coffee/libyuv
1
0
Fork 0
mirror of https://chromium.googlesource.com/libyuv/libyuv synced 2025-12-06 16:56:55 +08:00
Code Issues Packages Projects Releases Wiki Activity

ARGBToUV use AVX2 for 64 bit x86

Browse Source 
Skylake
Was ARGBToJ420_Opt (312 ms)
Now ARGBToJ420_Opt (242 ms)

Icelake
Was ARGBToJ420_Opt (302 ms)
Now ARGBToJ420_Opt (220 ms)

AMD Zen3 on Windows
Was ARGBToJ420_Opt (305 ms)
Now ARGBToJ420_Opt (216 ms)
32 bit x86 uses SSE
Now ARGBToJ420_Opt (326 ms)

MCA analysis of new AVX, SSE and old AVX
https://godbolt.org/z/37bdazWYr

Bug: None
Change-Id: I72f5504407751e164c3558aebe836dd15223d65f
Reviewed-on: https://chromium-review.googlesource.com/c/libyuv/libyuv/+/6957477
Reviewed-by: Justin Green <greenjustin@google.com>
This commit is contained in:
Frank Barchard 2025-09-17 13:42:12 -07:00
parent b33794a586
commit 142db12947
4 changed files with 159 additions and 44 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 Name: libyuv
URL: https://chromium.googlesource.com/libyuv/libyuv/ URL: https://chromium.googlesource.com/libyuv/libyuv/
Version: 1918 Version: 1919
License: BSD-3-Clause License: BSD-3-Clause
License File: LICENSE License File: LICENSE
Shipped: yes Shipped: yes

17
include/libyuv/row.h
View File

@ -346,11 +346,6 @@ extern "C" {
#define HAS_SPLITXRGBROW_AVX2 #define HAS_SPLITXRGBROW_AVX2
#define HAS_SWAPUVROW_AVX2 #define HAS_SWAPUVROW_AVX2
#define HAS_YUY2TONVUVROW_AVX2 #define HAS_YUY2TONVUVROW_AVX2
// TODO: Port SSSE3 to AVX2
// #define HAS_ABGRTOUVJROW_AVX2
// #define HAS_ABGRTOUVROW_AVX2
// #define HAS_ARGBTOUVJROW_AVX2
#define HAS_ARGBTOUVROW_AVX2
#if defined(__x86_64__) || !defined(__pic__) #if defined(__x86_64__) || !defined(__pic__)
// TODO(fbarchard): fix build error on android_full_debug=1 // TODO(fbarchard): fix build error on android_full_debug=1
@ -360,6 +355,18 @@ extern "C" {
#endif #endif
#endif #endif
// The following are available for AVX2 gcc/clang x64 platforms:
// TODO(fbarchard): Port to Visual C
// TODO(fbarchard): Port to x86 32 bit
#if !defined(LIBYUV_DISABLE_X86) && defined(__x86_64__) && \
(defined(CLANG_HAS_AVX2) || defined(GCC_HAS_AVX2)) && \
!defined(LIBYUV_ENABLE_ROWWIN)
#define HAS_ABGRTOUVJROW_AVX2
#define HAS_ABGRTOUVROW_AVX2
#define HAS_ARGBTOUVJROW_AVX2
#define HAS_ARGBTOUVROW_AVX2
#endif
// The following are available for AVX512 clang x86 platforms: // The following are available for AVX512 clang x86 platforms:
// TODO(fbarchard): Port to GCC and Visual C // TODO(fbarchard): Port to GCC and Visual C
// TODO(b/42280744): re-enable HAS_ARGBTORGB24ROW_AVX512VBMI. // TODO(b/42280744): re-enable HAS_ARGBTORGB24ROW_AVX512VBMI.

