1 //
2 // Copyright © 2017 Arm Ltd. All rights reserved.
3 // SPDX-License-Identifier: MIT
4 //
5 #pragma once
6
7 #include <armnn/Tensor.hpp>
8 #include <armnn/DescriptorsFwd.hpp>
9
10 #include <armnn/utility/NumericCast.hpp>
11
12 #include <arm_compute/core/ITensor.h>
13 #include <arm_compute/core/TensorInfo.h>
14 #include <arm_compute/core/Types.h>
15 #include <arm_compute/core/Size2D.h>
16
17 #include <Half.hpp>
18
19 namespace armnn
20 {
21 class ITensorHandle;
22
23 namespace armcomputetensorutils
24 {
25
26 /// Utility function to map an armnn::DataType to corresponding arm_compute::DataType.
27 arm_compute::DataType GetArmComputeDataType(armnn::DataType dataType, bool multiScales);
28
29 /// Utility function used to set up an arm_compute::Coordinates from a vector of ArmNN Axes for reduction functions
30 arm_compute::Coordinates BuildArmComputeReductionCoordinates(size_t inputDimensions,
31 unsigned int originalInputRank,
32 const std::vector<unsigned int>& armnnAxes);
33
34 /// Utility function used to setup an arm_compute::TensorShape object from an armnn::TensorShape.
35 arm_compute::TensorShape BuildArmComputeTensorShape(const armnn::TensorShape& tensorShape);
36
37 /// Utility function used to setup an arm_compute::ITensorInfo object whose dimensions are based on the given
38 /// armnn::ITensorInfo.
39 arm_compute::TensorInfo BuildArmComputeTensorInfo(const armnn::TensorInfo& tensorInfo);
40
41 /// Utility function used to setup an arm_compute::ITensorInfo object whose dimensions are based on the given
42 /// armnn::ITensorInfo.
43 /// armnn::DataLayout.
44 arm_compute::TensorInfo BuildArmComputeTensorInfo(const armnn::TensorInfo& tensorInfo,
45 armnn::DataLayout dataLayout);
46
47 /// Utility function used to convert armnn::DataLayout to arm_compute::DataLayout
48 /// armnn::DataLayout.
49 arm_compute::DataLayout ConvertDataLayout(armnn::DataLayout dataLayout);
50
51 /// Utility function used to setup an arm_compute::PoolingLayerInfo object from given
52 /// armnn::Pooling2dDescriptor
53 /// bool fpMixedPrecision
54 arm_compute::PoolingLayerInfo BuildArmComputePoolingLayerInfo(const Pooling2dDescriptor& descriptor,
55 bool fpMixedPrecision = false);
56
57 /// Utility function to setup an arm_compute::NormalizationLayerInfo object from an armnn::NormalizationDescriptor.
58 arm_compute::NormalizationLayerInfo BuildArmComputeNormalizationLayerInfo(const NormalizationDescriptor& desc);
59
60 /// Utility function used to setup an arm_compute::PermutationVector object from an armnn::PermutationVector.
61 arm_compute::PermutationVector BuildArmComputePermutationVector(const armnn::PermutationVector& vector);
62
63 /// Utility function used to setup an arm_compute::PermutationVector object from an armnn::PermutationVector.
64 arm_compute::PermutationVector BuildArmComputeTransposeVector(const armnn::PermutationVector& vector);
65
66 /// Utility function used to setup an arm_compute::Size2D object from width and height values.
67 arm_compute::Size2D BuildArmComputeSize2D(const unsigned int width, const unsigned int height);
68
69 /// Gets the appropriate PixelValue for the input DataType
70 arm_compute::PixelValue GetPixelValue(arm_compute::ITensor& input, float pixelValue);
71
72 /// Utility function used to setup an arm_compute::PadStrideInfo object from an armnn layer descriptor.
73 template <typename Descriptor>
BuildArmComputePadStrideInfo(const Descriptor & descriptor)74 arm_compute::PadStrideInfo BuildArmComputePadStrideInfo(const Descriptor &descriptor)
75 {
76 return arm_compute::PadStrideInfo(descriptor.m_StrideX,
77 descriptor.m_StrideY,
78 descriptor.m_PadLeft,
79 descriptor.m_PadRight,
80 descriptor.m_PadTop,
81 descriptor.m_PadBottom,
82 arm_compute::DimensionRoundingType::FLOOR);
83 }
84
85 /// Sets up the given ArmCompute tensor's dimensions based on the given ArmNN tensor.
