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

[AArch64] Add SVE2 implementation of ARGBToRGB565Row

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

Cortex-A510: -27.1%
Cortex-A720: -49.4%
  Cortex-X2: -67.9%

Bug: libyuv:973
Change-Id: I321dc080a6e89301cd959c2ee18bc6680f749312
Reviewed-on: https://chromium-review.googlesource.com/c/libyuv/libyuv/+/5505538
Reviewed-by: Frank Barchard <fbarchard@chromium.org>
Reviewed-by: Justin Green <greenjustin@google.com>
This commit is contained in:
George Steed 2024-04-15 11:42:26 +01:00 committed by Frank Barchard
parent 812b4955b2
commit bce3392830
3 changed files with 60 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
include/libyuv/row.h
View File

@ -573,6 +573,7 @@ extern "C" {
#if !defined(LIBYUV_DISABLE_SVE) && defined(__aarch64__)
#define HAS_ABGRTOUVJROW_SVE2
#define HAS_ABGRTOUVROW_SVE2
#define HAS_ARGBTORGB565ROW_SVE2
#define HAS_ARGBTOUVJROW_SVE2
#define HAS_ARGBTOUVROW_SVE2
#define HAS_BGRATOUVROW_SVE2
@ -3625,6 +3626,9 @@ void ARGBToRAWRow_NEON(const uint8_t* src_argb, uint8_t* dst_raw, int width);
void ARGBToRGB565Row_NEON(const uint8_t* src_argb,
uint8_t* dst_rgb565,
int width);
void ARGBToRGB565Row_SVE2(const uint8_t* src_argb,
uint8_t* dst_rgb565,
int width);
void ARGBToARGB1555Row_NEON(const uint8_t* src_argb,
uint8_t* dst_argb1555,
int width);

5
source/convert_from_argb.cc
View File

@ -2045,6 +2045,11 @@ int ARGBToRGB565(const uint8_t* src_argb,
}
}
#endif
#if defined(HAS_ARGBTORGB565ROW_SVE2)
if (TestCpuFlag(kCpuHasSVE2)) {
ARGBToRGB565Row = ARGBToRGB565Row_SVE2;
}
#endif
#if defined(HAS_ARGBTORGB565ROW_MSA)
if (TestCpuFlag(kCpuHasMSA)) {
ARGBToRGB565Row = ARGBToRGB565Row_Any_MSA;

51
source/row_sve.cc
View File

@ -499,6 +499,57 @@ void RGBAToUVRow_SVE2(const uint8_t* src_rgba,
kRGBAToUVCoefficients);
}
#define ARGBTORGB565_SVE \
/* Inputs: \
* z0: rrrrrxxxbbbbbxxx \
* z1: xxxxxxxxggggggxx \
* z3: 0000000000000011 (3, 0, 3, 0, ...) \
* z4: 0000011111100000 \
*/ \
"lsr z0.b, p0/m, z0.b, z3.b \n" \
"lsl z1.h, z1.h, #3 \n" \
"bsl z1.d, z1.d, z0.d, z4.d \n"
void ARGBToRGB565Row_SVE2(const uint8_t* src_argb,
uint8_t* dst_rgb,
int width) {
unsigned bsl_mask = 0x7e0;
uint64_t vl;
width *= 2;
asm("mov z3.h, #3 \n"
"dup z4.h, %w[bsl_mask] \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"
"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
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]
: "cc", "memory", "z0", "z1", "z3", "z4", "p0");
}
#endif // !defined(LIBYUV_DISABLE_SVE) && defined(__aarch64__)
#ifdef __cplusplus