cudav3.2 compute visual profiler under ubuntu 10.04 seeking help to resolve configuation issuein com

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi,

I am trying to get computerprof work under linux, I have described my system configuration below with error that I am getting. I would appreciate any help that I will get

System Configuration:

Graphic hardware: Nvidia GTX 465
CUDA tool kit version: 3.2
Graphic driver verion: linux 64 bit 260.19.21
Evironment variable and path setting:

PATH:
/opt/intel/composerxe-2011.1.107/bin/intel64:/opt/intel/composerxe-2011.1.107/mpirt/bin/intel64:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin:/usr/games:/usr/local/cuda/bin:/usr/local/java/bin:/opt/intel/bin

LD_LIBRARY_PATH:
/opt/intel/composerxe-2011.1.107/compiler/lib/intel64:/opt/intel/composerxe-2011.1.107/mpirt/lib/intel64:/opt/intel/composerxe-2011.1.107/ipp/…/compiler/lib/intel64:/opt/intel/composerxe-2011.1.107/ipp/lib/intel64:/opt/intel/composerxe-2011.1.107/compiler/lib/intel64:/opt/intel/composerxe-2011.1.107/mkl/lib/intel64:/opt/intel/composerxe-2011.1.107/tbb/lib/intel64//cc4.1.0_libc2.4_kernel2.6.16.21:/usr/local/cuda/lib:

CUDA_PROFILE = 1

CUDA_PROFILE_CSV=1

Earlier post tell that having these two environment variables should let me profile cuda applications. Please let me know what else do it need to do start profiling CUDA applications.

ERROR that I am getting is :

path to application.out file compiled using nvcc/…/temp_compute_profiler_0_0.csv’ for application run 0 not found.

Sincerely,
Vinay

cuda_profiler_env.png
cuda_profiler_env.png1024×768 160 KB

2

That sort of profiling error usually means that your application is not generating any profiler output. You should make sure that the application either calls cuContextDetach() or cudaThreadExit() (which depends which api you are using) at the end of execution, and you should also make sure that the code contains sufficient error checking so you will catch non fatal errors like launch failures and API function errors.

3

I am calling cudaThreadExit(); after my kernel is finished its job. Do I have to cuContextDetach() as well. When I included this function nvcc gave me error. I think its one of the utility functions and I am not able to use any utility library which comes with examples.

My code is compiling and running, however I am seeing expected speedup or want to find bottle necks to optimize more.

Thank you,
Vinay

4

No, as I indicated in my reply, that is what you would call if using the driver API rather than the runtime API. If you are calling that, then the application should flush the profiler output to disk.

I presume that all the obvious things, like write permissions in the directory in which you are running the application and where the profile session root directory is are correct? The application must be able to write the output csv files that the visual profiler uses for analysis.

5

Hi,

I made sure that directory and executable have all permission. Here is my code where I call cudaThreadExit(). I am not sure what more information will be helpful to resolve this.

extern “C” char* transferToGPU(int* pm_image, int pm_width, int pm_height, int* pt_image, int pt_width, int pt_height)
{
if(!InitCUDA())
{
return NULL;
}

// CUDA variables

// Allocate memory in CUDA to transfer

// Copy from host to device,




// Invoke kernel
if(pm_size < BLOCK_SIZE)
{
	dim_blocks = pm_size;
}
else{
	dim_blocks = BLOCK_SIZE;
	
}

dim3 dimBlock(dim_blocks, 1, 1);
dim3 dimGrid(num_blocks, 1);

compare<<< dimGrid, dimBlock >>>(...);

cudaMemcpy(result_host_s, result_device_s, sizeof(float)* s_array_size, cudaMemcpyDeviceToHost);

cudaFree(result_device_s);

cudaThreadExit();

return NULL;

}

6

There is no error checking in that code whatsoever. How do you know that the kernel ever ran? If it didn’t the profiler will produce no output.

7

Any idea to make sure I am running the kernel and profiler is generating output. I am compiling my code using following command, is there any command line switch to make sure profiler output is generated.

nvcc -L/usr/local/lib -I/usr/local/include,/home/vinay/NVIDIA_GPU_Computing_SDK/C/common/inc -lcuda -lcudart -lml -lhighgui -lcv -lcxcore -lcvaux src/main.cpp src/landmark_identification_kernel.cu -o Release/landmark_identification_gpu.out

8

Every API call returns a status. You should check that return status. Directly after the kernel launch call cudaGetLastError() and check the status that it returns. Only then can you be certain whether the program is acutally doing anything at all which will generate events for the profiler to log.

9

Hi all,
I am a very new person to GPU computing, cuda and Ubuntu. I think this will be the correct place to ask my issue in Nvidia driver installation in Linux 10.4 64bit.
I am practicing CUDA in my Notebook with GPU CUDA supported Nvidia graphics card NVS 3100M. I just follow the CUDA3.2 Toolkit guide to install the NVIDIA-Linux-x86_64-260.19.26.run driver for my Linux 10.4. I am trying all the possiblities to stop GUI and install the driver.(sudo /etc/init.d/gdm stop), then i get a black screen with blinking cursor, there # sh NVIDIA-Linux-x86_64-260.19.26.run is not working…I dont know how to install driver (260.19.26) after stop GUI…
I used hardware driver update option also, it take only 195.36.24.
But CUDA3.2 toolkit guide says i should have exact driver-260.19.26.
Please anyone help me to continue my CUDA learning, I see most of you are success cuda3.2 on Ubuntu 10.4.
Thanks
vishva

10

Hi I am looking for your support,
I could not install NVIDIA-Linux-x86_64-260.19.26.run driver in Ubuntu 10.4 64bit… to use CUDA3.2
Is anyone got this same issue?
please anyone can help me!!!