SpMV-Samples

A collection of sample implementations for Sparse Matrix-Vector Multiplication (SpMV) using the Compressed Sparse Row (CSR) sparse matrix format on GPU.

Current Supported SpMV Implementations:

• cuSparse SpMV
• CUSP CSR-Vector SpMV (supporting warp reduce)
• LightSpMV
• Merge-based SpMV in CUB (including a generalized SpMV implementation based on this method)

Requirements

GCC (version 9.4.0 and above)
CUDA (version 11.8 and above)
CMake (version 3.12 and above)

Setup

Execute the following instructions in the root directory of the project.

mkdir build && cd build
cmake .. 
make

Dataset

Support for Matrix Market format graph datasets.
You can download graph datasets from the website https://sparse.tamu.edu/.

Add Your SpMV

You can add your own CSR sparse matrix format SpMV implementation.
Just encapsulate it with the following interface:

// interface
template <typename index_t, typename offset_t, typename mat_value_t,
          typename vec_x_value_t, typename vec_y_value_t>
void SpMV(const std::string& kind_str,
    index_t n_rows,  index_t n_cols, offset_t nnz,
    const offset_t *Ap, const index_t *Aj, const mat_value_t *Ax, 
    const vec_x_value_t *x, vec_y_value_t *y); 

And define the corresponding label strings and function definitions in the spmv.h header file.

/// SPMV kind strings and its function
#define SPMV_KINDS                       \
    X("cusparse", SpMV_cusparse)         \
    X("YOUR_SPMV_LABEL", YOUR_SPMV_FUNC)