1 /*
2 * Copyright (c) 2019-2020 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 #include "arm_compute/runtime/CL/functions/CLCropResize.h"
25
26 #include "arm_compute/core/CL/CLHelpers.h"
27 #include "arm_compute/runtime/CL/CLScheduler.h"
28 #include "src/core/CL/kernels/CLCopyKernel.h"
29 #include "src/core/CL/kernels/CLCropKernel.h"
30 #include "src/core/CL/kernels/CLFillBorderKernel.h"
31 #include "src/core/CL/kernels/CLMemsetKernel.h"
32 #include "src/core/helpers/AutoConfiguration.h"
33 #include "src/core/helpers/WindowHelpers.h"
34
35 #include "support/MemorySupport.h"
36
37 #include <cstddef>
38
39 namespace arm_compute
40 {
41 namespace
42 {
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)43 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)
44 {
45 batch_index = *(reinterpret_cast<int32_t *>(box_ind->ptr_to_element(Coordinates(crop_box_ind))));
46
47 // _crop_box_ind is used to index crop_boxes and retrieve the appropriate crop box.
48 // The crop box is specified by normalized coordinates [y0, x0, y1, x1].
49 const float x0 = *reinterpret_cast<const float *>(crop_boxes->ptr_to_element(Coordinates(1, crop_box_ind)));
50 const float y0 = *reinterpret_cast<const float *>(crop_boxes->ptr_to_element(Coordinates(0, crop_box_ind)));
51 const float x1 = *reinterpret_cast<const float *>(crop_boxes->ptr_to_element(Coordinates(3, crop_box_ind)));
52 const float y1 = *reinterpret_cast<const float *>(crop_boxes->ptr_to_element(Coordinates(2, crop_box_ind)));
53 // The normalized coordinates are scaled to retrieve the floating point image coordinates which are rounded to integers.
54 start = Coordinates(std::floor(x0 * (input->info()->tensor_shape()[1] - 1) + 0.5f),
55 std::floor(y0 * (input->info()->tensor_shape()[2] - 1) + 0.5f));
56 end = Coordinates(std::floor(x1 * (input->info()->tensor_shape()[1] - 1) + 0.5f),
57 std::floor(y1 * (input->info()->tensor_shape()[2] - 1) + 0.5f));
58 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);
59 output->info()->set_tensor_shape(out_shape);
60 }
61 } // namespace
62
CLCropResize()63 CLCropResize::CLCropResize()
64 : _input(nullptr), _boxes(nullptr), _box_ind(nullptr), _output(nullptr), _num_boxes(0), _method(), _extrapolation_value(0), _scale(), _copy(), _crop_results(), _scaled_results(), _internal_kernels()
65 {
66 }
67
68 CLCropResize::~CLCropResize() = default;
69
validate(const ITensorInfo * input,ITensorInfo * boxes,ITensorInfo * box_ind,const ITensorInfo * output,Coordinates2D crop_size,InterpolationPolicy method,float extrapolation_value)70 Status CLCropResize::validate(const ITensorInfo *input, ITensorInfo *boxes, ITensorInfo *box_ind, const ITensorInfo *output,
71 Coordinates2D crop_size, InterpolationPolicy method, float extrapolation_value)
72 {
73 ARM_COMPUTE_RETURN_ERROR_ON(crop_size.x <= 0 || crop_size.y <= 0);
74 ARM_COMPUTE_RETURN_ERROR_ON(method == InterpolationPolicy::AREA);
75 ARM_COMPUTE_RETURN_ERROR_ON(boxes->tensor_shape()[0] != 4);
76 ARM_COMPUTE_RETURN_ERROR_ON(boxes->tensor_shape()[1] != box_ind->tensor_shape()[0]);
77 TensorInfo temp_info;
78 ARM_COMPUTE_RETURN_ON_ERROR(CLCropKernel::validate(input->clone().get(), &temp_info, { 0, 0 }, { 1, 1 }, input->dimension(3) - 1, extrapolation_value));
79 if(output->total_size() > 0)
80 {
81 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_NOT_IN(output, DataType::F32);
82 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(input, output);
83 TensorShape out_shape(input->tensor_shape()[0], crop_size.x, crop_size.y, boxes->tensor_shape()[1]);
84 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), out_shape);
85 }
86 return Status{};
87 }
88
configure(const ICLTensor * input,ICLTensor * boxes,ICLTensor * box_ind,ICLTensor * output,Coordinates2D crop_size,InterpolationPolicy method,float extrapolation_value)89 void CLCropResize::configure(const ICLTensor *input, ICLTensor *boxes, ICLTensor *box_ind, ICLTensor *output, Coordinates2D crop_size,
