coffee/libyuv
1
0
Fork 0
mirror of https://chromium.googlesource.com/libyuv/libyuv synced 2026-10-28 18:07:12 +08:00
Code Issues Packages Projects Releases Wiki Activity

I422ToRGB24Row_AVX512VBMI handle tail with mask

Browse Source 
I422ToRGB24Matrix (299x299, repeat=10000)
====================================================================
CPU                        Before       After            Improvement
--------------------------------------------------------------------
Intel Ice Lake             220 ms       181 ms       -17.7% (1.22x)
Intel Sapphire Rapids      181 ms       145 ms       -19.9% (1.25x)
Intel Emerald Rapids       169 ms       148 ms       -12.4% (1.14x)
Intel Granite Rapids       191 ms       141 ms       -26.2% (1.35x)
AMD Zen 4 (Genoa)          171 ms       136 ms       -20.5% (1.26x)
AMD Zen 5 (Turin)           94 ms        72 ms       -23.4% (1.31x)
AMD Zen 6 (Venice)          95 ms        66 ms       -30.5% (1.44x)
====================================================================

Bug: None
Change-Id: I4cd657410cc5253b5e8fd2caa78d17e2b3812393
Reviewed-on: https://chromium-review.googlesource.com/c/libyuv/libyuv/+/8445804
Reviewed-by: richard winterton <rrwinterton@gmail.com>
Commit-Queue: Frank Barchard <fbarchard@google.com>
This commit is contained in:
Frank Barchard 2026-09-23 21:40:00 -07:00 committed by libyuv-scoped@luci-project-accounts.iam.gserviceaccount.com
parent dab3111c8a
commit 41a6e684a6
4 changed files with 54 additions and 22 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

9
include/libyuv/row.h
View File

@ -375,7 +375,6 @@ extern "C" {
#define HAS_RGB24TOARGBROW_AVX512BW #define HAS_RGB24TOARGBROW_AVX512BW
#define HAS_RAWTORGB24ROW_AVX512BW #define HAS_RAWTORGB24ROW_AVX512BW
#define HAS_ARGBSHUFFLEROW_AVX512BW #define HAS_ARGBSHUFFLEROW_AVX512BW
#define HAS_I422TORGB24ROW_AVX512VBMI
#define HAS_RAWTORGB24ROW_AVX512VBMI #define HAS_RAWTORGB24ROW_AVX512VBMI
#define HAS_I422TORGB24ROW_AVX512BW #define HAS_I422TORGB24ROW_AVX512BW
#define HAS_I422TOAR30ROW_AVX512BW #define HAS_I422TOAR30ROW_AVX512BW
@ -418,6 +417,7 @@ extern "C" {
#define HAS_RAWTORGB24ROW_AVX512BW #define HAS_RAWTORGB24ROW_AVX512BW
#define HAS_RAWTORGB24ROW_AVX512VBMI #define HAS_RAWTORGB24ROW_AVX512VBMI
#define HAS_ARGBTORGB24ROW_AVX512VBMI #define HAS_ARGBTORGB24ROW_AVX512VBMI
#define HAS_I422TORGB24ROW_AVX512VBMI
#define HAS_CONVERT16TO8ROW_AVX512BW #define HAS_CONVERT16TO8ROW_AVX512BW
#define HAS_CONVERT8TO16ROW_AVX512BW #define HAS_CONVERT8TO16ROW_AVX512BW
#define HAS_HALFROW_16TO8_AVX512BW #define HAS_HALFROW_16TO8_AVX512BW
@ -436,7 +436,6 @@ extern "C" {
#define HAS_BLENDPLANEROW_AVX512BW #define HAS_BLENDPLANEROW_AVX512BW
#define HAS_RGBTOUV444MATRIXROW_AVX512BW #define HAS_RGBTOUV444MATRIXROW_AVX512BW
#define HAS_ARGBTOYMATRIXROW_AVX512BW #define HAS_ARGBTOYMATRIXROW_AVX512BW
#define HAS_I422TORGB24ROW_AVX512VBMI
#define HAS_RAWTORGB24ROW_AVX512VBMI #define HAS_RAWTORGB24ROW_AVX512VBMI
#define HAS_I422TORGB24ROW_AVX512BW #define HAS_I422TORGB24ROW_AVX512BW
#define HAS_ARGBTOUVMATRIXROW_AVX512BW #define HAS_ARGBTOUVMATRIXROW_AVX512BW
@ -5275,12 +5274,6 @@ void NV12ToRGB565Row_Any_AVX2(const uint8_t* y_buf,
uint8_t* dst_ptr, uint8_t* dst_ptr,
const struct YuvConstants* yuvconstants, const struct YuvConstants* yuvconstants,
int width); int width);
void I422ToRGB24Row_Any_AVX512VBMI(const uint8_t* y_buf,
const uint8_t* u_buf,
const uint8_t* v_buf,
uint8_t* dst_ptr,
const struct YuvConstants* yuvconstants,
int width);
void I422ToRGB24Row_Any_AVX512BW(const uint8_t* y_buf, void I422ToRGB24Row_Any_AVX512BW(const uint8_t* y_buf,
const uint8_t* u_buf, const uint8_t* u_buf,
const uint8_t* v_buf, const uint8_t* v_buf,

