Compute Library
 19.11
CLCropResize.cpp
Go to the documentation of this file.
1 /*
2  * Copyright (c) 2019 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  */
24 
26 
29 
30 #include <cstddef>
31 
32 namespace arm_compute
33 {
34 namespace
35 {
36 inline void configure_crop(const ICLTensor *input, ICLTensor *crop_boxes, ICLTensor *box_ind, ICLTensor *output, uint32_t crop_box_ind, Coordinates &start, Coordinates &end, uint32_t &batch_index)
37 {
38  batch_index = *(reinterpret_cast<int32_t *>(box_ind->ptr_to_element(Coordinates(crop_box_ind))));
39 
40  // _crop_box_ind is used to index crop_boxes and retrieve the appropriate crop box.
41  // The crop box is specified by normalized coordinates [y0, x0, y1, x1].
42  const float x0 = *reinterpret_cast<const float *>(crop_boxes->ptr_to_element(Coordinates(1, crop_box_ind)));
43  const float y0 = *reinterpret_cast<const float *>(crop_boxes->ptr_to_element(Coordinates(0, crop_box_ind)));
44  const float x1 = *reinterpret_cast<const float *>(crop_boxes->ptr_to_element(Coordinates(3, crop_box_ind)));
45  const float y1 = *reinterpret_cast<const float *>(crop_boxes->ptr_to_element(Coordinates(2, crop_box_ind)));
46  // The normalized coordinates are scaled to retrieve the floating point image coordinates which are rounded to integers.
47  start = Coordinates(std::floor(x0 * (input->info()->tensor_shape()[1] - 1) + 0.5f),
48  std::floor(y0 * (input->info()->tensor_shape()[2] - 1) + 0.5f));
49  end = Coordinates(std::floor(x1 * (input->info()->tensor_shape()[1] - 1) + 0.5f),
50  std::floor(y1 * (input->info()->tensor_shape()[2] - 1) + 0.5f));
51  const TensorShape out_shape(input->info()->tensor_shape()[0], static_cast<uint32_t>(abs(end[0] - start[0])) + 1, static_cast<uint32_t>(abs(end[1] - start[1])) + 1);
52  output->info()->set_tensor_shape(out_shape);
53 }
54 
55 inline void run_crop(const ICLTensor *input, ICLTensor *output, uint32_t batch_index, Coordinates start, Coordinates end, float extrapolation_value)
56 {
57  bool is_width_flipped = end[0] < start[0];
58  bool is_height_flipped = end[1] < start[1];
59  /** The number of rows out of bounds at the start and end of output. */
60  std::array<int32_t, 2> rows_out_of_bounds{ 0 };
61  /** The number of columns out of bounds at the start and end of output. */
62  std::array<int32_t, 2> cols_out_of_bounds{ 0 };
63  if(is_height_flipped)
64  {
65  rows_out_of_bounds[0] = start[1] >= static_cast<int32_t>(input->info()->dimension(2)) ? std::min(start[1] - input->info()->dimension(2) + 1, output->info()->dimension(2)) : 0;
66  rows_out_of_bounds[1] = end[1] < 0 ? std::min(-end[1], static_cast<int32_t>(output->info()->dimension(2))) : 0;
67  }
68  else
69  {
70  rows_out_of_bounds[0] = start[1] < 0 ? std::min(-start[1], static_cast<int32_t>(output->info()->dimension(2))) : 0;
71  rows_out_of_bounds[1] = end[1] >= static_cast<int32_t>(input->info()->dimension(2)) ? std::min(end[1] - input->info()->dimension(2) + 1, output->info()->dimension(2)) : 0;
72  }
73  if(is_width_flipped)
74  {
75  cols_out_of_bounds[0] = start[0] >= static_cast<int32_t>(input->info()->dimension(1)) ? std::min(start[0] - input->info()->dimension(1) + 1, output->info()->dimension(1)) : 0;
76  cols_out_of_bounds[1] = end[0] < 0 ? std::min(-end[0], static_cast<int32_t>(output->info()->dimension(1))) : 0;
77  }
78  else
79  {
80  cols_out_of_bounds[0] = start[0] < 0 ? std::min(-start[0], static_cast<int32_t>(output->info()->dimension(1))) : 0;
81  cols_out_of_bounds[1] = end[0] >= static_cast<int32_t>(input->info()->dimension(1)) ? std::min(end[0] - input->info()->dimension(1) + 1, output->info()->dimension(1)) : 0;
82  }
83 
84  Window full_window = calculate_max_window(*output->info());
85 
86  // Full output window:
87  // --------------------------------
88  // | Out of bounds |
89  // | rows before |
90  // |------------------------------|
91  // | Out of | In | Out of |
92  // | bounds | bounds | bounds |
93  // | cols | elements | cols |
94  // | before | copied | after |
95  // | | from input | |
96  // |------------------------------|
97  // | Out of bounds |
98  // | rows after |
99  // |------------------------------|
100  // Use a separate output window for each section of the full output window.
