MLIR, TOSA: Zero-size tensor crashes tosa-to-tensor

[mgehre-amd]

func.func @test_reshape_downrank_zero(%arg0: tensor<6x1x0xf32>) -> tensor<6x0xf32> {
  %0 = "tosa.reshape"(%arg0) {new_shape = array<i64: 6, 0>} : (tensor<6x1x0xf32>) -> tensor<6x0xf32>
  return %0 : tensor<6x0xf32>
}

run with
mlir-opt test.mlir --tosa-to-tensor
crashes in createReassociationMapsForCollapse due to

llvm-project/mlir/lib/Conversion/TosaToTensor/TosaToTensor.cpp

Line 107 in 9bccb5b

srcSize *= srcShape[currSrcDim];

being out of bounds.

I'm not sure how the correct tensor IR would look like for this case.
Tensors where a dimension have zero size seems to be allowed by the TOSA spec.

[eric-k256]

The TOSA spec requires each dimension to have size of at least 1: https://www.mlplatform.org/tosa/tosa_spec.html#_tensors

The tensor shape in each dimension must be greater than or equal to 1.

If you don't have that, you end up with a number of cases like this occurring. I'm not sure if we need something additional in the dialect to prevent zero size dimensions.

[mgehre-amd]

I guess I was mislead by the sentence The size of tensors along each of their dimensions is limited by the data type index_t. This type must be able to hold integers in the range 0 to, which I read as "size of a dimension can be 0 to ...".

Then the right way to fix this crash should be implementing proper op verifiers for TOSA.

[eric-k256]

Yes, I can see how that sentence would be confusing. I'll try to reword it to make it clear that 0 isn't a valid value. The index_t could hold a 0 as it is effectively an unsigned int of at least a certain size based on MAX_LOG2_SIZE, but 0 is reserved.

[Lewuathe]

I have modified verification to check the invalid dimension for TOSA op.

https://reviews.llvm.org/D154569

[rikhuijzer]

This was fixed in b76d8f7. Just to double check, this is current main:

$ mlir-opt temp.mlir --tosa-to-tensor
temp.mlir:2:8: error: 'tosa.reshape' op tensor has a dimension with size zero. Each dimension of a tensor must have size >= 1
  %0 = "tosa.reshape"(%arg0) {new_shape = array<i64: 6, 0>} : (tensor<6x1x0xf32>) -> tensor<6x0xf32>
       ^
temp.mlir:2:8: note: see current operation: %0 = "tosa.reshape"(%arg0) <{new_shape = array<i64: 6, 0>}> : (tensor<6x1x0xf32>) -> tensor<6x0xf32>

As expected, the verifier throws an error. I'll close this issue.