Fix circulate/reflection/replication padding and conv2d with different padding

[cccclai]

Differential Revision: D84071939

This PR make changes for the Pad and Conv2d with pad option

1. For Pad op, first of all, QNN can support it so we don't decompose it. circular mode is not supported by QNN so it has to be converted anyway. However, for replicate, ideally it should work with OpPad.Scheme.EDGE but I have been getting wrong result. In this PR, I convert pad with circular mode and replicate mode for the pass. We can follow up on why replicate didn't work with OpPad.Scheme.EDGE and remove the condition

2. Slice copy op needs to be fixed with conv2d node, because we convert pad to slice copy in step 1)

3. For Conv2d with pad option, it just work out of box once we fix the pad op

Add unit test for pad and conv2d with pad option, also add test for the new pass

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/14860

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

❌ 4 New Failures, 1 Cancelled Job, 1 Unrelated Failure

As of commit 3f83885 with merge base 2672dd3 ():

NEW FAILURES - The following jobs have failed:

• Lint / lintrunner / linux-job (gh)
  >>> Lint for backends/qualcomm/tests/test_qnn_delegate.py:
• pull / test-samsung-models-linux / linux-job (gh)
  RuntimeError: Command docker exec -t add54df038d14b24114cdf8968136c590addac5c57f45810b3da2df017adb61c /exec failed with exit code 1
• pull / unittest-arm-backend-with-no-fvp (test_pytest_ops) / linux-job (gh)
  RuntimeError: Command docker exec -t 58966ec5a3c50cbd52f19940001f0eebc4746087a1e671874a251a21f0df63d3 /exec failed with exit code 1
• Test CUDA Builds / test-voxtral-cuda-e2e / linux-job (gh)
  RuntimeError: Command docker exec -t 4b794a8848d284b7a831483c44b85ed6e3b7ac6c4dcb6504610ea1e0506fb896 /exec failed with exit code 2

CANCELLED JOB - The following job was cancelled. Please retry:

• Test CoreML Backend / test-coreml / test-backend-macos (coreml_static_int8, operators) / macos-job (gh)
  ##[error]The operation was canceled.

FLAKY - The following job failed but was likely due to flakiness present on trunk:

• pull / unittest-arm-backend-with-no-fvp (test_pytest_models) / linux-job (gh) (detected as infra flaky with no log or failing log classifier)

This comment was automatically generated by Dr. CI and updates every 15 minutes.

[meta-codesync]

@cccclai has exported this pull request. If you are a Meta employee, you can view the originating Diff in D84071939.

[github-actions]

This PR needs a release notes: label

If your change should be included in the release notes (i.e. would users of this library care about this change?), please use a label starting with release notes:. This helps us keep track and include your important work in the next release notes.

To add a label, you can comment to pytorchbot, for example
@pytorchbot label "release notes: none"

For more information, see
https://github.com/pytorch/pytorch/wiki/PyTorch-AutoLabel-Bot#why-categorize-for-release-notes-and-how-does-it-work.

[cccclai]

@winskuo-quic it seems like you're fixing failing. I have this pr fixing some of the conv2d with different mode but still missing circular mode.

[chenweng-quic]

Is meta['val'].shape correct for slice and cat?
I feel like their shapes cannot be copied directly and need some transform function.

[winskuo-quic]

Thanks for supporting the conv2d padding mode!

I noticed replicate mode generates correct output when running quantized flow, while FP has wrong output, which I am guessing is caused by QNN SDK.

If this is the case, maybe we don't need to convert replicate mode when running quant flow. We can use a flag like following:

executorch/backends/qualcomm/_passes/decompose_linalg_vector_norm.py

Line 37 in 8fbc42c

def __init__(self, quantization_capture=False) -> None:

to differentiate whether the pass is called during transform_for_annotation_pipeline or transform_for_export_pipeline.

To reproduce FP error, I have written a simple pad test:

def test_qnn_backend_pad_mode(self):
        sample_input = (torch.tensor([[[[1, 2, 3],
                               [4, 5, 6],
                               [7, 8, 9]]]], dtype=torch.float),)
        module = PadMode()  # noqa: F405
        self.lower_module_and_test_output(module, sample_input)

class PadMode(torch.nn.Module):
    def __init__(self):
        super().__init__()
 
    def forward(self, x):
        return torch.nn.functional.pad(x, pad=(1, 1, 1, 1), mode='replicate')

[winskuo-quic]

I think the function name is duplicated.

[winskuo-quic]

Could we possibly reuse the copy_meta under
from executorch.backends.qualcomm._passes.utils import copy_meta

[winskuo-quic]

I'm actually getting the UT passed even by commenting the ConvertPadToSliceConcat pass out when running quantized flow.

[winskuo-quic]

Actually, since we are now supporting 2 targets,
aten.constant_pad_nd.default and aten.pad.default,
the schema can be different.
I believe there's a change that the constant value is now stored in node.args[3].
Maybe we can add another condition like:

if len(node.args) >=3 and not isinstance(node.args[2], str):
                pad_value = node.args[2]
elif len(node.args) >= 4 and not isinstance(node.args[3], str):
                pad_value = node.args[3]

[winskuo-quic]

Could we also add aten.pad.default to LayoutTransform pass?

executorch/backends/qualcomm/_passes/layout_transform.py

Line 76 in 8fbc42c

exir_ops.edge.aten.constant_pad_nd.default,