Skip to content

intel/fs: Improve copy prop performance

Faith Ekstrand requested to merge gfxstrand/mesa:review/intel-fs-copy-prop-faster into master

This MR significantly improves the performance of the back-end compiler's copy propagation pass. My benchmark for this was the vs-isnan-dvec test in piglit which I ran on an unoptimized debug build with NIR_VALIDATE=0. At the start of this MR, it takes about 1:18.1 to run and by the end, we're down to 38.7 seconds. Given that that includes all the GLSL and NIR compilation as well as other back-end passes, 50% is a pretty good delta. 😄

Shader-db results are completely unaffected by this MR. I ran shader-db with and without this MR five times each and performance was, on average, 0.2% better though that may just be in the noise.

Edited by Faith Ekstrand

Merge request reports