Merge pull request #20725 from mologie:fix-dnn-tf-on-arm

* dnn: fix unaligned memory access crash on armv7

The getTensorContent function would return a Mat pointing to some
member of a Protobuf-encoded message. Protobuf does not make any
alignment guarantees, which results in a crash on armv7 when loading
models while bit 2 is set in /proc/cpu/alignment (or the relevant
kernel feature for alignment compatibility is disabled). Any read
attempt from the previously unaligned data member would send SIGBUS.

As workaround, this commit makes an aligned copy via existing clone
functionality in getTensorContent. The unsafe copy=false option is
removed. Unfortunately, a rather crude hack in PReLUSubgraph in fact
writes(!) to the Protobuf message. We limit ourselves to fixing the
alignment issues in this commit, and add getTensorContentRefUnaligned
to cover the write case with a safe memcpy. A FIXME marks the issue.

* dnn: reduce amount of .clone() calls

* dnn: update FIXME comment

Co-authored-by: Alexander Alekhin <alexander.a.alekhin@gmail.com>

modules/dnn/src/tensorflow/tf_graph_simplifier.cpp

@@ -19,6 +19,16 @@ CV__DNN_EXPERIMENTAL_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

@@ -122,8 +122,10 @@ 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());
 
@@ -2522,8 +2524,10 @@ 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());