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

NV21ToYUV24 replace ST3 with ST1. ARGBToAR64 replace ST2 with ST1

Browse Source 
On Samsung S8 Exynos M2
Was ST3 NV21ToYUV24_Opt (769 ms)
Now ST1 NV21ToYUV24_Opt (473 ms)
Was ST2 ARGBToAR64_Opt (1759 ms)
Now ST1 ARGBToAR64_Opt (987 ms)

Skylake Xeon, AVX2 version:
Was NV21ToYUV24_Opt (885 ms)
Now NV21ToYUV24_Opt (194 ms)

Bug: b/204562143, b/124413599
Change-Id: Icc9cb64d822cd11937789a4e04fbb773b3e33aa3
Reviewed-on: https://chromium-review.googlesource.com/c/libyuv/libyuv/+/3290664
Commit-Queue: Frank Barchard <fbarchard@chromium.org>
Reviewed-by: richard winterton <rrwinterton@gmail.com>
This commit is contained in:
Frank Barchard 2021-11-23 15:45:19 -08:00 committed by libyuv LUCI CQ
parent a04e4f87fb
commit 000806f373
7 changed files with 223 additions and 134 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: http://code.google.com/p/libyuv/ URL: http://code.google.com/p/libyuv/
Version: 1803 Version: 1805
License: BSD License: BSD
License File: LICENSE License File: LICENSE

13
include/libyuv/row.h
View File

@ -303,6 +303,7 @@ extern "C" {
#define HAS_MERGEXRGBROW_SSE2 #define HAS_MERGEXRGBROW_SSE2
#define HAS_MERGERGBROW_SSSE3 #define HAS_MERGERGBROW_SSSE3
#define HAS_MIRRORUVROW_SSSE3 #define HAS_MIRRORUVROW_SSSE3
#define HAS_NV21TOYUV24ROW_SSSE3
#define HAS_P210TOAR30ROW_SSSE3 #define HAS_P210TOAR30ROW_SSSE3
#define HAS_P210TOARGBROW_SSSE3 #define HAS_P210TOARGBROW_SSSE3
#define HAS_P410TOAR30ROW_SSSE3 #define HAS_P410TOAR30ROW_SSSE3
@ -319,6 +320,7 @@ extern "C" {
#define HAS_SPLITRGBROW_SSSE3 #define HAS_SPLITRGBROW_SSSE3
#define HAS_SWAPUVROW_SSSE3 #define HAS_SWAPUVROW_SSSE3
#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
// https://code.google.com/p/libyuv/issues/detail?id=517 // https://code.google.com/p/libyuv/issues/detail?id=517
@ -355,6 +357,7 @@ extern "C" {
#define HAS_MERGEXR64ROW_AVX2 #define HAS_MERGEXR64ROW_AVX2
#define HAS_MERGEXRGB16TO8ROW_AVX2 #define HAS_MERGEXRGB16TO8ROW_AVX2
#define HAS_MERGEXRGBROW_AVX2 #define HAS_MERGEXRGBROW_AVX2
#define HAS_NV21TOYUV24ROW_AVX2
#define HAS_I210TOAR30ROW_AVX2 #define HAS_I210TOAR30ROW_AVX2
#define HAS_I210TOARGBROW_AVX2 #define HAS_I210TOARGBROW_AVX2
#define HAS_I212TOAR30ROW_AVX2 #define HAS_I212TOAR30ROW_AVX2
@ -379,8 +382,6 @@ extern "C" {
#define HAS_SPLITXRGBROW_AVX2 #define HAS_SPLITXRGBROW_AVX2
#define HAS_SPLITUVROW_16_AVX2 #define HAS_SPLITUVROW_16_AVX2
#define HAS_SWAPUVROW_AVX2 #define HAS_SWAPUVROW_AVX2
// TODO(fbarchard): Fix AVX2 version of YUV24
// #define HAS_NV21TOYUV24ROW_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
@ -3227,6 +3228,10 @@ void NV21ToRGB24Row_AVX2(const uint8_t* src_y,
uint8_t* dst_rgb24, uint8_t* dst_rgb24,
const struct YuvConstants* yuvconstants, const struct YuvConstants* yuvconstants,
int width); int width);
void NV21ToYUV24Row_SSSE3(const uint8_t* src_y,
const uint8_t* src_vu,
uint8_t* dst_yuv24,
int width);
void NV21ToYUV24Row_AVX2(const uint8_t* src_y, void NV21ToYUV24Row_AVX2(const uint8_t* src_y,
const uint8_t* src_vu, const uint8_t* src_vu,
uint8_t* dst_yuv24, uint8_t* dst_yuv24,
@ -3568,6 +3573,10 @@ void NV21ToRGB24Row_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 NV21ToYUV24Row_Any_SSSE3(const uint8_t* src_y,
const uint8_t* src_vu,
uint8_t* dst_yuv24,
int width);
void NV21ToYUV24Row_Any_AVX2(const uint8_t* src_y, void NV21ToYUV24Row_Any_AVX2(const uint8_t* src_y,
const uint8_t* src_vu, const uint8_t* src_vu,
uint8_t* dst_yuv24, uint8_t* dst_yuv24,

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 1803 #define LIBYUV_VERSION 1805
#endif // INCLUDE_LIBYUV_VERSION_H_ #endif // INCLUDE_LIBYUV_VERSION_H_

8
source/convert_argb.cc
View File

@ -3925,6 +3925,14 @@ int NV21ToYUV24(const uint8_t* src_y,
} }
} }
#endif #endif
#if defined(HAS_NV21TOYUV24ROW_SSSE3)
if (TestCpuFlag(kCpuHasSSSE3)) {
NV21ToYUV24Row = NV21ToYUV24Row_Any_SSSE3;
if (IS_ALIGNED(width, 16)) {
NV21ToYUV24Row = NV21ToYUV24Row_SSSE3;
}
}
#endif
#if defined(HAS_NV21TOYUV24ROW_AVX2) #if defined(HAS_NV21TOYUV24ROW_AVX2)
if (TestCpuFlag(kCpuHasAVX2)) { if (TestCpuFlag(kCpuHasAVX2)) {
NV21ToYUV24Row = NV21ToYUV24Row_Any_AVX2; NV21ToYUV24Row = NV21ToYUV24Row_Any_AVX2;

33
source/row_any.cc
View File

@ -545,21 +545,21 @@ ANY31PT(MergeXRGB16To8Row_Any_NEON,
#undef ANY31PT #undef ANY31PT
// Any 2 planes to 1. // Any 2 planes to 1.
#define ANY21(NAMEANY, ANY_SIMD, UVSHIFT, SBPP, SBPP2, BPP, MASK) \ #define ANY21(NAMEANY, ANY_SIMD, UVSHIFT, SBPP, SBPP2, BPP, MASK) \
void NAMEANY(const uint8_t* y_buf, const uint8_t* uv_buf, uint8_t* dst_ptr, \ void NAMEANY(const uint8_t* y_buf, const uint8_t* uv_buf, uint8_t* dst_ptr, \
int width) { \ int width) { \
SIMD_ALIGNED(uint8_t temp[64 * 3]); \ SIMD_ALIGNED(uint8_t temp[128 * 3]); \
memset(temp, 0, 64 * 2); /* for msan */ \ memset(temp, 0, 128 * 2); /* for msan */ \
int r = width & MASK; \ int r = width & MASK; \
int n = width & ~MASK; \ int n = width & ~MASK; \
if (n > 0) { \ if (n > 0) { \
ANY_SIMD(y_buf, uv_buf, dst_ptr, n); \ ANY_SIMD(y_buf, uv_buf, dst_ptr, n); \
} \ } \
memcpy(temp, y_buf + n * SBPP, r * SBPP); \ memcpy(temp, y_buf + n * SBPP, r * SBPP); \
memcpy(temp + 64, uv_buf + (n >> UVSHIFT) * SBPP2, \ memcpy(temp + 128, uv_buf + (n >> UVSHIFT) * SBPP2, \
SS(r, UVSHIFT) * SBPP2); \ SS(r, UVSHIFT) * SBPP2); \
ANY_SIMD(temp, temp + 64, temp + 128, MASK + 1); \ ANY_SIMD(temp, temp + 128, temp + 256, MASK + 1); \
memcpy(dst_ptr + n * BPP, temp + 128, r * BPP); \ memcpy(dst_ptr + n * BPP, temp + 256, r * BPP); \
} }
// Merge functions. // Merge functions.
@ -581,6 +581,9 @@ ANY21(MergeUVRow_Any_MMI, MergeUVRow_MMI, 0, 1, 1, 2, 7)
#ifdef HAS_NV21TOYUV24ROW_NEON #ifdef HAS_NV21TOYUV24ROW_NEON
ANY21(NV21ToYUV24Row_Any_NEON, NV21ToYUV24Row_NEON, 1, 1, 2, 3, 15) ANY21(NV21ToYUV24Row_Any_NEON, NV21ToYUV24Row_NEON, 1, 1, 2, 3, 15)
#endif #endif
#ifdef HAS_NV21TOYUV24ROW_SSSE3
ANY21(NV21ToYUV24Row_Any_SSSE3, NV21ToYUV24Row_SSSE3, 1, 1, 2, 3, 15)
#endif
#ifdef HAS_NV21TOYUV24ROW_AVX2 #ifdef HAS_NV21TOYUV24ROW_AVX2
ANY21(NV21ToYUV24Row_Any_AVX2, NV21ToYUV24Row_AVX2, 1, 1, 2, 3, 31) ANY21(NV21ToYUV24Row_Any_AVX2, NV21ToYUV24Row_AVX2, 1, 1, 2, 3, 31)
#endif #endif

