Fix onnx.Pad constant

[josel-amd]

Looking at the GeneratedTorchOps.td. This leaves two options, either the operator specification is wrong or the code generating it.

I check a bit and it seems that indeed the op spec is right. References:

https://github.com/pytorch/pytorch/blob/543a870943120484db547382ed9ca9538a40f284/torch/csrc/api/include/torch/nn/functional/padding.h#L12

https://pytorch.org/docs/main/generated/torch.nn.functional.pad.html#torch.nn.functional.pad

def Torch_AtenPadOp : Torch_Op<"aten.pad", [
    AllowsTypeRefinement,
    HasValueSemantics,
    ReadOnly
  ]> {
  let summary = "Generated op for `aten::pad : (Tensor, int[], str, float?) -> (Tensor)`";
  let arguments = (ins
    AnyTorchTensorType:$self,
    AnyTorchListOfTorchIntType:$pad,
    Torch_StringType:$mode,
    AnyTorchOptionalFloatType:$value
  );
  let results = (outs
    AnyTorchTensorType:$result
  );
  let hasCustomAssemblyFormat = 1;
  let extraClassDefinition = [{
    ParseResult AtenPadOp::parse(OpAsmParser &parser, OperationState &result) {
      return parseDefaultTorchOp(parser, result, 4, 1);
    }
    void AtenPadOp::print(OpAsmPrinter &printer) {
      printDefaultTorchOp(printer, *this, 4, 1);
    }
  }];
}

[cferry-AMD]

To be sure I understood correctly, is the support of other data types than f32 what is wrong in TorchOnnxToTorch? Is there any test case where this appears?

[josel-amd]

To be sure I understood correctly, is the support of other data types than f32 what is wrong in TorchOnnxToTorch? Is there any test case where this appears?

Yes, only f32 is allowed in that specific argument. I don't think there is a test case for it. At least nothing started failing. I can add one, if there is general agreement about the fix.

[cferry-AMD]

Good, indeed if we look closely at what happens above, any constant scalar passed is turned into a float. Thanks!