RFE: cusparse mixed integer indexing with cusparseSpMV

I have a matrix with a small number of rows but where the numerical values of each element can be very large (and blow past 32 bit integer limitations). On the other hand, I have a lot of columns but the numerical values of the element are small.

In other words, the memory requirement of my matrix is driven by the columns, but the largest numerical values are for the rows. I would like to use 64 bits for the row indices but keep the column indices 32 bits to keep memory requirements down. This is not currently possible with cusparseSpMV.


How many iterations will you be running SpMV (i.e., are you running it 1000’s or just a single time)?

Hundreds to thousands with fixed matrix indices and values and varying vector values.

Okay. We will discuss at our next roadmap meeting. If you don’t hear back in a few weeks, ping this chat.