Compute Library
 21.11
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"
30 #include "src/common/utils/Log.h"
33 
34 namespace arm_compute
35 {
36 namespace
37 {
38 uint32_t last_padding_dimension(const PaddingList &padding)
39 {
40  int last_padding_dim = padding.size() - 1;
41  for(; last_padding_dim >= 0; --last_padding_dim)
42  {
43  if(padding[last_padding_dim].first > 0 || padding[last_padding_dim].second > 0)
44  {
45  break;
46  }
47  }
48  return static_cast<uint32_t>(last_padding_dim);
49 }
50 } // namespace
51 
52 NEPadLayer::~NEPadLayer() = default;
53 
55  : _copy_function(), _pad_kernel(), _mode(), _padding(), _num_dimensions(0), _slice_functions(), _concat_functions(), _slice_results(), _concat_results()
56 {
57 }
58 
59 void NEPadLayer::configure_constant_mode(ITensor *input, ITensor *output, const PaddingList &padding, const PixelValue constant_value)
60 {
61  _pad_kernel = std::make_unique<NEPadLayerKernel>();
62  _pad_kernel->configure(input, output, padding, constant_value, PaddingMode::CONSTANT);
63 }
64 
65 void NEPadLayer::configure_reflect_symmetric_mode(ITensor *input, ITensor *output)
66 {
67  // Reflecting can be performed by effectively unfolding the input as follows:
68  // For each dimension starting at DimX:
69  // For before and after:
70  // Use strided slice to extract and reverse the part of the
71  // input / previously produced tensor required for the padding.
72  // Concatenate the before and after padding with the input / previously
73  // produced tensor along the current dimension.
74 
75  // Two strided slice functions will be required for each dimension padded as well as a
76  // concatenate function and the tensors to hold the temporary results.
77  _slice_functions.resize(2 * _num_dimensions);
78  _slice_results.resize(2 * _num_dimensions);
79  _concat_functions.resize(_num_dimensions);
80  _concat_results.resize(_num_dimensions - 1);
81 
82  Coordinates starts_before{};
83  Coordinates ends_before{};
84  Coordinates starts_after{};
85  Coordinates ends_after{};
86  Coordinates strides{};
87  ITensor *prev = input;
88  for(uint32_t i = 0; i < _num_dimensions; ++i)
89  {
90  // Values in strides from the previous dimensions need to be set to 1 to avoid reversing again.
91  if(i > 0)
92  {
93  strides.set(i - 1, 1);
94  }
95 
96  if(_padding[i].first > 0 || _padding[i].second > 0)
97  {
98  // Set the starts, ends, and strides values for the current dimension.
99  // Due to the bit masks passed to strided slice, the values below the current dimension in
100  // starts and ends will be ignored so do not need to be modified.
101  if(_mode == PaddingMode::REFLECT)
102  {
103  starts_before.set(i, _padding[i].first);
104  ends_before.set(i, 0);
105  starts_after.set(i, input->info()->dimension(i) - 2);
106  ends_after.set(i, input->info()->dimension(i) - _padding[i].second - 2);
107  strides.set(i, -1);
108  }
109  else
110  {
111  starts_before.set(i, _padding[i].first - 1);
112  ends_before.set(i, -1);
113  starts_after.set(i, input->info()->dimension(i) - 1);
114  ends_after.set(i, input->info()->dimension(i) - _padding[i].second - 1);
115  strides.set(i, -1);
116  }
117 
118  // Strided slice wraps negative indexes around to the end of the range,
119  // instead this should indicate use of the full range and so the bit mask will be modified.
120  const int32_t begin_mask_before = starts_before[i] < 0 ? ~0 : ~(1u << i);
121  const int32_t end_mask_before = ends_before[i] < 0 ? ~0 : ~(1u << i);
122  const int32_t begin_mask_after = starts_after[i] < 0 ? ~0 : ~(1u << i);
123  const int32_t end_mask_after = ends_after[i] < 0 ? ~0 : ~(1u << i);
124 
125  // Reflect the input values for the padding before and after the input.
126  std::vector<const ITensor *> concat_vector;
127  if(_padding[i].first > 0)
128  {
129  if(i < prev->info()->num_dimensions())
130  {
131  _slice_functions[2 * i].configure(prev, &_slice_results[2 * i], starts_before, ends_before, strides, begin_mask_before, end_mask_before);
132  concat_vector.emplace_back(&_slice_results[2 * i]);
133  }
134  else
135  {
136  // Performing the slice is unnecessary if the result would simply be a copy of the tensor.
137  concat_vector.push_back(prev);
138  }
139  }
140  concat_vector.push_back(prev);
141  if(_padding[i].second > 0)
142  {
143  if(i < prev->info()->num_dimensions())
144  {
145  _slice_functions[2 * i + 1].configure(prev, &_slice_results[2 * i + 1], starts_after, ends_after, strides, begin_mask_after, end_mask_after);
146  concat_vector.emplace_back(&_slice_results[2 * i + 1]);
147  }
148  else
149  {
150  // Performing the slice is unnecessary if the result would simply be a copy of the tensor.