101  // Fill all output rows that have no elements that are within the input bounds
102  // with the extrapolation value using memset.
103  // First for the rows before the in bounds rows.
104  if(rows_out_of_bounds[0] > 0)
105  {
106  Window slice_fill_rows_before(full_window);
107  slice_fill_rows_before.set(2, Window::Dimension(0, rows_out_of_bounds[0], 1));
108  auto kernel = arm_compute::support::cpp14::make_unique<CLMemsetKernel>();
109  kernel->configure(output, extrapolation_value, &slice_fill_rows_before);
110  CLScheduler::get().enqueue(*kernel);
111  }
112 
113  Window slice_in(full_window);
114  slice_in.set(2, Window::Dimension(rows_out_of_bounds[0], output->info()->dimension(2) - rows_out_of_bounds[1], 1));
115  slice_in.set(1, Window::Dimension(cols_out_of_bounds[0], output->info()->dimension(1) - cols_out_of_bounds[1], 1));
116 
117  int rows_in_bounds = static_cast<int32_t>(output->info()->dimension(2)) - rows_out_of_bounds[0] - rows_out_of_bounds[1];
118  if(rows_in_bounds > 0)
119  {
120  // Fill all elements that share a row with an in bounds element with the extrapolation value.
121  if(cols_out_of_bounds[0] > 0)
122  {
123  Window slice_fill_cols_before(slice_in);
124  slice_fill_cols_before.set(1, Window::Dimension(0, cols_out_of_bounds[0], 1));
125  auto kernel = arm_compute::support::cpp14::make_unique<CLMemsetKernel>();
126  kernel->configure(output, extrapolation_value, &slice_fill_cols_before);
127  CLScheduler::get().enqueue(*kernel);
128  }
129 
130  if(cols_out_of_bounds[1] > 0)
131  {
132  Window slice_fill_cols_after(slice_in);
133  slice_fill_cols_after.set(1, Window::Dimension(output->info()->dimension(1) - cols_out_of_bounds[1], output->info()->dimension(1), 1));
134  auto kernel = arm_compute::support::cpp14::make_unique<CLMemsetKernel>();
135  kernel->configure(output, extrapolation_value, &slice_fill_cols_after);
136  CLScheduler::get().enqueue(*kernel);
137  }
138 
139  // Copy all elements within the input bounds from the input tensor.
140  int cols_in_bounds = static_cast<int32_t>(output->info()->dimension(1)) - cols_out_of_bounds[0] - cols_out_of_bounds[1];
141  if(cols_in_bounds > 0)
142  {
143  Coordinates2D start_in{ is_width_flipped ? start[0] - cols_out_of_bounds[0] : start[0] + cols_out_of_bounds[0],
144  is_height_flipped ? start[1] - rows_out_of_bounds[0] : start[1] + rows_out_of_bounds[0] };
145  Coordinates2D end_in{ is_width_flipped ? start_in.x - cols_in_bounds + 1 : start_in.x + cols_in_bounds - 1,
146  is_height_flipped ? start_in.y - rows_in_bounds + 1 : start_in.y + rows_in_bounds - 1 };
147  auto kernel = arm_compute::support::cpp14::make_unique<CLCropKernel>();
148 
149  kernel->configure(input, output, start_in, end_in, batch_index, extrapolation_value, &slice_in);
150  CLScheduler::get().enqueue(*kernel);
151  }
152  }
153 
154  // Fill all rows after the in bounds elements with the extrapolation value.
