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

[AArch64] Add SVE2 implementation of P210ToARGBRow

Browse Source 
Observed reduction in runtime compared to the existing Neon code:

Cortex-A510: -32.8%
Cortex-A520:  +8.7% (!)
Cortex-A715: -18.9%
Cortex-A720: -18.9%
  Cortex-X2:  -7.9%
  Cortex-X3:  -8.8%
  Cortex-X4:  +1.0% (!)
Cortex-X925:  -8.6%

Bug: b/42280942
Change-Id: Ibe557500c3788b4fb39372c92b2f42ba216e6fea
Reviewed-on: https://chromium-review.googlesource.com/c/libyuv/libyuv/+/5975320
Reviewed-by: Frank Barchard <fbarchard@chromium.org>
Reviewed-by: Justin Green <greenjustin@google.com>
This commit is contained in:
George Steed 2024-05-16 07:50:14 +01:00 committed by Frank Barchard
parent 721ad4aa18
commit b753822d47
3 changed files with 81 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
include/libyuv/row.h
View File

@ -562,6 +562,7 @@ extern "C" {
#define HAS_I444TOARGBROW_SVE2 #define HAS_I444TOARGBROW_SVE2
#define HAS_NV12TOARGBROW_SVE2 #define HAS_NV12TOARGBROW_SVE2
#define HAS_NV21TOARGBROW_SVE2 #define HAS_NV21TOARGBROW_SVE2
#define HAS_P210TOARGBROW_SVE2
#define HAS_RAWTOARGBROW_SVE2 #define HAS_RAWTOARGBROW_SVE2
#define HAS_RAWTORGB24ROW_SVE2 #define HAS_RAWTORGB24ROW_SVE2
#define HAS_RAWTORGBAROW_SVE2 #define HAS_RAWTORGBAROW_SVE2
@ -5376,6 +5377,11 @@ void P210ToARGBRow_NEON(const uint16_t* y_buf,
uint8_t* dst_argb, uint8_t* dst_argb,
const struct YuvConstants* yuvconstants, const struct YuvConstants* yuvconstants,
int width); int width);
void P210ToARGBRow_SVE2(const uint16_t* y_buf,
const uint16_t* uv_buf,
uint8_t* dst_argb,
const struct YuvConstants* yuvconstants,
int width);
void P410ToARGBRow_NEON(const uint16_t* y_buf, void P410ToARGBRow_NEON(const uint16_t* y_buf,
const uint16_t* uv_buf, const uint16_t* uv_buf,
uint8_t* dst_argb, uint8_t* dst_argb,

10
source/convert_argb.cc
View File

@ -1914,6 +1914,11 @@ int P010ToARGBMatrix(const uint16_t* src_y,
P210ToARGBRow = P210ToARGBRow_NEON; P210ToARGBRow = P210ToARGBRow_NEON;
} }
} }
#endif
#if defined(HAS_P210TOARGBROW_SVE2)
if (TestCpuFlag(kCpuHasSVE2)) {
P210ToARGBRow = P210ToARGBRow_SVE2;
}
#endif #endif
for (y = 0; y < height; ++y) { for (y = 0; y < height; ++y) {
P210ToARGBRow(src_y, src_uv, dst_argb, yuvconstants, width); P210ToARGBRow(src_y, src_uv, dst_argb, yuvconstants, width);
@ -1973,6 +1978,11 @@ int P210ToARGBMatrix(const uint16_t* src_y,
P210ToARGBRow = P210ToARGBRow_NEON; P210ToARGBRow = P210ToARGBRow_NEON;
} }
} }
#endif
#if defined(HAS_P210TOARGBROW_SVE2)
if (TestCpuFlag(kCpuHasSVE2)) {
P210ToARGBRow = P210ToARGBRow_SVE2;
}
#endif #endif
for (y = 0; y < height; ++y) { for (y = 0; y < height; ++y) {
P210ToARGBRow(src_y, src_uv, dst_argb, yuvconstants, width); P210ToARGBRow(src_y, src_uv, dst_argb, yuvconstants, width);