10
source/convert_argb.cc
View File

@ -5703,10 +5703,7 @@ int I420ToRGB24Matrix(const uint8_t* src_y,
#endif #endif
#if defined(HAS_I422TORGB24ROW_AVX512VBMI) #if defined(HAS_I422TORGB24ROW_AVX512VBMI)
if (TestCpuFlag(kCpuHasAVX512VBMI)) { if (TestCpuFlag(kCpuHasAVX512VBMI)) {
I422ToRGB24Row = I422ToRGB24Row_Any_AVX512VBMI; I422ToRGB24Row = I422ToRGB24Row_AVX512VBMI;
if (IS_ALIGNED(width, 32)) {
I422ToRGB24Row = I422ToRGB24Row_AVX512VBMI;
}
} }
#endif #endif
#if defined(HAS_I422TORGB24ROW_NEON) #if defined(HAS_I422TORGB24ROW_NEON)
@ -5924,10 +5921,7 @@ int I422ToRGB24Matrix(const uint8_t* src_y,
#endif #endif
#if defined(HAS_I422TORGB24ROW_AVX512VBMI) #if defined(HAS_I422TORGB24ROW_AVX512VBMI)
if (TestCpuFlag(kCpuHasAVX512VBMI)) { if (TestCpuFlag(kCpuHasAVX512VBMI)) {
I422ToRGB24Row = I422ToRGB24Row_Any_AVX512VBMI; I422ToRGB24Row = I422ToRGB24Row_AVX512VBMI;
if (IS_ALIGNED(width, 32)) {
I422ToRGB24Row = I422ToRGB24Row_AVX512VBMI;
}
} }
#endif #endif
#if defined(HAS_I422TORGB24ROW_NEON) #if defined(HAS_I422TORGB24ROW_NEON)

3
source/row_any.cc
View File

@ -420,9 +420,6 @@ ANY31C(I444ToRGB24Row_Any_SSSE3, I444ToRGB24Row_SSSE3, 0, 0, 3, 15)
#ifdef HAS_I422TORGB24ROW_AVX2 #ifdef HAS_I422TORGB24ROW_AVX2
ANY31C(I422ToRGB24Row_Any_AVX2, I422ToRGB24Row_AVX2, 1, 0, 3, 15) ANY31C(I422ToRGB24Row_Any_AVX2, I422ToRGB24Row_AVX2, 1, 0, 3, 15)
#endif #endif
#ifdef HAS_I422TORGB24ROW_AVX512VBMI
ANY31C(I422ToRGB24Row_Any_AVX512VBMI, I422ToRGB24Row_AVX512VBMI, 1, 0, 3, 31)
#endif
#ifdef HAS_I422TORGB24ROW_AVX512BW #ifdef HAS_I422TORGB24ROW_AVX512BW
ANY31C(I422ToRGB24Row_Any_AVX512BW, I422ToRGB24Row_AVX512BW, 1, 0, 3, 31) ANY31C(I422ToRGB24Row_Any_AVX512BW, I422ToRGB24Row_AVX512BW, 1, 0, 3, 31)
#endif #endif

