Performance and memory improvements to batched torch.linalg.solve (2nd attempt) #71756
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
Previously for single input matrix A and batched matrix B, the matrix A was expanded and cloned before computing the LU decomposition and solving the linear system. With this PR the LU decomposition is computed once for single matrix and then expanded&cloned if required by backend library call.
Implement broadcast linear indexing for lu and pivots.
IvanYashchuk
added
module: linear algebra
CI Flow Status⚛️ CI FlowRuleset - Version:
|
Contributor
🔗 Helpful links
💊 CI failures summary and remediationsAs of commit 262fe91 (more details on the Dr. CI page):
🕵️ 1 new failure recognized by patternsThe following CI failures do not appear to be due to upstream breakages:
|
| Job | Step | Action |
|---|---|---|
 ios-12-5-1-arm64-custom-ops / build |
Unknown | 🔁 rerun |
This comment was automatically generated by Dr. CI (expand for details).
Please report bugs/suggestions to the (internal) Dr. CI Users group.
lezcano
approved these changes
Jan 25, 2022
Collaborator
lezcano
left a comment
lezcano
left a comment
There was a problem hiding this comment.
A few comments, but none of them needs to be addressed.
| return at::arange(numel, options).view(original_shape).broadcast_to(broadcast_shape).contiguous(); | ||
| } | ||
|
|
||
| class BroadcastLinearIndices { |
Collaborator
There was a problem hiding this comment.
I believe this could be used in cusolver's gesvd, as we do a very similar thing there. Not for this PR ofc.
| auto b_stride = matrixStride(b); | ||
| auto lu_stride = matrixStride(lu); | ||
| auto pivots_stride = pivots_cpu.size(-1); | ||
| auto lu_stride = lu.dim() > 2 ? lu.stride(-3) : 0; |
Collaborator
There was a problem hiding this comment.
I really like this way to get the "matrix stride". This is a great idea!
Comment on lines
2894
to
2895
| TORCH_INTERNAL_ASSERT(batchCount(b) == batchCount(lu), "batch_size of b and lu must be the same"); | ||
| TORCH_INTERNAL_ASSERT(batchCount(lu) == batchCount(pivots.unsqueeze(-1)), "batch_size of lu and pivots must be the same"); |
Collaborator
There was a problem hiding this comment.
Just another note. I just realised that we can implement a faster batchCount doing t.dim() > 2 ? t.numel() / t.stride(-3) : 1 when we have a batch of column or row-major matrices.
| c10::MaybeOwned<Tensor> maybe_expand_lu(const Tensor& b, const Tensor& lu) { | ||
| if (batchCount(b) != batchCount(lu)) { | ||
| IntArrayRef b_batch_size(b.sizes().data(), b.dim() - 2); | ||
| std::vector<int64_t> expand_size = b_batch_size.vec(); |
Collaborator
There was a problem hiding this comment.
Nit. DimVector is better suited for these use cases.
Collaborator
Author
|
CI fails with AssertionError: False is not true : Tensors failed to compare as equal!With rtol=1e-05 and atol=1e-05, found 241 element(s) (out of 256) whose difference(s) exceeded the margin of error (including 0 nan comparisons). The greatest difference was 0.6182836257459539 (0.9761708184203711 vs. 0.3578871926744173), which occurred at index (5, 5).Tests pass locally. Maybe it's just bad seed + lstsq test being flaky? |
Collaborator
|
This PR does not touch |
mruberry
approved these changes
Jan 25, 2022
Collaborator
mruberry
left a comment
mruberry
left a comment
There was a problem hiding this comment.
Stamped! Thank you @IvanYashchuk
Contributor
|
@mruberry has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator. |
Collaborator
|
@IvanYashchuk Do you want to address @lezcano's comments and use it as an opportunity to retrigger CI and see if the lstsq error persists? If it does not we should file an issue for it being a flaky test. |
Collaborator
Author
|
I replaced Let's see whether that lstsq test passes now. |
Collaborator
Author
|
This PR has accumulated conflicts and I need to resolve them. |
Collaborator
Author
|
Hey @mruberry, could you please try importing this PR? |
Collaborator
|
Hi @mruberry. Could you import this one? As part of the LU stack that I'm working on, I plan to go and use all the new tools we have to optimise |
Contributor
|
@mruberry has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator. |
mruberry
approved these changes
Mar 15, 2022
facebook-github-bot
pushed a commit
that referenced
this pull request
Mar 15, 2022
…d attempt) (#71756) Summary: Previous PR with the same content: #69752. Opening a new PR by request: #69752 (comment). ------ Previously for single input matrix A and batched matrix B, matrix A was expanded and cloned before computing the LU decomposition and solving the linear system. With this PR the LU decomposition is computed once for a single matrix and then expanded&cloned if required by a backend library call for the linear system solving. Here's a basic comparison: ```python # BEFORE THE PR In [1]: import torch In [2]: a = torch.randn(256, 256) In [3]: b = torch.randn(1024, 256, 2) In [4]: %%timeit ...: torch.linalg.solve(a, b) ...: ...: 329 ms ± 17.4 ms per loop (mean ± std. dev. of 7 runs, 1 loop each) # WITH THIS PR In [1]: import torch In [2]: a = torch.randn(256, 256) In [3]: b = torch.randn(1024, 256, 2) In [4]: %%timeit ...: torch.linalg.solve(a, b) ...: ...: 21.4 ms ± 23 µs per loop (mean ± std. dev. of 7 runs, 10 loops each) ``` Fixes #71406, fixes #71610 Pull Request resolved: #71756 Reviewed By: ngimel Differential Revision: D33771981 Pulled By: mruberry fbshipit-source-id: 0917ee36a3eb622ff75d54787b1bffe26b41cb4a
Contributor
|
Hey @IvanYashchuk. |
IvanYashchuk
added
release notes: linalg_frontend
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
5 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.
Previous PR with the same content: #69752. Opening a new PR by request: #69752 (comment).
Previously for single input matrix A and batched matrix B, matrix A was expanded and cloned before computing the LU decomposition and solving the linear system.
With this PR the LU decomposition is computed once for a single matrix and then expanded&cloned if required by a backend library call for the linear system solving.
Here's a basic comparison:
Fixes #71406, fixes #71610