2
include/libyuv/version.h
View File

@ -11,6 +11,6 @@
#ifndef INCLUDE_LIBYUV_VERSION_H_ #ifndef INCLUDE_LIBYUV_VERSION_H_
#define INCLUDE_LIBYUV_VERSION_H_ #define INCLUDE_LIBYUV_VERSION_H_
#define LIBYUV_VERSION 1918 #define LIBYUV_VERSION 1919
#endif // INCLUDE_LIBYUV_VERSION_H_ #endif // INCLUDE_LIBYUV_VERSION_H_

182
source/row_gcc.cc
View File

@ -1747,9 +1747,8 @@ void ARGBToUVMatrixRow_SSSE3(const uint8_t* src_argb,
"movdqa %0,%%xmm6 \n" // ARGB to U "movdqa %0,%%xmm6 \n" // ARGB to U
"movdqa %1,%%xmm7 \n" // ARGB to V "movdqa %1,%%xmm7 \n" // ARGB to V
: :
: : "m"(rgbuvconstants->kRGBToU), // %0
"m"(rgbuvconstants->kRGBToU), // %0 "m"(rgbuvconstants->kRGBToV) // %1
"m"(rgbuvconstants->kRGBToV) // %1
: "memory", "cc"); : "memory", "cc");
asm volatile( asm volatile(
@ -1824,9 +1823,12 @@ void ARGBToUVMatrixRow_SSSE3(const uint8_t* src_argb,
#ifdef HAS_ARGBTOUVROW_AVX2 #ifdef HAS_ARGBTOUVROW_AVX2
// UYVY shuf 8 UV to 16 UV. // ARGBARGB to AARRGGBB shuffle
static const vec8 kShuffleAARRGGBB = {0, 4, 1, 5, 2, 6, 3, 7, static const lvec8 kShuffleAARRGGBB = {
8, 12, 9, 13, 10, 14, 11, 15}; 0, 4, 1, 5, 2, 6, 3, 7, 8, 12, 9, 13, 10, 14, 11, 15,
0, 4, 1, 5, 2, 6, 3, 7, 8, 12, 9, 13, 10, 14, 11, 15,
};
// 16x2 -> 8x1 ARGB pixels converted to 8 U and 8 V // 16x2 -> 8x1 ARGB pixels converted to 8 U and 8 V
// ARGBToUV does rounding average of 4 ARGB pixels // ARGBToUV does rounding average of 4 ARGB pixels
@ -1837,19 +1839,16 @@ void ARGBToUVMatrixRow_AVX2(const uint8_t* src_argb,
int width, int width,
const struct RgbUVConstants* rgbuvconstants) { const struct RgbUVConstants* rgbuvconstants) {
asm volatile( asm volatile(
"vbroadcastf128 %7,%%ymm15 \n" // kShuffleAARRGGBB "vmovdqa %7,%%ymm15 \n" // kShuffleAARRGGBB
"vpcmpeqb %%ymm14,%%ymm14,%%ymm14 \n" // 0x0101 "vpcmpeqb %%ymm14,%%ymm14,%%ymm14 \n" // 0x0101
"vpabsb %%ymm14,%%ymm14 \n" "vpabsb %%ymm14,%%ymm14 \n"
"vpxor %%ymm13,%%ymm13,%%ymm13 \n" // 0 for vpavgw "vpxor %%ymm13,%%ymm13,%%ymm13 \n" // 0 for vpavgw
"vbroadcastf128 %5,%%ymm6 \n" // RGBToU "vbroadcastf128 %5,%%ymm6 \n" // RGBToU
"vbroadcastf128 %6,%%ymm7 \n" // RGBToV "vbroadcastf128 %6,%%ymm7 \n" // RGBToV
"vpcmpeqb %%ymm5,%%ymm5,%%ymm5 \n" // 0x8000 "vpcmpeqb %%ymm5,%%ymm5,%%ymm5 \n" // 0x8000
"vpsllw $15,%%ymm5,%%ymm5 \n" "vpsllw $15,%%ymm5,%%ymm5 \n"
