Fix AMD by changing from GetTensorData<T>() to GetTensorMutableData<T>()

.claude/settings.local.json

@@ -117,7 +117,35 @@
       "Bash(NODE_PATH=\"C:/home/alex/.npm-global/node_modules\" node build_anslib_logging_guide.js)",
       "Bash(python \"C:/Users/nghia/AppData/Roaming/Claude/local-agent-mode-sessions/skills-plugin/d8e35aa4-a14e-4e20-b921-ba1b9a3cce86/cdda7cc8-a1c7-42ff-98b4-473ec3e8b9fb/skills/docx/scripts/office/validate.py\" \"C:/Projects/CLionProjects/ANSCORE/docs/ANSLIB_Logging_Guide.docx\")",
       "Bash(python \"C:/Users/nghia/AppData/Roaming/Claude/local-agent-mode-sessions/skills-plugin/d8e35aa4-a14e-4e20-b921-ba1b9a3cce86/cdda7cc8-a1c7-42ff-98b4-473ec3e8b9fb/skills/docx/scripts/office/soffice.py\" --headless --convert-to pdf --outdir \"C:/Projects/CLionProjects/ANSCORE/docs\" \"C:/Projects/CLionProjects/ANSCORE/docs/ANSLIB_Logging_Guide.docx\")",
-      "Bash(sort -t: -k1 -u)"
+      "Bash(sort -t: -k1 -u)",
+      "Bash(git -C C:/Projects/CLionProjects/ANSCORE diff core/ANSLicensingSystem/ANSLicense.cpp)",
+      "Bash(awk 'NR>=154022 && NR<=156428 {print}' \"C:/Users/nghia/Downloads/AVNET-8845HS2.log\")",
+      "Bash(awk 'NR>=156350 && NR<=156428 {print}' \"C:/Users/nghia/Downloads/AVNET-8845HS2.log\")",
+      "Bash(git -C C:/Projects/CLionProjects/ANSCORE log --oneline -10 -- modules/ANSFR/OpenVINODeviceConfig.h modules/ANSFR/OpenVINODeviceConfig.cpp)",
+      "Bash(git -C C:/Projects/CLionProjects/ANSCORE log --oneline -5 -- core/ANSLibsLoader/EPLoader.cpp)",
+      "Bash(git -C C:/Projects/CLionProjects/ANSCORE show --stat 69787b0)",
+      "Bash(git -C C:/Projects/CLionProjects/ANSCORE show --stat 97d8149)",
+      "Bash(git -C C:/Projects/CLionProjects/ANSCORE show 97d8149 -- engines/ONNXEngine/OpenVINODeviceConfig.h modules/ANSFR/ANSFR.cpp)",
+      "Bash(awk 'NR>=7106 && NR<=7250 && /ONNXYOLO|ANSCUSTOM|DetectObjects/ {print}' \"C:/Users/nghia/Downloads/AVNET-8845HS5.log\")",
+      "Bash(awk 'NR>=5850 && NR<=7115 && /ONNXYOLO|ANSCUSTOM_Infer|ANSONNXCL_pp|cls\\\\] calling/' \"C:/Users/nghia/Downloads/AVNET-8845HS5.log\")",
+      "Bash(git -C C:/Projects/CLionProjects/ANSCORE show --stat 3a527d2)",
+      "Bash(git -C C:/Projects/CLionProjects/ANSCORE show 3a527d2 -- modules/ANSODEngine/ANSCUSTOMDETECTOR.cpp modules/ANSODEngine/ANSONNXYOLO.cpp modules/ANSODEngine/ANSONNXCL.cpp)",
+      "Bash(git -C C:/Projects/CLionProjects/ANSCORE show 3a527d2 -- modules/ANSODEngine/ANSEngineCommon.h)",
+      "Bash(git -C C:/Projects/CLionProjects/ANSCORE diff HEAD modules/ANSODEngine/ANSONNXYOLO.cpp modules/ANSODEngine/ANSONNXCL.cpp)",
+      "Bash(git -C C:/Projects/CLionProjects/ANSCORE diff HEAD modules/ANSODEngine/ANSONNXYOLO.cpp)",
+      "Bash(git -C C:/Projects/CLionProjects/ANSCORE diff HEAD modules/ANSODEngine/ANSONNXCL.cpp)",
+      "Bash(git -C C:/Projects/CLionProjects/ANSCORE log --all --oneline --since='3 weeks ago' -- modules/ANSODEngine/ANSONNXYOLO.cpp modules/ANSODEngine/ANSONNXCL.cpp engines/ONNXEngine/ONNXEngine.cpp engines/ONNXEngine/ONNXEngine.h)",
+      "Bash(git -C C:/Projects/CLionProjects/ANSCORE show --stat 844d739)",
+      "Bash(git -C C:/Projects/CLionProjects/ANSCORE log -p --grep=\"mutex\\\\|lock\\\\|hang\\\\|stall\\\\|concurrent\\\\|thread\" --since=\"3 weeks ago\" -- modules/ANSODEngine/ANSONNXYOLO.cpp)",
+      "Bash(git -C C:/Projects/CLionProjects/ANSCORE show 844d739 -- modules/ANSODEngine/ANSONNXYOLO.cpp)",
+      "Bash(awk '$2>=835 && $2<=860 && /1853990010|ANSCUSTOM|ANSONNXCL|ANSONNXYOLO/' \"C:/Users/nghia/Downloads/AVNET-8845HS7.log\")",
+      "Bash(awk '$2>=836 && $2<=837 && /ANSCUSTOM|ANSONNXCL/' \"C:/Users/nghia/Downloads/AVNET-8845HS7.log\")",
+      "Bash(awk '$2>=836.0 && $2<=836.9' \"C:/Users/nghia/Downloads/AVNET-8845HS7.log\")",
+      "Bash(grep -nB1 -A3 \"DirectML\\\\|DisableMemPattern\\\\|SetExecutionMode\\\\|ExtendedOptimization\\\\|ORT_SEQUENTIAL\" \"C:/Projects/CLionProjects/ANSCORE/engines/ONNXEngine/ONNXEngine.cpp\")",
+      "Bash(grep -nE \"^\\\\}$|warmupModel\\\\\\(|#if 0|#endif|ANSONNXCL_legacy_Init\" C:/Projects/CLionProjects/ANSCORE/modules/ANSODEngine/ANSONNXCL.cpp)",
+      "Bash(git -C C:/Projects/CLionProjects/ANSCORE restore modules/ANSODEngine/ANSONNXCL.h modules/ANSODEngine/ANSONNXCL.cpp)",
+      "Bash(git -C C:/Projects/CLionProjects/ANSCORE status --short modules/)",
+      "Bash(git -C C:/Projects/CLionProjects/ANSCORE diff --stat HEAD modules/ANSODEngine/ANSONNXCL.cpp modules/ANSODEngine/ANSONNXOBB.cpp modules/ANSODEngine/ANSONNXPOSE.cpp modules/ANSODEngine/ANSONNXSEG.cpp modules/ANSODEngine/ANSYOLO12OD.cpp engines/ONNXEngine/ONNXEngine.cpp engines/ONNXEngine/ONNXSAM3.cpp)"
     ]
   }
 }

core/ANSLicensingSystem/ANSLicense.cpp

@@ -80,6 +80,17 @@ namespace {
         return s.substr(0, 4) + std::string(s.size() - 8, '*') + s.substr(s.size() - 4);
     }
 
+    // SDK identifier passed to ANSLSHelper for licensing-service auth. Despite
+    // the variable name "privateKey" used at call sites, this is an 11-byte
+    // SDK token that identifies the ANSCENTER product to the licensing
+    // service — NOT a cryptographic signing key. Centralised here so any
+    // future rotation (or move to runtime lookup via env var / config file)
+    // touches one place instead of duplicated literals scattered across the
+    // file. Narrow form is the source of truth; wide form is derived to
+    // avoid two literals drifting out of sync.
+    inline std::string  GetSdkKey()  { return "AQlSAiRTNtS7X20="; }
+    inline std::wstring GetSdkKeyW() { return String2WString(GetSdkKey()); }
+
     // Append both `\` and `/` slash flavours of `raw` to `dst`, lower-cased
     // and with a trailing separator forced. Empty / pathologically short
     // entries are skipped.
@@ -661,7 +672,7 @@ namespace ANSCENTER
 		try {
 			int _enableFeature;
 			int _productId;
-			std::string privateKey = "AQlSAiRTNtS7X20=";
+			std::string privateKey = GetSdkKey();
 			int activationResult = 0;
 			std::string licensingServiceURL = "https://licensingservice.anscenter.com/";
 			ANSCENTER::ANSLSHelper ansHelper(privateKey, licensingServiceURL);
@@ -775,7 +786,7 @@ namespace ANSCENTER
 		try {
 			int productId;
 			std::string registrationName;
-			std::string  privateKey = "AQlSAiRTNtS7X20=";
+			std::string  privateKey = GetSdkKey();
 			std::string licensingServiceURL = "https://licensingservice.anscenter.com/";
 			ANSCENTER::ANSLSHelper ansHelper(privateKey, licensingServiceURL);
 			ansHelper.SetupLicenseTemplate();
@@ -836,7 +847,7 @@ namespace ANSCENTER
 	int ANSLicenseHelper::DeactivateLicense(std::string productName) {
 		std::string licenseDirectory = GetLicenseDir();
 		std::vector<std::string> licenseKeyFiles = ListFilesInFolder(licenseDirectory);
-		std::string  privateKey = "AQlSAiRTNtS7X20=";
+		std::string  privateKey = GetSdkKey();
 		std::string licensingServiceURL = "https://licensingservice.anscenter.com/";
 		ANSCENTER::ANSLSHelper ansHelper(privateKey, licensingServiceURL);
 		ansHelper.SetupLicenseTemplate();
@@ -901,7 +912,7 @@ namespace ANSCENTER
 		return 0;
 	}
 	int ANSLicenseHelper::ActivateLicense(std::string productName, std::string licenseKey) {
-		std::string  privateKey = "AQlSAiRTNtS7X20=";
+		std::string  privateKey = GetSdkKey();
 		std::string licensingServiceURL = "https://licensingservice.anscenter.com/";
 		ANSCENTER::ANSLSHelper ansHelper(privateKey, licensingServiceURL);
 		ansHelper.SetupLicenseTemplate();
@@ -942,7 +953,7 @@ namespace ANSCENTER
 	}
 
 	int ANSLicenseHelper::ActivateLicenseWithCustomHWID(std::string productName, std::string licenseKey, std::string hwid, std::string &activationKey) {
-		std::string  privateKey = "AQlSAiRTNtS7X20=";
+		std::string  privateKey = GetSdkKey();
 		std::string licensingServiceURL = "https://licensingservice.anscenter.com/";
 		ANSCENTER::ANSLSHelper ansHelper(privateKey, licensingServiceURL);
 		ansHelper.SetupLicenseTemplate();
@@ -1020,7 +1031,7 @@ namespace ANSCENTER
 