232
source/row_gcc.cc
View File

@ -8894,127 +8894,135 @@ void ARGBLumaColorTableRow_SSSE3(const uint8_t* src_argb,
} }
#endif // HAS_ARGBLUMACOLORTABLEROW_SSSE3 #endif // HAS_ARGBLUMACOLORTABLEROW_SSSE3
#ifdef HAS_NV21TOYUV24ROW_AVX2 static const uvec8 kYUV24Shuffle[3] =
{{ 8, 9, 0, 8, 9, 1, 10, 11, 2, 10, 11, 3, 12, 13, 4, 12 },
{ 9, 1, 10, 11, 2, 10, 11, 3, 12, 13, 4, 12, 13, 5, 14, 15 },
{ 2, 10, 11, 3, 12, 13, 4, 12, 13, 5, 14, 15, 6, 14, 15, 7 }};
// begin NV21ToYUV24Row_C avx2 constants // Convert biplanar NV21 to packed YUV24
static const ulvec8 kBLEND0 = {0x80, 0x00, 0x80, 0x80, 0x00, 0x80, 0x80, 0x00, // NV21 has VU in memory for chroma.
0x80, 0x80, 0x00, 0x80, 0x80, 0x00, 0x80, 0x80, // YUV24 is VUY in memory
0x00, 0x80, 0x00, 0x00, 0x80, 0x00, 0x00, 0x80, void NV21ToYUV24Row_SSSE3(const uint8_t* src_y,
0x00, 0x00, 0x80, 0x00, 0x00, 0x80, 0x00, 0x00}; const uint8_t* src_vu,
uint8_t* dst_yuv24,
int width) {
asm volatile(
"sub %0,%1 \n"
"movdqa (%4),%%xmm4 \n" // 3 shuffler constants
"movdqa 16(%4),%%xmm5 \n"
"movdqa 32(%4),%%xmm6 \n"
"1: \n"
"movdqu (%0),%%xmm2 \n" // load 16 Y values
"movdqu (%0,%1),%%xmm3 \n" // load 8 VU values
"lea 16(%0),%0 \n"
"movdqa %%xmm2,%%xmm0 \n"
"movdqa %%xmm2,%%xmm1 \n"
"shufps $0x44,%%xmm3,%%xmm0 \n" // Y 0..7, UV 0..3
"shufps $0x99,%%xmm3,%%xmm1 \n" // Y 4..11, UV 2..5
"shufps $0xee,%%xmm3,%%xmm2 \n" // Y 8..15, UV 4..7
"pshufb %%xmm4, %%xmm0 \n" // weave into YUV24
"pshufb %%xmm5, %%xmm1 \n"
"pshufb %%xmm6, %%xmm2 \n"
"movdqu %%xmm0,(%2) \n"
"movdqu %%xmm1,16(%2) \n"
"movdqu %%xmm2,32(%2) \n"
"lea 48(%2),%2 \n"
"sub $16,%3 \n" // 16 pixels per loop
"jg 1b \n"
: "+r"(src_y), // %0
"+r"(src_vu), // %1
"+r"(dst_yuv24), // %2
"+r"(width) // %3
: "r"(&kYUV24Shuffle[0]) // %4
: "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6");
}
static const ulvec8 kBLEND1 = {0x00, 0x00, 0x80, 0x00, 0x00, 0x80, 0x00, 0x00, // Convert biplanar NV21 to packed YUV24
0x80, 0x00, 0x00, 0x80, 0x00, 0x00, 0x80, 0x00, // NV21 has VU in memory for chroma.
0x80, 0x00, 0x00, 0x80, 0x00, 0x00, 0x80, 0x00, // YUV24 is VUY in memory
0x00, 0x80, 0x00, 0x00, 0x80, 0x00, 0x00, 0x80};