86 template <typename Tensor>
BuildArmComputeTensor(Tensor & tensor,const armnn::TensorInfo & tensorInfo)87 void BuildArmComputeTensor(Tensor& tensor, const armnn::TensorInfo& tensorInfo)
88 {
89 tensor.allocator()->init(BuildArmComputeTensorInfo(tensorInfo));
90 }
91
92 /// Sets up the given ArmCompute tensor's dimensions based on the given ArmNN tensor.
93 template <typename Tensor>
BuildArmComputeTensor(Tensor & tensor,const armnn::TensorInfo & tensorInfo,DataLayout dataLayout)94 void BuildArmComputeTensor(Tensor& tensor, const armnn::TensorInfo& tensorInfo, DataLayout dataLayout)
95 {
96 tensor.allocator()->init(BuildArmComputeTensorInfo(tensorInfo, dataLayout));
97 }
98
99 template <typename Tensor>
InitialiseArmComputeTensorEmpty(Tensor & tensor)100 void InitialiseArmComputeTensorEmpty(Tensor& tensor)
101 {
102 tensor.allocator()->allocate();
103 }
104
105 /// Utility function to free unused tensors after a workload is configured and prepared
106 template <typename Tensor>
FreeTensorIfUnused(std::unique_ptr<Tensor> & tensor)107 void FreeTensorIfUnused(std::unique_ptr<Tensor>& tensor)
108 {
109 if (tensor && !tensor->is_used())
110 {
111 tensor.reset(nullptr);
112 }
113 }
114
115 // Helper function to obtain byte offset into tensor data
GetTensorOffset(const arm_compute::ITensorInfo & info,uint32_t depthIndex,uint32_t batchIndex,uint32_t channelIndex,uint32_t y,uint32_t x)116 inline size_t GetTensorOffset(const arm_compute::ITensorInfo& info,
117 uint32_t depthIndex,
118 uint32_t batchIndex,
119 uint32_t channelIndex,
120 uint32_t y,
121 uint32_t x)
122 {
123 arm_compute::Coordinates coords;
124 coords.set(4, static_cast<int>(depthIndex));
125 coords.set(3, static_cast<int>(batchIndex));
126 coords.set(2, static_cast<int>(channelIndex));
127 coords.set(1, static_cast<int>(y));
128 coords.set(0, static_cast<int>(x));
129 return armnn::numeric_cast<size_t>(info.offset_element_in_bytes(coords));
130 }
131
132 // Helper function to obtain element offset into data buffer representing tensor data (assuming no strides).
GetLinearBufferOffset(const arm_compute::ITensorInfo & info,uint32_t depthIndex,uint32_t batchIndex,uint32_t channelIndex,uint32_t y,uint32_t x)133 inline size_t GetLinearBufferOffset(const arm_compute::ITensorInfo& info,
134 uint32_t depthIndex,
135 uint32_t batchIndex,
136 uint32_t channelIndex,
137 uint32_t y,
138 uint32_t x)
139 {
140 const arm_compute::TensorShape& shape = info.tensor_shape();
141 uint32_t width = static_cast<uint32_t>(shape[0]);
142 uint32_t height = static_cast<uint32_t>(shape[1]);
143 uint32_t numChannels = static_cast<uint32_t>(shape[2]);
144 uint32_t numBatches = static_cast<uint32_t>(shape[3]);
145 return (((depthIndex * numBatches + batchIndex) * numChannels + channelIndex) * height + y) * width + x;
146 }
147
148 template <typename T>
CopyArmComputeITensorData(const arm_compute::ITensor & srcTensor,T * dstData)149 void CopyArmComputeITensorData(const arm_compute::ITensor& srcTensor, T* dstData)
150 {
151 // If MaxNumOfTensorDimensions is increased, this loop will need fixing.
152 static_assert(MaxNumOfTensorDimensions == 5, "Please update CopyArmComputeITensorData");
153 {
154 const arm_compute::ITensorInfo& info = *srcTensor.info();
155 const arm_compute::TensorShape& shape = info.tensor_shape();
156 const uint8_t* const bufferPtr = srcTensor.buffer();
157 uint32_t width = static_cast<uint32_t>(shape[0]);
158 uint32_t height = static_cast<uint32_t>(shape[1]);
159 uint32_t numChannels = static_cast<uint32_t>(shape[2]);
160 uint32_t numBatches = static_cast<uint32_t>(shape[3]);
161 uint32_t depth = static_cast<uint32_t>(shape[4]);
162
163 for (unsigned int depthIndex = 0; depthIndex < depth; ++depthIndex)
164 {
165 for (unsigned int batchIndex = 0; batchIndex < numBatches; ++batchIndex)
166 {
167 for (unsigned int channelIndex = 0; channelIndex < numChannels; ++channelIndex)
168 {
169 for (unsigned int y = 0; y < height; ++y)
170 {
171 // Copies one row from arm_compute tensor buffer to linear memory buffer.
