Skip to content

Conversation

@LoserCheems
Copy link
Collaborator

Changes default Clear_OOB_MN parameter to true for safer out-of-bounds handling.

Updates tensor dimension comments to use more descriptive MMA_M and MMA_N naming convention for better code clarity.

Changes default Clear_OOB_MN parameter to true for safer out-of-bounds handling.

Updates tensor dimension comments to use more descriptive MMA_M and MMA_N naming convention for better code clarity.
Copy link
Contributor

Copilot AI left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Pull Request Overview

This PR makes the out-of-bounds clearing behavior safer by enabling it by default and improves the clarity of tensor-dimension comments.

  • Change default Clear_OOB_MN parameter from false to true for safer out-of-bounds handling
  • Update comments to use descriptive MMA_M and MMA_N naming for tensor dimensions
Comments suppressed due to low confidence (3)

csrc/src/utils.h:500

  • Update the function's documentation/comments to mention that Clear_OOB_MN now defaults to true, ensuring that code comments stay in sync with the new behavior.
template <bool Is_even_MN=true, bool Clear_OOB_MN=true,

csrc/src/utils.h:500

  • Add or update unit tests to cover the new out-of-bounds clearing behavior, verifying that elements outside valid ranges are correctly zeroed when Clear_OOB_MN=true.
template <bool Is_even_MN=true, bool Clear_OOB_MN=true,

csrc/src/utils.h:500

  • Changing Clear_OOB_MN default to true may introduce additional clearing overhead; consider benchmarking the performance impact or making this behavior explicit in performance-critical contexts.
template <bool Is_even_MN=true, bool Clear_OOB_MN=true,

@LoserCheems LoserCheems merged commit 1da3d98 into main Jun 30, 2025
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

bug Something isn't working

Projects

None yet

Development

Successfully merging this pull request may close these issues.

5 participants