使用信号量控制线程执行顺序，进而控制不同视频流的解码顺序

#include <tchar.h>
#include "QueryPerformance.h" // time counter
#include "windows.h"
#include <stdio.h>
#include <process.h>
#include "cudaDecodeInterface.h"
#pragma comment(lib,"cudaDecode.lib")

typedef unsigned (WINAPI *PBEGINTHREADEX_THREADFUNC)
	(LPVOID lpThreadParameter);

typedef unsigned *PBEGINTHREADEX_THREADID;

HANDLE hand_cuda[6]  = {NULL, NULL, NULL, NULL,NULL,NULL};

bool cuda_decodeEnd[6] = {false, false, false, false,false,false};

static int counter[6] = {0, 0, 0, 0,0,0};

FILE *fp0 = fopen("../decodePic/cudadecode0.yuv", "wb");
FILE *fp1 = fopen("../decodePic/cudadecode1.yuv", "wb");
FILE *fp2 = fopen("../decodePic/cudadecode2.yuv", "wb");
FILE *fp3 = fopen("../decodePic/cudadecode3.yuv", "wb");
FILE *fp4 = fopen("../decodePic/cudadecode4.yuv", "wb");
FILE *fp5 = fopen("../decodePic/cudadecode5.yuv", "wb");

// front claim
int CudaDecodeNotify0(unsigned char* pDecodeData, int nDataSize, DWORD timestamp, int nEof);
int CudaDecodeNotify1(unsigned char* pDecodeData, int nDataSize, DWORD timestamp, int nEof);
int CudaDecodeNotify2(unsigned char* pDecodeData, int nDataSize, DWORD timestamp, int nEof);
int CudaDecodeNotify3(unsigned char* pDecodeData, int nDataSize, DWORD timestamp, int nEof);
int CudaDecodeNotify4(unsigned char* pDecodeData, int nDataSize, DWORD timestamp, int nEof);
int CudaDecodeNotify5(unsigned char* pDecodeData, int nDataSize, DWORD timestamp, int nEof);

DWORD WINAPI OnCuda0();
DWORD WINAPI OnCuda1();
DWORD WINAPI OnCuda2();
DWORD WINAPI OnCuda3();
DWORD WINAPI OnCuda4();
DWORD WINAPI OnCuda5();

HANDLE handle_video[6] = {NULL, NULL, NULL, NULL,NULL,NULL};

HANDLE hand_semaphore[6] = {NULL, NULL, NULL, NULL,NULL,NULL};

int main(int argc, char* argv[])
{
	printf("Cuda Decoder......\n");

	// Initialzie semaphore
	hand_semaphore[0] = CreateSemaphore(NULL, 1, 1, NULL);
	hand_semaphore[1] = CreateSemaphore(NULL, 0, 1, NULL);
	hand_semaphore[2] = CreateSemaphore(NULL, 0, 1, NULL);
	hand_semaphore[3] = CreateSemaphore(NULL, 0, 1, NULL);
	hand_semaphore[4] = CreateSemaphore(NULL, 0, 1, NULL);
	hand_semaphore[5] = CreateSemaphore(NULL, 0, 1, NULL);

	// Initialize hand
	//	DWORD begin_time = GetTickCount64();

	CStopwatch stopwatch;

	WaitForSingleObject(hand_semaphore[0], INFINITE);
	handle_video[0] = (HANDLE)_beginthreadex(NULL, 0, (PBEGINTHREADEX_THREADFUNC)OnCuda0, NULL, 0, NULL);

	WaitForSingleObject(hand_semaphore[1], INFINITE);
	handle_video[1] = (HANDLE)_beginthreadex(NULL, 0, (PBEGINTHREADEX_THREADFUNC)OnCuda1, NULL, 0, NULL);

	WaitForSingleObject(hand_semaphore[2], INFINITE);
	handle_video[2] = (HANDLE)_beginthreadex(NULL, 0, (PBEGINTHREADEX_THREADFUNC)OnCuda2, NULL, 0, NULL);

	WaitForSingleObject(hand_semaphore[3], INFINITE);
	handle_video[3] = (HANDLE)_beginthreadex(NULL, 0, (PBEGINTHREADEX_THREADFUNC)OnCuda3, NULL, 0, NULL);

	WaitForSingleObject(hand_semaphore[4], INFINITE);
	handle_video[4] = (HANDLE)_beginthreadex(NULL, 0, (PBEGINTHREADEX_THREADFUNC)OnCuda4, NULL, 0, NULL);

	WaitForSingleObject(hand_semaphore[5], INFINITE);
	handle_video[5] = (HANDLE)_beginthreadex(NULL, 0, (PBEGINTHREADEX_THREADFUNC)OnCuda5, NULL, 0, NULL);

	WaitForMultipleObjects(6, handle_video, TRUE, INFINITE);

	__int64 qwElapsedTime = stopwatch.Now();
	printf("总用时为：%d<ms>\n", qwElapsedTime);

	//	DWORD end_time = GetTickCount64();
	//	printf("总用时为：%d<ms>\n", end_time - begin_time);

	for (int i=0; i<6; i++)
	{
		CloseHandle(handle_video[i]);
	}

	fflush(stdin);
	system("pause");
	return 0;
}

// OnCuda函数
DWORD WINAPI OnCuda0()
{

	bool nstate = Cuda_SupportDecode();
	hand_cuda[0] = Cuda_CreateDecoder();

	DWORD StartTime = 0; // 起始的时间
	DWORD ElapsedTime = 0; //消耗的时间

	if( Cuda_OpenFile(hand_cuda[0], "../data/720P/video0.m2v", CudaDecodeNotify0)==1 ) 
	{
		StartTime = GetTickCount();
		printf("文件<1>打开成功\n");
	}
	else
		printf("文件<1>打开失败\n");

	while ( !cuda_decodeEnd[0] )
	{
		Sleep(1);
	}
	ElapsedTime = GetTickCount() - StartTime;
	//	printf("文件<1>解码完成   nTime1 = %d<ms>\n", ElapsedTime);

	Cuda_DestroyDecoder(hand_cuda[0]);

	return 0;
}

DWORD WINAPI OnCuda1()
{

	bool nstate = Cuda_SupportDecode();
	hand_cuda[1] = Cuda_CreateDecoder();

	DWORD StartTime = 0; // 起始的时间
	DWORD ElapsedTime = 0; //消耗的时间

	if( Cuda_OpenFile(hand_cuda[1], "../data/720P/video1.m2v", CudaDecodeNotify1)==1 ) 
	{
		StartTime = GetTickCount();
		printf("文件<2>打开成功\n");
	}
	else
		printf("文件<2>打开失败\n");

	while ( !cuda_decodeEnd[1] )
	{
		Sleep(1);
	}
	ElapsedTime = GetTickCount() - StartTime;
	//	printf("文件<2>解码完成   nTime2 = %d<ms>\n", ElapsedTime);

	Cuda_DestroyDecoder(hand_cuda[1]);

	return 0;
}

DWORD WINAPI OnCuda2()
{

	bool nstate = Cuda_SupportDecode();
	hand_cuda[2] = Cuda_CreateDecoder();

	DWORD StartTime = 0; // 起始的时间
	DWORD ElapsedTime = 0; //消耗的时间

	if( Cuda_OpenFile(hand_cuda[2], "../data/720P/video2.m2v", CudaDecodeNotify2)==1 ) 
	{
		StartTime = GetTickCount();
		printf("文件<3>打开成功\n");
	}
	else
		printf("文件<3>打开失败\n");

	while ( !cuda_decodeEnd[2] )
	{
		Sleep(1);
	}
	ElapsedTime = GetTickCount() - StartTime;
	//	printf("文件<3>解码完成   nTime3 = %d<ms>\n", ElapsedTime);