151  concat_vector.push_back(prev);
152  }
153  }
154  // Concatenate the padding before and after with the input.
155  ITensor *out = (i == _num_dimensions - 1) ? output : &_concat_results[i];
156  _concat_functions[i].configure(concat_vector, out, i);
157  if(i != _num_dimensions - 1)
158  {
159  _concat_results[i].allocator()->allocate();
160  }
161  prev = out;
162  }
163  _slice_results[2 * i].allocator()->allocate();
164  _slice_results[2 * i + 1].allocator()->allocate();
165  }
166 }
167 
168 void NEPadLayer::configure(ITensor *input, ITensor *output, const PaddingList &padding, const PixelValue constant_value, const PaddingMode mode)
169 {
170  ARM_COMPUTE_ERROR_THROW_ON(validate(input->info(), output->info(), padding, constant_value, mode));
171  ARM_COMPUTE_LOG_PARAMS(input, output, padding, constant_value, mode);
172 
173  _padding = padding;
174  _mode = mode;
175 
176  const TensorShape padded_shape = misc::shape_calculator::compute_padded_shape(input->info()->tensor_shape(), _padding);
177 
178  auto_init_if_empty(*output->info(), input->info()->clone()->set_tensor_shape(padded_shape));
179 
180  // Find the last dimension requiring padding so that it is known when to write to output and whether any padding is applied.
181  _num_dimensions = last_padding_dimension(padding) + 1;
182  if(_num_dimensions > 0)
183  {
184  switch(_mode)
185  {
187  {
188  configure_constant_mode(input, output, padding, constant_value);
189  break;
190  }
193  {
194  configure_reflect_symmetric_mode(input, output);
195  break;
196  }
197  default:
198  ARM_COMPUTE_ERROR("Padding mode not supported.");
199  }
200  }
201  else
202  {
203  // Copy the input to the whole output if no padding is applied
204  _copy_function.configure(input, output);
205  }
206 }
207 
208 Status NEPadLayer::validate(const ITensorInfo *input, const ITensorInfo *output, const PaddingList &padding, const PixelValue constant_value, const PaddingMode mode)
209 {
210  ARM_COMPUTE_UNUSED(constant_value);
211 
212  const TensorShape padded_shape = misc::shape_calculator::compute_padded_shape(input->tensor_shape(), padding);
213 
214  if(output->total_size() > 0)
215  {
218  }
219 
220  switch(mode)
221  {
223  {
224  return NEPadLayerKernel::validate(input, output, padding, constant_value, mode);
225  }
228  {
229  for(uint32_t i = 0; i < padding.size(); ++i)
230  {
231  if(mode == PaddingMode::REFLECT)
232  {
233  ARM_COMPUTE_RETURN_ERROR_ON(padding[i].first >= input->dimension(i));
234  ARM_COMPUTE_RETURN_ERROR_ON(padding[i].second >= input->dimension(i));
235  }
236  else
237  {
238  ARM_COMPUTE_RETURN_ERROR_ON(padding[i].first > input->dimension(i));
239  ARM_COMPUTE_RETURN_ERROR_ON(padding[i].second > input->dimension(i));
240  }
241  }
242  break;
243  }
244  default:
245  {
246  ARM_COMPUTE_ERROR("Invalid mode");
247  }
248  }
249  return Status{};
250 }
251 
253 {
254  if(_num_dimensions > 0)
255  {
256  switch(_mode)
257  {
259  {
260  NEScheduler::get().schedule(_pad_kernel.get(), Window::DimZ);
261  break;
262  }
265  {
266  for(uint32_t i = 0; i < _num_dimensions; ++i)
267  {
268  if(_padding[i].first > 0 || _padding[i].second > 0)
269  {
270  if(_padding[i].first > 0 && _slice_results[2 * i].info()->total_size() > 0)
271  {
272  _slice_functions[2 * i].run();
273  }
274  if(_padding[i].second > 0 && _slice_results[2 * i + 1].info()->total_size() > 0)
275  {
276  _slice_functions[2 * i + 1].run();
277  }
278  _concat_functions[i].run();
279  }
280  }
281  break;
282  }
283  default:
284  ARM_COMPUTE_ERROR("Padding mode not supported.");
285  }
286  }
287  else
288  {
289  _copy_function.run();
290  }
291 }
292 } // 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:440
#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 CPU tensor.
Definition: ITensor.h:36
#define ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(...)
Definition: Validate.h:284
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:165
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:208
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:54
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:541
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:252
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:168
#define ARM_COMPUTE_LOG_PARAMS(...)
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