Improve bmm performance on CPU by applying TensorAccessor #20266
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
Collaborator
Author
|
Happen to check on VTune yesterday and realized direct Tensor indexing has quite a heavy performance overhead on such scenario, 2/3 or even 3/4 of the time on bmm. I should have combined this with #19338 beforehand... |
Contributor
|
3x speedup sounds awesome :-) cc @BIT-silence as you are recently looking at the perf of a model that has bmm. |
Contributor
facebook-github-bot
left a comment
facebook-github-bot
left a comment
There was a problem hiding this comment.
@bddppq has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.
xiaomengy
approved these changes
May 8, 2019
Contributor
xiaomengy
left a comment
xiaomengy
left a comment
There was a problem hiding this comment.
Thanks for this improvement!
xiaomengy
reviewed
May 8, 2019
|
|
||
| const int lda = is_transposed(mat1[0]) ? mat1[0].stride(1) : mat1[0].stride(0); | ||
| const int ldb = is_transposed(mat2[0]) ? mat2[0].stride(1) : mat2[0].stride(0); | ||
| const int lda = is_transposed(mat1_acc[0]) ? mat1_acc[0].stride(1) : mat1_acc[0].stride(0); |
Contributor
|
travis lint failure in python file untouched by this |
Contributor
zdevito
pushed a commit
to zdevito/ATen
that referenced
this pull request
May 16, 2019
Summary: Currently `bmm()` has very heavy performance overhead on CPU due to construction/deconstruction of `TensorImpl`. Applying `TensorAccessor` when indexing tensor data can greatly improve the performance. I tested this on `fairseq` Transformer model. Results on Xeon 6148 (20*2 cores 2.5GHz) indicate this PR improves Transformer training performance by approximately **10%** (seconds per iteration reduced from **3.60** to **3.21**). Considering the fact that `bmm()` takes only **14%** of the total time, 10% overall improvement indicates `bmm()` itself improves by roughly **3x**. Before: ``` | epoch 001: 0%| | 43/25337 [02:34<25:17:11, 3.60s/it, loss=16.179, nll_loss=16.137, ppl=72045.59, wps=1320, ups=0, wpb=4758.767, bsz=136.558, num_updates=43, lr=6.45e-06, gnorm=6.88 ``` After: ``` | epoch 001: 0%| | 23/25337 [01:13<22:32:48, 3.21s/it, loss=17.072, nll_loss=17.068, ppl=137419.42, wps=1478, ups=0, wpb=4746.870, bsz=128.348, num_updates=23, lr=3.45e-06, gnorm=10. ``` Pull Request resolved: pytorch/pytorch#20266 Differential Revision: D15262201 Pulled By: cpuhrsch fbshipit-source-id: c2e4e406c06714b04cc7534f3da71e986eddca35
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
9 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.
Currently
bmm()has very heavy performance overhead on CPU due to construction/deconstruction ofTensorImpl. ApplyingTensorAccessorwhen indexing tensor data can greatly improve the performance.I tested this on
fairseqTransformer model. Results on Xeon 6148 (20*2 cores @2.5GHz) indicate this PR improves Transformer training performance by approximately 10% (seconds per iteration reduced from 3.60 to 3.21). Considering the fact thatbmm()takes only 14% of the total time, 10% overall improvement indicatesbmm()itself improves by roughly 3x.Before:
After: