pnnx: fix Conv2d padding tuple normalization

[Yeuvoir]

Related to #6463

This PR fixes Conv2d module padding normalization in pnnx and adds regression coverage for the asymmetric padding chain discussed in #6463.

TorchScript aten::pad stores 2D pads as (left, right, top, bottom), while nn.Conv2d padding params are (height, width). The previous logic only resized equal 4-element pads, which could preserve the wrong order for per-axis padding such as padding=(1, 2) with padding_mode=reflect/replicate.

Changes:

• normalize symmetric 4-element 2D pads to (top, left) before storing nn.Conv2d params
• reuse the normalization for aten::pad, reflection_pad2d, and replication_pad2d
• add regression coverage for ZeroPad2d((1, 2, 1, 2)) -> depthwise Conv2d(kernel=5, stride=2) -> BatchNorm2d
• verify the regression keeps pnnx output shape (1, 4, 28, 28), folds BN into Conv2d, and emits ncnn pads left=1, top=1, right=2, bottom=2
• extend test_nn_Conv2d with padding=(1, 2), padding_mode='reflect'

Verified on Windows with VS2022/v143 Release pnnx build:

• manual test_nn_Conv2d: passed
• manual test_pnnx_fuse_asymmetric_pad_conv2d with BatchNorm2d: passed
• issue EfficentNet模型转换错误，转出的中间过程有个size不正确 #6463 repro with BatchNorm2d outputs (1, 144, 28, 28), pnnx allclose passes, and ncnn param keeps pads left=1, top=1, right=2, bottom=2

Note: direct ctest on Windows fails before running pnnx because the existing test scripts call ../src/pnnx through cmd; I verified the same tests manually with the built pnnx.exe.

[tencent-adm]

Thank you for your submission, we really appreciate it. Like many open source projects, we ask that you all sign our Contributor License Agreement before we can accept your contribution.
1 out of 2 committers have signed the CLA.

✅ Yeuvoir
❌ nihui
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[nihui]

@codex review

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. Already looking forward to the next diff.

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[Copilot]

Pull request overview

This PR addresses incorrect padding tuple normalization when fusing PyTorch Conv2d with non-zero padding_mode in pnnx, specifically fixing the (left,right,top,bottom) vs (height,width) ordering mismatch that can break asymmetric-per-axis padding cases. It also adds regression coverage for the asymmetric pad → depthwise conv → batchnorm fusion scenario described in #6463.

Changes:

• Normalize 4-element 2D pad lists from aten::pad / reflection_pad2d / replication_pad2d into nn.Conv2d’s 2-tuple (height, width) form when symmetric per axis.
• Reuse the same normalization logic across the relevant padding-node cases in the Conv2d module fusion pass.
• Add/extend tests to cover asymmetric padding fusion and padding=(1,2) with padding_mode='reflect'.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 1 comment.

File	Description
tools/pnnx/src/pass_level1/nn_Conv2d.cpp	Adds a shared helper to normalize 4-element 2D pads into Conv2d padding parameters during module fusion.
tools/pnnx/tests/test_pnnx_fuse_asymmetric_pad_conv2d.py	New regression test for asymmetric ZeroPad2d → depthwise Conv2d(stride=2,k=5) → BatchNorm2d, including param checks.
tools/pnnx/tests/test_nn_Conv2d.py	Extends Conv2d test coverage with padding=(1,2) and padding_mode='reflect'.
tools/pnnx/tests/CMakeLists.txt	Registers the new asymmetric pad/conv2d fuse regression test in the test suite.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.