90 InterpolationPolicy method, float extrapolation_value)
91 {
92 configure(CLKernelLibrary::get().get_compile_context(), input, boxes, box_ind, output, crop_size, method, extrapolation_value);
93 }
94
configure(const CLCompileContext & compile_context,const ICLTensor * input,ICLTensor * boxes,ICLTensor * box_ind,ICLTensor * output,Coordinates2D crop_size,InterpolationPolicy method,float extrapolation_value)95 void CLCropResize::configure(const CLCompileContext &compile_context, const ICLTensor *input, ICLTensor *boxes, ICLTensor *box_ind, ICLTensor *output, Coordinates2D crop_size,
96 InterpolationPolicy method, float extrapolation_value)
97 {
98 ARM_COMPUTE_ERROR_ON_NULLPTR(input, output, boxes, box_ind);
99 ARM_COMPUTE_ERROR_THROW_ON(CLCropResize::validate(input->info(), boxes->info(), box_ind->info(), output->info(), crop_size, method, extrapolation_value));
100
101 TensorShape output_shape = TensorShape(input->info()->tensor_shape()[0], crop_size.x, crop_size.y, boxes->info()->tensor_shape()[1]);
102 auto_init_if_empty(*output->info(), output_shape, 1, DataType::F32);
103
104 _num_boxes = boxes->info()->tensor_shape()[1];
105 TensorShape out_shape(input->info()->tensor_shape()[0], crop_size.x, crop_size.y);
106
107 _input = input;
108 _boxes = boxes;
109 _box_ind = box_ind;
110 _output = output;
111 _method = method;
112 _extrapolation_value = extrapolation_value;
113
114 // For each crop box:
115 // - The initial cropped image is produced as specified by boxes[i] from the 3D image input[box_ind[i]].
116 // Possibly using a CLCropKernel and up to four CLMemsetKernels.
117 // - A tensor is required to hold this initial cropped image.
118 // - A scale function is used to resize the cropped image to the size specified by crop_size.
119 // - A tensor is required to hold the final scaled image before it is copied into the 4D output
120 // that will hold all final cropped and scaled 3D images using CLCopyKernel.
121
122 // The contents of _boxes and _box_ind are required to calculate the shape
123 // of the initial cropped image and thus are required to configure the
124 // kernels used for cropping and scaling.
125 _boxes->map(CLScheduler::get().queue());
126 _box_ind->map(CLScheduler::get().queue());
127 for(unsigned int num_box = 0; num_box < _num_boxes; ++num_box)
128 {
129 auto crop_tensor = support::cpp14::make_unique<CLTensor>();
130 TensorInfo crop_result_info(1, DataType::F32);
131 crop_result_info.set_data_layout(DataLayout::NHWC);
132 crop_tensor->allocator()->init(crop_result_info);
133 _crop_results.emplace_back(std::move(crop_tensor));
134
135 auto scale_tensor = support::cpp14::make_unique<CLTensor>();
136 TensorInfo scaled_result_info(out_shape, 1, DataType::F32);
137 scaled_result_info.set_data_layout(DataLayout::NHWC);
138 scale_tensor->allocator()->init(scaled_result_info);
139 _scaled_results.emplace_back(std::move(scale_tensor));
140
141 // Size of the crop box in _boxes has to be given before the configure
142 uint32_t batch_index;
143 Coordinates start{};
144 Coordinates end{};
145 configure_crop(_input, _boxes, _box_ind, _crop_results[num_box].get(), num_box, start, end, batch_index);
146
147 auto scale_kernel = support::cpp14::make_unique<CLScale>();
148 scale_kernel->configure(compile_context, _crop_results[num_box].get(), _scaled_results[num_box].get(), ScaleKernelInfo{ _method, BorderMode::CONSTANT, PixelValue(_extrapolation_value), SamplingPolicy::TOP_LEFT });
149 _scale.emplace_back(std::move(scale_kernel));
150
151 Window win = calculate_max_window(*_output->info());
152 win.set(3, Window::Dimension(num_box, num_box + 1, 1));
153
154 auto copy_kernel = support::cpp14::make_unique<CLCopyKernel>();
155 copy_kernel->configure(compile_context, _scaled_results[num_box].get(), _output, &win);
156 _copy.emplace_back(std::move(copy_kernel));
157