155  if(rows_out_of_bounds[1] > 0)
156  {
157  Window slice_fill_rows_after(full_window);
158  slice_fill_rows_after.set(2, Window::Dimension(output->info()->dimension(2) - rows_out_of_bounds[1], output->info()->dimension(2), 1));
159  auto kernel = arm_compute::support::cpp14::make_unique<CLMemsetKernel>();
160  kernel->configure(output, extrapolation_value, &slice_fill_rows_after);
161  CLScheduler::get().enqueue(*kernel);
162  }
163 }
164 } // namespace
165 
167  : _input(nullptr), _boxes(nullptr), _box_ind(nullptr), _output(nullptr), _num_boxes(0), _method(), _extrapolation_value(0), _scale(), _copy(), _crop_results(), _scaled_results()
168 {
169 }
170 
172  Coordinates2D crop_size, InterpolationPolicy method, float extrapolation_value)
173 {
174  ARM_COMPUTE_RETURN_ERROR_ON(crop_size.x <= 0 || crop_size.y <= 0);
176  ARM_COMPUTE_RETURN_ERROR_ON(boxes->tensor_shape()[0] != 4);
177  ARM_COMPUTE_RETURN_ERROR_ON(boxes->tensor_shape()[1] != box_ind->tensor_shape()[0]);
178  TensorInfo temp_info;
179  ARM_COMPUTE_RETURN_ON_ERROR(CLCropKernel::validate(input->clone().get(), &temp_info, { 0, 0 }, { 1, 1 }, input->dimension(3) - 1, extrapolation_value));
180  if(output->total_size() > 0)
181  {
184  TensorShape out_shape(input->tensor_shape()[0], crop_size.x, crop_size.y, boxes->tensor_shape()[1]);
186  }
187  return Status{};
188 }
189 
190 void CLCropResize::configure(const ICLTensor *input, ICLTensor *boxes, ICLTensor *box_ind, ICLTensor *output, Coordinates2D crop_size,
191  InterpolationPolicy method, float extrapolation_value)
192 {
194  ARM_COMPUTE_ERROR_THROW_ON(CLCropResize::validate(input->info(), boxes->info(), box_ind->info(), output->info(), crop_size, method, extrapolation_value));
195 
196  _num_boxes = boxes->info()->tensor_shape()[1];
197  TensorShape out_shape(input->info()->tensor_shape()[0], crop_size.x, crop_size.y);
198 
199  _input = input;
200  _boxes = boxes;
201  _box_ind = box_ind;
202  _output = output;
203  _method = method;
204  _extrapolation_value = extrapolation_value;
205 
206  // For each crop box:
207  // - The initial cropped image is produced as specified by boxes[i] from the 3D image input[box_ind[i]].
208  // Possibly using a CLCropKernel and up to four CLMemsetKernels.
209  // - A tensor is required to hold this initial cropped image.
210  // - A scale function is used to resize the cropped image to the size specified by crop_size.
211  // - A tensor is required to hold the final scaled image before it is copied into the 4D output
212  // that will hold all final cropped and scaled 3D images using CLCopyKernel.
213  for(unsigned int i = 0; i < _num_boxes; ++i)
214  {
215  auto crop_tensor = support::cpp14::make_unique<CLTensor>();
216  TensorInfo crop_result_info(1, DataType::F32);
217  crop_result_info.set_data_layout(DataLayout::NHWC);
218  crop_tensor->allocator()->init(crop_result_info);
219  _crop_results.emplace_back(std::move(crop_tensor));
220 
221  auto scale_tensor = support::cpp14::make_unique<CLTensor>();
222  TensorInfo scaled_result_info(out_shape, 1, DataType::F32);
223  scaled_result_info.set_data_layout(DataLayout::NHWC);
224  scale_tensor->allocator()->init(scaled_result_info);
225  _scaled_results.emplace_back(std::move(scale_tensor));
226  }
227 }
228 
230 {
231  ARM_COMPUTE_ERROR_ON_MSG(_output == nullptr, "Unconfigured function");
232  // The contents of _boxes and _box_ind are required to calculate the shape
233  // of the initial cropped image and thus are required to configure the
234  // kernels used for cropping and scaling.
235  _boxes->map(CLScheduler::get().queue());
236  _box_ind->map(CLScheduler::get().queue());
237  for(unsigned int i = 0; i < _num_boxes; ++i)
238  {
239  // Size of the crop box in _boxes and thus the shape of _crop_results[i]
240  // may not be known until run-time and so the kernels cannot be configured until then.
241  uint32_t batch_index;
242  Coordinates start{};
243  Coordinates end{};
244  configure_crop(_input, _boxes, _box_ind, _crop_results[i].get(), i, start, end, batch_index);
245 
246  auto scale_kernel = support::cpp14::make_unique<CLScale>();
248  _scale.emplace_back(std::move(scale_kernel));
249 
251  win.set(3, Window::Dimension(i, i + 1, 1));
252 
253  auto copy_kernel = support::cpp14::make_unique<CLCopyKernel>();
254  copy_kernel->configure(_scaled_results[i].get(), _output, PaddingList(), &win);
255  _copy.emplace_back(std::move(copy_kernel));
256 
257  _crop_results[i]->allocator()->allocate();
258  _scaled_results[i]->allocator()->allocate();
259 
260  run_crop(_input, _crop_results[i].get(), batch_index, start, end, _extrapolation_value);
261  }
262  _boxes->unmap(CLScheduler::get().queue());
263  _box_ind->unmap(CLScheduler::get().queue());
265  for(auto &kernel : _scale)
266  {
267  kernel->run();
268  }
270  for(auto &kernel : _copy)
271  {
272  CLScheduler::get().enqueue(*kernel, true);
273  }
275 }
276 } // namespace arm_compute
Class describing the value of a pixel for any image format.
Definition: PixelValue.h:34
InterpolationPolicy
Interpolation method.
