Help users avoid inefficient use of COOMatrix

[mmelnich]

A proper explanation is needed as to why the users are to convert the COO sparse matrices into CSC or CSR.

[rileyjmurray]

Giving a short answer for now. The dev notes for our sparse matrix functionality, here, explain that COO multiplication is handled in coo_spmm_impl.hh. Looking at that file we see a conversion to CSC (although without the CSCMatrix abstraction):

RandBLAS/RandBLAS/sparse_data/coo_spmm_impl.hh

Lines 98 to 122 in 1888089

	// Step 1: reduce to the case of CSC sort order.
	if (A0.sort != NonzeroSort::CSC) {
	auto orig_sort = A0.sort;
	sort_coo_data(NonzeroSort::CSC, A0);
	apply_coo_left_jki_p11(alpha, layout_B, layout_C, d, n, m, A0, ro_a, co_a, B, ldb, C, ldc);
	sort_coo_data(orig_sort, A0);
	return;
	}
	// Step 2: make a CSC-sort-order COOMatrix that represents the desired submatrix of A.
	// While we're at it, reduce to the case when alpha = 1.0 by scaling the values
	// of the matrix we just created.
	int64_t A_nnz;
	int64_t A0_nnz = A0.nnz;
	std::vector<sint_t> A_rows(A0_nnz, 0);
	std::vector<sint_t> A_colptr(std::max(A0_nnz, m + 1), 0);
	std::vector<T> A_vals(A0_nnz, 0.0);
	A_nnz = set_filtered_coo(
	A0.vals, A0.rows, A0.cols, A0.nnz,
	co_a, co_a + m,
	ro_a, ro_a + d,
	A_vals.data(), A_rows.data(), A_colptr.data()
	);
	blas::scal<T>(A_nnz, alpha, A_vals.data(), 1);
	sorted_nonzero_locations_to_pointer_array(A_nnz, A_colptr.data(), m);

That conversion is very expensive. Realistically, we should both make apply_coo_left_jki_p11 more efficient and add warnings to the documentation that one should use CSR or CSC format if you need to multiply with a sparse matrix more than once.

[rileyjmurray]

We could also raise a warning if a COOMatrix is passed to an spmm function in a way that asks for applying the entire matrix, rather than a submatrix.