Skip to content

Draft: Use cpu_cuda library as cpu backend

Arthur Marius Hennequinrequested to merge
ahennequ_cpucuda into master

Use https://gitlab.cern.ch/ahennequ/cpu_cuda as cpu backend

Goals:

  • Reimplement cuda behavior in c++, all cuda code should compile and behave exactly as on a gpu
  • Reduce friction when implementing cuda kernels in Allen, especially when using warp intrinsics (remove the need for cpu specialization)
  • Decouple the cpu backend from the Allen repository, increase modularity to allow reuse in other projects

TODO:

  • Dispatch blocks on multiple threads to increase true parallelism
  • Add ARM64 support
  • Fix all corner cases

FYI @gligorov @anmorris @godiazlo @dovombru

Merge request reports