commit f2e78e130434834e3815e4bdd2900d940ded9af3
author George Steed <george.steed@arm.com> Fri Mar 15 18:22:03 2024 +0000
committer Frank Barchard <fbarchard@chromium.org> Tue Apr 09 03:09:36 2024 +0000
tree 84a1ac2eb6fd0e00fc7b105740d6351e32e33786
parent 64061790630b8fab97bbc8ada7f558c665e13674

[AArch64] Use Neon dot-product instructions in ARGBToYMatrixRow

Using the dot-product instructions here allows us to avoid needing LD4
for loading individual colour channels, which gives a big benefit on
some micro-architectures where such instructions perform significantly
worse than LD1. In addition the dot-product instructions have higher
throughput compared to the Neon

Observed reduction in runtimes for selected kernels moving from *_NEON
to *_NEON_DotProd:

     Kernel | Cortex-A55 | Cortex-A510 | Cortex-A76 | Cortex-X2
ABGRToYJRow |      -6.5% |      -22.5% |     -43.5% |    -71.2%
 ABGRToYRow |      -6.5% |      -22.5% |     -43.5% |    -68.3%
ARGBToYJRow |      -6.5% |      -22.5% |     -43.5% |    -68.1%
 ARGBToYRow |      -6.5% |      -22.5% |     -43.5% |    -68.1%
 BGRAToYRow |      -6.5% |      -22.5% |     -42.3% |    -68.4%
RGBAToYJRow |      -6.5% |      -22.5% |     -42.2% |    -73.7%
 RGBAToYRow |      -6.5% |      -22.5% |     -42.3% |    -64.9%

Bug: libyuv:977
Change-Id: If244190a7bdacf7e6e6b16af7e6853ee13ff6585
Reviewed-on: https://chromium-review.googlesource.com/c/libyuv/libyuv/+/5424737
Reviewed-by: Frank Barchard <fbarchard@chromium.org>