EnzymeAD
Support differentiating through CUDA host code
In addition, we should at the same consider supporting hipStreams to maintain similar-ish support for both.
Or are hip-streams easier to handle @vchuravy ?
Was just looking at the source of the particle framework our smoothed particle hydrodynamics codes here are built on top of, and differentiating through it would depend on our ability to add support for the differentiation of host-code..
For reference, here is the base framework openFPM.
Kokkos does also seem to depend on CUDA streams for the launch of its Kokkos::CUDA execution space instances.
AMD has two complementary APIs HSA and HIP. HIP tries to follow CUDA so it should be straightforward HSA does not.
On Tue, Jun 14, 2022, 17:54 Ludger Paehler @.***> wrote:
In addition, we should at the same consider supporting hipStreams https://rocmdocs.amd.com/en/latest/ROCm_API_References/HIP_API/Stream-Management.html to maintain similar-ish support for both.
Or are hip-streams easier to handle @vchuravy https://github.com/vchuravy ?
— Reply to this email directly, view it on GitHub https://github.com/EnzymeAD/Enzyme/issues/535#issuecomment-1155747892, or unsubscribe https://github.com/notifications/unsubscribe-auth/AABDO2RGNEF5IPOXUDTR5FTVPD5RZANCNFSM5PRTNUDQ . You are receiving this because you were mentioned.Message ID: @.***>
