优化显存占用。当前在3080显卡上支持106路解码

Hu Chunming
1 parent 48330793
Showing 7 changed files with 242 additions and 22 deletions
ffmpeg-4.2.2/libavcodec/cuviddec.c
src/FFCuContextManager.cpp
src/FFCuContextManager.h
src/FFNvDecoder.cpp
src/FFNvDecoderManager.cpp
src/FFNvDecoderManager.h
src/main.cpp
@@ -1012,7 +1012,7 @@ static av_cold int cuvid_decode_init(AVCodecContext *avctx)
     }
     ctx->cuparseinfo.ulMaxNumDecodeSurfaces = ctx->nb_surfaces;
-    ctx->cuparseinfo.ulMaxDisplayDelay = 4;
+    ctx->cuparseinfo.ulMaxDisplayDelay = 2;
     ctx->cuparseinfo.pUserData = avctx;
     ctx->cuparseinfo.pfnSequenceCallback = cuvid_handle_video_sequence;
     ctx->cuparseinfo.pfnDecodePicture = cuvid_handle_picture_decode;
+#include "FFCuContextManager.h"
+#include <iostream>
+
+using namespace std;
+
+FFCuContextManager::~FFCuContextManager()
+{
+    for(auto iter = ctxMap.begin(); iter != ctxMap.end(); iter++){
+        av_buffer_unref(&iter->second);
+    }
+    ctxMap.clear();
+}
+
+AVBufferRef *FFCuContextManager::getCuCtx(string gpuid)
+{
+     AVBufferRef *hw_device_ctx = ctxMap[gpuid];
+     if (nullptr == hw_device_ctx)
+     {
+        // 初始化硬件解码器
+        if (av_hwdevice_ctx_create(&hw_device_ctx, AV_HWDEVICE_TYPE_CUDA, gpuid.c_str(), nullptr, 0) < 0) 
+        {
+            cout << "Failed to create specified HW device.";
+            return nullptr;
+        }
+        ctxMap[gpuid] = hw_device_ctx;
+     }
+     return hw_device_ctx;
+}
 \ No newline at end of file
+
+#include<map>
+#include<string>
+
+extern "C"
+{
+	#include <libavcodec/avcodec.h> 
+	#include <libavdevice/avdevice.h> 
+	#include <libavformat/avformat.h> 
+	#include <libavfilter/avfilter.h> 
+	#include <libavutil/avutil.h> 
+    #include <libavutil/pixdesc.h> 
+	#include <libswscale/swscale.h>
+}
+
+using namespace std;
+
+class FFCuContextManager{
+public:
+    static FFCuContextManager* getInstance(){
+		static FFCuContextManager* singleton = nullptr;
+		if (singleton == nullptr){
+			singleton = new FFCuContextManager();
+		}
+		return singleton;
+	}
+
+    AVBufferRef *getCuCtx(string gpuid);
+
+private:
+    FFCuContextManager(){}
+	~FFCuContextManager();
+
+private:
+    map<string,AVBufferRef *> ctxMap;
+
+};
 \ No newline at end of file
@@ -5,6 +5,8 @@
 #include <thread>
 #include <fstream>
+#include "FFCuContextManager.h"
+
 using namespace std;
 // 参考博客： https://blog.csdn.net/qq_40116098/article/details/120704340
@@ -67,7 +69,8 @@ bool FFNvDecoder::init(FFDecConfig&amp; cfg)
 bool FFNvDecoder::init(const char* uri, const char* gpuid, bool force_tcp)
 {
-	av_register_all();
+	// av_log_set_level(AV_LOG_DEBUG);
+
 	avformat_network_init();
 	// 打开输入视频文件
@@ -114,9 +117,17 @@ bool FFNvDecoder::init(const char* uri, const char* gpuid, bool force_tcp)
 	hw_pix_fmt = AV_PIX_FMT_CUDA;
+	FFCuContextManager* pCtxMgr = FFCuContextManager::getInstance();
+	avctx->hw_device_ctx = av_buffer_ref(pCtxMgr->getCuCtx(gpuid));
+	if (nullptr == avctx->hw_device_ctx)
+	{
+		return false;
+	}
+
 	// 打开解码器流
 	AVDictionary *op = nullptr;
 	av_dict_set( &op, "gpu", gpuid, 0 );
+	av_dict_set( &op, "surfaces", "3", 0 );
 	if (avcodec_open2(avctx, vcodec, &op) < 0) {
 		cout << "Failed to open codec for stream" << stream_index;
 		return false;
@@ -231,10 +242,6 @@ void FFNvDecoder::decode_finished()
 {
 	if (avctx)
 	{
-		if (avctx->hw_device_ctx)
-		{
-			av_buffer_unref(&avctx->hw_device_ctx);
-		}
 		avcodec_free_context(&avctx);
 	}
@@ -112,38 +112,52 @@ bool FFNvDecoderManager::closeDecoderByName(string name){
         return false;
     }
+    m_mutex_erase.lock();
     auto dec = decoderMap.find(name);
     if (dec != decoderMap.end())
     {
         dec->second->close();
         delete dec->second;
+        dec->second = nullptr;
         decoderMap.erase(dec);
+
+        m_mutex_erase.unlock();
         return true;
     }
-
+    m_mutex_erase.unlock();
     cout << "没有找到name为" << name << "的解码器!" << endl;
     return false;
 }
 void FFNvDecoderManager::closeAllDecoder()
 {
+    m_mutex_erase.lock();
     for(auto iter = decoderMap.begin(); iter != decoderMap.end(); iter++){
         iter->second->close();
         delete iter->second;
+        iter->second = nullptr;
     }
     decoderMap.clear();
+    m_mutex_erase.unlock();
 }
 void FFNvDecoderManager::closeAllFinishedDecoder()
 {
-     for(auto iter = decoderMap.begin(); iter != decoderMap.end(); iter++){
+    m_mutex_erase.lock();
+     for(auto iter = decoderMap.begin(); iter != decoderMap.end(); ){
         if (iter->second->isFinished())
         {
             delete iter->second;
-            decoderMap.erase(iter);
+            iter->second = nullptr;
+            iter = decoderMap.erase(iter);
+        }
+        else
+        {
+            iter++ ;
         }
     }
+    m_mutex_erase.unlock();
 }
 int FFNvDecoderManager::count()
@@ -3,6 +3,8 @@
 #include<vector>
 #include<map>
+#include <mutex>
+
 using namespace std;
 struct MgrDecConfig
@@ -55,4 +57,6 @@ private:
 private:
     map<string, FFNvDecoder*> decoderMap;
+
+    mutex m_mutex_erase;
 };
 \ No newline at end of file
@@ -5,6 +5,11 @@
 #include "NvJpegEncoder.h"
+#include <pthread.h>
+#include <thread>
+
+#include <chrono>
+
 unsigned char *pHwRgb = nullptr;
 /**
@@ -14,15 +19,21 @@ void postDecoded(const void * userPtr, AVFrame * gpuFrame){
     FFNvDecoder* decoder = (FFNvDecoder*)userPtr;
     if (decoder!= nullptr)
     {
-        cout << "decode name: " << decoder->getName() << endl;
+        // cout << "decode name: " << decoder->getName() << endl;
+
+        if (decoder->getName() == "dec1")
+        {
+            /* code */
+        }
+        
         // const char* gpu_pixfmt = av_get_pix_fmt_name((AVPixelFormat)gpuFrame->format);
         // cout << "pixfmt: " << gpu_pixfmt << endl;
-        cout << "keyframe: " << gpuFrame->key_frame << " width: " << gpuFrame->width << " height: "<< gpuFrame->height << endl;
+        // cout << "keyframe: " << gpuFrame->key_frame << " width: " << gpuFrame->width << " height: "<< gpuFrame->height << endl;
         // cout << "decode successed ✿✿ヽ(°▽°)ノ✿ " << endl;
         if (gpuFrame->format == AV_PIX_FMT_CUDA)
         {   
-            cout << "gpuid = " << atoi(decoder->m_cfg.gpuid.c_str()) << endl;
+            // cout << "gpuid = " << atoi(decoder->m_cfg.gpuid.c_str()) << endl;
             // cudaSetDevice(atoi(decoder->m_cfg.gpuid.c_str()));
             // cudaError_t cudaStatus;
             // if(pHwRgb == nullptr){
@@ -42,14 +53,75 @@ void postDecoded(const void * userPtr, AVFrame * gpuFrame){
     }
 }
+long start_time = 0;
+long end_time = 0;
+bool count_flag = false;
+int count = 0;
+int count_std = 100;
+
+long long get_cur_time(){
+    // 获取操作系统当前时间点（精确到微秒）
+    chrono::time_point<chrono::system_clock, chrono::microseconds> tpMicro
+        = chrono::time_point_cast<chrono::microseconds>(chrono::system_clock::now());
+    // (微秒精度的)时间点 => (微秒精度的)时间戳
+    time_t totalMicroSeconds = tpMicro.time_since_epoch().count();
+
+	long long currentTime = ((long long)totalMicroSeconds)/1000;
+
+    return currentTime;
+}
+
+int sum = 0;
+void postDecoded0(const void * userPtr, AVFrame * gpuFrame){
+    FFNvDecoder* decoder = (FFNvDecoder*)userPtr;
+    if (decoder!= nullptr)
+    {
+        // cout << "decode name: " << decoder->getName() << endl;
+        if (decoder->getName() == "dec")
+        {
+            if (! count_flag)
+            {
+                count_flag = true;
+                count = 0;
+                end_time = start_time = get_cur_time();
+            }
+            count++;
+            sum ++ ;
+            if (count >= count_std)
+            {
+                end_time = get_cur_time();
+                long time_using = end_time - start_time;
+                double time_per_frame = double(time_using)/count_std ;