static const ulvec8 kBLEND2 = {0x80, 0x00, 0x00, 0x80, 0x00, 0x00, 0x80, 0x00,
0x00, 0x80, 0x00, 0x00, 0x80, 0x00, 0x00, 0x80,
0x00, 0x00, 0x80, 0x00, 0x00, 0x80, 0x00, 0x00,
0x80, 0x00, 0x00, 0x80, 0x00, 0x00, 0x80, 0x00};
static const ulvec8 kSHUF0 = {0x00, 0x0b, 0x80, 0x01, 0x0c, 0x80, 0x02, 0x0d,
0x80, 0x03, 0x0e, 0x80, 0x04, 0x0f, 0x80, 0x05,
0x00, 0x0b, 0x80, 0x01, 0x0c, 0x80, 0x02, 0x0d,
0x80, 0x03, 0x0e, 0x80, 0x04, 0x0f, 0x80, 0x05};
static const ulvec8 kSHUF1 = {0x80, 0x00, 0x0b, 0x80, 0x01, 0x0c, 0x80, 0x02,
0x0d, 0x80, 0x03, 0x0e, 0x80, 0x04, 0x0f, 0x80,
0x80, 0x00, 0x0b, 0x80, 0x01, 0x0c, 0x80, 0x02,
0x0d, 0x80, 0x03, 0x0e, 0x80, 0x04, 0x0f, 0x80};
static const ulvec8 kSHUF2 = {0x0a, 0x80, 0x00, 0x0b, 0x80, 0x01, 0x0c, 0x80,
0x02, 0x0d, 0x80, 0x03, 0x0e, 0x80, 0x04, 0x0f,
0x0a, 0x80, 0x00, 0x0b, 0x80, 0x01, 0x0c, 0x80,
0x02, 0x0d, 0x80, 0x03, 0x0e, 0x80, 0x04, 0x0f};
static const ulvec8 kSHUF3 = {0x80, 0x80, 0x06, 0x80, 0x80, 0x07, 0x80, 0x80,
0x08, 0x80, 0x80, 0x09, 0x80, 0x80, 0x0a, 0x80,
0x80, 0x80, 0x06, 0x80, 0x80, 0x07, 0x80, 0x80,
0x08, 0x80, 0x80, 0x09, 0x80, 0x80, 0x0a, 0x80};
static const ulvec8 kSHUF4 = {0x05, 0x80, 0x80, 0x06, 0x80, 0x80, 0x07, 0x80,
0x80, 0x08, 0x80, 0x80, 0x09, 0x80, 0x80, 0x0a,
0x05, 0x80, 0x80, 0x06, 0x80, 0x80, 0x07, 0x80,
0x80, 0x08, 0x80, 0x80, 0x09, 0x80, 0x80, 0x0a};
static const ulvec8 kSHUF5 = {0x80, 0x05, 0x80, 0x80, 0x06, 0x80, 0x80, 0x07,
0x80, 0x80, 0x08, 0x80, 0x80, 0x09, 0x80, 0x80,
0x80, 0x05, 0x80, 0x80, 0x06, 0x80, 0x80, 0x07,
0x80, 0x80, 0x08, 0x80, 0x80, 0x09, 0x80, 0x80};
// NV21ToYUV24Row_AVX2
void NV21ToYUV24Row_AVX2(const uint8_t* src_y, void NV21ToYUV24Row_AVX2(const uint8_t* src_y,
const uint8_t* src_vu, const uint8_t* src_vu,
uint8_t* dst_yuv24, uint8_t* dst_yuv24,
int width) { int width) {
uint8_t* src_y_ptr;
uint64_t src_offset = 0;
uint64_t width64;
width64 = width;
src_y_ptr = (uint8_t*)src_y;
asm volatile( asm volatile(
"vmovdqu %5, %%ymm0 \n" // init blend value "sub %0,%1 \n"
"vmovdqu %6, %%ymm1 \n" // init blend value "vbroadcastf128 (%4),%%ymm4 \n" // 3 shuffler constants
