second-guess AVX-512 opts

We've had some reason to double back on these AVX-512
opts and second guess whether they have any impact.

There's not much reason to keep building them if they're
the same speed as the AVX2 code with the extra risk of
AVX-512 downclocking.  In this case it's pretty easy
to maintain the AVX-512-aware code in SkBlitRow_opts.h,
so it'd be easy to add back if we find benefit later.

I plan to watch

    https://perf.skia.org/e/?queries=cpu_or_gpu_value%3DAVX512%26source_type%3Dskp%26sub_result%3Dmin_ms%26config%3D8888

and the geometric mean summary

    https://perf.skia.org/e/?formulas=geo(filter(%22cpu_or_gpu_value%3DAVX512%26source_type%3Dskp%26sub_result%3Dmin_ms%26config%3D8888%22))

Looking at this AVX-512 vs. AVX2 chart, we wouldn't expect to see any regression,
and likely an improvement:

    https://perf.skia.org/e/?formulas=geo(filter(%22config%3D8888%26cpu_or_gpu_value%3DAVX2%26extra_config%3D!Fast%26extra_config%3D!SK_FORCE_RASTER_PIPELINE_BLITTER%26source_type%3Dskp%26sub_result%3Dmin_ms%22))&formulas=geo(filter(%22config%3D8888%26cpu_or_gpu_value%3DAVX512%26source_type%3Dskp%26sub_result%3Dmin_ms%22))

Change-Id: Ib5ef8ec8cb4af1ecab254c31751761ee2bd153f3
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/313809
Commit-Queue: Mike Klein <mtklein@google.com>
Reviewed-by: Herb Derby <herb@google.com>
Reviewed-by: Mike Reed <reed@google.com>
This commit is contained in:
Mike Klein 2020-08-28 09:54:15 -05:00 committed by Skia Commit-Bot
parent 442c57893c
commit fc1eb95f39

View File

@ -8,13 +8,11 @@
#include "src/core/SkOpts.h"
#define SK_OPTS_NS skx
#include "src/opts/SkBlitRow_opts.h"
#include "src/opts/SkSwizzler_opts.h"
#include "src/opts/SkVM_opts.h"
namespace SkOpts {
void Init_skx() {
blit_row_s32a_opaque = SK_OPTS_NS::blit_row_s32a_opaque;
interpret_skvm = SK_OPTS_NS::interpret_skvm;
RGBA_to_BGRA = SK_OPTS_NS::RGBA_to_BGRA;
RGBA_to_rgbA = SK_OPTS_NS::RGBA_to_rgbA;