Skip to content

Conversation

@rjoursler
Copy link
Contributor

@rjoursler rjoursler commented Oct 31, 2025

While investigating the regression identified in MFDNN-14350, I found two configurations resulted in the same score according to the performance model. Due to non-determinism in how these configurations are generated, different configurations can be chosen for that problem. This PR removes the non-determinism (although it happens to pick the bad case for MFDNN-14350, meaning more work will be required).

@rjoursler rjoursler requested a review from a team as a code owner October 31, 2025 21:42
@github-actions github-actions bot added the platform:gpu-intel Codeowner: @oneapi-src/onednn-gpu-intel label Oct 31, 2025
@rjoursler rjoursler changed the title xe: jit: remove non-determinism in blocking generator xe: jit: remove non-determinism between tool chains in blocking generator Oct 31, 2025
@rjoursler
Copy link
Contributor Author

make test
disable test_device_cpu

@rjoursler
Copy link
Contributor Author

make test perf-gpu

Prevents non-determinism in blocking choice for convolution when different
configurations achieve the same score by the model.
@rjoursler
Copy link
Contributor Author

make test perf-gpu
set primitive=gpu:conv

@rjoursler rjoursler force-pushed the rjoursle/fix_nondeterminism branch from f8e69e6 to 6473931 Compare November 5, 2025 00:36
@rjoursler
Copy link
Contributor Author

make test perf-gpu
set primitive=gpu:conv

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

Labels

platform:gpu-intel Codeowner: @oneapi-src/onednn-gpu-intel

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants