Timing module to be used in codes which use MPI and possibly CUDA/OpenACC using NVTX markers.
- Reports the average time per task for each tagged region by default, but more detailed reporting (minimum/maximum per call and/or per task) is supported too.
- Times tagged code regions a la MATLAB (timer_tic and timer_toc) using MPI_WTIME().
- A few different options for tagging the NVTX markers are available: prescribed color, random colored, or round-robin colormap sampling.
- Supports host-device (a)synchrony for applications accelerated with OpenACC or CUDA (optionally using CUDA streams/OpenACC queues).
- See the example program main.f90 illustrating the usage, and the Makefile illustrating compiling instructions.
The tool took inspiration from wcdawn/ftime, and adapted the NVTX bindings in maxcuda/NVTX_Example.