#include <cstdlib>
#include <sys/time.h>
#include <string>
#include <sstream>
#include <iostream>
#include <iomanip>
#include "cpu.h"
#include "gpu.h"

#define BSZ 4
#define NIT	100

using namespace std;

unsigned long bsize = 512;

double	gtime();
string	elapsed(double,double);
void	vinit(float* u,int n);

int main(int argc,char* argv[])
{
	double	ts,te;
	float	*u;
	int		bsz = BSZ;
	int		nit = NIT;
	int		sz;
	float	sample;

	if (argc>1)
		bsz = atoi(argv[1]);
	sz = bsz*1024*1024;

	cpu_init();
	gpu_init();
	cout << endl;

	u = new float[sz];

	cout << "Set size:   " << sz/(1024UL*1024UL) << " Mfloats" << endl;
	cout << "Data size:  " << (sz*sizeof(float))/(1024UL*1024UL) << " Mb" << endl;
	cout << "Iterations: " << NIT << endl;
	cout << "Block size: " << bsize << endl;
	cout << "Grid  size: " << sz/bsize << endl;
	cout << endl;

//	ts = gtime();
	cpu_run(u,sz,nit);
//	te = gtime();
//	cout << "CPU Total:    " << elapsed(ts,te) << endl;
//	cout << u[2] << endl;
	sample = u[2];

//	ts = gtime();
	gpu_run(u,sz,nit);
//	te = gtime();
//	cout << "GPU Total:    " << elapsed(ts,te) << endl;
//	cout << u[2] << endl;

//	if (u[2]!=sample)
//	{
//		cerr << "===========================" << endl;
//		cerr << "       Bad data!!          " << endl;
//		cerr << "===========================" << endl;
//	}

//	cout << endl;

	delete[] u;

	cpu_fini();
	gpu_fini();
	cout << endl;

	return 0;
}

void vinit(float* u,int n)
{
	for (int i=0; i<n; i++)
		u[i] = 0.0f;
	u[0]   = 1.0f;
	u[n-1] = 1.0f;
}

double gtime()
{
	struct timeval tv;
	gettimeofday(&tv,0);

	return 1000.0*tv.tv_sec+0.001*tv.tv_usec;
}

string elapsed(double ts,double te)
{
	std::ostringstream ss;

	// setfill('0')
	ss.setf(ios::fixed);
    ss << setw(10) << setprecision(2) << te-ts << " ms";
	ss.flush();

	string res(ss.str());
	return res;
}