Conv2d output shape mismatch

[azecevicTT]

Part of the generated MLIR from the model:

module {
  func.func @main(%arg0: tensor<1x224x224x3xf32>, %arg1: tensor<32x3x3x3xf32>) -> (tensor<1x111x112x32xf32>) {
    %0 = tensor.empty() : tensor<1x111x112x32xf32>
    %1 = "ttir.conv2d"(%arg0, %arg1, %0) <{dilation_height = 1 : si32, dilation_width = 1 : si32, groups = 1 : si32, padding_bottom = 1 : si32, padding_left = 0 : si32, padding_right = 1 : si32, padding_top = 0 : si32, stride_height = 2 : si32, stride_width = 2 : si32}> {channel_last = 1 : si32} : (tensor<1x224x224x3xf32>, tensor<32x3x3x3xf32>, tensor<1x111x112x32xf32>) -> tensor<1x111x112x32xf32>
    return %1 : tensor<1x111x112x32xf32>
  }
}

It should be checked if the output shape should be 1x111x112x32 instead of 1x112x112x32 for these arguments.

[meenakshiramanathan1]

@azecevicTT , can I get repro steps for this?

[azecevicTT]

I believe @nvukobratTT have repro steps for this, this is just a part of the MLIR that I got that seems like an error on FE side.

[nvukobratTT]

@meenakshiramanathan1 one of the MobileNet is problematic (e.g. v1 196). Feel free to checkout skipped test with this msg:

pytest.skip("Hitting segmentation fault in MLIR")

[meenakshiramanathan1]

@azecevicTT Tried reproducing the conv2d layer with required attributes , below given mlir module and the one given in description is same except for output shape for ttir.conv2d module

    %5 = "ttir.conv2d"(%3, %arg1, %arg2, %4) <{dilation_height = 1 : si32, dilation_width = 1 : si32, groups = 1 : si32, padding_bottom = 1 : si32, padding_left = 0 : si32, padding_right = 1 : si32, padding_top = 0 : si32, stride_height = 2 : si32, stride_width = 2 : si32}> {channel_last = 1 : si32} : (tensor<1x224x224x3xf32>, tensor<32x3x3x3xf32>, tensor<1x32x1x1xf32>, 

tensor<1x112x112x32xf32>) -> tensor<1x112x112x32xf32>

The output shape of this generated MLIR module is 1x112x112x32

Reproduce

git checkout mramanathan/conv2d
pytest forge/test/models/pytorch/vision/test_conv2d.py