Relax scratch space limits for HIP reductions #6029
Merged
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This change is motivated by profiling of some LJ reduction kernels in LAMMPS. Essentially we noticed that in some cases, we ended up severely occupancy limited by the "gridDim <= blockDim" restriction. Omniperf showed very low L1/L2 utilization, but the limiter was simply that we were not launching enough blocks.
This is a simple change to relax the scratch space limit slightly (from blockDim to 4096) to increase occupancy, but still keep the scratch limits well bounded. I was not particularly scientific about choosing the new limit, and I'm definitely open to better heuristics, but this increased RHODO (4^3) perf in LAMMPS by ~28% (the key kernel is a ParReduce)