ArmNN
 24.02
Conv3dImpl.cpp
Go to the documentation of this file.
1 //
2 // Copyright © 2021 Arm Ltd and Contributors. All rights reserved.
3 // SPDX-License-Identifier: MIT
4 //
5 
6 #include "Conv3dImpl.hpp"
7 
8 namespace armnn
9 {
10 
11 void Convolve3d(const TensorShape& rInputShape,
12  Decoder<float>& rInputDecoder,
13  const TensorShape& rOutputShape,
14  Encoder<float>& rOutputEncoder,
15  const TensorShape& rFilterShape,
16  Decoder<float>& rFilterDecoder,
17  bool biasEnabled,
18  Decoder<float>* pBiasDecoder,
19  DataLayout dataLayout,
20  unsigned int paddingTop,
21  unsigned int paddingLeft,
22  unsigned int paddingFront,
23  unsigned int xStride,
24  unsigned int yStride,
25  unsigned int zStride,
26  unsigned int xDilation,
27  unsigned int yDilation,
28  unsigned int zDilation)
29 {
30  if (biasEnabled && !pBiasDecoder)
31  {
32  throw InvalidArgumentException("Bias is enabled but the bias data is invalid");
33  }
34  const armnnUtils::DataLayoutIndexed dataLayoutIndexed(dataLayout);
35 
36  const unsigned int channelsIndex = dataLayoutIndexed.GetChannelsIndex();
37  const unsigned int heightIndex = dataLayoutIndexed.GetHeightIndex();
38  const unsigned int widthIndex = dataLayoutIndexed.GetWidthIndex();
39  const unsigned int depthIndex = dataLayoutIndexed.GetDepthIndex();
40 
41  const unsigned int inChannels = rInputShape[channelsIndex];
42  const unsigned int outChannels = rOutputShape[channelsIndex];
43 
44  const unsigned int batchSize = rOutputShape[0];
45  const unsigned int outputHeight = rOutputShape[heightIndex];
46  const unsigned int outputWidth = rOutputShape[widthIndex];
47  const unsigned int outputDepth = rOutputShape[depthIndex];
48  const unsigned int inputHeight = rInputShape[heightIndex];
49  const unsigned int inputWidth = rInputShape[widthIndex];
50  const unsigned int inputDepth = rInputShape[depthIndex];
51 
52  // Conv3d weights layout: [D,H,W,I,O]
53  const unsigned int filterDepth = rFilterShape[0];
54  const unsigned int filterHeight = rFilterShape[1];
55  const unsigned int filterWidth = rFilterShape[2];
56 
57  const std::vector<float> inputVec = rInputDecoder.DecodeTensor(rInputShape);
58  const std::vector<float> filterVec = rFilterDecoder.DecodeTensor(rFilterShape);
59 
60  const TensorShape biasShape{outChannels};
61  const std::vector<float> biasVec = biasEnabled ? pBiasDecoder->DecodeTensor(biasShape) : std::vector<float>();
62 
63  for (unsigned int batchIdx = 0; batchIdx < batchSize; batchIdx++)
64  {
65  for (unsigned int zOutput = 0; zOutput < outputDepth; zOutput++)
66  {
67  for (unsigned int xOutput = 0; xOutput < outputWidth; xOutput++)
68  {
69  for (unsigned int yOutput = 0; yOutput < outputHeight; yOutput++)
70  {
71  for (unsigned int cOutput = 0; cOutput < outChannels; cOutput++)
72  {
73  // This loop goes over each output element.
74  float sum = 0.0f;
75 
76  // Loop over each input channel.
77  for (unsigned int zFilter = 0; zFilter < filterDepth; zFilter++)
78  {
79  for (unsigned int yFilter = 0; yFilter < filterHeight; yFilter++)
80  {
81  for (unsigned int xFilter = 0; xFilter < filterWidth; xFilter++)
82  {
83  for (unsigned int cInput = 0; cInput < inChannels; cInput++)
84  {
85  // This loop goes over each input element for each output element.
86  unsigned int filterIndex = 0;
87 
88  // Conv3d weights layout: [D,H,W,I,O]
89  // Keep this implementation, as using DataLayoutIndexed::GetIndex
90  // causes large performance regression.
91  filterIndex = zFilter * filterHeight * filterWidth * inChannels * outChannels +
92  yFilter * filterWidth * inChannels * outChannels +
93  xFilter * inChannels * outChannels +
94  cInput * outChannels +
95  cOutput;
96 
97  unsigned int yInput = yOutput * yStride + yFilter * yDilation;
98  unsigned int xInput = xOutput * xStride + xFilter * xDilation;
99  unsigned int zInput = zOutput * zStride + zFilter * zDilation;
100 
101  float inputValue;
102 
103  // Check if we're in the padding.