158 _crop_results[num_box]->allocator()->allocate();
159 _scaled_results[num_box]->allocator()->allocate();
160
161 bool is_width_flipped = end[0] < start[0];
162 bool is_height_flipped = end[1] < start[1];
163 /** The number of rows out of bounds at the start and end of _crop_results[num_box].get(). */
164 std::array<int32_t, 2> rows_out_of_bounds{ 0 };
165 /** The number of columns out of bounds at the start and end of _crop_results[num_box].get(). */
166 std::array<int32_t, 2> cols_out_of_bounds{ 0 };
167 if(is_height_flipped)
168 {
169 rows_out_of_bounds[0] = start[1] >= static_cast<int32_t>(_input->info()->dimension(2)) ? std::min(start[1] - _input->info()->dimension(2) + 1, _crop_results[num_box].get()->info()->dimension(2)) : 0;
170 rows_out_of_bounds[1] = end[1] < 0 ? std::min(-end[1], static_cast<int32_t>(_crop_results[num_box].get()->info()->dimension(2))) : 0;
171 }
172 else
173 {
174 rows_out_of_bounds[0] = start[1] < 0 ? std::min(-start[1], static_cast<int32_t>(_crop_results[num_box].get()->info()->dimension(2))) : 0;
175 rows_out_of_bounds[1] = end[1] >= static_cast<int32_t>(_input->info()->dimension(2)) ? std::min(end[1] - _input->info()->dimension(2) + 1, _crop_results[num_box].get()->info()->dimension(2)) : 0;
176 }
177 if(is_width_flipped)
178 {
179 cols_out_of_bounds[0] = start[0] >= static_cast<int32_t>(_input->info()->dimension(1)) ? std::min(start[0] - _input->info()->dimension(1) + 1, _crop_results[num_box].get()->info()->dimension(1)) : 0;
180 cols_out_of_bounds[1] = end[0] < 0 ? std::min(-end[0], static_cast<int32_t>(_crop_results[num_box].get()->info()->dimension(1))) : 0;
181 }
182 else
183 {
184 cols_out_of_bounds[0] = start[0] < 0 ? std::min(-start[0], static_cast<int32_t>(_crop_results[num_box].get()->info()->dimension(1))) : 0;
185 cols_out_of_bounds[1] = end[0] >= static_cast<int32_t>(_input->info()->dimension(1)) ? std::min(end[0] - _input->info()->dimension(1) + 1, _crop_results[num_box].get()->info()->dimension(1)) : 0;
186 }
187
188 Window full_window = calculate_max_window(*_crop_results[num_box].get()->info());
189
190 // Full _crop_results[num_box].get() window:
191 // --------------------------------
192 // | Out of bounds |
193 // | rows before |
194 // |------------------------------|
195 // | Out of | In | Out of |
196 // | bounds | bounds | bounds |
197 // | cols | elements | cols |
198 // | before | copied | after |
199 // | | from input | |
200 // |------------------------------|
201 // | Out of bounds |
202 // | rows after |
203 // |------------------------------|
204 // Use a separate _crop_results[num_box].get() window for each section of the full _crop_results[num_box].get() window.
205 // Fill all _crop_results[num_box].get() rows that have no elements that are within the input bounds
206 // with the extrapolation value using memset.
207 // First for the rows before the in bounds rows.
208 if(rows_out_of_bounds[0] > 0)
209 {
210 Window slice_fill_rows_before(full_window);
211 slice_fill_rows_before.set(2, Window::Dimension(0, rows_out_of_bounds[0], 1));
212 auto kernel = arm_compute::support::cpp14::make_unique<CLMemsetKernel>();
213 kernel->configure(compile_context, _crop_results[num_box].get(), extrapolation_value, &slice_fill_rows_before);
214 _internal_kernels.push_back(std::move(kernel));
215 }
216
217 Window slice_in(full_window);
218 slice_in.set(2, Window::Dimension(rows_out_of_bounds[0], _crop_results[num_box].get()->info()->dimension(2) - rows_out_of_bounds[1], 1));
219 slice_in.set(1, Window::Dimension(cols_out_of_bounds[0], _crop_results[num_box].get()->info()->dimension(1) - cols_out_of_bounds[1], 1));
220
221 int rows_in_bounds = static_cast<int32_t>(_crop_results[num_box].get()->info()->dimension(2)) - rows_out_of_bounds[0] - rows_out_of_bounds[1];
222 if(rows_in_bounds > 0)
223 {
224 // Fill all elements that share a row with an in bounds element with the extrapolation value.
