Open pgrete opened 9 months ago
@pgrete We are using atomics and I believe our allocations are coarse grained - which would be compatible with -munsafe-fp-atomics
.
I looked at the allocs we use with RAJA + I asked the VTK-m team to check. Since VTK-m is are using Kokkos and default device allocations are coarse grained, I would expect them to be fine.
We haven't done any explicit performance testing with this flag (unless VTK-m or Kokkos is turning it on for us)
Great, thanks, I'll give this a try. Feel free to close the issue once confirmed with the other libs.
How many device atomics are used in Ascent and it's software stack and is the internal programming/memory model safe with regard to this, see https://docs.olcf.ornl.gov/systems/frontier_user_guide.html#floating-point-fp-atomic-operations-and-coarse-fine-grained-memory-allocations ?
We've been using
Kokkos::ScatterView
s on AMD and without hardware atomics performance is >100x worse than with so I'm wondering if it's safe to globally turn on when compiling Ascent (and not just for our downstream code that links against Ascent).