Hi,
I want to use cudaprof to profile cuda programs.
In the shell, I enter the executable /usr/local/cuda/cudaprof/bin/cudaprof
but it gives me an error and refuses to run:
/usr/local/cuda/cudaprof/bin/cudaprof: symbol lookup error: /usr/local/cuda/cudaprof/bin/cudaprof: undefined symbol: _ZN9QListData7detach2Ev
What could cause this problem? Thank you.
Probably missing dynamic libraries it requires to run (most likely QT). The toolkit actually ships with local versions of everything you need to run the profiler. If you add the full path of the profiler bin directory to your LD_LIBRARY_PATH variable, it should work.
Yes, this solves the problem. But I’ve got another problem: ERROR in reading profiler output. What could be wrong this time? Thank you.
Try adding a cudaThreadExit() call to the end of your program (or any point where you might have an abnormal termination point). That should force the CUDA data buffers to get flushed out to disk before the program exits. If that isn’t helping, I don’t know what to suggest.
It does not solve the problem. The profiler seems to be able to launch the program, but cannot read its output.
=== Start profiling for session ‘Session1’ ===
Start program xterm -e /home/grads/sgu1/gp/cuda/Gu/main.cu, run #1 …
Program run #1 completed.
Start program xterm -e /home/grads/sgu1/gp/cuda/Gu/main.cu, run #2 …
Program run #2 completed.
Start program xterm -e /home/grads/sgu1/gp/cuda/Gu/main.cu, run #3 …
Program run #3 completed.
Start program xterm -e /home/grads/sgu1/gp/cuda/Gu/main.cu, run #4 …
Program run #4 completed.
Start program xterm -e /home/grads/sgu1/gp/cuda/Gu/main.cu, run #5 …
Program run #5 completed.
Error in reading profiler output.
Hi again,
Finally I get the Visual Profiler running, but it only gives me the profile for one kernel, but I have three…;
Besides, the in the profiler counters tab, i cannot choose the gld uncoalesced, gld coalesced, gst uncoalesced, gst coalesced (it is disabled by cudaprof.), what would be the problem? Thank you.