[MPS] Gather sliced inputs to batch norm #133610
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/133610
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit 4b16953 with merge base 546c53b ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
hvaara
commented
Aug 15, 2024
Comment on lines
-156
to
-161
| bool executeGatherOp = true; | ||
| if (self.is_contiguous(memory_format)) { | ||
| memory_format = MemoryFormat::Contiguous; | ||
| executeGatherOp = false; | ||
| } | ||
|
|
There was a problem hiding this comment.
Does my assumption that the decision to gather is sufficiently handled in
hold?There was a problem hiding this comment.
Considering #94760, the issue is tricky I would say. The test expects input layout is preserved in output. Let's look at CI signals.
There was a problem hiding this comment.
Looks like everything passed in CI.
Comment on lines
+2542
to
+2545
| bn_cpu = nn.BatchNorm2d(100, affine=False, device='cpu') | ||
| bn_mps = nn.BatchNorm2d(100, affine=False, device='mps') | ||
|
|
||
| x_cpu = torch.randn(100, 100, 35, 45).to('cpu') |
There was a problem hiding this comment.
Are these values fine, or should I consider lowering them to cover the minimal case? timeit results suggests a single loop is sub-millisecond.
There was a problem hiding this comment.
If it's sub-millisecond this is fine
|
@pytorchbot label "module: mps" "ciflow/mps" |
|
Can't add following labels to PR: ciflow/mps. Please ping one of the reviewers for help. |
|
cc @bauerwer for visibility |
|
@hvaara Quick check passed on my system. I integrated the Normalization.mm changes into my torch main, built locally and was able to run flux without noise in the image. I ran a few other checkpoints and workflows as well without problems. |
qqaatw
reviewed
Aug 17, 2024
Comment on lines
-156
to
-161
| bool executeGatherOp = true; | ||
| if (self.is_contiguous(memory_format)) { | ||
| memory_format = MemoryFormat::Contiguous; | ||
| executeGatherOp = false; | ||
| } | ||
|
|
There was a problem hiding this comment.
Considering #94760, the issue is tricky I would say. The test expects input layout is preserved in output. Let's look at CI signals.
|
@kulinseth @malfet @qqaatw are either of you available for a review? 🙏 |
malfet
approved these changes
Aug 20, 2024
Comment on lines
+2542
to
+2545
| bn_cpu = nn.BatchNorm2d(100, affine=False, device='cpu') | ||
| bn_mps = nn.BatchNorm2d(100, affine=False, device='mps') | ||
|
|
||
| x_cpu = torch.randn(100, 100, 35, 45).to('cpu') |
There was a problem hiding this comment.
If it's sub-millisecond this is fine
|
@pytorchbot merge -f "MPS tests are green" |
Merge startedYour change will be merged immediately since you used the force (-f) flag, bypassing any CI checks (ETA: 1-5 minutes). Please use Learn more about merging in the wiki. Questions? Feedback? Please reach out to the PyTorch DevX Team |
|
Thanks for the review! Following up in #133520 to address comments and future work. |
|
@pytorchbot cherry-pick --onto release/2.4 -c critical --fixes #133520 |
pytorchbot
pushed a commit
that referenced
this pull request
Aug 21, 2024
