Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2017 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/core/CL/kernels/CLCannyEdgeKernel.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/CL/OpenCL.h" |
| 30 | #include "arm_compute/core/Helpers.h" |
| 31 | #include "arm_compute/core/Validate.h" |
| 32 | |
| 33 | using namespace arm_compute; |
| 34 | |
| 35 | CLGradientKernel::CLGradientKernel() |
| 36 | : _gx(nullptr), _gy(nullptr), _magnitude(nullptr), _phase(nullptr) |
| 37 | { |
| 38 | } |
| 39 | |
| 40 | void CLGradientKernel::configure(const ICLTensor *gx, const ICLTensor *gy, ICLTensor *magnitude, ICLTensor *phase, int32_t norm_type) |
| 41 | { |
| 42 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(gx, 1, DataType::S16, DataType::S32); |
| 43 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(gy, 1, DataType::S16, DataType::S32); |
| 44 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(magnitude, 1, DataType::U16, DataType::U32); |
| 45 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(phase, 1, DataType::U8); |
| 46 | ARM_COMPUTE_ERROR_ON_MSG(data_size_from_type(gx->info()->data_type()) != data_size_from_type(gy->info()->data_type()), |
| 47 | "Gx and Gy must have the same pixel size"); |
| 48 | ARM_COMPUTE_ERROR_ON_MSG(data_size_from_type(gx->info()->data_type()) != data_size_from_type(magnitude->info()->data_type()), |
| 49 | "Mag must have the same pixel size as Gx and Gy"); |
| 50 | |
| 51 | _gx = gx; |
| 52 | _gy = gy; |
| 53 | _magnitude = magnitude; |
| 54 | _phase = phase; |
| 55 | |
| 56 | // Create build opts |
| 57 | std::set<std::string> built_opts; |
| 58 | built_opts.emplace("-DDATA_TYPE_IN=" + get_cl_type_from_data_type(gx->info()->data_type())); |
| 59 | built_opts.emplace("-DDATA_TYPE_OUT=" + get_cl_type_from_data_type(gx->info()->data_type())); |
| 60 | |
| 61 | // Create kernel |
| 62 | const std::string kernel_name = (norm_type == 1) ? std::string("combine_gradients_L1") : std::string("combine_gradients_L2"); |
| 63 | _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name, built_opts)); |
| 64 | |
| 65 | // Configure kernel window |
| 66 | constexpr unsigned int num_elems_processed_per_iteration = 4; |
| 67 | |
| 68 | Window win = calculate_max_window(*_gx->info(), Steps(num_elems_processed_per_iteration)); |
| 69 | |
| 70 | AccessWindowHorizontal gx_access(_gx->info(), 0, num_elems_processed_per_iteration); |
| 71 | AccessWindowHorizontal gy_access(_gy->info(), 0, num_elems_processed_per_iteration); |
| 72 | AccessWindowHorizontal mag_access(_magnitude->info(), 0, num_elems_processed_per_iteration); |
| 73 | AccessWindowHorizontal phase_access(_phase->info(), 0, num_elems_processed_per_iteration); |
| 74 | |
| 75 | update_window_and_padding(win, gx_access, gy_access, mag_access, phase_access); |
| 76 | |
| 77 | mag_access.set_valid_region(win, _gx->info()->valid_region()); |
| 78 | phase_access.set_valid_region(win, _gx->info()->valid_region()); |
| 79 | |
| 80 | ICLKernel::configure(win); |
| 81 | } |
| 82 | |
| 83 | void CLGradientKernel::run(const Window &window, cl::CommandQueue &queue) |
| 84 | { |
| 85 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 86 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IKernel::window(), window); |
| 87 | |
| 88 | Window slice = window.first_slice_window_2D(); |
| 89 | do |
| 90 | { |
| 91 | unsigned int idx = 0; |
| 92 | add_2D_tensor_argument(idx, _gx, slice); |
| 93 | add_2D_tensor_argument(idx, _gy, slice); |
| 94 | add_2D_tensor_argument(idx, _magnitude, slice); |
| 95 | add_2D_tensor_argument(idx, _phase, slice); |
| 96 | enqueue(queue, *this, slice); |
| 97 | } |
| 98 | while(window.slide_window_slice_2D(slice)); |
| 99 | } |
| 100 | |
| 101 | CLEdgeNonMaxSuppressionKernel::CLEdgeNonMaxSuppressionKernel() |
| 102 | : _magnitude(nullptr), _phase(nullptr), _output(nullptr) |
| 103 | { |
| 104 | } |
| 105 | |
| 106 | BorderSize CLEdgeNonMaxSuppressionKernel::border_size() const |
| 107 | { |
| 108 | return BorderSize(1); |
| 109 | } |
| 110 | |
| 111 | void CLEdgeNonMaxSuppressionKernel::configure(const ICLTensor *magnitude, const ICLTensor *phase, ICLTensor *output, int32_t lower_thr, bool border_undefined) |
| 112 | { |
| 113 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(magnitude, 1, DataType::U16, DataType::U32); |
| 114 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(phase, 1, DataType::U8); |
| 115 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::U16, DataType::U32); |
| 116 | |
| 117 | _magnitude = magnitude; |
| 118 | _phase = phase; |
| 119 | _output = output; |
| 120 | |
| 121 | // Create build opts |
| 122 | std::set<std::string> built_opts; |
| 123 | built_opts.emplace("-DDATA_TYPE_IN=" + get_cl_type_from_data_type(magnitude->info()->data_type())); |
| 124 | built_opts.emplace("-DDATA_TYPE_OUT=" + get_cl_type_from_data_type(output->info()->data_type())); |
| 125 | |
| 126 | // Create kernel |
| 127 | _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("suppress_non_maximum", built_opts)); |
| 128 | |
| 129 | // Set minimum threshold argument |
| 130 | unsigned int idx = 3 * num_arguments_per_2D_tensor(); //Skip the input and output parameters |
| 131 | _kernel.setArg(idx++, lower_thr); |
| 132 | |
| 133 | // Configure kernel window |
| 134 | constexpr unsigned int num_elems_processed_per_iteration = 1; |
| 135 | constexpr unsigned int num_elems_read_written_per_iteration = 3; |
| 136 | |
| 137 | Window win = calculate_max_window(*_magnitude->info(), Steps(num_elems_processed_per_iteration), border_undefined, border_size()); |
| 138 | |
| 139 | AccessWindowRectangle mag_access(_magnitude->info(), -border_size().left, -border_size().top, |
| 140 | num_elems_read_written_per_iteration, num_elems_read_written_per_iteration); |
| 141 | AccessWindowHorizontal phase_access(_phase->info(), 0, num_elems_processed_per_iteration); |
| 142 | AccessWindowHorizontal output_access(_output->info(), 0, num_elems_processed_per_iteration); |
| 143 | |
| 144 | update_window_and_padding(win, mag_access, phase_access, output_access); |
| 145 | |
| 146 | output_access.set_valid_region(win, _magnitude->info()->valid_region(), border_undefined, border_size()); |
| 147 | |
| 148 | ICLKernel::configure(win); |
| 149 | } |
| 150 | |
| 151 | void CLEdgeNonMaxSuppressionKernel::run(const Window &window, cl::CommandQueue &queue) |
| 152 | { |
| 153 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 154 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IKernel::window(), window); |
| 155 | |
| 156 | Window slice = window.first_slice_window_2D(); |
| 157 | do |
| 158 | { |
| 159 | unsigned int idx = 0; |
| 160 | add_2D_tensor_argument(idx, _magnitude, slice); |
| 161 | add_2D_tensor_argument(idx, _phase, slice); |
| 162 | add_2D_tensor_argument(idx, _output, slice); |
| 163 | enqueue(queue, *this, slice); |
| 164 | } |
| 165 | while(window.slide_window_slice_2D(slice)); |
| 166 | } |
| 167 | |
| 168 | CLEdgeTraceKernel::CLEdgeTraceKernel() |
| 169 | : _input(nullptr), _output(nullptr), _lower_thr(0), _upper_thr(0), _visited(nullptr), _recorded(nullptr), _l1_stack(nullptr), _l1_stack_counter(nullptr) |
| 170 | { |
| 171 | } |
| 172 | |
| 173 | void CLEdgeTraceKernel::configure(const ICLTensor *input, ICLTensor *output, int32_t upper_thr, int32_t lower_thr, |
| 174 | ICLTensor *visited, ICLTensor *recorded, ICLTensor *l1_stack, ICLTensor *l1_stack_counter) |
| 175 | { |
| 176 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U16, DataType::U32); |
| 177 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::U8); |
| 178 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(visited, 1, DataType::U32); |
| 179 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(recorded, 1, DataType::U32); |
| 180 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(l1_stack, 1, DataType::S32); |
| 181 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(l1_stack_counter, 1, DataType::U8); |
| 182 | |
| 183 | _input = input; |
| 184 | _output = output; |
| 185 | _lower_thr = lower_thr; |
| 186 | _upper_thr = upper_thr; |
| 187 | _visited = visited; |
| 188 | _recorded = recorded; |
| 189 | _l1_stack = l1_stack; |
| 190 | _l1_stack_counter = l1_stack_counter; |
| 191 | |
| 192 | // Create build opts |
| 193 | std::set<std::string> built_opts; |
| 194 | built_opts.emplace("-DDATA_TYPE_IN=" + get_cl_type_from_data_type(input->info()->data_type())); |
| 195 | built_opts.emplace("-DDATA_TYPE_OUT=" + get_cl_type_from_data_type(output->info()->data_type())); |
| 196 | |
| 197 | // Create kernel |
| 198 | _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("hysteresis", built_opts)); |
| 199 | |
| 200 | // Set constant kernel args |
| 201 | unsigned int width = _input->info()->dimension(0); |
| 202 | unsigned int height = _input->info()->dimension(1); |
| 203 | unsigned int idx = 6 * num_arguments_per_2D_tensor(); //Skip the input and output parameters |
| 204 | _kernel.setArg(idx++, static_cast<cl_uint>(_lower_thr)); |
| 205 | _kernel.setArg(idx++, static_cast<cl_uint>(_upper_thr)); |
| 206 | _kernel.setArg(idx++, static_cast<cl_uint>(width)); |
| 207 | _kernel.setArg(idx++, static_cast<cl_uint>(height)); |
| 208 | |
| 209 | // Configure kernel window |
| 210 | constexpr unsigned int num_elems_processed_per_iteration = 1; |
| 211 | Window win = calculate_max_window(*_input->info(), Steps(num_elems_processed_per_iteration)); |
| 212 | |
| 213 | AccessWindowHorizontal output_access(_output->info(), 0, num_elems_processed_per_iteration); |
| 214 | AccessWindowHorizontal visited_access(_visited->info(), 0, num_elems_processed_per_iteration); |
| 215 | AccessWindowHorizontal recorded_access(_recorded->info(), 0, num_elems_processed_per_iteration); |
| 216 | AccessWindowHorizontal l1_stack_access(_l1_stack->info(), 0, num_elems_processed_per_iteration); |
| 217 | AccessWindowHorizontal l1_stack_counter_access(_l1_stack_counter->info(), 0, num_elems_processed_per_iteration); |
| 218 | |
| 219 | update_window_and_padding(win, |
| 220 | AccessWindowHorizontal(_input->info(), 0, num_elems_processed_per_iteration), |
| 221 | output_access, |
| 222 | visited_access, |
| 223 | recorded_access, |
| 224 | l1_stack_access, |
| 225 | l1_stack_counter_access); |
| 226 | |
| 227 | output_access.set_valid_region(win, _input->info()->valid_region()); |
| 228 | visited_access.set_valid_region(win, _input->info()->valid_region()); |
| 229 | recorded_access.set_valid_region(win, _input->info()->valid_region()); |
| 230 | l1_stack_access.set_valid_region(win, _input->info()->valid_region()); |
| 231 | l1_stack_counter_access.set_valid_region(win, _input->info()->valid_region()); |
| 232 | |
| 233 | ICLKernel::configure(win); |
| 234 | } |
| 235 | |
| 236 | void CLEdgeTraceKernel::run(const Window &window, cl::CommandQueue &queue) |
| 237 | { |
| 238 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 239 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IKernel::window(), window); |
| 240 | |
| 241 | Window slice = window.first_slice_window_2D(); |
| 242 | do |
| 243 | { |
| 244 | unsigned int idx = 0; |
| 245 | add_2D_tensor_argument(idx, _input, slice); |
| 246 | add_2D_tensor_argument(idx, _output, slice); |
| 247 | add_2D_tensor_argument(idx, _visited, slice); |
| 248 | add_2D_tensor_argument(idx, _recorded, slice); |
| 249 | add_2D_tensor_argument(idx, _l1_stack, slice); |
| 250 | add_2D_tensor_argument(idx, _l1_stack_counter, slice); |
| 251 | |
| 252 | enqueue(queue, *this, slice); |
| 253 | } |
| 254 | while(window.slide_window_slice_2D(slice)); |
| 255 | } |