Best way to do a sparse matrix update?

I have two relatively big matrices in dense format (size 100,000 * 2000), the second one being filled by zeros, except for a few hundred rows which are full (and I know which ones), and I need to add the second matrix to the first one in an efficient way.

I had a look at cusparse, but I feel that it would be more convenient to just do a loop of standard cublas rather than to bother changing the way I store the matrices to then do the updates.

What is the recommended solution in that case?

what would be the basic operation to sum matrices in dense format:

a) find all unique entries in both matrices and write them away,
b) sum all overlapping entries, and write the result away


if the above is true, and if you then know which rows are non-zero, both a) and b) would be simplified, i would think