pnnx ncnn legalize global pooling layout

[nihui]

No description provided.

[tencent-adm]

Thank you for your submission, we really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[nihui]

Supplementary notes for this PR:

This PR changes pnnx ncnn conversion around global pooling layout, without changing ncnn runtime Pooling behavior.

Design split:

• insert_reshape_numpy_binaryop_broadcast still runs first and keeps torch/numpy broadcast semantics explicit. This is important for cases like [4] * [4,5,4], which is last-dim torch broadcast and must not be treated as ncnn channel broadcast.
• legalize_global_pooling_layout then rewrites only proven global-pooling outputs into ncnn compact vector shape in Operand::shape:
  • [N,C,1,1] / [N,C,1,1,1] -> [N,C]
  • [C,1,1] / [C,1,1,1] -> [C]
    No extra layout metadata is introduced.
• Global-pool branches stay compact through activation, Linear, and valid Conv2d 1x1 groups=1. Such convs are converted to nn.Linear, with weight shape changed from [out,in,1,1] to [out,in].
• Rank-sensitive consumers such as permute, slice, non-vector reshape, depthwise/grouped/non-1x1 conv, and model output get a restore Tensor.reshape only at the boundary where torch rank is needed.
• A later eliminate_reshape_binaryop_broadcast pass runs after ncnn layer rewriting. It only removes no-batch channel broadcast adapters that are already equivalent under ncnn BinaryOp, such as [C] -> [C,1,1] before [C,H,W]. It intentionally keeps last-dim broadcast adapters like [4] -> [1,1,4], and also keeps explicit batched model-input adapters [N,C] -> [N,C,1,1] for now.

Expected generated graph examples:

• classifier head: Pooling -> InnerProduct
• SE-style branch: Pooling -> InnerProduct -> activation -> InnerProduct -> sigmoid -> BinaryOp
• rank-sensitive fallback: Pooling -> Reshape -> Permute/Slice/Convolution...

Tests added/updated:

• test_ncnn_global_pooling_layout.py
  • direct global-pool output
  • no-batch global-pool output
  • avg/max global-pool -> flatten -> linear
  • 3D global-pool -> linear
  • F.linear
  • SE conv1x1 chain and BinaryOp broadcast
  • permute/slice/reshape fallback
  • depthwise and non-1x1 conv fallback
  • conversion without explicit inputshape
• test_ncnn_binaryop_broadcast_cleanup.py
  • removable no-batch channel adapter
  • non-removable last-dim broadcast adapter
  • non-removable explicit batched adapter
• test_ncnn_numpy_binaryop_broadcast.py
  • adds [4] with [4,5,4] regression coverage

Focused validation run locally:

cd tools/pnnx/build
make -j16 pnnx
ctest --output-on-failure -R "AdaptiveAvgPool|AdaptiveMaxPool|Linear|Conv2d|global_pooling|binaryop_broadcast_cleanup|numpy_binaryop_broadcast|batch_layout|solve_batch_index"

Result: 21/21 passed.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 05fbcc10ba

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".

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 2ee2f629e6

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".

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: fb079900f4

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".

[nihui]

link #6777