Reuse OneDNN handler for SGD and SUM for SelectedRows input tensors. #35510
Conversation
This file contains 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
This makes it possible to call it multiple times without recreating the oneDNN primitives every time.
* For dense param and sparse grad.
|
Thanks for your contribution! |
* Stop use PP cache mechanims to store mem and primitive obj. * Handler obj store and reuse needed desc & prim
|
@wzzju Please tak a look at changes in this PR |
|
@wzzju Hi, please take a look at this PR, use oneDNN sum and SGD |
|
@tsocha Please review this PR |
lidanqing-intel
approved these changes
Sep 18, 2021
chenwhql
approved these changes
Sep 18, 2021
|
@chenwhql could you please approve PR-CI-APPROVAL ? |
tsocha
approved these changes
Sep 20, 2021
ghost
pushed a commit
to piotrekobi/Paddle
that referenced
this pull request
Sep 24, 2021
…addlePaddle#35510) * Create stateful OneDNNAXPYHandler object. This makes it possible to call it multiple times without recreating the oneDNN primitives every time. * Prepare SGDOpKernel to reuse its implementation from OneDNN kernel. * OneDNN SGD kernel. * Update call to use new OneDNNAXPYHandler object api. * Setup seed in proper place. * Enable OneDNN kernel only for single case. * For dense param and sparse grad. * Small refactor. * Enable oneDNN by op attr or by cmd line flag. * Use int64_t type for number of elements. * Support dense param and grad from OneDNN kernel. * Enable SGD OneDNN kernel when use MP BF16 optimizer. * Force non-copyable/movable OneDNNAXPYHandler. * Reuse OneDNNAXPYHandler for spare tensors in SUM op. * Fix SFINAE rules. * Remove recording event inside AXPY. * Get rid of internal primitive caching. * Stop use PP cache mechanims to store mem and primitive obj. * Handler obj store and reuse needed desc & prim * Do not derive from MKLDNNHandlerT
AnnaTrainingG
pushed a commit
to AnnaTrainingG/Paddle
that referenced
this pull request
Sep 29, 2021
…addlePaddle#35510) * Create stateful OneDNNAXPYHandler object. This makes it possible to call it multiple times without recreating the oneDNN primitives every time. * Prepare SGDOpKernel to reuse its implementation from OneDNN kernel. * OneDNN SGD kernel. * Update call to use new OneDNNAXPYHandler object api. * Setup seed in proper place. * Enable OneDNN kernel only for single case. * For dense param and sparse grad. * Small refactor. * Enable oneDNN by op attr or by cmd line flag. * Use int64_t type for number of elements. * Support dense param and grad from OneDNN kernel. * Enable SGD OneDNN kernel when use MP BF16 optimizer. * Force non-copyable/movable OneDNNAXPYHandler. * Reuse OneDNNAXPYHandler for spare tensors in SUM op. * Fix SFINAE rules. * Remove recording event inside AXPY. * Get rid of internal primitive caching. * Stop use PP cache mechanims to store mem and primitive obj. * Handler obj store and reuse needed desc & prim * Do not derive from MKLDNNHandlerT
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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.
PR types
Performance optimization
PR changes
OPs
Describe
This PR optimize current implementations for SGD and SUM operators for BF16 data type (mainly) when SelectedRows (sparse) tensors were used by reusing OneDNN handler.
The performance results on word2vec model are as follows on CPX 6348 machine with single thread:
This gives ~15% speedup.
From profiling: