This is a clone of Merrill and Garland's Merge-based Parallel Sparse Matrix-Vector Multiplication. We integrated our CSRLenGoto method as the kernel SpMV method to their code. Once the matrix is partitioned using the merge-path method, we convert the CSR data to CSRLen, and then use the CSRLenGoto SpMV function on each partition. This way, CSRLenGoto method is parallelized.
The starting point where our code differentiates from the merge-based SpMV code base is this: https://github.com/aktemur/CSRLenGoto/blob/master/cpu_spmv.cpp#L529
For a version of CSRLenGoto where we used runtime code generation to produce the SpMV code dynamically, see Thundercat.
Type make cpu_spmv to compile. This will produce, compile, and link the CSRLenGoto
assembly file, csrlengoto.s. If you're only interested in generating this file, type
make csrlengoto.s.
Currently, the generated file will work for matrices
whose max row length is smaller than 25. To handle matrices with larger
max row lengths, change the line below to e.g. BODY_50K.
https://github.com/aktemur/CSRLenGoto/blob/master/csrlengoto-body-gen.s#L22: