giuros01 | 164a272 | 2018-11-20 18:34:46 +0000 | [diff] [blame] | 1 | /* |
Michalis Spyrou | bcfd09a | 2019-05-01 13:03:59 +0100 | [diff] [blame] | 2 | * Copyright (c) 2018-2019 ARM Limited. |
giuros01 | 164a272 | 2018-11-20 18:34:46 +0000 | [diff] [blame] | 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/CLElementwiseOperationKernel.h" |
| 25 | |
| 26 | #include "arm_compute/core/CL/CLHelpers.h" |
| 27 | #include "arm_compute/core/CL/CLValidate.h" |
| 28 | #include "arm_compute/core/CL/ICLTensor.h" |
| 29 | #include <map> |
| 30 | |
| 31 | namespace arm_compute |
| 32 | { |
| 33 | namespace |
| 34 | { |
| 35 | constexpr unsigned int num_elems_processed_per_iteration = 16; |
| 36 | |
| 37 | std::map<ArithmeticOperation, std::string> supported_arithmetic_ops = |
| 38 | { |
| 39 | { ArithmeticOperation::ADD, "ADD" }, |
| 40 | { ArithmeticOperation::SUB, "SUB" }, |
| 41 | { ArithmeticOperation::DIV, "DIV" }, |
| 42 | { ArithmeticOperation::SQUARED_DIFF, "SQUARED_DIFF" }, |
| 43 | { ArithmeticOperation::MIN, "MIN" }, |
| 44 | { ArithmeticOperation::MAX, "MAX" }, |
Usama Arif | 52c54f6 | 2019-05-14 10:22:36 +0100 | [diff] [blame] | 45 | { ArithmeticOperation::POWER, "POWER" }, |
giuros01 | 1e6e1b8 | 2019-05-14 16:12:53 +0100 | [diff] [blame] | 46 | { ArithmeticOperation::PRELU, "PRELU" }, |
giuros01 | 164a272 | 2018-11-20 18:34:46 +0000 | [diff] [blame] | 47 | }; |
| 48 | |
| 49 | std::map<ArithmeticOperation, std::string> supported_sat_arithmetic_ops = |
| 50 | { |
| 51 | { ArithmeticOperation::ADD, "ADD" }, |
| 52 | { ArithmeticOperation::SUB, "SUB" }, |
| 53 | }; |
| 54 | |
| 55 | std::string generate_id_for_tuning_common(const std::string &kernel_name, const ITensorInfo &input1, const ITensorInfo &output) |
| 56 | { |
| 57 | std::string config_id; |
| 58 | // Set config_id for enabling LWS tuning |
| 59 | config_id = kernel_name; |
| 60 | config_id += "_"; |
| 61 | config_id += lower_string(string_from_data_type(input1.data_type())); |
| 62 | config_id += "_"; |
| 63 | config_id += support::cpp11::to_string(output.dimension(0)); |
| 64 | config_id += "_"; |
| 65 | config_id += support::cpp11::to_string(output.dimension(1)); |
| 66 | return config_id; |
| 67 | } |
| 68 | |
Usama Arif | 52c54f6 | 2019-05-14 10:22:36 +0100 | [diff] [blame] | 69 | Status validate_arguments_with_float_only_supported_rules(const ITensorInfo &input1, const ITensorInfo &input2, const ITensorInfo &output) |
giuros01 | 49f7c02 | 2018-12-03 19:25:22 +0000 | [diff] [blame] | 70 | { |
| 71 | ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(&input1, &input2, &output); |
| 72 | ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(&input1); |
| 73 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&input1, 1, DataType::F16, DataType::F32); |
| 74 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(&input1, &input2); |
| 75 | |
| 76 | const TensorShape out_shape = TensorShape::broadcast_shape(input1.tensor_shape(), input2.tensor_shape()); |
| 77 | |
| 78 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(out_shape.total_size() == 0, "Inputs are not broadcast compatible"); |
| 79 | |
| 80 | // Validate in case of configured output |
| 81 | if(output.total_size() > 0) |
| 82 | { |
| 83 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&output, 1, DataType::F16, DataType::F32); |
| 84 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(&input1, &output); |
| 85 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(detail::have_different_dimensions(out_shape, output.tensor_shape(), 0), |
