + s[2] * 1) >> 1 + "urhadd v1.8b, v1.8b, v2.8b \n" + + // a2 = (src[2] * 1 + s[3] * 3) >> 2 + "ushll v16.8h, v2.8b, #0 \n" + "umlal v16.8h, v3.8b, v20.8b \n" + "uqrshrn v2.8b, v16.8h, #2 \n" + + MEMACCESS(1) + "st3 {v0.8b,v1.8b,v2.8b}, [%1], #24 \n" + + "b.gt 1b \n" + : "+r"(src_ptr), // %0 + "+r"(dst_ptr), // %1 + "+r"(dst_width), // %2 + "+r"(src_stride) // %3 + : + : "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16", "v17", "v18", "v19", + "v20", "memory", "cc" + ); +} + +void ScaleRowDown34_1_Box_NEON(const uint8* src_ptr, + ptrdiff_t src_stride, + uint8* dst_ptr, int dst_width) { + asm volatile ( + "movi v20.8b, #3 \n" + "add %3, %3, %0 \n" + "1: \n" + MEMACCESS(0) + "ld4 {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n" // src line 0 + MEMACCESS(3) + "ld4 {v4.8b,v5.8b,v6.8b,v7.8b}, [%3], #32 \n" // src line 1 + "subs %w2, %w2, #24 \n" + // average src line 0 with src line 1 + "urhadd v0.8b, v0.8b, v4.8b \n" + "urhadd v1.8b, v1.8b, v5.8b \n" + "urhadd v2.8b, v2.8b, v6.8b \n" + "urhadd v3.8b, v3.8b, v7.8b \n" + + // a0 = (src[0] * 3 + s[1] * 1) >> 2 + "ushll v4.8h, v1.8b, #0 \n" + "umlal v4.8h, v0.8b, v20.8b \n" + "uqrshrn v0.8b, v4.8h, #2 \n" + + // a1 = (src[1] * 1 + s[2] * 1) >> 1 + "urhadd v1.8b, v1.8b, v2.8b \n" + + // a2 = (src[2] * 1 + s[3] * 3) >> 2 + "ushll v4.8h, v2.8b, #0 \n" + "umlal v4.8h, v3.8b, v20.8b \n" + "uqrshrn v2.8b, v4.8h, #2 \n" + + MEMACCESS(1) + "st3 {v0.8b,v1.8b,v2.8b}, [%1], #24 \n" + "b.gt 1b \n" + : "+r"(src_ptr), // %0 + "+r"(dst_ptr), // %1 + "+r"(dst_width), // %2 + "+r"(src_stride) // %3 + : + : "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20", "memory", "cc" + ); +} + +static uvec8 kShuf38 = + { 0, 3, 6, 8, 11, 14, 16, 19, 22, 24, 27, 30, 0, 0, 0, 0 }; +static uvec8 kShuf38_2 = + { 0, 16, 32, 2, 18, 33, 4, 20, 34, 6, 22, 35, 0, 0, 0, 0 }; +static vec16 kMult38_Div6 = + { 65536 / 12, 65536 / 12, 65536 / 12, 65536 / 12, + 65536 / 12, 65536 / 12, 65536 / 12, 65536 / 12 }; +static vec16 kMult38_Div9 = + { 65536 / 18, 65536 / 18, 65536 / 18, 65536 / 18, + 65536 / 18, 65536 / 18, 65536 / 18, 65536 / 18 }; + +// 32 -> 12 +void ScaleRowDown38_NEON(const uint8* src_ptr, + ptrdiff_t src_stride, + uint8* dst_ptr, int dst_width) { + asm volatile ( + MEMACCESS(3) + "ld1 {v3.16b}, [%3] \n" + "1: \n" + MEMACCESS(0) + "ld1 {v0.16b,v1.16b}, [%0], #32 \n" + "subs %w2, %w2, #12 \n" + "tbl v2.16b, {v0.16b,v1.16b}, v3.16b \n" + MEMACCESS(1) + "st1 {v2.8b}, [%1], #8 \n" + MEMACCESS(1) + "st1 {v2.s}[2], [%1], #4 \n" + "b.gt 1b \n" + : "+r"(src_ptr), // %0 + "+r"(dst_ptr), // %1 + "+r"(dst_width) // %2 + : "r"(&kShuf38) // %3 + : "v0", "v1", "v2", "v3", "memory", "cc" + ); +} + +// 32x3 -> 12x1 +void OMITFP ScaleRowDown38_3_Box_NEON(const uint8* src_ptr, + ptrdiff_t src_stride, + uint8* dst_ptr, int dst_width) { + const uint8* src_ptr1 = src_ptr + src_stride * 2; + ptrdiff_t tmp_src_stride = src_stride; + + asm volatile ( + MEMACCESS(5) + "ld1 {v29.8h}, [%5] \n" + MEMACCESS(6) + "ld1 {v30.16b}, [%6] \n" + MEMACCESS(7) + "ld1 {v31.8h}, [%7] \n" + "add %2, %2, %0 \n" + "1: \n" + + // 00 40 01 41 02 42 03 43 + // 10 50 11 51 12 52 13 53 + // 20 60 21 61 22 62 23 63 + // 30 70 31 71 32 72 33 73 + MEMACCESS(0) + "ld4 {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n" + MEMACCESS(3) + "ld4 {v4.8b,v5.8b,v6.8b,v7.8b}, [%2], #32 \n" + MEMACCESS(4) + "ld4 {v16.8b,v17.8b,v18.8b,v19.8b}, [%3], #32 \n" + "subs %w4, %w4, #12 \n" + + // Shuffle the input data around to get align the data + // so adjacent data can be added. 0,1 - 2,3 - 4,5 - 6,7 + // 00 10 01 11 02 12 03 13 + // 40 50 41 51 42 52 43 53 + "trn1 v20.8b, v0.8b, v1.8b \n" + "trn2 v21.8b, v0.8b, v1.8b \n" + "trn1 v22.8b, v4.8b, v5.8b \n" + "trn2 v23.8b, v4.8b, v5.8b \n" + "trn1 v24.8b, v16.8b, v17.8b \n" + "trn2 v25.8b, v16.8b, v17.8b \n" + + // 20 30 21 31 22 32 23 33 + // 60 70 61 71 62 72 63 73 + "trn1 v0.8b, v2.8b, v3.8b \n" + "trn2 v1.8b, v2.8b, v3.8b \n" + "trn1 v4.8b, v6.8b, v7.8b \n" + "trn2 v5.8b, v6.8b, v7.8b \n" + "trn1 v16.8b, v18.8b, v19.8b \n" + "trn2 v17.8b, v18.8b, v19.8b \n" + + // 00+10 01+11 02+12 03+13 + // 40+50 41+51 42+52 43+53 + "uaddlp v20.4h, v20.8b \n" + "uaddlp v21.4h, v21.8b \n" + "uaddlp v22.4h, v22.8b \n" + "uaddlp v23.4h, v23.8b \n" + "uaddlp v24.4h, v24.8b \n" + "uaddlp v25.4h, v25.8b \n" + + // 60+70 61+71 62+72 63+73 + "uaddlp v1.4h, v1.8b \n" + "uaddlp v5.4h, v5.8b \n" + "uaddlp v17.4h, v17.8b \n" + + // combine source lines + "add v20.4h, v20.4h, v22.4h \n" + "add v21.4h, v21.4h, v23.4h \n" + "add v20.4h, v20.4h, v24.4h \n" + "add v21.4h, v21.4h, v25.4h \n" + "add v2.4h, v1.4h, v5.4h \n" + "add v2.4h, v2.4h, v17.4h \n" + + // dst_ptr[3] = (s[6 + st * 0] + s[7 + st * 0] + // + s[6 + st * 1] + s[7 + st * 1] + // + s[6 + st * 2] + s[7 + st * 2]) / 6 + "sqrdmulh v2.8h, v2.8h, v29.8h \n" + "xtn v2.8b, v2.8h \n" + + // Shuffle 2,3 reg around so that 2 can be added to the + // 0,1 reg and 3 can be added to the 4,5 reg. This + // requires expanding from u8 to u16 as the 0,1 and 4,5 + // registers are already expanded. Then do transposes + // to get aligned. + // xx 20 xx 30 xx 21 xx 31 xx 22 xx 32 xx 23 xx 33 + "ushll v16.8h, v16.8b, #0 \n" + "uaddl v0.8h, v0.8b, v4.8b \n" + + // combine source lines + "add v0.8h, v0.8h, v16.8h \n" + + // xx 20 xx 21 xx 22 xx 23 + // xx 30 xx 31 xx 32 xx 33 + "trn1 v1.8h, v0.8h, v0.8h \n" + "trn2 v4.8h, v0.8h, v0.8h \n" + "xtn v0.4h, v1.4s \n" + "xtn v4.4h, v4.4s \n" + + // 0+1+2, 3+4+5 + "add v20.8h, v20.8h, v0.8h \n" + "add v21.8h, v21.8h, v4.8h \n" + + // Need to divide, but can't downshift as the the value + // isn't a power of 2. So multiply by 65536 / n + // and take the upper 16 bits. + "sqrdmulh v0.8h, v20.8h, v31.8h \n" + "sqrdmulh v1.8h, v21.8h, v31.8h \n" + + // Align for table lookup, vtbl requires registers to + // be adjacent + "tbl v3.16b, {v0.16b, v1.16b, v2.16b}, v30.16b \n" + + MEMACCESS(1) + "st1 {v3.8b}, [%1], #8 \n" + MEMACCESS(1) + "st1 {v3.s}[2], [%1], #4 \n" + "b.gt 1b \n" + : "+r"(src_ptr), // %0 + "+r"(dst_ptr), // %1 + "+r"(tmp_src_stride), // %2 + "+r"(src_ptr1), // %3 + "+r"(dst_width) // %4 + : "r"(&kMult38_Div6), // %5 + "r"(&kShuf38_2), // %6 + "r"(&kMult38_Div9) // %7 + : "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16", "v17", + "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v29", + "v30", "v31", "memory", "cc" + ); +} + +// 32x2 -> 12x1 +void ScaleRowDown38_2_Box_NEON(const uint8* src_ptr, + ptrdiff_t src_stride, + uint8* dst_ptr, int dst_width) { + // TODO(fbarchard): use src_stride directly for clang 3.5+. + ptrdiff_t tmp_src_stride = src_stride; + asm volatile ( + MEMACCESS(4) + "ld1 {v30.8h}, [%4] \n" + MEMACCESS(5) + "ld1 {v31.16b}, [%5] \n" + "add %2, %2, %0 \n" + "1: \n" + + // 00 40 01 41 02 42 03 43 + // 10 50 11 51 12 52 13 53 + // 20 60 21 61 22 62 23 63 + // 30 70 31 71 32 72 33 73 + MEMACCESS(0) + "ld4 {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n" + MEMACCESS(3) + "ld4 {v4.8b,v5.8b,v6.8b,v7.8b}, [%2], #32 \n" + "subs %w3, %w3, #12 \n" + + // Shuffle the input data around to get align the data + // so adjacent data can be added. 0,1 - 2,3 - 4,5 - 6,7 + // 00 10 01 11 02 12 03 13 + // 40 50 41 51 42 52 43 53 + "trn1 v16.8b, v0.8b, v1.8b \n" + "trn2 v17.8b, v0.8b, v1.8b \n" + "trn1 v18.8b, v4.8b, v5.8b \n" + "trn2 v19.8b, v4.8b, v5.8b \n" + + // 20 30 21 31 22 32 23 33 + // 60 70 61 71 62 72 63 73 + "trn1 v0.8b, v2.8b, v3.8b \n" + "trn2 v1.8b, v2.8b, v3.8b \n" + "trn1 v4.8b, v6.8b, v7.8b \n" + "trn2 v5.8b, v6.8b, v7.8b \n" + + // 00+10 01+11 02+12 03+13 + // 40+50 41+51 42+52 43+53 + "uaddlp v16.4h, v16.8b \n" + "uaddlp v17.4h, v17.8b \n" + "uaddlp v18.4h, v18.8b \n" + "uaddlp v19.4h, v19.8b \n" + + // 60+70 61+71 62+72 63+73 + "uaddlp v1.4h, v1.8b \n" + "uaddlp v5.4h, v5.8b \n" + + // combine source lines + "add v16.4h, v16.4h, v18.4h \n" + "add v17.4h, v17.4h, v19.4h \n" + "add v2.4h, v1.4h, v5.4h \n" + + // dst_ptr[3] = (s[6] + s[7] + s[6+st] + s[7+st]) / 4 + "uqrshrn v2.8b, v2.8h, #2 \n" + + // Shuffle 2,3 reg around so that 2 can be added to the + // 0,1 reg and 3 can be added to the 4,5 reg. This + // requires expanding from u8 to u16 as the 0,1 and 4,5 + // registers are already expanded. Then do transposes + // to get aligned. + // xx 20 xx 30 xx 21 xx 31 xx 22 xx 32 xx 23 xx 33 + + // combine source lines + "uaddl v0.8h, v0.8b, v4.8b \n" + + // xx 20 xx 21 xx 22 xx 23 + // xx 30 xx 31 xx 32 xx 33 + "trn1 v1.8h, v0.8h, v0.8h \n" + "trn2 v4.8h, v0.8h, v0.8h \n" + "xtn v0.4h, v1.4s \n" + "xtn v4.4h, v4.4s \n" + + // 0+1+2, 3+4+5 + "add v16.8h, v16.8h, v0.8h \n" + "add v17.8h, v17.8h, v4.8h \n" + + // Need to divide, but can't downshift as the the value + // isn't a power of 2. So multiply by 65536 / n + // and take the upper 16 bits. + "sqrdmulh v0.8h, v16.8h, v30.8h \n" + "sqrdmulh v1.8h, v17.8h, v30.8h \n" + + // Align for table lookup, vtbl requires registers to + // be adjacent + + "tbl v3.16b, {v0.16b, v1.16b, v2.16b}, v31.16b \n" + + MEMACCESS(1) + "st1 {v3.8b}, [%1], #8 \n" + MEMACCESS(1) + "st1 {v3.s}[2], [%1], #4 \n" + "b.gt 1b \n" + : "+r"(src_ptr), // %0 + "+r"(dst_ptr), // %1 + "+r"(tmp_src_stride), // %2 + "+r"(dst_width) // %3 + : "r"(&kMult38_Div6), // %4 + "r"(&kShuf38_2) // %5 + : "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16", "v17", + "v18", "v19", "v30", "v31", "memory", "cc" + ); +} + +void ScaleAddRows_NEON(const uint8* src_ptr, ptrdiff_t src_stride, + uint16* dst_ptr, int src_width, int src_height) { + const uint8* src_tmp = NULL; + asm volatile ( + "1: \n" + "mov %0, %1 \n" + "mov w12, %w5 \n" + "eor v2.16b, v2.16b, v2.16b \n" + "eor v3.16b, v3.16b, v3.16b \n" + "2: \n" + // load 16 pixels into q0 + MEMACCESS(0) + "ld1 {v0.16b}, [%0], %3 \n" + "uaddw2 v3.8h, v3.8h, v0.16b \n" + "uaddw v2.8h, v2.8h, v0.8b \n" + "subs w12, w12, #1 \n" + "b.gt 2b \n" + MEMACCESS(2) + "st1 {v2.8h, v3.8h}, [%2], #32 \n" // store pixels + "add %1, %1, #16 \n" + "subs %w4, %w4, #16 \n" // 16 processed per loop + "b.gt 1b \n" + : "+r"(src_tmp), // %0 + "+r"(src_ptr), // %1 + "+r"(dst_ptr), // %2 + "+r"(src_stride), // %3 + "+r"(src_width), // %4 + "+r"(src_height) // %5 + : + : "memory", "cc", "w12", "v0", "v1", "v2", "v3" // Clobber List + ); +} + +// TODO(Yang Zhang): Investigate less load instructions for +// the x/dx stepping +#define LOAD2_DATA8_LANE(n) \ + "lsr %5, %3, #16 \n" \ + "add %6, %1, %5 \n" \ + "add %3, %3, %4 \n" \ + MEMACCESS(6) \ + "ld2 {v4.b, v5.b}["#n"], [%6] \n" + +void ScaleFilterCols_NEON(uint8* dst_ptr, const uint8* src_ptr, + int dst_width, int x, int dx) { + int dx_offset[4] = {0, 1, 2, 3}; + int* tmp = dx_offset; + const uint8* src_tmp = src_ptr; + int64 dst_width64 = (int64) dst_width; // Work around ios 64 bit warning. + int64 x64 = (int64) x; + int64 