"vmovdqu %7, %%ymm2 \n" // init blend value "vbroadcastf128 16(%4),%%ymm5 \n"
// "sub $0x20, %3 \n" //sub 32 from "vbroadcastf128 32(%4),%%ymm6 \n"
// width for final loop
LABELALIGN "1: \n"
"1: \n" // label 1 "vmovdqu (%0),%%ymm2 \n" // load 32 Y values
"vmovdqu (%0,%4), %%ymm3 \n" // src_y "vmovdqu (%0,%1),%%ymm3 \n" // load 16 VU values
"vmovdqu 1(%1,%4), %%ymm4 \n" // src_uv+1 "lea 32(%0),%0 \n"
"vmovdqu (%1), %%ymm5 \n" // src_uv "vshufps $0x44,%%ymm3,%%ymm2,%%ymm0 \n" // Y 0..7, UV 0..3
"vpshufb %8, %%ymm3, %%ymm13 \n" // y, kSHUF0 for shuf "vshufps $0x99,%%ymm3,%%ymm2,%%ymm1 \n" // Y 4..11, UV 2..5
"vpshufb %9, %%ymm4, %%ymm14 \n" // uv+1, kSHUF1 for "vshufps $0xee,%%ymm3,%%ymm2,%%ymm2 \n" // Y 8..15, UV 4..7
// shuf "vpshufb %%ymm4,%%ymm0,%%ymm0 \n" // weave into YUV24
"vpshufb %10, %%ymm5, %%ymm15 \n" // uv, kSHUF2 for "vpshufb %%ymm5,%%ymm1,%%ymm1 \n"
// shuf "vpshufb %%ymm6,%%ymm2,%%ymm2 \n"
"vpshufb %11, %%ymm3, %%ymm3 \n" // y kSHUF3 for shuf "vperm2i128 $0x20,%%ymm1,%%ymm0,%%ymm3 \n"
"vpshufb %12, %%ymm4, %%ymm4 \n" // uv+1 kSHUF4 for "vperm2i128 $0x30,%%ymm0,%%ymm2,%%ymm0 \n"
// shuf "vperm2i128 $0x31,%%ymm2,%%ymm1,%%ymm1 \n"
"vpblendvb %%ymm0, %%ymm14, %%ymm13, %%ymm12 \n" // blend 0 "vmovdqu %%ymm3,(%2) \n"
"vpblendvb %%ymm0, %%ymm13, %%ymm14, %%ymm14 \n" // blend 0 "vmovdqu %%ymm0,32(%2) \n"
"vpblendvb %%ymm2, %%ymm15, %%ymm12, %%ymm12 \n" // blend 2 "vmovdqu %%ymm1,64(%2) \n"
"vpblendvb %%ymm1, %%ymm15, %%ymm14, %%ymm13 \n" // blend 1 "lea 96(%2),%2 \n"
"vpshufb %13, %%ymm5, %%ymm15 \n" // shuffle const "sub $32,%3 \n" // 32 pixels per loop
"vpor %%ymm4, %%ymm3, %%ymm5 \n" // get results "jg 1b \n"
"vmovdqu %%ymm12, 0x20(%2) \n" // store dst_yuv+20h "vzeroupper \n"
"vpor %%ymm15, %%ymm5, %%ymm3 \n" // get results : "+r"(src_y), // %0
"add $0x20, %4 \n" // add to src buffer "+r"(src_vu), // %1
// ptr "+r"(dst_yuv24), // %2
"vinserti128 $0x1, %%xmm3, %%ymm13, %%ymm4 \n" // insert "+r"(width) // %3
"vperm2i128 $0x31, %%ymm13, %%ymm3, %%ymm5 \n" // insert : "r"(&kYUV24Shuffle[0]) // %4
"vmovdqu %%ymm4, (%2) \n" // store dst_yuv : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6");
