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

Merge OpenAI Triton commit 716a521 #3361

Draft
wants to merge 8 commits into
base: main
Choose a base branch
from
Draft

Conversation

whitneywhtsang
Copy link
Contributor

This PR change the Triton base from 1e0e51c to 716a521 (Feb 4).
Pass rate: 98.19%

Please do not squash and merge this PR.

Mogball and others added 8 commits February 3, 2025 22:31
Hoisting layout conversions into ifs relies on the assumption that the
if infrequently executes, but this assumption only makes sense in a
loop. A single top-level if in a kernel either executes or it doesn't,
and if the hoist is incorrect, it can lead to a slowdown.
Since we load data in the column major format with `ldmatrix.trans`,
pre-blackwell hardware seems difficult to support the transpose case.
Simplify pipelining by removing outer loop pipelining transformation.
Performance benefits of it are smaller than pipelining fused persistent
loops, while making the pipeliner harder to maintain and refactor.
The folder for TransOp was a bit too aggresive. Sometimes it would
change the representation of a layout for an equivalent one, and that's
not allowed in the current state of things. We move the optimisation we
had to a different canonicalizer.
If we expect a warning, we need to use a fresh cache dir; otherwise, no
warning will be thrown when the cache is hit.
Also take out interpreter related code from this file.
@whitneywhtsang whitneywhtsang self-assigned this Feb 5, 2025
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.

5 participants