TaiChi is a hybrid format for binary sparse matrix and fully exploits the data distribution of non-zero elements. Input matrices are firstly partitioned into relative dense and ultra-sparse areas, then the dense areas are encoded inversely by marking “0”s, while the ultra-sparse area is encoded using popular CSR5 format by marking “1”s. We also design a new SpMV algorithm just using addition and subtraction for binary matrices based on our partition and encoding format. We evaluate our SpMV algorithm using some real-world binary sparse matrices from the SuiteSparse Matrix Collection. Evaluation results show that the speedup of CSR5-NV-based SpMV to CSR5-based SpMV is up to 3.26x on GTX 1080 Ti and 2.20x on Tesla V100-SXM2, and our hybrid encoding for binary matrix significantly compresses the original matrix and obtains the highest speedup of 6.89x and 4.21x on GTX 1080 Ti and Tesla V100-SXM2 respectively.
We run the script https://github.com/SC-Tech-Program/Author-Kit/blob/master/collect_environment.sh and collect the output to the file environment_info.txt, including OS, compilers, CPU, GPU, memory, disk, and so on.
- change the CUDA path in Makefile
- change the ARCH: 61 in GTX 1080 Ti, 70 in Telsa V100-SXM2.
make./spmv matrixPath sdfPath- matrixPath is the directory of binary sparse matrices, which require to be downloaded by yourself from the SuiteSparse Matrix Collection. We uploaded some example matrices to
dataset/matrixPath/. - unzip the
dataset/sdfFile.zip, and sdfPath is the directory of sdf file:./dataset/sdfFile/
- matrixPath is the directory of binary sparse matrices, which require to be downloaded by yourself from the SuiteSparse Matrix Collection. We uploaded some example matrices to
The timing codes of different parts are included in the program, the total SpMV time should be equal to the sum of time_add_and_sub, CSR5_transfer, CSR5-based SpMV time, and time_gather.