225 if(cols_out_of_bounds[0] > 0)
226 {
227 Window slice_fill_cols_before(slice_in);
228 slice_fill_cols_before.set(1, Window::Dimension(0, cols_out_of_bounds[0], 1));
229 auto kernel = arm_compute::support::cpp14::make_unique<CLMemsetKernel>();
230 kernel->configure(compile_context, _crop_results[num_box].get(), extrapolation_value, &slice_fill_cols_before);
231 _internal_kernels.push_back(std::move(kernel));
232 }
233
234 if(cols_out_of_bounds[1] > 0)
235 {
236 Window slice_fill_cols_after(slice_in);
237 slice_fill_cols_after.set(1, Window::Dimension(_crop_results[num_box].get()->info()->dimension(1) - cols_out_of_bounds[1], _crop_results[num_box].get()->info()->dimension(1), 1));
238 auto kernel = arm_compute::support::cpp14::make_unique<CLMemsetKernel>();
239 kernel->configure(compile_context, _crop_results[num_box].get(), extrapolation_value, &slice_fill_cols_after);
240 _internal_kernels.push_back(std::move(kernel));
241 }
242
243 // Copy all elements within the input bounds from the input tensor.
244 int cols_in_bounds = static_cast<int32_t>(_crop_results[num_box].get()->info()->dimension(1)) - cols_out_of_bounds[0] - cols_out_of_bounds[1];
245 if(cols_in_bounds > 0)
246 {
247 Coordinates2D start_in{ is_width_flipped ? start[0] - cols_out_of_bounds[0] : start[0] + cols_out_of_bounds[0],
248 is_height_flipped ? start[1] - rows_out_of_bounds[0] : start[1] + rows_out_of_bounds[0] };
249 Coordinates2D end_in{ is_width_flipped ? start_in.x - cols_in_bounds + 1 : start_in.x + cols_in_bounds - 1,
250 is_height_flipped ? start_in.y - rows_in_bounds + 1 : start_in.y + rows_in_bounds - 1 };
251 auto kernel = arm_compute::support::cpp14::make_unique<CLCropKernel>();
252
253 kernel->configure(compile_context, _input, _crop_results[num_box].get(), start_in, end_in, batch_index, extrapolation_value, &slice_in);
254 _internal_kernels.push_back(std::move(kernel));
255 }
256 }
257
258 // Fill all rows after the in bounds elements with the extrapolation value.
259 if(rows_out_of_bounds[1] > 0)
260 {
261 Window slice_fill_rows_after(full_window);
262 slice_fill_rows_after.set(2, Window::Dimension(_crop_results[num_box].get()->info()->dimension(2) - rows_out_of_bounds[1], _crop_results[num_box].get()->info()->dimension(2), 1));
263 auto kernel = arm_compute::support::cpp14::make_unique<CLMemsetKernel>();
264 kernel->configure(compile_context, _crop_results[num_box].get(), extrapolation_value, &slice_fill_rows_after);
265 _internal_kernels.push_back(std::move(kernel));
266 }
267 }
268 _boxes->unmap(CLScheduler::get().queue());
269 _box_ind->unmap(CLScheduler::get().queue());
270 CLScheduler::get().sync();
271 }
272
run()273 void CLCropResize::run()
274 {
275 ARM_COMPUTE_ERROR_ON_MSG(_output == nullptr, "Unconfigured function");
276
277 for(unsigned int i = 0; i < _internal_kernels.size(); ++i)
278 {
279 CLScheduler::get().enqueue(*(_internal_kernels[i]));
280 }
281
282 CLScheduler::get().sync();
283 for(auto &kernel : _scale)
284 {
285 kernel->run();
286 }
287 CLScheduler::get().sync();
288 for(auto &kernel : _copy)
289 {
290 CLScheduler::get().enqueue(*kernel, true);
291 }
292 CLScheduler::get().sync();
293 }
294 } // namespace arm_compute