Compute Library
 21.02
NEPadLayer.cpp
Go to the documentation of this file.
1 /*
2  * Copyright (c) 2018-2021 Arm Limited.
3  *
4  * SPDX-License-Identifier: MIT
5  *
6  * Permission is hereby granted, free of charge, to any person obtaining a copy
7  * of this software and associated documentation files (the "Software"), to
8  * deal in the Software without restriction, including without limitation the
9  * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
10  * sell copies of the Software, and to permit persons to whom the Software is
11  * furnished to do so, subject to the following conditions:
12  *
13  * The above copyright notice and this permission notice shall be included in all
14  * copies or substantial portions of the Software.
15  *
16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
19  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
22  * SOFTWARE.
23  */
25 
27 
28 #include "arm_compute/core/Types.h"
32 
33 namespace arm_compute
34 {
35 namespace
36 {
37 uint32_t last_padding_dimension(const PaddingList &padding)
38 {
39  int last_padding_dim = padding.size() - 1;
40  for(; last_padding_dim >= 0; --last_padding_dim)
41  {
42  if(padding[last_padding_dim].first > 0 || padding[last_padding_dim].second > 0)
43  {
44  break;
45  }
46  }
47  return static_cast<uint32_t>(last_padding_dim);
48 }
49 } // namespace
50 
51 NEPadLayer::~NEPadLayer() = default;
52 
54  : _copy_function(), _pad_kernel(), _mode(), _padding(), _num_dimensions(0), _slice_functions(), _concat_functions(), _slice_results(), _concat_results()
55 {
56 }
57 
58 void NEPadLayer::configure_constant_mode(ITensor *input, ITensor *output, const PaddingList &padding, const PixelValue constant_value)
59 {
60  _pad_kernel = std::make_unique<NEPadLayerKernel>();
61  _pad_kernel->configure(input, output, padding, constant_value, PaddingMode::CONSTANT);
62 }
63 
64 void NEPadLayer::configure_reflect_symmetric_mode(ITensor *input, ITensor *output)
65 {
66  // Reflecting can be performed by effectively unfolding the input as follows:
67  // For each dimension starting at DimX:
68  // For before and after:
69  // Use strided slice to extract and reverse the part of the
70  // input / previously produced tensor required for the padding.
71  // Concatenate the before and after padding with the input / previously
72  // produced tensor along the current dimension.
73 
74  // Two strided slice functions will be required for each dimension padded as well as a
75  // concatenate function and the tensors to hold the temporary results.
76  _slice_functions.resize(2 * _num_dimensions);
77  _slice_results.resize(2 * _num_dimensions);
78  _concat_functions.resize(_num_dimensions);
79  _concat_results.resize(_num_dimensions - 1);
80 
81  Coordinates starts_before{};
82  Coordinates ends_before{};
83  Coordinates starts_after{};
84  Coordinates ends_after{};
85  Coordinates strides{};
86  ITensor *prev = input;
87  for(uint32_t i = 0; i < _num_dimensions; ++i)
88  {
89  // Values in strides from the previous dimensions need to be set to 1 to avoid reversing again.
90  if(i > 0)
91  {
92  strides.set(i - 1, 1);
93  }
94 
95  if(_padding[i].first > 0 || _padding[i].second > 0)
96  {
97  // Set the starts, ends, and strides values for the current dimension.
98  // Due to the bit masks passed to strided slice, the values below the current dimension in
99  // starts and ends will be ignored so do not need to be modified.
100  if(_mode == PaddingMode::REFLECT)
101  {
102  starts_before.set(i, _padding[i].first);
103  ends_before.set(i, 0);
104  starts_after.set(i, input->info()->dimension(i) - 2);
105  ends_after.set(i, input->info()->dimension(i) - _padding[i].second - 2);
106  strides.set(i, -1);
107  }
108  else
109  {
110  starts_before.set(i, _padding[i].first - 1);
111  ends_before.set(i, -1);
112  starts_after.set(i, input->info()->dimension(i) - 1);
113  ends_after.set(i, input->info()->dimension(i) - _padding[i].second - 1);
114  strides.set(i, -1);
115  }
116 
117  // Strided slice wraps negative indexes around to the end of the range,
118  // instead this should indicate use of the full range and so the bit mask will be modified.
119  const int32_t begin_mask_before = starts_before[i] < 0 ? ~0 : ~(1u << i);
120  const int32_t end_mask_before = ends_before[i] < 0 ? ~0 : ~(1u << i);
121  const int32_t begin_mask_after = starts_after[i] < 0 ? ~0 : ~(1u << i);
122  const int32_t end_mask_after = ends_after[i] < 0 ? ~0 : ~(1u << i);
123 
124  // Reflect the input values for the padding before and after the input.
