diff --git a/README.chromium b/README.chromium index ec25d3903..95f46b384 100644 --- a/README.chromium +++ b/README.chromium @@ -1,6 +1,6 @@ Name: libyuv URL: http://code.google.com/p/libyuv/ -Version: 286 +Version: 287 License: BSD License File: LICENSE diff --git a/include/libyuv/version.h b/include/libyuv/version.h index 6ae848321..528de2bf8 100644 --- a/include/libyuv/version.h +++ b/include/libyuv/version.h @@ -11,7 +11,7 @@ #ifndef INCLUDE_LIBYUV_VERSION_H_ #define INCLUDE_LIBYUV_VERSION_H_ -#define LIBYUV_VERSION 286 +#define LIBYUV_VERSION 287 #endif // INCLUDE_LIBYUV_VERSION_H_ diff --git a/source/row_common.cc b/source/row_common.cc index 080c54895..85cb70d3b 100644 --- a/source/row_common.cc +++ b/source/row_common.cc @@ -588,76 +588,166 @@ void UYVYToYRow_C(const uint8* src_yuy2, uint8* dst_y, int width) { } } -#define BLENDER(f, b, a) (((256 - a) * b) >> 8) + f +#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) { for (int 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]; - if (a == 0) { - *reinterpret_cast(dst_argb) = - *reinterpret_cast(src_argb1); - } else if (a == 255) { - *reinterpret_cast(dst_argb) = - *reinterpret_cast(src_argb0); - } else { - const uint32 fb = src_argb0[0]; - const uint32 fg = src_argb0[1]; - const uint32 fr = src_argb0[2]; - const uint32 bb = src_argb1[0]; - const uint32 bg = src_argb1[1]; - const uint32 br = src_argb1[2]; - dst_argb[0] = BLENDER(fb, bb, a); - dst_argb[1] = BLENDER(fg, bg, a); - dst_argb[2] = BLENDER(fr, br, a); - dst_argb[3] = 255u; - } + 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]; - if (a == 0) { - *reinterpret_cast(dst_argb + 4) = - *reinterpret_cast(src_argb1 + 4); - } else if (a == 255) { - *reinterpret_cast(dst_argb + 4) = - *reinterpret_cast(src_argb0 + 4); - } else { - const uint32 fb = src_argb0[4 + 0]; - const uint32 fg = src_argb0[4 + 1]; - const uint32 fr = src_argb0[4 + 2]; - const uint32 bb = src_argb1[4 + 0]; - const uint32 bg = src_argb1[4 + 1]; - const uint32 br = src_argb1[4 + 2]; - dst_argb[4 + 0] = BLENDER(fb, bb, a); - dst_argb[4 + 1] = BLENDER(fg, bg, a); - dst_argb[4 + 2] = BLENDER(fr, br, a); - dst_argb[4 + 3] = 255u; - } + 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]; - if (a == 0) { - *reinterpret_cast(dst_argb) = - *reinterpret_cast(src_argb1); - } else if (a == 255) { - *reinterpret_cast(dst_argb) = - *reinterpret_cast(src_argb0); - } else { - const uint32 fb = src_argb0[0]; - const uint32 fg = src_argb0[1]; - const uint32 fr = src_argb0[2]; - const uint32 bb = src_argb1[0]; - const uint32 bg = src_argb1[1]; - const uint32 br = src_argb1[2]; - dst_argb[0] = BLENDER(fb, bb, a); - dst_argb[1] = BLENDER(fg, bg, a); - dst_argb[2] = BLENDER(fr, br, a); - dst_argb[3] = 255u; + 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) { + for (int 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 +// 8.16 fixed point inverse table +#define T(a) 0x10000 / a +uint32 fixed_invtbl8[256] = { + 0x0100, T(0x01), 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), 0x0100 }; +#undef T + +void ARGBUnattenuateRow_C(const uint8* src_argb, uint8* dst_argb, int width) { + for (int 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]; + if (a) { + const uint32 ia = fixed_invtbl8[a]; // 8.16 fixed point + b = (b * ia) >> 8; + g = (g * ia) >> 8; + r = (r * ia) >> 8; + // Clamping should not be necessary but is free in assembly. + if (b > 255) { + b = 255; + } + if (g > 255) { + g = 255; + } + if (r > 255) { + r = 