Profiling OpenCV Cuda code on Jetson TX2

Autonomous Machines Jetson & Embedded Systems Jetson TX2
1

Hi, I have made program that has a number of OpenCV cuda calls. I am trying to profile the GPU performance using nvprof but I get the following error

Warning: Unified Memory Profiling is not supported on the underlying platform. System requirements for unified memory can be found at: http://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#um-requirements
^C==3414== Profiling application: ./video_reader
==3414== Profiling result:
No kernels were profiled.

==3414== API calls:
No API activities were profiled.
==3414== Warning: Some profiling data are not recorded. Make sure cudaProfilerStop() or cuProfilerStop() is called before application exit to flush profile data.

How can I get around this? Or are there other profiling tools I can use on the Jetson TX2 (locally, not on another host PC) to measure the performance of the GPU?

Thanks

2

You may try nsight.
You may also use tegrastats for a global view of ressources usage.

3

Nsight needs another machine (host). Is there a way to install it locally on the TX2 ?

4

Nsight does not run directly on the Jetson (not supported in arm64 architecture).

5

Hi,

From the error message, there is no CUDA kernel code in your program.
Could you share your source for us checking?

Thanks.

6

Apologies for the late reply, here is my code

#include <stdio.h>
#include <string.h>
#include <sys/ioctl.h>
#include <sys/mman.h>
#include <unistd.h>
#include <opencv2/core/core.hpp>
#include "opencv2/imgproc/imgproc.hpp"
#include "opencv2/highgui/highgui.hpp"
#include <opencv2/opencv.hpp>

int main(int argc, char * argv[])
{
    cv::VideoCapture cap(0);
    
    for (;;)
    {
    cv::Mat img, y;
    cap >> img;
    cv::cuda::GpuMat img_g, x;
    img_g.upload(img);
    cv::cuda::cvtColor(img_g, x, cv::COLOR_RGB2BGR);
    x.download(y);
    imshow("y", y);
    cv::waitKey(1);
    
    
}
    
    
    
    return 0;
    
    
}
7

You can also get this message off with something like:

/usr/local/cuda/bin/nvprof --unified-memory-profiling off your_cvt

Not sure how much opencv does/can use unified memory by default. Other forum users may share their knowledge about this.

8

Furthermore, you may try to use runtime API:

#include <opencv2/opencv.hpp>
#include <opencv2/videoio.hpp>
#include <opencv2/cudaimgproc.hpp>
#include <opencv2/highgui.hpp>
#include <cuda_runtime.h>
#include <cuda_profiler_api.h>

int main(int argc, char * argv[])
{
	cv::VideoCapture input(0);
	if(!input.isOpened()) {
		std::cout<<"Failed to open camera."<<std::endl;
		return -1;
	}

	cv::Mat          img,   y;
	cv::cuda::GpuMat img_g, x;
	[b]cudaProfilerStart();
[/b]	for (int loop=0; loop <100; ++loop)
	{ 
		if (!input.read(img))
			break;

		img_g.upload(img);
		cv::cuda::cvtColor(img_g, x, cv::COLOR_RGB2BGR);
		x.download(y);
		imshow("y", y);
		cv::waitKey(1);
	}
[b]	cudaProfilerStop();
[/b]
	return 0;
}

Compile with something like: (I’m using opencv-3.3.0 installed in /usr/local/opencv-3.3.0)

g++ -Wall -I/usr/local/opencv-3.3.0/include your_cvt.cpp <b>-I/usr/local/cuda/targets/aarch64-linux/include</b> -L/usr/local/opencv-3.3.0/lib -lopencv_core -lopencv_videoio -lopencv_highgui -lopencv_cudaimgproc <b>-L/usr/local/cuda/targets/aarch64-linux/lib -lcudart</b> -o your_cvt

and run:

export LD_LIBRARY_PATH=/usr/local/cuda/lib64:/usr/local/opencv-3.3.0/lib   # /usr/local/cuda/targets/aarch64-linux/lib should be a default path, but if not found you may add it
/usr/local/cuda/bin/nvprof  --print-gpu-trace --unified-memory-profiling off your_cvt
9

It worked. Thank you very much.