ThreadPool simple example

#define DELAY_COUNT 20
#define CACHE_LINE_SIZE 64


void CALLBACK Worker(PTP_CALLBACK_INSTANCE, PVOID, PTP_WORK);
int workerDelay = DELAY_COUNT;
DWORD tid[256]={0,};

__declspec(align(CACHE_LINE_SIZE))
	typedef struct _THARG{
		SRWLOCK SRWL;
		int threadNumber;
		unsigned int tasksToComplete;
		unsigned int tasksComplete;
} THARG;


int _tmain(DWORD argc, LPTSTR argv[])
{
	int nThread, iThread;
	HANDLE* pWorkObjects;
	SRWLOCK srwl;
	unsigned int tasksPerThread, totalTasksComplete;
	THARG ** pThreadArgsArray, *pThreadArg;
	TP_CALLBACK_ENVIRON cbe;
	BOOL traceFlag = FALSE;

	nThread = _ttoi(argv[1]);
	tasksPerThread = _ttoi(argv[2]);

	InitializeSRWLock(&srwl);

	pWorkObjects = malloc ( nThread * sizeof(PTP_WORK));
	pThreadArgsArray = malloc ( nThread * sizeof(THARG*));

	InitializeThreadpoolEnvironment(&cbe);

	for ( iThread = 0; iThread < nThread; iThread++ )
	{
		pThreadArg = (pThreadArgsArray[iThread] = _aligned_malloc(sizeof(THARG), CACHE_LINE_SIZE));
		pThreadArg->threadNumber = iThread;
		pThreadArg->tasksToComplete = tasksPerThread;
		pThreadArg->tasksComplete = 0;
		pThreadArg->SRWL = srwl;
		pWorkObjects[iThread] = CreateThreadpoolWork(Worker, pThreadArg, &cbe);
		SubmitThreadpoolWork(pWorkObjects[iThread]);
	}

	for(iThread=0; iThreadtasksComplete);
		_tprintf(_T("Tasks completed by thread %5d: %6d\n"), iThread, tid[iThread]);
		totalTasksComplete += pThreadArg->tasksComplete;
		_aligned_free((pThreadArg));
	}


	free (pThreadArgsArray);
	_tprintf(_T("TotalWork performed : %d\n"), totalTasksComplete);
	return 0;
}

void CALLBACK Worker(PTP_CALLBACK_INSTANCE Instance, PVOID Context, PTP_WORK Work)
{
	THARG *threadArgs;
	threadArgs = (THARG*) Context;
	


	tid[threadArgs->threadNumber] = GetCurrentThreadId();
	while(threadArgs->tasksComplete < threadArgs->tasksToComplete){
		delay_cpu(workerDelay);
		AcquireSRWLockExclusive(&(threadArgs->SRWL));
		(threadArgs->tasksComplete)++;
		ReleaseSRWLockExclusive(&(threadArgs->SRWL));

	}
	return;

}