[llvm] [Offload] Add olLaunchKernelSuggestedGroupSize (PR #142130)

[Ross Brunton via llvm-commits]

RossBrunton wrote:

@jhuber6 @callumfare This in response to some internal discussions while implementing the UR adapter. Basically, in UR you can specify a null workgroup size to have the implementation determine the best one. In theory this could be done via various queries to kernel and device info, but to avoid duplicating work and API calls, it might be worth adding this as a feature to Offload itself.

This could be either a new entry point or by setting the group size to zero in the "regular" launch method. I think a new entry point is probably more in line with how we've been designing liboffload.

This MR is just to get feedback on the proposed design, and to see if we want to go in this direction. And also maybe come up with a better name than `LaunchKernelSuggestedGroupSize`.

https://github.com/llvm/llvm-project/pull/142130