/* Check cf5-opt.vim defs.
VIM: let g:lcppflags="-std=c++11 -O2 -pthread"
VIM: let g:wcppflags="/O2 /EHsc /DWIN32"
VIM: let g:cppflags=g:Iboost
VIM: let g:ldflags=g:Lboost
VIM: let g:ldlibpath=g:Bboost
VIM: let g:argv=""
VIM-: let g:cf5output=0
*/
#include <iostream>
#include <iomanip>
#include <exception>
#include <chrono>
#include <algorithm>

typedef long long duration_type;
static const size_t ce = 1024L*1024L*1024L;

duration_type scan_read( char * p, const size_t c, const size_t s )
{
	auto b = std::chrono::high_resolution_clock::now();
	
	int sum = 0;
	char * qe = p+c;
	for ( int n =0, ne=ce/c; n < ne; ++n ) 
		for ( size_t i = 0; i < s; ++i )
			for ( char * q = p+i; q < qe; q+=s )
				sum += *q;
	volatile int no_optimization = sum;
	auto e = std::chrono::high_resolution_clock::now();
	return std::chrono::nanoseconds(e-b).count();
}

duration_type scan_write( char * p, const size_t c, const size_t s )
{
	int a = rand();

	auto b = std::chrono::high_resolution_clock::now();
	
	char * qe = p+c;
	for ( int n =0, ne=ce/c; n < ne; ++n ) 
		for ( size_t i = 0; i < s; ++i )
			for ( char * q = p+i; q < qe; q+=s )
				*q = a;
	
	auto e = std::chrono::high_resolution_clock::now();
	return std::chrono::nanoseconds(e-b).count();
}

template <class F>
void test_steps( F f , char * p, const size_t c, char * nm )
{
	f( p, c, 1 ); // First run ignore.

	std::cout << std::setw(10) << "steps";
	std::cout << std::setw(16) << nm;
	std::cout << std::setw(16) << nm;
	std::cout << std::setw(16) << nm;
	std::cout << std::setw(16) << "average";
	std::cout << std::setw(16) << "deviation" << std::endl;

	for ( size_t s = 1; s < c; s<<=1 )
	{
		std::cout << std::setw(10) << s << std::flush;
		auto d1 = f( p, c, s );
		std::cout << std::setw(16) << d1 << std::flush;
		auto d2 = f( p, c, s );
		std::cout << std::setw(16) << d2 << std::flush;
		auto d3 = f( p, c, s );
		std::cout << std::setw(16) << d3 << std::flush;
		auto a = (d1+d2+d3)/3;
		std::cout << std::setw(16) << a;
		auto dev = (std::abs(d1-a)+std::abs(d2-a)+std::abs(d3-a))/3;
		std::cout << std::setw(16) << dev << std::endl;
	}
}

void test_workingset( const size_t c )
{
	char * p = (char*)malloc( c );
	
	std::cout << "Working set " << c << std::endl ;
	test_steps( scan_read, p, c, "read" );
	test_steps( scan_write, p, c, "write" );
}

int main ( void )
{try{


	for ( size_t c = 4096; c <= ce; c <<=1 )
		test_workingset( c );

	return 0;
}
catch ( const std::exception& e )
{
	std::cerr << std::endl
			<< "std::exception(\"" << e.what() << "\")." << std::endl;
	return 2;
}
catch ( ... )
{
	std::cerr  << std::endl
			<< "unknown exception." << std::endl;
	return 1;
}}