RecursiveSparseBlocks

Julia interface to librsb, a general purpose sparse matrix library that claims to be broadly competitive with MKL on the subset of functions it implements but is open source and more permissively licensed (GPL3).

Usage

RecursiveSparseBlocks exports a SparseMatrixRSB{T} type which works similarly to SparseMatrixCSC, but with a smaller subset of features. T can be Float32, Float64, Complex64, or Complex128.

The following functionality is implemented:

• Conversion to and from SparseMatrixCSC.
• Element-wise indexing (only non-zero indexes can be set)
• Sparse matrix by dense vector multiplication
• Sparse matrix by dense matrix multiplication
• full, findnz, nnz, size

Obviously, there's more left to implement.

Benchamrks

For large matrices SparseMatrixRSB is much faster that Julia's built-in SparseMatrixCSC (in no small part because the former is multithreaded).

The following plots show elapsed time multiplying random sparse matrices by random dense vectors, each with a density of 1e-4. The first shown Float32 performance, the second Float64. These benchmarks were run on an 8-core cpu.