I've got a complex Cuda C++ application that runs fine until I build the cuda code -g -G
. When built debug, it crashes with cudaErrorIllegalAddress
, so I ran cuda-memcheck
, but that finds no issues. I also ran valgrind
and it found no issues.
When cuda-memcheck
doesn't find anything, what are some strategies to figure out where the illegal memory read or write is occurring in the Cuda code? Is there a way with the cuda-gdb
to get to the line of code with invalid memory access?