On Linux:
sudo apt-get update
sudo apt-get upgrade
sudo apt-get install libopencv-dev
To compile: make
To run serial program ./output
To run parallel program ./p_output
Navigate into gpu directory: cd gpu
To compile: make
To show an image of what the result of the Canny algorithm is with GPU code execute: ./canny_show_img
To run total time benchmark for unoptimized Canny: ./canny_data_uo
To run total time benchmark for optimized Canny: ./canny_data_o
To run detailed timing analysis for various stages of unoptimized Canny: ./canny_test_analysis_uo
To run detailed timing analysis for various stages of optimized Canny: ./canny_test_analysis_o
The GPU code can only run for images that are of size less than or equal to 1024x1024
The optimized GPU code will not have speedup if one is using more than 16 Streaming Multiprocessors.
A bug in the GPU code that has not been resolved is running optimized canny on 1024x1024 images in a dataset.