© 2017 Boitumelo Ruf. All rights reserved.
Example program performing a grayscale conversion on a GPU with CUDA. The CUDA kernel expects a cudaTextureObject_t with 4 bytes per pixel, which hold the color coded in RGBA. It uses an unsigned int to fetch the pixel data from the texture and references each channel as a uchar array of size 4. After computing the grayscale value it saves the result into the data array of the output image.
On host side the program uses the cv::Mat container of OpenCV to load, store and visualize the input and output images. The program incoorporates a custom function that uploads the cv::Mat into a cudaTextureObject_t and downloads the uchar output array into a cv::Mat.
Note that it is common practise in the context of GPGPU to use a vectorized image data where each pixel holds 4 bytes of image data, i.e. RGBA. Since OpenCV by default uses the BGR format it is required to pretransform the image from BGR to RGBA.
This program is developed and tested on Ubuntu 16.04 with GCC 5.4 and depends on following 3rd party libraries:
- Qt >5.0
- OpenCV >3.0
- CUDA >7.5
- Download or clone repository.
- Copy
src/cuda.pri.tplandsrc/opencv.pri.tplintosrc/cuda.priandsrc/opencv.pri. - Replace tags
<edit-here>insrc/cuda.priandsrc/opencv.priwith corresponding information such as version number and root directory of OpenCV and CUDA installations. - Open
CUDA-GrayscaleCvt.prowith QtCreator or runqmake CUDA-GrayscaleCvt.pro. - Build and run with image as command line argument.
CUDA-GrayscaleCvt ../assets/tsukuba.png. Build output is located in bin/.
NOTE: You might need to add the path to the .so files of OpenCV and CUDA to the LD_LIBRARY_PATH (see qmake output) if you want to start the program from the terminal.
Input [Martull et al. (2012)]:
Output:
Martull et al. (2012)
Martull, S.; Peris, M. & Fukui, K. "Realistic cg stereo image dataset with ground truth disparity maps", Proceedings of the International Conference on Pattern Recognition (ICPR), 2012, vol. 111, pp. 117-118.