fix convolution bug by retain primary context for device (#894)

* fix convolution bug by retain primary context for device
* add release to destructor

src/ngraph/runtime/gpu/gpu_cuda_context_manager.cpp

@@ -31,6 +31,11 @@ runtime::gpu::CudaContextManager::CudaContextManager()
 {
     CUDA_SAFE_CALL(cuInit(0));
     CUDA_SAFE_CALL(cuDeviceGet(&m_device, 0));
-    CUDA_SAFE_CALL(cuCtxCreate(&m_context, 0, m_device));
+    CUDA_SAFE_CALL(cuDevicePrimaryCtxRetain(&m_context, m_device));
     m_context_ptr = std::make_shared<CUcontext>(m_context);
 }
+
+runtime::gpu::CudaContextManager::~CudaContextManager()
+{
+    CUDA_SAFE_CALL(cuDevicePrimaryCtxRelease(m_device));
+}

src/ngraph/runtime/gpu/gpu_cuda_context_manager.hpp

@@ -39,7 +39,7 @@ namespace ngraph
                 std::shared_ptr<CUcontext> get_context() { return m_context_ptr; }
             protected:
                 CudaContextManager();
-                ~CudaContextManager() {}
+                ~CudaContextManager();
                 CUdevice m_device;
                 CUcontext m_context;
                 std::shared_ptr<CUcontext> m_context_ptr;