October 27, 2023
Conference Paper
Design and Evaluation of GPU-FPX: A Low-Overhead tool for Floating-Point Exception Detection in NVIDIA GPUs
Abstract
Despite the growing uptake of GPUs, the fact that they do not support floating-point exceptions in hardware is a significant concern. In this work, we develop a technique to efficiently mine these exceptions from GPU programs through a novel tool, GPU-FPX. In our unique study of 115 programs, we discovered 22 unhandled exceptions including NaN, Infinity, and divide-by-0. We also investigate how these exceptions flow through the application logic, how exceptions are affected by compiler optimizations, and also result-reproducibility. GPU-FPX uses a binary-instrumentation method that allows it to collect and process exceptions on the GPU even from libraries without sources. Our work paves the way to efficient floating-point exception checking essential for trustworthy accelerated applications.Published: October 27, 2023