This PR removes the `executeGatherOp` flag from batch norm in favor of relying on the logic in https://github.com/pytorch/pytorch/blob/4aa66f68a803927ddd127ceaaa1521b8d6e90e5f/aten/src/ATen/native/mps/OperationUtils.mm#L372 to decide if gathering is necessary. It's not the most efficient way to solve this issue, but it assures correctness for sliced inputs. ### Performance impact #### With fix ``` python -m timeit -n 100 -s "import torch; import torch.nn as nn; bn = nn.BatchNorm2d(100, affine=False, device='mps');x = torch.randn(100, 100, 35, 45).to('mps')" "bn(x)" 100 loops, best of 5: 282 usec per loop python -m timeit -n 100 -s "import torch; import torch.nn as nn; bn = nn.BatchNorm2d(100, affine=False, device='mps');x = torch.randn(100, 100, 35, 45).to('mps')" "bn(x[5:])" 100 loops, best of 5: 448 usec per loop python -m timeit -n 1000 -s "import torch; import torch.nn as nn; bn = nn.BatchNorm2d(100, affine=False, device='mps');x = torch.randn(100, 100, 35, 45).to('mps')" "bn(x)" 1000 loops, best of 5: 705 usec per loop python -m timeit -n 1000 -s "import torch; import torch.nn as nn; bn = nn.BatchNorm2d(100, affine=False, device='mps');x = torch.randn(100, 100, 35, 45).to('mps')" "bn(x[5:])" 1000 loops, best of 5: 1.11 msec per loop python -m timeit -n 1000 -s "import torch; import torch.nn as nn; bn = nn.BatchNorm2d(100, affine=False, device='mps');x = torch.randn(1000, 100, 35, 45).to('mps')" "bn(x)" 1000 loops, best of 5: 7.16 msec per loop python -m timeit -n 1000 -s "import torch; import torch.nn as nn; bn = nn.BatchNorm2d(100, affine=False, device='mps');x = torch.randn(1000, 100, 35, 45).to('mps')" "bn(x[5:])" 1000 l A93C oops, best of 5: 11.7 msec per loop ``` #### Without fix ``` python -m timeit -n 100 -s "import torch; import torch.nn as nn; bn = nn.BatchNorm2d(100, affine=False, device='mps');x = torch.randn(100, 100, 35, 45).to('mps')" "bn(x)" 100 loops, best of 5: 284 usec per loop python -m timeit -n 100 -s "import torch; import torch.nn as nn; bn = nn.BatchNorm2d(100, affine=False, device='mps');x = torch.randn(100, 100, 35, 45).to('mps')" "bn(x[5:])" 100 loops, best of 5: 265 usec per loop python -m timeit -n 1000 -s "import torch; import torch.nn as nn; bn = nn.BatchNorm2d(100, affine=False, device='mps');x = torch.randn(100, 100, 35, 45).to('mps')" "bn(x)" 1000 loops, best of 5: 715 usec per loop python -m timeit -n 1000 -s "import torch; import torch.nn as nn; bn = nn.BatchNorm2d(100, affine=False, device='mps');x = torch.randn(100, 100, 35, 45).to('mps')" "bn(x[5:])" 1000 loops, best of 5: 675 usec per loop python -m timeit -n 1000 -s "import torch; import torch.nn as nn; bn = nn.BatchNorm2d(100, affine=False, device='mps');x = torch.randn(1000, 100, 35, 45).to('mps')" "bn(x)" 1000 loops, best of 5: 7.19 msec per loop python -m timeit -n 1000 -s "import torch; import torch.nn as nn; bn = nn.BatchNorm2d(100, affine=False, device='mps');x = torch.randn(1000, 100, 35, 45).to('mps')" "bn(x[5:])" 1000 loops, best of 5: 7.13 msec per loop ``` Please feel free to push back or request changes. Fixes #133520 Pull Request resolved: #133610 Approved by: https://github.com/malfet (cherry picked from commit 43f78bf)
Cherry picking #133610The cherry pick PR is at #134121 and it is linked with issue #133520. The following tracker issues are updated: Details for Dev Infra teamRaised by workflow job |
atalman
pushed a commit
that referenced
this pull request
Aug 22, 2024
