Expand description
Sampled Dense-Dense Matrix Multiply (SDDMM).
Computes C_ij = alpha * (A @ B)_ij * spy(S)_ij + beta * S_ij
where the result is only computed at positions where the sparse matrix S
has non-zero entries. This is a key primitive in graph neural networks and
sparse attention mechanisms.
§Strategy
Each thread handles one non-zero entry of S. For that entry at position
(row, col), the thread computes the dot product A[row, :] . B[:, col]
(inner dimension K), scales by alpha, and blends with the existing
value via beta.
Functions§
- sddmm
- Sampled Dense-Dense Matrix Multiply.