Sparse blas matrix multiplication - speed vs memory

[joao-araujo]

I was doing some comparison tests and noticed that a general matrix multiplication function is faster than a sparse blas specific function (cblas_?gemm and mkl_sparse_?_mm). The objective of sparse blas' matrix multiplication functions, at least, is to be less expensive in terms of memory, but not necessarily in terms of execution time, am I right?

[mecej4]

As an extreme counterexample, consider the outer product of two n-vectors. You may use a general matrix multiply routine to perform this operation, although other routines may be better suited. The general matrix routine need not use more temporary storage than that needed for a single real/double.