Support for GPU sparse matrix-matrix multiply

[bpasley]

It seems matrix multiply is not implemented for GSMat and GSMat. Are there any plans to support this (e.g., cuSPARSE's csrGEMM)?

[jcanny]

There are a few issues with making this work: First is to decide what form of output you want (dense or sparse). Sparse is probably easy (but I'm not sure that's what you need), because most sparse-sparse operators return a sparse result. Then there's wrapping the CUDA routine, adding a C native code equivalent, adding operator logic for conversion etc. We havent really needed it, so we've avoided biulding it. What's your use case?

[bpasley]

the main use case is building large (sparse) similarity matrices from sparse features (e.g., tfidf vectors). This works fine using the CUDA routine directly. of course it's more convenient to have that process integrated directly into BidMat but not strictly necessary