| 86 | "Wrong shape for output"); |
| 87 | } |
| 88 | |
| 89 | return Status{}; |
| 90 | } |
| 91 | |
giuros01 | 164a272 | 2018-11-20 18:34:46 +0000 | [diff] [blame] | 92 | Status validate_arguments_with_arithmetic_rules(const ITensorInfo &input1, const ITensorInfo &input2, const ITensorInfo &output) |
| 93 | { |
| 94 | ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(&input1); |
| 95 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&input1, 1, DataType::U8, DataType::QASYMM8, DataType::S16, DataType::F16, DataType::F32); |
| 96 | ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(&input2); |
| 97 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&input2, 1, DataType::U8, DataType::QASYMM8, DataType::S16, DataType::F16, DataType::F32); |
| 98 | |
| 99 | const bool is_qasymm = is_data_type_quantized_asymmetric(input1.data_type()) || is_data_type_quantized_asymmetric(input2.data_type()); |
| 100 | if(is_qasymm) |
| 101 | { |
| 102 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(&input1, &input2); |
| 103 | } |
| 104 | |
| 105 | const TensorShape out_shape = TensorShape::broadcast_shape(input1.tensor_shape(), input2.tensor_shape()); |
| 106 | |
| 107 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(out_shape.total_size() == 0, "Inputs are not broadcast compatible"); |
| 108 | |
| 109 | // Validate in case of configured output |
| 110 | if(output.total_size() > 0) |
| 111 | { |
| 112 | ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(&output); |
| 113 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&output, 1, DataType::U8, DataType::QASYMM8, DataType::S16, DataType::F16, DataType::F32); |
| 114 | ARM_COMPUTE_RETURN_ERROR_ON_MSG((output.data_type() == DataType::U8) && ((input1.data_type() != DataType::U8) || (input2.data_type() != DataType::U8)), |
| 115 | "Output can only be U8 if both inputs are U8"); |
| 116 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(detail::have_different_dimensions(out_shape, output.tensor_shape(), 0), |
| 117 | "Wrong shape for output"); |
| 118 | if(is_qasymm) |
| 119 | { |
| 120 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(&input1, &output); |
| 121 | } |
| 122 | } |
| 123 | return Status{}; |
| 124 | } |
| 125 | |
| 126 | CLBuildOptions generate_build_options_with_arithmetic_rules(const ITensorInfo &input1, const ITensorInfo &input2, const ITensorInfo &output, const std::string &operation_string) |
| 127 | { |
| 128 | CLBuildOptions build_opts; |
| 129 | |
| 130 | build_opts.add_option("-DDATA_TYPE_IN1=" + get_cl_type_from_data_type(input1.data_type())); |
| 131 | build_opts.add_option("-DDATA_TYPE_IN2=" + get_cl_type_from_data_type(input2.data_type())); |
| 132 | build_opts.add_option("-DDATA_TYPE_OUT=" + get_cl_type_from_data_type(output.data_type())); |
| 133 | build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration)); |
| 134 | build_opts.add_option("-DOP=" + operation_string); |
| 135 | if(is_data_type_quantized_asymmetric(input1.data_type())) |
| 136 | { |
Georgios Pinitas | 4c5469b | 2019-05-21 13:32:43 +0100 | [diff] [blame] | 137 | const UniformQuantizationInfo iq1info = input1.quantization_info().uniform(); |
| 138 | const UniformQuantizationInfo iq2info = input2.quantization_info().uniform(); |
| 139 | const UniformQuantizationInfo oqinfo = output.quantization_info().uniform(); |
| 140 | |
| 141 | build_opts.add_option("-DOFFSET_IN1=" + support::cpp11::to_string(iq1info.offset)); |
