[Caffe2] Improve AddMomentsVec and UpdateMomentsVec #167664
Closed
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters. Learn more about bidirectional Unicode characters
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/167664
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit c23615b with merge base 0d7ba97 ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
pytorch-bot bot added the module: cpu 
Nicoshev force-pushed the export-D86805648 branch from a595fb5 to d9aa729 Compare Nicoshev added a commit to Nicoshev/pytorch that referenced this pull request Nov 13, 2025
Summary: RowwiseMomentsImpl accounts for about 0.4% cpu time of AdRanker: https://fburl.com/strobelight/ywf79nw3. It primarily calls AddMomentsVec and UpdateMomentsVec. These two routines are written using Pytorch's VecLib, meaning the utilized operators translate into intrinsics. Unfortunately, the compiler makes less transformations and optimizations when intrinsics are used. Therefore, if we carefully decouple and re-order operations, the emitted instruction sequence improves. Here we can see the dissassembly for the old and new AddMomentsVec: https://godbolt.org/z/83fxYvKfv We can see a much better instruction sequence is achieved in the new implementation. Test Plan: AdRanker ServiceLab Reviewed By: mcfi Differential Revision: D86805648
Nicoshev added topic: improvements Nicoshev force-pushed the export-D86805648 branch from d9aa729 to 6f164b8 Compare Nicoshev added a commit to Nicoshev/pytorch that referenced this pull request Nov 14, 2025
Summary: RowwiseMomentsImpl accounts for about 0.4% cpu time of AdRanker: https://fburl.com/strobelight/ywf79nw3. It primarily calls AddMomentsVec and UpdateMomentsVec. These two routines are written using Pytorch's VecLib, meaning the utilized operators translate into intrinsics. Unfortunately, the compiler makes less transformations and optimizations when intrinsics are used. Therefore, if we carefully decouple and re-order operations, the emitted instruction sequence improves. Here we can see the dissassembly for the old and new AddMomentsVec: https://godbolt.org/z/83fxYvKfv We can see a much better instruction sequence is achieved in the new implementation. Test Plan: AdRanker ServiceLab Reviewed By: mcfi Differential Revision: D86805648
Nicoshev force-pushed the export-D86805648 branch from 6f164b8 to 14dd41f Compare Nicoshev added a commit to Nicoshev/pytorch that referenced this pull request Nov 14, 2025
Summary: RowwiseMomentsImpl accounts for about 0.4% cpu time of AdRanker: https://fburl.com/strobelight/ywf79nw3. It primarily calls AddMomentsVec and UpdateMomentsVec. These two routines are written using Pytorch's VecLib, meaning the utilized operators translate into intrinsics. Unfortunately, the compiler makes less transformations and optimizations when intrinsics are used. Therefore, if we carefully decouple and re-order operations, the emitted instruction sequence improves. Here we can see the dissassembly for the old and new AddMomentsVec: https://godbolt.org/z/83fxYvKfv We can see a much better instruction sequence is achieved in the new implementation. Test Plan: We ran three AdRanker ServiceLabs, one x86-based and two on aarch64. We are also launching an x86 prod canary. The three ServiceLabs showed reduced cpu time on RowwiseMoments, as well as no numerical errors. https://www.internalfb.com/servicelab/experiment/6402421173/complete https://www.internalfb.com/servicelab/experiment/6102396760/complete https://www.internalfb.com/servicelab/experiment/6202396546/complete Reviewed By: mcfi Differential Revision: D86805648
Nicoshev force-pushed the export-D86805648 branch from 14dd41f to b034270 Compare pytorch-bot bot pushed a commit that referenced this pull request Nov 17, 2025
