commit | 501e8e1a2c9fe73275c49f104dacd1f5397e5404 | [log] [tgz] |
---|---|---|
author | Herb Derby <herb@google.com> | Tue Sep 28 14:51:04 2021 -0400 |
committer | Herb Derby <herb@google.com> | Wed Sep 29 18:25:16 2021 +0000 |
tree | 023997f9d8f91a35c049b562a1298c8094853296 | |
parent | 6abafd681ee56f6652d0d48b16b3480af7c5739a [diff] |
fix rcp_precise for compiling with AVX This will move highp AVX users from AVX to SSE41. This is about 7.5%[1] of users for software. The underlying problem is that AVX works well with floating point, it has little support for integer operations. So for lowp, AVX works as SSE41 and in highp it works as HSW and SKX. The rcp_precise has the distinction as being the only function that works as highp and lowp. So, when compiled with -mavx it doesn't compile because in highp it has 8 elements and in lowp it has 4 elements. So, move it to be with SSE41 for highp. [1] https://uma.googleplex.com/p/chrome/histograms?sid=ec095beddef4b2771a6a3f12a78f32c1 Bug: skia:12453 Change-Id: I2929312ef36cc0c0a68515ac74e5d953c0b7fd09 Reviewed-on: https://skia-review.googlesource.com/c/skia/+/453716 Reviewed-by: Brian Osman <brianosman@google.com>