 			std::string licenseDirectory = GetLicenseDir();
 			std::vector<std::string> licenseKeyFiles = ListFilesInFolder(licenseDirectory);
-			std::string  privateKey = "AQlSAiRTNtS7X20=";
+			std::string  privateKey = GetSdkKey();
 			std::string licensingServiceURL = "https://licensingservice.anscenter.com/";
 			ANSCENTER::ANSLSHelper ansHelper(privateKey, licensingServiceURL);
 			ansHelper.SetupLicenseTemplate();
@@ -1224,7 +1235,7 @@ namespace ANSCENTER
 	ANSLSHelper::ANSLSHelper()
 	{
 		try {
-			this->_privateKey = _T("AQlSAiRTNtS7X20=");
+			this->_privateKey = GetSdkKeyW();
 			this->_licenseServiceURL = _T("https://licensingservice.anscenter.com/");
 			this->_sdkLicenseKey = _T("MYNSU-GBQ2Q-SF5U5-S3RVF-5ZKFD");
 			SDKRegistration::SetLicenseKey(_sdkLicenseKey.c_str());
@@ -1239,7 +1250,7 @@ namespace ANSCENTER
 	ANSLSHelper::ANSLSHelper(std::string licenseServiceURL)
 	{
 		try {
-			this->_privateKey = _T("AQlSAiRTNtS7X20=");
+			this->_privateKey = GetSdkKeyW();
 			this->_licenseServiceURL = String2WString(licenseServiceURL);
 			this->_sdkLicenseKey = _T("MYNSU-GBQ2Q-SF5U5-S3RVF-5ZKFD");
 			SDKRegistration::SetLicenseKey(_sdkLicenseKey.c_str());
@@ -1255,7 +1266,7 @@ namespace ANSCENTER
 	{
 		try {
 			if (privateKey.empty()) {
-				this->_privateKey = _T("AQlSAiRTNtS7X20=");
+				this->_privateKey = GetSdkKeyW();
 			}
 			else {
 				this->_privateKey = String2WString(privateKey);
@@ -1278,7 +1289,7 @@ namespace ANSCENTER
 	{
 		try {
 			if (privateKey.empty()) {
-				this->_privateKey = _T("AQlSAiRTNtS7X20=");
+				this->_privateKey = GetSdkKeyW();
 			}
 			else {
 				this->_privateKey = String2WString(privateKey);
@@ -1517,7 +1528,7 @@ namespace ANSCENTER
 			this->_licenseTemplate->SetDataSize(36);						//16 bits for ProductId, 4 bits for FeatureBitMask, and 16 bits for ExpirationData (16+4+16 =36)
 			this->_licenseTemplate->SetTemplateId(_T("24880"));
 			if(this->_privateKey.empty())
-				this->_licenseTemplate->SetPrivateKey(_T("AQlSAiRTNtS7X20="));
+				this->_licenseTemplate->SetPrivateKey(GetSdkKeyW().c_str());
 			else this->_licenseTemplate->SetPrivateKey(this->_privateKey.c_str());
 			
 			if (this->_publicKey.empty())

engines/ONNXEngine/ONNXEngine.cpp

@@ -819,7 +819,10 @@ namespace ANSCENTER {
             input_values_handler.clear();
             input_values_handler.shrink_to_fit();
 
-            const float* vals = output_tensors[0].GetTensorData<float>();
+            // GetTensorMutableData on DirectML — the const GetTensorData
+            // triggers a per-call host-readable mapping that on AMD DML
+            // exhausts a staging-buffer pool after ~8 calls and hangs.
+            const float* vals = output_tensors[0].GetTensorMutableData<float>();
             const unsigned int  hidden_dim =
                 static_cast<unsigned int>(output_node_dims.at(0).at(1));
 
@@ -1377,7 +1380,9 @@ namespace ANSCENTER {
             Ort::RunOptions{ nullptr },
             in_names, &input_tensor, 1, out_names, 1);
 
-        out_result = post_processing(outputs[0].GetTensorData<float>());
+        // GetTensorMutableData (not GetTensorData) — see comment in this
+        // file's other output-read sites; const GetTensorData hangs on AMD DML.
+        out_result = post_processing(outputs[0].GetTensorMutableData<float>());
     }
 
     Ort::Value MOVINET::transform(const cv::Mat& mat)

engines/ONNXEngine/ONNXSAM3.cpp

@@ -463,7 +463,9 @@ namespace ANSCENTER
             auto info = outputs[maskIdx].GetTensorTypeAndShapeInfo();
             m_cachedLangMaskShape = info.GetShape();
             size_t count = info.GetElementCount();
-            const bool* data = outputs[maskIdx].GetTensorData<bool>();
+            // GetTensorMutableData not GetTensorData on DML — const variant
+            // hangs after ~8 calls.  Read-only despite the "Mutable" name.
+            const bool* data = outputs[maskIdx].GetTensorMutableData<bool>();
             m_cachedLangMask.resize(count);
             for (size_t i = 0; i < count; ++i)
                 m_cachedLangMask[i] = data[i] ? 1 : 0;
@@ -474,7 +476,7 @@ namespace ANSCENTER
             auto info = outputs[featIdx].GetTensorTypeAndShapeInfo();
             m_cachedLangFeaturesShape = info.GetShape();
             size_t count = info.GetElementCount();
-            const float* data = outputs[featIdx].GetTensorData<float>();
+            const float* data = outputs[featIdx].GetTensorMutableData<float>();
             m_cachedLangFeatures.assign(data, data + count);
         }
 
