coffee/libyuv
1
0
Fork 0
mirror of https://chromium.googlesource.com/libyuv/libyuv synced 2025-12-07 17:26:49 +08:00
Code Issues Packages Projects Releases Wiki Activity

Remove sse2 functions that also have ssse3

Browse Source 
ARGBBlendRow_SSE2, ARGBAttenuateRow_SSE2, and MirrorRow_SSE2
Since vast majority of CPUs have SSSE3 now, removing the SSE2
improves the performance of CPU dispatching.

R=harryjin@google.com
BUG=none

Review URL: https://codereview.chromium.org/1377053003 .
This commit is contained in:
Frank Barchard 2015-09-30 14:24:44 -07:00
parent d039ad6e9b
commit 2cc1a2b233
9 changed files with 16 additions and 413 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
README.chromium
View File

@ -1,6 +1,6 @@
Name: libyuv
URL: http://code.google.com/p/libyuv/
Version: 1495
Version: 1496
License: BSD
License File: LICENSE

18
include/libyuv/row.h
View File

@ -121,7 +121,6 @@ extern "C" {
#define HAS_H422TOARGBROW_SSSE3
#define HAS_H422TOABGRROW_SSSE3
#define HAS_MERGEUVROW_SSE2
#define HAS_MIRRORROW_SSE2
#define HAS_MIRRORROW_SSSE3
#define HAS_MIRRORROW_UV_SSSE3
#define HAS_MIRRORUVROW_SSSE3
@ -181,8 +180,7 @@ extern "C" {
#define HAS_SOBELYROW_SSE2
#endif
// The following are available on x64 Visual C and clangcl.
// TODO(fbarchard): Port to gcc.
// The following are also available on x64 Visual C.
#if !defined(LIBYUV_DISABLE_X86) && defined (_M_X64) && \
(!defined(__clang__) || defined(__SSSE3__))
#define HAS_I422TOARGBROW_SSSE3
@ -262,16 +260,6 @@ extern "C" {
#define HAS_ARGBUNATTENUATEROW_AVX2
#endif
// The following are disabled when SSSE3 is available:
// TODO(fbarchard): remove sse2. ssse3 is faster and well supported.
#if !defined(LIBYUV_DISABLE_X86) && \
(defined(_M_IX86) || defined(__x86_64__) || defined(__i386__)) && \
!defined(LIBYUV_SSSE3_ONLY)
#define HAS_ARGBATTENUATEROW_SSE2
#define HAS_ARGBBLENDROW_SSE2
#define HAS_MIRRORROW_SSE2
#endif
// The following are available on Neon platforms:
#if !defined(LIBYUV_DISABLE_NEON) && \
(defined(__aarch64__) || defined(__ARM_NEON__) || defined(LIBYUV_NEON))
@ -822,7 +810,6 @@ void ARGBToUVJ422Row_C(const uint8* src_argb,
void MirrorRow_AVX2(const uint8* src, uint8* dst, int width);
void MirrorRow_SSSE3(const uint8* src, uint8* dst, int width);
void MirrorRow_SSE2(const uint8* src, uint8* dst, int width);
void MirrorRow_NEON(const uint8* src, uint8* dst, int width);
void MirrorRow_MIPS_DSPR2(const uint8* src, uint8* dst, int width);
void MirrorRow_C(const uint8* src, uint8* dst, int width);
@ -1620,8 +1607,6 @@ void I400ToARGBRow_Any_NEON(const uint8* src_y, uint8* dst_argb, int width);
// ARGB preattenuated alpha blend.
void ARGBBlendRow_SSSE3(const uint8* src_argb, const uint8* src_argb1,
uint8* dst_argb, int width);
void ARGBBlendRow_SSE2(const uint8* src_argb, const uint8* src_argb1,
uint8* dst_argb, int width);
void ARGBBlendRow_NEON(const uint8* src_argb, const uint8* src_argb1,
uint8* dst_argb, int width);
void ARGBBlendRow_C(const uint8* src_argb, const uint8* src_argb1,
@ -1941,7 +1926,6 @@ void I422ToUYVYRow_Any_NEON(const uint8* src_y,
// Effects related row functions.
void ARGBAttenuateRow_C(const uint8* src_argb, uint8* dst_argb, int width);
void ARGBAttenuateRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width);
void ARGBAttenuateRow_SSSE3(const uint8* src_argb, uint8* dst_argb, int width);
void ARGBAttenuateRow_AVX2(const uint8* src_argb, uint8* dst_argb, int width);
void ARGBAttenuateRow_NEON(const uint8* src_argb, uint8* dst_argb, int width);