65
source/row_sve.cc
View File

@ -78,6 +78,15 @@ extern "C" {
"trn1 z0.b, z0.b, z0.b \n" /* YYYY */ \ "trn1 z0.b, z0.b, z0.b \n" /* YYYY */ \
"tbl z1.b, {z1.b}, z22.b \n" /* UVUV */ "tbl z1.b, {z1.b}, z22.b \n" /* UVUV */
#define READP210_SVE \
"ld1h {z0.h}, p1/z, [%[src_y]] \n" \
"ld1h {z1.h}, p2/z, [%[src_uv]] \n" \
"incb %[src_y] \n" \
"incb %[src_uv] \n" \
"prfm pldl1keep, [%[src_y], 448] \n" \
"prfm pldl1keep, [%[src_uv], 256] \n" \
"tbl z1.b, {z1.b}, z22.b \n"
#define READYUY2_SVE \ #define READYUY2_SVE \
"ld1w {z0.s}, p2/z, [%[src_yuy2]] \n" /* YUYV */ \ "ld1w {z0.s}, p2/z, [%[src_yuy2]] \n" /* YUYV */ \
"incb %[src_yuy2] \n" \ "incb %[src_yuy2] \n" \
@ -1789,6 +1798,62 @@ void HalfFloat1Row_SVE2(const uint16_t* src,
: "cc", "memory", "z0", "z1", "p0", "p1"); : "cc", "memory", "z0", "z1", "p0", "p1");
} }
// P210 has 10 bits in msb of 16 bit NV12 style layout.
void P210ToARGBRow_SVE2(const uint16_t* src_y,
const uint16_t* src_uv,
uint8_t* dst_argb,
const struct YuvConstants* yuvconstants,
int width) {
uint64_t vl;
asm("cnth %0" : "=r"(vl));
int width_last_y = width & (vl - 1);
width_last_y = width_last_y == 0 ? vl : width_last_y;
int width_last_uv = width_last_y + (width_last_y & 1);
uint32_t nv_uv_start = 0x03010301U;
uint32_t nv_uv_step = 0x04040404U;
asm volatile(
"ptrue p0.b \n" YUVTORGB_SVE_SETUP
"index z22.s, %w[nv_uv_start], %w[nv_uv_step] \n"
"dup z19.b, #255 \n" // A
"subs %w[width], %w[width], %w[vl] \n"
"b.lt 2f \n"
// Run bulk of computation with an all-true predicate to avoid predicate
// generation overhead.
"ptrue p1.h \n"
"ptrue p2.h \n"
"1: \n" //
READP210_SVE NVTORGB_SVE RGBTOARGB8_SVE
"subs %w[width], %w[width], %w[vl] \n"
"st2h {z16.h, z17.h}, p1, [%[dst_argb]] \n"
"add %[dst_argb], %[dst_argb], %[vl], lsl #2 \n"
"b.ge 1b \n"
"2: \n"
"adds %w[width], %w[width], %w[vl] \n"
"b.eq 99f \n"
// Calculate a predicate for the final iteration to deal with the tail.
"whilelt p1.h, wzr, %w[width_last_y] \n"
"whilelt p2.h, wzr, %w[width_last_uv] \n" //
READP210_SVE NVTORGB_SVE RGBTOARGB8_SVE
"st2h {z16.h, z17.h}, p1, [%[dst_argb]] \n"
"99: \n"
: [src_y] "+r"(src_y), // %[src_y]
[src_uv] "+r"(src_uv), // %[src_uv]
[dst_argb] "+r"(dst_argb), // %[dst_argb]
[width] "+r"(width) // %[width]
: [vl] "r"(vl), // %[vl]
[kUVCoeff] "r"(&yuvconstants->kUVCoeff), // %[kUVCoeff]
[kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias), // %[kRGBCoeffBias]
[nv_uv_start] "r"(nv_uv_start), // %[nv_uv_start]
[nv_uv_step] "r"(nv_uv_step), // %[nv_uv_step]
[width_last_y] "r"(width_last_y), // %[width_last_y]
[width_last_uv] "r"(width_last_uv) // %[width_last_uv]
: "cc", "memory", YUVTORGB_SVE_REGS);
}
#endif // !defined(LIBYUV_DISABLE_SVE) && defined(__aarch64__) #endif // !defined(LIBYUV_DISABLE_SVE) && defined(__aarch64__)
#ifdef __cplusplus #ifdef __cplusplus