fix null stream initialization for multi-gpu systems

modules/core/src/cuda_stream.cpp

@@ -330,14 +330,11 @@ namespace cv { namespace cuda
         void initStreams();
         void initPools();
 
+        std::vector<Ptr<Stream> > streams_;
         Mutex streams_mtx_;
-        volatile bool streams_initialized_;
-
-        Mutex pools_mtx_;
-        volatile bool pools_initialized_;
 
-        std::vector<Ptr<Stream> > streams_;
         std::vector<MemoryPool> pools_;
+        Mutex pools_mtx_;
     };
 
     DefaultDeviceInitializer::DefaultDeviceInitializer()
@@ -358,60 +355,44 @@ namespace cv { namespace cuda
     }
 
     Stream& DefaultDeviceInitializer::getNullStream(int deviceId)
-    {
-        initStreams();
-
-        CV_DbgAssert( deviceId >= 0 && deviceId < static_cast<int>(streams_.size()) );
-
-        return *streams_[deviceId];
-    }
-
-    MemoryPool* DefaultDeviceInitializer::getMemoryPool(int deviceId)
-    {
-        initPools();
-
-        CV_DbgAssert( deviceId >= 0 && deviceId < static_cast<int>(pools_.size()) );
-
-        return &pools_[deviceId];
-    }
-
-    void DefaultDeviceInitializer::initStreams()
     {
         AutoLock lock(streams_mtx_);
 
-        if (!streams_initialized_)
+        if (streams_.empty())
         {
             int deviceCount = getCudaEnabledDeviceCount();
 
             if (deviceCount > 0)
-            {
                 streams_.resize(deviceCount);
+        }
+
+        CV_DbgAssert( deviceId >= 0 && deviceId < static_cast<int>(streams_.size()) );
 
-                for (int i = 0; i < deviceCount; ++i)
+        if (streams_[deviceId].empty())
         {
             cudaStream_t stream = NULL;
             Ptr<Stream::Impl> impl = makePtr<Stream::Impl>(stream);
-                    streams_[i] = Ptr<Stream>(new Stream(impl));
+            streams_[deviceId] = Ptr<Stream>(new Stream(impl));
-                }
         }
 
-            streams_initialized_ = true;
+        return *streams_[deviceId];
-        }
     }
 
-    void DefaultDeviceInitializer::initPools()
+    MemoryPool* DefaultDeviceInitializer::getMemoryPool(int deviceId)
     {
         AutoLock lock(pools_mtx_);
 
-        if (!pools_initialized_)
+        if (pools_.empty())
         {
             int deviceCount = getCudaEnabledDeviceCount();
 
             if (deviceCount > 0)
                 pools_.resize(deviceCount);
-
-            pools_initialized_ = true;
         }
+
+        CV_DbgAssert( deviceId >= 0 && deviceId < static_cast<int>(pools_.size()) );
+
+        return &pools_[deviceId];
     }
 
     DefaultDeviceInitializer initializer;