Enable cusolver potrf batched for Cholesky decomposition when cuda >= 11.3#57788
Closed
Enable cusolver potrf batched for Cholesky decomposition when cuda >= 11.3#57788
Conversation
Contributor
💊 CI failures summary and remediationsAs of commit 003fe38 (more details on the Dr. CI page): 💚 💚 Looks good so far! There are no failures yet. 💚 💚 This comment was automatically generated by Dr. CI (expand for details).Follow this link to opt-out of these comments for your Pull Requests.Please report bugs/suggestions to the (internal) Dr. CI Users group. |
Collaborator
Author
|
reserved |
Codecov Report
@@ Coverage Diff @@
## master #57788 +/- ##
==========================================
- Coverage 76.83% 76.83% -0.01%
==========================================
Files 1986 1986
Lines 197430 197430
==========================================
- Hits 151691 151690 -1
- Misses 45739 45740 +1 |
Collaborator
|
If both cuSOLVER and MAGMA are available and CUDA version is < 11.3 we should continue using batched MAGMA as it has better performance, than single input cuSOLVER variant called in a loop, right? This PR modifies the behavior for < 11.3 versions to use looped cuSOLVER instead of batched MAGMA. Besides that dispatch issue, everything looks good. |
Collaborator
Author
|
Ohhh, yes, you're right. Let me fix that dispatch logic. 😄 |
…holesky-batched_cuda11.3
mruberry
reviewed
May 11, 2021
|
|
||
| // Implementation of Cholesky decomposition using batched cusolverDn<T>potrfBatched | ||
| // Warning: cusolverDn<T>potrfBatched doesn't work quite well when matrix size or batch size is zero. | ||
| // If you write your own C++ extension and use this function, make sure you do a zero numel check for the input. |
mruberry
reviewed
May 11, 2021
| #define USE_CUSOLVER | ||
| #endif | ||
|
|
||
| // cusolverDn<T>potrfBatched may have numerical issue before cuda 11.3 release, |
Contributor
|
@mruberry has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator. |
Contributor
krshrimali
pushed a commit
to krshrimali/pytorch
that referenced
this pull request
May 19, 2021
… 11.3 (pytorch#57788) Summary: This PR enables the usage of cusolver potrf batched as the backend of Cholesky decomposition (`torch.linalg.cholesky` and `torch.linalg.cholesky_ex`) when cuda version is greater than or equal to 11.3. Benchmark available at https://github.com/xwang233/code-snippet/tree/master/linalg/cholesky-new. It is seen that cusolver potrf batched performs better than magma potrf batched in most cases. ## cholesky dispatch heuristics: ### before: - batch size == 1: cusolver potrf - batch size > 1: magma xpotrf batched ### after: cuda >= 11.3: - batch size == 1: cusolver potrf - batch size > 1: cusolver potrf batched cuda < 11.3 (not changed): - batch size == 1: cusolver potrf - batch size > 1: magma xpotrf batched --- See also pytorch#42666 pytorch#47953 pytorch#53104 pytorch#53879 Pull Request resolved: pytorch#57788 Reviewed By: ngimel Differential Revision: D28345530 Pulled By: mruberry fbshipit-source-id: 3022cf73b2750e1953c0e00a9e8b093dfc551f61
laurentdupin
pushed a commit
to laurentdupin/pytorch
that referenced
this pull request
Apr 25, 2026
… 11.3 (pytorch#57788) Summary: This PR enables the usage of cusolver potrf batched as the backend of Cholesky decomposition (`torch.linalg.cholesky` and `torch.linalg.cholesky_ex`) when cuda version is greater than or equal to 11.3. Benchmark available at https://github.com/xwang233/code-snippet/tree/master/linalg/cholesky-new. It is seen that cusolver potrf batched performs better than magma potrf batched in most cases. ## cholesky dispatch heuristics: ### before: - batch size == 1: cusolver potrf - batch size > 1: magma xpotrf batched ### after: cuda >= 11.3: - batch size == 1: cusolver potrf - batch size > 1: cusolver potrf batched cuda < 11.3 (not changed): - batch size == 1: cusolver potrf - batch size > 1: magma xpotrf batched --- See also pytorch#42666 pytorch#47953 pytorch#53104 pytorch#53879 Pull Request resolved: pytorch#57788 Reviewed By: ngimel Differential Revision: D28345530 Pulled By: mruberry fbshipit-source-id: 3022cf73b2750e1953c0e00a9e8b093dfc551f61
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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
6 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.
This PR enables the usage of cusolver potrf batched as the backend of Cholesky decomposition (
torch.linalg.choleskyandtorch.linalg.cholesky_ex) when cuda version is greater than or equal to 11.3.Benchmark available at https://github.com/xwang233/code-snippet/tree/master/linalg/cholesky-new. It is seen that cusolver potrf batched performs better than magma potrf batched in most cases.
cholesky dispatch heuristics:
before:
after:
cuda >= 11.3:
cuda < 11.3 (not changed):
See also #42666 #47953 #53104 #53879