Skip to content

OpenCL: correctly query Intel sub-group sizes

Andrey Alekseenko requested to merge aa-fix-opencl-subgroup-query into release-2023

Due to incorrect usage of clDeviceGetInfo, we were:

  • ignoring the result if the device supported >2 sub-group sizes,
  • reading garbage if the device supported 1 or 2 sub-group sizes.

In practice, it was always the former since only Intel GPUs have this extension, and they all support at least three sub-group sizes.

Our fallback values were good enough for it to work correctly on current hardware, so it never caused any problems.

Introduced in 5efa6802.

Merge request reports