rvv-evaluation

RVV-Eval is a benchmark suite for studying performance portability on RISC-V Vector (RVV) processors. It contains three benchmark groups:

1. Assembly microbenchmarks for instruction-level throughput studies.
2. Proxy benchmarks for application kernels and larger benchmark codes.
3. Applications for end-to-end workloads ported to or evaluated on RVV.

Group 1: Assembly Microbenchmarks

The assembly benchmarks generate small C++ kernels with inline assembly and measure low-level scalar/RVV instruction behavior.

• Arithmetic instructions: assembly_microbench/arith_instr/README.md
• Load/store instructions: assembly_microbench/ldst_instr/README.md

Group 2: Proxy Benchmarks

The proxy benchmarks compare non-vectorized, auto-vectorized, and LMUL-controlled RVV builds for larger kernels and applications.

• Darknet: proxy_bench/darknet/README.rvv.md
• GEMM: proxy_bench/gemm/README.rvv.md
• SPMV-v2: proxy_bench/spmv-v2/README.rvv.md

Group 3: Applications

The application benchmarks contain larger end-to-end workloads evaluated or ported for RVV.

• qsim_rvv: apps_bench/qsim_rvv/README.md

Citation

Please cite:

@misc{shi2026closergapportableperformance,
      title={Closer in the Gap: Towards Portable Performance on RISC-V Vector Processors}, 
      author={Ruimin Shi and Maya Gokhale and Pei-Hung Lin and Xavier Teruel and Ivy Peng},
      year={2026},
      eprint={2605.10860},
      archivePrefix={arXiv},
      primaryClass={cs.DC},
      url={https://arxiv.org/abs/2605.10860}, 
}

Contact

Ruimin Shi: ruimins@kth.se.