@@ -649,7 +651,7 @@ namespace ANSCENTER
                 if (elemType == ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT && !shape.empty()) {
                     size_t numElems = info.GetElementCount();
                     if (numElems > 0 && numElems < 100000000) {
-                        const float* data = decInputs[di].GetTensorData<float>();
+                        const float* data = decInputs[di].GetTensorMutableData<float>();
                         double sum = 0;
                         for (size_t k = 0; k < numElems; ++k) sum += data[k];
                         double mean = sum / numElems;
@@ -661,14 +663,14 @@ namespace ANSCENTER
                 // Print bool tensor values (for language_mask)
                 else if (elemType == ONNX_TENSOR_ELEMENT_DATA_TYPE_BOOL && !shape.empty()) {
                     size_t numElems = info.GetElementCount();
-                    const bool* data = decInputs[di].GetTensorData<bool>();
+                    const bool* data = decInputs[di].GetTensorMutableData<bool>();
                     std::cout << " vals:";
                     for (size_t k = 0; k < std::min(numElems, (size_t)32); ++k)
                         std::cout << " " << (int)data[k];
                 }
                 // Print int64 scalar value
                 else if (elemType == ONNX_TENSOR_ELEMENT_DATA_TYPE_INT64 && shape.empty()) {
-                    const int64_t* data = decInputs[di].GetTensorData<int64_t>();
+                    const int64_t* data = decInputs[di].GetTensorMutableData<int64_t>();
                     std::cout << " value=" << data[0];
                 }
                 std::cout << std::endl;
@@ -702,10 +704,10 @@ namespace ANSCENTER
         auto boxInfo = decOutputs[boxesIdx].GetTensorTypeAndShapeInfo();
         auto boxShape = boxInfo.GetShape();
         int numBoxes = (boxShape.size() >= 1) ? static_cast<int>(boxShape[0]) : 0;
-        const float* boxesData = decOutputs[boxesIdx].GetTensorData<float>();
+        const float* boxesData = decOutputs[boxesIdx].GetTensorMutableData<float>();
 
         // Get scores
-        const float* scoresData = decOutputs[scoresIdx].GetTensorData<float>();
+        const float* scoresData = decOutputs[scoresIdx].GetTensorMutableData<float>();
 
         // Get masks
         auto maskInfo = decOutputs[masksIdx].GetTensorTypeAndShapeInfo();
@@ -713,7 +715,7 @@ namespace ANSCENTER
         // masks shape: [N, 1, H, W]
         int maskH = (maskShape.size() >= 3) ? static_cast<int>(maskShape[2]) : 0;
         int maskW = (maskShape.size() >= 4) ? static_cast<int>(maskShape[3]) : 0;
-        const bool* masksData = decOutputs[masksIdx].GetTensorData<bool>();
+        const bool* masksData = decOutputs[masksIdx].GetTensorMutableData<bool>();
 
         m_maskH = maskH;
         m_maskW = maskW;

modules/ANSODEngine/ANSONNXCL.cpp

@@ -575,7 +575,7 @@ namespace ANSCENTER
             return false;
         }
     }
-    std::vector<Object> ANSONNXCL::postprocess(const std::vector<Ort::Value>& outputTensors, const std::string& camera_id) {
+    std::vector<Object> ANSONNXCL::postprocess(std::vector<Ort::Value>& outputTensors, const std::string& camera_id) {
         ANS_DBG("ANSONNXCL_pp", "ENTRY tensors=%zu cam=%s this=%p",
             outputTensors.size(), camera_id.c_str(), (void*)this);
         std::lock_guard<std::recursive_mutex> lock(_mutex);
@@ -589,8 +589,16 @@ namespace ANSCENTER
                 return {};
             }
 
-            ANS_DBG("ANSONNXCL_pp", "GetTensorData<float>");
+            ANS_DBG("ANSONNXCL_pp", "GetTensorMutableData<float>");
-            const float* rawOutput = outputTensors[0].GetTensorData<float>();
+            // GetTensorMutableData (not GetTensorData) on DirectML.  The const
+            // GetTensorData triggers a per-call host-readable mapping that on
+            // AMD DML exhausts a small staging-buffer pool after ~8 calls and
+            // blocks indefinitely.  GetTensorMutableData returns the existing
+            // host-accessible pointer directly with no per-call mapping cost.
+            // Same pattern used by every output-tensor read in ANSONNXYOLO
+            // and engines/ONNXEngine.  Safe on all EPs (CUDA/OpenVINO/CPU);
+            // we read the data only, never mutate it.
+            const float* rawOutput = outputTensors[0].GetTensorMutableData<float>();
             if (!rawOutput) {
                 ANS_DBG("ANSONNXCL_pp", "EARLY-RETURN rawOutput=null");
                 this->_logger.LogError("ANSONNXCL::postprocess", "rawOutput pointer is null", __FILE__, __LINE__);

modules/ANSODEngine/ANSONNXCL.h

@@ -28,7 +28,11 @@ namespace ANSCENTER {
         void warmupModel();
         bool Init(const std::string& modelPath, const cv::Size& targetInputShape, bool useGPU = true);
         bool preprocess(const cv::Mat& image, float*& blob, std::vector<int64_t>& inputTensorShape);
-        std::vector<Object> postprocess(const std::vector<Ort::Value>& outputTensors, const std::string& camera_id);
+        // outputTensors is non-const because GetTensorMutableData() (the
+        // ORT API that doesn't hang on AMD DirectML) requires a non-const
+        // Ort::Value receiver.  See comment at the GetTensorMutableData
+        // call site in postprocess() for the full rationale.
+        std::vector<Object> postprocess(std::vector<Ort::Value>& outputTensors, const std::string& camera_id);
         std::vector<Object> classify(const cv::Mat& image, const std::string& camera_id);
 
     private:

modules/ANSODEngine/ANSONNXOBB.cpp

@@ -1089,7 +1089,7 @@ namespace ANSCENTER {
 	std::vector<Object> ANSONNXOBB::postprocess(
 		const cv::Size& originalImageSize,
 		const cv::Size& resizedImageShape,
-		const std::vector<Ort::Value>& outputTensors,
+		std::vector<Ort::Value>& outputTensors,
 		int topk,
 		const std::string& camera_id)
 	{
@@ -1103,8 +1103,10 @@ namespace ANSCENTER {
 				return {};
 			}
 
