Merge pull request #10466 from dkurt:reduce_umat_try_2

* UMat blobs are wrapped

* Replace getUMat and getMat at OpenCLBackendWrapper

modules/dnn/src/layers/convolution_layer.cpp

@@ -709,6 +709,10 @@ public:
         inps.getUMatVector(inputs);
         outs.getUMatVector(outputs);
 
+        CV_Assert(outputs.size() == 1);
+        for (int i = 0; i < inputs.size(); ++i)
+            CV_Assert(inputs[i].u != outputs[0].u);
+
         int group = inputs[0].size[1] / umat_blobs[0].size[1];
 
         if (convolutionOp.empty())
@@ -913,7 +917,9 @@ public:
                name.c_str(), inputs[0]->size[0], inputs[0]->size[1], inputs[0]->size[2], inputs[0]->size[3],
                kernel.width, kernel.height, pad.width, pad.height,
                stride.width, stride.height, dilation.width, dilation.height);*/
-        CV_Assert(inputs.size() == (size_t)1 && inputs[0]->size[1] % blobs[0].size[1] == 0);
+        CV_Assert(inputs.size() == (size_t)1, inputs[0]->size[1] % blobs[0].size[1] == 0,
+                  outputs.size() == 1, inputs[0]->data != outputs[0].data);
+
         int ngroups = inputs[0]->size[1]/blobs[0].size[1];
         CV_Assert(outputs[0].size[1] % ngroups == 0);
         int k, outCn = blobs[0].size[0];

modules/dnn/test/test_caffe_importer.cpp

@@ -303,6 +303,14 @@ OCL_TEST(Reproducibility_ResNet50, Accuracy)
 
     Mat ref = blobFromNPY(_tf("resnet50_prob.npy"));
     normAssert(ref, out);
+
+    UMat out_umat;
+    net.forward(out_umat);
+    normAssert(ref, out_umat, "out_umat");
+
+    std::vector<UMat> out_umats;
+    net.forward(out_umats);
+    normAssert(ref, out_umats[0], "out_umat_vector");
 }
 
 TEST(Reproducibility_SqueezeNet_v1_1, Accuracy)
@@ -331,9 +339,14 @@ OCL_TEST(Reproducibility_SqueezeNet_v1_1, Accuracy)
     Mat input = blobFromImage(imread(_tf("googlenet_0.png")), 1.0f, Size(227,227), Scalar(), false);
     ASSERT_TRUE(!input.empty());
 
-    net.setInput(input);
+    // Firstly set a wrong input blob and run the model to receive a wrong output.
+    net.setInput(input * 2.0f);
     Mat out = net.forward();
 
+    // Then set a correct input blob to check CPU->GPU synchronization is working well.
+    net.setInput(input);
+    out = net.forward();
+
     Mat ref = blobFromNPY(_tf("squeezenet_v1.1_prob.npy"));
     normAssert(ref, out);
 }