Should graphblas.matrix_select simply be graphblas.select?

[paul-tqh-nguyen]

Should graphblas.matrix_select simply be graphblas.select?

It would make sense that the verifier would enforce that tril and triu only work with matrices, but for a selector like gt, it would make sense for this selector to work with vectors.

Also, should we merge graphblas.matrix_select_random into graphblas.matrix_select? Or is it sufficiently different? If it is sufficiently different, then what makes (graphblas.matrix_select+gt) not sufficiently different than (graphblas.matrix_select+tril)?

[seibert]

Yes, I think it would be good to remove terms like matrix and vector from ops whenever the semantics are clear, especially since we know the types of the inputs at code generation type and can switch between vector and matrix specific implementations very easily.

select_random is very different than select, though, because the selection condition is not independent for every element. (Something like gt0 or tril can be evaluated in parallel on every element if you want to, for example.) In fact, I think select_random is not a good name for it, but I'm not sure what to call it. It is more of a select_random_from_rows, so it can process each row in a CSR independently, but within the row the selection is not independent (you want exactly N elements, with no duplication).

[jim22k]

In grblas, we have:

• select (for vector or matrix)
• selectk (for vector, similar to select_random)
• selectk_rowwise (for matrix, same as the current matrix_select_random)
• selectk_colwise (for matrix, does the same but along columns)

I think we should do something similar in mlir-graphblas. We should also group our selectors based on whether they compare:

• rows & cols (only applies to matrix)
• index (only applies to vector)
• value (applies to both matrix and vector)

Also, I would like to propose that all selectors take a thunk. Even something like triu can take a thunk indicating how far off-diagonal the selection is limited to. This would make it much easier to handle multiple selectors in the dialect signature.