"vmovdqu %%ymm5, 0x40(%2) \n" // store dst_yuv+40h
"add $0x60,%2 \n" // add to dst buffer
// ptr
// "cmp %3, %4 \n" //(width64 -
// 32 bytes) and src_offset
"sub $0x20,%3 \n" // 32 pixels per loop
"jg 1b \n"
"vzeroupper \n" // sse-avx2
// transistions
: "+r"(src_y), //%0
"+r"(src_vu), //%1
"+r"(dst_yuv24), //%2
"+r"(width64), //%3
"+r"(src_offset) //%4
: "m"(kBLEND0), //%5
"m"(kBLEND1), //%6
"m"(kBLEND2), //%7
"m"(kSHUF0), //%8
"m"(kSHUF1), //%9
"m"(kSHUF2), //%10
"m"(kSHUF3), //%11
"m"(kSHUF4), //%12
"m"(kSHUF5) //%13
: "memory", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm12",
"xmm13", "xmm14", "xmm15");
} }
#endif // HAS_NV21TOYUV24ROW_AVX2
#ifdef HAS_NV21ToYUV24ROW_AVX512
// The following VMBI VEX256 code tests okay with the intelsde emulator.
static const lvec8 kYUV24Perm[3] =
{{ 32, 33, 0, 32, 33, 1, 34, 35, 2, 34, 35, 3, 36, 37, 4, 36,
37, 5, 38, 39, 6, 38, 39, 7, 40, 41, 8, 40, 41, 9, 42, 43 },
{ 10, 42, 43, 11, 44, 45, 12, 44, 45, 13, 46, 47, 14, 46, 47, 15,
48, 49, 16, 48, 49, 17, 50, 51, 18, 50, 51, 19, 52, 53, 20, 52 },
{ 53, 21, 54, 55, 22, 54, 55, 23, 56, 57, 24, 56, 57, 25, 58, 59,
26, 58, 59, 27, 60, 61, 28, 60, 61, 29, 62, 63, 30, 62, 63, 31 }};
void NV21ToYUV24Row_AVX512(const uint8_t* src_y,
const uint8_t* src_vu,
uint8_t* dst_yuv24,
int width) {
asm volatile(
"sub %0,%1 \n"
"vmovdqa (%4),%%ymm4 \n" // 3 shuffler constants
"vmovdqa 32(%4),%%ymm5 \n"
"vmovdqa 64(%4),%%ymm6 \n"
LABELALIGN
"1: \n"
"vmovdqu (%0),%%ymm2 \n" // load 32 Y values
"vmovdqu (%0,%1),%%ymm3 \n" // load 16 VU values
"lea 32(%0),%0 \n"
"vmovdqa %%ymm2, %%ymm0 \n"
"vmovdqa %%ymm2, %%ymm1 \n"
"vpermt2b %%ymm3,%%ymm4,%%ymm0 \n"
"vpermt2b %%ymm3,%%ymm5,%%ymm1 \n"
"vpermt2b %%ymm3,%%ymm6,%%ymm2 \n"
"vmovdqu %%ymm0,(%2) \n"
"vmovdqu %%ymm1,32(%2) \n"
"vmovdqu %%ymm2,64(%2) \n"
"lea 96(%2),%2 \n"
"sub $32,%3 \n"
"jg 1b \n"
"vzeroupper \n"
: "+r"(src_y), // %0
"+r"(src_vu), // %1
"+r"(dst_yuv24), // %2
"+r"(width) // %3
: "r"(&kYUV24Perm[0]) // %4
: "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6");
}
#endif // HAS_NV21ToYUV24ROW_AVX512
#ifdef HAS_SWAPUVROW_SSSE3 #ifdef HAS_SWAPUVROW_SSSE3

