I have installed the driver, toolkit and the SDK for cuda 4.1 on Ubuntu 11.10, but I can’t find the visual profiler. Is it included in the linux release? If so, where is it?
Try looking under: /usr/local/cuda/computeprof/
I went crazy looking at computeprof in the the new 4.1 toolkit.
Normal, computeprof did not exist anymore. Now the profiler application is nvvp and is located in cuda/bin/.
The new Visual Profiler is called NVVP. You’ll find it under $CUDA/bin/nvvp.
By itself, bin/nvvp is just a script calling the executable in libnvvp.
I’ve try the NVVP but failed, it seems the tool can not locate the cuda library, neither could it generate the timeline. When I tried to generate timeline, it always complain the error:CUPTI_NOT_INITIALIZED. seems not good test performed before release. I suggest this tool to add some configure memu to locate the libcuda.so or CUPTI libraries.
I got the same behavior. The profiler acts as it collect information (up to 21 launches), but nothing is displayed.
I’m running on Ubuntu 10.04 x64
I got the following behavior:
I can load nvvp, and it runs my executable, but when it tries to generate a timeline (postprocessing) it stops with an error message:
Unable to locate CUDA libraries and establish connection with CUDA driver.
Make sure CUDA libraries are on your library path.
The visual profiler will exit now
com.nvidia.viper.jni.CuException: CUDA_ERROR_OUT_OF_MEMORY
Visual Profiler worked a few times for me, but now when generating the timeline the profiled program runs but in the end the timeline generation stalls. It keeps running forever without any error messages, so i have to cancel it, but then off course there is no profiling info. That’s really a shame, because it is a great tool which provides insight for internal CUDA operations.
Same here. Same error message:
Unable to locate CUDA libraries and establish connection with CUDA driver.
Make sure CUDA libraries are on your library path.
This is on a SUSE Linux Enterprise (SLES) 11 SP1 x64 system with 2 tesla M2050.
My program using the GPU runs fine without the visual profiler.
I am using the nvvp from the 4.1 rc2 cuda toolkit.
oclDeviceQuery returns CUDA version 4.1.1
Same here, on Debian.
nvvp exit with this error:
"Unable to locate CUDA libraries and establish connection with CUDA driver.
Make sure CUDA libraries are on your library path.
The visual profiler will exit now
com.nvidia.viper.jni.CuException: CUDA_ERROR_OUT_OF_MEMORY"
Did anyone solve the problem??
It is a known bug that will be fixed in the official release
thank u mfatica. do u know when it will be released?
Still throws the same error in the final release (4.1.28 cudatoolkit)… Anyone knows how to fix this?
Aah, my bad. Forgot to reinstall the driver as well. The visual profiler works now!
Cheers!
That doesn’t seem to have worked for me. Did you do anything else to make it work?
Oops, my mistake. I’d installed the development driver that shipped with the RC, and hadn’t noticed that it had also been updated with the release.
The same error (“Unable to locate CUDA libraries … CUDA_ERROR_OUT_OF_MEMORY”) on Ubuntu 11.04 with release 4.1 and driver 290.10. Profiler works only with the development driver 285.05.32.
285.05.32 is the official 4.1 CUDA driver and the one containing the fixes for this problem.
Higher driver number does not mean “newer/better”. 290.10 is probably older than 285.05.32.
I ran visual profiler but there is no timeline appears.
There is a document called “Compute visual profiler” for older visual profiler (computeprof).
Does anyone know where I can find the newest “Compute visual profiler” or user manual for latest visual profiler(nvvp)?
Thanks in advance~
I want to use nvvp,but I can’t find it $CUDA/bin/nvvp.I just find nvprof.
Did I miss something?