dx64 = (int64) dx; + asm volatile ( + "dup v0.4s, %w3 \n" // x + "dup v1.4s, %w4 \n" // dx + "ld1 {v2.4s}, [%5] \n" // 0 1 2 3 + "shl v3.4s, v1.4s, #2 \n" // 4 * dx + "mul v1.4s, v1.4s, v2.4s \n" + // x , x + 1 * dx, x + 2 * dx, x + 3 * dx + "add v1.4s, v1.4s, v0.4s \n" + // x + 4 * dx, x + 5 * dx, x + 6 * dx, x + 7 * dx + "add v2.4s, v1.4s, v3.4s \n" + "shl v0.4s, v3.4s, #1 \n" // 8 * dx + "1: \n" + LOAD2_DATA8_LANE(0) + LOAD2_DATA8_LANE(1) + LOAD2_DATA8_LANE(2) + LOAD2_DATA8_LANE(3) + LOAD2_DATA8_LANE(4) + LOAD2_DATA8_LANE(5) + LOAD2_DATA8_LANE(6) + LOAD2_DATA8_LANE(7) + "mov v6.16b, v1.16b \n" + "mov v7.16b, v2.16b \n" + "uzp1 v6.8h, v6.8h, v7.8h \n" + "ushll v4.8h, v4.8b, #0 \n" + "ushll v5.8h, v5.8b, #0 \n" + "ssubl v16.4s, v5.4h, v4.4h \n" + "ssubl2 v17.4s, v5.8h, v4.8h \n" + "ushll v7.4s, v6.4h, #0 \n" + "ushll2 v6.4s, v6.8h, #0 \n" + "mul v16.4s, v16.4s, v7.4s \n" + "mul v17.4s, v17.4s, v6.4s \n" + "shrn v6.4h, v16.4s, #16 \n" + "shrn2 v6.8h, v17.4s, #16 \n" + "add v4.8h, v4.8h, v6.8h \n" + "xtn v4.8b, v4.8h \n" + + MEMACCESS(0) + "st1 {v4.8b}, [%0], #8 \n" // store pixels + "add v1.4s, v1.4s, v0.4s \n" + "add v2.4s, v2.4s, v0.4s \n" + "subs %w2, %w2, #8 \n" // 8 processed per loop + "b.gt 1b \n" + : "+r"(dst_ptr), // %0 + "+r"(src_ptr), // %1 + "+r"(dst_width64), // %2 + "+r"(x64), // %3 + "+r"(dx64), // %4 + "+r"(tmp), // %5 + "+r"(src_tmp) // %6 + : + : "memory", "cc", "v0", "v1", "v2", "v3", + "v4", "v5", "v6", "v7", "v16", "v17" + ); +} + +#undef LOAD2_DATA8_LANE + +// 16x2 -> 16x1 +void ScaleFilterRows_NEON(uint8* dst_ptr, + const uint8* src_ptr, ptrdiff_t src_stride, + int dst_width, int source_y_fraction) { + int y_fraction = 256 - source_y_fraction; + asm volatile ( + "cmp %w4, #0 \n" + "b.eq 100f \n" + "add %2, %2, %1 \n" + "cmp %w4, #64 \n" + "b.eq 75f \n" + "cmp %w4, #128 \n" + "b.eq 50f \n" + "cmp %w4, #192 \n" + "b.eq 25f \n" + + "dup v5.8b, %w4 \n" + "dup v4.8b, %w5 \n" + // General purpose row blend. + "1: \n" + MEMACCESS(1) + "ld1 {v0.16b}, [%1], #16 \n" + MEMACCESS(2) + "ld1 {v1.16b}, [%2], #16 \n" + "subs %w3, %w3, #16 \n" + "umull v6.8h, v0.8b, v4.8b \n" + "umull2 v7.8h, v0.16b, v4.16b \n" + "umlal v6.8h, v1.8b, v5.8b \n" + "umlal2 v7.8h, v1.16b, v5.16b \n" + "rshrn v0.8b, v6.8h, #8 \n" + "rshrn2 v0.16b, v7.8h, #8 \n" + MEMACCESS(0) + "st1 {v0.16b}, [%0], #16 \n" + "b.gt 1b \n" + "b 99f \n" + + // Blend 25 / 75. + "25: \n" + MEMACCESS(1) + "ld1 {v0.16b}, [%1], #16 \n" + MEMACCESS(2) + "ld1 {v1.16b}, [%2], #16 \n" + "subs %w3, %w3, #16 \n" + "urhadd v0.16b, v0.16b, v1.16b \n" + "urhadd v0.16b, v0.16b, v1.16b \n" + MEMACCESS(0) + "st1 {v0.16b}, [%0], #16 \n" + "b.gt 25b \n" + "b 99f \n" + + // Blend 50 / 50. + "50: \n" + MEMACCESS(1) + "ld1 {v0.16b}, [%1], #16 \n" + MEMACCESS(2) + "ld1 {v1.16b}, [%2], #16 \n" + "subs %w3, %w3, #16 \n" + "urhadd v0.16b, v0.16b, v1.16b \n" + MEMACCESS(0) + "st1 {v0.16b}, [%0], #16 \n" + "b.gt 50b \n" + "b 99f \n" + + // Blend 75 / 25. + "75: \n" + MEMACCESS(1) + "ld1 {v1.16b}, [%1], #16 \n" + MEMACCESS(2) + "ld1 {v0.16b}, [%2], #16 \n" + "subs %w3, %w3, #16 \n" + "urhadd v0.16b, v0.16b, v1.16b \n" + "urhadd v0.16b, v0.16b, v1.16b \n" + MEMACCESS(0) + "st1 {v0.16b}, [%0], #16 \n" + "b.gt 75b \n" + "b 99f \n" + + // Blend 100 / 0 - Copy row unchanged. + "100: \n" + MEMACCESS(1) + "ld1 {v0.16b}, [%1], #16 \n" + "subs %w3, %w3, #16 \n" + MEMACCESS(0) + "st1 {v0.16b}, [%0], #16 \n" + "b.gt 100b \n" + + "99: \n" + MEMACCESS(0) + "st1 {v0.b}[15], [%0] \n" + : "+r"(dst_ptr), // %0 + "+r"(src_ptr), // %1 + "+r"(src_stride), // %2 + "+r"(dst_width), // %3 + "+r"(source_y_fraction),// %4 + "+r"(y_fraction) // %5 + : + : "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "memory", "cc" + ); +} + +void ScaleARGBRowDown2_NEON(const uint8* src_ptr, ptrdiff_t src_stride, + uint8* dst, int dst_width) { + asm volatile ( + "1: \n" + // load even pixels into q0, odd into q1 + MEMACCESS (0) + "ld2 {v0.4s, v1.4s}, [%0], #32 \n" + MEMACCESS (0) + "ld2 {v2.4s, v3.4s}, [%0], #32 \n" + "subs %w2, %w2, #8 \n" // 8 processed per loop + MEMACCESS (1) + "st1 {v1.16b}, [%1], #16 \n" // store odd pixels + MEMACCESS (1) + "st1 {v3.16b}, [%1], #16 \n" + "b.gt 1b \n" + : "+r" (src_ptr), // %0 + "+r" (dst), // %1 + "+r" (dst_width) // %2 + : + : "memory", "cc", "v0", "v1", "v2", "v3" // Clobber List + ); +} + +void ScaleARGBRowDown2Linear_NEON(const uint8* src_argb, ptrdiff_t src_stride, + uint8* dst_argb, int dst_width) { + asm volatile ( + "1: \n" + MEMACCESS (0) + // load 8 ARGB pixels. + "ld4 {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64 \n" + "subs %w2, %w2, #8 \n" // 8 processed per loop. + "uaddlp v0.8h, v0.16b \n" // B 16 bytes -> 8 shorts. + "uaddlp v1.8h, v1.16b \n" // G 16 bytes -> 8 shorts. + "uaddlp v2.8h, v2.16b \n" // R 16 bytes -> 8 shorts. + "uaddlp v3.8h, v3.16b \n" // A 16 bytes -> 8 shorts. + "rshrn v0.8b, v0.8h, #1 \n" // downshift, round and pack + "rshrn v1.8b, v1.8h, #1 \n" + "rshrn v2.8b, v2.8h, #1 \n" + "rshrn v3.8b, v3.8h, #1 \n" + MEMACCESS (1) + "st4 {v0.8b,v1.8b,v2.8b,v3.8b}, [%1], #32 \n" + "b.gt 1b \n" + : "+r"(src_argb), // %0 + "+r"(dst_argb), // %1 + "+r"(dst_width) // %2 + : + : "memory", "cc", "v0", "v1", "v2", "v3" // Clobber List + ); +} + +void ScaleARGBRowDown2Box_NEON(const uint8* src_ptr, ptrdiff_t src_stride, + uint8* dst, int dst_width) { + asm volatile ( + // change the stride to row 2 pointer + "add %1, %1, %0 \n" + "1: \n" + MEMACCESS (0) + "ld4 {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64 \n" // load 8 ARGB pixels. + "subs %w3, %w3, #8 \n" // 8 processed per loop. + "uaddlp v0.8h, v0.16b \n" // B 16 bytes -> 8 shorts. + "uaddlp v1.8h, v1.16b \n" // G 16 bytes -> 8 shorts. + "uaddlp v2.8h, v2.16b \n" // R 16 bytes -> 8 shorts. + "uaddlp v3.8h, v3.16b \n" // A 16 bytes -> 8 shorts. + MEMACCESS (1) + "ld4 {v16.16b,v17.16b,v18.16b,v19.16b}, [%1], #64 \n" // load 8 more ARGB pixels. + "uadalp v0.8h, v16.16b \n" // B 16 bytes -> 8 shorts. + "uadalp v1.8h, v17.16b \n" // G 16 bytes -> 8 shorts. + "uadalp v2.8h, v18.16b \n" // R 16 bytes -> 8 shorts. + "uadalp v3.8h, v19.16b \n" // A 16 bytes -> 8 shorts. + "rshrn v0.8b, v0.8h, #2 \n" // downshift, round and pack + "rshrn v1.8b, v1.8h, #2 \n" + "rshrn v2.8b, v2.8h, #2 \n" + "rshrn v3.8b, v3.8h, #2 \n" + MEMACCESS (2) + "st4 {v0.8b,v1.8b,v2.8b,v3.8b}, [%2], #32 \n" + "b.gt 1b \n" + : "+r" (src_ptr), // %0 + "+r" (src_stride), // %1 + "+r" (dst), // %2 + "+r" (dst_width) // %3 + : + : "memory", "cc", "v0", "v1", "v2", "v3", "v16", "v17", "v18", "v19" + ); +} + +// Reads 4 pixels at a time. +// Alignment requirement: src_argb 4 byte aligned. +void ScaleARGBRowDownEven_NEON(const uint8* src_argb, ptrdiff_t src_stride, + int src_stepx, uint8* dst_argb, int dst_width) { + asm volatile ( + "1: \n" + MEMACCESS(0) + "ld1 {v0.s}[0], [%0], %3 \n" + MEMACCESS(0) + "ld1 {v0.s}[1], [%0], %3 \n" + MEMACCESS(0) + "ld1 {v0.s}[2], [%0], %3 \n" + MEMACCESS(0) + "ld1 {v0.s}[3], [%0], %3 \n" + "subs %w2, %w2, #4 \n" // 4 pixels per loop. + MEMACCESS(1) + "st1 {v0.16b}, [%1], #16 \n" + "b.gt 1b \n" + : "+r"(src_argb), // %0 + "+r"(dst_argb), // %1 + "+r"(dst_width) // %2 + : "r"((int64)(src_stepx * 4)) // %3 + : "memory", "cc", "v0" + ); +} + +// Reads 4 pixels at a time. +// Alignment requirement: src_argb 4 byte aligned. +// TODO(Yang Zhang): Might be worth another optimization pass in future. +// It could be upgraded to 8 pixels at a time to start with. +void ScaleARGBRowDownEvenBox_NEON(const uint8* src_argb, ptrdiff_t src_stride, + int src_stepx, + uint8* dst_argb, int dst_width) { + asm volatile ( + "add %1, %1, %0 \n" + "1: \n" + MEMACCESS(0) + "ld1 {v0.8b}, [%0], %4 \n" // Read 4 2x2 blocks -> 2x1 + MEMACCESS(1) + "ld1 {v1.8b}, [%1], %4 \n" + MEMACCESS(0) + "ld1 {v2.8b}, [%0], %4 \n" + MEMACCESS(1) + "ld1 {v3.8b}, [%1], %4 \n" + MEMACCESS(0) + "ld1 {v4.8b}, [%0], %4 \n" + MEMACCESS(1) + "ld1 {v5.8b}, [%1], %4 \n" + MEMACCESS(0) + "ld1 {v6.8b}, [%0], %4 \n" + MEMACCESS(1) + "ld1 {v7.8b}, [%1], %4 \n" + "uaddl v0.8h, v0.8b, v1.8b \n" + "uaddl v2.8h, v2.8b, v3.8b \n" + "uaddl v4.8h, v4.8b, v5.8b \n" + "uaddl v6.8h, v6.8b, v7.8b \n" + "mov v16.d[1], v0.d[1] \n" // ab_cd -> ac_bd + "mov v0.d[1], v2.d[0] \n" + "mov v2.d[0], v16.d[1] \n" + "mov v16.d[1], v4.d[1] \n" // ef_gh -> eg_fh + "mov v4.d[1], v6.d[0] \n" + "mov v6.d[0], v16.d[1] \n" + "add v0.8h, v0.8h, v2.8h \n" // (a+b)_(c+d) + "add v4.8h, v4.8h, v6.8h \n" // (e+f)_(g+h) + "rshrn v0.8b, v0.8h, #2 \n" // first 2 pixels. + "rshrn2 v0.16b, v4.8h, #2 \n" // next 2 pixels. + "subs %w3, %w3, #4 \n" // 4 pixels per loop. + MEMACCESS(2) + "st1 {v0.16b}, [%2], #16 \n" + "b.gt 1b \n" + : "+r"(src_argb), // %0 + "+r"(src_stride), // %1 + "+r"(dst_argb), // %2 + "+r"(dst_width) // %3 + : "r"((int64)(src_stepx * 4)) // %4 + : "memory", "cc", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16" + ); +} + +// TODO(Yang Zhang): Investigate less load instructions for +// the x/dx stepping +#define LOAD1_DATA32_LANE(vn, n) \ + "lsr %5, %3, #16 \n" \ + "add %6, %1, %5, lsl #2 \n" \ + "add %3, %3, %4 \n" \ + MEMACCESS(6) \ + "ld1 {"#vn".s}["#n"], [%6] \n" + +void ScaleARGBCols_NEON(uint8* dst_argb, const uint8* src_argb, + int dst_width, int x, int dx) { + const uint8* src_tmp = src_argb; + int64 dst_width64 = (int64) dst_width; // Work around ios 64 bit warning. + int64 x64 = (int64) x; + int64 dx64 = (int64) dx; + int64 tmp64 = 0; + asm volatile ( + "1: \n" + LOAD1_DATA32_LANE(v0, 0) + LOAD1_DATA32_LANE(v0, 1) + LOAD1_DATA32_LANE(v0, 2) + LOAD1_DATA32_LANE(v0, 3) + LOAD1_DATA32_LANE(v1, 0) + LOAD1_DATA32_LANE(v1, 1) + LOAD1_DATA32_LANE(v1, 2) + LOAD1_DATA32_LANE(v1, 3) + + MEMACCESS(0) + "st1 {v0.4s, v1.4s}, [%0], #32 \n" // store pixels + "subs %w2, %w2, #8 \n" // 8 processed per loop + "b.gt 1b \n" + : "+r"(dst_argb), // %0 + "+r"(src_argb), // %1 + "+r"(dst_width64), // %2 + "+r"(x64), // %3 + "+r"(dx64), // %4 + "+r"(tmp64), // %5 + "+r"(src_tmp) // %6 + : + : "memory", "cc", "v0", "v1" + ); +} + +#undef LOAD1_DATA32_LANE + +// TODO(Yang Zhang): Investigate less load instructions for +// the x/dx stepping +#define LOAD2_DATA32_LANE(vn1, vn2, n) \ + "lsr %5, %3, #16 \n" \ + "add %6, %1, %5, lsl #2 \n" \ + "add %3, %3, %4 \n" \ + MEMACCESS(6) \ + "ld2 {"#vn1".s, "#vn2".s}["#n"], [%6] \n" + +void ScaleARGBFilterCols_NEON(uint8* dst_argb, const uint8* src_argb, + int dst_width, int x, int dx) { + int dx_offset[4] = {0, 1, 2, 3}; + int* tmp = dx_offset; + const uint8* src_tmp = src_argb; + int64 dst_width64 = (int64) dst_width; // Work around ios 64 bit warning. + int64 x64 = (int64) x; + int64 dx64 = (int64) dx; + asm volatile ( + "dup v0.4s, %w3 \n" // x + "dup v1.4s, %w4 \n" // dx + "ld1 {v2.4s}, [%5] \n" // 0 1 2 3 + "shl v6.4s, v1.4s, #2 \n" // 4 * dx + "mul