[Rewriter] Add optimizer to fold Pad operators into Conv

[Johansmm]

Following (#2301), fuse_pad_into_conv rule set is introduced to reduce the following list of operators:

• Conv ∘ Pad -> Conv
• ConvInteger ∘ Pad -> ConvInteger

Additionally, NormalizePadFormat is introduced in order to change auto_pads Conv attribute in its explicit pads list (ref: https://onnx.ai/onnx/operators/onnx__Conv.html).

[codecov]

Codecov Report

Attention: Patch coverage is 95.48611% with 13 lines in your changes missing coverage. Please review.

Project coverage is 69.83%. Comparing base (e63a16b) to head (7f7d17e).

Files with missing lines	Patch %	Lines
onnxscript/rewriter/fuse_pad_into_conv_test.py	94.44%	4 Missing and 4 partials ⚠️
onnxscript/rewriter/fuse_pad_into_conv.py	96.52%	3 Missing and 2 partials ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #2363      +/-   ##
==========================================
+ Coverage   69.55%   69.83%   +0.28%     
==========================================
  Files         210      212       +2     
  Lines       26120    26408     +288     
  Branches     2721     2764      +43     
==========================================
+ Hits        18168    18443     +275     
- Misses       7017     7024       +7     
- Partials      935      941       +6     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[Johansmm]

Push force rebasing on main and fixing conflicts.

[Johansmm]

@justinchuby I forgot to explain in the previous message that the changes were not ready yet (just fixing rebase with main).
In the last commits I update the code with all the suggestions.
Now this work is ready to be reviewed.

[justinchuby]

Could you update the PR title and description? Thanks

[Johansmm]

@justinchuby is ok the new description and title ?

[justinchuby]

Thanks - maybe @gramalingam or @titaiwangms for another review?

[titaiwangms]

Please run "lintrunner -a"

[titaiwangms]

I suppose this does not fallback when the rewritten conv still does not match FusePad rules? Do we still benefit from converting auto_pad attribute into 'NOTSET' if it's still a no match?

[Johansmm]

You are right, always NormalizePadFormat is applied, even if FusePad is not applied.

On the other hand, I consider that having a single pad format is coherent, since it makes it easier to be interpreted by different accelerators. (bug example in onnxruntime due to the auto_pad type).

Let me know if you still disagree with this information. What do you think @justinchuby.

[titaiwangms]

cc @gramalingam

[Johansmm]

Last commit with @titaiwangms suggestions

[Johansmm]

You are right, always NormalizePadFormat is applied, even if FusePad is not applied.

On the other hand, I consider that having a single pad format is coherent, since it makes it easier to be interpreted by different accelerators. (bug example in onnxruntime due to the auto_pad type).

Let me know if you still disagree with this information. What do you think @justinchuby.

[Copilot]

Pull Request Overview

Adds rewrite rules that fold preceding Pad operators into Conv/ConvInteger nodes and normalize Conv auto_pad attributes to explicit pads.

• Define fuse_pad_into_conv and fuse_pad_into_conv_integer rules to merge Pad into Conv/ConvInteger
• Introduce normalize_pad_format_conv/normalize_pad_format_conv_integer to convert auto_pad into explicit pads
• Add and register comprehensive unit tests for fusion and normalization, and update __init__.py to expose the new rules

Reviewed Changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 2 comments.

File	Description
onnxscript/rewriter/fuse_pad_into_conv.py	New rewrite rules for fusing Pad into Conv/ConvInteger and normalizing pad formats
onnxscript/rewriter/fuse_pad_into_conv_test.py	Unit tests for both fusion and normalization behaviors
onnxscript/rewriter/init.py	Imported and included the new rules in the global rule set

Comments suppressed due to low confidence (2)

onnxscript/rewriter/fuse_pad_into_conv.py:202

• [nitpick] The error message in NotImplementedError has a grammar issue; consider updating it to 'Subclasses must implement this function' for clarity.

        raise NotImplementedError("Child have to implement this function")

onnxscript/rewriter/fuse_pad_into_conv_test.py:405

• There are tests for NormalizePadFormatConv but none for NormalizePadFormatConvInteger; consider adding similar tests to cover the ConvInteger normalization rule.

if __name__ == "__main__":

[Johansmm]

Last push force rebasing on main and addig @justinchuby suggestions

[gramalingam]

nit: I suggest if not check_result: ... a MatchResult can be treated as a boolean, with True meaning success, and False meaning failure.

[gramalingam]

minor nit: I think this condition can be simplified to conv_node.attributes.get_string("auto_pad", "NOTSET") != "NOTSET" .

[gramalingam]

The above line needs to handle various special-conditions and error-situations. I suggest using onnxscript.rewriter._ir_utils.get_numpy_value(pad_node.inputs[1]) which will handle the cases where the input is None or is not a constant. And check that the value is not None before using it.

[gramalingam]

Oh, I see, the check is being done in the check method done below.

[gramalingam]

If this is the same as the pad_pads used in the rewrite method, you can save it as self._pads_list and use it in the rewrite method to avoid redoing the entire computation. It involves implicitly passed state, but works fine.