internet time @ ???

rnd's website |

cuda debugging notes

So, turns out, when you run a program that uses CUDA, either directly or indirectly (my program runs on CUDA, but uses OpenCL instead), there are issues preventing libasan from working. More specifically, if you run a program with libasan, the CUDA libraries will not work, and the OpenCL ones will not display an NVIDIA platform as available in the first place.

From my experience, seems like there are similar issues preventing Valgrind from working, either.

Also, when a program is being run using mpirun or another such utility, the environmental variable that shows the process’s MPI rank is confusingly named $PMI_RANK. This variable seems to be used by the Intel and Microsoft implementations of MPI.