[libcxx-commits] [libcxx] [RFC] Offloading C++ standard parallel algorithms to GPUs using OpenMP (PR #66465)

[Johannes Doerfert via libcxx-commits]

jdoerfert wrote:

> In your current patch, you are reusing the CPU backend machinery. Originally, we designed the backend so that you could write your own GPU backend that would chunk work in a way that makes the most sense for it (which may not be something hierarchical like we do for the CPU, where we often do a map-reduce kind of pattern). Do you have thoughts on that? Is there a reason why you went for reusing the CPU backend? Is it just because the simd pragmas currently live there and it seemed easier to do it that way?

The last part was certainly a driver for this PoC. We should, most likely, not split work hierarchically for the GPU backend, thus, a new one, probably based on the SIMD one, is going to be in the next PR.

https://github.com/llvm/llvm-project/pull/66465