125  std::vector<const ITensor *> concat_vector;
126  if(_padding[i].first > 0)
127  {
128  if(i < prev->info()->num_dimensions())
129  {
130  _slice_functions[2 * i].configure(prev, &_slice_results[2 * i], starts_before, ends_before, strides, begin_mask_before, end_mask_before);
131  concat_vector.emplace_back(&_slice_results[2 * i]);
132  }
133  else
134  {
135  // Performing the slice is unnecessary if the result would simply be a copy of the tensor.
136  concat_vector.push_back(prev);
137  }
138  }
139  concat_vector.push_back(prev);
140  if(_padding[i].second > 0)
141  {
142  if(i < prev->info()->num_dimensions())
143  {
144  _slice_functions[2 * i + 1].configure(prev, &_slice_results[2 * i + 1], starts_after, ends_after, strides, begin_mask_after, end_mask_after);
145  concat_vector.emplace_back(&_slice_results[2 * i + 1]);
146  }
147  else
148  {
149  // Performing the slice is unnecessary if the result would simply be a copy of the tensor.
150  concat_vector.push_back(prev);
151  }
152  }
153  // Concatenate the padding before and after with the input.
154  ITensor *out = (i == _num_dimensions - 1) ? output : &_concat_results[i];
155  _concat_functions[i].configure(concat_vector, out, i);
156  if(i != _num_dimensions - 1)
157  {
158  _concat_results[i].allocator()->allocate();
159  }
160  prev = out;
161  }
162  _slice_results[2 * i].allocator()->allocate();
163  _slice_results[2 * i + 1].allocator()->allocate();
164  }
165 }
166 
167 void NEPadLayer::configure(ITensor *input, ITensor *output, const PaddingList &padding, const PixelValue constant_value, const PaddingMode mode)
168 {
169  ARM_COMPUTE_ERROR_THROW_ON(validate(input->info(), output->info(), padding, constant_value, mode));
170 
171  _padding = padding;
172  _mode = mode;
173 
174  const TensorShape padded_shape = misc::shape_calculator::compute_padded_shape(input->info()->tensor_shape(), _padding);
175 
176  auto_init_if_empty(*output->info(), input->info()->clone()->set_tensor_shape(padded_shape));
177 
178  // Find the last dimension requiring padding so that it is known when to write to output and whether any padding is applied.
179  _num_dimensions = last_padding_dimension(padding) + 1;
180  if(_num_dimensions > 0)
181  {
182  switch(_mode)
183  {
185  {
186  configure_constant_mode(input, output, padding, constant_value);
187  break;
188  }
191  {
192  configure_reflect_symmetric_mode(input, output);
193  break;
194  }
195  default:
196  ARM_COMPUTE_ERROR("Padding mode not supported.");
197  }
198  }
199  else
200  {
201  // Copy the input to the whole output if no padding is applied
202  _copy_function.configure(input, output);
203  }
204 }
205 
206 Status NEPadLayer::validate(const ITensorInfo *input, const ITensorInfo *output, const PaddingList &padding, const PixelValue constant_value, const PaddingMode mode)
207 {
208  ARM_COMPUTE_UNUSED(constant_value);
209 
210  const TensorShape padded_shape = misc::shape_calculator::compute_padded_shape(input->tensor_shape(), padding);
211 
212  if(output->total_size() > 0)
213  {
216  }
217 
218  switch(mode)
219  {
221  {
222  return NEPadLayerKernel::validate(input, output, padding, constant_value, mode);
223  }
226  {
227  for(uint32_t i = 0; i < padding.size(); ++i)
228  {
229  if(mode == PaddingMode::REFLECT)
230  {
231  ARM_COMPUTE_RETURN_ERROR_ON(padding[i].first >= input->dimension(i));
232  ARM_COMPUTE_RETURN_ERROR_ON(padding[i].second >= input->dimension(i));
233  }
234  else
235  {
236  ARM_COMPUTE_RETURN_ERROR_ON(padding[i].first > input->dimension(i));
237  ARM_COMPUTE_RETURN_ERROR_ON(padding[i].second > input->dimension(i));
238  }
239  }
240  break;
241  }
242  default:
243  {
244  ARM_COMPUTE_ERROR("Invalid mode");
245  }
246  }
247  return Status{};
248 }
249 
251 {
252  if(_num_dimensions > 0)
253  {
254  switch(_mode)
255  {
257  {
258  NEScheduler::get().schedule(_pad_kernel.get(), Window::DimZ);
259  break;
260  }
263  {
264  for(uint32_t i = 0; i < _num_dimensions; ++i)
265  {
266  if(_padding[i].first > 0 || _padding[i].second > 0)
267  {
268  if(_padding[i].first > 0 && _slice_results[2 * i].info()->total_size() > 0)
269  {
270  _slice_functions[2 * i].run();
271  }
272  if(_padding[i].second > 0 && _slice_results[2 * i + 1].info()->total_size() > 0)
273  {
274  _slice_functions[2 * i + 1].run();
275  }
276  _concat_functions[i].run();
277  }
278  }
279  break;
280  }
281  default:
282  ARM_COMPUTE_ERROR("Padding mode not supported.");
283  }
284  }
285  else
286  {
287  _copy_function.run();
288  }
289 }
290 } // namespace arm_compute
Class describing the value of a pixel for any image format.