-			// Extract output tensor data and shape [1, num_features, num_detections]
+			// Extract output tensor data and shape [1, num_features, num_detections].
-			const float* rawOutput = outputTensors[0].GetTensorData<float>();
+			// GetTensorMutableData (not GetTensorData) on DML — const variant
+			// hangs on AMD after ~8 calls.  Read-only despite the name.
+			const float* rawOutput = outputTensors[0].GetTensorMutableData<float>();
 			const std::vector<int64_t> outputShape = outputTensors[0].GetTensorTypeAndShapeInfo().GetShape();
 
 			if (outputShape.size() < 3) {

modules/ANSODEngine/ANSONNXOBB.h

@@ -74,10 +74,12 @@ namespace ANSCENTER {
         void warmupModel();
         bool Init(const std::string& modelPath, bool useGPU=true, int deviceId = 0);
         cv::Mat preprocess(const cv::Mat& image, float*& blob, std::vector<int64_t>& inputTensorShape);
+        // outputTensors is non-const because GetTensorMutableData() requires
+        // a non-const Ort::Value receiver — see ANSONNXCL.h for full note.
         std::vector<Object> postprocess(
             const cv::Size& originalImageSize,
             const cv::Size& resizedImageShape,
-            const std::vector<Ort::Value>& outputTensors, int topk,
+            std::vector<Ort::Value>& outputTensors, int topk,
             const std::string& camera_id);
         std::vector<Object> detect(const cv::Mat& image, const std::string& camera_id);
     private:

modules/ANSODEngine/ANSONNXPOSE.cpp

@@ -759,7 +759,7 @@ namespace ANSCENTER {
 	std::vector<Object> ANSONNXPOSE::postprocess(
 		const cv::Size& originalImageSize,
 		const cv::Size& resizedImageShape,
-		const std::vector<Ort::Value>& outputTensors,
+		std::vector<Ort::Value>& outputTensors,
 		const std::string& camera_id)
 	{
 		std::lock_guard<std::recursive_mutex> lock(_mutex);
@@ -773,7 +773,9 @@ namespace ANSCENTER {
 				return {};
 			}
 
-			const float* rawOutput = outputTensors[0].GetTensorData<float>();
+			// GetTensorMutableData (not GetTensorData) on DML — const variant
+			// hangs on AMD after ~8 calls.  Read-only despite the name.
+			const float* rawOutput = outputTensors[0].GetTensorMutableData<float>();
 			if (!rawOutput) {
 				this->_logger.LogError("ANSONNXPOSE::postprocess", "rawOutput pointer is null", __FILE__, __LINE__);
 				return {};

modules/ANSODEngine/ANSONNXPOSE.h

@@ -41,8 +41,10 @@ namespace ANSCENTER {
         void warmupModel();
         bool Init(const std::string& modelPath, bool useGPU=true, int deviceId = 0);
         cv::Mat preprocess(const cv::Mat& image, float*& blob, std::vector<int64_t>& inputTensorShape);
+        // outputTensors is non-const because GetTensorMutableData() requires
+        // a non-const Ort::Value receiver — see ANSONNXCL.h for full note.
         std::vector<Object> postprocess(const cv::Size& originalImageSize, const cv::Size& resizedImageShape,
-                                        const std::vector<Ort::Value>& outputTensors, const std::string& camera_id);
+                                        std::vector<Ort::Value>& outputTensors, const std::string& camera_id);
         std::vector<Object> detect(const cv::Mat& image, const std::string& camera_id);
     private:
         static std::atomic<int> instanceCounter_;  // Thread-safe counter

modules/ANSODEngine/ANSONNXSEG.cpp

@@ -726,7 +726,7 @@ namespace ANSCENTER {
 	std::vector<Object> ANSONNXSEG::postprocess(
 		const cv::Size& origSize,
 		const cv::Size& letterboxSize,
-		const std::vector<Ort::Value>& outputs,
+		std::vector<Ort::Value>& outputs,
 		const std::string& camera_id)
 	{
 		std::lock_guard<std::recursive_mutex> lock(_mutex);
@@ -738,9 +738,11 @@ namespace ANSCENTER {
 					std::to_string(outputs.size()));
 			}
 
-			// Extract output tensors
+			// Extract output tensors.  GetTensorMutableData (not GetTensorData)
-			const float* detections = outputs[0].GetTensorData<float>();
+			// on DML — const variant hangs on AMD after ~8 calls.  Read-only
-			const float* prototypes = outputs[1].GetTensorData<float>();
+			// despite the name.
+			const float* detections = outputs[0].GetTensorMutableData<float>();
+			const float* prototypes = outputs[1].GetTensorMutableData<float>();
 