"sub %1,%2 \n" "sub %1,%2 \n"
LABELALIGN
"1: \n" "1: \n"
"vmovdqu (%0),%%ymm8 \n" // Read 16x2 ARGB Pixels "vmovdqu (%0),%%ymm8 \n" // Read 16x2 ARGB Pixels
"vmovdqu 0x20(%0),%%ymm9 \n" "vmovdqu 0x20(%0),%%ymm9 \n"
@ -1886,7 +1885,7 @@ void ARGBToUVMatrixRow_AVX2(const uint8_t* src_argb,
"lea 0x40(%0),%0 \n" "lea 0x40(%0),%0 \n"
"lea 0x8(%1),%1 \n" "lea 0x8(%1),%1 \n"
"subl $0x10,%3 \n" "subl $0x10,%3 \n"
"jg 1b \n" "jg 1b \n"
"vzeroupper \n" "vzeroupper \n"
: "+r"(src_argb), // %0 : "+r"(src_argb), // %0
"+r"(dst_u), // %1 "+r"(dst_u), // %1
@ -1899,10 +1898,85 @@ void ARGBToUVMatrixRow_AVX2(const uint8_t* src_argb,
: "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm5", "xmm6", "xmm7", : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm5", "xmm6", "xmm7",
"xmm8", "xmm9", "xmm10", "xmm11", "xmm13", "xmm14", "xmm15"); "xmm8", "xmm9", "xmm10", "xmm11", "xmm13", "xmm14", "xmm15");
} }
// 16x2 -> 8x1 ARGB pixels converted to 8 U and 8 V
// ARGBToUV does rounding average of 4 ARGB pixels
// TODO: unroll to 32x2 or update caller/any to 16 pixels
void ARGBToUVMatrixRow_AVX2_MADDWD(const uint8_t* src_argb,
int src_stride_argb,
uint8_t* dst_u,
uint8_t* dst_v,
int width,
const struct RgbUVConstants* rgbuvconstants) {
asm volatile(
"vmovdqa %7,%%ymm15 \n" // kShuffleAARRGGBB
"vpcmpeqb %%ymm14,%%ymm14,%%ymm14 \n" // 0x0101
"vpabsb %%ymm14,%%ymm14 \n"
"vpxor %%ymm13,%%ymm13,%%ymm13 \n" // 0 for vpavgw
"vpcmpeqw %%ymm12,%%ymm12,%%ymm12 \n" // -1
"vbroadcastf128 %5,%%ymm6 \n" // RGBToU
"vbroadcastf128 %6,%%ymm7 \n" // RGBToV
"vpcmpeqd %%ymm5,%%ymm5,%%ymm5 \n" // 0x00008000
"vpabsd %%ymm5,%%ymm5 \n"
"vpslld $15,%%ymm5,%%ymm5 \n"
"sub %1,%2 \n"
LABELALIGN
"1: \n"
"vmovdqu (%0),%%ymm8 \n" // Read 16x2 ARGB Pixels
"vmovdqu 0x20(%0),%%ymm9 \n"
"vmovdqu 0x00(%0,%4,1),%%ymm10 \n"
"vmovdqu 0x20(%0,%4,1),%%ymm11 \n"
"vpshufb %%ymm15,%%ymm8,%%ymm8 \n" // aarrggbb
"vpshufb %%ymm15,%%ymm9,%%ymm9 \n"
"vpshufb %%ymm15,%%ymm10,%%ymm10 \n"
"vpshufb %%ymm15,%%ymm11,%%ymm11 \n"
"vpmaddubsw %%ymm14,%%ymm8,%%ymm8 \n" // 16x2 -> 8x2
"vpmaddubsw %%ymm14,%%ymm9,%%ymm9 \n"
"vpmaddubsw %%ymm14,%%ymm10,%%ymm10 \n"
"vpmaddubsw %%ymm14,%%ymm11,%%ymm11 \n"
"vpaddw %%ymm8,%%ymm10,%%ymm8 \n" // 8x2 -> 8x1
"vpaddw %%ymm9,%%ymm11,%%ymm9 \n"
"vpsrlw $1,%%ymm8,%%ymm8 \n"
"vpsrlw $1,%%ymm9,%%ymm9 \n"
"vpavgw %%ymm13,%%ymm8,%%ymm8 \n"