v1.4s, v1.4s, v2.4s \n" + "movi v3.16b, #0x7f \n" // 0x7F + "movi v4.8h, #0x7f \n" // 0x7F + // x , x + 1 * dx, x + 2 * dx, x + 3 * dx + "add v5.4s, v1.4s, v0.4s \n" + "1: \n" + // d0, d1: a + // d2, d3: b + LOAD2_DATA32_LANE(v0, v1, 0) + LOAD2_DATA32_LANE(v0, v1, 1) + LOAD2_DATA32_LANE(v0, v1, 2) + LOAD2_DATA32_LANE(v0, v1, 3) + "shrn v2.4h, v5.4s, #9 \n" + "and v2.8b, v2.8b, v4.8b \n" + "dup v16.8b, v2.b[0] \n" + "dup v17.8b, v2.b[2] \n" + "dup v18.8b, v2.b[4] \n" + "dup v19.8b, v2.b[6] \n" + "ext v2.8b, v16.8b, v17.8b, #4 \n" + "ext v17.8b, v18.8b, v19.8b, #4 \n" + "ins v2.d[1], v17.d[0] \n" // f + "eor v7.16b, v2.16b, v3.16b \n" // 0x7f ^ f + "umull v16.8h, v0.8b, v7.8b \n" + "umull2 v17.8h, v0.16b, v7.16b \n" + "umull v18.8h, v1.8b, v2.8b \n" + "umull2 v19.8h, v1.16b, v2.16b \n" + "add v16.8h, v16.8h, v18.8h \n" + "add v17.8h, v17.8h, v19.8h \n" + "shrn v0.8b, v16.8h, #7 \n" + "shrn2 v0.16b, v17.8h, #7 \n" + + MEMACCESS(0) + "st1 {v0.4s}, [%0], #16 \n" // store pixels + "add v5.4s, v5.4s, v6.4s \n" + "subs %w2, %w2, #4 \n" // 4 processed per loop + "b.gt 1b \n" + : "+r"(dst_argb), // %0 + "+r"(src_argb), // %1 + "+r"(dst_width64), // %2 + "+r"(x64), // %3 + "+r"(dx64), // %4 + "+r"(tmp), // %5 + "+r"(src_tmp) // %6 + : + : "memory", "cc", "v0", "v1", "v2", "v3", "v4", "v5", + "v6", "v7", "v16", "v17", "v18", "v19" + ); +} + +#undef LOAD2_DATA32_LANE + +#endif // !defined(LIBYUV_DISABLE_NEON) && defined(__aarch64__) + +#ifdef __cplusplus +} // extern "C" +} // namespace libyuv +#endif diff --git a/media/libaom/src/third_party/libyuv/source/scale_win.cc b/media/libaom/src/third_party/libyuv/source/scale_win.cc new file mode 100644 index 000000000..c3896ebad --- /dev/null +++ b/media/libaom/src/third_party/libyuv/source/scale_win.cc @@ -0,0 +1,1354 @@ +/* + * Copyright 2013 The LibYuv Project Authors. All rights reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#include "libyuv/row.h" +#include "libyuv/scale_row.h" + +#ifdef __cplusplus +namespace libyuv { +extern "C" { +#endif + +// This module is for Visual C x86. +#if !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86) && \ + defined(_MSC_VER) && !defined(__clang__) + +// Offsets for source bytes 0 to 9 +static uvec8 kShuf0 = + { 0, 1, 3, 4, 5, 7, 8, 9, 128, 128, 128, 128, 128, 128, 128, 128 }; + +// Offsets for source bytes 11 to 20 with 8 subtracted = 3 to 12. +static uvec8 kShuf1 = + { 3, 4, 5, 7, 8, 9, 11, 12, 128, 128, 128, 128, 128, 128, 128, 128 }; + +// Offsets for source bytes 21 to 31 with 16 subtracted = 5 to 31. +static uvec8 kShuf2 = + { 5, 7, 8, 9, 11, 12, 13, 15, 128, 128, 128, 128, 128, 128, 128, 128 }; + +// Offsets for source bytes 0 to 10 +static uvec8 kShuf01 = + { 0, 1, 1, 2, 2, 3, 4, 5, 5, 6, 6, 7, 8, 9, 9, 10 }; + +// Offsets for source bytes 10 to 21 with 8 subtracted = 3 to 13. +static uvec8 kShuf11 = + { 2, 3, 4, 5, 5, 6, 6, 7, 8, 9, 9, 10, 10, 11, 12, 13 }; + +// Offsets for source bytes 21 to 31 with 16 subtracted = 5 to 31. +static uvec8 kShuf21 = + { 5, 6, 6, 7, 8, 9, 9, 10, 10, 11, 12, 13, 13, 14, 14, 15 }; + +// Coefficients for source bytes 0 to 10 +static uvec8 kMadd01 = + { 3, 1, 2, 2, 1, 3, 3, 1, 2, 2, 1, 3, 3, 1, 2, 2 }; + +// Coefficients for source bytes 10 to 21 +static uvec8 kMadd11 = + { 1, 3, 3, 1, 2, 2, 1, 3, 3, 1, 2, 2, 1, 3, 3, 1 }; + +// Coefficients for source bytes 21 to 31 +static uvec8 kMadd21 = + { 2, 2, 1, 3, 3, 1, 2, 2, 1, 3, 3, 1, 2, 2, 1, 3 }; + +// Coefficients for source bytes 21 to 31 +static vec16 kRound34 = + { 2, 2, 2, 2, 2, 2, 2, 2 }; + +static uvec8 kShuf38a = + { 0, 3, 6, 8, 11, 14, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128 }; + +static uvec8 kShuf38b = + { 128, 128, 128, 128, 128, 128, 0, 3, 6, 8, 11, 14, 128, 128, 128, 128 }; + +// Arrange words 0,3,6 into 0,1,2 +static uvec8 kShufAc = + { 0, 1, 6, 7, 12, 13, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128 }; + +// Arrange words 0,3,6 into 3,4,5 +static uvec8 kShufAc3 = + { 128, 128, 128, 128, 128, 128, 0, 1, 6, 7, 12, 13, 128, 128, 128, 128 }; + +// Scaling values for boxes of 3x3 and 2x3 +static uvec16 kScaleAc33 = + { 65536 / 9, 65536 / 9, 65536 / 6, 65536 / 9, 65536 / 9, 65536 / 6, 0, 0 }; + +// Arrange first value for pixels 0,1,2,3,4,5 +static uvec8 kShufAb0 = + { 0, 128, 3, 128, 6, 128, 8, 128, 11, 128, 14, 128, 128, 128, 128, 128 }; + +// Arrange second value for pixels 0,1,2,3,4,5 +static uvec8 kShufAb1 = + { 1, 128, 4, 128, 7, 128, 9, 128, 12, 128, 15, 128, 128, 128, 128, 128 }; + +// Arrange third value for pixels 0,1,2,3,4,5 +static uvec8 kShufAb2 = + { 2, 128, 5, 128, 128, 128, 10, 128, 13, 128, 128, 128, 128, 128, 128, 128 }; + +// Scaling values for boxes of 3x2 and 2x2 +static uvec16 kScaleAb2 = + { 65536 / 3, 65536 / 3, 65536 / 2, 65536 / 3, 65536 / 3, 65536 / 2, 0, 0 }; + +// Reads 32 pixels, throws half away and writes 16 pixels. +__declspec(naked) +void ScaleRowDown2_SSE2(const uint8* src_ptr, ptrdiff_t src_stride, + uint8* dst_ptr, int dst_width) { + __asm { + mov eax, [esp + 4] // src_ptr + // src_stride ignored + mov edx, [esp + 12] // dst_ptr + mov ecx, [esp + 16] // dst_width + + wloop: + movdqu xmm0, [eax] + movdqu xmm1, [eax + 16] + lea eax, [eax + 32] + psrlw xmm0, 8 // isolate odd pixels. + psrlw xmm1, 8 + packuswb xmm0, xmm1 + movdqu [edx], xmm0 + lea edx, [edx + 16] + sub ecx, 16 + jg wloop + + ret + } +} + +// Blends 32x1 rectangle to 16x1. +__declspec(naked) +void ScaleRowDown2Linear_SSE2(const uint8* src_ptr, ptrdiff_t src_stride, + uint8* dst_ptr, int dst_width) { + __asm { + mov eax, [esp + 4] // src_ptr + // src_stride + mov edx, [esp + 12] // dst_ptr + mov ecx, [esp + 16] // dst_width + pcmpeqb xmm5, xmm5 // generate mask 0x00ff00ff + psrlw xmm5, 8 + + wloop: + movdqu xmm0, [eax] + movdqu xmm1, [eax + 16] + lea eax, [eax + 32] + + movdqa xmm2, xmm0 // average columns (32 to 16 pixels) + psrlw xmm0, 8 + movdqa xmm3, xmm1 + psrlw xmm1, 8 + pand xmm2, xmm5 + pand xmm3, xmm5 + pavgw xmm0, xmm2 + pavgw xmm1, xmm3 + packuswb xmm0, xmm1 + + movdqu [edx], xmm0 + lea edx, [edx + 16] + sub ecx, 16 + jg wloop + + ret + } +} + +// Blends 32x2 rectangle to 16x1. +__declspec(naked) +void ScaleRowDown2Box_SSE2(const uint8* src_ptr, ptrdiff_t src_stride, + uint8* dst_ptr, int dst_width) { + __asm { + push esi + mov eax, [esp + 4 + 4] // src_ptr + mov esi, [esp + 4 + 8] // src_stride + mov edx, [esp + 4 + 12] // dst_ptr + mov ecx, [esp + 4 + 16] // dst_width + pcmpeqb xmm5, xmm5 // generate mask 0x00ff00ff + psrlw xmm5, 8 + + wloop: + movdqu xmm0, [eax] + movdqu xmm1, [eax + 16] + movdqu xmm2, [eax + esi] + movdqu xmm3, [eax + esi + 16] + lea eax, [eax + 32] + pavgb xmm0, xmm2 // average rows + pavgb xmm1, xmm3 + + movdqa xmm2, xmm0 // average columns (32 to 16 pixels) + psrlw xmm0, 8 + movdqa xmm3, xmm1 + psrlw xmm1, 8 + pand xmm2, xmm5 + pand xmm3, xmm5 + pavgw xmm0, xmm2 + pavgw xmm1, xmm3 + packuswb xmm0, xmm1 + + movdqu [edx], xmm0 + lea edx, [edx + 16] + sub ecx, 16 + jg wloop + + pop esi + ret + } +} + +#ifdef HAS_SCALEROWDOWN2_AVX2 +// Reads 64 pixels, throws half away and writes 32 pixels. +__declspec(naked) +void ScaleRowDown2_AVX2(const uint8* src_ptr, ptrdiff_t src_stride, + uint8* dst_ptr, int dst_width) { + __asm { + mov eax, [esp + 4] // src_ptr + // src_stride ignored + mov edx, [esp + 12] // dst_ptr + mov ecx, [esp + 16] // dst_width + + wloop: + vmovdqu ymm0, [eax] + vmovdqu ymm1, [eax + 32] + lea eax, [eax + 64] + vpsrlw ymm0, ymm0, 8 // isolate odd pixels. + vpsrlw ymm1, ymm1, 8 + vpackuswb ymm0, ymm0, ymm1 + vpermq ymm0, ymm0, 0xd8 // unmutate vpackuswb + vmovdqu [edx], ymm0 + lea edx, [edx + 32] + sub ecx, 32 + jg wloop + + vzeroupper + ret + } +} + +// Blends 64x1 rectangle to 32x1. +__declspec(naked) +void ScaleRowDown2Linear_AVX2(const uint8* src_ptr, ptrdiff_t src_stride, + uint8* dst_ptr, int dst_width) { + __asm { + mov eax, [esp + 4] // src_ptr + // src_stride + mov edx, [esp + 12] // dst_ptr + mov ecx, [esp + 16] // dst_width + + vpcmpeqb ymm4, ymm4, ymm4 // '1' constant, 8b + vpsrlw ymm4, ymm4, 15 + vpackuswb ymm4, ymm4, ymm4 + vpxor ymm5, ymm5, ymm5 // constant 0 + + wloop: + vmovdqu ymm0, [eax] + vmovdqu ymm1, [eax + 32] + lea eax, [eax + 64] + + vpmaddubsw ymm0, ymm0, ymm4 // average horizontally + vpmaddubsw ymm1, ymm1, ymm4 + vpavgw ymm0, ymm0, ymm5 // (x + 1) / 2 + vpavgw ymm1, ymm1, ymm5 + vpackuswb ymm0, ymm0, ymm1 + vpermq ymm0, ymm0, 0xd8 // unmutate vpackuswb + + vmovdqu [edx], ymm0 + lea edx, [edx + 32] + sub ecx, 32 + jg wloop + + vzeroupper + ret + } +} + +// Blends 64x2 rectangle to 32x1. +__declspec(naked) +void ScaleRowDown2Box_AVX2(const uint8* src_ptr, ptrdiff_t src_stride, + uint8* dst_ptr, int dst_width) { + __asm { + push esi + mov eax, [esp + 4 + 4] // src_ptr + mov esi, [esp + 4 + 8] // src_stride + mov edx, [esp + 4 + 12] // dst_ptr + mov ecx, [esp + 4 + 16] // dst_width + + vpcmpeqb ymm4, ymm4, ymm4 // '1' constant, 8b + vpsrlw ymm4, ymm4, 15 + vpackuswb ymm4, ymm4, ymm4 + vpxor ymm5, ymm5, ymm5 // constant 0 + + wloop: + vmovdqu ymm0, [eax] // average rows + vmovdqu ymm1, [eax + 32] + vpavgb ymm0, ymm0, [eax + esi] + vpavgb ymm1, ymm1, [eax + esi + 32] + lea eax, [eax + 64] + + vpmaddubsw ymm0, ymm0, ymm4 // average horizontally + vpmaddubsw ymm1, ymm1, ymm4 + vpavgw ymm0, ymm0, ymm5 // (x + 1) / 2 + vpavgw ymm1, ymm1, ymm5 + vpackuswb ymm0, ymm0, ymm1 + vpermq ymm0, ymm0, 0xd8 // unmutate vpackuswb + + vmovdqu [edx], ymm0 + lea edx, [edx + 32] + sub ecx, 32 + jg wloop + + pop esi + vzeroupper + ret + } +} +#endif // HAS_SCALEROWDOWN2_AVX2 + +// Point samples 32 pixels to 8 pixels. +__declspec(naked) +void ScaleRowDown4_SSE2(const uint8* src_ptr, ptrdiff_t src_stride, + uint8* dst_ptr, int dst_width) { + __asm { + mov eax, [esp + 4] // src_ptr + // src_stride ignored + mov edx, [esp + 12] // dst_ptr + mov ecx, [esp + 16] // dst_width + pcmpeqb xmm5, xmm5 // generate mask 0x00ff0000 + psrld xmm5, 24 + pslld xmm5, 16 + + wloop: + movdqu xmm0, [eax] + movdqu xmm1, [eax + 16] + lea eax, [eax + 32] + pand xmm0, xmm5 + pand xmm1, xmm5 + packuswb xmm0, xmm1 + psrlw xmm0, 8 + packuswb xmm0, xmm0 + movq qword ptr [edx], xmm0 + lea edx, [edx + 8] + sub ecx, 8 + jg wloop + + ret + } +} + +// Blends 32x4 rectangle to 8x1. +__declspec(naked) +void ScaleRowDown4Box_SSE2(const uint8* src_ptr, ptrdiff_t src_stride, + uint8* dst_ptr, int dst_width) { + __asm { + push esi + push edi + mov eax, [esp + 8 + 4] // src_ptr + mov esi, [esp + 8 + 8] // src_stride + mov edx, [esp + 8 + 12] // dst_ptr + mov ecx, [esp + 8 + 16] // dst_width + lea edi, [esi + esi * 2] // src_stride * 3 + pcmpeqb xmm7, xmm7 // generate mask 0x00ff00ff + psrlw xmm7, 8 + + wloop: + movdqu xmm0, [eax] // average rows + movdqu xmm1, [eax + 16] + movdqu xmm2, [eax + esi] + movdqu xmm3, [eax + esi + 16] + pavgb xmm0, xmm2 + pavgb xmm1, xmm3 + movdqu xmm2, [eax + esi * 2] + movdqu xmm3, [eax + esi * 2 + 16] + movdqu xmm4, [eax + edi] + movdqu xmm5, [eax + edi + 16] + lea eax, [eax + 32] + pavgb xmm2, xmm4 + pavgb xmm3, xmm5 + pavgb xmm0, xmm2 + pavgb xmm1, xmm3 + + movdqa xmm2, xmm0 // average columns (32 to 16 pixels) + psrlw xmm0, 8 + movdqa xmm3, xmm1 + psrlw xmm1, 8 + pand xmm2, xmm7 + pand xmm3, xmm7 + pavgw xmm0, xmm2 + pavgw xmm1, xmm3 + packuswb xmm0, xmm1 + + movdqa xmm2, xmm0 // average columns (16 to 8 pixels) + psrlw xmm0, 8 + pand xmm2, xmm7 + pavgw xmm0, xmm2 + packuswb xmm0, xmm0 + + movq qword ptr [edx], xmm0 + lea edx, [edx + 8] + sub ecx, 8 + jg wloop + + pop edi + pop esi + ret + } +} + +#ifdef HAS_SCALEROWDOWN4_AVX2 +// Point samples 64 pixels to 16 pixels. +__declspec(naked) +void ScaleRowDown4_AVX2(const uint8* src_ptr, ptrdiff_t src_stride, + uint8* dst_ptr, int dst_width) { + __asm { + mov eax, [esp + 4] // src_ptr + // src_stride ignored + mov edx, [esp + 12] // dst_ptr + mov ecx, [esp + 16] // dst_width + vpcmpeqb ymm5, ymm5, ymm5 // generate mask 0x00ff0000 + vpsrld ymm5, ymm5, 24 + vpslld ymm5, ymm5, 16 + + wloop: + vmovdqu ymm0, [eax] + vmovdqu ymm1, [eax + 32] + lea eax, [eax + 64] + vpand ymm0, ymm0, ymm5 + vpand ymm1, ymm1, ymm5 + vpackuswb ymm0, ymm0, ymm1 + vpermq ymm0, ymm0, 0xd8 // unmutate vpackuswb + vpsrlw ymm0, ymm0, 8 + vpackuswb ymm0, ymm0, ymm0 + vpermq ymm0, ymm0, 0xd8 // unmutate vpackuswb + vmovdqu [edx], xmm0 + lea edx, [edx + 16] + sub ecx, 16 + jg wloop + + vzeroupper + ret + } +} + +// Blends 64x4 rectangle to 16x1. +__declspec(naked) +void ScaleRowDown4Box_AVX2(const uint8* src_ptr, ptrdiff_t src_stride, + uint8* dst_ptr, int dst_width) { + __asm { + push esi + push edi + mov eax, [esp + 8 + 4] // src_ptr + mov esi, [esp + 8 + 8] // src_stride + mov edx, [esp + 8 + 12] // dst_ptr + mov ecx, [esp + 8 + 16] // dst_width + lea edi, [esi + esi * 2] // src_stride * 3 + vpcmpeqb ymm7, ymm7, ymm7 // generate mask 0x00ff00ff + vpsrlw ymm7, ymm7, 8 + + wloop: + vmovdqu ymm0, [eax] // average rows + vmovdqu ymm1, [eax + 32] + vpavgb ymm0, ymm0, [eax + esi] + vpavgb ymm1, ymm1, [eax + esi + 32] + vmovdqu ymm2, [eax + esi * 2] + vmovdqu ymm3, [eax + esi * 2 + 32] + vpavgb ymm2, ymm2, [eax + edi] + vpavgb ymm3, ymm3, [eax + edi + 32] + lea eax, [eax + 64] + vpavgb ymm0, ymm0, ymm2 + vpavgb ymm1, ymm1, ymm3 + + vpand ymm2, ymm0, ymm7 // average columns (64 to 32 pixels) + vpand ymm3, ymm1, ymm7 + vpsrlw ymm0, ymm0, 8 + vpsrlw ymm1, ymm1, 8 + vpavgw ymm0, ymm0, ymm2 + vpavgw ymm1, ymm1, ymm3 + vpackuswb ymm0, ymm0, ymm1 + vpermq ymm0, ymm0, 0xd8 // unmutate vpackuswb + + vpand ymm2, ymm0, ymm7 // average columns (32 to 16 pixels) + vpsrlw ymm0, ymm0, 8 + vpavgw ymm0, ymm0, ymm2 + vpackuswb ymm0, ymm0, ymm0 + vpermq ymm0, ymm0, 0xd8 // unmutate vpackuswb + + vmovdqu [edx], xmm0 + lea edx, [edx + 16] + sub ecx, 16 + jg wloop + + pop edi + pop esi + vzeroupper + ret + } +} +#endif // HAS_SCALEROWDOWN4_AVX2 + +// Point samples 32 pixels to 24 pixels. +// Produces three 8 byte values. For each 8 bytes, 16 bytes are read. +// Then shuffled to do the scaling. + +__declspec(naked) +void ScaleRowDown34_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride, + uint8* dst_ptr, int dst_width) { + __asm { + mov eax, [esp + 4] // src_ptr + // src_stride ignored + mov edx, [esp + 12] // dst_ptr + mov ecx, [esp + 16] // dst_width + movdqa xmm3, kShuf0 + movdqa xmm4, kShuf1 + movdqa xmm5, kShuf2 + + wloop: + movdqu xmm0, [eax] + movdqu xmm1, [eax + 16] + lea eax, [eax + 32] + movdqa xmm2, xmm1 + palignr xmm1, xmm0, 8 + pshufb xmm0, xmm3 + pshufb xmm1, xmm4 + pshufb xmm2, xmm5 + movq qword ptr [edx], xmm0 + movq qword ptr [edx + 8], xmm1 + movq qword ptr [edx + 16], xmm2 + lea edx, [edx + 24] + sub ecx, 24 + jg wloop + + ret + } +} + +// Blends 32x2 rectangle to 24x1 +// Produces three 8 byte values. For each 8 bytes, 16 bytes are read. +// Then shuffled to do the scaling. + +// Register usage: +// xmm0 src_row 0 +// xmm1 src_row 1 +// xmm2 shuf 0 +// xmm3 shuf 1 +// xmm4 shuf 2 +// xmm5 madd 0 +// xmm6 madd 1 +// xmm7 kRound34 + +// Note that movdqa+palign may be better than movdqu. +__declspec(naked) +void ScaleRowDown34_1_Box_SSSE3(const uint8* src_ptr, + ptrdiff_t src_stride, + uint8* dst_ptr, int dst_width) { + __asm { + push esi + mov eax, [esp + 4 + 4] // src_ptr + mov esi, [esp + 4 + 8] // src_stride + mov edx, [esp + 4 + 12] // dst_ptr + mov ecx, [esp + 4 + 16] // dst_width + movdqa xmm2, kShuf01 + movdqa xmm3, kShuf11 + movdqa xmm4, kShuf21 + movdqa xmm5, kMadd01 + movdqa xmm6, kMadd11 + movdqa xmm7, kRound34 + + wloop: + movdqu xmm0, [eax] // pixels 0..7 + movdqu xmm1, [eax + esi] + pavgb xmm0, xmm1 + pshufb xmm0, xmm2 + pmaddubsw xmm0, xmm5 + paddsw xmm0, xmm7 + psrlw xmm0, 2 + packuswb xmm0, xmm0 + movq qword ptr [edx], xmm0 + movdqu xmm0, [eax + 8] // pixels 8..15 + movdqu xmm1, [eax + esi + 8] + pavgb xmm0, xmm1 + pshufb xmm0, xmm3 + pmaddubsw xmm0, xmm6 + paddsw xmm0, xmm7 + psrlw xmm0, 2 + packuswb xmm0, xmm0 + movq qword ptr [edx + 8], xmm0 + movdqu xmm0, [eax + 16] // pixels 16..23 + movdqu xmm1, [eax + esi + 16] + lea eax, [eax + 32] + pavgb xmm0, xmm1 + pshufb xmm0, xmm4 + movdqa xmm1, kMadd21 + pmaddubsw xmm0, xmm1 + paddsw xmm0, xmm7 + psrlw xmm0, 2 + packuswb xmm0, xmm0 + movq qword ptr [edx + 16], xmm0 + lea edx, [edx + 24] + sub ecx, 24 + jg wloop + + pop esi + ret + } +} + +// Note that movdqa+palign may be better than movdqu. +__declspec(naked) +void ScaleRowDown34_0_Box_SSSE3(const uint8* src_ptr, + ptrdiff_t src_stride, + uint8* dst_ptr, int dst_width) { + __asm { + push esi + mov eax, [esp + 4 + 4] // src_ptr + mov esi, [esp + 4 + 8] // src_stride + mov edx, [esp + 4 + 12] // dst_ptr + mov ecx, [esp + 4 + 16] // dst_width + movdqa xmm2, kShuf01 + movdqa xmm3, kShuf11 + movdqa xmm4, kShuf21 + movdqa xmm5, kMadd01 + movdqa xmm6, kMadd11 + movdqa xmm7, kRound34 + + wloop: + movdqu xmm0, [eax] // pixels 0..7 + movdqu xmm1, [eax + esi] + pavgb xmm1, xmm0 + pavgb xmm0, xmm1 + pshufb xmm0, xmm2 + pmaddubsw xmm0, xmm5 + paddsw xmm0, xmm7 + psrlw xmm0, 2 + packuswb xmm0, xmm0 + movq qword ptr [edx], xmm0 + movdqu xmm0, [eax + 8] // pixels 8..15 + movdqu xmm1, [eax + esi + 8] + pavgb xmm1, xmm0 + pavgb xmm0, xmm1 + pshufb xmm0, xmm3 + pmaddubsw xmm0, xmm6 + paddsw xmm0, xmm7 + psrlw xmm0, 2 + packuswb xmm0, xmm0 + movq qword ptr [edx + 8], xmm0 + movdqu xmm0, [eax + 16] // pixels 16..23 + movdqu xmm1, [eax + esi + 16] + lea eax, [eax + 32] + pavgb xmm1, xmm0 + pavgb xmm0, xmm1 + pshufb xmm0, xmm4 + movdqa xmm1, kMadd21 + pmaddubsw xmm0, xmm1 + paddsw xmm0, xmm7 + psrlw xmm0, 2 + packuswb xmm0, xmm0 + movq qword ptr [edx + 16], xmm0 + lea edx, [edx+24] + sub ecx, 24 + jg wloop + + pop esi + ret + } +} + +// 3/8 point sampler + +// Scale 32 pixels to 12 +__declspec(naked) +void ScaleRowDown38_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride, + uint8* dst_ptr, int dst_width) { + __asm { + mov eax, [esp + 4] // src_ptr + // src_stride ignored + mov edx, [esp + 12] // dst_ptr + mov ecx, [esp + 16] // dst_width + movdqa xmm4, kShuf38a + movdqa xmm5, kShuf38b + + xloop: + movdqu xmm0, [eax] // 16 pixels -> 0,1,2,3,4,5 + movdqu xmm1, [eax + 16] // 16 pixels -> 6,7,8,9,10,11 + lea eax, [eax + 32] + pshufb xmm0, xmm4 + pshufb xmm1, xmm5 + paddusb xmm0, xmm1 + + movq qword ptr [edx], xmm0 // write 12 pixels + movhlps xmm1, xmm0 + movd [edx + 8], xmm1 + lea edx, [edx + 12] + sub ecx, 12 + jg xloop + + ret + } +} + +// Scale 16x3 pixels to 6x1 with interpolation +__declspec(naked) +void ScaleRowDown38_3_Box_SSSE3(const uint8* src_ptr, + ptrdiff_t src_stride, + uint8* dst_ptr, int dst_width) { + __asm { + push esi + mov eax, [esp + 4 + 4] // src_ptr + mov esi, [esp + 4 + 8] // src_stride + mov edx, [esp + 4 + 12] // dst_ptr + mov ecx, [esp + 4 + 16] // dst_width + movdqa xmm2, kShufAc + movdqa xmm3, kShufAc3 + movdqa xmm4, kScaleAc33 + pxor xmm5, xmm5 + + xloop: + movdqu xmm0, [eax] // sum up 3 rows into xmm0/1 + movdqu xmm6, [eax + esi] + movhlps xmm1, xmm0 + movhlps xmm7, xmm6 + punpcklbw xmm0, xmm5 + punpcklbw xmm1, xmm5 + punpcklbw xmm6, xmm5 + punpcklbw xmm7, xmm5 + paddusw xmm0, xmm6 + paddusw xmm1, xmm7 + movdqu xmm6, [eax + esi * 2] + lea eax, [eax + 16] + movhlps xmm7, xmm6 + punpcklbw xmm6, xmm5 + punpcklbw xmm7, xmm5 + paddusw xmm0, xmm6 + paddusw xmm1, xmm7 + + movdqa xmm6, xmm0 // 8 pixels -> 0,1,2 of xmm6 + psrldq xmm0, 2 + paddusw xmm6, xmm0 + psrldq xmm0, 2 + paddusw xmm6, xmm0 + pshufb xmm6, xmm2 + + movdqa xmm7, xmm1 // 8 pixels -> 3,4,5 of xmm6 + psrldq xmm1, 2 + paddusw xmm7, xmm1 + psrldq xmm1, 2 + paddusw xmm7, xmm1 + pshufb xmm7, xmm3 + paddusw xmm6, xmm7 + + pmulhuw xmm6, xmm4 // divide by 9,9,6, 9,9,6 + packuswb xmm6, xmm6 + + movd [edx], xmm6 // write 6 pixels + psrlq xmm6, 16 + movd [edx + 2], xmm6 + lea edx, [edx + 6] + sub ecx, 6 + jg xloop + + pop esi + ret + } +} + +// Scale 16x2 pixels to 6x1 with interpolation +__declspec(naked) +void ScaleRowDown38_2_Box_SSSE3(const uint8* src_ptr, + ptrdiff_t src_stride, + uint8* dst_ptr, int dst_width) { + __asm { + push esi + mov eax, [esp + 4 + 4] // src_ptr + mov esi, [esp + 4 + 8] // src_stride + mov edx, [esp + 4 + 12] // dst_ptr + mov ecx, [esp + 4 + 16] // dst_width + movdqa xmm2, kShufAb0 + movdqa xmm3, kShufAb1 + movdqa xmm4, kShufAb2 + movdqa xmm5, kScaleAb2 + + xloop: + movdqu xmm0, [eax] // average 2 rows into xmm0 + movdqu xmm1, [eax + esi] + lea eax, [eax + 16] + pavgb xmm0, xmm1 + + movdqa xmm1, xmm0 // 16 pixels -> 0,1,2,3,4,5 of xmm1 + pshufb xmm1, xmm2 + movdqa xmm6, xmm0 + pshufb xmm6, xmm3 + paddusw xmm1, xmm6 + pshufb xmm0, xmm4 + paddusw xmm1, xmm0 + + pmulhuw xmm1, xmm5 // divide by 3,3,2, 3,3,2 + packuswb xmm1, xmm1 + + movd [edx], xmm1 // write 6 pixels + psrlq xmm1, 16 + movd [edx + 2], xmm1 + lea edx, [edx + 6] + sub ecx, 6 + jg xloop + + pop esi + ret + } +} + +// Reads 16 bytes and accumulates to 16 shorts at a time. +__declspec(naked) +void ScaleAddRow_SSE2(const uint8* src_ptr, uint16* dst_ptr, int src_width) { + __asm { + mov eax, [esp + 4] // src_ptr + mov edx, [esp + 8] // dst_ptr + mov ecx, [esp + 12] // src_width + pxor xmm5, xmm5 + + // sum rows + xloop: + movdqu xmm3, [eax] // read 16 bytes + lea eax, [eax + 16] + movdqu xmm0, [edx] // read 16 words from destination + movdqu xmm1, [edx + 16] + movdqa xmm2, xmm3 + punpcklbw xmm2, xmm5 + punpckhbw xmm3, xmm5 + paddusw xmm0, xmm2 // sum 16 words + paddusw xmm1, xmm3 + movdqu [edx], xmm0 // write 16 words to destination + movdqu [edx + 16], xmm1 + lea edx, [edx + 32] + sub ecx, 16 + jg xloop + ret + } +} + +#ifdef HAS_SCALEADDROW_AVX2 +// Reads 32 bytes and accumulates to 32 shorts at a time. +__declspec(naked) +void