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 */
24 #include "src/core/CL/kernels/CLPriorBoxLayerKernel.h"
25
26 #include "arm_compute/core/CL/CLHelpers.h"
27 #include "arm_compute/core/CL/CLKernelLibrary.h"
28 #include "arm_compute/core/CL/ICLTensor.h"
29 #include "arm_compute/core/Helpers.h"
30 #include "arm_compute/core/TensorInfo.h"
31 #include "arm_compute/core/Utils.h"
32 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
33 #include "src/core/CL/CLValidate.h"
34 #include "src/core/helpers/AutoConfiguration.h"
35 #include "src/core/helpers/WindowHelpers.h"
36
37 #include "support/StringSupport.h"
38
39 using namespace arm_compute::misc::shape_calculator;
40
41 namespace arm_compute
42 {
43 namespace
44 {
validate_arguments(const ITensorInfo * input1,const ITensorInfo * input2,const ITensorInfo * output,const PriorBoxLayerInfo & info)45 Status validate_arguments(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, const PriorBoxLayerInfo &info)
46 {
47 ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input1, input2, output);
48 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input1, 1, DataType::F32);
49 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(input1, input2);
50 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input1, input2);
51
52 // Check variances
53 const int var_size = info.variances().size();
54 if(var_size > 1)
55 {
56 ARM_COMPUTE_RETURN_ERROR_ON_MSG(var_size != 4, "Must provide 4 variance values");
57 for(int i = 0; i < var_size; ++i)
58 {
59 ARM_COMPUTE_RETURN_ERROR_ON_MSG(var_size <= 0, "Must be greater than 0");
60 }
61 }
62 ARM_COMPUTE_RETURN_ERROR_ON_MSG(info.steps()[0] < 0.f, "Step x should be greater or equal to 0");
63 ARM_COMPUTE_RETURN_ERROR_ON_MSG(info.steps()[1] < 0.f, "Step y should be greater or equal to 0");
64
65 if(!info.max_sizes().empty())
66 {
67 ARM_COMPUTE_RETURN_ERROR_ON_MSG(info.max_sizes().size() != info.min_sizes().size(), "Max and min sizes dimensions should match");
68 }
69
70 for(unsigned int i = 0; i < info.max_sizes().size(); ++i)
71 {
72 ARM_COMPUTE_RETURN_ERROR_ON_MSG(info.max_sizes()[i] < info.min_sizes()[i], "Max size should be greater than min size");
73 }
74
75 if(output != nullptr && output->total_size() != 0)
76 {
77 ARM_COMPUTE_RETURN_ERROR_ON(output->dimension(1) != 2);
78 }
79
80 return Status{};
81 }
82
validate_and_configure_window(const ITensorInfo * input1,const ITensorInfo * input2,ITensorInfo * output,const PriorBoxLayerInfo & info,int num_priors)83 std::pair<Status, Window> validate_and_configure_window(const ITensorInfo *input1, const ITensorInfo *input2, ITensorInfo *output, const PriorBoxLayerInfo &info, int num_priors)
84 {
85 ARM_COMPUTE_UNUSED(input2);
86 // Output tensor auto initialization if not yet initialized
87 TensorShape output_shape = compute_prior_box_shape(*input1, info);
88 auto_init_if_empty(*output, output_shape, 1, input1->data_type());
89
90 const unsigned int num_elems_processed_per_iteration = 4 * num_priors;
91 Window win = calculate_max_window(*output, Steps(num_elems_processed_per_iteration));
92 AccessWindowHorizontal output_access(output, 0, num_elems_processed_per_iteration);
93 bool window_changed = update_window_and_padding(win, output_access);
94 Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
95 return std::make_pair(err, win);
96 }
97 } // namespace
98
CLPriorBoxLayerKernel()99 CLPriorBoxLayerKernel::CLPriorBoxLayerKernel()
100 : _input1(nullptr), _input2(nullptr), _output(nullptr), _info(), _num_priors(), _min(), _max(), _aspect_ratios()
101 {
102 _type = CLKernelType::ELEMENTWISE;
103 }
104
configure(const ICLTensor * input1,const ICLTensor * input2,ICLTensor * output,const PriorBoxLayerInfo & info,cl::Buffer * min,cl::Buffer * max,cl::Buffer * aspect_ratios)105 void CLPriorBoxLayerKernel::configure(const ICLTensor *input1, const ICLTensor *input2, ICLTensor *output, const PriorBoxLayerInfo &info, cl::Buffer *min, cl::Buffer *max, cl::Buffer *aspect_ratios)
106 {
107 configure(CLKernelLibrary::get().get_compile_context(), input1, input2, output, info, min, max, aspect_ratios);
108 }
109
configure(const CLCompileContext & compile_context,const ICLTensor * input1,const ICLTensor * input2,ICLTensor * output,const PriorBoxLayerInfo & info,cl::Buffer * min,cl::Buffer * max,cl::Buffer * aspect_ratios)110 void CLPriorBoxLayerKernel::configure(const CLCompileContext &compile_context, const ICLTensor *input1, const ICLTensor *input2, ICLTensor *output, const PriorBoxLayerInfo &info, cl::Buffer *min,
111 cl::Buffer *max, cl::Buffer *aspect_ratios)
112 {
113 ARM_COMPUTE_ERROR_ON_NULLPTR(input1, input2, output);
114
115 _input1 = input1;
116 _input2 = input2;
117 _output = output;
118 _info = info;
119 _min = min;
120 _max = max;
121 _aspect_ratios = aspect_ratios;
122
123 ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input1->info(), input2->info(), output->info(), info));