67
source/row_neon64.cc
View File

@ -15,6 +15,9 @@ namespace libyuv {
extern "C" { extern "C" {
#endif #endif
// Enable LIBYUV_USE_ST2 and LIBYUV_USE_ST3 for CPUs that prefer them.
// Exynos M1, M2, M3 are slow with ST2, ST3 and ST4 instructions.
// This module is for GCC Neon armv8 64 bit. // This module is for GCC Neon armv8 64 bit.
#if !defined(LIBYUV_DISABLE_NEON) && defined(__aarch64__) #if !defined(LIBYUV_DISABLE_NEON) && defined(__aarch64__)
@ -1683,6 +1686,7 @@ void ARGBToARGB4444Row_NEON(const uint8_t* src_argb,
: "cc", "memory", "v0", "v1", "v16", "v17", "v18", "v19", "v23"); : "cc", "memory", "v0", "v1", "v16", "v17", "v18", "v19", "v23");
} }
#if LIBYUV_USE_ST2
void ARGBToAR64Row_NEON(const uint8_t* src_argb, void ARGBToAR64Row_NEON(const uint8_t* src_argb,
uint16_t* dst_ar64, uint16_t* dst_ar64,
int width) { int width) {
@ -1702,6 +1706,28 @@ void ARGBToAR64Row_NEON(const uint8_t* src_argb,
: :
: "cc", "memory", "v0", "v1", "v2", "v3"); : "cc", "memory", "v0", "v1", "v2", "v3");
} }
#else
void ARGBToAR64Row_NEON(const uint8_t* src_argb,
uint16_t* dst_ar64,
int width) {
asm volatile(
"1: \n"
"ldp q0, q1, [%0], #32 \n" // load 8 ARGB pixels
"subs %w2, %w2, #8 \n" // 8 processed per loop.
"zip1 v2.16b, v0.16b, v0.16b \n"
"zip2 v3.16b, v0.16b, v0.16b \n"
"prfm pldl1keep, [%0, 448] \n"
"zip1 v4.16b, v1.16b, v1.16b \n"
"zip2 v5.16b, v1.16b, v1.16b \n"
"st1 {v2.16b, v3.16b, v4.16b, v5.16b}, [%1], #64 \n" // 8 AR64
"b.gt 1b \n"
: "+r"(src_argb), // %0
"+r"(dst_ar64), // %1
"+r"(width) // %2
:
: "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5");
}
#endif // LIBYUV_USE_ST2
static const uvec8 kShuffleARGBToABGR = {2, 1, 0, 3, 6, 5, 4, 7, static const uvec8 kShuffleARGBToABGR = {2, 1, 0, 3, 6, 5, 4, 7,
10, 9, 8, 11, 14, 13, 12, 15}; 10, 9, 8, 11, 14, 13, 12, 15};
@ -3669,6 +3695,7 @@ void GaussRow_F32_NEON(const float* src, float* dst, int width) {
: "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8"); : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8");
} }
#if LIBYUV_USE_ST3
// Convert biplanar NV21 to packed YUV24 // Convert biplanar NV21 to packed YUV24
void NV21ToYUV24Row_NEON(const uint8_t* src_y, void NV21ToYUV24Row_NEON(const uint8_t* src_y,
const uint8_t* src_vu, const uint8_t* src_vu,
@ -3692,8 +3719,42 @@ void NV21ToYUV24Row_NEON(const uint8_t* src_y,
: :
: "cc", "memory", "v0", "v1", "v2"); : "cc", "memory", "v0", "v1", "v2");
} }
#else
static const uvec8 kYUV24Shuffle[3] =
{{ 16, 17, 0, 16, 17, 1, 18, 19, 2, 18, 19, 3, 20, 21, 4, 20 },
{ 21, 5, 22, 23, 6, 22, 23, 7, 24, 25, 8, 24, 25, 9, 26, 27 },
{ 10, 26, 27, 11, 28, 29, 12, 28, 29, 13, 30, 31, 14, 30, 31, 15 }};
// AYUV is YVUA in memory. UV for NV12 is UV order in memory. // Convert biplanar NV21 to packed YUV24