[MPS] Gather sliced inputs to batch norm (#133610) This PR removes the `executeGatherOp` flag from batch norm in favor of relying on the logic in https://github.com/pytorch/pytorch/blob/4aa66f68a803927ddd127ceaaa1521b8d6e90e5f/aten/src/ATen/native/mps/OperationUtils.mm#L372 to decide if gathering is necessary. It's not the most efficient way to solve this issue, but it assures correctness for sliced inputs. ### Performance impact #### With fix ``` python -m timeit -n 100 -s "import torch; import torch.nn as nn; bn = nn.BatchNorm2d(100, affine=False, device='mps');x = torch.randn(100, 100, 35, 45).to('mps')" "bn(x)" 100 loops, best of 5: 282 usec per loop python -m timeit -n 100 -s "import torch; import torch.nn as nn; bn = nn.BatchNorm2d(100, affine=False, device='mps');x = torch.randn(100, 100, 35, 45).to('mps')" "bn(x[5:])" 100 loops, best of 5: 448 usec per loop python -m timeit -n 1000 -s "import torch; import torch.nn as nn; bn = nn.BatchNorm2d(100, affine=False, device='mps');x = torch.randn(100, 100, 35, 45).to('mps')" "bn(x)" 1000 loops, best of 5: 705 usec per loop python -m timeit -n 1000 -s "import torch; import torch.nn as nn; bn = nn.BatchNorm2d(100, affine=False, device='mps');x = torch.randn(100, 100, 35, 45).to('mps')" "bn(x[5:])" 1000 loops, best of 5: 1.11 msec per loop python -m timeit -n 1000 -s "import torch; import torch.nn as nn; bn = nn.BatchNorm2d(100, affine=False, device='mps');x = torch.randn(1000, 100, 35, 45).to('mps')" "bn(x)" 1000 loops, best of 5: 7.16 msec per loop python -m timeit -n 1000 -s "import torch; import torch.nn as nn; bn = nn.BatchNorm2d(100, affine=False, device='mps');x = torch.randn(1000, 100, 35, 45).to('mps')" "bn(x[5:])" 1000 loops, best of 5: 11.7 msec per loop ``` #### Without fix ``` python -m timeit -n 100 -s "import torch; import torch.nn as nn; bn = nn.BatchNorm2d(100, affine=False, device='mps');x = torch.randn(100, 100, 35, 45).to('mps')" "bn(x)" 100 loops, best of 5: 284 usec per loop python -m timeit -n 100 -s "import torch; import torch.nn as nn; bn = nn.BatchNorm2d(100, affine=False, device='mps');x = torch.randn(100, 100, 35, 45).to('mps')" "bn(x[5:])" 100 loops, best of 5: 265 usec per loop python -m timeit -n 1000 -s "import torch; import torch.nn as nn; bn = nn.BatchNorm2d(100, affine=False, device='mps');x = torch.randn(100, 100, 35, 45).to('mps')" "bn(x)" 1000 loops, best of 5: 715 usec per loop python -m timeit -n 1000 -s "import torch; import torch.nn as nn; bn = nn.BatchNorm2d(100, affine=False, device='mps');x = torch.randn(100, 100, 35, 45).to('mps')" "bn(x[5:])" 1000 loops, best of 5: 675 usec per loop python -m timeit -n 1000 -s "import torch; import torch.nn as nn; bn = nn.BatchNorm2d(100, affine=False, device='mps');x = torch.randn(1000, 100, 35, 45).to('mps')" "bn(x)" 1000 loops, best of 5: 7.19 msec per loop python -m timeit -n 1000 -s "import torch; import torch.nn as nn; bn = nn.BatchNorm2d(100, affine=False, device='mps');x = torch.randn(1000, 100, 35, 45).to('mps')" "bn(x[5:])" 1000 loops, best of 5: 7.13 msec per loop ``` Please feel free to push back or request changes. Fixes #133520 Pull Request resolved: #133610 Approved by: https://github.com/malfet (cherry picked from commit 43f78bf) Co-authored-by: Roy Hvaara <roy@lightyear.no>
40 tasks
pytorch-bot bot
pushed a commit
that referenced
this pull request
Sep 13, 2024
