feat 优化edge和blob耗时

AIEngineModule/include/AIModel_Impl.h

@@ -12,6 +12,7 @@
 #include <vector>
 #include <thread>
 #include <mutex>
+#include <atomic>
 #include "NvInfer.h"
 #include "cuda_runtime_api.h"
 #include "AI_Factory.h"
@@ -213,8 +214,8 @@ private:
     int m_nALLStreamNum;
 
 private:
-    // 上次使用的GPU stream Idx;
-    int m_nLast_GPUStreamIdx;
+    // 上次使用的GPU stream Idx（多线程并发调用，需原子自增避免竞争）;
+    std::atomic<int> m_nLast_GPUStreamIdx;
 };
 
 #endif

AIEngineModule/src/AIModel_Impl.cpp

@@ -20,7 +20,7 @@ AIModel_Impl::AIModel_Impl()
     m_pNode_output_1 = NULL;
     m_pNode_output_2 = NULL;
     m_DetGPUStream.clear();
-    m_nLast_GPUStreamIdx = 0;
+    m_nLast_GPUStreamIdx.store(0);
     m_nALLStreamNum = 0;
 }
 AIModel_Impl::~AIModel_Impl()
@@ -492,7 +492,11 @@ int AIModel_Impl::AI_Det_In_1_Out_1(Node_Config *pConfig_in, Node_Config *pConfi
 {
     // printf("=== s1 ");
     std::shared_ptr<Det_GPU_Stram> pdetStream;
-    GetStream(pdetStream);
+    if (GetStream(pdetStream) != 0 || !pdetStream)
+    {
+        printf("AI_Det_In_1_Out_1: GetStream error \n");
+        return 3;
+    }
     //  printf("=== s2 ");
     std::lock_guard<std::mutex> lock(pdetStream->AI_mutex);
     // printf("  ss g %d s %d --  ", pdetStream->nGPUIdx, pdetStream->cuda_stream->nstreamIdx);
@@ -521,22 +525,40 @@ int AIModel_Impl::AI_Det_In_1_Out_1(Node_Config *pConfig_in, Node_Config *pConfi
 
 int AIModel_Impl::GetStream(std::shared_ptr<Det_GPU_Stram> &pdetStream)
 {
-    int sidx = m_nLast_GPUStreamIdx;
+    if (m_nALLStreamNum <= 0 || m_DetGPUStream.empty())
+    {
+        return 1;
+    }
 
-    sidx++;
-    if (sidx >= m_nALLStreamNum)
+    // 多线程并发时原子轮询分配 stream：
+    int sidx = m_nLast_GPUStreamIdx.load();
+    while (true)
     {
-        sidx = 0;
+        int next = sidx + 1;
+        if (next >= m_nALLStreamNum)
+        {
+            next = 0;
+        }
+        if (m_nLast_GPUStreamIdx.compare_exchange_weak(sidx, next))
+        {
+            sidx = next;
+            break;
+        }
+        // CAS 失败时 sidx 已被更新为当前实际值，重新计算 next
     }
-    pdetStream = m_DetGPUStream.at(sidx);
-    m_nLast_GPUStreamIdx = sidx;
 
+    pdetStream = m_DetGPUStream.at(sidx);
     return 0;
 }
 int AIModel_Impl::AI_Det_In_1_Out_1_class(unsigned char *p_indata_0, float *fmaxScore)
 {
     std::shared_ptr<Det_GPU_Stram> pdetStream;
-    GetStream(pdetStream);
+    if (GetStream(pdetStream) != 0 || !pdetStream)
+    {
+        printf("AI_Det_In_1_Out_1_class: GetStream error \n");
+        *fmaxScore = 0.0f;
+        return -3;
+    }
     //  printf("=== s2 ");
     std::lock_guard<std::mutex> lock(pdetStream->AI_mutex);
     // printf("  ss g %d s %d --  ", pdetStream->nGPUIdx, pdetStream->cuda_stream->nstreamIdx);

AIEngineModule/src/AI_Factory.cpp

@@ -45,7 +45,7 @@ int AIFactory::InitALLAIModle(GPU_Config gupconfig)
         edge_config.strPath = "/home/aidlux/BOE/UseModel_ZB/BM/DefectLight.engine";
         edge_config.strName = "Defect";
         edge_config.inputType = AIModel_Base::Input_CHW;
-        edge_config.Stream_num = 2;
+        edge_config.Stream_num = 4;
         Defect->Init(edge_config);
     }
     if (!Class)
@@ -67,7 +67,7 @@ int AIFactory::InitALLAIModle(GPU_Config gupconfig)
         edge_config.strPath = "/home/aidlux/BOE/UseModel_ZB/BM/BadLabel.engine";
         edge_config.strName = "Tag_Loc";
         edge_config.inputType = AIModel_Base::Input_CHW;
-        edge_config.Stream_num = 2;
+        edge_config.Stream_num = 4;
         Tag_Loc->Init(edge_config);
     }
     if (!Align_Outer)
@@ -78,6 +78,7 @@ int AIFactory::InitALLAIModle(GPU_Config gupconfig)
         edge_config.strPath = "/home/aidlux/BOE/UseModel_ZB/BM/ProductLoc.engine";
         edge_config.strName = "Align_Outer";
         edge_config.inputType = AIModel_Base::Input_CHW;
+        edge_config.Stream_num = 2;
         Align_Outer->Init(edge_config);
     }
     if (!Align_Inner)
@@ -88,6 +89,7 @@ int AIFactory::InitALLAIModle(GPU_Config gupconfig)
         edge_config.strPath = "/home/aidlux/BOE/UseModel_ZB/BM/ClampLoc.engine";
         edge_config.strName = "Align_Inner";
         edge_config.inputType = AIModel_Base::Input_CHW;
+        edge_config.Stream_num = 2;
         Align_Inner->Init(edge_config);
     }
     m_bInitSucc = true;

