Merge remote-tracking branch 'upstream/3.4' into merge-3.4

modules/dnn/misc/python/test/test_dnn.py

@@ -391,20 +391,22 @@ class dnn_test(NewOpenCVTests):
             raise unittest.SkipTest("Missing DNN test files (dnn/onnx/data/{input/output}_hidden_lstm.npy). "
                                     "Verify OPENCV_DNN_TEST_DATA_PATH configuration parameter.")
 
-        net = cv.dnn.readNet(model)
         input = np.load(input_file)
         # we have to expand the shape of input tensor because Python bindings cut 3D tensors to 2D
         # it should be fixed in future. see : https://github.com/opencv/opencv/issues/19091
         # please remove `expand_dims` after that
         input = np.expand_dims(input, axis=3)
         gold_output = np.load(output_file)
-        net.setInput(input)
 
         for backend, target in self.dnnBackendsAndTargets:
             printParams(backend, target)
 
+            net = cv.dnn.readNet(model)
+
             net.setPreferableBackend(backend)
             net.setPreferableTarget(target)
+
+            net.setInput(input)
             real_output = net.forward()
 
             normAssert(self, real_output, gold_output, "", getDefaultThreshold(target))

modules/dnn/src/tensorflow/tf_graph_simplifier.cpp

@@ -19,6 +19,16 @@ CV__DNN_INLINE_NS_BEGIN
 using ::google::protobuf::RepeatedField;
 using ::google::protobuf::MapPair;
 
+static Mat getTensorContentRef_(const tensorflow::TensorProto& tensor);
+static inline
+bool isAlignedMat(const Mat& m)
+{
+    int depth = m.depth();
+    int alignment = CV_ELEM_SIZE1(depth);
+    return (((size_t)m.data) & (alignment - 1)) == 0;
+}
+
+
 class TFNodeWrapper : public ImportNodeWrapper
 {
 public:
@@ -719,8 +729,19 @@ public:
     {
         if (!negativeScales)
         {
-            Mat scales = getTensorContent(inputNodes[1]->attr().at("value").tensor(), /*copy*/false);
-            scales *= -1;
+            Mat scalesRef = getTensorContentRef_(inputNodes[1]->attr().at("value").tensor());
+            // FIXME: This breaks the const guarantees of tensor() by writing to scalesRef
+            if (isAlignedMat(scalesRef))
+            {
+                scalesRef *= -1;
+            }
+            else
+            {
+                Mat scales = scalesRef.clone() * -1;
+                CV_Assert(scalesRef.isContinuous());
+                CV_Assert(scales.isContinuous());
+                memcpy(scalesRef.data, scales.data, scales.total() * scales.elemSize());
+            }
         }
     }
 
@@ -832,7 +853,8 @@ void RemoveIdentityOps(tensorflow::GraphDef& net)
     }
 }
 
-Mat getTensorContent(const tensorflow::TensorProto &tensor, bool copy)
+// NB: returned Mat::data pointer may be unaligned
+Mat getTensorContentRef_(const tensorflow::TensorProto& tensor)
 {
     const std::string& content = tensor.tensor_content();
     Mat m;
@@ -904,7 +926,18 @@ Mat getTensorContent(const tensorflow::TensorProto &tensor, bool copy)
             CV_Error(Error::StsError, "Tensor's data type is not supported");
             break;
     }
-    return copy ? m.clone() : m;
+
+    return m;
+}
+
+Mat getTensorContent(const tensorflow::TensorProto& tensor, bool forceCopy)
+{
+    // If necessary clone m to have aligned data pointer
+    Mat m = getTensorContentRef_(tensor);
+    if (forceCopy || !isAlignedMat(m))
+        return m.clone();
+    else
+        return m;
 }
 
 void releaseTensor(tensorflow::TensorProto* tensor)

modules/dnn/src/tensorflow/tf_graph_simplifier.hpp

@@ -21,7 +21,7 @@ void RemoveIdentityOps(tensorflow::GraphDef& net);
 
 void simplifySubgraphs(tensorflow::GraphDef& net);
 
-Mat getTensorContent(const tensorflow::TensorProto &tensor, bool copy = true);
+Mat getTensorContent(const tensorflow::TensorProto& tensor, bool forceCopy = true);
 
 void releaseTensor(tensorflow::TensorProto* tensor);
 

modules/dnn/src/tensorflow/tf_importer.cpp

@@ -124,8 +124,10 @@ void parseTensor(const tensorflow::TensorProto &tensor, Mat &dstBlob)
     }
 
     dstBlob.create(shape, CV_32F);
+    CV_Assert(dstBlob.isContinuous());
 
     Mat tensorContent = getTensorContent(tensor, /*no copy*/false);
+    CV_Assert(tensorContent.isContinuous());
     int size = tensorContent.total();
     CV_Assert(size == (int)dstBlob.total());
 
@@ -2671,8 +2673,10 @@ void TFImporter::kernelFromTensor(const tensorflow::TensorProto &tensor, Mat &ds
     out_c = shape[0]; input_c = shape[1];
 
     dstBlob.create(shape, CV_32F);
+    CV_Assert(dstBlob.isContinuous());
 
     Mat tensorContent = getTensorContent(tensor, /*no copy*/false);
+    CV_Assert(tensorContent.isContinuous());
     int size = tensorContent.total();
     CV_Assert(size == (int)dstBlob.total());