	Cuda_DestroyDecoder(hand_cuda[2]);

	return 0;
}

DWORD WINAPI OnCuda3()
{

	bool nstate = Cuda_SupportDecode();
	hand_cuda[3] = Cuda_CreateDecoder();

	DWORD StartTime = 0; // 起始的时间
	DWORD ElapsedTime = 0; //消耗的时间

	if( Cuda_OpenFile(hand_cuda[3], "../data/720P/video3.m2v", CudaDecodeNotify3)==1 ) 
	{
		StartTime = GetTickCount();
		printf("文件<4>打开成功\n");
	}
	else
		printf("文件<4>打开失败\n");

	while ( !cuda_decodeEnd[3] )
	{
		Sleep(1);
	}
	ElapsedTime = GetTickCount() - StartTime;
	//	printf("文件<4>解码完成   nTime4 = %d<ms>\n", ElapsedTime);

	Cuda_DestroyDecoder(hand_cuda[3]);

	return 0;
}

DWORD WINAPI OnCuda4()
{

	bool nstate = Cuda_SupportDecode();
	hand_cuda[4] = Cuda_CreateDecoder();

	DWORD StartTime = 0; // 起始的时间
	DWORD ElapsedTime = 0; //消耗的时间

	if( Cuda_OpenFile(hand_cuda[4], "../data/720P/video4.m2v", CudaDecodeNotify4)==1 ) 
	{
		StartTime = GetTickCount();
		printf("文件<1>打开成功\n");
	}
	else
		printf("文件<1>打开失败\n");

	while ( !cuda_decodeEnd[4] )
	{
		Sleep(1);
	}
	ElapsedTime = GetTickCount() - StartTime;
	//	printf("文件<1>解码完成   nTime1 = %d<ms>\n", ElapsedTime);

	Cuda_DestroyDecoder(hand_cuda[4]);

	return 0;
}

DWORD WINAPI OnCuda5()
{

	bool nstate = Cuda_SupportDecode();
	hand_cuda[5] = Cuda_CreateDecoder();

	DWORD StartTime = 0; // 起始的时间
	DWORD ElapsedTime = 0; //消耗的时间

	if( Cuda_OpenFile(hand_cuda[5], "../data/720P/video5.m2v", CudaDecodeNotify5)==1 ) 
	{
		StartTime = GetTickCount();
		printf("文件<1>打开成功\n");
	}
	else
		printf("文件<1>打开失败\n");

	while ( !cuda_decodeEnd[5] )
	{
		Sleep(1);
	}
	ElapsedTime = GetTickCount() - StartTime;
	//	printf("文件<1>解码完成   nTime1 = %d<ms>\n", ElapsedTime);

	Cuda_DestroyDecoder(hand_cuda[5]);

	return 0;
}

// cudaDecodeNotify函数，拷贝解码后的数据到硬盘
int CudaDecodeNotify0(unsigned char* pDecodeData, int nDataSize, DWORD timestamp, int nEof)
{
	counter[0]++;
	//	fwrite(pDecodeData,1,nDataSize,fp0);
	if(nEof)
	{
		OutputDebugString("thread 1 读包结束！！！\n");
		printf("====== thread 1 读包结束！！！\n");
		//		fclose(fp0);
		cuda_decodeEnd[0] = true;
	}

#if 0
	printf(" \n------ video 1：%d ------\n", counter[0]);
	FILE *file = fopen("../order.txt", "at");
	fprintf(file, " ------ 视频1：%d ------\n", counter[0]);
	fclose(file);
#endif

	ReleaseSemaphore(hand_semaphore[1], 1, NULL);
	Sleep(1);
	return 0;
}

int CudaDecodeNotify1(unsigned char* pDecodeData, int nDataSize, DWORD timestamp, int nEof)
{
	counter[1]++;
	//	fwrite(pDecodeData,1,nDataSize,fp1);
	if(nEof)
	{
		OutputDebugString("thread 2 读包结束！！！\n");
		printf("====== thread 2 读包结束！！！\n");
		//				fclose(fp4);
		cuda_decodeEnd[1] = true;
	}