 			// Get tensor shapes
 			auto detectionShape = outputs[0].GetTensorTypeAndShapeInfo().GetShape(); // [1, 116, N]

modules/ANSODEngine/ANSONNXSEG.h

@@ -51,8 +51,10 @@ namespace ANSCENTER {
         void warmupModel();
         bool Init(const std::string& modelPath, bool useGPU=true, int deviceId = 0);
         cv::Mat preprocess(const cv::Mat& image,float*& blobPtr,std::vector<int64_t>& inputTensorShape);
+        // outputs is non-const because GetTensorMutableData() requires a
+        // non-const Ort::Value receiver — see ANSONNXCL.h for full note.
         std::vector<Object> postprocess(const cv::Size& origSize,const cv::Size& letterboxSize,
-                                       const std::vector<Ort::Value>& outputs, const std::string& camera_id);
+                                       std::vector<Ort::Value>& outputs, const std::string& camera_id);
         std::vector<Object> segment(const cv::Mat& image, const std::string& camera_id);
         std::vector<cv::Point2f> maskToPolygon(const cv::Mat& binaryMask,
             const cv::Rect& boundingBox,

modules/ANSODEngine/ANSYOLO12OD.cpp

@@ -518,14 +518,16 @@ namespace ANSCENTER {
         }
     }
     std::vector<Object> YOLO12OD::postprocess(const cv::Size& originalImageSize, const cv::Size& resizedImageShape,
-                                            const std::vector<Ort::Value>& outputTensors,
+                                            std::vector<Ort::Value>& outputTensors,
                                             float confThreshold, float iouThreshold)
     {
 		std::lock_guard<std::recursive_mutex> lock(_mutex);
         try {
 
             std::vector<Object> detections;
-            const float* rawOutput = outputTensors[0].GetTensorData<float>(); // Extract raw output data from the first output tensor
+            // GetTensorMutableData (not GetTensorData) on DML — const variant
+            // hangs on AMD after ~8 calls.  Read-only despite the name.
+            const float* rawOutput = outputTensors[0].GetTensorMutableData<float>();
             const std::vector<int64_t> outputShape = outputTensors[0].GetTensorTypeAndShapeInfo().GetShape();
 
             // Determine the number of features and detections

modules/ANSODEngine/ANSYOLO12OD.h

@@ -49,8 +49,11 @@ namespace ANSCENTER {
             std::vector<Object> detect(const cv::Mat& image, float confThreshold = 0.4f, float iouThreshold = 0.45f);
             //cv::Mat preprocess(const cv::Mat& image, float*& blob, std::vector<int64_t>& inputTensorShape);
             cv::Mat preprocess(const cv::Mat& image, std::vector<float>& blob, std::vector<int64_t>& inputTensorShape);
+            // outputTensors is non-const because GetTensorMutableData()
+            // requires a non-const Ort::Value receiver — see ANSONNXCL.h
+            // for full note.
             std::vector<Object> postprocess(const cv::Size& originalImageSize, const cv::Size& resizedImageShape,
-                                            const std::vector<Ort::Value>& outputTensors,
+                                            std::vector<Ort::Value>& outputTensors,
                                             float confThreshold, float iouThreshold);
 
         private:

tests/ANSODEngine-UnitTest/ANSODEngine-UnitTest.cpp

@@ -1867,12 +1867,13 @@ int main()
     //YOLO26POSEYolo11Test();
     //YOLO26CLYolo11Test();
      //YOLO26ODYolo12Test();
-    YOLO26ODYolo11Test();
+    //YOLO26ODYolo11Test();
     //YOLO26ODYolo10Test();
     //YOLO26OBBYolo11Test();
     //SAM3ONNX_ImageTest();   // ORT reference — runs first, prints decoder input stats
     //SAM3TRT_ImageTest();    // TRT under test — compare decoder input stats with above
-    //CustomModel_StressTest_FilePlayer();  // Multi-task stress test (LabVIEW flow)
+    //CustomModel_StressTest_FilePlayer();      // Multi-task stress test (LabVIEW flow)
+    CustomModel_SingleStream_FilePlayer();    // 1 cam + 1 task — isolates concurrency from per-instance bugs
     //SAM3TRT_UnitTest();  // TensorRT SAM3 test (in ANSSAM3-UnitTest.cpp)
     //TensorRT10Test();
     //FireNSmokeCustomDetection();

tests/ANSODEngine-UnitTest/ANSODTest.h

@@ -117,3 +117,4 @@ int TestYOLOV12();
 int PPETest();
 int RVATest();
 int CustomModel_StressTest_FilePlayer();
+int CustomModel_SingleStream_FilePlayer();   // 1 camera + 1 task — isolates concurrency from per-instance bugs

tests/ANSODEngine-UnitTest/CustomModel-StressTest.cpp

@@ -292,15 +292,15 @@ int CustomModel_StressTest_FilePlayer() {
 
     // Video files (one per stream)
     const std::string videoFiles[NUM_STREAMS] = {
-        "E:\\Programs\\DemoAssets\\Videos\\Helmet\\HM1.mp4",
+        "C:\\ProgramData\\ANSCENTER\\Shared\\HM1.mp4",
-        "E:\\Programs\\DemoAssets\\Videos\\Helmet\\HM2.mp4",
+        "C:\\ProgramData\\ANSCENTER\\Shared\\HM2.mp4",
     };
 
     // Which stream each task uses
     const int taskStreamMap[NUM_TASKS] = { 0, 0, 1, 1 };
 
