Skip to content

Unroll dynamic indexing in SYCL gather kernel

Szilárd Páll requested to merge sz_fix-sycl-gather-performance-on-amd into release-2022

Thread ID-based dynamic indexing of constant memory data in the SYCL gather kernel caused a large amount of register spills and poor performance of the gather kernel on AMD. Avoiding dynamic indexing eliminates spills and improves performance of this kernel >10x.

Refs #3927

Merge request reports