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

Remove gemminc kernels #8246

Open
wants to merge 1 commit into
base: master
Choose a base branch
from
Open

Remove gemminc kernels #8246

wants to merge 1 commit into from

Conversation

copybara-service[bot]
Copy link
Contributor

Remove gemminc kernels

These are unused currently, not hooked up to any configs.

We are very interested in tiling the K dimension of gemms, which these kernels would be helpful with. However, it's not clear that these are on the critical path to doing this:

  • This approach doesn't generalize beyond fp32 accumulators.
  • There are some issues with these kernels that are likely undiscovered because this code is unused (e.g. min/max is applied at every tile of K, could use multiple different params to work around of course, but I think this just hasn't been thought about yet, and there are probably many other things like that).
  • Packing that exists currently is not compatible with existing packing approaches, and would need quite a bit of work to get it up to speed.

If we do end up wanting these kernels, we can always revert this change. But as it stands, these kernels have been unused for years, and are unlikely to be used in the future.

These are unused currently, not hooked up to any configs.

We are very interested in tiling the K dimension of gemms, which these kernels would be helpful with. However, it's not clear that these are on the critical path to doing this:
- This approach doesn't generalize beyond fp32 accumulators.
- There are some issues with these kernels that are likely undiscovered because this code is unused (e.g. min/max is applied at every tile of K, could use multiple different params to work around of course, but I think this just hasn't been thought about yet, and there are probably many other things like that).
- Packing that exists currently is not compatible with existing packing approaches, and would need quite a bit of work to get it up to speed.

If we do end up wanting these kernels, we can always revert this change. But as it stands, these kernels have been unused for years, and are unlikely to be used in the future.

PiperOrigin-RevId: 746084904
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