[llvm-branch-commits] [llvm] [offload] Write small unordered host-to-device transfers directly on AMDGPU (PR #231569)

[Robert Imschweiler via llvm-branch-commits]

ro-i wrote:

hm yea, whether there is such a direct wide option probably needs to be a device attribute 🤔 

Btw, forgot to mention: the 40%+ came from testing an empty ompx_bare kernel with KLE: https://github.com/ro-i/omp-microtest/blob/main/src/omp_launch_latency.cpp
That kernel's complete runtime reduced by that percentage, so the latency reduction might even be a tad more, I guess

https://github.com/llvm/llvm-project/pull/231569