2
include/libyuv/version.h
View File

@ -11,6 +11,6 @@
#ifndef INCLUDE_LIBYUV_VERSION_H_ // NOLINT
#define INCLUDE_LIBYUV_VERSION_H_
#define LIBYUV_VERSION 1495
#define LIBYUV_VERSION 1496
#endif // INCLUDE_LIBYUV_VERSION_H_ NOLINT

22
source/convert_argb.cc
View File

@ -390,14 +390,6 @@ int I420AlphaToARGB(const uint8* src_y, int src_stride_y,
I422AlphaToARGBRow = I422AlphaToARGBRow_MIPS_DSPR2;
}
#endif
#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;
@ -424,7 +416,8 @@ int I420AlphaToARGB(const uint8* src_y, int src_stride_y,
#endif
for (y = 0; y < height; ++y) {
I422AlphaToARGBRow(src_y, src_u, src_v, src_a, dst_argb, &kYuvConstants, width);
I422AlphaToARGBRow(src_y, src_u, src_v, src_a, dst_argb, &kYuvConstants,
width);
if (attenuate) {
ARGBAttenuateRow(dst_argb, dst_argb, width);
}
@ -500,14 +493,6 @@ int I420AlphaToABGR(const uint8* src_y, int src_stride_y,
I422AlphaToABGRRow = I422AlphaToABGRRow_MIPS_DSPR2;
}
#endif
#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;
@ -534,7 +519,8 @@ int I420AlphaToABGR(const uint8* src_y, int src_stride_y,
#endif
for (y = 0; y < height; ++y) {
I422AlphaToABGRRow(src_y, src_u, src_v, src_a, dst_abgr, &kYuvConstants, width);
I422AlphaToABGRRow(src_y, src_u, src_v, src_a, dst_abgr, &kYuvConstants,
width);
if (attenuate) {
ARGBAttenuateRow(dst_abgr, dst_abgr, width);
}

21
source/planar_functions.cc
View File

@ -237,14 +237,6 @@ void MirrorPlane(const uint8* src_y, int src_stride_y,
}
}
#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;
@ -541,11 +533,6 @@ ARGBBlendRow GetARGBBlend() {
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;
@ -1267,14 +1254,6 @@ int ARGBAttenuate(const uint8* src_argb, int src_stride_argb,
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;

8
source/rotate.cc
View File

@ -117,14 +117,6 @@ void RotatePlane180(const uint8* src, int src_stride,
}
}
#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;

6
source/row_any.cc
View File

@ -443,9 +443,6 @@ ANY11(ARGB4444ToARGBRow_Any_NEON, ARGB4444ToARGBRow_NEON, 0, 2, 4, 7)
#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
@ -617,9 +614,6 @@ ANY11M(MirrorRow_Any_AVX2, MirrorRow_AVX2, 1, 31)
#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

187
source/row_gcc.cc
View File

@ -1679,7 +1679,7 @@ void OMITFP I422AlphaToARGBRow_SSSE3(const uint8* y_buf,
[width]"+m"(width) // %[width]
#else
[width]"+rm"(width) // %[width]
#endif
#endif
: [yuvconstants]"r"(yuvconstants) // %[yuvconstants]
: "memory", "cc", NACL_R14
"xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
@ -1711,7 +1711,7 @@ void OMITFP I422AlphaToABGRRow_SSSE3(const uint8* y_buf,
[width]"+m"(width) // %[width]
#else
[width]"+rm"(width) // %[width]
#endif
#endif
: [yuvconstants]"r"(yuvconstants) // %[yuvconstants]
: "memory", "cc", NACL_R14
"xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
@ -1987,8 +1987,8 @@ void OMITFP I422ToRGBARow_SSSE3(const uint8* y_buf,
"vpunpcklwd %%ymm2,%%ymm0,%%ymm1 \n" \
"vpunpckhwd %%ymm2,%%ymm0,%%ymm0 \n" \
"vmovdqu %%ymm1," MEMACCESS([dst_argb]) " \n" \
"vmovdqu %%ymm0," MEMACCESS2(0x20,[dst_argb]) " \n" \
"lea " MEMLEA(0x40,[dst_argb]) ",%[dst_argb] \n"
"vmovdqu %%ymm0," MEMACCESS2(0x20, [dst_argb]) " \n" \
"lea " MEMLEA(0x40, [dst_argb]) ", %[dst_argb] \n"
// Store 16 ABGR values.
#define STOREABGR_AVX2 \
@ -1999,8 +1999,8 @@ void OMITFP I422ToRGBARow_SSSE3(const uint8* y_buf,
"vpunpcklwd %%ymm2,%%ymm1,%%ymm0 \n" \
"vpunpckhwd %%ymm2,%%ymm1,%%ymm1 \n" \
"vmovdqu %%ymm0," MEMACCESS([dst_abgr]) " \n" \
"vmovdqu %%ymm1," MEMACCESS2(0x20,[dst_abgr]) " \n" \
"lea " MEMLEA(0x40,[dst_abgr]) ",%[dst_abgr] \n"
"vmovdqu %%ymm1," MEMACCESS2(0x20, [dst_abgr]) " \n" \
"lea " MEMLEA(0x40, [dst_abgr]) ", %[dst_abgr] \n"
#if defined(HAS_I422TOBGRAROW_AVX2)
// 16 pixels
@ -2026,7 +2026,6 @@ void OMITFP I422ToBGRARow_AVX2(const uint8* y_buf,
"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"
@ -2106,7 +2105,7 @@ void OMITFP I422AlphaToARGBRow_AVX2(const uint8* y_buf,
[width]"+m"(width) // %[width]
#else
[width]"+rm"(width) // %[width]
#endif
#endif
: [yuvconstants]"r"(yuvconstants) // %[yuvconstants]
: "memory", "cc", NACL_R14
"xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
@ -2143,7 +2142,7 @@ void OMITFP I422AlphaToABGRRow_AVX2(const uint8* y_buf,
[width]"+m"(width) // %[width]
#else
[width]"+rm"(width) // %[width]
#endif
#endif
: [yuvconstants]"r"(yuvconstants) // %[yuvconstants]
: "memory", "cc", NACL_R14
"xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
@ -2233,7 +2232,6 @@ void OMITFP NV12ToARGBRow_AVX2(const uint8* y_buf,
uint8* dst_argb,
struct YuvConstants* yuvconstants,
int width) {
asm volatile (
"vpcmpeqb %%ymm5,%%ymm5,%%ymm5 \n"
LABELALIGN
@ -2263,7 +2261,6 @@ void OMITFP YUY2ToARGBRow_AVX2(const uint8* yuy2_buf,
uint8* dst_argb,
struct YuvConstants* yuvconstants,
int width) {
asm volatile (
"vpcmpeqb %%ymm5,%%ymm5,%%ymm5 \n"
LABELALIGN
@ -2293,7 +2290,6 @@ void OMITFP UYVYToARGBRow_AVX2(const uint8* uyvy_buf,
uint8* dst_argb,
struct YuvConstants* yuvconstants,
int width) {
asm volatile (
"vpcmpeqb %%ymm5,%%ymm5,%%ymm5 \n"
LABELALIGN
@ -2461,34 +2457,6 @@ void MirrorRow_AVX2(const uint8* src, uint8* dst, int width) {
}
#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 = {
@ -3333,92 +3301,6 @@ void UYVYToUV422Row_AVX2(const uint8* src_uyvy,
}
#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 = {
@ -3427,15 +3309,6 @@ static uvec8 kShuffleAlpha = {
};
// 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 (
@ -3516,50 +3389,6 @@ void ARGBBlendRow_SSSE3(const uint8* src_argb0, const uint8* src_argb1,
}
#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 = {

163
source/row_win.cc
View File

@ -3460,32 +3460,6 @@ void MirrorRow_AVX2(const uint8* src, uint8* dst, int width) {
}
#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 = {
@ -4382,107 +4356,14 @@ void UYVYToUV422Row_SSE2(const uint8* src_uyvy,
}
#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.
// Blend 8 pixels at a time.
__declspec(naked)
void ARGBBlendRow_SSSE3(const uint8* src_argb0, const uint8* src_argb1,
uint8* dst_argb, int width) {
@ -4564,48 +4445,6 @@ void ARGBBlendRow_SSSE3(const uint8* src_argb0, const uint8* src_argb1,
}
#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 = {