Workshops Archive

TSVC-RVV: A Diagnostic Extension of TSVC for Evaluating RISC-V Auto-Vectorization


Workshop: International Workshop on RISC-V for HPC (RISCVHPC)

Authors: Hung-Ming Lai (National Tsing Hua University, Lawrence Livermore National Laboratory (LLNL)); Hsuan-Yu Lin, Wei-Chun Wang, Yu-Yi Lee, and Jenq-Kuen Lee (National Tsing Hua University); and Pei-Hung Lin and Maya B. Gokhale (Lawrence Livermore National Laboratory)

Abstract: RISC-V is attracting growing interest in high-performance computing (HPC), with the RISC-V Vector Extension (RVV) providing a vector-length-agnostic architecture for accelerating data-parallel workloads. Realizing this potential requires compiler toolchains that reliably generate efficient RVV code. Assessing and improving RVV auto-vectorization is therefore important to the RISC-V HPC software ecosystem.

The Test Suite for Vectorizing Compilers (TSVC) is widely used to evaluate auto-vectorization. It contains 151 loops, each designed to exercise a specific vectorization pattern. However, effective TSVC evaluation requires substantial manual effort, as the suite provides no reference vector implementations and each loop may admit multiple valid vectorization strategies.

We extend TSVC with two complementary RVV-focused references: (1) RVV intrinsic implementations for all 151 loops, providing vectorization strategies; and (2) source-code variants that expose limitations in RISC-V compiler toolchains. Together, these references enable more informative TSVC evaluation and provide actionable guidance for improving RVV support for HPC workloads.


Back to International Workshop on RISC-V for HPC (RISCVHPC) Archive Listing

Back to Full Workshop Archive Listing