CUDA Sparse-Matrix Vector Multiplication using the Sliced Coordinate format (and others in cusp)

prep: to partition the input matrix, store the partitioned matrix in binary for faster access

spmv: perform spmv with prepared files from prep

cusp: SpMV using cusp library matrix formats

./prep matrix=./samples/test.mtx datatype=float
./spmv matrix=./samples/test.mtx datatype=float vector=./samples/vector.txt outfile=/tmp/output 

Required:

• matrix: path to the input matrix file stored in Matrix Market Format
• datatype: float or double

Optional:

• vector: path to the input vector, default is a unit vector
• outfile: output file for spmv, will not output by default
• wdir: working directory (all temporary files and caches are stored here), default is the current directory
• gpu: id of the GPU to execute, default is 0
• times: number of SpMV iterations, default is 1, if more than 1 sortcol should be enabled to ensure the same result at the end
• sortcol: 0|1 set as 1 to permute columns the same as rows (only useful in prep), default is 0
• sortrow: 0|1 set as 1 force sorting rows, default is 0, however will still sort if rows weights differ by a lot (only useful in prep)

If you use this software for a publication, please cite: https://dl.acm.org/citation.cfm?id=2562941

@article{dang2013cuda,
 title={{CUDA}-enabled Sparse Matrix--Vector Multiplication on {GPUs} using atomic operations},
 author={Dang, Hoang-Vu and Schmidt, Bertil},
 journal={Parallel Computing},
 volume={39},
 number={11},
 pages={737--750},
 year={2013},
 publisher={Elsevier}
}