when I run cudaprof on a machine equipped with a cuda device, I get “unable to load the cuda library” dialog.
CudaVisualProfiler is version 1.1.5
cuda driver is release 2.1, V0.2.1221
The operating system is SuSE Linux Enterprise Server 10 SP1.
I did have CudaVisualProfiler 1.0.11 on the same system, According to changelog it did not perform this check, however this version of the profiler always failed with Error -94 on binaries
created with cuda 2.1.
I have a similar problem on Ubuntu 8.04 with those visual profiler and cuda versions.
I have found that I can get around it by starting up the profiler version 1.0.11 and using it to execute the program I wish to profile. When the Error -94 appears, I copy the temp_cuda_profiler_*.csv files somewhere else (they get deleted after you click OK). Then, I launch the visual profiler 1.1.5 and import the CSV files.
This is a bit troublesome, but it seems to get around the problem for now.
I wonder if I can get away without using the Visual Profiler GUI at all. The profiler does not seem to add any instrumentation to the executable, so it probably is just launching the executable with some special environmental variables set to get it to dump all the interesting parameters, such as the Instruction count, number of CTAs, local loads, glds and gsts.
When I just set CUDA_PROFILE and CUDA_PROFILE_CSV variables and run the code outside of the profiler I get only the timings in the cuda_profile.csv.
If I could get it to dump all the extra numbers I would’ve just use gnuplot or openoffice, to get whatever visualization there is.