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

ARGBToRGB24 AVX512 with tail masking

Browse Source 
Bug: None
Change-Id: I7590c66940b5354ca0fdd8ff120392434d2785d7
Reviewed-on: https://chromium-review.googlesource.com/c/libyuv/libyuv/+/8490011
Reviewed-by: richard winterton <rrwinterton@gmail.com>
Commit-Queue: Frank Barchard <fbarchard@google.com>
This commit is contained in:
Frank Barchard 2026-09-29 15:20:53 -07:00 committed by libyuv-scoped@luci-project-accounts.iam.gserviceaccount.com
parent b25fa89920
commit 18835bfb35
6 changed files with 120 additions and 61 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: https://chromium.googlesource.com/libyuv/libyuv/
Version: 1973
Version: 1974
Revision: DEPS
License: BSD-3-Clause
License File: LICENSE

10
include/libyuv/row.h
View File

@ -411,7 +411,6 @@ extern "C" {
// The following are available for AVX512 clang x64 platforms:
// TODO(fbarchard): Port to GCC and Visual C
// TODO(b/42280744): re-enable HAS_ARGBTORGB24ROW_AVX512VBMI.
#if !defined(LIBYUV_DISABLE_X86) && \
(defined(__x86_64__) || defined(_M_X64)) && defined(CLANG_HAS_AVX512) && \
!defined(LIBYUV_ENABLE_ROWWIN)
@ -420,7 +419,8 @@ extern "C" {
#define HAS_RGB24TOARGBROW_AVX512BW
#define HAS_RAWTORGB24ROW_AVX512BW
#define HAS_RAWTORGB24ROW_AVX512VBMI
#define HAS_ARGBTORGB24ROW_AVX512VBMI
#define HAS_ARGBTORGB24ROW_AVX512BW
#define HAS_ARGBTORAWROW_AVX512BW
#define HAS_CONVERT16TO8ROW_AVX512BW
#define HAS_CONVERT8TO16ROW_AVX512BW
#define HAS_HALFROW_16TO8_AVX512BW
@ -4174,7 +4174,8 @@ void ARGBToAR30Row_SSSE3(const uint8_t* src, uint8_t* dst, int width);
void ARGBToRAWRow_AVX2(const uint8_t* src, uint8_t* dst, int width);
void ARGBToRGB24Row_AVX2(const uint8_t* src, uint8_t* dst, int width);
void ARGBToRGB24Row_AVX512VBMI(const uint8_t* src, uint8_t* dst, int width);
void ARGBToRAWRow_AVX512BW(const uint8_t* src, uint8_t* dst, int width);
void ARGBToRGB24Row_AVX512BW(const uint8_t* src, uint8_t* dst, int width);
void ARGBToRGB565DitherRow_C(const uint8_t* src_argb,
uint8_t* dst_rgb,
@ -5584,9 +5585,6 @@ void ARGBToRAWRow_Any_AVX2(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
void ARGBToRGB24Row_Any_AVX2(const uint8_t* src_ptr,
uint8_t* dst_ptr,
int width);
void ARGBToRGB24Row_Any_AVX512VBMI(const uint8_t* src_ptr,
uint8_t* dst_ptr,
int width);
void ARGBToRGB565DitherRow_Any_AVX2(const uint8_t* src_ptr,
uint8_t* dst_ptr,

2
include/libyuv/version.h
View File

@ -11,6 +11,6 @@
#ifndef INCLUDE_LIBYUV_VERSION_H_
#define INCLUDE_LIBYUV_VERSION_H_
#define LIBYUV_VERSION 1973
#define LIBYUV_VERSION 1974
#endif // INCLUDE_LIBYUV_VERSION_H_

