-
Notifications
You must be signed in to change notification settings - Fork 12
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
CUDA: for each PR, show delta register usage for GPU kernels #765
Labels
Comments
BenWibking
added
CI
enhancement
New feature or request
github_actions
Pull requests that update GitHub Actions code
labels
Oct 1, 2024
BenWibking
changed the title
CUDA: show delta register usage for GPU kernels
CUDA: for each PR, show delta register usage for GPU kernels
Oct 1, 2024
Perl one-liner from Weiqun:
|
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Labels
Describe the proposal
List the GPU kernels with changed register usage as a comment in each PR.
This can done by using the
--ptxas-options=v
compiler flag, then parsing the compiler output withsed
or another text parser, e.g.:We can parse the above into a *.csv file that consists of two columns: kernel name and register usage:
and then diff it against a reference file computed using the current development branch.
Describe alternatives you've considered
Profile GPU performance directly for each PR. This is tricky to do on a GPU kernel-by-kernel basis.
Additional context
GPU performance for our code is exquisitely sensitive to register usage. Conversely, register pressure is a good predictor of performance for our code. This metric should tell us whether there are major performance regressions on GPU.
See also: https://stackoverflow.com/questions/12388207/interpreting-the-verbose-output-of-ptxas-part-i
The text was updated successfully, but these errors were encountered: