call* functions in GraphBLASUtils.cpp cause errors when called on fixed-size tensors of different size

[paul-tqh-nguyen]

This (let's call it Example A) lowers through graphblas-opt --graphblas-lower with no problem:

#CSR64 = #sparse_tensor.encoding<{
  dimLevelType = [ "dense", "compressed" ],
  dimOrdering = affine_map<(i,j) -> (i,j)>,
  pointerBitWidth = 64,
  indexBitWidth = 64
}>

module {

    func @matrix_select_triu_100(%sparse_tensor: tensor<100x100xf64, #CSR64>) -> tensor<100x100xf64, #CSR64> {
        %answer = graphblas.matrix_select %sparse_tensor { selectors = ["triu"] } : tensor<100x100xf64, #CSR64> to tensor<100x100xf64, #CSR64>
        return %answer : tensor<100x100xf64, #CSR64>
    }

}

This (let's call it Example B) also lowers through graphblas-opt --graphblas-lower with no problem:

#CSR64 = #sparse_tensor.encoding<{
  dimLevelType = [ "dense", "compressed" ],
  dimOrdering = affine_map<(i,j) -> (i,j)>,
  pointerBitWidth = 64,
  indexBitWidth = 64
}>

module {

    func @matrix_select_triu_200(%sparse_tensor: tensor<200x200xf64, #CSR64>) -> tensor<200x200xf64, #CSR64> {
        %answer = graphblas.matrix_select %sparse_tensor { selectors = ["triu"] } : tensor<200x200xf64, #CSR64> to tensor<200x200xf64, #CSR64>
        return %answer : tensor<200x200xf64, #CSR64>
    }

}

This (let's call it Example C) does not:

#CSR64 = #sparse_tensor.encoding<{
  dimLevelType = [ "dense", "compressed" ],
  dimOrdering = affine_map<(i,j) -> (i,j)>,
  pointerBitWidth = 64,
  indexBitWidth = 64
}>

module {

    func @matrix_select_triu_100(%sparse_tensor: tensor<100x100xf64, #CSR64>) -> tensor<100x100xf64, #CSR64> {
        %answer = graphblas.matrix_select %sparse_tensor { selectors = ["triu"] } : tensor<100x100xf64, #CSR64> to tensor<100x100xf64, #CSR64>
        return %answer : tensor<100x100xf64, #CSR64>
    }

    func @matrix_select_triu_200(%sparse_tensor: tensor<200x200xf64, #CSR64>) -> tensor<200x200xf64, #CSR64> {
        %answer = graphblas.matrix_select %sparse_tensor { selectors = ["triu"] } : tensor<200x200xf64, #CSR64> to tensor<200x200xf64, #CSR64>
        return %answer : tensor<200x200xf64, #CSR64>
    }

}

This problem stems from the following:

• Example A declares this function: builtin.func private @dup_matrix(tensor<100x100xf64, #CSX64>) -> tensor<100x100xf64, #CSX64>
• Example B declares this function: builtin.func private @dup_matrix(tensor<200x200xf64, #CSX64>) -> tensor<200x200xf64, #CSX64>
• Example C attempts to declare and use both of the above, which causes problems.

A valid (partial) solution would be to:

• Always declare: builtin.func private @dup_matrix(tensor<?x?xf64, #CSX64>) -> tensor<?x?xf64, #CSX64>
• When a call* function from GraphBLASUtils.cpp is used in the lowering code, the function should insert tensor.cast before and after the MLIR assembly that calls dup_matrix (or whatever function name is relevant) where dup_matrix takes a tensor<?x?xf64, #CSX64>.

This is only a partial solution since this same problem might occur if we tried use callDupTensor to duplicate an f64 tensor and i64 tensor in the same MLIR module or if we tried to use callDupTensor to duplicate a CSR tensor and CSC tensor in the same MLIR module.

[jim22k]

This should no longer be a problem because we always declare top-level C calls using !llvm.ptr<i8>. That change (PR#166) was done specifically to handle the case of multiple dtypes (f64 vs i64) and different rank (?x64 vs ?x?xf64).

I haven't checked if it works with fixed dimensions, but there's no reason it shouldn't.