coffee/libyuv
1
0
Fork 0
mirror of https://chromium.googlesource.com/libyuv/libyuv synced 2025-12-07 01:06:46 +08:00
Code Issues Packages Projects Releases Wiki Activity

[AArch64] Add SVE2 implementation of ARGBToRGB565DitherRow

Browse Source 
Observed performance improvements compared to the existing Neon
implementation:

Cortex-A510: -21.7%
Cortex-A720: -49.2%
  Cortex-X2: -62.6%

Bug: libyuv:973
Change-Id: I2c7ae483c0b488a122bb3b80a745412ed44622df
Reviewed-on: https://chromium-review.googlesource.com/c/libyuv/libyuv/+/5505539
Reviewed-by: Frank Barchard <fbarchard@chromium.org>
Reviewed-by: Justin Green <greenjustin@google.com>
This commit is contained in:
George Steed 2024-04-15 15:11:38 +01:00 committed by Frank Barchard
parent dff7bad43d
commit 250e1e1ba3
4 changed files with 62 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

5
include/libyuv/row.h
View File

@ -575,6 +575,7 @@ extern "C" {
#if !defined(LIBYUV_DISABLE_SVE) && defined(__aarch64__) #if !defined(LIBYUV_DISABLE_SVE) && defined(__aarch64__)
#define HAS_ABGRTOUVJROW_SVE2 #define HAS_ABGRTOUVJROW_SVE2
#define HAS_ABGRTOUVROW_SVE2 #define HAS_ABGRTOUVROW_SVE2
#define HAS_ARGBTORGB565DITHERROW_SVE2
#define HAS_ARGBTORGB565ROW_SVE2 #define HAS_ARGBTORGB565ROW_SVE2
#define HAS_ARGBTOUVJROW_SVE2 #define HAS_ARGBTOUVJROW_SVE2
#define HAS_ARGBTOUVROW_SVE2 #define HAS_ARGBTOUVROW_SVE2
@ -3653,6 +3654,10 @@ void ARGBToRGB565DitherRow_NEON(const uint8_t* src_argb,
uint8_t* dst_rgb, uint8_t* dst_rgb,
uint32_t dither4, uint32_t dither4,
int width); int width);
void ARGBToRGB565DitherRow_SVE2(const uint8_t* src_argb,
uint8_t* dst_rgb,
uint32_t dither4,
int width);
void ARGBToRGB24Row_MSA(const uint8_t* src_argb, uint8_t* dst_rgb, int width); void ARGBToRGB24Row_MSA(const uint8_t* src_argb, uint8_t* dst_rgb, int width);
void ARGBToRAWRow_MSA(const uint8_t* src_argb, uint8_t* dst_rgb, int width); void ARGBToRAWRow_MSA(const uint8_t* src_argb, uint8_t* dst_rgb, int width);
void ARGBToRGB565Row_MSA(const uint8_t* src_argb, uint8_t* dst_rgb, int width); void ARGBToRGB565Row_MSA(const uint8_t* src_argb, uint8_t* dst_rgb, int width);

5
source/convert_argb.cc
View File

@ -6088,6 +6088,11 @@ int I420ToRGB565Dither(const uint8_t* src_y,
} }
} }
#endif #endif
#if defined(HAS_ARGBTORGB565DITHERROW_SVE2)
if (TestCpuFlag(kCpuHasSVE2)) {
ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_SVE2;
}
#endif
#if defined(HAS_ARGBTORGB565DITHERROW_MSA) #if defined(HAS_ARGBTORGB565DITHERROW_MSA)
if (TestCpuFlag(kCpuHasMSA)) { if (TestCpuFlag(kCpuHasMSA)) {
ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_Any_MSA; ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_Any_MSA;

5
source/convert_from_argb.cc
View File

@ -1960,6 +1960,11 @@ int ARGBToRGB565Dither(const uint8_t* src_argb,
} }
} }
#endif #endif
#if defined(HAS_ARGBTORGB565DITHERROW_SVE2)
if (TestCpuFlag(kCpuHasSVE2)) {
ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_SVE2;
}
#endif
#if defined(HAS_ARGBTORGB565DITHERROW_MSA) #if defined(HAS_ARGBTORGB565DITHERROW_MSA)
if (TestCpuFlag(kCpuHasMSA)) { if (TestCpuFlag(kCpuHasMSA)) {
ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_Any_MSA; ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_Any_MSA;

47
source/row_sve.cc
View File

@ -550,6 +550,53 @@ void ARGBToRGB565Row_SVE2(const uint8_t* src_argb,
: "cc", "memory", "z0", "z1", "z3", "z4", "p0"); : "cc", "memory", "z0", "z1", "z3", "z4", "p0");
} }
void ARGBToRGB565DitherRow_SVE2(const uint8_t* src_argb,
uint8_t* dst_rgb,
uint32_t dither4,
int width) {
unsigned bsl_mask = 0x7e0;
uint64_t vl;
width *= 2;
asm("mov z3.h, #3 \n"
"dup z4.h, %w[bsl_mask] \n"
"dup z2.s, %w[dither4] \n"
"zip1 z2.b, z2.b, z2.b \n"
"cntb %[vl] \n"
"subs %w[width], %w[width], %w[vl] \n"
"b.lt 2f \n"
"ptrue p0.b \n"
"1: \n"
"ld2b {z0.b, z1.b}, p0/z, [%[src]] \n" // BR, GA
"incb %[src], all, mul #2 \n"
"uqadd z0.b, z0.b, z2.b \n"
"uqadd z1.b, z1.b, z2.b \n"
"subs %w[width], %w[width], %w[vl] \n" ARGBTORGB565_SVE
"st1b {z1.b}, p0, [%[dst]] \n"
"incb %[dst] \n"
"b.ge 1b \n"
"2: \n"
"adds %w[width], %w[width], %w[vl] \n"
"b.eq 99f \n"
"whilelt p0.b, wzr, %w[width] \n"
"ld2b {z0.b, z1.b}, p0/z, [%[src]] \n" // BR, GA
"uqadd z0.b, z0.b, z2.b \n"
"uqadd z1.b, z1.b, z2.b \n" ARGBTORGB565_SVE
"st1b {z1.b}, p0, [%[dst]] \n"
"99: \n"
: [src] "+r"(src_argb), // %[src]
[dst] "+r"(dst_rgb), // %[dst]
[width] "+r"(width), // %[width]
[vl] "=&r"(vl) // %[vl]
: [bsl_mask] "r"(bsl_mask), // %[bsl_mask]
[dither4] "r"(dither4) // %[dither4]
: "cc", "memory", "z0", "z1", "z3", "z4", "p0");
}
#endif // !defined(LIBYUV_DISABLE_SVE) && defined(__aarch64__) #endif // !defined(LIBYUV_DISABLE_SVE) && defined(__aarch64__)
#ifdef __cplusplus #ifdef __cplusplus