| commit | fa64774820cb42594d3f5bc2059953510f038636 | [log] [tgz] |
|---|---|---|
| author | Mike Klein <mtklein@chromium.org> | Thu Feb 16 06:21:54 2017 -0500 |
| committer | Mike Klein <mtklein@chromium.org> | Thu Feb 16 12:54:04 2017 +0000 |
| tree | 46b4bdee3eaa607c5455b03da125563016d83ab0 | |
| parent | 8729e5bbf7c436fd7c7c13182adbbfb419f566b5 [diff] |
Flush to zero when loading f16 with sse2/sse4.1. The multiply by 0x77800000 is quite slow when the input is denormalized. We don't mind flushing those values (in the range of 1e-5) to zero. Implement portable load_f16() / store_f16() too. Change-Id: I125cff1c79ca71d9abe22ac7877136d86707cb56 Reviewed-on: https://skia-review.googlesource.com/8467 Reviewed-by: Mike Klein <mtklein@chromium.org> Commit-Queue: Mike Klein <mtklein@chromium.org>