Summary: RowwiseMomentsImpl accounts for about 0.4% cpu time of AdRanker: https://fburl.com/strobelight/ywf79nw3. It primarily calls AddMomentsVec and UpdateMomentsVec. These two routines are written using Pytorch's VecLib, meaning the utilized operators translate into intrinsics. Unfortunately, the compiler makes less transformations and optimizations when intrinsics are used. Therefore, if we carefully decouple and re-order operations, the emitted instruction sequence improves. Here we can see the dissassembly for the old and new AddMomentsVec: https://godbolt.org/z/83fxYvKfv We can see a much better instruction sequence is achieved in the new implementation. Test Plan: We ran three AdRanker ServiceLabs, one x86-based and two on aarch64. We are also launching an x86 prod canary. The three ServiceLabs showed reduced cpu time on RowwiseMoments, as well as no numerical errors. https://www.internalfb.com/servicelab/experiment/6402421173/complete https://www.internalfb.com/servicelab/experiment/6102396760/complete https://www.internalfb.com/servicelab/experiment/6202396546/complete Reviewed By: mcfi Differential Revision: D86805648
Nicoshev added a commit to Nicoshev/pytorch that referenced this pull request Nov 17, 2025
Summary: RowwiseMomentsImpl accounts for about 0.4% cpu time of AdRanker: https://fburl.com/strobelight/ywf79nw3. It primarily calls AddMomentsVec and UpdateMomentsVec. These two routines are written using Pytorch's VecLib, meaning the utilized operators translate into intrinsics. Unfortunately, the compiler makes less transformations and optimizations when intrinsics are used. Therefore, if we carefully decouple and re-order operations, the emitted instruction sequence improves. Here we can see the dissassembly for the old and new AddMomentsVec: https://godbolt.org/z/83fxYvKfv We can see a much better instruction sequence is achieved in the new implementation. Test Plan: We ran three AdRanker ServiceLabs, one x86-based and two on aarch64. We are also launching an x86 prod canary. The three ServiceLabs showed reduced cpu time on RowwiseMoments, as well as no numerical errors. https://www.internalfb.com/servicelab/experiment/6402421173/complete https://www.internalfb.com/servicelab/experiment/6102396760/complete https://www.internalfb.com/servicelab/experiment/6202396546/complete Reviewed By: mcfi Differential Revision: D86805648
Nicoshev force-pushed the export-D86805648 branch 2 times, most recently from 997a550 to 76aac10 Compare pytorch-bot bot pushed a commit that referenced this pull request Nov 18, 2025
Summary: RowwiseMomentsImpl accounts for about 0.4% cpu time of AdRanker: https://fburl.com/strobelight/ywf79nw3. It primarily calls AddMomentsVec and UpdateMomentsVec. These two routines are written using Pytorch's VecLib, meaning the utilized operators translate into intrinsics. Unfortunately, the compiler makes less transformations and optimizations when intrinsics are used. Therefore, if we carefully decouple and re-order operations, the emitted instruction sequence improves. Here we can see the dissassembly for the old and new AddMomentsVec: https://godbolt.org/z/83fxYvKfv We can see a much better instruction sequence is achieved in the new implementation. Test Plan: We ran three AdRanker ServiceLabs, one x86-based and two on aarch64. We are also launching an x86 prod canary. The three ServiceLabs showed reduced cpu time on RowwiseMoments, as well as no numerical errors. https://www.internalfb.com/servicelab/experiment/6402421173/complete https://www.internalfb.com/servicelab/experiment/6102396760/complete https://www.internalfb.com/servicelab/experiment/6202396546/complete Reviewed By: mcfi Differential Revision: D86805648
Nicoshev added a commit to Nicoshev/pytorch that referenced this pull request Nov 18, 2025