#if 0
	printf(" \n------ video 2 : %d ------\n", counter[1]);
	FILE *file = fopen("../order_new.txt", "at");
	fprintf(file, " ------ 视频2：%d ------\n", counter[1]);
	fclose(file);
#endif

	ReleaseSemaphore(hand_semaphore[2], 1, NULL);
	Sleep(1);

	return 0;
}

int CudaDecodeNotify2(unsigned char* pDecodeData, int nDataSize, DWORD timestamp, int nEof)
{
	counter[2]++;
	//	fwrite(pDecodeData,1,nDataSize,fp2);
	if(nEof)
	{
		OutputDebugString("thread 3 读包结束！！！\n");
		printf("====== thread 3 读包结束！！！\n");
		//				fclose(fp4);
		cuda_decodeEnd[2] = true;
	}

#if 0
	printf(" \n------ video 3：%d ------\n", counter[2]);
	FILE *file = fopen("../order.txt", "at");
	fprintf(file, " ------ 视频3：%d ------\n", counter[2]);
	fclose(file);

#endif

	ReleaseSemaphore(hand_semaphore[3], 1, NULL);
	Sleep(1);

	return 0;
}

int CudaDecodeNotify3(unsigned char* pDecodeData, int nDataSize, DWORD timestamp, int nEof)
{
	counter[3]++;
	//	fwrite(pDecodeData,1,nDataSize,fp3);
	if(nEof)
	{
		OutputDebugString("thread 4 读包结束！！！\n");
		printf("====== thread 4 读包结束！！！\n");
		//				fclose(fp4);
		cuda_decodeEnd[3] = true;
	}

#if 0
	printf(" \n------ video 4 :%d ------\n\n", counter[3]);
	FILE *file = fopen("../order.txt", "at");
	fprintf(file, " ------ 视频4：%d ------\n", counter[3]);
	fclose(file);
#endif

	ReleaseSemaphore(hand_semaphore[4], 1, NULL);
	Sleep(1);

	return 0;
}

int CudaDecodeNotify4(unsigned char* pDecodeData, int nDataSize, DWORD timestamp, int nEof)
{
	counter[4]++;
	//	fwrite(pDecodeData,1,nDataSize,fp0);
	if(nEof)
	{
		OutputDebugString("thread 5 读包结束！！！\n");
		printf("====== thread 5 读包结束！！！\n");
		//		fclose(fp0);
		cuda_decodeEnd[4] = true;
	}

#if 0
	printf(" \n------ video 1：%d ------\n", counter[0]);
	FILE *file = fopen("../order.txt", "at");
	fprintf(file, " ------ 视频1：%d ------\n", counter[0]);
	fclose(file);
#endif

	ReleaseSemaphore(hand_semaphore[5], 1, NULL);
	Sleep(1);
	return 0;
}

int CudaDecodeNotify5(unsigned char* pDecodeData, int nDataSize, DWORD timestamp, int nEof)
{
	counter[5]++;
	//	fwrite(pDecodeData,1,nDataSize,fp0);
	if(nEof)
	{
		OutputDebugString("thread 6 读包结束！！！\n");
		printf("====== thread 6 读包结束！！！\n");
		//		fclose(fp0);
		cuda_decodeEnd[5] = true;
	}

#if 0
	printf(" \n------ video 1：%d ------\n", counter[0]);
	FILE *file = fopen("../order.txt", "at");
	fprintf(file, " ------ 视频1：%d ------\n", counter[0]);
	fclose(file);
#endif

	ReleaseSemaphore(hand_semaphore[0], 1, NULL);
	Sleep(1);
	return 0;
}