Use CPU resize before upload to GPU to remove PCIe bottleneck

engines/TensorRTAPI/include/engine/EngineRunInference.inl

@@ -284,7 +284,13 @@ bool Engine<T>::runInference(const std::vector<std::vector<cv::cuda::GpuMat>>& i
     // fatal "illegal memory access" that permanently corrupts the CUDA context.
     //
     // Pool-mode slots have their own busy-flag dispatch so they do NOT need this.
+    auto _mutexWaitStart = std::chrono::steady_clock::now();
     std::lock_guard<std::mutex> inferenceLock(m_inferenceMutex);
+    auto _mutexAcquired = std::chrono::steady_clock::now();
+    double _mutexWaitMs = std::chrono::duration<double, std::milli>(_mutexAcquired - _mutexWaitStart).count();
+    if (_mutexWaitMs > 50.0) {
+        ANS_DBG("TRT_Engine", "MUTEX WAIT: %.1fms (queued behind another inference)", _mutexWaitMs);
+    }
 
     // ============================================================================
     // THREAD-SAFE GPU CONTEXT
@@ -955,6 +961,20 @@ bool Engine<T>::runInference(const std::vector<std::vector<cv::cuda::GpuMat>>& i
         }
     }
 
+    // ============================================================================
+    // Per-inference total timing breakdown (mutex wait + preprocess + GPU)
+    // ============================================================================
+    {
+        double totalMs = std::chrono::duration<double, std::milli>(
+            std::chrono::steady_clock::now() - _mutexWaitStart).count();
+        double gpuMs = totalMs - _mutexWaitMs;  // Everything after mutex acquired
+        // Log every inference that takes >100ms total (including mutex wait)
+        if (totalMs > 100.0) {
+            ANS_DBG("TRT_Timing", "total=%.1fms (mutex=%.1fms gpu=%.1fms) batch=%d active=%d",
+                    totalMs, _mutexWaitMs, gpuMs, batchSize, s_globalActiveInf.load());
+        }
+    }
+
     // ============================================================================
     // SM=100% DETECTOR — end-of-inference timing
     // ============================================================================

engines/TensorRTAPI/include/engine/EngineUtilities.inl

@@ -23,6 +23,29 @@ void Engine<T>::transformOutput(std::vector<std::vector<std::vector<T>>> &input,
     }
     output = std::move(input[0][0]);
 }
+// CPU letterbox resize — same logic as the GPU version but runs on CPU.
+// Used in Preprocess to resize BEFORE GPU upload, reducing PCIe transfer
+// from 25 MB (4K) to 1.2 MB (640×640) — 20x less bandwidth.
+template <typename T>
+cv::Mat Engine<T>::cpuResizeKeepAspectRatioPadRightBottom(const cv::Mat& input,
+                                                          size_t height, size_t width,
+                                                          const cv::Scalar& bgcolor) {
+    if (input.empty()) return cv::Mat();
+
+    float r = std::min(static_cast<float>(width) / input.cols,
+                       static_cast<float>(height) / input.rows);
+    int unpad_w = static_cast<int>(r * input.cols);
+    int unpad_h = static_cast<int>(r * input.rows);
+
+    cv::Mat re;
+    cv::resize(input, re, cv::Size(unpad_w, unpad_h), 0, 0, cv::INTER_LINEAR);
+
+    cv::Mat out(static_cast<int>(height), static_cast<int>(width), input.type(), bgcolor);
+    re.copyTo(out(cv::Rect(0, 0, re.cols, re.rows)));
+
+    return out;
+}
+
 template <typename T>
 cv::cuda::GpuMat Engine<T>::resizeKeepAspectRatioPadRightBottom(const cv::cuda::GpuMat& input,
                                                                 size_t height, size_t width,