Workshop: 10th International Workshop on Software Correctness for HPC Applications (Correctness '26)
Authors: Alishba Gul and Sreepathi Pai (University of Rochester)
Abstract: Floating-point exceptions, including NaN, infinity, and division by zero, present significant challenges to GPU-accelerated programs. These exceptions can silently propagate in the programs, leading to inaccurate results that might go unnoticed. Existing tools like FPChecker and GPU-FPX detect these exceptions dynamically but impose significant run-time overhead, making them impractical for large-scale GPU applications. We propose a static analysis framework that allows selective instrumentation of only high-risk floating-point instructions in GPU programs. We achieve this selective instrumentation in two ways: first, by analyzing SASS files on a per-instruction basis, comparing IEEE 754 bit pattern exception probabilities with thresholds; second, by comparing propagated exception probabilities within basic blocks while traversing the data flow graphs. We evaluate our framework by testing 19 programs across four CUDA benchmark suites and demonstrate that our approach can eliminate certain checks with at least 90% exception detection coverage.
Back to 10th International Workshop on Software Correctness for HPC Applications (Correctness '26) Archive Listing
Back to Full Workshop Archive Listing