Definition: PixelValue.h:34
Shape of a tensor.
Definition: TensorShape.h:39
virtual size_t dimension(size_t index) const =0
Return the size of the requested dimension.
std::vector< PaddingInfo > PaddingList
List of padding information.
Definition: Types.h:481
#define ARM_COMPUTE_ERROR(msg)
Print the given message then throw an std::runtime_error.
Definition: Error.h:352
Store the tensor&#39;s metadata.
Definition: ITensorInfo.h:40
#define ARM_COMPUTE_ERROR_THROW_ON(status)
Definition: Error.h:455
Status class.
Definition: Error.h:52
#define ARM_COMPUTE_RETURN_ERROR_ON(cond)
If the condition is true, an error is returned.
Definition: Error.h:296
Interface for Neon tensor.
Definition: ITensor.h:36
#define ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(...)
Definition: Validate.h:288
Copyright (c) 2017-2021 Arm Limited.
#define ARM_COMPUTE_UNUSED(...)
To avoid unused variables warnings.
Definition: Error.h:152
virtual const TensorShape & tensor_shape() const =0
Size for each dimension of the tensor.
PaddingMode
Padding mode to use for PadLayer.
Definition: Types.h:169
static Status validate(const ITensorInfo *input, const ITensorInfo *output, const PaddingList &padding, const PixelValue constant_value=PixelValue(), const PaddingMode mode=PaddingMode::CONSTANT)
Static function to check if given info will lead to a valid configuration of NEPadLayer.
Definition: NEPadLayer.cpp:206
Coordinates of an item.
Definition: Coordinates.h:37
~NEPadLayer()
Default destructor.
bool auto_init_if_empty(ITensorInfo &info, const TensorShape &shape, int num_channels, DataType data_type, QuantizationInfo quantization_info=QuantizationInfo())
Auto initialize the tensor info (shape, number of channels and data type) if the current assignment i...
virtual std::unique_ptr< T > clone() const =0
Provide a clone of the current object of class T.
virtual ITensorInfo * info() const =0
Interface to be implemented by the child class to return the tensor&#39;s metadata.
TensorShape compute_padded_shape(const TensorShape &input_shape, const PaddingList &padding)
Calculate the padded shape of a tensor.
static Status validate(const ITensorInfo *input, const ITensorInfo *output, const PaddingList &padding, const PixelValue constant_value=PixelValue(), const PaddingMode mode=PaddingMode::CONSTANT)
Static function to check if given info will lead to a valid configuration of NEPadLayer.
NEPadLayer()
Default Constructor.
Definition: NEPadLayer.cpp:53
ScaleKernelInfo info(interpolation_policy, default_border_mode, PixelValue(), sampling_policy, false)
virtual size_t total_size() const =0
Returns the total size of the tensor in bytes.
virtual void schedule(ICPPKernel *kernel, const Hints &hints)=0
Runs the kernel in the same thread as the caller synchronously.
static constexpr size_t DimZ
Alias for dimension 2 also known as Z dimension.
Definition: Window.h:47
#define ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(...)
Definition: Validate.h:545
void run() override
Run the kernels contained in the function.
Definition: NECopy.cpp:66
void run() override
Run the kernels contained in the function.
Definition: NEPadLayer.cpp:250
void configure(ITensor *input, ITensor *output, const PaddingList &padding, const PixelValue constant_value=PixelValue(), const PaddingMode mode=PaddingMode::CONSTANT)
Initialize the function.
Definition: NEPadLayer.cpp:167
void configure(ITensor *input, ITensor *output)
Initialise the function&#39;s source and destination.
Definition: NECopy.cpp:48
static IScheduler & get()
Access the scheduler singleton.
Definition: Scheduler.cpp:94