14
source/convert_from_argb.cc
View File

@ -2964,12 +2964,9 @@ int ARGBToRGB24(const uint8_t* src_argb,
}
}
#endif
#if defined(HAS_ARGBTORGB24ROW_AVX512VBMI)
if (TestCpuFlag(kCpuHasAVX512VBMI)) {
ARGBToRGB24Row = ARGBToRGB24Row_Any_AVX512VBMI;
if (IS_ALIGNED(width, 32)) {
ARGBToRGB24Row = ARGBToRGB24Row_AVX512VBMI;
}
#if defined(HAS_ARGBTORGB24ROW_AVX512BW)
if (TestCpuFlag(kCpuHasAVX512BW)) {
ARGBToRGB24Row = ARGBToRGB24Row_AVX512BW;
}
#endif
#if defined(HAS_ARGBTORGB24ROW_NEON)
@ -3057,6 +3054,11 @@ int ARGBToRAW(const uint8_t* src_argb,
}
}
#endif
#if defined(HAS_ARGBTORAWROW_AVX512BW)
if (TestCpuFlag(kCpuHasAVX512BW)) {
ARGBToRAWRow = ARGBToRAWRow_AVX512BW;
}
#endif
#if defined(HAS_ARGBTORAWROW_NEON)
if (TestCpuFlag(kCpuHasNEON)) {
ARGBToRAWRow = ARGBToRAWRow_Any_NEON;

3
source/row_any.cc
View File

@ -1021,9 +1021,6 @@ ANY11(ARGBToRAWRow_Any_SSSE3, ARGBToRAWRow_SSSE3, 0, 4, 3, 15)
#if defined(HAS_ARGBTORGB24ROW_AVX2)
ANY11(ARGBToRGB24Row_Any_AVX2, ARGBToRGB24Row_AVX2, 0, 4, 3, 31)
#endif
#if defined(HAS_ARGBTORGB24ROW_AVX512VBMI)
ANY11(ARGBToRGB24Row_Any_AVX512VBMI, ARGBToRGB24Row_AVX512VBMI, 0, 4, 3, 31)
#endif
#if defined(HAS_ARGBTORAWROW_AVX2)
ANY11(ARGBToRAWRow_Any_AVX2, ARGBToRAWRow_AVX2, 0, 4, 3, 31)
#endif

