Fix Intel ONNX runtime

modules/ANSODEngine/ANSONNXYOLO.cpp

@@ -4,6 +4,7 @@
 #include "NV12PreprocessHelper.h"   // tl_currentGpuFrame()
 #include <numeric>   // std::iota
 #include <cmath>
+#include <chrono>    // WarmUpEngine() timing
 
 namespace ANSCENTER {
 
@@ -1587,6 +1588,64 @@ namespace ANSCENTER {
         }
     }
 
+    // ========================================================================
+    // WarmUpEngine — run 2 dummy inferences after session creation
+    //
+    // On AMD RDNA2 iGPUs (e.g. Radeon 680M on Ryzen 6000-series APUs), the
+    // very first detect() call triggers DirectML shader compile + GPU kernel
+    // cache population for the entire YOLO graph.  That first pass can
+    // legitimately take several seconds of sustained GPU work, which is long
+    // enough to coincide with TDR watchdog firing and has triggered
+    // amdkmdag.sys bugchecks at +0xf03d under DirectML 1.15.4 (the latest).
+    //
+    // Running 2 dummy inferences at startup burns the compile cost under
+    // controlled conditions so that the first real frame is already fast.
+    // The second call should always be quick and confirms the cache is warm.
+    //
+    // Non-fatal on failure: if warm-up itself crashes, regular inference may
+    // still succeed, or will fail with a clearer error message.
+    // ========================================================================
+    void ANSONNXYOLO::WarmUpEngine() {
+        if (!m_ortEngine) return;
+
+        // Warm-up exists solely to pre-compile DirectML shaders on AMD RDNA2
+        // iGPUs (Radeon 680M).  It has no benefit on CPU / OpenVINO / CUDA
+        // and running detect() at load time has been observed to expose
+        // latent heap-corruption bugs (ntdll +0x1176e5 / STATUS_HEAP_CORRUPTION
+        // 0xc0000374) on Intel machines.  Gate strictly on AMD_GPU.
+        if (m_ortEngine->getEngineType() != EngineType::AMD_GPU) {
+            ANS_DBG("ONNXYOLO", "Warm-up skipped (non-AMD EP)");
+            return;
+        }
+
+        try {
+            const int w = _modelConfig.inpWidth  > 0 ? _modelConfig.inpWidth  : 640;
+            const int h = _modelConfig.inpHeight > 0 ? _modelConfig.inpHeight : 640;
+
+            // Mid-gray BGR image matches the letterbox fill colour used in
+            // preprocessing (114,114,114 ~ 128) and avoids degenerate inputs.
+            cv::Mat dummy(h, w, CV_8UC3, cv::Scalar(128, 128, 128));
+
+            ANS_DBG("ONNXYOLO", "Warm-up: running 2 dummy inferences (%dx%d)", w, h);
+
+            for (int i = 0; i < 2; ++i) {
+                auto t0 = std::chrono::steady_clock::now();
+                (void)m_ortEngine->detect(dummy, _classes,
+                                          PROBABILITY_THRESHOLD,
+                                          NMS_THRESHOLD,
+                                          NUM_KPS);
+                auto t1 = std::chrono::steady_clock::now();
+                auto ms = std::chrono::duration_cast<std::chrono::milliseconds>(t1 - t0).count();
+                ANS_DBG("ONNXYOLO", "Warm-up #%d: %lld ms", i, (long long)ms);
+            }
+        }
+        catch (const std::exception& e) {
+            _logger.LogError("ANSONNXYOLO::WarmUpEngine",
+                std::string("Warm-up failed (non-fatal): ") + e.what(),
+                __FILE__, __LINE__);
+        }
+    }
+
     bool ANSONNXYOLO::Initialize(std::string licenseKey, ModelConfig modelConfig,
                                    const std::string& modelZipFilePath,
                                    const std::string& modelZipPassword,
@@ -1651,6 +1710,10 @@ namespace ANSCENTER {
                 }
             }
 
+            // Pre-compile DirectML shaders / kernel cache before first real
+            // frame (mitigates amdkmdag TDR on Radeon 680M).  Non-fatal.
+            WarmUpEngine();
+
             _modelLoadValid = true;
             _isInitialized  = true;
             return true;
@@ -1708,6 +1771,10 @@ namespace ANSCENTER {
                 }
             }
 
+            // Pre-compile DirectML shaders / kernel cache before first real
+            // frame (mitigates amdkmdag TDR on Radeon 680M).  Non-fatal.
+            WarmUpEngine();
+
             _modelLoadValid = true;
             _isInitialized  = true;
             return true;
@@ -1778,6 +1845,10 @@ namespace ANSCENTER {
                 }
             }
 
+            // Pre-compile DirectML shaders / kernel cache before first real
+            // frame (mitigates amdkmdag TDR on Radeon 680M).  Non-fatal.
+            WarmUpEngine();
+
             _modelLoadValid = true;
             _isInitialized  = true;
             return true;