"vpavgw %%ymm13,%%ymm9,%%ymm9 \n"
"vpackuswb %%ymm9,%%ymm8,%%ymm0 \n" // mutates
"vpermq $0xd8,%%ymm0,%%ymm0 \n" // 8 ARGB Pixels
"vpmaddubsw %%ymm7,%%ymm0,%%ymm1 \n" // 8 V
"vpmaddubsw %%ymm6,%%ymm0,%%ymm0 \n" // 8 U
"vpmaddwd %%ymm12,%%ymm1,%%ymm1 \n" // negate + hadd
"vpmaddwd %%ymm12,%%ymm0,%%ymm0 \n"
"vpaddd %%ymm1,%%ymm5,%%ymm1 \n" // +0x8000 = 0 to 0xFFFF
"vpaddd %%ymm0,%%ymm5,%%ymm0 \n"
"vpackusdw %%ymm1,%%ymm0,%%ymm0 \n" // mutates
"vpermq $0xd8,%%ymm0,%%ymm0 \n" // uuuuuuuu vvvvvvvv
"vpsrlw $0x8,%%ymm0,%%ymm0 \n"
"vpackuswb %%ymm0,%%ymm0,%%ymm0 \n" // mutates
"vmovq %%xmm0,(%1) \n" // Write 8 U's
"vextractf128 $0x1,%%ymm0,%%xmm0 \n" // Copy V to low 8 bytes
"vmovq %%xmm0,0x00(%1,%2,1) \n" // Write 8 V's
"lea 0x40(%0),%0 \n"
"lea 0x8(%1),%1 \n"
"subl $0x10,%3 \n"
"jg 1b \n"
"vzeroupper \n"
: "+r"(src_argb), // %0
"+r"(dst_u), // %1
"+r"(dst_v), // %2
"+r"(width) // %3
: "r"((intptr_t)(src_stride_argb)), // %4
"m"(rgbuvconstants->kRGBToU), // %5
"m"(rgbuvconstants->kRGBToV), // %6
"m"(kShuffleAARRGGBB) // %7
: "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm5", "xmm6", "xmm7", "xmm8",
"xmm9", "xmm10", "xmm11", "xmm12", "xmm13", "xmm14", "xmm15");
}
#endif // HAS_ARGBTOUVROW_AVX2 #endif // HAS_ARGBTOUVROW_AVX2
#if defined(HAS_ARGBTOUV444ROW_SSSE3) || defined(HAS_ARGBTOUVROW_AVX2)
#ifdef HAS_ARGBTOUV444ROW_SSSE3
// RGB to BT601 coefficients // RGB to BT601 coefficients
// UB 0.875 coefficient = 112 // UB 0.875 coefficient = 112
@ -1916,6 +1990,20 @@ static const struct RgbUVConstants kARGBI601UVConstants = {
{-112, 74, 38, 0, -112, 74, 38, 0, -112, 74, 38, 0, -112, 74, 38, 0}, {-112, 74, 38, 0, -112, 74, 38, 0, -112, 74, 38, 0, -112, 74, 38, 0},
{18, 94, -112, 0, 18, 94, -112, 0, 18, 94, -112, 0, 18, 94, -112, 0}}; {18, 94, -112, 0, 18, 94, -112, 0, 18, 94, -112, 0, 18, 94, -112, 0}};
static const struct RgbUVConstants kABGRI601UVConstants = {
{38, 74, -112, 0, 38, 74, -112, 0, 38, 74, -112, 0, 38, 74, -112, 0},
{-112, 94, 18, 0, -112, 94, 18, 0, -112, 94, 18, 0, -112, 94, 18, 0}};
static const struct RgbUVConstants kBGRAI601UVConstants = {
{0, 38, 74, -112, 0, 38, 74, -112, 0, 38, 74, -112, 0, 38, 74, -112},
{0, -112, 94, 18, 0, -112, 94, 18, 0, -112, 94, 18, 0, -112, 94, 18}};
static const struct RgbUVConstants kRGBAI601UVConstants = {