150
source/row_gcc.cc
View File

@ -91,14 +91,18 @@ static const uint32_t kPermd2_AVX2[8] = {0, 0, 0, 0, 0, 1, 2, 4};
static const uint32_t kPermd3_AVX2[8] = {5, 6, 0, 0, 0, 0, 0, 0};
static const uint32_t kPermd4_AVX2[8] = {0, 0, 0, 1, 2, 4, 5, 6};
#endif // defined(HAS_RAWTORGB24ROW_AVX2)
#if defined(HAS_RAWTORGB24ROW_AVX512BW)
#if defined(HAS_RAWTORGB24ROW_AVX512BW) || \
defined(HAS_ARGBTORGB24ROW_AVX512BW) || \
defined(HAS_ARGBTORAWROW_AVX512BW)
static const uint32_t kPermdRGB24_0_AVX512BW[16] = {
0, 1, 2, 4, 5, 6, 8, 9, 10, 12, 13, 14, 16, 17, 18, 20};
static const uint32_t kPermdRGB24_1_AVX512BW[16] = {
5, 6, 8, 9, 10, 12, 13, 14, 16, 17, 18, 20, 21, 22, 24, 25};
static const uint32_t kPermdRGB24_2_AVX512BW[16] = {
10, 12, 13, 14, 16, 17, 18, 20, 21, 22, 24, 25, 26, 28, 29, 30};
#endif // defined(HAS_RAWTORGB24ROW_AVX512BW)
#endif // defined(HAS_RAWTORGB24ROW_AVX512BW) ||
// defined(HAS_ARGBTORGB24ROW_AVX512BW) ||
// defined(HAS_ARGBTORAWROW_AVX512BW)
#if defined(HAS_RAWTORGB24ROW_AVX512VBMI)
static const uint8_t kPermRAWToRGB24_0[64] = {
0x02, 0x01, 0x00, 0x05, 0x04, 0x03, 0x08, 0x07, 0x06, 0x0b, 0x0a,
@ -1023,54 +1027,106 @@ void ARGBToRGB24Row_AVX2(const uint8_t* src, uint8_t* dst, int width) {
}
#endif // HAS_ARGBTORGB24ROW_AVX2
#ifdef HAS_ARGBTORGB24ROW_AVX512VBMI
// Shuffle table for converting ARGBToRGB24
static const ulvec8 kPermARGBToRGB24_0 = {
0u, 1u, 2u, 4u, 5u, 6u, 8u, 9u, 10u, 12u, 13u,
14u, 16u, 17u, 18u, 20u, 21u, 22u, 24u, 25u, 26u, 28u,
29u, 30u, 32u, 33u, 34u, 36u, 37u, 38u, 40u, 41u};
static const ulvec8 kPermARGBToRGB24_1 = {
10u, 12u, 13u, 14u, 16u, 17u, 18u, 20u, 21u, 22u, 24u,
25u, 26u, 28u, 29u, 30u, 32u, 33u, 34u, 36u, 37u, 38u,
40u, 41u, 42u, 44u, 45u, 46u, 48u, 49u, 50u, 52u};
static const ulvec8 kPermARGBToRGB24_2 = {
21u, 22u, 24u, 25u, 26u, 28u, 29u, 30u, 32u, 33u, 34u,
36u, 37u, 38u, 40u, 41u, 42u, 44u, 45u, 46u, 48u, 49u,
50u, 52u, 53u, 54u, 56u, 57u, 58u, 60u, 61u, 62u};
void ARGBToRGB24Row_AVX512VBMI(const uint8_t* src, uint8_t* dst, int width) {
#if defined(HAS_ARGBTORGB24ROW_AVX512BW) || defined(HAS_ARGBTORAWROW_AVX512BW)
static inline void ARGBToRGBRow_AVX512BW(const uint8_t* src,
uint8_t* dst,
int width,
const uvec8& shuffler) {
uintptr_t temp;
asm volatile(
"vmovdqa32 %3,%%ymm5 \n"
"vmovdqa32 %4,%%ymm16 \n"
"vmovdqa32 %5,%%ymm17 \n"
"vbroadcasti32x4 %[shuffler],%%zmm4 \n"
"vmovdqu32 %[kPermdRGB24_0],%%zmm5 \n"
"vmovdqu32 %[kPermdRGB24_1],%%zmm16 \n"
"vmovdqu32 %[kPermdRGB24_2],%%zmm17 \n"
"sub $0x40,%[width] \n"
"jl 2f \n"
LABELALIGN
"1: \n"
"vmovdqu (%0),%%ymm0 \n"
"vmovdqu 0x20(%0),%%ymm1 \n"
"vmovdqu 0x40(%0),%%ymm2 \n"
"vmovdqu 0x60(%0),%%ymm3 \n"
"lea 0x80(%0),%0 \n"
"vpermt2b %%ymm1,%%ymm5,%%ymm0 \n"
"vpermt2b %%ymm2,%%ymm16,%%ymm1 \n"
"vpermt2b %%ymm3,%%ymm17,%%ymm2 \n"
"vmovdqu %%ymm0,(%1) \n"
"vmovdqu %%ymm1,0x20(%1) \n"
"vmovdqu %%ymm2,0x40(%1) \n"
"lea 0x60(%1),%1 \n"
"sub $0x20,%2 \n"
"jg 1b \n"
"vmovdqu8 (%[src]),%%zmm0 \n"
"vmovdqu8 0x40(%[src]),%%zmm1 \n"
"vmovdqu8 0x80(%[src]),%%zmm2 \n"
"vmovdqu8 0xc0(%[src]),%%zmm3 \n"
"lea 0x100(%[src]),%[src] \n"
"vpshufb %%zmm4,%%zmm0,%%zmm0 \n"
