Merge pull request #3070 from asmorkalov:as/hw_optflow_destructor

modules/cudaoptflow/src/nvidiaOpticalFlow.cpp

@@ -314,15 +314,11 @@ private:
 
     NV_OF_CUDA_API_FUNCTION_LIST* GetAPI()
     {
-        std::lock_guard<std::mutex> lock(m_lock);
         return  m_ofAPI.get();
     }
 
     NvOFHandle GetHandle() { return m_hOF; }
 
-protected:
-    std::mutex m_lock;
-
 public:
     NvidiaOpticalFlowImpl(cv::Size imageSize, NV_OF_PERF_LEVEL perfPreset, bool bEnableTemporalHints,
         bool bEnableExternalHints, bool bEnableCostBuffer, int gpuId, Stream inputStream, Stream outputStream);
@@ -338,6 +334,8 @@ public:
         int gridSize, InputOutputArray upsampledFlow);
 
     virtual int getGridSize() const { return m_gridSize; }
+
+    ~NvidiaOpticalFlowImpl();
 };
 
 NvidiaOpticalFlowImpl::NvidiaOpticalFlowImpl(
@@ -599,20 +597,24 @@ void NvidiaOpticalFlowImpl::collectGarbage()
     if (m_hInputBuffer)
     {
         NVOF_API_CALL(GetAPI()->nvOFDestroyGPUBufferCuda(m_hInputBuffer));
+        m_hInputBuffer = nullptr;
     }
     if (m_hReferenceBuffer)
     {
         NVOF_API_CALL(GetAPI()->nvOFDestroyGPUBufferCuda(m_hReferenceBuffer));
+        m_hReferenceBuffer = nullptr;
     }
     if (m_hOutputBuffer)
     {
         NVOF_API_CALL(GetAPI()->nvOFDestroyGPUBufferCuda(m_hOutputBuffer));
+        m_hOutputBuffer = nullptr;
     }
     if (m_enableExternalHints)
     {
         if (m_hHintBuffer)
         {
             NVOF_API_CALL(GetAPI()->nvOFDestroyGPUBufferCuda(m_hHintBuffer));
+            m_hHintBuffer = nullptr;
         }
     }
     if (m_enableCostBuffer)
@@ -620,6 +622,7 @@ void NvidiaOpticalFlowImpl::collectGarbage()
         if (m_hCostBuffer)
         {
             NVOF_API_CALL(GetAPI()->nvOFDestroyGPUBufferCuda(m_hCostBuffer));
+            m_hCostBuffer = nullptr;
         }
     }
     if (m_inputStream)
@@ -633,9 +636,15 @@ void NvidiaOpticalFlowImpl::collectGarbage()
     if (m_hOF)
     {
         NVOF_API_CALL(GetAPI()->nvOFDestroy(m_hOF));
+        m_hOF = nullptr;
     }
 }
 
+NvidiaOpticalFlowImpl::~NvidiaOpticalFlowImpl()
+{
+    collectGarbage();
+}
+
 void NvidiaOpticalFlowImpl::upSampler(InputArray _flow, cv::Size imageSize,
     int gridSize, InputOutputArray upsampledFlow)
 {
@@ -754,15 +763,11 @@ private:
 
     NV_OF_CUDA_API_FUNCTION_LIST* GetAPI()
     {
-        std::lock_guard<std::mutex> lock(m_lock);
         return  m_ofAPI.get();
     }
 
     NvOFHandle GetHandle() { return m_hOF; }
 
-protected:
-    std::mutex m_lock;
-
 public:
     NvidiaOpticalFlowImpl_2(cv::Size imageSize, NV_OF_PERF_LEVEL perfPreset,
         NV_OF_OUTPUT_VECTOR_GRID_SIZE outputGridSize, NV_OF_HINT_VECTOR_GRID_SIZE hintGridSize,
@@ -778,6 +783,8 @@ public:
     virtual void convertToFloat(InputArray flow, InputOutputArray floatFlow);
 
     virtual int getGridSize() const { return m_gridSize; }
+
+    ~NvidiaOpticalFlowImpl_2();
 };
 
 NvidiaOpticalFlowImpl_2::NvidiaOpticalFlowImpl_2(
@@ -1149,24 +1156,29 @@ void NvidiaOpticalFlowImpl_2::collectGarbage()
     if (m_hInputBuffer)
     {
         NVOF_API_CALL(GetAPI()->nvOFDestroyGPUBufferCuda(m_hInputBuffer));
+        m_hInputBuffer = nullptr;
     }
     if (m_hReferenceBuffer)
     {
         NVOF_API_CALL(GetAPI()->nvOFDestroyGPUBufferCuda(m_hReferenceBuffer));
+        m_hReferenceBuffer = nullptr;
     }
     if (m_hOutputBuffer)
     {
         NVOF_API_CALL(GetAPI()->nvOFDestroyGPUBufferCuda(m_hOutputBuffer));
+        m_hOutputBuffer = nullptr;
     }
     if (m_scaleFactor > 1 && m_hOutputUpScaledBuffer)
     {
         NVOF_API_CALL(GetAPI()->nvOFDestroyGPUBufferCuda(m_hOutputUpScaledBuffer));
+        m_hOutputUpScaledBuffer = nullptr;
     }
     if (m_enableExternalHints)
     {
         if (m_hHintBuffer)
         {
             NVOF_API_CALL(GetAPI()->nvOFDestroyGPUBufferCuda(m_hHintBuffer));
+            m_hHintBuffer = nullptr;
         }
     }
     if (m_enableCostBuffer)
@@ -1174,6 +1186,7 @@ void NvidiaOpticalFlowImpl_2::collectGarbage()
         if (m_hCostBuffer)
         {
             NVOF_API_CALL(GetAPI()->nvOFDestroyGPUBufferCuda(m_hCostBuffer));
+            m_hCostBuffer = nullptr;
         }
     }
     if (m_inputStream)
@@ -1187,9 +1200,15 @@ void NvidiaOpticalFlowImpl_2::collectGarbage()
     if (m_hOF)
     {
         NVOF_API_CALL(GetAPI()->nvOFDestroy(m_hOF));
+        m_hOF = nullptr;
     }
 }
 
+NvidiaOpticalFlowImpl_2::~NvidiaOpticalFlowImpl_2()
+{
+    collectGarbage();
+}
+
 void NvidiaOpticalFlowImpl_2::convertToFloat(InputArray _flow, InputOutputArray floatFlow)
 {
     Mat flow;

modules/cudaoptflow/test/test_optflow.cpp

@@ -605,7 +605,6 @@ CUDA_TEST_P(NvidiaOpticalFlow_2_0, Regression)
     ASSERT_FALSE(golden.empty());
 
     EXPECT_MAT_SIMILAR(golden, upsampledFlow, 1e-10);
-    d_nvof->collectGarbage();
 }
 
 CUDA_TEST_P(NvidiaOpticalFlow_2_0, OpticalFlowNan)
@@ -640,7 +639,6 @@ CUDA_TEST_P(NvidiaOpticalFlow_2_0, OpticalFlowNan)
 
     EXPECT_TRUE(cv::checkRange(flowx));
     EXPECT_TRUE(cv::checkRange(flowy));
-    d_nvof->collectGarbage();
 };
 
 INSTANTIATE_TEST_CASE_P(CUDA_OptFlow, NvidiaOpticalFlow_2_0, ALL_DEVICES);