Add lowering of aten.Int.Tensor op.

[pashu123]

The lowering of aten.Int.Tensor op has been added.
The changes has been made as a part of convert-torch-to-linalg pass.

[pashu123]

Hi @ALL, I have tried to come up with the lowering seeing pytorch/TensorRT#513. I was not able to find the respective torch python function that can emit aten.Int.Tensor op.

[ramiro050]

I was not able to find the respective torch python function that can emit aten.Int.Tensor op.

The python function that generates the aten.Int.Tensor op is int(..). If you do something like

def forward(self, x):
    return int(x)

where x is a torch tensor, you get the following MLIR:

func private @__torch__.MyModule.forward(%arg0: !torch.nn.Module<"__torch__.MyModule">, %arg1: !torch.tensor {torch.type_bound = !torch.vtensor<[],f32>}) -> !torch.int {
    %1 = torch.operator "aten.Int.Tensor"(%arg1) : (!torch.tensor) -> !torch.int
    return %1 : !torch.int
  }

Although, I don't know if the e2e_testing currently supports returning ints. You could re-wrap the value in a tensor again before returning, if this is an issue.

[pashu123]

I was not able to find the respective torch python function that can emit aten.Int.Tensor op.

The python function that generates the aten.Int.Tensor op is int(..). If you do something like

def forward(self, x):
    return int(x)

where x is a torch tensor, you get the following MLIR:

func private @__torch__.MyModule.forward(%arg0: !torch.nn.Module<"__torch__.MyModule">, %arg1: !torch.tensor {torch.type_bound = !torch.vtensor<[],f32>}) -> !torch.int {
    %1 = torch.operator "aten.Int.Tensor"(%arg1) : (!torch.tensor) -> !torch.int
    return %1 : !torch.int
  }

Although, I don't know if the e2e_testing currently supports returning ints. You could re-wrap the value in a tensor again before returning, if this is an issue.

Hey, thanks for this. Sure, let me check.

[pashu123]

I was not able to find the respective torch python function that can emit aten.Int.Tensor op.

The python function that generates the aten.Int.Tensor op is int(..). If you do something like

def forward(self, x):
    return int(x)

where x is a torch tensor, you get the following MLIR:

func private @__torch__.MyModule.forward(%arg0: !torch.nn.Module<"__torch__.MyModule">, %arg1: !torch.tensor {torch.type_bound = !torch.vtensor<[],f32>}) -> !torch.int {
    %1 = torch.operator "aten.Int.Tensor"(%arg1) : (!torch.tensor) -> !torch.int
    return %1 : !torch.int
  }

Although, I don't know if the e2e_testing currently supports returning ints. You could re-wrap the value in a tensor again before returning if this is an issue.

Yes, this seems to be an issue. The return type should be Memref type.

        Lowering Linalg-on-Tensors IR to LLVM with RefBackend failed with the following diagnostics:                                                                                  
        error: return value must be memref type                                                
        note: see current operation: "std.return"(%1) : (i64) -> () 

I can rewrap the integer into a tensor and return, but there is another op that pops up "torch.aten.tensor.int"(%8, %6, %2, %5) : (!torch.int, !torch.int, !torch.none, !torch.bool) -> !torch.vtensor<[1],si64> and its lowering is missing.

[ramiro050]

this is pretty clever :)

[pashu123]

@ramiro050 All thanks to @cathyzhyi.

[cathyzhyi]

It was originally @silvasean 's idea!