Repository navigation
Add a benchmark for cuda::memcpy_async #6511
Description
With the merge of #6431, we have device side benchmarking available now. We should add a benchmark for cuda::memcpy_async.
cuda::memcpy_async #6511With the merge of #6431, we have device side benchmarking available now. We should add a benchmark for cuda::memcpy_async.
Activity
bernhardmgruber commentedon Nov 11, 2025
Ideally, we should also have a test comparing the generated SASS.
fbusato commentedon Nov 11, 2025
my feeling is that a direct comparison could be hard due to register renaming and code block moving.
It could work for small code. @wmaxey did something similar for atomic. Otherwise, instruction counts could be a tradeoff.