I’m having an odd problem with cuda-gdb. I have a kernel like this:
__device__ void kernel ()
{
float q;
...do initialization...
q = compute (args);
...do something with q...
}
inline __device__ float compute (args)
{
float q;
q = result of computation;
return (q);
}
When I run this in cuda-gdb, set “b kernel” and run, I get something like
Breakpoint 1: (pc) in kernel () at /tmp/tmpxft_…kernel_cudafe1.stub.c:3
/tmp/tmpxft_…kernel_cudafe1.stub.c: No such file or directory.
I can’t step into the kernel - single-stepping takes me to the end of another kernel - and doing a “cuda kernel n” says the kernel id is not known.
If I comment out the one line that has the call “q = compute ()”, I can step through the kernel just fine. Any suggestions?
---- PS: Upgrading to the latest compiler/SDK version seems to fix it.