Tool Description
The BinFPE tool can be used as an alternative to FPChecker to detect exceptions in NVIDIA GPUs: https://github.com/LLNL/BinFPE. This tool can detect NaN, infinity (positive and negative), and subnormal quantities in NVIDIA GPU applications using binary analysis. As a result, it requires no re-compilation of the application and can analyze libraries. The tool uses the NVBit tool, which is provided by NVIDIA Labs.
How to Use it
Simply preload the shared library detect_fp_exceptions.so before running the application using the Linux LD_PRELOAD variable method:
$ LD_PRELOAD=/path/detect_fp_exceptions.so ./application input
You should see the following output:
#FPCHECKER: Initializing...
#FPCHECKER: kernel void RAJA::internal::CudaKernelLauncherFixed...
#FPCHECKER: kernel...
Error Reports
If a calculation that results in a NaN or infinity is found, you should see the following output:
#FPCHECKER: NaN found @ /tests/my_code/dot_product.cu:18
#FPCHECKER: INF found @ /tests/my_code/mult_matrix.cu:26
Requirements
The GPU code must be compiled with the --generate-line-info flag (see the nvcc options here).
Other requirements:
- SM compute capability: >= 3.5 && <= 8.6
- GCC version: >= 5.3.0
- CUDA version: >= 8.0 && <= 11.x
- nvcc version for tool compilation >= 10.2