124
125 // Calculate number of aspect ratios
126 _num_priors = info.aspect_ratios().size() * info.min_sizes().size() + info.max_sizes().size();
127
128 const DataLayout data_layout = input1->info()->data_layout();
129
130 const int width_idx = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH);
131 const int height_idx = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT);
132
133 const int layer_width = input1->info()->dimension(width_idx);
134 const int layer_height = input1->info()->dimension(height_idx);
135
136 int img_width = info.img_size().x;
137 int img_height = info.img_size().y;
138 if(img_width == 0 || img_height == 0)
139 {
140 img_width = input2->info()->dimension(width_idx);
141 img_height = input2->info()->dimension(height_idx);
142 }
143
144 float step_x = info.steps()[0];
145 float step_y = info.steps()[0];
146 if(step_x == 0.f || step_y == 0.f)
147 {
148 step_x = static_cast<float>(img_width) / layer_width;
149 step_y = static_cast<float>(img_height) / layer_height;
150 }
151
152 // Set build options
153 CLBuildOptions build_opts;
154 build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input1->info()->data_type()));
155 build_opts.add_option("-DWIDTH=" + support::cpp11::to_string(img_width));
156 build_opts.add_option("-DHEIGHT=" + support::cpp11::to_string(img_height));
157 build_opts.add_option("-DLAYER_WIDTH=" + support::cpp11::to_string(layer_width));
158 build_opts.add_option("-DLAYER_HEIGHT=" + support::cpp11::to_string(layer_height));
159 build_opts.add_option("-DSTEP_X=" + support::cpp11::to_string(step_x));
160 build_opts.add_option("-DSTEP_Y=" + support::cpp11::to_string(step_y));
161 build_opts.add_option("-DNUM_PRIORS=" + support::cpp11::to_string(_num_priors));
162 build_opts.add_option("-DOFFSET=" + support::cpp11::to_string(info.offset()));
163 build_opts.add_option_if(info.clip(), "-DIN_PLACE");
164
165 if(info.variances().size() > 1)
166 {
167 for(unsigned int i = 0; i < info.variances().size(); ++i)
168 {
169 build_opts.add_option("-DVARIANCE_" + support::cpp11::to_string(i) + "=" + support::cpp11::to_string(info.variances().at(i)));
170 }
171 }
172 else
173 {
174 for(unsigned int i = 0; i < 4; ++i)
175 {
176 build_opts.add_option("-DVARIANCE_" + support::cpp11::to_string(i) + "=" + support::cpp11::to_string(info.variances().at(0)));
177 }
178 }
179
180 unsigned int idx = num_arguments_per_2D_tensor();
181 _kernel = create_kernel(compile_context, "prior_box_layer_nchw", build_opts.options());
182
183 _kernel.setArg(idx++, *_min);
184 _kernel.setArg(idx++, *_max);
185 _kernel.setArg(idx++, *_aspect_ratios);
186 _kernel.setArg<unsigned int>(idx++, info.min_sizes().size());
187 _kernel.setArg<unsigned int>(idx++, info.max_sizes().size());
188 _kernel.setArg<unsigned int>(idx++, info.aspect_ratios().size());
189
190 // Configure kernel window
191 auto win_config = validate_and_configure_window(input1->info(), input2->info(), output->info(), info, _num_priors);
192
193 ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
194 ICLKernel::configure_internal(win_config.second);
195 }
196
validate(const ITensorInfo * input1,const ITensorInfo * input2,const ITensorInfo * output,const PriorBoxLayerInfo & info)197 Status CLPriorBoxLayerKernel::validate(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, const PriorBoxLayerInfo &info)
198 {
199 ARM_COMPUTE_ERROR_ON_NULLPTR(input1, input2, output);
200 ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input1, input2, output, info));
201 const int num_priors = info.aspect_ratios().size() * info.min_sizes().size() + info.max_sizes().size();
202 ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input1->clone().get(), input2->clone().get(), output->clone().get(), info, num_priors)
203 .first);
204
205 return Status{};
206 }
207
run(const Window & window,cl::CommandQueue & queue)208 void CLPriorBoxLayerKernel::run(const Window &window, cl::CommandQueue &queue)
209 {
210 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
211 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IKernel::window(), window);
212
213 queue.enqueueWriteBuffer(*_min, CL_TRUE, 0, _info.min_sizes().size() * sizeof(float), _info.min_sizes().data());
214 queue.enqueueWriteBuffer(*_aspect_ratios, CL_TRUE, 0, _info.aspect_ratios().size() * sizeof(float), _info.aspect_ratios().data());
215 if(!_info.max_sizes().empty())
216 {
217 queue.enqueueWriteBuffer(*_max, CL_TRUE, 0, _info.max_sizes().size() * sizeof(float), _info.max_sizes().data());
218 }
219
220 Window slice = window.first_slice_window_2D();
221 slice.set(Window::DimY, Window::Dimension(0, _output->info()->dimension(1), 2));
222
223 unsigned int idx = 0;
224 add_2D_tensor_argument(idx, _output, slice);
225 enqueue(queue, *this, slice, lws_hint());
226 }
227 } // namespace arm_compute
228