Skip to content

Draft: Extend Support for CUDA and HIP kernel invocations

Frederik Hennig requested to merge fhennig/cuda-invoke into master

This MR extends the support for CUDA and HIP kernel invocations and API reflection in pystencils-sfg, reflecting the concepts introduced in pystencils!449 (merged).

  • Move GPU-related stuff into new composer mix-in SfgGpuComposer
  • Replace cuda_invoke() by gpu_invoke()
  • Implement three separate signatures of gpu_invoke(), depending on the kernel indexing scheme selected by the user
  • Replace extensions.gpu by lang.gpu
  • Add explanations on GPU kernels to user manual
  • Fix some errors and clarify some interface descriptions in the user manual

This MR will already be using Target.HIP and therefore will have to wait for pystencils!458.

Edited by Frederik Hennig

Merge request reports