FPChecker: Floating-Point Exception Detection Tool and Benchmark for Parallel and Distributed HPC

Floating-point arithmetic is fundamental to many areas including high-performance computing and machine learning. In order to ensure the numerical integrity of the overall computation, numerical exceptions (such as NaNs) must be detected and suitably reported to the user. Unfortunately, today’s best available methods and tools are not general-enough. Moreover, there are no comprehensive benchmarks today that can be compiled to intermediate representations such as LLVM and analyzed at that level. In this paper, we contribute the first such benchmark suite that spans five HPC proxy applications plus eight public benchmarks that run on CPU multicores under MPI, OpenMP, and performance portability layers. We also release our tool (also called LLFPX) that is up to 7.9× faster on many important benchmarks and overall more comprehensive with respect to exception variety coverage. This paper presents the tool, its design, and evaluation on our benchmarks, with the tool, benchmarks, and a facility to examine results on the web available for public use upon acceptance. Other result highlights include the effect of compiler optimizations on exceptions.