ScaleAddRow_AVX2(const uint8* src_ptr, uint16* dst_ptr, int src_width) { + __asm { + mov eax, [esp + 4] // src_ptr + mov edx, [esp + 8] // dst_ptr + mov ecx, [esp + 12] // src_width + vpxor ymm5, ymm5, ymm5 + + // sum rows + xloop: + vmovdqu ymm3, [eax] // read 32 bytes + lea eax, [eax + 32] + vpermq ymm3, ymm3, 0xd8 // unmutate for vpunpck + vpunpcklbw ymm2, ymm3, ymm5 + vpunpckhbw ymm3, ymm3, ymm5 + vpaddusw ymm0, ymm2, [edx] // sum 16 words + vpaddusw ymm1, ymm3, [edx + 32] + vmovdqu [edx], ymm0 // write 32 words to destination + vmovdqu [edx + 32], ymm1 + lea edx, [edx + 64] + sub ecx, 32 + jg xloop + + vzeroupper + ret + } +} +#endif // HAS_SCALEADDROW_AVX2 + +// Bilinear column filtering. SSSE3 version. +__declspec(naked) +void ScaleFilterCols_SSSE3(uint8* dst_ptr, const uint8* src_ptr, + int dst_width, int x, int dx) { + __asm { + push ebx + push esi + push edi + mov edi, [esp + 12 + 4] // dst_ptr + mov esi, [esp + 12 + 8] // src_ptr + mov ecx, [esp + 12 + 12] // dst_width + movd xmm2, [esp + 12 + 16] // x + movd xmm3, [esp + 12 + 20] // dx + mov eax, 0x04040000 // shuffle to line up fractions with pixel. + movd xmm5, eax + pcmpeqb xmm6, xmm6 // generate 0x007f for inverting fraction. + psrlw xmm6, 9 + pextrw eax, xmm2, 1 // get x0 integer. preroll + sub ecx, 2 + jl xloop29 + + movdqa xmm0, xmm2 // x1 = x0 + dx + paddd xmm0, xmm3 + punpckldq xmm2, xmm0 // x0 x1 + punpckldq xmm3, xmm3 // dx dx + paddd xmm3, xmm3 // dx * 2, dx * 2 + pextrw edx, xmm2, 3 // get x1 integer. preroll + + // 2 Pixel loop. + xloop2: + movdqa xmm1, xmm2 // x0, x1 fractions. + paddd xmm2, xmm3 // x += dx + movzx ebx, word ptr [esi + eax] // 2 source x0 pixels + movd xmm0, ebx + psrlw xmm1, 9 // 7 bit fractions. + movzx ebx, word ptr [esi + edx] // 2 source x1 pixels + movd xmm4, ebx + pshufb xmm1, xmm5 // 0011 + punpcklwd xmm0, xmm4 + pxor xmm1, xmm6 // 0..7f and 7f..0 + pmaddubsw xmm0, xmm1 // 16 bit, 2 pixels. + pextrw eax, xmm2, 1 // get x0 integer. next iteration. + pextrw edx, xmm2, 3 // get x1 integer. next iteration. + psrlw xmm0, 7 // 8.7 fixed point to low 8 bits. + packuswb xmm0, xmm0 // 8 bits, 2 pixels. + movd ebx, xmm0 + mov [edi], bx + lea edi, [edi + 2] + sub ecx, 2 // 2 pixels + jge xloop2 + + xloop29: + + add ecx, 2 - 1 + jl xloop99 + + // 1 pixel remainder + movzx ebx, word ptr [esi + eax] // 2 source x0 pixels + movd xmm0, ebx + psrlw xmm2, 9 // 7 bit fractions. + pshufb xmm2, xmm5 // 0011 + pxor xmm2, xmm6 // 0..7f and 7f..0 + pmaddubsw xmm0, xmm2 // 16 bit + psrlw xmm0, 7 // 8.7 fixed point to low 8 bits. + packuswb xmm0, xmm0 // 8 bits + movd ebx, xmm0 + mov [edi], bl + + xloop99: + + pop edi + pop esi + pop ebx + ret + } +} + +// Reads 16 pixels, duplicates them and writes 32 pixels. +__declspec(naked) +void ScaleColsUp2_SSE2(uint8* dst_ptr, const uint8* src_ptr, + int dst_width, int x, int dx) { + __asm { + mov edx, [esp + 4] // dst_ptr + mov eax, [esp + 8] // src_ptr + mov ecx, [esp + 12] // dst_width + + wloop: + movdqu xmm0, [eax] + lea eax, [eax + 16] + movdqa xmm1, xmm0 + punpcklbw xmm0, xmm0 + punpckhbw xmm1, xmm1 + movdqu [edx], xmm0 + movdqu [edx + 16], xmm1 + lea edx, [edx + 32] + sub ecx, 32 + jg wloop + + ret + } +} + +// Reads 8 pixels, throws half away and writes 4 even pixels (0, 2, 4, 6) +__declspec(naked) +void ScaleARGBRowDown2_SSE2(const uint8* src_argb, + ptrdiff_t src_stride, + uint8* dst_argb, int dst_width) { + __asm { + mov eax, [esp + 4] // src_argb + // src_stride ignored + mov edx, [esp + 12] // dst_argb + mov ecx, [esp + 16] // dst_width + + wloop: + movdqu xmm0, [eax] + movdqu xmm1, [eax + 16] + lea eax, [eax + 32] + shufps xmm0, xmm1, 0xdd + movdqu [edx], xmm0 + lea edx, [edx + 16] + sub ecx, 4 + jg wloop + + ret + } +} + +// Blends 8x1 rectangle to 4x1. +__declspec(naked) +void ScaleARGBRowDown2Linear_SSE2(const uint8* src_argb, + ptrdiff_t src_stride, + uint8* dst_argb, int dst_width) { + __asm { + mov eax, [esp + 4] // src_argb + // src_stride ignored + mov edx, [esp + 12] // dst_argb + mov ecx, [esp + 16] // dst_width + + wloop: + movdqu xmm0, [eax] + movdqu xmm1, [eax + 16] + lea eax, [eax + 32] + movdqa xmm2, xmm0 + shufps xmm0, xmm1, 0x88 // even pixels + shufps xmm2, xmm1, 0xdd // odd pixels + pavgb xmm0, xmm2 + movdqu [edx], xmm0 + lea edx, [edx + 16] + sub ecx, 4 + jg wloop + + ret + } +} + +// Blends 8x2 rectangle to 4x1. +__declspec(naked) +void ScaleARGBRowDown2Box_SSE2(const uint8* src_argb, + ptrdiff_t src_stride, + uint8* dst_argb, int dst_width) { + __asm { + push esi + mov eax, [esp + 4 + 4] // src_argb + mov esi, [esp + 4 + 8] // src_stride + mov edx, [esp + 4 + 12] // dst_argb + mov ecx, [esp + 4 + 16] // dst_width + + wloop: + movdqu xmm0, [eax] + movdqu xmm1, [eax + 16] + movdqu xmm2, [eax + esi] + movdqu xmm3, [eax + esi + 16] + lea eax, [eax + 32] + pavgb xmm0, xmm2 // average rows + pavgb xmm1, xmm3 + movdqa xmm2, xmm0 // average columns (8 to 4 pixels) + shufps xmm0, xmm1, 0x88 // even pixels + shufps xmm2, xmm1, 0xdd // odd pixels + pavgb xmm0, xmm2 + movdqu [edx], xmm0 + lea edx, [edx + 16] + sub ecx, 4 + jg wloop + + pop esi + ret + } +} + +// Reads 4 pixels at a time. +__declspec(naked) +void ScaleARGBRowDownEven_SSE2(const uint8* src_argb, ptrdiff_t src_stride, + int src_stepx, + uint8* dst_argb, int dst_width) { + __asm { + push ebx + push edi + mov eax, [esp + 8 + 4] // src_argb + // src_stride ignored + mov ebx, [esp + 8 + 12] // src_stepx + mov edx, [esp + 8 + 16] // dst_argb + mov ecx, [esp + 8 + 20] // dst_width + lea ebx, [ebx * 4] + lea edi, [ebx + ebx * 2] + + wloop: + movd xmm0, [eax] + movd xmm1, [eax + ebx] + punpckldq xmm0, xmm1 + movd xmm2, [eax + ebx * 2] + movd xmm3, [eax + edi] + lea eax, [eax + ebx * 4] + punpckldq xmm2, xmm3 + punpcklqdq xmm0, xmm2 + movdqu [edx], xmm0 + lea edx, [edx + 16] + sub ecx, 4 + jg wloop + + pop edi + pop ebx + ret + } +} + +// Blends four 2x2 to 4x1. +__declspec(naked) +void ScaleARGBRowDownEvenBox_SSE2(const uint8* src_argb, + ptrdiff_t src_stride, + int src_stepx, + uint8* dst_argb, int dst_width) { + __asm { + push ebx + push esi + push edi + mov eax, [esp + 12 + 4] // src_argb + mov esi, [esp + 12 + 8] // src_stride + mov ebx, [esp + 12 + 12] // src_stepx + mov edx, [esp + 12 + 16] // dst_argb + mov ecx, [esp + 12 + 20] // dst_width + lea esi, [eax + esi] // row1 pointer + lea ebx, [ebx * 4] + lea edi, [ebx + ebx * 2] + + wloop: + movq xmm0, qword ptr [eax] // row0 4 pairs + movhps xmm0, qword ptr [eax + ebx] + movq xmm1, qword ptr [eax + ebx * 2] + movhps xmm1, qword ptr [eax + edi] + lea eax, [eax + ebx * 4] + movq xmm2, qword ptr [esi] // row1 4 pairs + movhps xmm2, qword ptr [esi + ebx] + movq xmm3, qword ptr [esi + ebx * 2] + movhps xmm3, qword ptr [esi + edi] + lea esi, [esi + ebx * 4] + pavgb xmm0, xmm2 // average rows + pavgb xmm1, xmm3 + movdqa xmm2, xmm0 // average columns (8 to 4 pixels) + shufps xmm0, xmm1, 0x88 // even pixels + shufps xmm2, xmm1, 0xdd // odd pixels + pavgb xmm0, xmm2 + movdqu [edx], xmm0 + lea edx, [edx + 16] + sub ecx, 4 + jg wloop + + pop edi + pop esi + pop ebx + ret + } +} + +// Column scaling unfiltered. SSE2 version. +__declspec(naked) +void ScaleARGBCols_SSE2(uint8* dst_argb, const uint8* src_argb, + int dst_width, int x, int dx) { + __asm { + push edi + push esi + mov edi, [esp + 8 + 4] // dst_argb + mov esi, [esp + 8 + 8] // src_argb + mov ecx, [esp + 8 + 12] // dst_width + movd xmm2, [esp + 8 + 16] // x + movd xmm3, [esp + 8 + 20] // dx + + pshufd xmm2, xmm2, 0 // x0 x0 x0 x0 + pshufd xmm0, xmm3, 0x11 // dx 0 dx 0 + paddd xmm2, xmm0 + paddd xmm3, xmm3 // 0, 0, 0, dx * 2 + pshufd xmm0, xmm3, 0x05 // dx * 2, dx * 2, 0, 0 + paddd xmm2, xmm0 // x3 x2 x1 x0 + paddd xmm3, xmm3 // 0, 0, 0, dx * 4 + pshufd xmm3, xmm3, 0 // dx * 4, dx * 4, dx * 4, dx * 4 + + pextrw eax, xmm2, 1 // get x0 integer. + pextrw edx, xmm2, 3 // get x1 integer. + + cmp ecx, 0 + jle xloop99 + sub ecx, 4 + jl xloop49 + + // 4 Pixel loop. + xloop4: + movd xmm0, [esi + eax * 4] // 1 source x0 pixels + movd xmm1, [esi + edx * 4] // 1 source x1 pixels + pextrw eax, xmm2, 5 // get x2 integer. + pextrw edx, xmm2, 7 // get x3 integer. + paddd xmm2, xmm3 // x += dx + punpckldq xmm0, xmm1 // x0 x1 + + movd xmm1, [esi + eax * 4] // 1 source x2 pixels + movd xmm4, [esi + edx * 4] // 1 source x3 pixels + pextrw eax, xmm2, 1 // get x0 integer. next iteration. + pextrw edx, xmm2, 3 // get x1 integer. next iteration. + punpckldq xmm1, xmm4 // x2 x3 + punpcklqdq xmm0, xmm1 // x0 x1 x2 x3 + movdqu [edi], xmm0 + lea edi, [edi + 16] + sub ecx, 4 // 4 pixels + jge xloop4 + + xloop49: + test ecx, 2 + je xloop29 + + // 2 Pixels. + movd xmm0, [esi + eax * 4] // 1 source x0 pixels + movd xmm1, [esi + edx * 4] // 1 source x1 pixels + pextrw eax, xmm2, 5 // get x2 integer. + punpckldq xmm0, xmm1 // x0 x1 + + movq qword ptr [edi], xmm0 + lea edi, [edi + 8] + + xloop29: + test ecx, 1 + je xloop99 + + // 1 Pixels. + movd xmm0, [esi + eax * 4] // 1 source x2 pixels + movd dword ptr [edi], xmm0 + xloop99: + + pop esi + pop edi + ret + } +} + +// Bilinear row filtering combines 2x1 -> 1x1. SSSE3 version. +// TODO(fbarchard): Port to Neon + +// Shuffle table for arranging 2 pixels into pairs for pmaddubsw +static uvec8 kShuffleColARGB = { + 0u, 4u, 1u, 5u, 2u, 6u, 3u, 7u, // bbggrraa 1st pixel + 8u, 12u, 9u, 13u, 10u, 14u, 11u, 15u // bbggrraa 2nd pixel +}; + +// Shuffle table for duplicating 2 fractions into 8 bytes each +static uvec8 kShuffleFractions = { + 0u, 0u, 0u, 0u, 0u, 0u, 0u, 0u, 4u, 4u, 4u, 4u, 4u, 4u, 4u, 4u, +}; + +__declspec(naked) +void ScaleARGBFilterCols_SSSE3(uint8* dst_argb, const uint8* src_argb, + int dst_width, int x, int dx) { + __asm { + push esi + push edi + mov edi, [esp + 8 + 4] // dst_argb + mov esi, [esp + 8 + 8] // src_argb + mov ecx, [esp + 8 + 12] // dst_width + movd xmm2, [esp + 8 + 16] // x + movd xmm3, [esp + 8 + 20] // dx + movdqa xmm4, kShuffleColARGB + movdqa xmm5, kShuffleFractions + pcmpeqb xmm6, xmm6 // generate 0x007f for inverting fraction. + psrlw xmm6, 9 + pextrw eax, xmm2, 1 // get x0 integer. preroll + sub ecx, 2 + jl xloop29 + + movdqa xmm0, xmm2 // x1 = x0 + dx + paddd xmm0, xmm3 + punpckldq xmm2, xmm0 // x0 x1 + punpckldq xmm3, xmm3 // dx dx + paddd xmm3, xmm3 // dx * 2, dx * 2 + pextrw edx, xmm2, 3 // get x1 integer. preroll + + // 2 Pixel loop. + xloop2: + movdqa xmm1, xmm2 // x0, x1 fractions. + paddd xmm2, xmm3 // x += dx + movq xmm0, qword ptr [esi + eax * 4] // 2 source x0 pixels + psrlw xmm1, 9 // 7 bit fractions. + movhps xmm0, qword ptr [esi + edx * 4] // 2 source x1 pixels + pshufb xmm1, xmm5 // 0000000011111111 + pshufb xmm0, xmm4 // arrange pixels into pairs + pxor xmm1, xmm6 // 0..7f and 7f..0 + pmaddubsw xmm0, xmm1 // argb_argb 16 bit, 2 pixels. + pextrw eax, xmm2, 1 // get