A Way to Identify Potential Functions for Vectorization for RISC-V

Vectorization is crucial for exploiting the RISC-V Vector Extension to accelerate data-parallel workloads. However, compilers do not always automatically vectorize critical code sections, especially for newer length-agnostic instruction set architectures like RISC-V. In this paper, we propose a methodology to identify functions for vectorization by combining compiler autovectorization analysis with runtime performance profiling. Using the HEVC video encoder as a case study, we compile the code with vectorization diagnostics, and profile its execution on a Banana Pi BPI-F3 board. By correlating compiler reports of which loops were vectorized with perf collected cycle counts for each function, we pinpoint hotspot functions. Impact of RVV extension on performance is highlighted. We discuss how guided interventions improve the performance when using RVV instructions. This approach provides a guideline for performance tuning, and we highlight its generality for other architectures.