     // Model config — EDIT for your custom model
-    const std::string modelFolder = "C:\\Projects\\ANSVIS\\Models\\ANS_Helmet_v2.0.zip";
+    const std::string modelFolder = "C:\\ProgramData\\ANSCENTER\\ANSVIS Server\\Models\\B-IN_ANS_Helmet_v2.0_102728911.zip";
     //const char* modelName = "detector";
     //const char* className = "detector.names";
     const int modelType = 16;   // 16 = CustomDetector, 31 = RTYOLO, 30 = ONNXYOLO
@@ -550,3 +550,237 @@ int CustomModel_StressTest_FilePlayer() {
 
     return 0;
 }
+
+
+// =============================================================================
+//  CustomModel_SingleStream_FilePlayer
+//
+//  ISOLATION TEST — 1 camera, 1 OD handle, 1 worker thread.  No concurrent
+//  inference whatsoever.  Same flow as CustomModel_StressTest_FilePlayer
+//  (FilePlayer → CloneImage → RunInferenceComplete_CPP → ReleaseImage), just
+//  without the multi-stream / multi-task fan-out.
+//
+//  Use this to determine whether a hang is per-instance (will still hang here)
+//  or only triggered by cross-session DML contention (will NOT hang here).
+//  If THIS test runs cleanly for an extended period but the multi-stream
+//  stress test hangs after a few inferences, the issue is concurrent DML
+//  submissions on the AMD iGPU — not a bug in the engine code itself.
+//
+//  Reuses helpers from CustomModel_StressTest_FilePlayer:
+//    LoadANSCV / UnloadANSCV, ODWorkerThread, GetPerGpuFreeMiB.
+// =============================================================================
+int CustomModel_SingleStream_FilePlayer() {
+    printf("\n");
+    printf("============================================================\n");
+    printf("  Custom Model SINGLE-STREAM Isolation Test (FilePlayer)\n");
+    printf("  1 camera + 1 model + 1 worker thread\n");
+    printf("  Press ESC to stop\n");
+    printf("============================================================\n\n");
+
+    // --- Load ANSCV.dll at runtime (same helper as stress test) ---
+    if (!LoadANSCV()) return -1;
+    if (pInitCameraNetwork) pInitCameraNetwork();
+
+    // =====================================================================
+    //  CONFIGURATION — EDIT THESE FOR YOUR TEST
+    // =====================================================================
+    const std::string videoFile =
+        "C:\\ProgramData\\ANSCENTER\\Shared\\HM1.mp4";
+
+    const std::string modelFolder =
+        "C:\\ProgramData\\ANSCENTER\\ANSVIS Server\\Models\\B-IN_ANS_Helmet_v2.0_102728911.zip";
+    const int   modelType    = 16;     // 16 = CustomDetector (same as stress test)
+    const int   detectorType = 1;      // Detection
+    const float scoreThresh  = 0.5f;
+    const float confThresh   = 0.5f;
+    const float nmsThresh    = 0.45f;
+    // =====================================================================
+
+    // Reset shared run flag (it's a static at file scope shared with stress test)
+    g_stressRunning.store(true);
+
+    std::cout << "\n--- Single-stream isolation test (no concurrency) ---\n" << std::endl;
+    // NOTE: deliberately NOT calling OptimizeModelStr here.  OptimizeModelStr
+    // creates a separate "warmup" ANSCUSTOM instance whose detector and
+    // classifier sessions stay loaded for the lifetime of the process — even
+    // though that instance never runs inference, its 2 DML sessions hold AMD
+    // GPU resources and were suspected of contributing to a hang in the
+    // active session's GetTensorData<float>.  Skipping it here leaves exactly
+    // 1 ANSCUSTOM = 2 DML sessions (detector + classifier) in the process,
+    // for the cleanest possible single-session isolation.
+    (void)detectorType;  // unused without the OptimizeModelStr call
+
+    // --- Per-task state (just one) ---
+    StressTaskState taskState;
+
+    // --- Create FilePlayer (single stream) ---
+    void* fpClient = nullptr;
+    {
+        printf("[Stream0] Creating FilePlayer: %s\n", videoFile.c_str());
+        int result = pCreateFilePlayer(&fpClient, "", videoFile.c_str());
+        if (result != 1 || !fpClient) {
+            printf("[Stream0] FAILED to create FilePlayer (result=%d)\n", result);
+            UnloadANSCV();
+            return -2;
+        }
+        if (pSetFilePlayerDisplayRes) {
+            pSetFilePlayerDisplayRes(&fpClient, 1920, 1080);
+        }
+        printf("[Stream0] FilePlayer created (display: 1920x1080)\n");
+    }
+
+    // --- Create OD handle (single instance) ---
+    ANSCENTER::ANSODBase* odHandle = nullptr;
+    {
+        printf("[Task0] Creating OD handle (modelType=%d)...\n", modelType);
+        auto loadStart  = std::chrono::steady_clock::now();
+        auto vramBefore = GetPerGpuFreeMiB();
+
+        std::string labelMap = CreateANSODHandle(
+            &odHandle,
+            "",                  // licenseKey
+            modelFolder.c_str(), // modelFilePath (zip or folder)
+            "",                  // modelZipFilePassword
+            scoreThresh,
+            confThresh,
+            nmsThresh,
+            1,                   // autoDetectEngine
+            modelType,
+            1,                   // detectionType (1 = Detection)
+            1);                  // loadEngineOnCreation
+
+        auto loadEnd = std::chrono::steady_clock::now();
+        double loadMs = std::chrono::duration<double, std::milli>(loadEnd - loadStart).count();
+
+        if (!odHandle) {
+            printf("[Task0] FAILED to create OD handle\n");
+            pStopFilePlayer(&fpClient);
