search

hpcgarage / spatter

Benchmark for measuring the performance of sparse and irregular memory access.
Other
75 stars 15 forks source link

Option for atomics for Scatter with OpenMP and CUDA backends #178

Closed JDTruj2018 closed 8 months ago

JDTruj2018 commented 8 months ago

Add atomic operations for scatter kernels for the OpenMP and CUDA backends to prevent race conditions and properly order writes.

Should this be a command line option that can be toggled? Or just the default/only option?

jyoung3131 commented 8 months ago

I am trying to think of a scenario where we would not want this to be the default, but maybe @plavin can also weigh in.

My thought is that we would want this enabled by default unless we were explicitly trying to find issues with how a memory system orders writes at the HW level.