AlgorithmModule/src/ALLImgCheckAnalysisy.cpp

@@ -450,7 +450,8 @@ int ALLImgCheckAnalysisy::InitData()
 
 int ALLImgCheckAnalysisy::Det_Product(std::shared_ptr<Product> &product)
 {
-
+	string cur_time_s = CheckUtil::getCurTimeHMS();
+	printf("[%s]>>>>>>>>>>>>>>>Det_Product****************det Start************\n", cur_time_s.c_str());
 	// 处理每个相机
 	while (true)
 	{

AlgorithmModule/src/CameraCheckAnalysisy.cpp

@@ -10,6 +10,7 @@
 #include "ImgCheckAnalysisy.hpp"
 #include "CheckUtil.hpp"
 #include "Define.h"
+#include <atomic>
 
 CameraCheckAnalysisy::CameraCheckAnalysisy()
 {
@@ -675,10 +676,18 @@ int CameraCheckAnalysisy::InitRun()
 int CameraCheckAnalysisy::InitCheckAnalysisy()
 {
 	int re = 0;
+	// 全局递增的 CPU 核偏移：
+	static std::atomic<int> s_cpuOffset{0};
 	for (int i = 0; i < IMGCHECKANALYSISY_NUM; i++)
 	{
 		RunInfoST RunConfig;
 		RunConfig.nThreadIdx = i;
+		{
+			int range = (m_RunConfig.nCpu_num > 0) ? m_RunConfig.nCpu_num : 1;
+			int offset = s_cpuOffset.fetch_add(1) % range;
+			RunConfig.nCpu_start_Idx = m_RunConfig.nCpu_start_Idx + offset;
+		}
+		RunConfig.nCpu_num = 1;
 
 		m_pImgCheckAnalysisy[i] = ImgCheckBase::GetInstance();
 		re = m_pImgCheckAnalysisy[i]->UpdateConfig((void *)&RunConfig, CHECK_CONFIG_Run);

example/deal.cpp

@@ -371,28 +371,20 @@ int deal::preCheck()
     setReadThreadStart();
     // 遍历所有图片 开始读图处理
     AllImgNum = 0;
+    // 每张图对应一个通道，独立读入送检；多张图（多通道）分别处理，不能合并为 AB 双相机
     for (const auto pcam : product->camera_list)
     {
-        std::shared_ptr<JC_IMAGE_INFO_> tem = std::make_shared<JC_IMAGE_INFO_>();
-        tem->strCamID = pcam->strCamName;
         for (const auto pimage : pcam->image_list)
         {
+            std::shared_ptr<JC_IMAGE_INFO_> tem = std::make_shared<JC_IMAGE_INFO_>();
+            tem->strCamID = pcam->strCamName;
             tem->strchannelName = pimage->strchannelName;
             tem->strName = pimage->strName;
             tem->strProductID = pimage->strProductID;
-
-            if (tem->strPath == "")
-            {
             tem->strPath = pimage->strPath;
-            }
-            else
-            {
-                tem->strPath_B = pimage->strPath;
-            }
-        }
             InsertReadImgInfo(tem);
             AllImgNum++;
-        // break;
+        }
     }
     // AllImgNum = 1;
     {
@@ -1023,14 +1015,18 @@ int deal::SendImgToCheck(std::shared_ptr<JC_IMAGE_INFO_> pDetImageInfo, IN_IMG_S
         return -1;
     }
     if (status == IN_IMG_Status_End)
+    {
+        // 一张图上下拆分成两个产品（_0 上半、_1 下半），给两个产品分别发送结束标志
+        for (int pi = 0; pi < 2; pi++)
         {
             std::shared_ptr<shareImage> tem = std::make_shared<shareImage>();
-        tem->strImgProductID = pDetImageInfo->strProductID;
+            tem->strImgProductID = pDetImageInfo->strProductID + "_" + std::to_string(pi);
             tem->Status = -1;
             if (m_pALLImgCheckAnalysisy)
             {
                 m_pALLImgCheckAnalysisy->SetDataRun_SharePtr(tem);
             }
+        }
         return 0;
     }
 
@@ -1100,7 +1096,8 @@ int deal::SendImgToCheck(std::shared_ptr<JC_IMAGE_INFO_> pDetImageInfo, IN_IMG_S
             *sub = *tem;
             sub->img = splitList[i];
             sub->strImgProductID = tem->strImgProductID + "_" + std::to_string(i);
-            sub->Status = IN_IMG_Status_OneImg;
+            // 同一产品（上半/下半）可能包含多个通道的图，子图作为中间图，待所有通道送完后再统一发送结束标志
+            sub->Status = IN_IMG_Status_Other;
             if (m_pALLImgCheckAnalysisy)
             {
                 re = m_pALLImgCheckAnalysisy->SetDataRun_SharePtr(sub);