llvm
[RFC] Implementing aten::conv_transpose1d/2d/3d
Next op I'm going to work on is the aten::conv_transpose*d family. I figure decomposing them to aten::convolution and implementing the transposed case for that op is going to be the best way forward. From my research, it looks like padding and striding the input and using a normal linalg convolution is going to be the most efficient way to implement that.
An issue I'm running into with the initial prep work:
RuntimeError:
Expected a default value of type List[int] on parameter "stride".:
File ".../mlir-npcomp/build/tools/torch-mlir/python_packages/torch_mlir/torch_mlir/dialects/torch/importer/jit_ir/build_tools/shape_lib_gen.py", line 778
def aten〇conv_transpose1d(input: List[int], weight: List[int], bias: Optional[List[int]] = None, stride: List[int] = (1), padding: List[int] = (0), output_padding: List[int] = (0), groups: i
nt = 1, dilation: List[int] = (1)) -> List[int]:
Changing the default values to list(1) leads to shape lib complaining that the signature doesn't match.
Not sure if this was fixed already, but I think the problem is due to the fact that in Python a single element tuple needs to have a comma after the element. For example: (1,). You could also just do [1], since the expected type is a list.
