Workshops Archive

Evaluating Function-Level Accuracy and Performance of SIMD Math Libraries with FP32 Inputs


Workshop: 10th International Workshop on Software Correctness for HPC Applications (Correctness '26)

Authors: Xingjian Qi and Robert Strzodka (Heidelberg University)

Abstract: SIMD math libraries are important building blocks for high perfor- mance computer arithmetic. We evaluate unary IEEE 754 binary32 (FP32) AVX2 functions implementations from AOCL-LibM, GNU libmvec, SLEEF, and Intel SVML. We exhaustively test every fi- nite FP32 encoding against a 256-bit MPFR oracle. Most of the functions achieve smaller error than the library-level bounds, but several AOCL and SLEEF functions exceed stated accuracy bounds. For performance evaluation, we identify and test data-dependent performance paths from cheap SIMD execution to expensive fall- backs. Cheap SIMD paths outperform matched scalar calls, but expensive corrections can make vector functions slower than scalar routines. This work fills a gap by providing the community with a per-function, per-input reference for the accuracy and performance of FP32 AVX2 math libraries.


Back to 10th International Workshop on Software Correctness for HPC Applications (Correctness '26) Archive Listing

Back to Full Workshop Archive Listing