[Tracking Issue] Dense-sparse Unstructured Matrix Multiplication (Column-major SpMM)

[yzh119]

The network pruning requires dense-sparse matrix multiplication, which is equivalent to column-major SpMM.

This setting is more challenging for utilizing Tensor Cores for unstructured Sparse Matrix, because the memory access to the dense-matrix would not be coalesced. One solution is to pre-load the tiles of dense-matrix into shared mem.

This issue depends on #52 .