🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/137795

• 📄 Preview Python docs built from this PR
• 📄 Preview C++ docs built from this PR
• ❓ Need help or want to give feedback on the CI? Visit the bot commands wiki or our office hours

Note: Links to docs will display an error until the docs builds have been completed.

✅ No Failures

As of commit 5628514 with merge base 9aaf3a0 ():
💚 Looks good so far! There are no failures yet. 💚

This comment was automatically generated by Dr. CI and updates every 15 minutes.

@pytorchbot label "topic: not user facing"

@pytorchbot label "topic: not user facing"

I disagree that this is a not user facing change, it should be properly marked as bugfix

Please seek CI approval before scheduling CIFlow labels

I've converted this to draft to prevent it from being merged. Whilst this fixed the problem empirically, I think a safer fix is to disable gcc's tree-vectorizor and tree-loop-vectorizer in older versions of gcc for sve supported platforms.

I've changed the fix to disable tree-vectorize optimizations accross the whole of vec_base.h for gcc versions from 10.0 up to 10.2 where the compiler error is known to occur.
This is a more broad and inclusive fix than I had proposed previously.

#137775 (comment)

I have propsed to target a specific version of gcc for this because when I made some quick performance comparisons there with GCC 11 with and without the optimizations for SVE there was a significant performance hit on inductor micro benchmarks.

I've changed the fix to disable tree-vectorize optimizations accross the whole of vec_base.h for gcc versions from 10.0 up to 10.2 where the compiler error is known to occur. This is a more broad and inclusive fix than I had proposed previously.

#137775 (comment)

I have propsed to target a specific version of gcc for this because when I made some quick performance comparisons there with GCC 11 with and without the optimizations for SVE there was a significant performance hit on inductor micro benchmarks.

I agree, optimizations for SVE is critical to get performance boost.

Summary of Key Developments w.r.t GCC for SVE:
GCC 8.x: Experimental support, minimal SVE features.
GCC 9.x: Introduced -march=armv8.2-a+sve support, better but still developing vectorization and SVE handling.
GCC 10.x: Solid stability and improvements in optimizations for SVE workloads.
GCC 11.x: Focused on enhancing performance, debug support, and SVE optimizations.
GCC 12.x and 13.x: Full maturity with best-in-class support for modern ARM cores, optimizations, and feature completeness for production use.

Summary of GCC versions in the stable Ubuntu releases:
Ubuntu 22.04 LTS: GCC 11.2 (default)
Ubuntu 22.10: GCC 12.2 (default)
Ubuntu 23.04: GCC 13.x (default)

Most of the Linux CI pipelines in PyTorch comes with Ubuntu 22.04 OS which has GCC 11.2 as default compiler and now we plan to move PyTorch CI to use GCC 12. I feel, starting from GCC 11, the SVE path of the compiler is more stable.

"I would suggest we should disable SVE build if the GCC compiler version is <11 (GCC < 11). This would avoid the bugs that we encounter mostly with gcc 10."

Since GCC is already at 13.x, I think we will still be supporting good range of GCC compiler versions for SVE.

CC @robert-hardwick @malfet

@pytorchbot rebase

@pytorchbot started a rebase job onto refs/remotes/origin/viable/strict. Check the current status here

Successfully rebased gcc_compiler_fix onto refs/remotes/origin/viable/strict, please pull locally before adding more changes (for example, via git checkout gcc_compiler_fix && git pull --rebase)

@pytorchbot merge -r

@pytorchbot started a rebase job onto refs/remotes/origin/viable/strict. Check the current status here

Tried to rebase and push PR #137795, but it was already up to date. Try rebasing against main by issuing:
@pytorchbot rebase -b main

The merge job was canceled or timed out. This most often happen if two merge requests were issued for the same PR, or if merge job was waiting for more than 6 hours for tests to finish. In later case, please do not hesitate to reissue the merge command
For more information see pytorch-bot wiki.

@pytorchbot merge -r

@pytorchbot started a rebase job onto refs/remotes/origin/viable/strict. Check the current status here

Successfully rebased gcc_compiler_fix onto refs/remotes/origin/viable/strict, please pull locally before adding more changes (for example, via git checkout gcc_compiler_fix && git pull --rebase)

Merge started

Your change will be merged once all checks pass (ETA 0-4 Hours).

Learn more about merging in the wiki.

Questions? Feedback? Please reach out to the PyTorch DevX Team