"vpshufb %%zmm4,%%zmm1,%%zmm1 \n"
"vpshufb %%zmm4,%%zmm2,%%zmm2 \n"
"vpshufb %%zmm4,%%zmm3,%%zmm3 \n"
"vpermt2d %%zmm1,%%zmm5,%%zmm0 \n"
"vpermt2d %%zmm2,%%zmm16,%%zmm1 \n"
"vpermt2d %%zmm3,%%zmm17,%%zmm2 \n"
"vmovdqu8 %%zmm0,(%[dst]) \n"
"vmovdqu8 %%zmm1,0x40(%[dst]) \n"
"vmovdqu8 %%zmm2,0x80(%[dst]) \n"
"lea 0xc0(%[dst]),%[dst] \n"
"sub $0x40,%[width] \n"
"jge 1b \n"
"2: \n"
"add $0x40,%[width] \n"
"je 99f \n"
"mov $-1,%[temp] \n"
"bzhi %q[width],%[temp],%[temp] \n"
"kmovw %k[temp],%%k1 \n"
"shr $0x10,%[temp] \n"
"kmovw %k[temp],%%k2 \n"
"shr $0x10,%[temp] \n"
"kmovw %k[temp],%%k3 \n"
"shr $0x10,%[temp] \n"
"kmovw %k[temp],%%k4 \n"
"vmovdqu32 (%[src]),%%zmm0%{%%k1%}%{z%} \n"
"vmovdqu32 0x40(%[src]),%%zmm1%{%%k2%}%{z%} \n"
"vmovdqu32 0x80(%[src]),%%zmm2%{%%k3%}%{z%} \n"
"vmovdqu32 0xc0(%[src]),%%zmm3%{%%k4%}%{z%} \n"
"mov $-1,%[temp] \n"
"lea (%q[width],%q[width],2),%k[width] \n"
"bzhi %q[width],%[temp],%[src] \n"
"kmovq %[src],%%k1 \n"
"xor %k[src],%k[src] \n"
"sub $0x40,%k[width] \n"
"cmovg %[temp],%[src] \n"
"bzhi %q[width],%[src],%[src] \n"
"kmovq %[src],%%k2 \n"
"xor %k[src],%k[src] \n"
"sub $0x40,%k[width] \n"
"cmovg %[temp],%[src] \n"
"bzhi %q[width],%[src],%[src] \n"
"kmovq %[src],%%k3 \n"
"vpshufb %%zmm4,%%zmm0,%%zmm0 \n"
"vpshufb %%zmm4,%%zmm1,%%zmm1 \n"
"vpshufb %%zmm4,%%zmm2,%%zmm2 \n"
"vpshufb %%zmm4,%%zmm3,%%zmm3 \n"
"vpermt2d %%zmm1,%%zmm5,%%zmm0 \n"
"vpermt2d %%zmm2,%%zmm16,%%zmm1 \n"
"vpermt2d %%zmm3,%%zmm17,%%zmm2 \n"
"vmovdqu8 %%zmm0,(%[dst])%{%%k1%} \n"
"vmovdqu8 %%zmm1,0x40(%[dst])%{%%k2%} \n"
"vmovdqu8 %%zmm2,0x80(%[dst])%{%%k3%} \n"
"99: \n"
"vzeroupper \n"
: "+r"(src), // %0
"+r"(dst), // %1
"+r"(width) // %2
: "m"(kPermARGBToRGB24_0), // %3
"m"(kPermARGBToRGB24_1), // %4
"m"(kPermARGBToRGB24_2) // %5
: "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm5", "xmm16",
"xmm17");
: [src] "+r"(src), // %[src]
[dst] "+r"(dst), // %[dst]
[width] "+r"(width), // %[width]
[temp] "=&r"(temp) // %[temp]
: [shuffler] "m"(shuffler),
[kPermdRGB24_0] "m"(kPermdRGB24_0_AVX512BW),
[kPermdRGB24_1] "m"(kPermdRGB24_1_AVX512BW),
[kPermdRGB24_2] "m"(kPermdRGB24_2_AVX512BW)
: "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm16",
"xmm17", "k1", "k2", "k3", "k4");
}
#endif // HAS_ARGBTORGB24ROW_AVX512VBMI
#endif // defined(HAS_ARGBTORGB24ROW_AVX512BW) ||
// defined(HAS_ARGBTORAWROW_AVX512BW)
#ifdef HAS_ARGBTORGB24ROW_AVX512BW
void ARGBToRGB24Row_AVX512BW(const uint8_t* src, uint8_t* dst, int width) {
ARGBToRGBRow_AVX512BW(src, dst, width, kShuffleMaskARGBToRGB24[0]);
}
#endif // HAS_ARGBTORGB24ROW_AVX512BW
#ifdef HAS_ARGBTORAWROW_AVX2
void ARGBToRAWRow_AVX2(const uint8_t* src, uint8_t* dst, int width) {
@ -1118,6 +1174,12 @@ void ARGBToRAWRow_AVX2(const uint8_t* src, uint8_t* dst, int width) {
}
#endif // HAS_ARGBTORAWROW_AVX2
#ifdef HAS_ARGBTORAWROW_AVX512BW
void ARGBToRAWRow_AVX512BW(const uint8_t* src, uint8_t* dst, int width) {
ARGBToRGBRow_AVX512BW(src, dst, width, kShuffleMaskARGBToRAW);
}
#endif // HAS_ARGBTORAWROW_AVX512BW
#ifdef HAS_ARGBTORGB565DITHERROW_AVX2
void ARGBToRGB565DitherRow_AVX2(const uint8_t* src,
uint8_t* dst,