255; + } } + dst_argb[0] = b; + dst_argb[1] = g; + dst_argb[2] = r; + dst_argb[3] = a; + src_argb += 4; + dst_argb += 4; } } @@ -757,115 +847,6 @@ UVANY(UYVYToUVRow_Any_SSE2, UYVYToUVRow_Unaligned_SSE2, UYVYToUVRow_C, 2) #undef UVANY #endif -// Multiply source RGB by alpha and store to destination. -// b = (b * a + 127) / 255; -void ARGBAttenuateRow_C(const uint8* src_argb, uint8* dst_argb, int width) { - for (int 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] = (b * a + 255) >> 8; - dst_argb[1] = (g * a + 255) >> 8; - dst_argb[2] = (r * a + 255) >> 8; - dst_argb[3] = a; - b = src_argb[4]; - g = src_argb[5]; - r = src_argb[6]; - a = src_argb[7]; - dst_argb[4] = (b * a + 255) >> 8; - dst_argb[5] = (g * a + 255) >> 8; - dst_argb[6] = (r * a + 255) >> 8; - 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] = (b * a + 255) >> 8; - dst_argb[1] = (g * a + 255) >> 8; - dst_argb[2] = (r * a + 255) >> 8; - dst_argb[3] = a; - } -} - -// 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 -// 8.16 fixed point inverse table -#define T(a) 0x10000 / a -uint32 fixed_invtbl8[256] = { - 0x0100, T(0x01), 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), 0x0100 }; -#undef T - -void ARGBUnattenuateRow_C(const uint8* src_argb, uint8* dst_argb, int width) { - for (int 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]; - if (a) { - const uint32 ia = fixed_invtbl8[a]; // 8.16 fixed point - b = (b * ia) >> 8; - g = (g * ia) >> 8; - r = (r * ia) >> 8; - // Clamping should not be necessary but is free in assembly. - if (b > 255) { - b = 255; - } - if (g > 255) { - g = 255; - } - if (r > 255) { - r = 255; - } - } - dst_argb[0] = b; - dst_argb[1] = g; - dst_argb[2] = 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}; diff --git a/source/row_posix.cc b/source/row_posix.cc index 4d071fc61..e69779d46 100644 --- a/source/row_posix.cc +++ b/source/row_posix.cc @@ -2375,60 +2375,39 @@ void ARGBBlendRow_SSE2(const uint8* src_argb0, const uint8* src_argb1, "add $1-4,%3 \n" "jl 49f \n" - // 8 pixel loop. + // 4 pixel loop. ".p2align 2 \n" "41: \n" "movdqu (%0),%%xmm3 \n" + "lea 0x10(%0),%0 \n" "movdqa %%xmm3,%%xmm0 \n" "pxor %%xmm4,%%xmm3 \n" + "movdqu (%1),%%xmm2 \n" "psrlw $0x8,%%xmm3 \n" "pshufhw $0xf5,%%xmm3,%%xmm3 \n" "pshuflw $0xf5,%%xmm3,%%xmm3 \n" - "movdqu (%1),%%xmm2 \n" "pand %%xmm6,%%xmm2 \n" "paddw %%xmm7,%%xmm3 \n" "pmullw %%xmm3,%%xmm2 \n" "movdqu (%1),%%xmm1 \n" + "lea 0x10(%1),%1 \n" "psrlw $0x8,%%xmm1 \n" "por %%xmm4,%%xmm0 \n" "pmullw %%xmm3,%%xmm1 \n" - "movdqu 0x10(%0),%%xmm3 \n" - "lea 0x20(%0),%0 \n" "psrlw $0x8,%%xmm2 \n" "paddusb %%xmm2,%%xmm0 \n" "pand %%xmm5,%%xmm1 \n" "paddusb %%xmm1,%%xmm0 \n" "sub $0x4,%3 \n" "movdqa %%xmm0,(%2) \n" - "jl 49f \n" - "movdqa %%xmm3,%%xmm0 \n" - "pxor %%xmm4,%%xmm3 \n" - "movdqu 0x10(%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 0x10(%1),%%xmm1 \n" - "lea 0x20(%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" - "sub $0x4,%3 \n" - "movdqa %%xmm0,0x10(%2) \n" - "lea 0x20(%2),%2 \n" + "lea 0x10(%2),%2 \n" "jge 41b \n" "49: \n" "add $0x3,%3 \n" "jl 99f \n" - // 1 pixel loop. + // 1 pixel loop. "91: \n" "movd (%0),%%xmm3 \n" "lea 0x4(%0),%0 \n" @@ -2531,56 +2510,37 @@ void ARGBBlendRow_SSSE3(const uint8* src_argb0, const uint8* src_argb1, "add $1-4,%3 \n" "jl 49f \n" - // 8 pixel loop. + // 4 pixel loop. ".p2align 2 \n" "41: \n" "movdqu (%0),%%xmm3 \n" + "lea 0x10(%0),%0 \n" "movdqa %%xmm3,%%xmm0 \n" "pxor %%xmm4,%%xmm3 \n" - "pshufb %4,%%xmm3 \n" "movdqu (%1),%%xmm2 \n" + "pshufb %4,%%xmm3 \n" "pand %%xmm6,%%xmm2 \n" "paddw %%xmm7,%%xmm3 \n" "pmullw %%xmm3,%%xmm2 \n" "movdqu (%1),%%xmm1 \n" + "lea 0x10(%1),%1 \n" "psrlw $0x8,%%xmm1 \n" "por %%xmm4,%%xmm0 \n" "pmullw %%xmm3,%%xmm1 \n" - "movdqu 0x10(%0),%%xmm3 \n" - "lea 0x20(%0),%0 \n" "psrlw $0x8,%%xmm2 \n" "paddusb %%xmm2,%%xmm0 \n" "pand %%xmm5,%%xmm1 \n" "paddusb %%xmm1,%%xmm0 \n" "sub $0x4,%3 \n" "movdqa %%xmm0,(%2) \n" - "jl 49f \n" - "movdqa %%xmm3,%%xmm0 \n" - "pxor %%xmm4,%%xmm3 \n" - "movdqu 0x10(%1),%%xmm2 \n" - "pshufb %4,%%xmm3 \n" - "pand %%xmm6,%%xmm2 \n" - "paddw %%xmm7,%%xmm3 \n" - "pmullw %%xmm3,%%xmm2 \n" - "movdqu 0x10(%1),%%xmm1 \n" - "lea 0x20(%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" - "sub $0x4,%3 \n" - "movdqa %%xmm0,0x10(%2) \n" - "lea 0x20(%2),%2 \n" + "lea 0x10(%2),%2 \n" "jge 41b \n" "49: \n" "add $0x3,%3 \n" "jl 99f \n" - // 1 pixel loop. + // 1 pixel loop. "91: \n" "movd (%0),%%xmm3 \n" "lea 0x4(%0),%0 \n" @@ -2629,7 +2589,7 @@ void ARGBAttenuateRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width) { "pcmpeqb %%xmm5,%%xmm5 \n" "psrld $0x8,%%xmm5 \n" - // 4 pixel loop + // 4 pixel loop ".p2align 4 \n" "1: \n" "movdqa (%0),%%xmm0 \n" diff --git a/source/row_win.cc b/source/row_win.cc index 832c4fd78..6a710672a 100644 --- a/source/row_win.cc +++ b/source/row_win.cc @@ -2474,54 +2474,31 @@ void ARGBBlendRow_SSE2(const uint8* src_argb0, const uint8* src_argb1, add ecx, 1 - 4 jl convertloop4b - // 8 pixel loop. - align 4 + // 4 pixel loop. convertloop4: - movdqu xmm3, [eax] + 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 - movdqu xmm2, [esi] // _r_b 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 - movdqu xmm3, [eax + 16] - lea eax, [eax + 32] 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 sub ecx, 4 movdqa [edx], xmm0 - jl convertloop4b - - movdqa xmm0, xmm3 // src argb - pxor xmm3, xmm4 // ~alpha - movdqu xmm2, [esi + 16] // _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 + 16] // _a_g - lea esi, [esi + 32] - 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 - sub ecx, 4 - movdqa [edx + 16], xmm0 - lea edx, [edx + 32] + lea edx, [edx + 16] jge convertloop4 convertloop4b: @@ -2530,7 +2507,7 @@ void ARGBBlendRow_SSE2(const uint8* src_argb0, const uint8* src_argb1, // 1 pixel loop. convertloop1: - movd xmm3, [eax] + movd xmm3, [eax] // src argb lea eax, [eax + 4] movdqa xmm0, xmm3 // src argb pxor xmm3, xmm4 // ~alpha @@ -2629,50 +2606,29 @@ void ARGBBlendRow_SSSE3(const uint8* src_argb0, const uint8* src_argb1, add ecx, 1 - 4 jl convertloop4b - // 8 pixel loop. - align 4 + // 4 pixel loop. convertloop4: - movdqu xmm3, [eax] + movdqu xmm3, [eax] // src argb + lea eax, [eax + 16] movdqa xmm0, xmm3 // src argb pxor xmm3, xmm4 // ~alpha - pshufb xmm3, kShuffleAlpha // 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 - movdqu xmm3, [eax + 16] - lea eax, [eax + 32] 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 sub ecx, 4 movdqa [edx], xmm0 - jl convertloop4b - - movdqa xmm0, xmm3 // src argb - pxor xmm3, xmm4 // ~alpha - movdqu xmm2, [esi + 16] // _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 + 16] // _a_g - lea esi, [esi + 32] - 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 - sub ecx, 4 - movdqa [edx + 16], xmm0 - lea edx, [edx + 32] + lea edx, [edx + 16] jge convertloop4 convertloop4b: @@ -2681,7 +2637,7 @@ void ARGBBlendRow_SSSE3(const uint8* src_argb0, const uint8* src_argb1, // 1 pixel loop. convertloop1: - movd xmm3, [eax] + movd xmm3, [eax] // src argb lea eax, [eax + 4] movdqa xmm0, xmm3 // src argb pxor xmm3, xmm4 // ~alpha