x0 integer. next iteration. + pextrw edx, xmm2, 3 // get x1 integer. next iteration. + psrlw xmm0, 7 // argb 8.7 fixed point to low 8 bits. + packuswb xmm0, xmm0 // argb_argb 8 bits, 2 pixels. + movq qword ptr [edi], xmm0 + lea edi, [edi + 8] + sub ecx, 2 // 2 pixels + jge xloop2 + + xloop29: + + add ecx, 2 - 1 + jl xloop99 + + // 1 pixel remainder + psrlw xmm2, 9 // 7 bit fractions. + movq xmm0, qword ptr [esi + eax * 4] // 2 source x0 pixels + pshufb xmm2, xmm5 // 00000000 + pshufb xmm0, xmm4 // arrange pixels into pairs + pxor xmm2, xmm6 // 0..7f and 7f..0 + pmaddubsw xmm0, xmm2 // argb 16 bit, 1 pixel. + psrlw xmm0, 7 + packuswb xmm0, xmm0 // argb 8 bits, 1 pixel. + movd [edi], xmm0 + + xloop99: + + pop edi + pop esi + ret + } +} + +// Reads 4 pixels, duplicates them and writes 8 pixels. +__declspec(naked) +void ScaleARGBColsUp2_SSE2(uint8* dst_argb, const uint8* src_argb, + int dst_width, int x, int dx) { + __asm { + mov edx, [esp + 4] // dst_argb + mov eax, [esp + 8] // src_argb + mov ecx, [esp + 12] // dst_width + + wloop: + movdqu xmm0, [eax] + lea eax, [eax + 16] + movdqa xmm1, xmm0 + punpckldq xmm0, xmm0 + punpckhdq xmm1, xmm1 + movdqu [edx], xmm0 + movdqu [edx + 16], xmm1 + lea edx, [edx + 32] + sub ecx, 8 + jg wloop + + ret + } +} + +// Divide num by div and return as 16.16 fixed point result. +__declspec(naked) +int FixedDiv_X86(int num, int div) { + __asm { + mov eax, [esp + 4] // num + cdq // extend num to 64 bits + shld edx, eax, 16 // 32.16 + shl eax, 16 + idiv dword ptr [esp + 8] + ret + } +} + +// Divide num by div and return as 16.16 fixed point result. +__declspec(naked) +int FixedDiv1_X86(int num, int div) { + __asm { + mov eax, [esp + 4] // num + mov ecx, [esp + 8] // denom + cdq // extend num to 64 bits + shld edx, eax, 16 // 32.16 + shl eax, 16 + sub eax, 0x00010001 + sbb edx, 0 + sub ecx, 1 + idiv ecx + ret + } +} +#endif // !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86) + +#ifdef __cplusplus +} // extern "C" +} // namespace libyuv +#endif diff --git a/media/libaom/src/third_party/libyuv/source/video_common.cc b/media/libaom/src/third_party/libyuv/source/video_common.cc new file mode 100644 index 000000000..379a0669a --- /dev/null +++ b/media/libaom/src/third_party/libyuv/source/video_common.cc @@ -0,0 +1,64 @@ +/* + * Copyright 2011 The LibYuv Project Authors. All rights reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + + +#include "libyuv/video_common.h" + +#ifdef __cplusplus +namespace libyuv { +extern "C" { +#endif + +#define ARRAY_SIZE(x) (int)(sizeof(x) / sizeof(x[0])) + +struct FourCCAliasEntry { + uint32 alias; + uint32 canonical; +}; + +static const struct FourCCAliasEntry kFourCCAliases[] = { + {FOURCC_IYUV, FOURCC_I420}, + {FOURCC_YU16, FOURCC_I422}, + {FOURCC_YU24, FOURCC_I444}, + {FOURCC_YUYV, FOURCC_YUY2}, + {FOURCC_YUVS, FOURCC_YUY2}, // kCMPixelFormat_422YpCbCr8_yuvs + {FOURCC_HDYC, FOURCC_UYVY}, + {FOURCC_2VUY, FOURCC_UYVY}, // kCMPixelFormat_422YpCbCr8 + {FOURCC_JPEG, FOURCC_MJPG}, // Note: JPEG has DHT while MJPG does not. + {FOURCC_DMB1, FOURCC_MJPG}, + {FOURCC_BA81, FOURCC_BGGR}, // deprecated. + {FOURCC_RGB3, FOURCC_RAW }, + {FOURCC_BGR3, FOURCC_24BG}, + {FOURCC_CM32, FOURCC_BGRA}, // kCMPixelFormat_32ARGB + {FOURCC_CM24, FOURCC_RAW }, // kCMPixelFormat_24RGB + {FOURCC_L555, FOURCC_RGBO}, // kCMPixelFormat_16LE555 + {FOURCC_L565, FOURCC_RGBP}, // kCMPixelFormat_16LE565 + {FOURCC_5551, FOURCC_RGBO}, // kCMPixelFormat_16LE5551 +}; +// TODO(fbarchard): Consider mapping kCMPixelFormat_32BGRA to FOURCC_ARGB. +// {FOURCC_BGRA, FOURCC_ARGB}, // kCMPixelFormat_32BGRA + +LIBYUV_API +uint32 CanonicalFourCC(uint32 fourcc) { + int i; + for (i = 0; i < ARRAY_SIZE(kFourCCAliases); ++i) { + if (kFourCCAliases[i].alias == fourcc) { + return kFourCCAliases[i].canonical; + } + } + // Not an alias, so return it as-is. + return fourcc; +} + +#ifdef __cplusplus +} // extern "C" +} // namespace libyuv +#endif + diff --git a/media/libaom/src/third_party/libyuv/source/x86inc.asm b/media/libaom/src/third_party/libyuv/source/x86inc.asm new file mode 100644 index 000000000..cb5c32df3 --- /dev/null +++ b/media/libaom/src/third_party/libyuv/source/x86inc.asm @@ -0,0 +1,1136 @@ +;***************************************************************************** +;* x86inc.asm: x264asm abstraction layer +;***************************************************************************** +;* Copyright (C) 2005-2012 x264 project +;* +;* Authors: Loren Merritt <lorenm@u.washington.edu> +;* Anton Mitrofanov <BugMaster@narod.ru> +;* Jason Garrett-Glaser <darkshikari@gmail.com> +;* Henrik Gramner <hengar-6@student.ltu.se> +;* +;* Permission to use, copy, modify, and/or distribute this software for any +;* purpose with or without fee is hereby granted, provided that the above +;* copyright notice and this permission notice appear in all copies. +;* +;* THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES +;* WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF +;* MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR +;* ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES +;* WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN +;* ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF +;* OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE. +;***************************************************************************** + +; This is a header file for the x264ASM assembly language, which uses +; NASM/YASM syntax combined with a large number of macros to provide easy +; abstraction between different calling conventions (x86_32, win64, linux64). +; It also has various other useful features to simplify writing the kind of +; DSP functions that are most often used in x264. + +; Unlike the rest of x264, this file is available under an ISC license, as it +; has significant usefulness outside of x264 and we want it to be available +; to the largest audience possible. Of course, if you modify it for your own +; purposes to add a new feature, we strongly encourage contributing a patch +; as this feature might be useful for others as well. Send patches or ideas +; to x264-devel@videolan.org . + +; Local changes for libyuv: +; remove %define program_name and references in labels +; rename cpus to uppercase + +%define WIN64 0 +%define UNIX64 0 +%if ARCH_X86_64 + %ifidn __OUTPUT_FORMAT__,win32 + %define WIN64 1 + %elifidn __OUTPUT_FORMAT__,win64 + %define WIN64 1 + %else + %define UNIX64 1 + %endif +%endif + +%ifdef PREFIX + %define mangle(x) _ %+ x +%else + %define mangle(x) x +%endif + +; Name of the .rodata section. +; Kludge: Something on OS X fails to align .rodata even given an align attribute, +; so use a different read-only section. +%macro SECTION_RODATA 0-1 16 + %ifidn __OUTPUT_FORMAT__,macho64 + SECTION .text align=%1 + %elifidn __OUTPUT_FORMAT__,macho + SECTION .text align=%1 + fakegot: + %elifidn __OUTPUT_FORMAT__,aout + section .text + %else + SECTION .rodata align=%1 + %endif +%endmacro + +; aout does not support align= +%macro SECTION_TEXT 0-1 16 + %ifidn __OUTPUT_FORMAT__,aout + SECTION .text + %else + SECTION .text align=%1 + %endif +%endmacro + +%if WIN64 + %define PIC +%elif ARCH_X86_64 == 0 +; x86_32 doesn't require PIC. +; Some distros prefer shared objects to be PIC, but nothing breaks if +; the code contains a few textrels, so we'll skip that complexity. + %undef PIC +%endif +%ifdef PIC + default rel +%endif + +; Always use long nops (reduces 0x90 spam in disassembly on x86_32) +CPU amdnop + +; Macros to eliminate most code duplication between x86_32 and x86_64: +; Currently this works only for leaf functions which load all their arguments +; into registers at the start, and make no other use of the stack. Luckily that +; covers most of x264's asm. + +; PROLOGUE: +; %1 = number of arguments. loads them from stack if needed. +; %2 = number of registers used. pushes callee-saved regs if needed. +; %3 = number of xmm registers used. pushes callee-saved xmm regs if needed. +; %4 = list of names to define to registers +; PROLOGUE can also be invoked by adding the same options to cglobal + +; e.g. +; cglobal foo, 2,3,0, dst, src, tmp +; declares a function (foo), taking two args (dst and src) and one local variable (tmp) + +; TODO Some functions can use some args directly from the stack. If they're the +; last args then you can just not declare them, but if they're in the middle +; we need more flexible macro. + +; RET: +; Pops anything that was pushed by PROLOGUE, and returns. + +; REP_RET: +; Same, but if it doesn't pop anything it becomes a 2-byte ret, for athlons +; which are slow when a normal ret follows a branch. + +; registers: +; rN and rNq are the native-size register holding function argument N +; rNd, rNw, rNb are dword, word, and byte size +; rNh is the high 8 bits of the word size +; rNm is the original location of arg N (a register or on the stack), dword +; rNmp is native size + +%macro DECLARE_REG 2-3 + %define r%1q %2 + %define r%1d %2d + %define r%1w %2w + %define r%1b %2b + %define r%1h %2h + %if %0 == 2 + %define r%1m %2d + %define r%1mp %2 + %elif ARCH_X86_64 ; memory + %define r%1m [rsp + stack_offset + %3] + %define r%1mp qword r %+ %1m + %else + %define r%1m [esp + stack_offset + %3] + %define r%1mp dword r %+ %1m + %endif + %define r%1 %2 +%endmacro + +%macro DECLARE_REG_SIZE 3 + %define r%1q r%1 + %define e%1q r%1 + %define r%1d e%1 + %define e%1d e%1 + %define r%1w %1 + %define e%1w %1 + %define r%1h %3 + %define e%1h %3 + %define r%1b %2 + %define e%1b %2 +%if ARCH_X86_64 == 0 + %define r%1 e%1 +%endif +%endmacro + +DECLARE_REG_SIZE ax, al, ah +DECLARE_REG_SIZE bx, bl, bh +DECLARE_REG_SIZE cx, cl, ch +DECLARE_REG_SIZE dx, dl, dh +DECLARE_REG_SIZE si, sil, null +DECLARE_REG_SIZE di, dil, null +DECLARE_REG_SIZE bp, bpl, null + +; t# defines for when per-arch register allocation is more complex than just function arguments + +%macro DECLARE_REG_TMP 1-* + %assign %%i 0 + %rep %0 + CAT_XDEFINE t, %%i, r%1 + %assign %%i %%i+1 + %rotate 1 + %endrep +%endmacro + +%macro DECLARE_REG_TMP_SIZE 0-* + %rep %0 + %define t%1q t%1 %+ q + %define t%1d t%1 %+ d + %define t%1w t%1 %+ w + %define t%1h t%1 %+ h + %define t%1b t%1 %+ b + %rotate 1 + %endrep +%endmacro + +DECLARE_REG_TMP_SIZE 0,1,2,3,4,5,6,7,8,9,10,11,12,13,14 + +%if ARCH_X86_64 + %define gprsize 8 +%else + %define gprsize 4 +%endif + +%macro PUSH 1 + push %1 + %assign stack_offset stack_offset+gprsize +%endmacro + +%macro POP 1 + pop %1 + %assign stack_offset stack_offset-gprsize +%endmacro + +%macro PUSH_IF_USED 1-* + %rep %0 + %if %1 < regs_used + PUSH r%1 + %endif + %rotate 1 + %endrep +%endmacro + +%macro POP_IF_USED 1-* + %rep %0 + %if %1 < regs_used + pop r%1 + %endif + %rotate 1 + %endrep +%endmacro + +%macro LOAD_IF_USED 1-* + %rep %0 + %if %1 < num_args + mov r%1, r %+ %1 %+ mp + %endif + %rotate 1 + %endrep +%endmacro + +%macro SUB 2 + sub %1, %2 + %ifidn %1, rsp + %assign stack_offset stack_offset+(%2) + %endif +%endmacro + +%macro ADD 2 + add %1, %2 + %ifidn %1, rsp + %assign stack_offset stack_offset-(%2) + %endif +%endmacro + +%macro movifnidn 2 + %ifnidn %1, %2 + mov %1, %2 + %endif +%endmacro + +%macro movsxdifnidn 2 + %ifnidn %1, %2 + movsxd %1, %2 + %endif +%endmacro + +%macro ASSERT 1 + %if (%1) == 0 + %error assert failed + %endif +%endmacro + +%macro DEFINE_ARGS 0-* + %ifdef n_arg_names + %assign %%i 0 + %rep n_arg_names + CAT_UNDEF arg_name %+ %%i, q + CAT_UNDEF arg_name %+ %%i, d + CAT_UNDEF arg_name %+ %%i, w + CAT_UNDEF arg_name %+ %%i, h + CAT_UNDEF arg_name %+ %%i, b + CAT_UNDEF arg_name %+ %%i, m + CAT_UNDEF arg_name %+ %%i, mp + CAT_UNDEF arg_name, %%i + %assign %%i %%i+1 + %endrep + %endif + + %xdefine %%stack_offset stack_offset + %undef stack_offset ; so that the current value of stack_offset doesn't get baked in by xdefine + %assign %%i 0 + %rep %0 + %xdefine %1q r %+ %%i %+ q + %xdefine %1d r %+ %%i %+ d + %xdefine %1w r %+ %%i %+ w + %xdefine %1h r %+ %%i %+ h + %xdefine %1b r %+ %%i %+ b + %xdefine %1m r %+ %%i %+ m + %xdefine %1mp r %+ %%i %+ mp + CAT_XDEFINE arg_name, %%i, %1 + %assign %%i %%i+1 + %rotate 1 + %endrep + %xdefine stack_offset %%stack_offset + %assign n_arg_names %0 +%endmacro + +%if WIN64 ; Windows x64 ;================================================= + +DECLARE_REG 0, rcx +DECLARE_REG 1, rdx +DECLARE_REG 2, R8 +DECLARE_REG 3, R9 +DECLARE_REG 4, R10, 40 +DECLARE_REG 5, R11, 48 +DECLARE_REG 6, rax, 56 +DECLARE_REG 7, rdi, 64 +DECLARE_REG 8, rsi, 72 +DECLARE_REG 9, rbx, 80 +DECLARE_REG 10, rbp, 88 +DECLARE_REG 11, R12, 96 +DECLARE_REG 12, R13, 104 +DECLARE_REG 13, R14, 112 +DECLARE_REG 14, R15, 120 + +%macro PROLOGUE 2-4+ 0 ; #args, #regs, #xmm_regs, arg_names... + %assign num_args %1 + %assign regs_used %2 + ASSERT regs_used >= num_args + ASSERT regs_used <= 15 + PUSH_IF_USED 7, 8, 9, 10, 11, 12, 13, 14 + %if mmsize == 8 + %assign xmm_regs_used 0 + %else + WIN64_SPILL_XMM %3 + %endif + LOAD_IF_USED 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14 + DEFINE_ARGS %4 +%endmacro + +%macro WIN64_SPILL_XMM 1 + %assign xmm_regs_used %1 + ASSERT xmm_regs_used <= 16 + %if xmm_regs_used > 6 + SUB rsp, (xmm_regs_used-6)*16+16 + %assign %%i xmm_regs_used + %rep (xmm_regs_used-6) + %assign %%i %%i-1 + movdqa [rsp + (%%i-6)*16+(~stack_offset&8)], xmm %+ %%i + %endrep + %endif +%endmacro + +%macro WIN64_RESTORE_XMM_INTERNAL 1 + %if xmm_regs_used > 6 + %assign %%i xmm_regs_used + %rep (xmm_regs_used-6) + %assign %%i %%i-1 + movdqa xmm %+ %%i, [%1 + (%%i-6)*16+(~stack_offset&8)] + %endrep + add %1, (xmm_regs_used-6)*16+16 + %endif +%endmacro + +%macro WIN64_RESTORE_XMM 1 + WIN64_RESTORE_XMM_INTERNAL %1 + %assign stack_offset stack_offset-(xmm_regs_used-6)*16+16 + %assign xmm_regs_used 0 +%endmacro + +%define has_epilogue regs_used > 7 || xmm_regs_used > 6 || mmsize == 32 + +%macro RET 0 + WIN64_RESTORE_XMM_INTERNAL rsp + POP_IF_USED 14, 13, 12, 11, 10, 9, 8, 7 +%if mmsize == 32 + vzeroupper +%endif + ret +%endmacro + +%elif ARCH_X86_64 ; *nix x64 ;============================================= + +DECLARE_REG 0, rdi +DECLARE_REG 1, rsi +DECLARE_REG 2, rdx +DECLARE_REG 3, rcx +DECLARE_REG 4, R8 +DECLARE_REG 5, R9 +DECLARE_REG 6, rax, 8 +DECLARE_REG 7, R10, 16 +DECLARE_REG 8, R11, 24 +DECLARE_REG 9, rbx, 32 +DECLARE_REG 10, rbp, 40 +DECLARE_REG 11, R12, 48 +DECLARE_REG 12, R13, 56 +DECLARE_REG 13, R14, 64 +DECLARE_REG 14, R15, 72 + +%macro PROLOGUE 2-4+ ; #args, #regs, #xmm_regs, arg_names... + %assign num_args %1 + %assign regs_used %2 + ASSERT regs_used >= num_args + ASSERT regs_used <= 15 + PUSH_IF_USED 9, 10, 11, 12, 13, 14 + LOAD_IF_USED 6, 7, 8, 9, 10, 11, 12, 13, 14 + DEFINE_ARGS %4 +%endmacro + +%define has_epilogue regs_used > 9 || mmsize == 32 + +%macro RET 0 + POP_IF_USED 14, 13, 12, 11, 10, 9 +%if mmsize == 32 + vzeroupper +%endif + ret +%endmacro + +%else ; X86_32 ;============================================================== + +DECLARE_REG 0, eax, 4 +DECLARE_REG 1, ecx, 8 +DECLARE_REG 2, edx, 12 +DECLARE_REG 3, ebx, 16 +DECLARE_REG 4, esi, 20 +DECLARE_REG 5, edi, 24 +DECLARE_REG 6, ebp, 28 +%define rsp esp + +%macro DECLARE_ARG 1-* + %rep %0 + %define r%1m [esp + stack_offset + 4*%1 + 4] + %define r%1mp dword r%1m + %rotate 1 + %endrep +%endmacro + +DECLARE_ARG 7, 8, 9, 10, 11, 12, 13, 14 + +%macro PROLOGUE 2-4+ ; #args, #regs, #xmm_regs, arg_names... + %assign num_args %1 + %assign regs_used %2 + %if regs_used > 7 + %assign regs_used 7 + %endif + ASSERT regs_used >= num_args + PUSH_IF_USED 3, 4, 5, 6 + LOAD_IF_USED 0, 1, 2, 3, 4, 5, 6 + DEFINE_ARGS %4 +%endmacro + +%define has_epilogue regs_used > 3 || mmsize == 32 + +%macro RET 0 + POP_IF_USED 6, 5, 4, 3 +%if mmsize == 32 + vzeroupper +%endif + ret +%endmacro + +%endif ;====================================================================== + +%if WIN64 == 0 +%macro WIN64_SPILL_XMM 1 +%endmacro +%macro WIN64_RESTORE_XMM 1 +%endmacro +%endif + +%macro REP_RET 0 + %if has_epilogue + RET + %else + rep ret + %endif +%endmacro + +%macro TAIL_CALL 2 ; callee, is_nonadjacent + %if has_epilogue + call %1 + RET + %elif %2 + jmp %1 + %endif +%endmacro + +;============================================================================= +; arch-independent part +;============================================================================= + +%assign function_align 16 + +; Begin a function. +; Applies any symbol mangling needed for C linkage, and sets up a define such that +; subsequent uses of the function name automatically refer to the mangled version. +; Appends cpuflags to the function name if cpuflags has been specified. +%macro cglobal 1-2+ ; name, [PROLOGUE args] +%if %0 == 1 + cglobal_internal %1 %+ SUFFIX +%else + cglobal_internal %1 %+ SUFFIX, %2 +%endif +%endmacro +%macro cglobal_internal 1-2+ + %ifndef cglobaled_%1 + %xdefine %1 mangle(%1) + %xdefine %1.skip_prologue %1 %+ .skip_prologue + CAT_XDEFINE cglobaled_, %1, 1 + %endif + %xdefine current_function %1 + %ifidn __OUTPUT_FORMAT__,elf + global %1:function hidden + %else + global %1 + %endif + align function_align + %1: + RESET_MM_PERMUTATION ; not really needed, but makes disassembly somewhat nicer + %assign stack_offset 0 + %if %0 > 1 + PROLOGUE %2 + %endif +%endmacro + +%macro cextern 1 + %xdefine %1 mangle(%1) + CAT_XDEFINE cglobaled_, %1, 1 + extern %1 +%endmacro + +; like cextern, but without the prefix +%macro cextern_naked 1 + %xdefine %1 mangle(%1) + CAT_XDEFINE cglobaled_, %1, 1 + extern %1 +%endmacro + +%macro const 2+ + %xdefine %1 mangle(%1) + global %1 + %1: %2 +%endmacro + +; This is needed for ELF, otherwise the GNU linker assumes the stack is +; executable by default. +%ifidn __OUTPUT_FORMAT__,elf +SECTION .note.GNU-stack noalloc noexec nowrite progbits +%endif +%ifidn __OUTPUT_FORMAT__,elf32 +section .note.GNU-stack noalloc noexec nowrite progbits +%endif +%ifidn __OUTPUT_FORMAT__,elf64 +section .note.GNU-stack noalloc noexec nowrite progbits +%endif + +; cpuflags + +%assign cpuflags_MMX (1<<0) +%assign cpuflags_MMX2 (1<<1) | cpuflags_MMX +%assign cpuflags_3dnow (1<<2) | cpuflags_MMX +%assign cpuflags_3dnow2 (1<<3) | cpuflags_3dnow +%assign cpuflags_SSE (1<<4) | cpuflags_MMX2 +%assign cpuflags_SSE2 (1<<5) | cpuflags_SSE +%assign cpuflags_SSE2slow (1<<6) | cpuflags_SSE2 +%assign cpuflags_SSE3 (1<<7) | cpuflags_SSE2 +%assign cpuflags_SSSE3 (1<<8) | cpuflags_SSE3 +%assign cpuflags_SSE4 (1<<9) | cpuflags_SSSE3 +%assign cpuflags_SSE42 (1<<10)| cpuflags_SSE4 +%assign cpuflags_AVX (1<<11)| cpuflags_SSE42 +%assign cpuflags_xop (1<<12)| cpuflags_AVX +%assign cpuflags_fma4 (1<<13)| cpuflags_AVX +%assign cpuflags_AVX2 (1<<14)| cpuflags_AVX +%assign cpuflags_fma3 (1<<15)| cpuflags_AVX + +%assign cpuflags_cache32 (1<<16) +%assign cpuflags_cache64 (1<<17) +%assign cpuflags_slowctz (1<<18) +%assign cpuflags_lzcnt (1<<19) +%assign cpuflags_misalign (1<<20) +%assign cpuflags_aligned (1<<21) ; not a cpu feature, but a function variant +%assign cpuflags_atom (1<<22) +%assign cpuflags_bmi1 (1<<23) +%assign cpuflags_bmi2 (1<<24)|cpuflags_bmi1 +%assign cpuflags_tbm (1<<25)|cpuflags_bmi1 + +%define cpuflag(x) ((cpuflags & (cpuflags_ %+ x)) == (cpuflags_ %+ x)) +%define notcpuflag(x) ((cpuflags & (cpuflags_ %+ x)) != (cpuflags_ %+ x)) + +; Takes up to 2 cpuflags from the above list. +; All subsequent functions (up to the next INIT_CPUFLAGS) is built for the specified cpu. +; You shouldn't need to invoke this macro directly, it's a subroutine for INIT_MMX &co. +%macro INIT_CPUFLAGS 0-2 + %if %0 >= 1 + %xdefine cpuname %1 + %assign cpuflags cpuflags_%1 + %if %0 >= 2 + %xdefine cpuname %1_%2 + %assign cpuflags cpuflags | cpuflags_%2 + %endif + %xdefine SUFFIX _ %+ cpuname + %if cpuflag(AVX) + %assign AVX_enabled 1 + %endif + %if mmsize == 16 && notcpuflag(SSE2) + %define mova movaps + %define movu movups + %define movnta movntps + %endif + %if cpuflag(aligned) + %define movu mova + %elifidn %1, SSE3 + %define movu lddqu + %endif + %else + %xdefine SUFFIX + %undef cpuname + %undef cpuflags + %endif +%endmacro + +; merge MMX and SSE* + +%macro CAT_XDEFINE 3 + %xdefine %1%2 %3 +%endmacro + +%macro CAT_UNDEF 2 + %undef %1%2 +%endmacro + +%macro INIT_MMX 0-1+ + %assign AVX_enabled 0 + %define RESET_MM_PERMUTATION INIT_MMX %1 + %define mmsize 8 + %define num_mmregs 8 + %define mova movq + %define movu movq + %define movh movd + %define movnta movntq + %assign %%i 0 + %rep 8 + CAT_XDEFINE m, %%i, mm %+ %%i + CAT_XDEFINE nmm, %%i, %%i + %assign %%i %%i+1 + %endrep + %rep 8 + CAT_UNDEF m, %%i + CAT_UNDEF nmm, %%i + %assign %%i %%i+1 + %endrep + INIT_CPUFLAGS %1 +%endmacro + +%macro INIT_XMM 0-1+ + %assign AVX_enabled 0 + %define RESET_MM_PERMUTATION INIT_XMM %1 + %define mmsize 16 + %define num_mmregs 8 + %if ARCH_X86_64 + %define num_mmregs 16 + %endif + %define mova movdqa + %define movu movdqu + %define movh movq + %define movnta movntdq + %assign %%i 0 + %rep num_mmregs + CAT_XDEFINE m, %%i, xmm %+ %%i + CAT_XDEFINE nxmm, %%i, %%i + %assign %%i %%i+1 + %endrep + INIT_CPUFLAGS %1 +%endmacro + +%macro INIT_YMM 0-1+ + %assign AVX_enabled 1 + %define RESET_MM_PERMUTATION INIT_YMM %1 + %define mmsize 32 + %define num_mmregs 8 + %if ARCH_X86_64 + %define num_mmregs 16 + %endif + %define mova vmovaps + %define movu vmovups + %undef movh + %define movnta vmovntps + %assign %%i 0 + %rep num_mmregs + CAT_XDEFINE m, %%i, ymm %+ %%i + CAT_XDEFINE nymm, %%i, %%i + %assign %%i %%i+1 + %endrep + INIT_CPUFLAGS %1 +%endmacro + +INIT_XMM + +; I often want to use macros that permute their