{0, -112, 74, 38, 0, -112, 74, 38, 0, -112, 74, 38, 0, -112, 74, 38},
{0, 18, 94, -112, 0, 18, 94, -112, 0, 18, 94, -112, 0, 18, 94, -112}};
#endif
#ifdef HAS_ARGBTOUV444ROW_SSSE3
void ARGBToUV444Row_SSSE3(const uint8_t* src_argb, void ARGBToUV444Row_SSSE3(const uint8_t* src_argb,
uint8_t* dst_u, uint8_t* dst_u,
uint8_t* dst_v, uint8_t* dst_v,
@ -1935,18 +2023,6 @@ void ARGBToUV444Row_AVX2(const uint8_t* src_argb,
} }
#endif // HAS_ARGBTOUV444ROW_AVX2 #endif // HAS_ARGBTOUV444ROW_AVX2
#ifdef HAS_ARGBTOUVROW_AVX2
void ARGBToUVRow_AVX2(const uint8_t* src_argb,
int src_stride_argb,
uint8_t* dst_u,
uint8_t* dst_v,
int width) {
ARGBToUVMatrixRow_AVX2(src_argb, src_stride_argb, dst_u, dst_v, width,
&kARGBI601UVConstants);
}
#endif // HAS_ARGBTOUVROW_AVX2
#ifdef HAS_ARGBTOUVROW_SSSE3 #ifdef HAS_ARGBTOUVROW_SSSE3
void ARGBToUVRow_SSSE3(const uint8_t* src_argb, void ARGBToUVRow_SSSE3(const uint8_t* src_argb,
int src_stride_argb, int src_stride_argb,
@ -1957,10 +2033,6 @@ void ARGBToUVRow_SSSE3(const uint8_t* src_argb,
&kARGBI601UVConstants); &kARGBI601UVConstants);
} }
static const struct RgbUVConstants kABGRI601UVConstants = {
{38, 74, -112, 0, 38, 74, -112, 0, 38, 74, -112, 0, 38, 74, -112, 0},
{-112, 94, 18, 0, -112, 94, 18, 0, -112, 94, 18, 0, -112, 94, 18, 0}};
void ABGRToUVRow_SSSE3(const uint8_t* src_abgr, void ABGRToUVRow_SSSE3(const uint8_t* src_abgr,
int src_stride_abgr, int src_stride_abgr,
uint8_t* dst_u, uint8_t* dst_u,
@ -1970,10 +2042,6 @@ void ABGRToUVRow_SSSE3(const uint8_t* src_abgr,
&kABGRI601UVConstants); &kABGRI601UVConstants);
} }
static const struct RgbUVConstants kBGRAI601UVConstants = {
{0, 38, 74, -112, 0, 38, 74, -112, 0, 38, 74, -112, 0, 38, 74, -112},
{0, -112, 94, 18, 0, -112, 94, 18, 0, -112, 94, 18, 0, -112, 94, 18}};
void BGRAToUVRow_SSSE3(const uint8_t* src_bgra, void BGRAToUVRow_SSSE3(const uint8_t* src_bgra,
int src_stride_bgra, int src_stride_bgra,
uint8_t* dst_u, uint8_t* dst_u,
@ -1983,10 +2051,6 @@ void BGRAToUVRow_SSSE3(const uint8_t* src_bgra,
&kBGRAI601UVConstants); &kBGRAI601UVConstants);
} }
static const struct RgbUVConstants kRGBAI601UVConstants = {
{0, -112, 74, 38, 0, -112, 74, 38, 0, -112, 74, 38, 0, -112, 74, 38},
{0, 18, 94, -112, 0, 18, 94, -112, 0, 18, 94, -112, 0, 18, 94, -112}};
void RGBAToUVRow_SSSE3(const uint8_t* src_rgba, void RGBAToUVRow_SSSE3(const uint8_t* src_rgba,