Summary: RowwiseMomentsImpl accounts for about 0.4% cpu time of AdRanker: https://fburl.com/strobelight/ywf79nw3. It primarily calls AddMomentsVec and UpdateMomentsVec. These two routines are written using Pytorch's VecLib, meaning the utilized operators translate into intrinsics. Unfortunately, the compiler makes less transformations and optimizations when intrinsics are used. Therefore, if we carefully decouple and re-order operations, the emitted instruction sequence improves. Here we can see the dissassembly for the old and new AddMomentsVec: https://godbolt.org/z/83fxYvKfv We can see a much better instruction sequence is achieved in the new implementation. Test Plan: We ran three AdRanker ServiceLabs, one x86-based and two on aarch64. We are also launching an x86 prod canary. The three ServiceLabs showed reduced cpu time on RowwiseMoments, as well as no numerical errors. https://www.internalfb.com/servicelab/experiment/6402421173/complete https://www.internalfb.com/servicelab/experiment/6102396760/complete https://www.internalfb.com/servicelab/experiment/6202396546/complete Reviewed By: mcfi Differential Revision: D86805648
Nicoshev force-pushed the export-D86805648 branch from 76aac10 to b900fd5 Compare 
mcfi approved these changes Nov 18, 2025
Nicoshev added a commit to Nicoshev/pytorch that referenced this pull request Nov 18, 2025
Summary: RowwiseMomentsImpl accounts for about 0.4% cpu time of AdRanker: https://fburl.com/strobelight/ywf79nw3. It primarily calls AddMomentsVec and UpdateMomentsVec. These two routines are written using Pytorch's VecLib, meaning the utilized operators translate into intrinsics. Unfortunately, the compiler makes less transformations and optimizations when intrinsics are used. Therefore, if we carefully decouple and re-order operations, the emitted instruction sequence improves. Here we can see the dissassembly for the old and new AddMomentsVec: https://godbolt.org/z/83fxYvKfv We can see a much better instruction sequence is achieved in the new implementation. Test Plan: We ran three AdRanker ServiceLabs, one x86-based and two on aarch64. We are also launching an x86 prod canary. The three ServiceLabs showed reduced cpu time on RowwiseMoments, as well as no numerical errors. https://www.internalfb.com/servicelab/experiment/6402421173/complete https://www.internalfb.com/servicelab/experiment/6102396760/complete https://www.internalfb.com/servicelab/experiment/6202396546/complete Reviewed By: mcfi Differential Revision: D86805648
Nicoshev force-pushed the export-D86805648 branch 2 times, most recently from 1da2777 to 94ff965 Compare pytorch-bot bot pushed a commit that referenced this pull request Nov 19, 2025
Summary: RowwiseMomentsImpl accounts for about 0.4% cpu time of AdRanker: https://fburl.com/strobelight/ywf79nw3. It primarily calls AddMomentsVec and UpdateMomentsVec. These two routines are written using Pytorch's VecLib, meaning the utilized operators translate into intrinsics. Unfortunately, the compiler makes less transformations and optimizations when intrinsics are used. Therefore, if we carefully decouple and re-order operations, the emitted instruction sequence improves. Here we can see the dissassembly for the old and new AddMomentsVec: https://godbolt.org/z/83fxYvKfv We can see a much better instruction sequence is achieved in the new implementation. Test Plan: We ran three AdRanker ServiceLabs, one x86-based and two on aarch64. We are also launching an x86 prod canary. The three ServiceLabs showed reduced cpu time on RowwiseMoments, as well as no numerical errors. https://www.internalfb.com/servicelab/experiment/6402421173/complete https://www.internalfb.com/servicelab/experiment/6102396760/complete https://www.internalfb.com/servicelab/experiment/6202396546/complete Prod canary: https://www.internalfb.com/intern/ads/canary/473290286595708266 Reviewed By: mcfi Differential Revision: D86805648
Nicoshev added a commit to Nicoshev/pytorch that referenced this pull request Nov 19, 2025
