Skip to content

xe: remove with_punning#5067

Open
rjoursler wants to merge 1 commit intomainfrom
rjoursle/remove_punning
Open

xe: remove with_punning#5067
rjoursler wants to merge 1 commit intomainfrom
rjoursle/remove_punning

Conversation

@rjoursler
Copy link
Copy Markdown
Contributor

In MFDNN-14925, we found multiple primitive failure due to inconsistent use of WITH_PUNNING across buffers. This PR completely removes this knob so that this failure mode can no longer occur. This fixes MFDNN-14925 along with a few other existing issues I found during testing.

@rjoursler rjoursler requested a review from a team as a code owner April 22, 2026 15:28
@github-actions github-actions Bot added the platform:gpu-intel Codeowner: @oneapi-src/onednn-gpu-intel label Apr 22, 2026
@rjoursler rjoursler force-pushed the rjoursle/remove_punning branch 3 times, most recently from e7580e5 to ac6c830 Compare April 22, 2026 15:35
@rjoursler
Copy link
Copy Markdown
Contributor Author

make test
set test_scope=NIGHTLY
disable test_device_cpu
disable build_cpu_runtime_omp
disable build_cpu_runtime_sycl
disable build_cpu_runtime_tbb

@rjoursler rjoursler force-pushed the rjoursle/remove_punning branch 5 times, most recently from 547e732 to 2294e07 Compare April 27, 2026 22:24
@rjoursler
Copy link
Copy Markdown
Contributor Author

make test
set test_scope=NIGHTLY
disable test_device_cpu
disable build_cpu_runtime_omp
disable build_cpu_runtime_sycl
disable build_cpu_runtime_tbb

This prevents mixing punning/non-punning APIs causing incorrect kernel
generation. Additionally fixes the following existing errors:

  benchdnn --mode-modifier=P --global-impl=ocl --engine=gpu     --bnorm --dt=bf16 --tag=axb ic32iw32n
  benchdnn --mode-modifier=P --global-impl=ocl --engine=gpu     --softmax --ddt=bf16 16x16
  benchdnn --mode-modifier=P --global-impl=ocl --engine=gpu     --prelu --dir=BWD_DW --sdt=bf16:bf16 --stag=axb:any 4x16x3x4:1x1x1x1
  benchdnn --mode-modifier=P --global-impl=ocl --engine=gpu     --binary --sdt=bf16:bf16 --ddt=bf16 15x12x3x5:1x1x1x1
  benchdnn --mode-modifier=P --global-impl=ocl --engine=gpu     --binary --sdt=bf16:bf16 --ddt=bf16 --alg=min --inplace=true     3x5x6x9:3x5x6x9
@rjoursler rjoursler force-pushed the rjoursle/remove_punning branch from 2294e07 to 7708ca6 Compare April 27, 2026 23:21
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