| 142 | build_opts.add_option("-DOFFSET_IN2=" + support::cpp11::to_string(iq2info.offset)); |
| 143 | build_opts.add_option("-DOFFSET_OUT=" + support::cpp11::to_string(oqinfo.offset)); |
| 144 | build_opts.add_option("-DSCALE_IN1=" + float_to_string_with_full_precision(iq1info.scale)); |
| 145 | build_opts.add_option("-DSCALE_IN2=" + float_to_string_with_full_precision(iq2info.scale)); |
| 146 | build_opts.add_option("-DSCALE_OUT=" + float_to_string_with_full_precision(oqinfo.scale)); |
giuros01 | 164a272 | 2018-11-20 18:34:46 +0000 | [diff] [blame] | 147 | } |
| 148 | return build_opts; |
| 149 | } |
| 150 | |
giuros01 | 49f7c02 | 2018-12-03 19:25:22 +0000 | [diff] [blame] | 151 | std::pair<Status, Window> configure_window_arithmetic_common(const ValidRegion &valid_region, ITensorInfo &input1, ITensorInfo &input2, ITensorInfo &output) |
| 152 | { |
| 153 | Window win = calculate_max_window(valid_region, Steps(num_elems_processed_per_iteration)); |
| 154 | Window win_input1 = win.broadcast_if_dimension_le_one(input1); |
| 155 | Window win_input2 = win.broadcast_if_dimension_le_one(input2); |
| 156 | |
| 157 | AccessWindowHorizontal input1_access(&input1, 0, num_elems_processed_per_iteration); |
| 158 | AccessWindowHorizontal input2_access(&input2, 0, num_elems_processed_per_iteration); |
| 159 | AccessWindowHorizontal output_access(&output, 0, num_elems_processed_per_iteration); |
| 160 | |
| 161 | bool window_changed = update_window_and_padding(win_input1, input1_access) |
| 162 | || update_window_and_padding(win_input2, input2_access) |
| 163 | || update_window_and_padding(win, output_access); |
| 164 | |
| 165 | output_access.set_valid_region(win, valid_region); |
| 166 | |
| 167 | Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{}; |
| 168 | return std::make_pair(err, win); |
| 169 | } |
| 170 | |
giuros01 | 164a272 | 2018-11-20 18:34:46 +0000 | [diff] [blame] | 171 | std::pair<Status, Window> validate_and_configure_window_for_arithmetic_operators(ITensorInfo &input1, ITensorInfo &input2, ITensorInfo &output) |
| 172 | { |
| 173 | const std::pair<TensorShape, ValidRegion> broadcast_pair = ITensorInfo::broadcast_shape_and_valid_region(input1, input2); |
| 174 | const TensorShape &out_shape = broadcast_pair.first; |
| 175 | const ValidRegion &valid_region = broadcast_pair.second; |
| 176 | |
| 177 | set_shape_if_empty(output, out_shape); |
| 178 | |
| 179 | if(input1.data_type() == DataType::S16 || input2.data_type() == DataType::S16) |
| 180 | { |
| 181 | set_format_if_unknown(output, Format::S16); |
| 182 | } |
| 183 | else if(input1.data_type() == DataType::F16 && input2.data_type() == DataType::F16) |
| 184 | { |
| 185 | set_format_if_unknown(output, Format::F16); |
| 186 | } |
| 187 | else if(input1.data_type() == DataType::F32 || input2.data_type() == DataType::F32) |
| 188 | { |
| 189 | set_format_if_unknown(output, Format::F32); |
| 190 | } |
| 191 | |
giuros01 | 49f7c02 | 2018-12-03 19:25:22 +0000 | [diff] [blame] | 192 | return configure_window_arithmetic_common(valid_region, input1, input2, output); |
| 193 | } |
giuros01 | 164a272 | 2018-11-20 18:34:46 +0000 | [diff] [blame] | 194 | |
giuros01 | 49f7c02 | 2018-12-03 19:25:22 +0000 | [diff] [blame] | 195 | std::pair<Status, Window> validate_and_configure_window_for_division(ITensorInfo &input1, ITensorInfo &input2, ITensorInfo &output) |
| 196 | { |
| 197 | const std::pair<TensorShape, ValidRegion> broadcast_pair = ITensorInfo::broadcast_shape_and_valid_region(input1, input2); |
| 198 | const TensorShape &out_shape = broadcast_pair.first; |
| 199 | const ValidRegion &valid_region = broadcast_pair.second; |
| 200 | auto_init_if_empty(output, out_shape, 1, input1.data_type()); |
| 201 | return configure_window_arithmetic_common(valid_region, input1, input2, output); |
giuros01 | 164a272 | 2018-11-20 18:34:46 +0000 | [diff] [blame] | 202 | } |
| 203 | } // namespace |
| 204 | |
| 205 | CLElementwiseOperationKernel::CLElementwiseOperationKernel() |
| 206 | : _input1(nullptr), _input2(nullptr), _output(nullptr) |
| 207 | { |
| 208 | } |
| 209 | |
| 210 | void CLElementwiseOperationKernel::configure_common(const ICLTensor *input1, const ICLTensor *input2, ICLTensor *output) |
| 211 | { |
| 212 | ARM_COMPUTE_ERROR_ON_NULLPTR(input1, input2, output); |
| 213 | ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(*input1->info(), *input2->info(), *output->info())); |
| 214 | |
| 215 | // Configure kernel window |
| 216 | auto win_config = validate_and_configure_window(*input1->info(), *input2->info(), *output->info()); |
| 217 | ARM_COMPUTE_ERROR_THROW_ON(win_config.first); |
| 218 | |
| 219 | _input1 = input1; |
| 220 | _input2 = input2; |
| 221 | _output = output; |
| 222 | |
| 223 | std::string kernel_name = "elementwise_operation_" + name(); |
| 224 | if(is_data_type_quantized_asymmetric(input1->info()->data_type())) |
| 225 | { |
| 226 | kernel_name += "_quantized"; |
| 227 | } |
| 228 | |
| 229 | // Set kernel build options |
| 230 | CLBuildOptions build_opts = generate_build_options(*input1->info(), *input2->info(), *output->info()); |
| 231 | |
| 232 | // Create kernel |
| 233 | _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name, build_opts.options())); |
| 234 | |
| 235 | ICLKernel::configure_internal(win_config.second); |
| 236 | |
| 237 | _config_id = generate_id_for_tuning(kernel_name, *input1->info(), *output->info()); |
| 238 | } |
| 239 | |
| 240 | void CLElementwiseOperationKernel::run(const Window &window, cl::CommandQueue &queue) |
| 241 | { |
| 242 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 243 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window); |
| 244 | |
| 245 | const TensorShape &in_shape1 = _input1->info()->tensor_shape(); |
| 246 | const TensorShape &in_shape2 = _input2->info()->tensor_shape(); |
| 247 | const TensorShape &out_shape = _output->info()->tensor_shape(); |
| 248 | |
| 249 | bool can_collapse = true; |
| 250 | const bool is_vector = in_shape1.num_dimensions() == 1 || in_shape2.num_dimensions() == 1; |
| 251 | if(std::min(in_shape1.total_size(), in_shape2.total_size()) > 1 && !is_vector) |
| 252 | { |
| 253 | can_collapse = (std::min(in_shape1.num_dimensions(), in_shape2.num_dimensions()) > Window::DimZ); |
| 254 | for(size_t d = Window::DimZ; can_collapse && (d < out_shape.num_dimensions()); d++) |
| 255 | { |
| 256 | can_collapse = (in_shape1[d] == in_shape2[d]); |
| 257 | } |
| 258 | } |
| 259 | |
| 260 | bool has_collapsed = false; |
| 261 | Window collapsed = can_collapse ? window.collapse_if_possible(ICLKernel::window(), Window::DimZ, &has_collapsed) : window; |
| 262 | |
| 263 | const TensorShape &in_shape1_collapsed = has_collapsed ? in_shape1.collapsed_from(Window::DimZ) : in_shape1; |
| 264 | const TensorShape &in_shape2_collapsed = has_collapsed ? in_shape2.collapsed_from(Window::DimZ) : in_shape2; |
| 265 | |
| 266 | Window slice = collapsed.first_slice_window_3D(); |
| 267 | Window slice_input1 = slice.broadcast_if_dimension_le_one(in_shape1_collapsed); |
| 268 | Window slice_input2 = slice.broadcast_if_dimension_le_one(in_shape2_collapsed); |
| 269 | |
| 270 | do |
| 271 | { |
| 272 | unsigned int idx = 0; |
| 273 | |
| 274 | add_3D_tensor_argument(idx, _input1, slice_input1); |
| 275 | add_3D_tensor_argument(idx, _input2, slice_input2); |
| 276 | add_3D_tensor_argument(idx, _output, slice); |
| 277 | |
| 278 | enqueue(queue, *this, slice, lws_hint()); |
| 279 | |
| 280 | collapsed.slide_window_slice_3D(slice_input1); |
| 281 | collapsed.slide_window_slice_3D(slice_input2); |
| 282 | } |
| 283 | while(collapsed.slide_window_slice_3D(slice)); |
| 284 | } |
| 285 | |
| 286 | BorderSize CLElementwiseOperationKernel::border_size() const |
| 287 | { |
| 288 | const unsigned int replicateSize = _output->info()->dimension(0) - std::min(_input1->info()->dimension(0), _input2->info()->dimension(0)); |
| 289 | const unsigned int border = std::min<unsigned int>(num_elems_processed_per_iteration - 1U, replicateSize); |
Michalis Spyrou | bcfd09a | 2019-05-01 13:03:59 +0100 | [diff] [blame] | 290 | return BorderSize{ 0, border, 0, 0 }; |
giuros01 | 164a272 | 2018-11-20 18:34:46 +0000 | [diff] [blame] | 291 | } |
| 292 | |
| 293 | /** Arithmetic operations with saturation*/ |
| 294 | |
| 295 | void CLSaturatedArithmeticOperationKernel::configure(ArithmeticOperation op, const ICLTensor *input1, const ICLTensor *input2, ICLTensor *output, const ConvertPolicy &policy) |
| 296 | { |
| 297 | _policy = policy; |
| 298 | _op = op; |
| 299 | configure_common(input1, input2, output); |
| 300 | } |
| 301 | |
| 302 | Status CLSaturatedArithmeticOperationKernel::validate(ArithmeticOperation op, const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, const ConvertPolicy &policy) |
| 303 | { |
| 304 | ARM_COMPUTE_UNUSED(op, policy); |
| 305 | ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input1, input2, output); |
| 306 | ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments_with_arithmetic_rules(*input1, *input2, *output)); |
| 307 | ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window_for_arithmetic_operators(*input1->clone(), *input2->clone(), *output->clone()).first); |
| 308 | |
| 309 | return Status{}; |
| 310 | } |
| 311 | |
| 312 | std::pair<Status, Window> CLSaturatedArithmeticOperationKernel::validate_and_configure_window(ITensorInfo &input1, ITensorInfo &input2, ITensorInfo &output) |
| 313 | { |
| 314 | return validate_and_configure_window_for_arithmetic_operators(input1, input2, output); |
| 315 | } |
| 316 | |
| 317 | Status CLSaturatedArithmeticOperationKernel::validate_arguments(const ITensorInfo &input1, const ITensorInfo &input2, const ITensorInfo &output) |
| 318 | { |
| 319 | return validate_arguments_with_arithmetic_rules(input1, input2, output); |
| 320 | } |
| 321 | |
| 322 | CLBuildOptions CLSaturatedArithmeticOperationKernel::generate_build_options(const ITensorInfo &input1, const ITensorInfo &input2, const ITensorInfo &output) |
| 323 | { |
| 324 | const bool has_float_out = is_data_type_float(output.data_type()); |
| 325 | auto build_options = generate_build_options_with_arithmetic_rules(input1, input2, output, name()); |
| 326 | build_options.add_option((_policy == ConvertPolicy::WRAP || has_float_out) ? "-DWRAP" : "-DSATURATE"); |
| 327 | return build_options; |
| 328 | } |
