#pragma warning(disable:4996)
#include <windows.h>
#include <winhttp.h>
#include <stdio.h>
#include <stdlib.h>
#include <process.h>

#include "mytypes.h"
#include "tpool.h"
#include "bitlock.h"

typedef struct {
	U64 sem;
	U32 n;
	union {
		U64 version;
		struct { U16 counts[4]; };
	};
} ARGS;


void worker0( void *arg ) {
	ARGS *v = (ARGS*)arg;
	volatile U16 dummy[ 4 ];
	U32 i;

	wait( &v->sem, 0, 1 );
	for( i = 0; i < v->n; ++i ) ++dummy[ rand() % 4 ];
	return;
}

void worker1( void *arg ) {
	ARGS *v = (ARGS*)arg;
	U32 i;

	wait( &v->sem, 0, 1 );
	for( i = 0; i < v->n; ++i ) _InterlockedIncrement16( &v->counts[ rand() % 4 ] );
	return;
}

void worker2( void *arg ) {
	ARGS *v = (ARGS*)arg;
	U32 i;

	wait( &v->sem, 0, 1 );
	for( i = 0; i < v->n; ++i ) ++v->counts[ rand() % 4 ];
	return;
}

int main( int argc, char **argv ) {
	U32 workers = argc > 1 ? atoi( argv[ 1 ] ) : 4;
	U32 n       = argc > 2 ? atoi( argv[ 2 ] ) : 100000000;
	ARGS a;
    HANDLE pool;
	LARGE_INTEGER freq, start, stop;

	QueryPerformanceFrequency( &freq );

	a.sem = 0ull; a.n = n; a.version = 0ull;
	lock( &a.sem, 0, 0 );
	pool = tpool( &worker0, workers, &a );
	QueryPerformanceCounter( &start );
	unlock( &a.sem, 0 );
	WaitForSingleObject( pool, INFINITE );
	QueryPerformanceCounter( &stop );
	printf( "%4u threads x %10u iterations of normal increment non-shared took %14.9f seconds\n",
		workers, n, (double)( stop.QuadPart - start.QuadPart ) / freq.QuadPart
	);

	a.sem = 0ull; a.n = n; a.version = 0ull;
	lock( &a.sem, 0, 0 );
	pool = tpool( &worker1, workers, &a );
	QueryPerformanceCounter( &start );
	unlock( &a.sem, 0 );
	WaitForSingleObject( pool, INFINITE );
	QueryPerformanceCounter( &stop );
	printf( "%4u threads x %10u iterations of    interlockedIncrement16() took %14.9f seconds\n",
		workers, n, (double)( stop.QuadPart - start.QuadPart ) / freq.QuadPart
	);

	a.sem = 0ull; a.n = n; a.version = 0ull;
	lock( &a.sem, 0, 0 );
	pool = tpool( &worker2, workers, &a );
	QueryPerformanceCounter( &start );
	unlock( &a.sem, 0 );
	WaitForSingleObject( pool, INFINITE );
	QueryPerformanceCounter( &stop );
	printf( "%4u threads x %10u iterations of normal increment shared     took %14.9f seconds\n",
		workers, n, (double)( stop.QuadPart - start.QuadPart ) / freq.QuadPart
	);
	return 0;
}