Skip to content

Conversation

@sebvince
Copy link
Contributor

@sebvince sebvince commented Oct 24, 2025

This PR improves a bit the scheduling of the pingpong matmul Epilogue. Expected performance gain is 1-2% depending on the shape.

Before change:
image
The last iteration of second wave is stilling doing nothing and we get LDS issue stall on the next iteration.

After change:
image
MFMA unit are always busy until we write to global memory.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant