根据解码用时优化数据送给解码器的频率，避免解码器内部占用过多的显存

Hu Chunming
1 parent 08c1b61e
Showing 3 changed files with 43 additions and 14 deletions
src/decoder/dvpp/DvppDecoder.cpp
src/decoder/dvpp/DvppDecoder.h
src/demo/demo.cpp
@@ -16,7 +16,7 @@
 struct Vdec_CallBack_UserData {
     uint64_t frameId;
-    unsigned long long frame_nb;
+    uint64_t frame_nb;
     long startTime;
     long sendTime;
 	DvppDecoder* self;
@@ -49,7 +49,7 @@ static void VdecCallback(acldvppStreamDesc *input, acldvppPicDesc *output, void 
     if(nullptr != userData){
         DvppDecoder* self = userData->self;
         if(self != nullptr){
-            self->doVdppVdecCallBack(input, output, userData->frame_nb);
+            self->doVdppVdecCallBack(input, output, userData);
         }
         delete userData;
 	    userData = nullptr;
@@ -387,8 +387,8 @@ DeviceMemory* DvppDecoder::snapshot(){
         snapshot_mem = new DvppDataMemory(mem);
         m_decoded_data_queue_mtx.unlock();
-        snap_count++;
-        LOG_INFO("[{}]- snap_count:{} ", m_dec_name, snap_count);
+        // snap_count++;
+        // LOG_INFO("[{}]- snap_count:{} ", m_dec_name, snap_count);
         break;
     }
@@ -518,6 +518,12 @@ void DvppDecoder::read_thread() {
                     m_recoderManager.cache_pkt(pkt, frame_nb);
     #endif
                     nSended = sendPkt(vdecChannelDesc, pkt, frame_nb);
+
+                    if(!m_bReal && frame_nb > 20 && m_avg_decode_time > 0) 
+                    {
+                        // 针对文件，根据解码时间做延时，避免占用过多显存
+                        std::this_thread::sleep_for(std::chrono::milliseconds(10));
+                    }
                 }
                 if(nSended < 0) {
@@ -542,8 +548,6 @@ void DvppDecoder::read_thread() {
     } while (0);
-    LOG_INFO("[{}]- release.", m_dec_name);
-
     if (vdecChannelDesc) {
         CHECK_NOT_RETURN(aclvdecDestroyChannel(vdecChannelDesc), "aclvdecDestroyChannel failed");
         CHECK_NOT_RETURN(aclvdecDestroyChannelDesc(vdecChannelDesc), "aclvdecDestroyChannelDesc failed");
@@ -691,10 +695,31 @@ void DvppDecoder::doProcessReport(){
     LOG_INFO("doProcessReport exit.");
 }
-void DvppDecoder::doVdppVdecCallBack(acldvppStreamDesc *input, acldvppPicDesc *output, unsigned long long frame_nb){
+void DvppDecoder::doVdppVdecCallBack(acldvppStreamDesc *input, acldvppPicDesc *output, void *pUserData){
+
+    if(nullptr == pUserData){
+        return;
+    }
+
+    Vdec_CallBack_UserData *userData = (Vdec_CallBack_UserData *) pUserData;
+    uint64_t frame_nb = userData->frame_nb;
     m_out_count++;
+    if (m_out_count % 20 == 0)
+    {
+        m_decode_20_time -= m_avg_decode_time*20;   // 减去等待用的时间就是实际解码用时
+        m_avg_decode_time = m_decode_20_time / 20;
+        // LOG_INFO("[{}]- m_avg_decode_time: {}", m_dec_name, m_avg_decode_time);
+        if (m_avg_decode_time <= 0) {
+            m_avg_decode_time = 1;
+        }
+        m_decode_20_time = 0;
+    }
+    m_decode_20_time += UtilTools::get_cur_time_ms() - userData->sendTime; 
+    // LOG_INFO("[{}]- m_decode_20_time: {}", m_dec_name, m_decode_20_time);
+
+
     CHECK_AND_RETURN_NOVALUE(aclrtSetCurrentContext(m_context), "aclrtSetCurrentContext failed");
     void *inputDataDev = acldvppGetStreamDescData(input);
@@ -59,7 +59,7 @@ public:
     void set_mq_callback(mq_callback_t cb);
 public:
-    void doVdppVdecCallBack(acldvppStreamDesc *input, acldvppPicDesc *output, unsigned long long frame_nb);
+    void doVdppVdecCallBack(acldvppStreamDesc *input, acldvppPicDesc *output, void *pUserData);
     void doProcessReport();
 private:
@@ -124,6 +124,9 @@ private:
     long long last_ts {0};
-    int m_in_count {0};
-    int m_out_count {0};
+    uint64_t m_in_count {0};
+    uint64_t m_out_count {0};
+
+    int m_avg_decode_time{1};
+    int m_decode_20_time{0};
 };
 \ No newline at end of file
@@ -63,9 +63,10 @@ void init_mq_conn(void *handle) {
 #endif // #ifdef POST_USE_RABBITMQ
 void set_task_params(task_param &tparam, const unsigned &idx, const algorithm_type_t &algor_type) {
-  switch (algor_type) {
   auto algor_init_params = new algor_init_config_param_t;
+  switch (algor_type) {
+  
   case algorithm_type_t::NONMOTOR_VEHICLE_NOHELMET: {
     auto algor_params = new algor_config_param_manned_incident;
     {
@@ -1135,7 +1136,7 @@ void test_gpu(int gpuID){
     // createTask(handle, algor_vec, 20, false); 
     // createTask(handle, algor_vec, 20, false); 
     // createTask(handle, algor_vec, 20, false); 
-    // createTask(handle, algor_vec, 0);  
+    createTask(handle, algor_vec, 0);  
     // createTask(handle, algor_vec, 1);  
     // createTask(handle, algor_vec, 2);  
     // createTask(handle, algor_vec, 3);  
@@ -1189,9 +1190,9 @@ void test_gpu(int gpuID){
     switch (ch)
     {
     case 'a':
-      createTask(handle, algor_vec3, 4, false);  
+      // createTask(handle, algor_vec3, 4, false);  
       // createTask(handle, algor_vec3, 5, false);  
-      createTask(handle, algor_vec3, 6, false);  
+      // createTask(handle, algor_vec3, 6, false);  
       // createTask(handle, algor_vec3, 7, false);  
       // createTask(handle, algor_vec3, 8, false);  
       break;