Fix ANSOCR TensorRT Release

.claude/settings.local.json

@@ -2,7 +2,10 @@
   "permissions": {
     "allow": [
       "Bash(find C:ProjectsCLionProjectsANSCORE -type f \\\\\\(-name *.h -o -name *.hpp -o -name *.cpp -o -name *.cc \\\\\\))",
-      "Bash(xargs grep:*)"
+      "Bash(xargs grep:*)",
+      "Bash(grep -n \"~Engine\\\\|TRTEngineCache::release\\\\|EnginePoolManager::release\\\\|destructor\" /c/Projects/CLionProjects/ANSCORE/engines/TensorRTAPI/include/engine/*.inl /c/Projects/CLionProjects/ANSCORE/modules/ANSODEngine/*.h)",
+      "Bash(grep -l \"EnginePoolManager\" /c/Projects/CLionProjects/ANSCORE/modules/ANSODEngine/*.cpp)",
+      "Bash(grep -n \"g_processExiting\" /c/Projects/CLionProjects/ANSCORE/engines/TensorRTAPI/include/engine/*.h /c/Projects/CLionProjects/ANSCORE/modules/ANSODEngine/engine.h)"
     ]
   }
 }

engines/TensorRTAPI/include/engine/EnginePoolManager.h

@@ -323,7 +323,19 @@ private:
             // unreliable.  Don't destroy Engine objects (their destructors
             // call cudaFree, thread::join, etc. which deadlock or crash).
             // The OS reclaims all memory, VRAM, and handles at process exit.
+            //
+            // Intentionally leak Engine shared_ptrs: after the explicit dtor
+            // body returns, the compiler still runs implicit member dtors for
+            // m_pools.  That would destroy shared_ptr<Engine<T>>, triggering
+            // TRT/CUDA cleanup on a dead context.  Detach the shared_ptrs
+            // first so the map destructor only frees empty entries.
             m_sweeperRunning.store(false);
+            // Leak Engine objects: bump refcount so shared_ptr dtor won't
+            // actually delete them when m_pools is implicitly destroyed.
+            for (auto& [_, entry] : m_pools) {
+                auto* leaked = new std::shared_ptr<Engine<T>>(std::move(entry.engine));
+                (void)leaked;  // intentional leak — OS reclaims at exit
+            }
             return;
         }
         // Normal FreeLibrary path: threads are alive, safe to clean up.

engines/TensorRTAPI/include/engine/TRTEngineCache.h

@@ -150,6 +150,18 @@ public:
 
 private:
     TRTEngineCache() = default;
+    ~TRTEngineCache() {
+        if (g_processExiting().load(std::memory_order_relaxed)) {
+            // ExitProcess path: CUDA context is dead.  Leak ICudaEngine and
+            // IRuntime shared_ptrs so their destructors don't call into a
+            // destroyed CUDA driver.  The OS reclaims everything at exit.
+            for (auto& [_, entry] : m_cache) {
+                auto* le = new std::shared_ptr<nvinfer1::ICudaEngine>(std::move(entry.engine));
+                auto* lr = new std::shared_ptr<nvinfer1::IRuntime>(std::move(entry.runtime));
+                (void)le; (void)lr;  // intentional leak
+            }
+        }
+    }
     TRTEngineCache(const TRTEngineCache&) = delete;
     TRTEngineCache& operator=(const TRTEngineCache&) = delete;
 

modules/ANSOCR/dllmain.cpp

@@ -8,6 +8,8 @@
 #include "ANSGpuFrameRegistry.h"
 #include <json.hpp>
 #include "NV12PreprocessHelper.h"
+#include "engine/TRTEngineCache.h"
+#include "engine/EnginePoolManager.h"
 #include <unordered_map>
 #include <condition_variable>
 #include <cstdint>
@@ -93,7 +95,32 @@ BOOL APIENTRY DllMain(HMODULE hModule,
 	case DLL_PROCESS_ATTACH:
 	case DLL_THREAD_ATTACH:
 	case DLL_THREAD_DETACH:
+		break;
 	case DLL_PROCESS_DETACH:
+		// When lpReserved != NULL, the process is terminating via ExitProcess.
+		// The OS has already killed all worker threads (idle timers, CUDA
+		// threads, etc.).  Set the global flag so atexit destructors skip
+		// thread joins and CUDA/TRT cleanup that would fail on a dead context.
+		if (lpReserved != nullptr) {
+			g_processExiting().store(true, std::memory_order_relaxed);
+			break;
+		}
+
+		// Dynamic FreeLibrary — threads are still alive, safe to clean up.
+		try {
+			std::vector<ANSCENTER::ANSOCRBase*> leakedHandles;
+			{
+				std::lock_guard<std::mutex> lk(OCRHandleRegistryMutex());
+				for (auto& [h, _] : OCRHandleRegistry())
+					leakedHandles.push_back(h);
+				OCRHandleRegistry().clear();
+			}
+			for (auto* h : leakedHandles) {
+				try { h->Destroy(); delete h; } catch (...) {}
+			}
+			try { EnginePoolManager<float>::instance().clearAll(); } catch (...) {}
+			try { TRTEngineCache::instance().clearAll(); } catch (...) {}
+		} catch (...) {}
 		break;
 	}
 	return TRUE;

tests/ANSOCR-UnitTest/ANSOCR-UnitTest.cpp

@@ -304,10 +304,10 @@ int TestOCRv5mage() {
     std::cout << "Current working directory: " << currentPath << std::endl;
     std::string licenseKey = "";
     std::string modelFilePath = "C:\\Projects\\ANSVIS\\Models\\ANS_GenericOCR_v2.0.zip";
-    std::string imagePath = "C:\\Programs\\ModelTraining\\JALPR\\data\\20260329_174127_834.jpg";//"E:\\Programs\\DemoAssets\\Images\\OCR\\ref3_000.bmp";
+    std::string imagePath = "C:\\Programs\\ModelTraining\\JLPD\\data\\0b9b013343f0bd8c7809653dfab16eac_jpeg.rf.1438e1237023ad7a254605942193df99.jpg";//"E:\\Programs\\DemoAssets\\Images\\OCR\\ref3_000.bmp";
 
     int language = 0;  // CUSTOM
-    int engine = 0;// GPU
+    int engine = 1;// GPU
 
     // For high-resolution images with PP-OCRv5 server models, use higher limitSideLen
     // (default 960 downscales large images too aggressively, missing small text)