Cilk and OpenMP version of scene detection algorithm

cilk_scene.cpp

#include <math.h>

#include <iostream>
#include <vector>
#include <cstdlib>

#include <sys/time.h>
#include <sys/types.h>

#include <cilk/cilk.h>
#include <cilk/reducer_opadd.h>

#define PIXEL_SIZE 3

using namespace std;

// this is just a stub
unsigned char* generateVideo(int numberOfFrames, int frameWidth, int frameHeight) {
	unsigned int sizeOfFrame = frameWidth * frameHeight * PIXEL_SIZE;

	unsigned char* video = new unsigned char[sizeOfFrame * numberOfFrames];

	for (unsigned int i = 0; i < numberOfFrames; i++) {
		for (int j = 0; j < sizeOfFrame; j++) {
			video[i * sizeOfFrame + j] = i % 256;
		}
	}

	return video;
}


void foo() {
	sleep(3);
}

void prepareTest(int argc, char *argv[]) {
	cilk_spawn foo();
	cout << "Intel Cilk Plus version of video scene extraction algorithm" << endl;
	cout << "Number of workers: " << __cilkrts_get_nworkers() << endl;
}

void computeMAFDs(unsigned char* video, int numberOfFrames, int width, int height, float* MAFDs) {
	int frameSize = width * height;
	int pixelsCount = frameSize * PIXEL_SIZE;


	cilk_for(unsigned int k = 1; k < numberOfFrames; k++) {
		unsigned char* frame1 = video + (k - 1) * pixelsCount;
		unsigned char* frame2 = frame1 + pixelsCount;

		cilk::reducer_opadd<float> MAFD;
		MAFD.set_value(0.0);

		cilk_for (int i = 0; i < pixelsCount; i += PIXEL_SIZE) {
			float blue = frame1[i] - frame2[i];
			float green = frame1[i + 1] - frame2[i + 1];
			float red = frame1[i + 2] - frame2[i + 2];

			MAFD += sqrt(blue * blue + green * green + red * red);
		}

		MAFDs[k - 1] = MAFD.get_value() / frameSize;
	}
}

int main(int argc, char *argv[]) {
	int numberOfFrames;
	int width;
	int height;

	if (argc != 4) {
		numberOfFrames = 250;
		width = 720;
		height = 480;
	} else {
		numberOfFrames = std::atoi(argv[1]);
		width = std::atoi(argv[2]);
		height = std::atoi(argv[3]);
	}

	if (numberOfFrames <= 0 || width <= 0 || height <= 0) {
		cout << "Incorrect parameters used" << endl;
		return 1;
	}

	cout << "Number of frames: " << numberOfFrames << ", width: " << width << ", height: " << height << endl;

	float* MAFDs = new float[numberOfFrames - 1];

	for (int i = 0; i < numberOfFrames - 1; i++) {
		MAFDs[i] = 0;
	}

	unsigned char* video = generateVideo(numberOfFrames, width, height);

	prepareTest(argc, argv);

	cout << "start " << endl;

	struct timeval tp;
	double sec, usec, start, end;

	gettimeofday( &tp, NULL );
	sec = static_cast<double>( tp.tv_sec );
	usec = static_cast<double>( tp.tv_usec )/1E6;
	start = sec + usec;

	computeMAFDs(video, numberOfFrames, width, height, MAFDs);

	gettimeofday( &tp, NULL );
	sec = static_cast<double>( tp.tv_sec );
	usec = static_cast<double>( tp.tv_usec )/1E6;
	end = sec + usec;

	double time = end - start;

	cout << "Time " << time << "s" << endl;

	delete[] video;
	delete[] MAFDs;

	return 0;
}

openmp_scene.cpp

#include <math.h>

#include <iostream>
#include <vector>
#include <cstdlib>

#include <sys/time.h>
#include <sys/types.h>

#include <omp.h>

#define PIXEL_SIZE 3

using namespace std;

// this is just a stub
unsigned char* generateVideo(int numberOfFrames, int frameWidth, int frameHeight) {
	unsigned int sizeOfFrame = frameWidth * frameHeight * PIXEL_SIZE;

	unsigned char* video = new unsigned char[sizeOfFrame * numberOfFrames];

	for (unsigned int i = 0; i < numberOfFrames; i++) {
		for (int j = 0; j < sizeOfFrame; j++) {
			video[i * sizeOfFrame + j] = i % 256;
		}
	}

	return video;
}


void prepareTest(int argc, char *argv[]) {
	out << "OpenMP version of video scene extraction algorithm" << endl;
}

void computeMAFDs(unsigned char* video, int numberOfFrames, int width, int height, float* MAFDs) {
	int frameSize = width * height;
	int pixelsCount = frameSize * PIXEL_SIZE;

	#pragma omp parallel for
	for (int k = 1; k < numberOfFrames; k++) {
		unsigned char* frame1 = video + (k - 1) * pixelsCount;
		unsigned char* frame2 = frame1 + pixelsCount;

		float MAFD = 0.0;

		// this is commented because nested parallelism is not efficient in OpenMP
		//#pragma omp parallel for reduction(+:MAFD)
		for (int i = 0; i < pixelsCount; i += PIXEL_SIZE) {
			float blue = frame1[i] - frame2[i];
			float green = frame1[i + 1] - frame2[i + 1];
			float red = frame1[i + 2] - frame2[i + 2];

			MAFD += sqrt(blue * blue + green * green + red * red);
		}

		MAFDs[k - 1] = MAFD / frameSize;
	}
}

int main(int argc, char *argv[]) {
	int numberOfFrames;
	int width;
	int height;

	if (argc != 4) {
		numberOfFrames = 250;
		width = 720;
		height = 480;
	} else {
		numberOfFrames = std::atoi(argv[1]);
		width = std::atoi(argv[2]);
		height = std::atoi(argv[3]);
	}

	if (numberOfFrames <= 0 || width <= 0 || height <= 0) {
		cout << "Incorrect parameters used" << endl;
		return 1;
	}

	cout << "Number of frames: " << numberOfFrames << ", width: " << width << ", height: " << height << endl;

	float* MAFDs = new float[numberOfFrames - 1];

	for (int i = 0; i < numberOfFrames - 1; i++) {
		MAFDs[i] = 0;
	}

	unsigned char* video = generateVideo(numberOfFrames, width, height);

	prepareTest(argc, argv);

	cout << "start " << endl;

	struct timeval tp;
	double sec, usec, start, end;

	gettimeofday( &tp, NULL );
	sec = static_cast<double>( tp.tv_sec );
	usec = static_cast<double>( tp.tv_usec )/1E6;
	start = sec + usec;

	computeMAFDs(video, numberOfFrames, width, height, MAFDs);

	gettimeofday( &tp, NULL );
	sec = static_cast<double>( tp.tv_sec );
	usec = static_cast<double>( tp.tv_usec )/1E6;
	end = sec + usec;

	double time = end - start;

	cout << "Time " << time << "s" << endl;

	delete[] video;
	delete[] MAFDs;

	return 0;
}