opencl gemm_kernel error: CL_INVALID_WORK_GROUP_SIZE for the function gemm_offset_gpu() in blas_kernels.c

Question

opencl gemm_kernel error: CL_INVALID_WORK_GROUP_SIZE for the function gemm_offset_gpu() in blas_kernels.c

TaihuLight opened this issue 4 years ago · 4 comments

When I run darknet using the function gemm_offset_gpu() in blas_kernels.c, it occurs the following error:
FATAL ERROR: CL_INVALID_WORK_GROUP_SIZE opencl gemm_kernel error: CL_INVALID_WORK_GROUP_SIZE

How to fix it?

Answer 1 · 2021-05-23T10:20:04.000Z

Yes, I was trying to dynamically adjust tuning parameters (in 7 methods) but I decided to put it as a const. pls try to re-clone now and build. thanks! :).

Answer 2 · 2021-06-02T14:02:19.000Z

Hi, is it anything better now?

Answer 3 · 2021-06-20T09:57:45.000Z

no comments... seems the issue can be closed... feel free to reopen...

Answer 4 · 2022-02-05T06:57:29.000Z

Hi @sowson , I cloned the code two or three weeks ago but I got the same problem FATAL ERROR: CL_INVALID_WORK_GROUP_SIZE opencl gemm_kernel error: CL_INVALID_WORK_GROUP_SIZE. It seems that the problem is not solved?

The config is cifar.cfg. I checked the passed arguments of clEnqueueNDRangeKernel in opencl_kernel_local. The globalOffset is 0, 0, globalItems is 32, 100352, localItems is 32, 1. Do you have any idea why these sizes are invalid?