+            pReleaseFilePlayer(&fpClient);
+            UnloadANSCV();
+            return -3;
+        }
+
+        auto vramAfter = GetPerGpuFreeMiB();
+        int bestGpu = 0;
+        size_t maxDelta = 0;
+        for (size_t g = 0; g < vramBefore.size() && g < vramAfter.size(); g++) {
+            size_t delta = (vramBefore[g] > vramAfter[g]) ? vramBefore[g] - vramAfter[g] : 0;
+            if (delta > maxDelta) { maxDelta = delta; bestGpu = (int)g; }
+        }
+
+        printf("[Task0] Model loaded in %.0f ms | GPU[%d] | VRAM: %zu MiB | Labels: %s\n",
+               loadMs, bestGpu, maxDelta,
+               labelMap.empty() ? "(none)" : labelMap.substr(0, 80).c_str());
+
+        std::lock_guard<std::mutex> lk(taskState.mtx);
+        taskState.engineLoaded = true;
+        taskState.statusMsg    = "Running";
+        taskState.gpuDeviceId  = bestGpu;
+        taskState.vramUsedMiB  = maxDelta;
+    }
+
+    // --- Start playback ---
+    pStartFilePlayer(&fpClient);
+    printf("[Stream0] Playback started\n");
+    std::this_thread::sleep_for(std::chrono::milliseconds(500));
+
+    // --- Single worker thread (reuse ODWorkerThread from stress test) ---
+    std::thread worker(ODWorkerThread, /*taskId=*/0, fpClient, odHandle, std::ref(taskState));
+
+    // --- Display loop (single cell) ---
+    const int cellW = 1280, cellH = 720;
+    const char* windowName = "Custom Model — Single Stream Isolation";
+    cv::namedWindow(windowName, cv::WINDOW_NORMAL);
+    cv::resizeWindow(windowName, cellW, cellH + 40);
+
+    auto testStart = std::chrono::steady_clock::now();
+
+    while (g_stressRunning.load()) {
+        cv::Mat canvas(cellH + 40, cellW, CV_8UC3, cv::Scalar(30, 30, 30));
+
+        cv::Mat cell;
+        double fps = 0, infMs = 0, grabMs = 0;
+        int    fCount = 0, dCount = 0, gpuId = -1;
+        std::string statusMsg, lastDet;
+        bool engineLoaded = false;
+        {
+            std::lock_guard<std::mutex> lk(taskState.mtx);
+            if (!taskState.displayFrame.empty()) {
+                cv::resize(taskState.displayFrame, cell, cv::Size(cellW, cellH));
+            }
+            fps          = taskState.fps;
+            infMs        = taskState.inferenceMs;
+            grabMs       = taskState.grabMs;
+            fCount       = taskState.frameCount;
+            dCount       = taskState.detectionCount;
+            gpuId        = taskState.gpuDeviceId;
+            statusMsg    = taskState.statusMsg;
+            lastDet      = taskState.lastDetection;
+            engineLoaded = taskState.engineLoaded;
+        }
+
+        if (cell.empty()) {
+            cell = cv::Mat(cellH, cellW, CV_8UC3, cv::Scalar(40, 40, 40));
+            cv::putText(cell, "Task 0: " + statusMsg,
+                        cv::Point(20, cellH / 2),
+                        cv::FONT_HERSHEY_SIMPLEX, 0.8, cv::Scalar(100, 100, 255), 2);
+        }
+
+        cv::rectangle(cell, cv::Rect(0, cellH - 45, cellW, 45),
+                      cv::Scalar(0, 0, 0), cv::FILLED);
+        char bar1[256], bar2[128];
+        snprintf(bar1, sizeof(bar1),
+                 "%.1f FPS | inf:%.0fms grab:%.0fms | Frames:%d | Det:%d",
+                 fps, infMs, grabMs, fCount, dCount);
+        snprintf(bar2, sizeof(bar2), "GPU[%d] | last:%s",
+                 gpuId, lastDet.empty() ? "-" : lastDet.c_str());
+        cv::Scalar barColor = engineLoaded ? cv::Scalar(0, 255, 0) : cv::Scalar(0, 100, 255);
+        cv::putText(cell, bar1, cv::Point(5, cellH - 25),
+                    cv::FONT_HERSHEY_SIMPLEX, 0.5, barColor, 1);
+        cv::putText(cell, bar2, cv::Point(5, cellH - 5),
+                    cv::FONT_HERSHEY_SIMPLEX, 0.5, cv::Scalar(0, 200, 255), 1);
+
+        cell.copyTo(canvas(cv::Rect(0, 0, cellW, cellH)));
+
+        double elapsed = std::chrono::duration<double>(
+            std::chrono::steady_clock::now() - testStart).count();
+        char bottomBar[256];
+        snprintf(bottomBar, sizeof(bottomBar),
+                 "Single-stream | Elapsed: %.0fs | %.1f FPS | Press ESC to stop",
+                 elapsed, fps);
+        cv::putText(canvas, bottomBar, cv::Point(10, cellH + 25),
+                    cv::FONT_HERSHEY_SIMPLEX, 0.5, cv::Scalar(200, 200, 200), 1);
+
+        cv::imshow(windowName, canvas);
+        int key = cv::waitKey(30);
+        if (key == 27) {
+            printf("\nESC pressed - stopping...\n");
+            g_stressRunning.store(false);
+        }
+    }
+
+    printf("Waiting for worker thread...\n");
+    if (worker.joinable()) worker.join();
+
+    double totalElapsed = std::chrono::duration<double>(
+        std::chrono::steady_clock::now() - testStart).count();
+    printf("\n============================================================\n");
+    printf("  SINGLE-STREAM SUMMARY (runtime: %.0fs)\n", totalElapsed);
+    printf("============================================================\n");
+    printf("  GPU[%d] | %d frames | %d detections | %.1f FPS | Inf: %.0fms\n",
+           taskState.gpuDeviceId, taskState.frameCount, taskState.detectionCount,
+           taskState.fps, taskState.inferenceMs);
+    printf("============================================================\n");
+
+    if (odHandle) ReleaseANSODHandle(&odHandle);
+    if (fpClient) {
+        pStopFilePlayer(&fpClient);
+        pReleaseFilePlayer(&fpClient);
+    }
+
+    cv::destroyAllWindows();
+    if (pDeinitCameraNetwork) pDeinitCameraNetwork();
+    UnloadANSCV();
+
+    return 0;
+}