Summary: RowwiseMomentsImpl accounts for about 0.4% cpu time of AdRanker: https://fburl.com/strobelight/ywf79nw3. It primarily calls AddMomentsVec and UpdateMomentsVec. These two routines are written using Pytorch's VecLib, meaning the utilized operators translate into intrinsics. Unfortunately, the compiler makes less transformations and optimizations when intrinsics are used. Therefore, if we carefully decouple and re-order operations, the emitted instruction sequence improves. Here we can see the dissassembly for the old and new AddMomentsVec: https://godbolt.org/z/83fxYvKfv We can see a much better instruction sequence is achieved in the new implementation. Test Plan: We ran three AdRanker ServiceLabs, one x86-based and two on aarch64. We are also launching an x86 prod canary. The three ServiceLabs showed reduced cpu time on RowwiseMoments, as well as no numerical errors. https://www.internalfb.com/servicelab/experiment/6402421173/complete https://www.internalfb.com/servicelab/experiment/6102396760/complete https://www.internalfb.com/servicelab/experiment/6202396546/complete Prod canary: https://www.internalfb.com/intern/ads/canary/473290286595708266 Reviewed By: mcfi Differential Revision: D86805648
Nicoshev force-pushed the export-D86805648 branch from 94ff965 to d3b6d28 Compare 
Summary: RowwiseMomentsImpl accounts for about 0.4% cpu time of AdRanker: https://fburl.com/strobelight/ywf79nw3. It primarily calls AddMomentsVec and UpdateMomentsVec. These two routines are written using Pytorch's VecLib, meaning the utilized operators translate into intrinsics. Unfortunately, the compiler makes less transformations and optimizations when intrinsics are used. Therefore, if we carefully decouple and re-order operations, the emitted instruction sequence improves. Here we can see the dissassembly for the old and new AddMomentsVec: https://godbolt.org/z/83fxYvKfv We can see a much better instruction sequence is achieved in the new implementation. Test Plan: We ran three AdRanker ServiceLabs, one x86-based and two on aarch64. We are also launching an x86 prod canary. The three ServiceLabs showed reduced cpu time on RowwiseMoments, as well as no numerical errors. https://www.internalfb.com/servicelab/experiment/6402421173/complete https://www.internalfb.com/servicelab/experiment/6102396760/complete https://www.internalfb.com/servicelab/experiment/6202396546/complete Prod canary: https://www.internalfb.com/intern/ads/canary/473290286595708266 Reviewed By: mcfi Differential Revision: D86805648
Nicoshev force-pushed the export-D86805648 branch from d3b6d28 to c23615b Compare 
Contributor
| @pytorchbot merge (Initiating merge automatically since Phabricator Diff has merged) |
Collaborator
Merge startedYour 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 |
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
4 participants
Add this suggestion to a batch that can be applied as a single commit. This suggestion is invalid because no changes were made to the code. Suggestions cannot be applied while the pull request is closed. Suggestions cannot be applied while viewing a subset of changes. Only one suggestion per line can be applied in a batch. Add this suggestion to a batch that can be applied as a single commit. Applying suggestions on deleted lines is not supported. You must change the existing code in this line in order to create a valid suggestion. Outdated suggestions cannot be applied. This suggestion has been applied or marked resolved. Suggestions cannot be applied from pending reviews. Suggestions cannot be applied on multi-line comments. Suggestions cannot be applied while the pull request is queued to merge. Suggestion cannot be applied right now. Please check back later.
Summary:
RowwiseMomentsImpl accounts for about 0.4% cpu time of AdRanker: https://fburl.com/strobelight/ywf79nw3. It primarily calls AddMomentsVec and UpdateMomentsVec.
These two routines are written using Pytorch's VecLib, meaning the utilized operators translate into intrinsics.
Unfortunately, the compiler makes less transformations and optimizations when intrinsics are used.
Therefore, if we carefully decouple and re-order operations, the emitted instruction sequence improves.
Here we can see the dissassembly for the old and new AddMomentsVec: https://godbolt.org/z/83fxYvKfv
We can see a much better instruction sequence is achieved in the new implementation.
Test Plan: AdRanker ServiceLab
Differential Revision: D86805648
cc @jgong5 @mingfeima @XiaobingSuper @sanchitintel @ashokei @jingxu10 @jerryzh168 @aditew01