// NV21 has VU in memory for chroma.
// YUV24 is VUY in memory
void NV21ToYUV24Row_NEON(const uint8_t* src_y,
const uint8_t* src_vu,
uint8_t* dst_yuv24,
int width) {
asm volatile(
"ld1 {v5.16b,v6.16b,v7.16b}, [%4]\n" // 3 shuffler constants
"1: \n"
"ld1 {v0.16b}, [%0], #16 \n" // load 16 Y values
"ld1 {v1.16b}, [%1], #16 \n" // load 8 VU values
"tbl v2.16b, {v0.16b,v1.16b}, v5.16b\n" // weave into YUV24
"prfm pldl1keep, [%0, 448] \n"
"tbl v3.16b, {v0.16b,v1.16b}, v6.16b\n"
"prfm pldl1keep, [%1, 448] \n"
"tbl v4.16b, {v0.16b,v1.16b}, v7.16b\n"
"subs %w3, %w3, #16 \n" // 16 pixels per loop
"st1 {v2.16b,v3.16b,v4.16b}, [%2], #48\n" // store 16 YUV pixels
"b.gt 1b \n"
: "+r"(src_y), // %0
"+r"(src_vu), // %1
"+r"(dst_yuv24), // %2
"+r"(width) // %3
: "r"(&kYUV24Shuffle[0]) // %4
: "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7");
}
#endif // LIBYUV_USE_ST3
// AYUV is VUYA in memory. UV for NV12 is UV order in memory.
void AYUVToUVRow_NEON(const uint8_t* src_ayuv, void AYUVToUVRow_NEON(const uint8_t* src_ayuv,
int src_stride_ayuv, int src_stride_ayuv,
uint8_t* dst_uv, uint8_t* dst_uv,
@ -3708,8 +3769,8 @@ void AYUVToUVRow_NEON(const uint8_t* src_ayuv,
"uaddlp v1.8h, v1.16b \n" // U 16 bytes -> 8 shorts. "uaddlp v1.8h, v1.16b \n" // U 16 bytes -> 8 shorts.
"ld4 {v4.16b,v5.16b,v6.16b,v7.16b}, [%1], #64 \n" // load next 16 "ld4 {v4.16b,v5.16b,v6.16b,v7.16b}, [%1], #64 \n" // load next 16
"uadalp v0.8h, v4.16b \n" // V 16 bytes -> 8 shorts. "uadalp v0.8h, v4.16b \n" // V 16 bytes -> 8 shorts.
"prfm pldl1keep, [%1, 448] \n"
"uadalp v1.8h, v5.16b \n" // U 16 bytes -> 8 shorts. "uadalp v1.8h, v5.16b \n" // U 16 bytes -> 8 shorts.
"prfm pldl1keep, [%1, 448] \n"
"uqrshrn v3.8b, v0.8h, #2 \n" // 2x2 average "uqrshrn v3.8b, v0.8h, #2 \n" // 2x2 average
"uqrshrn v2.8b, v1.8h, #2 \n" "uqrshrn v2.8b, v1.8h, #2 \n"
"subs %w3, %w3, #16 \n" // 16 processed per loop. "subs %w3, %w3, #16 \n" // 16 processed per loop.
@ -3737,8 +3798,8 @@ void AYUVToVURow_NEON(const uint8_t* src_ayuv,
"uaddlp v1.8h, v1.16b \n" // U 16 bytes -> 8 shorts. "uaddlp v1.8h, v1.16b \n" // U 16 bytes -> 8 shorts.
"ld4 {v4.16b,v5.16b,v6.16b,v7.16b}, [%1], #64 \n" // load next 16 "ld4 {v4.16b,v5.16b,v6.16b,v7.16b}, [%1], #64 \n" // load next 16
"uadalp v0.8h, v4.16b \n" // V 16 bytes -> 8 shorts. "uadalp v0.8h, v4.16b \n" // V 16 bytes -> 8 shorts.
"prfm pldl1keep, [%1, 448] \n"
"uadalp v1.8h, v5.16b \n" // U 16 bytes -> 8 shorts. "uadalp v1.8h, v5.16b \n" // U 16 bytes -> 8 shorts.
"prfm pldl1keep, [%1, 448] \n"
"uqrshrn v0.8b, v0.8h, #2 \n" // 2x2 average "uqrshrn v0.8b, v0.8h, #2 \n" // 2x2 average
"uqrshrn v1.8b, v1.8h, #2 \n" "uqrshrn v1.8b, v1.8h, #2 \n"
"subs %w3, %w3, #16 \n" // 16 processed per loop. "subs %w3, %w3, #16 \n" // 16 processed per loop.