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()
@@ -169,6 +169,19 @@ int AIModel_Impl::Init(AIModelRun_Config config)
     }
 
     m_bInitSuccess = true;
+
+    // warm-up：首次推理会触发 TensorRT/CUDA 懒初始化（kernel 加载、workspace 分配、首次 enqueueV3 建图等），
+    // 提前用一张空图预热，避免首张检测图耗时飙升。
+    {
+        cv::Mat warmIn = InitMat(m_pNode_input_0->channel, m_pNode_input_0->width, m_pNode_input_0->height);
+        cv::Mat warmOut;
+        int warmRe = AIDet(warmIn, warmOut);
+        if (warmRe != 0)
+        {
+            printf("%s warm-up error = %d\n", m_modelRun_Config.strName.c_str(), warmRe);
+        }
+    }
+
     return 0;
 }
 cv::Mat AIModel_Impl::InitMat(int channel, int w, int h)
@@ -479,7 +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);
@@ -508,22 +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

@@ -89,6 +89,7 @@ int AIFactory::InitALLAIModle(GPU_Config gupconfig)
         edge_config.strPath = "/home/aidlux/BOE/UseModel_ZB/KK/AlignOuter.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)
@@ -99,6 +100,7 @@ int AIFactory::InitALLAIModle(GPU_Config gupconfig)
         edge_config.strPath = "/home/aidlux/BOE/UseModel_ZB/KK/AlignInner.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/CameraCheckAnalysisy.cpp

@@ -10,6 +10,7 @@
 #include "ImgCheckAnalysisy.hpp"
 #include "CheckUtil.hpp"
 #include "Define.h"
+#include <atomic>
 
 CameraCheckAnalysisy::CameraCheckAnalysisy()
 {
@@ -676,10 +677,18 @@ 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);

AlgorithmModule/src/ImageResultJudge.cpp

@@ -559,9 +559,7 @@ int ImageResultJudge::ResultJudge(std::shared_ptr<ImageAllResult> pImageResult)
 
 				GetAIDetImg(pImageResult, pCenter, tem.AI_in_Img, tem.AI_out_img);
 				m_CheckResult_shareP->qxImageResult.push_back(tem);
-				pQxLog->bPrintStr = true;
 				pQxLog->AddCheckstr(PrintLevel_4, DET_LOG_LEVEL_3, " result ", " ---name: %s ---code: %s  ---srcImgroi.x: %d  ---srcImgroi.y: %d\n", tem.strTypeName.c_str(), tem.qx_Code.c_str(), tem.srcImgroi.x, tem.srcImgroi.y);
-				pQxLog->bPrintStr = false;
 			}
 			else
 			{

AlgorithmModule/src/ImgCheckAnalysisy.cpp

@@ -427,7 +427,7 @@ int ImgCheckAnalysisy::CheckRun()
 	{
 		m_pdetlog->bPrintStr = true;
 	}
-	m_pdetlog->bPrintStr = true;
+	// m_pdetlog->bPrintStr = true;
 	
 	m_pdetlog->AddCheckstr(PrintLevel_0, "1、basic Info", "---------------------------1、basic Info---------------------------------");
 	m_pdetlog->AddCheckstr(PrintLevel_0, "Version", "%s", GetVersion().c_str());