104  if (yInput < paddingTop || yInput >= inputHeight + paddingTop ||
105  xInput < paddingLeft || xInput >= inputWidth + paddingLeft ||
106  zInput < paddingFront || zInput >= inputDepth + paddingFront)
107  {
108  inputValue = 0.0f;
109  }
110  else
111  {
112  unsigned int inputIndex = 0;
113 
114  // Keep this implementation, as using DataLayoutIndexed::GetIndex
115  // causes large performance regression.
116  if (dataLayoutIndexed.GetDataLayout() == DataLayout::NDHWC)
117  {
118  inputIndex =
119  batchIdx * inputDepth * inputHeight * inputWidth * inChannels +
120  (zInput-paddingFront) * inputHeight * inputWidth * inChannels +
121  (yInput-paddingTop) * inputWidth * inChannels +
122  (xInput-paddingLeft) * inChannels +
123  cInput;
124  }
125  else
126  {
127  // NCDHW DataLayout
128  inputIndex =
129  batchIdx * inputDepth * inputHeight * inputWidth * inChannels +
130  inputDepth * inputHeight * inputWidth * cInput +
131  (zInput-paddingFront) * inputHeight * inputWidth +
132  (yInput-paddingTop) * inputWidth +
133  xInput-paddingLeft;
134  }
135 
136  inputValue = inputVec[inputIndex];
137  }
138 
139  sum += filterVec[filterIndex] * inputValue;
140  }
141  }
142  }
143  }
144 
145  if (biasEnabled)
146  {
147  sum += biasVec[cOutput];
148  }
149 
150  unsigned int outIdx;
151  if (dataLayoutIndexed.GetDataLayout() == DataLayout::NDHWC)
152  {
153  outIdx = batchIdx * outputDepth * outputHeight * outputWidth * outChannels +
154  zOutput * outputHeight * outputWidth * outChannels +
155  yOutput * outputWidth * outChannels +
156  xOutput * outChannels +
157  cOutput;
158  }
159  else
160  {
161  // NCDHW DataLayout
162  outIdx = batchIdx * outputDepth * outputHeight * outputWidth * outChannels +
163  cOutput * outputDepth * outputHeight * outputWidth +
164  zOutput * outputHeight * outputWidth +
165  yOutput * outputWidth +
166  xOutput;
167  }
168 
169  rOutputEncoder[outIdx];
170  rOutputEncoder.Set(sum);
171  }
172  }
173  }
174  }
175  }
176 }
177 
178 } // namespace armnn
armnn::Decoder< float >
armnn::Encoder::Set
virtual void Set(IType right)=0
armnn::DataLayout
DataLayout
Definition: Types.hpp:62
Conv3dImpl.hpp
armnnUtils::DataLayoutIndexed
Provides access to the appropriate indexes for Channels, Height and Width based on DataLayout.
Definition: DataLayoutIndexed.hpp:17
armnnUtils::DataLayoutIndexed::GetDataLayout
armnn::DataLayout GetDataLayout() const
Definition: DataLayoutIndexed.hpp:22
armnnUtils::DataLayoutIndexed::GetHeightIndex
unsigned int GetHeightIndex() const
Definition: DataLayoutIndexed.hpp:24
armnn::DataLayout::NDHWC
@ NDHWC
armnn::TensorShape
Definition: Tensor.hpp:20
armnn::Encoder< float >
armnn::InvalidArgumentException
Definition: Exceptions.hpp:80
armnn::Decoder::DecodeTensor
virtual std::vector< float > DecodeTensor(const TensorShape &tensorShape, bool isDepthwise=false)=0
armnnUtils::DataLayoutIndexed::GetWidthIndex
unsigned int GetWidthIndex() const
Definition: DataLayoutIndexed.hpp:25
armnn::Convolve3d
void Convolve3d(const TensorShape &rInputShape, Decoder< float > &rInputDecoder, const TensorShape &rOutputShape, Encoder< float > &rOutputEncoder, const TensorShape &rFilterShape, Decoder< float > &rFilterDecoder, bool biasEnabled, Decoder< float > *pBiasDecoder, DataLayout dataLayout, unsigned int paddingTop, unsigned int paddingLeft, unsigned int paddingFront, unsigned int xStride, unsigned int yStride, unsigned int zStride, unsigned int xDilation, unsigned int yDilation, unsigned int zDilation)
Definition: Conv3dImpl.cpp:11
armnn
Copyright (c) 2021 ARM Limited and Contributors.
Definition: 01_00_quick_start.dox:6
armnnUtils::DataLayoutIndexed::GetChannelsIndex
unsigned int GetChannelsIndex() const
Definition: DataLayoutIndexed.hpp:23
armnnUtils::DataLayoutIndexed::GetDepthIndex
unsigned int GetDepthIndex() const
Definition: DataLayoutIndexed.hpp:26