+                cout << count_std << "帧用时:" << time_using << "ms 每帧用时：" << time_per_frame << "ms" << endl;
+                cout << "keyframe: " << gpuFrame->key_frame << " width: " << gpuFrame->width << " height: "<< gpuFrame->height << endl;
+                cout << gpuFrame->pts << endl;
+
+                count_flag = false;
+            }
+        }
+    }
+}
+
+// string test_uri = "rtmp://192.168.10.56:1935/objecteye/1";
+string test_uri = "/home/cmhu/data/test.mp4";
+
 void createDecode(int index){
     FFNvDecoderManager* pDecManager = FFNvDecoderManager::getInstance();
     MgrDecConfig config;
     config.name = "dec" + to_string(index);
-    config.cfg.uri = "rtsp://176.10.0.4:8554/stream";
+    config.cfg.uri = test_uri;
     config.cfg.post_decoded_cbk = postDecoded;
     config.cfg.force_tcp = true;
-    config.cfg.gpuid = "1";
+
+    if (index % 2 == 0)
+    {
+        config.cfg.gpuid = "2";
+    }
+    else
+    {
+        config.cfg.gpuid = "1";
+    }
+    
     FFNvDecoder* decoder = pDecManager->createDecoder(config);
     if (!decoder)
     {
@@ -59,21 +131,57 @@ void createDecode(int index){
     pDecManager->startDecodeByName(config.name);
 }
+#define checkCudaErrors(S) do {CUresult  status; \
+        status = S; \
+        if (status != CUDA_SUCCESS ) std::cout << __LINE__ <<" checkCudaErrors - status = " << status << std::endl; \
+        } while (false)
+
+int CheckCUDAProperty( int devId )
+{
+    cuInit(0);
+
+	CUdevice dev = devId;
+	size_t memSize = 0;
+	char devName[256] = {0};
+	int major = 0, minor = 0;
+	CUresult rlt = CUDA_SUCCESS;
+
+	rlt = cuDeviceComputeCapability( &major, &minor, dev );
+	checkCudaErrors( rlt );
+
+	rlt = cuDeviceGetName( devName, sizeof( devName ), dev );
+	checkCudaErrors( rlt );
+
+	printf( "Using GPU Device %d: %s has SM %d.%d compute capability\n",
+		    dev, devName, major, minor );
+
+	rlt = cuDeviceTotalMem( &memSize, dev );
+	checkCudaErrors( rlt );
+
+	printf( "Total amount of global memory:   %4.4f MB\n",
+		   (float)memSize / ( 1024 * 1024 ) );
+
+	return 0;
+}
+
 int main(){
+    CheckCUDAProperty(1);
+
     FFNvDecoderManager* pDecManager = FFNvDecoderManager::getInstance();
-    // for (size_t i = 0; i < 20; i++)
-    // {
-    //     createDecode(i);
-    // }
+    int count = 105;
+    for (size_t i = 0; i < count ; i++)
+    {
+        createDecode(i);
+    }
     MgrDecConfig config;
-    config.name = "dec2";
-    config.cfg.uri = "/home/cmhu/data/test.mp4";
-     config.cfg.post_decoded_cbk = postDecoded;
+    config.name = "dec";
+    config.cfg.uri = test_uri;
+    config.cfg.post_decoded_cbk = postDecoded0;
     config.cfg.force_tcp = true;
-    config.cfg.gpuid = "2";
+    config.cfg.gpuid = "1";
     FFNvDecoder* dec2 = pDecManager->createDecoder(config);
     if (!dec2)
     {
@@ -82,6 +190,26 @@ int main(){
     pDecManager->setUserPtr(config.name, dec2);
     pDecManager->startDecodeByName(config.name);
+    pthread_t m_decode_thread;
+    pthread_create(&m_decode_thread,0,
+        [](void* arg)
+        {
+            while (true)
+            {
+                std::this_thread::sleep_for(std::chrono::milliseconds(5000));
+                FFNvDecoderManager* pDecManager = FFNvDecoderManager::getInstance();
+                int count = pDecManager->count();
+                cout << "当前运行路数： " << pDecManager->count() << endl;
+                if (count <= 0)
+                {
+                    break;
+                }
+            }  
+
+            return (void*)0;
+        }
+    ,nullptr);
+
     // config.name = "dec0";
     // config.cfg.uri = "rtmp://192.168.10.56:1935/objecteye/1";
@@ -116,6 +244,8 @@ int main(){
     // // pDecManager->resumeDecoder("dec1");
     // pDecManager->resumeDecoder("dec2");
+    cout << "总共帧数：" << sum << endl;
+
     while (getchar() != 'q');
     pDecManager->closeAllDecoder();