[ATen][CUDA] Optimize 128 bit vectorization #152967
Conversation
Fixes #147376. As per request: #145746 (review) This PR omits sm80 or older of using vec8 kernels due to long compilation and large binary size. Pull Request resolved: #148320 Approved by: https://github.com/eqy, https://github.com/malfet, https://github.com/atalman (cherry picked from commit 72337bd)
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/152967
Note: Links to docs will display an error until the docs builds have been completed. ❌ 2 New Failures, 4 PendingAs of commit 26f4937 with merge base 924a247 ( NEW FAILURES - The following jobs have failed:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
|
Hm, I don't see large decrease in size here after merging the cherry-pick: 2.7.0 Release Nightly build seems better: Looks like improvement on nightly happened on May 01. Must be one of these changes: Nightly CUDA arches : 5.0;6.0;7.0;7.5;8.0;8.6,9.0 When original PR landed May 02 and was include in May 03 nightly I see only a slight improvement: |
|
Should we revert this PR then, if it doesn't improve build time/binary size? It adds considerable complexity, and we shouldn't add complexity for nothing. |
Fixes #147376.
As per request: #145746 (review)
This PR omits sm80 or older of using vec8 kernels due to long compilation and large binary size.
cc @ptrblck @msaroufim @eqy @jerryzh168 @manuelcandales @SherlockNoMad @angelayi