54
source/row_gcc.cc
View File

@ -4120,12 +4120,15 @@ void OMITFP I422ToRGB24Row_AVX512VBMI(const uint8_t* y_buf,
uint8_t* dst_rgb24, uint8_t* dst_rgb24,
const struct YuvConstants* yuvconstants, const struct YuvConstants* yuvconstants,
int width) { int width) {
uintptr_t temp;
asm volatile ( asm volatile (
YUVTORGB_SETUP_AVX512BW(yuvconstants) YUVTORGB_SETUP_AVX512BW(yuvconstants)
"vmovdqu32 %[kMaskBG],%%zmm20 \n" "vmovdqu32 %[kMaskBG],%%zmm20 \n"
"vmovdqu32 %[kMaskDST0],%%zmm21 \n" "vmovdqu32 %[kMaskDST0],%%zmm21 \n"
"vmovdqu32 %[kMaskDST1],%%zmm22 \n" "vmovdqu32 %[kMaskDST1],%%zmm22 \n"
"sub %[u_buf],%[v_buf] \n" "sub %[u_buf],%[v_buf] \n"
"sub $0x20,%[width] \n"
"jl 2f \n"
LABELALIGN LABELALIGN
"1: \n" "1: \n"
@ -4133,13 +4136,57 @@ void OMITFP I422ToRGB24Row_AVX512VBMI(const uint8_t* y_buf,
YUVTORGB_AVX512BW(yuvconstants) YUVTORGB_AVX512BW(yuvconstants)
STORERGB24_AVX512VBMI STORERGB24_AVX512VBMI
"sub $0x20,%[width] \n" "sub $0x20,%[width] \n"
"jg 1b \n" "jge 1b \n"
"2: \n"
"add $0x20,%[width] \n"
"je 99f \n"
// Calculate a mask for the final iteration to deal with the tail.
"mov $-1,%[temp] \n"
"bzhi %q[width],%[temp],%[temp] \n"
"kmovd %k[temp],%%k1 \n"
"vpmovzxbw (%[y_buf]),%%zmm4%{%%k1%}%{z%}\n"
"mov $-1,%[temp] \n"
"lea 1(%q[width]),%k[y_buf] \n"
"shr $1,%k[y_buf] \n"
"bzhi %q[y_buf],%[temp],%[y_buf] \n"
"kmovw %k[y_buf],%%k2 \n"
"vmovdqu8 (%[u_buf]),%%xmm3%{%%k2%}%{z%}\n"
"vmovdqu8 0x00(%[u_buf],%[v_buf],1),%%xmm1%{%%k2%}%{z%}\n"
"lea (%q[width],%q[width],2),%k[width]\n"
"bzhi %q[width],%[temp],%[y_buf] \n"
"kmovq %[y_buf],%%k3 \n"
"xor %k[y_buf],%k[y_buf] \n"
"sub $0x40,%k[width] \n"
"cmovg %[temp],%[y_buf] \n"
"bzhi %q[width],%[y_buf],%[y_buf] \n"
"kmovd %k[y_buf],%%k4 \n"
"vpunpcklbw %%xmm1,%%xmm3,%%xmm2 \n"
"vpunpckhbw %%xmm1,%%xmm3,%%xmm3 \n"
"vinserti64x2 $1,%%xmm3,%%ymm2,%%ymm3 \n"
"vmovdqa64 %%zmm16,%%zmm1 \n"
"vpermi2w %%zmm3,%%zmm3,%%zmm1 \n"
"vpsllw $8,%%zmm4,%%zmm3 \n"
"vpord %%zmm3,%%zmm4,%%zmm4 \n"
"vmovdqa64 %%zmm1,%%zmm3 \n"
YUVTORGB_AVX512BW(yuvconstants)
"vpermt2b %%zmm1,%%zmm20,%%zmm0 \n"
"vmovdqa64 %%zmm0,%%zmm3 \n"
"vpermt2b %%zmm2,%%zmm21,%%zmm3 \n"
"vpermt2b %%zmm2,%%zmm22,%%zmm0 \n"
"vmovdqu8 %%zmm3,(%[dst_rgb24])%{%%k3%} \n"
"vmovdqu8 %%ymm0,0x40(%[dst_rgb24])%{%%k4%}\n"
"99: \n"
"vzeroupper \n" "vzeroupper \n"
: [y_buf]"+r"(y_buf), // %[y_buf] : [y_buf]"+r"(y_buf), // %[y_buf]
[u_buf]"+r"(u_buf), // %[u_buf] [u_buf]"+r"(u_buf), // %[u_buf]
[v_buf]"+r"(v_buf), // %[v_buf] [v_buf]"+r"(v_buf), // %[v_buf]
[dst_rgb24]"+r"(dst_rgb24), // %[dst_rgb24] [dst_rgb24]"+r"(dst_rgb24), // %[dst_rgb24]
[width]"+rm"(width) // %[width] [width]"+r"(width), // %[width]
[temp]"=&r"(temp) // %[temp]
: [yuvconstants]"r"(yuvconstants), // %[yuvconstants] : [yuvconstants]"r"(yuvconstants), // %[yuvconstants]
[quadsplitperm]"r"(kSplitQuadWords), // %[quadsplitperm] [quadsplitperm]"r"(kSplitQuadWords), // %[quadsplitperm]
[dquadsplitperm]"r"(kSplitDoubleQuadWords), // %[dquadsplitperm] [dquadsplitperm]"r"(kSplitDoubleQuadWords), // %[dquadsplitperm]
@ -4149,7 +4196,8 @@ void OMITFP I422ToRGB24Row_AVX512VBMI(const uint8_t* y_buf,
[kMaskDST1]"m"(kMaskDST1) // %[kMaskDST1] [kMaskDST1]"m"(kMaskDST1) // %[kMaskDST1]
: "memory", "cc", YUVTORGB_REGS_AVX512BW : "memory", "cc", YUVTORGB_REGS_AVX512BW
"xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5",
"xmm20", "xmm21", "xmm22" "xmm20", "xmm21", "xmm22",
"k1", "k2", "k3", "k4"
); );
} }
#endif // defined(HAS_I422TORGB24ROW_AVX512VBMI) #endif // defined(HAS_I422TORGB24ROW_AVX512VBMI)