search

oneapi-src / SYCLomatic

Other
222 stars 91 forks source link

migration of stream-ordered functions #1872

Open jinz2014 opened 4 months ago

jinz2014 commented 4 months ago

Could you please bring the SYCL support to the compiler people for cudaDeviceGetDefaultMemPool, cudaMallocAsync, cudaFreeAsync ?

Thanks

Reference https://developer.nvidia.com/blog/using-cuda-stream-ordered-memory-allocator-part-1/ https://github.com/NVIDIA/cuda-samples/tree/master/Samples/2_Concepts_and_Techniques/streamOrderedAllocation

oscarbg commented 4 months ago

Interesting.. +1

tomflinda commented 3 months ago

@jinz2014

  1. The compiler team is aware of this language gap.
  2. As a WA in SYCLomatic side, we can consider removing the cudaDeviceGetDefaultMemPool call. For cudaMallocAsync and cudaFreeAsync, we can map them to the sync APIs on the SYCL side.