RVVLASparse is an open-source, high-performance C++ library designed for efficient sparse matrix operations on RISC-V processors. The library utilizes RVV 1.0 intrinsics and OpenMP threading.
This repository accompanies the research paper: "Rethinking Sparse Formats for RISC-V: A Hierarchical Approach to High-Performance SpMV".
- Sparse matrix-dense vector multiplication
- HCSR (Hierarchical CSR — our modification of CSR)
- CSR
- SELL-C-σ
- CSR5
- LAV
- VNEC
- VHCC
- CVR
The library has been evaluated on SpacemiT K1 and K3 RISC-V boards. Benchmarks on 8 cores demonstrate that the HCSR format provides a significant speedup over the scalar CSR baseline:
- Double Precision: An average speedup of 19% on K1 and 13% on K3.
- Single Precision: An average speedup of 36% on K1 and 34% on K3.