#include 
#include 
#include 
#include 
#include 
#include 

#ifdef WIN32
#include 
#define aligned_alloc(align, size) _aligned_malloc((size), (align))
#endif

#ifdef __EMSCRIPTEN__
#include 
#endif

#include "tick.h"

char dst[1024*1024*64+16] = {};
char src[1024*1024*64+16] = {};

uint8_t resultCheckSum = 0;

void __attribute__((noinline)) test_memcpy(int numTimes, int copySize)
{
	for(int i = 0; i < numTimes - 8; i += 8)
	{
		memcpy(dst, src, copySize); resultCheckSum += dst[copySize >> 1];
		memcpy(dst, src, copySize); resultCheckSum += dst[copySize >> 1];
		memcpy(dst, src, copySize); resultCheckSum += dst[copySize >> 1];
		memcpy(dst, src, copySize); resultCheckSum += dst[copySize >> 1];
		memcpy(dst, src, copySize); resultCheckSum += dst[copySize >> 1];
		memcpy(dst, src, copySize); resultCheckSum += dst[copySize >> 1];
		memcpy(dst, src, copySize); resultCheckSum += dst[copySize >> 1];
		memcpy(dst, src, copySize); resultCheckSum += dst[copySize >> 1];
	}
	numTimes &= 15;
	for(int i = 0; i < numTimes; ++i)
	{
		memcpy(dst, src, copySize); resultCheckSum += dst[copySize >> 1];
	}
}

std::vector copySizes;
std::vector results;

std::vector testCases;

double totalTimeSecs = 0.0;

void test_case(int copySize)
{
	const int minimumCopyBytes = 1024*1024*64;

	int numTimes = (minimumCopyBytes + copySize-1) / copySize;
	if (numTimes < 8) numTimes = 8;

	tick_t bestResult = 1e9;

#ifndef NUM_TRIALS
#define NUM_TRIALS 5
#endif

	for(int i = 0; i < NUM_TRIALS; ++i)
	{
		double t0 = tick();
		test_memcpy(numTimes, copySize);
		double t1 = tick();
		if (t1 - t0 < bestResult) bestResult = t1 - t0;
		totalTimeSecs += (double)(t1 - t0) / ticks_per_sec();
	}
	unsigned long long totalBytesTransferred = numTimes * copySize;

	copySizes.push_back(copySize);

	tick_t ticksElapsed = bestResult;
	if (ticksElapsed > 0)
	{
		double seconds = (double)ticksElapsed / ticks_per_sec();
		double bytesPerSecond = totalBytesTransferred / seconds;
		double mbytesPerSecond = bytesPerSecond / (1024.0*1024.0);
		results.push_back(mbytesPerSecond);
	}
	else
	{
		results.push_back(0.0);
	}
}

void print_results()
{
	std::cout