OpenCL M1 CPU bugs
Dear Apple OpenCL maintenance team,
I would like to share the following bugs I encountered with M1 CPU's OpenCL runtime.
- erf function produce wrong results while I have to use 1 - erfc in the kernel to produce the correct results.
- max_workitems_size is 1024x1x1 which make many existing kernels very inflexible to use since they normally use get_local_id(1) or even (2) to get flexible work-item indexing. Could you guys please modified it to 1024x1024x1024? The M1 GPU implementation has no such problem where the max_workgroup_size is 256 while max_workitems_size is 256x256x256. Besides, the Intel CPU runtime for Core i7-1165G7 ( driver = 2022.13.3.0.16_160000) has the max_workgroup_size 8192 while max_workitems_size = 8192x8192x8192.
Thank you very much!
Best wishes,
Jinchuan Tang
MacBook Air 13″, macOS 13.0