Milestone2.1: Partition to_dim_order_copy op in XNN delegate #12220
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/pytorch/executorch/12220
Note: Links to docs will display an error until the docs builds have been completed. ❌ 3 New FailuresAs of commit 365de21 with merge base a8070ec ( NEW FAILURES - The following jobs have failed:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
facebook-github-bot
added
the
CLA Signed
Contributor
Author
|
@pytorchbot label "release notes: none" |
pytorch-bot
bot
added
the
release notes: none
digantdesai
reviewed
Jul 11, 2025
| # The node requires nchw inputs | ||
| for input_node in node.all_input_nodes: | ||
| self.input_to_nchw(graph_module, input_node, node) | ||
| elif node.target == exir_ops.edge.aten._to_copy.default: |
Contributor
There was a problem hiding this comment.
so the reason we still have to_copy even after we partition to_dim_order_copy is because we revert it back to to_copy. So when we add a node visitor next for the to_dim_order_copy we should remove the revert pass.
Contributor
Author
There was a problem hiding this comment.
I see, should I make those changes in a follow up pr or would it be better to keep them here?
| @@ -0,0 +1,85 @@ | |||
| # Copyright (c) Meta Platforms, Inc. and affiliates. | |||
Contributor
There was a problem hiding this comment.
As I said earlier, using to_copy is OK but we can just as easily move to to_dim_order_copy and remove the dim_order ops revert pass.
leafs1
force-pushed
the
milestone2.1
branch
3 times, most recently
from
bbc194c to
c26c56b
Compare
mcr229
reviewed
Jul 11, 2025
| return True | ||
|
|
||
| def supported_precision_types(self) -> List[ConfigPrecisionType]: | ||
| return [ConfigPrecisionType.FP32] |
Contributor
There was a problem hiding this comment.
add ConfigPrecisionType.STATIC_QUANT
Contributor
|
lets wait for CI, but looking good! |
mcr229
approved these changes
Jul 14, 2025
SS-JIA
added a commit
that referenced
this pull request
Jul 16, 2025
SS-JIA
added a commit
that referenced
this pull request
Jul 16, 2025
…12220)" (#12542) This reverts commit dd6caa3. The imported diff is breaking an internal test: [D78368033](https://www.internalfb.com/diff/D78368033). Please see the diff for more details.
lucylq
pushed a commit
that referenced
this pull request
Jul 17, 2025
### Summary This PR adds support for the `to_dim_order_copy` operation in the XNNPACK delegate partitioner, enabling direct handling of memory format conversions initiated by users via `.to(memory_format=)` calls. This enhancement significantly improves performance by producing more compressed graphs that avoid unnecessary partitioning boundaries at memory format conversion points. By delegating these operations directly to XNNPACK, we eliminate the overhead of context switching between the runtime and delegate, reducing both execution time and memory footprint. The implementation leverages XNNPACK's highly optimized memory format conversion routines, which are specifically designed for efficient tensor layout transformations on various hardware targets. ### Test plan Confirmed expected output when having user specified dim order conversions as well as appropriate partitioning. I did this by writing individual tests for the to_copy op ensuring it changes dim order and dtype when appropriate. Also added test module to confirm that the to copy nodes are partitioned and not in another partition
lucylq
pushed a commit
that referenced
this pull request
Jul 17, 2025
…12220)" (#12542) This reverts commit dd6caa3. The imported diff is breaking an internal test: [D78368033](https://www.internalfb.com/diff/D78368033). Please see the diff for more details.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
4 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.
Summary
This PR adds support for the
to_dim_order_copyoperation in the XNNPACK delegate partitioner, enabling direct handling of memory format conversions initiated by users via.to(memory_format=)calls. This enhancement significantly improves performance by producing more compressed graphs that avoid unnecessary partitioning boundaries at memory format conversion points. By delegating these operations directly to XNNPACK, we eliminate the overhead of context switching between the runtime and delegate, reducing both execution time and memory footprint. The implementation leverages XNNPACK's highly optimized memory format conversion routines, which are specifically designed for efficient tensor layout transformations on various hardware targets.Test plan
Confirmed expected output when having user specified dim order conversions as well as appropriate partitioning. I did this by writing individual tests for the to_copy op ensuring it changes dim order and dtype when appropriate. Also added test module to confirm that the to copy nodes are partitioned and not in another partition