arguments. e.g. there's no +; efficient way to implement butterfly or transpose or dct without swapping some +; arguments. +; +; I would like to not have to manually keep track of the permutations: +; If I insert a permutation in the middle of a function, it should automatically +; change everything that follows. For more complex macros I may also have multiple +; implementations, e.g. the SSE2 and SSSE3 versions may have different permutations. +; +; Hence these macros. Insert a PERMUTE or some SWAPs at the end of a macro that +; permutes its arguments. It's equivalent to exchanging the contents of the +; registers, except that this way you exchange the register names instead, so it +; doesn't cost any cycles. + +%macro PERMUTE 2-* ; takes a list of pairs to swap +%rep %0/2 + %xdefine tmp%2 m%2 + %xdefine ntmp%2 nm%2 + %rotate 2 +%endrep +%rep %0/2 + %xdefine m%1 tmp%2 + %xdefine nm%1 ntmp%2 + %undef tmp%2 + %undef ntmp%2 + %rotate 2 +%endrep +%endmacro + +%macro SWAP 2-* ; swaps a single chain (sometimes more concise than pairs) +%rep %0-1 +%ifdef m%1 + %xdefine tmp m%1 + %xdefine m%1 m%2 + %xdefine m%2 tmp + CAT_XDEFINE n, m%1, %1 + CAT_XDEFINE n, m%2, %2 +%else + ; If we were called as "SWAP m0,m1" rather than "SWAP 0,1" infer the original numbers here. + ; Be careful using this mode in nested macros though, as in some cases there may be + ; other copies of m# that have already been dereferenced and don't get updated correctly. + %xdefine %%n1 n %+ %1 + %xdefine %%n2 n %+ %2 + %xdefine tmp m %+ %%n1 + CAT_XDEFINE m, %%n1, m %+ %%n2 + CAT_XDEFINE m, %%n2, tmp + CAT_XDEFINE n, m %+ %%n1, %%n1 + CAT_XDEFINE n, m %+ %%n2, %%n2 +%endif + %undef tmp + %rotate 1 +%endrep +%endmacro + +; If SAVE_MM_PERMUTATION is placed at the end of a function, then any later +; calls to that function will automatically load the permutation, so values can +; be returned in mmregs. +%macro SAVE_MM_PERMUTATION 0-1 + %if %0 + %xdefine %%f %1_m + %else + %xdefine %%f current_function %+ _m + %endif + %assign %%i 0 + %rep num_mmregs + CAT_XDEFINE %%f, %%i, m %+ %%i + %assign %%i %%i+1 + %endrep +%endmacro + +%macro LOAD_MM_PERMUTATION 1 ; name to load from + %ifdef %1_m0 + %assign %%i 0 + %rep num_mmregs + CAT_XDEFINE m, %%i, %1_m %+ %%i + CAT_XDEFINE n, m %+ %%i, %%i + %assign %%i %%i+1 + %endrep + %endif +%endmacro + +; Append cpuflags to the callee's name iff the appended name is known and the plain name isn't +%macro call 1 + call_internal %1, %1 %+ SUFFIX +%endmacro +%macro call_internal 2 + %xdefine %%i %1 + %ifndef cglobaled_%1 + %ifdef cglobaled_%2 + %xdefine %%i %2 + %endif + %endif + call %%i + LOAD_MM_PERMUTATION %%i +%endmacro + +; Substitutions that reduce instruction size but are functionally equivalent +%macro add 2 + %ifnum %2 + %if %2==128 + sub %1, -128 + %else + add %1, %2 + %endif + %else + add %1, %2 + %endif +%endmacro + +%macro sub 2 + %ifnum %2 + %if %2==128 + add %1, -128 + %else + sub %1, %2 + %endif + %else + sub %1, %2 + %endif +%endmacro + +;============================================================================= +; AVX abstraction layer +;============================================================================= + +%assign i 0 +%rep 16 + %if i < 8 + CAT_XDEFINE sizeofmm, i, 8 + %endif + CAT_XDEFINE sizeofxmm, i, 16 + CAT_XDEFINE sizeofymm, i, 32 +%assign i i+1 +%endrep +%undef i + +%macro CHECK_AVX_INSTR_EMU 3-* + %xdefine %%opcode %1 + %xdefine %%dst %2 + %rep %0-2 + %ifidn %%dst, %3 + %error non-AVX emulation of ``%%opcode'' is not supported + %endif + %rotate 1 + %endrep +%endmacro + +;%1 == instruction +;%2 == 1 if float, 0 if int +;%3 == 1 if 4-operand (xmm, xmm, xmm, imm), 0 if 2- or 3-operand (xmm, xmm, xmm) +;%4 == number of operands given +;%5+: operands +%macro RUN_AVX_INSTR 6-7+ + %ifid %6 + %define %%sizeofreg sizeof%6 + %elifid %5 + %define %%sizeofreg sizeof%5 + %else + %define %%sizeofreg mmsize + %endif + %if %%sizeofreg==32 + %if %4>=3 + v%1 %5, %6, %7 + %else + v%1 %5, %6 + %endif + %else + %if %%sizeofreg==8 + %define %%regmov movq + %elif %2 + %define %%regmov movaps + %else + %define %%regmov movdqa + %endif + + %if %4>=3+%3 + %ifnidn %5, %6 + %if AVX_enabled && %%sizeofreg==16 + v%1 %5, %6, %7 + %else + CHECK_AVX_INSTR_EMU {%1 %5, %6, %7}, %5, %7 + %%regmov %5, %6 + %1 %5, %7 + %endif + %else + %1 %5, %7 + %endif + %elif %4>=3 + %1 %5, %6, %7 + %else + %1 %5, %6 + %endif + %endif +%endmacro + +; 3arg AVX ops with a memory arg can only have it in src2, +; whereas SSE emulation of 3arg prefers to have it in src1 (i.e. the mov). +; So, if the op is symmetric and the wrong one is memory, swap them. +%macro RUN_AVX_INSTR1 8 + %assign %%swap 0 + %if AVX_enabled + %ifnid %6 + %assign %%swap 1 + %endif + %elifnidn %5, %6 + %ifnid %7 + %assign %%swap 1 + %endif + %endif + %if %%swap && %3 == 0 && %8 == 1 + RUN_AVX_INSTR %1, %2, %3, %4, %5, %7, %6 + %else + RUN_AVX_INSTR %1, %2, %3, %4, %5, %6, %7 + %endif +%endmacro + +;%1 == instruction +;%2 == 1 if float, 0 if int +;%3 == 1 if 4-operand (xmm, xmm, xmm, imm), 0 if 2- or 3-operand (xmm, xmm, xmm) +;%4 == 1 if symmetric (i.e. doesn't matter which src arg is which), 0 if not +%macro AVX_INSTR 4 + %macro %1 2-9 fnord, fnord, fnord, %1, %2, %3, %4 + %ifidn %3, fnord + RUN_AVX_INSTR %6, %7, %8, 2, %1, %2 + %elifidn %4, fnord + RUN_AVX_INSTR1 %6, %7, %8, 3, %1, %2, %3, %9 + %elifidn %5, fnord + RUN_AVX_INSTR %6, %7, %8, 4, %1, %2, %3, %4 + %else + RUN_AVX_INSTR %6, %7, %8, 5, %1, %2, %3, %4, %5 + %endif + %endmacro +%endmacro + +AVX_INSTR addpd, 1, 0, 1 +AVX_INSTR addps, 1, 0, 1 +AVX_INSTR addsd, 1, 0, 1 +AVX_INSTR addss, 1, 0, 1 +AVX_INSTR addsubpd, 1, 0, 0 +AVX_INSTR addsubps, 1, 0, 0 +AVX_INSTR andpd, 1, 0, 1 +AVX_INSTR andps, 1, 0, 1 +AVX_INSTR andnpd, 1, 0, 0 +AVX_INSTR andnps, 1, 0, 0 +AVX_INSTR blendpd, 1, 0, 0 +AVX_INSTR blendps, 1, 0, 0 +AVX_INSTR blendvpd, 1, 0, 0 +AVX_INSTR blendvps, 1, 0, 0 +AVX_INSTR cmppd, 1, 0, 0 +AVX_INSTR cmpps, 1, 0, 0 +AVX_INSTR cmpsd, 1, 0, 0 +AVX_INSTR cmpss, 1, 0, 0 +AVX_INSTR cvtdq2ps, 1, 0, 0 +AVX_INSTR cvtps2dq, 1, 0, 0 +AVX_INSTR divpd, 1, 0, 0 +AVX_INSTR divps, 1, 0, 0 +AVX_INSTR divsd, 1, 0, 0 +AVX_INSTR divss, 1, 0, 0 +AVX_INSTR dppd, 1, 1, 0 +AVX_INSTR dpps, 1, 1, 0 +AVX_INSTR haddpd, 1, 0, 0 +AVX_INSTR haddps, 1, 0, 0 +AVX_INSTR hsubpd, 1, 0, 0 +AVX_INSTR hsubps, 1, 0, 0 +AVX_INSTR maxpd, 1, 0, 1 +AVX_INSTR maxps, 1, 0, 1 +AVX_INSTR maxsd, 1, 0, 1 +AVX_INSTR maxss, 1, 0, 1 +AVX_INSTR minpd, 1, 0, 1 +AVX_INSTR minps, 1, 0, 1 +AVX_INSTR minsd, 1, 0, 1 +AVX_INSTR minss, 1, 0, 1 +AVX_INSTR movhlps, 1, 0, 0 +AVX_INSTR movlhps, 1, 0, 0 +AVX_INSTR movsd, 1, 0, 0 +AVX_INSTR movss, 1, 0, 0 +AVX_INSTR mpsadbw, 0, 1, 0 +AVX_INSTR mulpd, 1, 0, 1 +AVX_INSTR mulps, 1, 0, 1 +AVX_INSTR mulsd, 1, 0, 1 +AVX_INSTR mulss, 1, 0, 1 +AVX_INSTR orpd, 1, 0, 1 +AVX_INSTR orps, 1, 0, 1 +AVX_INSTR pabsb, 0, 0, 0 +AVX_INSTR pabsw, 0, 0, 0 +AVX_INSTR pabsd, 0, 0, 0 +AVX_INSTR packsswb, 0, 0, 0 +AVX_INSTR packssdw, 0, 0, 0 +AVX_INSTR packuswb, 0, 0, 0 +AVX_INSTR packusdw, 0, 0, 0 +AVX_INSTR paddb, 0, 0, 1 +AVX_INSTR paddw, 0, 0, 1 +AVX_INSTR paddd, 0, 0, 1 +AVX_INSTR paddq, 0, 0, 1 +AVX_INSTR paddsb, 0, 0, 1 +AVX_INSTR paddsw, 0, 0, 1 +AVX_INSTR paddusb, 0, 0, 1 +AVX_INSTR paddusw, 0, 0, 1 +AVX_INSTR palignr, 0, 1, 0 +AVX_INSTR pand, 0, 0, 1 +AVX_INSTR pandn, 0, 0, 0 +AVX_INSTR pavgb, 0, 0, 1 +AVX_INSTR pavgw, 0, 0, 1 +AVX_INSTR pblendvb, 0, 0, 0 +AVX_INSTR pblendw, 0, 1, 0 +AVX_INSTR pcmpestri, 0, 0, 0 +AVX_INSTR pcmpestrm, 0, 0, 0 +AVX_INSTR pcmpistri, 0, 0, 0 +AVX_INSTR pcmpistrm, 0, 0, 0 +AVX_INSTR pcmpeqb, 0, 0, 1 +AVX_INSTR pcmpeqw, 0, 0, 1 +AVX_INSTR pcmpeqd, 0, 0, 1 +AVX_INSTR pcmpeqq, 0, 0, 1 +AVX_INSTR pcmpgtb, 0, 0, 0 +AVX_INSTR pcmpgtw, 0, 0, 0 +AVX_INSTR pcmpgtd, 0, 0, 0 +AVX_INSTR pcmpgtq, 0, 0, 0 +AVX_INSTR phaddw, 0, 0, 0 +AVX_INSTR phaddd, 0, 0, 0 +AVX_INSTR phaddsw, 0, 0, 0 +AVX_INSTR phsubw, 0, 0, 0 +AVX_INSTR phsubd, 0, 0, 0 +AVX_INSTR phsubsw, 0, 0, 0 +AVX_INSTR pmaddwd, 0, 0, 1 +AVX_INSTR pmaddubsw, 0, 0, 0 +AVX_INSTR pmaxsb, 0, 0, 1 +AVX_INSTR pmaxsw, 0, 0, 1 +AVX_INSTR pmaxsd, 0, 0, 1 +AVX_INSTR pmaxub, 0, 0, 1 +AVX_INSTR pmaxuw, 0, 0, 1 +AVX_INSTR pmaxud, 0, 0, 1 +AVX_INSTR pminsb, 0, 0, 1 +AVX_INSTR pminsw, 0, 0, 1 +AVX_INSTR pminsd, 0, 0, 1 +AVX_INSTR pminub, 0, 0, 1 +AVX_INSTR pminuw, 0, 0, 1 +AVX_INSTR pminud, 0, 0, 1 +AVX_INSTR pmovmskb, 0, 0, 0 +AVX_INSTR pmulhuw, 0, 0, 1 +AVX_INSTR pmulhrsw, 0, 0, 1 +AVX_INSTR pmulhw, 0, 0, 1 +AVX_INSTR pmullw, 0, 0, 1 +AVX_INSTR pmulld, 0, 0, 1 +AVX_INSTR pmuludq, 0, 0, 1 +AVX_INSTR pmuldq, 0, 0, 1 +AVX_INSTR por, 0, 0, 1 +AVX_INSTR psadbw, 0, 0, 1 +AVX_INSTR pshufb, 0, 0, 0 +AVX_INSTR pshufd, 0, 1, 0 +AVX_INSTR pshufhw, 0, 1, 0 +AVX_INSTR pshuflw, 0, 1, 0 +AVX_INSTR psignb, 0, 0, 0 +AVX_INSTR psignw, 0, 0, 0 +AVX_INSTR psignd, 0, 0, 0 +AVX_INSTR psllw, 0, 0, 0 +AVX_INSTR pslld, 0, 0, 0 +AVX_INSTR psllq, 0, 0, 0 +AVX_INSTR pslldq, 0, 0, 0 +AVX_INSTR psraw, 0, 0, 0 +AVX_INSTR psrad, 0, 0, 0 +AVX_INSTR psrlw, 0, 0, 0 +AVX_INSTR psrld, 0, 0, 0 +AVX_INSTR psrlq, 0, 0, 0 +AVX_INSTR psrldq, 0, 0, 0 +AVX_INSTR psubb, 0, 0, 0 +AVX_INSTR psubw, 0, 0, 0 +AVX_INSTR psubd, 0, 0, 0 +AVX_INSTR psubq, 0, 0, 0 +AVX_INSTR psubsb, 0, 0, 0 +AVX_INSTR psubsw, 0, 0, 0 +AVX_INSTR psubusb, 0, 0, 0 +AVX_INSTR psubusw, 0, 0, 0 +AVX_INSTR ptest, 0, 0, 0 +AVX_INSTR punpckhbw, 0, 0, 0 +AVX_INSTR punpckhwd, 0, 0, 0 +AVX_INSTR punpckhdq, 0, 0, 0 +AVX_INSTR punpckhqdq, 0, 0, 0 +AVX_INSTR punpcklbw, 0, 0, 0 +AVX_INSTR punpcklwd, 0, 0, 0 +AVX_INSTR punpckldq, 0, 0, 0 +AVX_INSTR punpcklqdq, 0, 0, 0 +AVX_INSTR pxor, 0, 0, 1 +AVX_INSTR shufps, 1, 1, 0 +AVX_INSTR subpd, 1, 0, 0 +AVX_INSTR subps, 1, 0, 0 +AVX_INSTR subsd, 1, 0, 0 +AVX_INSTR subss, 1, 0, 0 +AVX_INSTR unpckhpd, 1, 0, 0 +AVX_INSTR unpckhps, 1, 0, 0 +AVX_INSTR unpcklpd, 1, 0, 0 +AVX_INSTR unpcklps, 1, 0, 0 +AVX_INSTR xorpd, 1, 0, 1 +AVX_INSTR xorps, 1, 0, 1 + +; 3DNow instructions, for sharing code between AVX, SSE and 3DN +AVX_INSTR pfadd, 1, 0, 1 +AVX_INSTR pfsub, 1, 0, 0 +AVX_INSTR pfmul, 1, 0, 1 + +; base-4 constants for shuffles +%assign i 0 +%rep 256 + %assign j ((i>>6)&3)*1000 + ((i>>4)&3)*100 + ((i>>2)&3)*10 + (i&3) + %if j < 10 + CAT_XDEFINE q000, j, i + %elif j < 100 + CAT_XDEFINE q00, j, i + %elif j < 1000 + CAT_XDEFINE q0, j, i + %else + CAT_XDEFINE q, j, i + %endif +%assign i i+1 +%endrep +%undef i +%undef j + +%macro FMA_INSTR 3 + %macro %1 4-7 %1, %2, %3 + %if cpuflag(xop) + v%5 %1, %2, %3, %4 + %else + %6 %1, %2, %3 + %7 %1, %4 + %endif + %endmacro +%endmacro + +FMA_INSTR pmacsdd, pmulld, paddd +FMA_INSTR pmacsww, pmullw, paddw +FMA_INSTR pmadcswd, pmaddwd, paddd + +; tzcnt is equivalent to "rep bsf" and is backwards-compatible with bsf. +; This lets us use tzcnt without bumping the yasm version requirement yet. +%define tzcnt rep bsf |