profiling opencl using nvprof commandline

#1

Hello

When profiling an opencl application with compute capability 6.1 and cuda 8.0 toolkit, using the commandline nvprof.exe, I run into the following error:
==9652== Profiling application: myopencl.exe
==9652== Profiling result:
No kernels were profiled.

I couldn’t find anything online that seemed to work. Please note this is for OPENCL and not CUDA.

Any suggestion would be appreciated.

thank you

#2

You might try older NVIDIA tools. They’ve been slowly removing OpenCL support in newer versions of the tools.

#3

I tried an older version of the toolkit: 7.5. That did not work either. Do you know if OpenCL support is going to be coming back in the newer versions of the toolkit with OpenCL 2.0?

1 Like
#4

Hi, minidrive

Do you know if OpenCL support is going to be coming back in the newer versions of the toolkit with OpenCL 2.0?

I think nVidia will have not any plans to support OpenCL in nVidia Tools, they haven’t supported OpenCL 2.0 in full yet.

#5

Interestingly, I can get the GUI version to work with 8.0 , just not the commandline version. Does anyone have any idea if the GUI version saves the numbers somewhere. I have looked around in the nvreport , nvactivity etc. None of them hold any useful text data.

#6

It may not give you quite as much detail as the IHV-provided tools, but the “Device Performance Timing” capabilities of the Intercept Layer for OpenCL Applications will give you some profiling information about your OpenCL application, such as the total / min / max / average time for each kernel, and it works on any OpenCL implementation:

https://github.com/intel/opencl-intercept-layer/blob/master/docs/controls.md#deviceperformancetiming-bool

Is this sufficient for your needs?

#7

do you mind sharing the steps/settings how you got nvvp to profile OpenCL?

I could not even get that to work - my program does run in nvvp, but it does not produce a timeline, or provide any useful information (I am interested in the PC sampling profiling like for my cuda code).

Here is my screenshot

1 Like
#8

@fangqq: were you able to profile an OpenCL program on Nvidia devices?
If so, could you please share it? Thanks!