coffee/libyuv
1
0
Fork 0
mirror of https://chromium.googlesource.com/libyuv/libyuv synced 2026-01-01 03:12:16 +08:00
Code Issues Packages Projects Releases Wiki Activity

[AArch64] Use LD1/ST1 rather than LD4/ST4 in ARGBShadeRow_NEON

Browse Source 
The use of LD4 and ST4 to de-interleave ARGB color channels is
unnecessary here since we can just adjust the scale multiplicand to
match the interleaved layout. LD4 and ST4 are known to perform poorly on
some micro-architectures so using LD1 and ST1 here should be preferred.

Reduction in runtime for ARGBShadeRow_NEON:

  Cortex-A55: -19.9%
 Cortex-A510: -50.8%
  Cortex-A76: -36.0%
   Cortex-X2: -46.4%

Bug: libyuv:976
Change-Id: I10a0e6a0a62242826d39b1e963063770f084226a
Reviewed-on: https://chromium-review.googlesource.com/c/libyuv/libyuv/+/5494093
Commit-Queue: Frank Barchard <fbarchard@chromium.org>
Reviewed-by: Frank Barchard <fbarchard@chromium.org>
This commit is contained in:
George Steed 2024-03-14 06:58:22 +00:00 committed by libyuv LUCI CQ
parent ce32eb773f
commit 1eae2efbc7
1 changed files with 23 additions and 23 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

46
source/row_neon64.cc
View File

@ -3613,36 +3613,36 @@ void ARGBQuantizeRow_NEON(uint8_t* dst_argb,
}
// Shade 8 pixels at a time by specified value.
// NOTE vqrdmulh.s16 q10, q10, d0[0] must use a scaler register from 0 to 8.
// Rounding in vqrdmulh does +1 to high if high bit of low s16 is set.
// sqrdmulh is a rounding instruction, so +1 if high bit of low half of
// multiply result is set.
void ARGBShadeRow_NEON(const uint8_t* src_argb,
uint8_t* dst_argb,
int width,
uint32_t value) {
asm volatile(
"dup v0.4s, %w3 \n" // duplicate scale value.
"zip1 v0.8b, v0.8b, v0.8b \n" // v0.8b aarrggbb.
"ushr v0.8h, v0.8h, #1 \n" // scale / 2.
"dup v0.4s, %w3 \n" // duplicate scale value.
"zip1 v0.16b, v0.16b, v0.16b \n" // v0.16b aarrggbbaarrggbb.
"ushr v0.8h, v0.8h, #1 \n" // scale / 2.
// 8 pixel loop.
"1: \n"
"ld4 {v4.8b,v5.8b,v6.8b,v7.8b}, [%0], #32 \n" // load 8 ARGB
"subs %w2, %w2, #8 \n" // 8 processed per loop.
"uxtl v4.8h, v4.8b \n" // b (0 .. 255)
"prfm pldl1keep, [%0, 448] \n"
"uxtl v5.8h, v5.8b \n"
"uxtl v6.8h, v6.8b \n"
"uxtl v7.8h, v7.8b \n"
"sqrdmulh v4.8h, v4.8h, v0.h[0] \n" // b * scale * 2
"sqrdmulh v5.8h, v5.8h, v0.h[1] \n" // g
"sqrdmulh v6.8h, v6.8h, v0.h[2] \n" // r
"sqrdmulh v7.8h, v7.8h, v0.h[3] \n" // a
"uqxtn v4.8b, v4.8h \n"
"uqxtn v5.8b, v5.8h \n"
"uqxtn v6.8b, v6.8h \n"
"uqxtn v7.8b, v7.8h \n"
"st4 {v4.8b,v5.8b,v6.8b,v7.8b}, [%1], #32 \n" // store 8 ARGB
"b.gt 1b \n"
"1: \n"
"ld1 {v4.8b,v5.8b,v6.8b,v7.8b}, [%0], #32 \n" // load 8 ARGB
"subs %w2, %w2, #8 \n" // 8 processed per loop.
"uxtl v4.8h, v4.8b \n"
"prfm pldl1keep, [%0, 448] \n"
"uxtl v5.8h, v5.8b \n"
"uxtl v6.8h, v6.8b \n"
"uxtl v7.8h, v7.8b \n"
"sqrdmulh v4.8h, v4.8h, v0.8h \n" // argb * scale * 2
"sqrdmulh v5.8h, v5.8h, v0.8h \n"
"sqrdmulh v6.8h, v6.8h, v0.8h \n"
"sqrdmulh v7.8h, v7.8h, v0.8h \n"
"uqxtn v4.8b, v4.8h \n"
"uqxtn v5.8b, v5.8h \n"
"uqxtn v6.8b, v6.8h \n"
"uqxtn v7.8b, v7.8h \n"
"st1 {v4.8b,v5.8b,v6.8b,v7.8b}, [%1], #32 \n" // store 8 ARGB
"b.gt 1b \n"
: "+r"(src_argb), // %0
"+r"(dst_argb), // %1
"+r"(width) // %2