src/armnnUtils/TensorUtils.cpp

   1 //
   2 // Copyright © 2017 Arm Ltd. All rights reserved.
   3 // SPDX-License-Identifier: MIT
   4 //
   5
   6 #include "TensorUtils.hpp"
   7 #include <backendsCommon/ITensorHandle.hpp>
   8
   9 #include <boost/assert.hpp>
  10 #include <boost/format.hpp>
  11 #include <boost/numeric/conversion/cast.hpp>
  12
  13 namespace armnnUtils
  14 {
  15
  16 armnn::TensorShape GetTensorShape(unsigned int numberOfBatches,
  17                                   unsigned int numberOfChannels,
  18                                   unsigned int height,
  19                                   unsigned int width,
  20                                   const armnn::DataLayout dataLayout)
  21 {
  22     switch (dataLayout)
  23     {
  24         case armnn::DataLayout::NCHW:
  25             return armnn::TensorShape({numberOfBatches, numberOfChannels, height, width});
  26         case armnn::DataLayout::NHWC:
  27             return armnn::TensorShape({numberOfBatches, height, width, numberOfChannels});
  28         default:
  29             throw armnn::InvalidArgumentException("Unknown data layout ["
  30                                                   + std::to_string(static_cast<int>(dataLayout)) +
  31                                                   "]", CHECK_LOCATION());
  32     }
  33 }
  34
  35 armnn::TensorInfo GetTensorInfo(unsigned int numberOfBatches,
  36                                 unsigned int numberOfChannels,
  37                                 unsigned int height,
  38                                 unsigned int width,
  39                                 const armnn::DataLayout dataLayout,
  40                                 const armnn::DataType dataType)
  41 {
  42     switch (dataLayout)
  43     {
  44         case armnn::DataLayout::NCHW:
  45             return armnn::TensorInfo({numberOfBatches, numberOfChannels, height, width}, dataType);
  46         case armnn::DataLayout::NHWC:
  47             return armnn::TensorInfo({numberOfBatches, height, width, numberOfChannels}, dataType);
  48         default:
  49             throw armnn::InvalidArgumentException("Unknown data layout ["
  50                                                   + std::to_string(static_cast<int>(dataLayout)) +
  51                                                   "]", CHECK_LOCATION());
  52     }
  53 }
  54
  55 std::pair<float, float> FindMinMax(armnn::ITensorHandle* tensorHandle)
  56 {
  57     auto tensor_data = static_cast<const float *>(tensorHandle->Map(true));
  58     auto tensor_size = tensorHandle->GetShape().GetNumElements();
  59
  60     // Set min/max initially to first value in tensor
  61     float min = tensor_data[0];
  62     float max = tensor_data[0];
  63
  64     // Loop over rest of tensor and update min/max if necessary
  65     for (unsigned int val = 1; val < tensor_size; val++)
  66     {
  67         if (tensor_data[val] < min)
  68         {
  69             min = tensor_data[val];
  70         }
  71         else if (tensor_data[val] > max)
  72         {
  73             max = tensor_data[val];
  74         }
  75     }
  76
  77     tensorHandle->Unmap();
  78
  79     return std::make_pair(min, max);
  80 }
  81
  82 armnn::TensorShape ExpandDims(const armnn::TensorShape& tensorShape, int axis)
  83 {
  84     unsigned int outputDim = tensorShape.GetNumDimensions() + 1;
  85
  86     if (axis < -boost::numeric_cast<int>(outputDim) || axis > boost::numeric_cast<int>(tensorShape.GetNumDimensions()))
  87     {
  88         throw armnn::InvalidArgumentException(
  89             boost::str(boost::format("Invalid expansion axis %1% for %2%D input tensor. %3%") %
  90                        axis %
  91                        tensorShape.GetNumDimensions() %
  92                        CHECK_LOCATION().AsString()));
  93     }
  94
  95     if (axis < 0)
  96     {
  97         axis = boost::numeric_cast<int>(outputDim) + axis;
  98     }
  99
 100     std::vector<unsigned int> outputShape;
 101     for (unsigned int i = 0; i < tensorShape.GetNumDimensions(); ++i)
 102     {
 103         outputShape.push_back(tensorShape[i]);
 104     }
 105     outputShape.insert(outputShape.begin() + axis, 1);
 106
 107     return armnn::TensorShape(outputDim, outputShape.data());
 108 }
 109
 110 }