From ea5646030053a179071379a77d5f3833676f2a11 Mon Sep 17 00:00:00 2001 From: George Steed Date: Sun, 24 Mar 2024 07:01:15 +0000 Subject: [PATCH] [AArch64] Use LD1/ST1 rather than LD4/ST4 in ARGBMultiplyRow_NEON There is no need to de-interleave channels here since we are applying the same operation across all lanes. LD4 and ST4 are known to be significantly slower than LD1/ST1 on some micro-architectures so we should prefer to avoid them where possible. Reduction in runtimes observed for ARGBMultiplyRow_NEON: Cortex-A55: -22.3% Cortex-A510: -56.6% Cortex-A76: -45.5% Cortex-X2: -54.6% Change-Id: I9103111a109a4d87d358e06eb513746314aaf66a Reviewed-on: https://chromium-review.googlesource.com/c/libyuv/libyuv/+/5454832 Commit-Queue: Frank Barchard Reviewed-by: Frank Barchard --- source/row_neon64.cc | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/source/row_neon64.cc b/source/row_neon64.cc index d34827991..8dd9af7e4 100644 --- a/source/row_neon64.cc +++ b/source/row_neon64.cc @@ -3779,8 +3779,8 @@ void ARGBMultiplyRow_NEON(const uint8_t* src_argb, asm volatile( // 8 pixel loop. "1: \n" - "ld4 {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n" // load 8 ARGB - "ld4 {v4.8b,v5.8b,v6.8b,v7.8b}, [%1], #32 \n" // load 8 more + "ld1 {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n" // load 8 ARGB + "ld1 {v4.8b,v5.8b,v6.8b,v7.8b}, [%1], #32 \n" // load 8 more "subs %w3, %w3, #8 \n" // 8 processed per loop. "umull v0.8h, v0.8b, v4.8b \n" // multiply B "prfm pldl1keep, [%0, 448] \n" @@ -3792,7 +3792,7 @@ void ARGBMultiplyRow_NEON(const uint8_t* src_argb, "rshrn v1.8b, v1.8h, #8 \n" // 16 bit to 8 bit G "rshrn v2.8b, v2.8h, #8 \n" // 16 bit to 8 bit R "rshrn v3.8b, v3.8h, #8 \n" // 16 bit to 8 bit A - "st4 {v0.8b,v1.8b,v2.8b,v3.8b}, [%2], #32 \n" // store 8 ARGB + "st1 {v0.8b,v1.8b,v2.8b,v3.8b}, [%2], #32 \n" // store 8 ARGB "b.gt 1b \n" : "+r"(src_argb), // %0 "+r"(src_argb1), // %1