Herbst has improved the workgroup size selection for Rusticl for coming up with the best block size. He noted in the merge request that this change should " [increase] hardware utilization by a significant amount " by better filling workgroups and subgroups. In turn this should improve performance for applications requesting arbitrary global sizes. No real-world workload performance numbers were shared as part of this merge request, but on a fill rate basis there is indeed a huge improvement at least for better GPU utilization:
Those making use of the Rusticl OpenCL driver and wanting to learn all the details on this improvement can see this Mesa merge request . This is now in Mesa 26.2 for its stable release in August.