172 // A row is the largest contiguous region we can copy, as the tensor data may be using strides.
173 memcpy(
174 dstData + GetLinearBufferOffset(info, depthIndex, batchIndex, channelIndex, y, 0),
175 bufferPtr + GetTensorOffset(info, depthIndex, batchIndex, channelIndex, y, 0),
176 width * sizeof(T));
177 }
178 }
179 }
180 }
181 }
182 }
183
184 template <typename T>
CopyArmComputeITensorData(const T * srcData,arm_compute::ITensor & dstTensor)185 void CopyArmComputeITensorData(const T* srcData, arm_compute::ITensor& dstTensor)
186 {
187 // If MaxNumOfTensorDimensions is increased, this loop will need fixing.
188 static_assert(MaxNumOfTensorDimensions == 5, "Please update CopyArmComputeITensorData");
189 {
190 const arm_compute::ITensorInfo& info = *dstTensor.info();
191 const arm_compute::TensorShape& shape = info.tensor_shape();
192 uint8_t* const bufferPtr = dstTensor.buffer();
193 uint32_t width = static_cast<uint32_t>(shape[0]);
194 uint32_t height = static_cast<uint32_t>(shape[1]);
195 uint32_t numChannels = static_cast<uint32_t>(shape[2]);
196 uint32_t numBatches = static_cast<uint32_t>(shape[3]);
197 uint32_t depth = static_cast<uint32_t>(shape[4]);
198
199 for (unsigned int depthIndex = 0; depthIndex < depth; ++depthIndex)
200 {
201 for (unsigned int batchIndex = 0; batchIndex < numBatches; ++batchIndex)
202 {
203 for (unsigned int channelIndex = 0; channelIndex < numChannels; ++channelIndex)
204 {
205 for (unsigned int y = 0; y < height; ++y)
206 {
207 // Copies one row from linear memory buffer to arm_compute tensor buffer.
208 // A row is the largest contiguous region we can copy, as the tensor data may be using strides.
209 memcpy(
210 bufferPtr + GetTensorOffset(info, depthIndex, batchIndex, channelIndex, y, 0),
211 srcData + GetLinearBufferOffset(info, depthIndex, batchIndex, channelIndex, y, 0),
212 width * sizeof(T));
213 }
214 }
215 }
216 }
217 }
218 }
219
220 /// Construct a TensorShape object from an ArmCompute object based on arm_compute::Dimensions.
221 /// \tparam ArmComputeType Any type that implements the Dimensions interface
222 /// \tparam T Shape value type
223 /// \param shapelike An ArmCompute object that implements the Dimensions interface
224 /// \param initial A default value to initialise the shape with
225 /// \return A TensorShape object filled from the Acl shapelike object.
226 template<typename ArmComputeType, typename T>
GetTensorShape(const ArmComputeType & shapelike,T initial)227 TensorShape GetTensorShape(const ArmComputeType& shapelike, T initial)
228 {
229 std::vector<unsigned int> s(MaxNumOfTensorDimensions, initial);
230 for (unsigned int i=0; i < shapelike.num_dimensions(); ++i)
231 {
232 s[(shapelike.num_dimensions()-1)-i] = armnn::numeric_cast<unsigned int>(shapelike[i]);
233 }
234 return TensorShape(armnn::numeric_cast<unsigned int>(shapelike.num_dimensions()), s.data());
235 };
236
237 /// Get the strides from an ACL strides object
GetStrides(const arm_compute::Strides & strides)238 inline TensorShape GetStrides(const arm_compute::Strides& strides)
239 {
240 return GetTensorShape(strides, 0U);
241 }
242
243 /// Get the shape from an ACL shape object
GetShape(const arm_compute::TensorShape & shape)244 inline TensorShape GetShape(const arm_compute::TensorShape& shape)
245 {
246 return GetTensorShape(shape, 1U);
247 }
248
249 } // namespace armcomputetensorutils
250 } // namespace armnn
251