search

oneapi-src / oneMKL

oneAPI Math Kernel Library (oneMKL) Interfaces
Apache License 2.0
620 stars 158 forks source link

Some tests hang/slow down significantly when running with portBLAS #418

Open toxicscum opened 11 months ago

toxicscum commented 11 months ago

It was found that some tests hang/slow down significantly when running with portBLAS. Tested on Intel CPU/GPU, NVIDIA GPU with DPC++ compiler from oneAPI Base Toolkit 2023.2 and opensource implementation https://github.com/intel/llvm. intel_cpu_dpcpp_rhel9.txt intel_gpu_dpcpp_ubuntu22.txt nvidia_gpu_dpcpp_ubuntu22.txt

Could someone please check?

muhammad-tanvir-1211 commented 11 months ago

Hi @toxicscum, Thank you for creating the issue here. We are currently looking at the portblas tests hanging through oneMKL. So far, we have not been able to recreate the hangs on any of the GPU devices that we are testing on locally. There seems to be some tests taking longer to execute on CPU but almost all of them require GEMM in some capacity, so we are investigating those now. Could you please share the command you use to build portblas backend in oneMKL for all the above devices? Thanks.