Fix NV12 crash issue when recreate camera object

modules/ANSODEngine/nv12_to_rgb.cu

@@ -8,6 +8,9 @@
 
 #include <cuda_runtime.h>
 #include <cstdint>
+#ifdef _WIN32
+#include <windows.h>  // Sleep()
+#endif
 #include <cstdio>
 
 // ── Shared YUV→RGB computation ───────────────────────────────────────────
@@ -651,7 +654,24 @@ int ANSGpuNV12ToBGR(
                       width * 3, height,
                       cudaMemcpyDeviceToHost, t_bufs.stream);
 
-    cudaStreamSynchronize(t_bufs.stream);
+    // Use polling sync instead of cudaStreamSynchronize to avoid
+    // holding nvcuda64 SRW lock continuously (WDDM deadlock prevention).
+    // Short Sleep(0) fast path for sub-ms kernels, then Sleep(1) to give
+    // cleanup operations (cuArrayDestroy, cuMemFree) a window to acquire
+    // the exclusive SRW lock.
+    {
+        cudaError_t qerr = cudaStreamQuery(t_bufs.stream);
+        if (qerr == cudaErrorNotReady) {
+            for (int i = 0; i < 10 && qerr == cudaErrorNotReady; ++i) {
+                Sleep(0);
+                qerr = cudaStreamQuery(t_bufs.stream);
+            }
+            while (qerr == cudaErrorNotReady) {
+                Sleep(1);
+                qerr = cudaStreamQuery(t_bufs.stream);
+            }
+        }
+    }
 
     // Check for errors
     cudaError_t err = cudaGetLastError();