Definition: Types.h:365
std::vector< std::unique_ptr< CLTensor > > _crop_results
Definition: CLCropResize.h:110
CLCropResize()
Default constructor.
void map(cl::CommandQueue &q, bool blocking=true)
Enqueue a map operation of the allocated buffer on the given queue.
Definition: ICLTensor.cpp:35
Shape of a tensor.
Definition: TensorShape.h:39
std::vector< PaddingInfo > PaddingList
List of padding information.
Definition: Types.h:454
static CLScheduler & get()
Access the scheduler singleton.
Definition: CLScheduler.cpp:99
static Status validate(const ITensorInfo *input, const ITensorInfo *output, Coordinates2D start, Coordinates2D end, uint32_t batch_index, float extrapolation_value=0, Window *output_window=nullptr)
Static function to check if given info will lead to a valid configuration of CLStridedSliceKernel.
#define ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(...)
Definition: Validate.h:494
2D Coordinates structure
Definition: types.h:28
#define ARM_COMPUTE_RETURN_ON_ERROR(status)
Checks if a status contains an error and returns it.
Definition: Error.h:204
1 channel, 1 F32 per channel
Store the tensor's metadata.
Definition: ITensorInfo.h:40
#define ARM_COMPUTE_ERROR_THROW_ON(status)
Definition: Error.h:455
Describe one of the image's dimensions with a start, end and step.
Definition: Window.h:75
std::vector< std::unique_ptr< CLCopyKernel > > _copy
Definition: CLCropResize.h:109
Status class.
Definition: Error.h:52
#define ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_NOT_IN(t,...)
Definition: Validate.h:694
#define ARM_COMPUTE_RETURN_ERROR_ON(cond)
If the condition is true, an error is returned.
Definition: Error.h:296
Window calculate_max_window(const ValidRegion &valid_region, const Steps &steps=Steps(), bool skip_border=false, BorderSize border_size=BorderSize())
Calculate the maximum window for a given tensor shape and border setting.
Definition: Helpers.cpp:28
Copyright (c) 2017-2019 ARM Limited.
int32_t x
X coordinates.
Definition: Types.h:438
std::vector< std::unique_ptr< CLTensor > > _scaled_results
Definition: CLCropResize.h:111
InterpolationPolicy _method
Definition: CLCropResize.h:105
ITensorInfo & set_data_layout(const DataLayout &data_layout) override
Set the data layout of the tensor.
Definition: TensorInfo.cpp:378
int32_t y
Y coordinates.
Definition: Types.h:439
virtual const TensorShape & tensor_shape() const =0
Size for each dimension of the tensor.
#define ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(...)
Definition: Validate.h:288
void unmap(cl::CommandQueue &q)
Enqueue an unmap operation of the allocated and mapped buffer on the given queue.
Definition: ICLTensor.cpp:40
#define ARM_COMPUTE_ERROR_ON_MSG(cond, msg)
Definition: Error.h:456
Coordinates of an item.
Definition: Coordinates.h:37
static Status validate(const ITensorInfo *input, ITensorInfo *boxes, ITensorInfo *box_ind, const ITensorInfo *output, Coordinates2D crop_size, InterpolationPolicy method, float extrapolation_value)
Static function to check if given info will lead to a valid configuration of NESlice.
void run() override
Run the kernels contained in the function.
virtual ITensorInfo * info() const =0
Interface to be implemented by the child class to return the tensor's metadata.
Samples are taken at pixel top left corner.
const ICLTensor * _input
Definition: CLCropResize.h:100
void set(size_t dimension, const Dimension &dim)
Set the values of a given dimension.
Definition: Window.inl:49
int x
The x coordinate.
Definition: types.h:30
Output values are determined by averaging the source pixels whose areas fall under the area of the de...
void enqueue(ICLKernel &kernel, bool flush=true)
Schedule the execution of the passed kernel if possible.
void sync()
Blocks until all commands in the associated command queue have finished.
Definition: CLScheduler.cpp:67
#define ARM_COMPUTE_ERROR_ON_NULLPTR(...)
Definition: Validate.h:161
Interface for OpenCL tensor.
Definition: ICLTensor.h:42
virtual size_t total_size() const =0
Returns the total size of the tensor in bytes.
Coordinate type.
Definition: Types.h:436
Num samples, height, width, channels.
void configure(const ICLTensor *input, ICLTensor *boxes, ICLTensor *box_ind, ICLTensor *output, Coordinates2D crop_size, InterpolationPolicy method=InterpolationPolicy::BILINEAR, float extrapolation_value=0)
Configure kernel.
std::vector< std::unique_ptr< CLScale > > _scale
Definition: CLCropResize.h:108
Store the tensor's metadata.
Definition: TensorInfo.h:45
Describe a multidimensional execution window.
Definition: Window.h:39