Accurate CUDA performance modeling for sparse matrix-vector multiplication

This paper presents an integrated analytical and profile-based CUDA performance modeling approach to accurately predict the kernel execution times of sparse matrix-vector multiplication for CSR, ELL, COO, and HYB SpMV CUDA kernels. Based on our experiments conducted on a collection of 8 widely-used testing matrices on NVIDIA Tesla C2050, the execution times predicted by our model match the measured execution times of NVIDIA´s SpMV implementations very well. Specifically, for 29 out of 32 test cases, the performance differences are under or around 7%. For the rest 3 test cases, the differences are between 8% and 10%. For CSR, ELL, COO, and HYB SpMV kernels, the differences are 4.2%, 5.2%, 1.0%, and 5.7% on the average, respectively.