Skip to content
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

[QST] in implicit gemm conv, why does not support split-k when group !=1 ? #2049

Open
preFiredman opened this issue Jan 21, 2025 · 3 comments

Comments

@preFiredman
Copy link

preFiredman commented Jan 21, 2025

Image
the file is include/cutlass/conv/device/implicit_gemm_convolution.h
in implicit gemm conv, why does not support split-k when group !=1 ?
is this for performance ?

@preFiredman preFiredman closed this as not planned Won't fix, can't repro, duplicate, stale Jan 21, 2025
@preFiredman preFiredman changed the title [QST] implicit gemm conv does not support split-k currently ? [QST] typo Jan 21, 2025
@preFiredman preFiredman changed the title [QST] typo [QST] why does not support split-k when group !=1 ? Jan 22, 2025
@preFiredman preFiredman changed the title [QST] why does not support split-k when group !=1 ? [QST] in implicit gemm conv, why does not support split-k when group !=1 ? Jan 22, 2025
@preFiredman preFiredman reopened this Jan 22, 2025
@hwu36
Copy link
Collaborator

hwu36 commented Feb 6, 2025

@Junkai-Wu

@Junkai-Wu
Copy link
Contributor

For group fprop/dgrad, usually C/K per group is small, e.g. < 64, and RS dimension of filter is not large, e.g. 3x3, so there is no need to do split k. In other words, group is already used to split the C/K dimension in fprop/dgrad, there is no need to use split_k to split it again.

@preFiredman
Copy link
Author

thank you.
does wgrad support splitk, which seems necessary in scenario like depthwise conv ?

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Projects
None yet
Development

No branches or pull requests

3 participants