int src_stride_rgba, int src_stride_rgba,
uint8_t* dst_u, uint8_t* dst_u,
@ -1997,6 +2061,26 @@ void RGBAToUVRow_SSSE3(const uint8_t* src_rgba,
} }
#endif // HAS_ARGBTOUVROW_SSSE3 #endif // HAS_ARGBTOUVROW_SSSE3
#ifdef HAS_ARGBTOUVROW_AVX2
void ARGBToUVRow_AVX2(const uint8_t* src_argb,
int src_stride_argb,
uint8_t* dst_u,
uint8_t* dst_v,
int width) {
ARGBToUVMatrixRow_AVX2(src_argb, src_stride_argb, dst_u, dst_v, width,
&kARGBI601UVConstants);
}
void ABGRToUVRow_AVX2(const uint8_t* src_abgr,
int src_stride_abgr,
uint8_t* dst_u,
uint8_t* dst_v,
int width) {
ARGBToUVMatrixRow_AVX2(src_abgr, src_stride_abgr, dst_u, dst_v, width,
&kABGRI601UVConstants);
}
#endif // HAS_ARGBTOUVROW_AVX2
#ifdef HAS_ARGBTOUVJ444ROW_SSSE3 #ifdef HAS_ARGBTOUVJ444ROW_SSSE3
// RGB to JPEG coefficients // RGB to JPEG coefficients
// UB 0.500 coefficient = 128 // UB 0.500 coefficient = 128
@ -2043,7 +2127,9 @@ void ARGBToUVJRow_SSSE3(const uint8_t* src_argb,
ARGBToUVMatrixRow_SSSE3(src_argb, src_stride_argb, dst_u, dst_v, width, ARGBToUVMatrixRow_SSSE3(src_argb, src_stride_argb, dst_u, dst_v, width,
&kARGBJPEGUVConstants); &kARGBJPEGUVConstants);
} }
#endif // HAS_ARGBTOUVJROW_SSSE3
#ifdef HAS_ABGRTOUVJROW_SSSE3
void ABGRToUVJRow_SSSE3(const uint8_t* src_abgr, void ABGRToUVJRow_SSSE3(const uint8_t* src_abgr,
int src_stride_abgr, int src_stride_abgr,
uint8_t* dst_u, uint8_t* dst_u,
@ -2054,6 +2140,28 @@ void ABGRToUVJRow_SSSE3(const uint8_t* src_abgr,
} }
#endif // HAS_ABGRTOUVJROW_SSSE3 #endif // HAS_ABGRTOUVJROW_SSSE3
#ifdef HAS_ARGBTOUVJROW_AVX2
void ARGBToUVJRow_AVX2(const uint8_t* src_argb,
int src_stride_argb,
uint8_t* dst_u,
uint8_t* dst_v,
int width) {
ARGBToUVMatrixRow_AVX2(src_argb, src_stride_argb, dst_u, dst_v, width,
&kARGBJPEGUVConstants);
}
#endif // HAS_ARGBTOUVJROW_AVX2
#ifdef HAS_ABGRTOUVJROW_AVX2
void ABGRToUVJRow_AVX2(const uint8_t* src_abgr,
int src_stride_abgr,
uint8_t* dst_u,
uint8_t* dst_v,
int width) {
ARGBToUVMatrixRow_AVX2(src_abgr, src_stride_abgr, dst_u, dst_v, width,
&kABGRJPEGUVConstants);
}
#endif // HAS_ABGRTOUVJROW_AVX2
void BGRAToYRow_SSSE3(const uint8_t* src_bgra, uint8_t* dst_y, int width) { void BGRAToYRow_SSSE3(const uint8_t* src_bgra, uint8_t* dst_y, int width) {
asm volatile( asm volatile(
"movdqa %3,%%xmm4 \n" "movdqa %3,%%xmm4 \n"