This PR removes the `executeGatherOp` flag from batch norm in favor of relying on the logic in https://github.com/pytorch/pytorch/blob/4aa66f68a803927ddd127ceaaa1521b8d6e90e5f/aten/src/ATen/native/mps/OperationUtils.mm#L372 to decide if gathering is necessary. It's not the most efficient way to solve this issue, but it assures correctness for sliced inputs. ### Performance impact #### With fix ``` python -m timeit -n 100 -s "import torch; import torch.nn as nn; bn = nn.BatchNorm2d(100, affine=False, device='mps');x = torch.randn(100, 100, 35, 45).to('mps')" "bn(x)" 100 loops, best of 5: 282 usec per loop python -m timeit -n 100 -s "import torch; import torch.nn as nn; bn = nn.BatchNorm2d(100, affine=False, device='mps');x = torch.randn(100, 100, 35, 45).to('mps')" "bn(x[5:])" 100 loops, best of 5: 448 usec per loop python -m timeit -n 1000 -s "import torch; import torch.nn as nn; bn = nn.BatchNorm2d(100, affine=False, device='mps');x = torch.randn(100, 100, 35, 45).to('mps')" "bn(x)" 1000 loops, best of 5: 705 usec per loop python -m timeit -n 1000 -s "import torch; import torch.nn as nn; bn = nn.BatchNorm2d(100, affine=False, device='mps');x = torch.randn(100, 100, 35, 45).to('mps')" "bn(x[5:])" 1000 loops, best of 5: 1.11 msec per loop python -m timeit -n 1000 -s "import torch; import torch.nn as nn; bn = nn.BatchNorm2d(100, affine=False, device='mps');x = torch.randn(1000, 100, 35, 45).to('mps')" "bn(x)" 1000 loops, best of 5: 7.16 msec per loop python -m timeit -n 1000 -s "import torch; import torch.nn as nn; bn = nn.BatchNorm2d(100, affine=False, device='mps');x = torch.randn(1000, 100, 35, 45).to('mps')" "bn(x[5:])" 1000 loops, best of 5: 11.7 msec per loop ``` #### Without fix ``` python -m timeit -n 100 -s "import torch; import torch.nn as nn; bn = nn.BatchNorm2d(100, affine=False, device='mps');x = torch.randn(100, 100, 35, 45).to('mps')" "bn(x)" 100 loops, best of 5: 284 usec per loop python -m timeit -n 100 -s "import torch; import torch.nn as nn; bn = nn.BatchNorm2d(100, affine=False, device='mps');x = torch.randn(100, 100, 35, 45).to('mps')" "bn(x[5:])" 100 loops, best of 5: 265 usec per loop python -m timeit -n 1000 -s "import torch; import torch.nn as nn; bn = nn.BatchNorm2d(100, affine=False, device='mps');x = torch.randn(100, 100, 35, 45).to('mps')" "bn(x)" 1000 loops, best of 5: 715 usec per loop python -m timeit -n 1000 -s "import torch; import torch.nn as nn; bn = nn.BatchNorm2d(100, affine=False, device='mps');x = torch.randn(100, 100, 35, 45).to('mps')" "bn(x[5:])" 1000 loops, best of 5: 675 usec per loop python -m timeit -n 1000 -s "import torch; import torch.nn as nn; bn = nn.BatchNorm2d(100, affine=False, device='mps');x = torch.randn(1000, 100, 35, 45).to('mps')" "bn(x)" 1000 loops, best of 5: 7.19 msec per loop python -m timeit -n 1000 -s "import torch; import torch.nn as nn; bn = nn.BatchNorm2d(100, affine=False, device='mps');x = torch.randn(1000, 100, 35, 45).to('mps')" "bn(x[5:])" 1000 loops, best of 5: 7.13 msec per loop ``` Please feel free to push back or request changes. Fixes #133520 Pull Request resolved: #133610 Approved by: https://github.com/malfet
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.
This PR removes the
executeGatherOpflag from batch norm in favor of relying on the logic inpytorch/aten/src/ATen/native/mps/OperationUtils.mm
Line 372 in 4aa66f6
It's not the most efficient way to solve this issue, but it assures correctness for sliced inputs.
Performance impact
With fix
Without fix
Please feel free to push back or request changes.
Fixes #133520