XNN Conv2d won't lower with single stride/dilation/etc.

[GregoryComer]

🐛 Describe the bug

When trying to lower a Conv2d operator with a single-element stride array, the Conv1dUsqueezePass errors out. This is because it is using stride to determine if a convolution is 1d (there is a single core op for all convs). From a quick look at the op set, it looks like we should maybe use weight rank instead.

In the conv2d node visitor logic, it also assumes that stride/dilation/etc. have length equal to the conv dimension, which isn't always the case. Using the repro below and adding dilation=[2] to the Conv2d module will cause lowering to fail in the node visitor.

Repro:

import torch
from executorch.backends.xnnpack.partition.xnnpack_partitioner import XnnpackPartitioner
from executorch.exir import to_edge_transform_and_lower

class Model(torch.nn.Module):
    def __init__(self):
        super().__init__()
        self.conv = torch.nn.Conv2d(3, 3, 5, [2])
    
    def forward(self, x):
        return self.conv(x)

inputs = (
    torch.randn(1, 3, 16, 16),
)

ep = torch.export.export(Model(), inputs)
et_program = to_edge_transform_and_lower(
    ep,  
    partitioner=[XnnpackPartitioner()]
).to_executorch()

Output:

File ~/miniconda3/envs/pytorch/lib/python3.11/site-packages/executorch/backends/xnnpack/_passes/conv1d_unsqueeze_pass.py:191, in Conv1dUnsqueezePass.call(self, graph_module)
    189 # Since we are overriding "call", we need to call the parent's "call"
    190 # to retrace the graph and regenerate metadata
--> 191 graph_module = super().call(graph_module).graph_module
    193 return PassResult(graph_module, True)
...
File ~/miniconda3/envs/pytorch/lib/python3.11/site-packages/torch/_subclasses/fake_impls.py:762, in conv(fake_mode, func, *args, **kwargs)
    761 if func is aten.convolution.default:
--> 762     conv_backend = torch._C._select_conv_backend(**kwargs)
    763 else:

RuntimeError: expected stride to be a single integer value or a list of 3 values to match the convolution dimensions, but got stride=[2, 1]

...

Exception: An error occurred when running the 'Conv1dUnsqueezePass' pass after the following passes: []

Versions

N/A

cc @digantdesai @mcr229 @cbilgin