| 329 | std::string CLSaturatedArithmeticOperationKernel::generate_id_for_tuning(const std::string &kernel_name, const ITensorInfo &input1, const ITensorInfo &output) |
| 330 | { |
| 331 | auto config_id = generate_id_for_tuning_common(kernel_name, input1, output); |
| 332 | config_id += (_policy == ConvertPolicy::WRAP) ? "_wrap_" : "_saturate_"; |
| 333 | config_id += lower_string(string_from_data_layout(input1.data_layout())); |
| 334 | return config_id; |
| 335 | } |
| 336 | |
| 337 | std::string CLSaturatedArithmeticOperationKernel::name() |
| 338 | { |
| 339 | return supported_sat_arithmetic_ops[_op]; |
| 340 | } |
| 341 | |
| 342 | /** Arithmetic operations*/ |
| 343 | |
| 344 | void CLArithmeticOperationKernel::configure(ArithmeticOperation op, const ICLTensor *input1, const ICLTensor *input2, ICLTensor *output) |
| 345 | { |
| 346 | _op = op; |
| 347 | configure_common(input1, input2, output); |
| 348 | } |
| 349 | |
| 350 | Status CLArithmeticOperationKernel::validate(ArithmeticOperation op, const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output) |
| 351 | { |
giuros01 | 164a272 | 2018-11-20 18:34:46 +0000 | [diff] [blame] | 352 | ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input1, input2, output); |
Usama Arif | 52c54f6 | 2019-05-14 10:22:36 +0100 | [diff] [blame] | 353 | if(op == ArithmeticOperation::DIV || op == ArithmeticOperation::POWER) |
giuros01 | 49f7c02 | 2018-12-03 19:25:22 +0000 | [diff] [blame] | 354 | { |
Usama Arif | 52c54f6 | 2019-05-14 10:22:36 +0100 | [diff] [blame] | 355 | // Division and Power operators don't support integer arithmetic |
| 356 | ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments_with_float_only_supported_rules(*input1, *input2, *output)); |
giuros01 | 49f7c02 | 2018-12-03 19:25:22 +0000 | [diff] [blame] | 357 | ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window_for_division(*input1->clone(), *input2->clone(), *output->clone()).first); |
| 358 | } |
| 359 | else |
| 360 | { |
| 361 | ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments_with_arithmetic_rules(*input1, *input2, *output)); |
| 362 | ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window_for_arithmetic_operators(*input1->clone(), *input2->clone(), *output->clone()).first); |
| 363 | } |
| 364 | |
giuros01 | 164a272 | 2018-11-20 18:34:46 +0000 | [diff] [blame] | 365 | return Status{}; |
| 366 | } |
| 367 | std::pair<Status, Window> CLArithmeticOperationKernel::validate_and_configure_window(ITensorInfo &input1, ITensorInfo &input2, ITensorInfo &output) |
| 368 | { |
Usama Arif | 52c54f6 | 2019-05-14 10:22:36 +0100 | [diff] [blame] | 369 | if(_op == ArithmeticOperation::DIV || _op == ArithmeticOperation::POWER) |
giuros01 | 49f7c02 | 2018-12-03 19:25:22 +0000 | [diff] [blame] | 370 | { |
Usama Arif | 52c54f6 | 2019-05-14 10:22:36 +0100 | [diff] [blame] | 371 | // Division and Power operators don't support integer arithmetic |
giuros01 | 49f7c02 | 2018-12-03 19:25:22 +0000 | [diff] [blame] | 372 | return validate_and_configure_window_for_division(input1, input2, output); |
| 373 | } |
| 374 | else |
| 375 | { |
| 376 | return validate_and_configure_window_for_arithmetic_operators(input1, input2, output); |
| 377 | } |
giuros01 | 164a272 | 2018-11-20 18:34:46 +0000 | [diff] [blame] | 378 | } |
| 379 | Status CLArithmeticOperationKernel::validate_arguments(const ITensorInfo &input1, const ITensorInfo &input2, const ITensorInfo &output) |
| 380 | { |
Usama Arif | 52c54f6 | 2019-05-14 10:22:36 +0100 | [diff] [blame] | 381 | if(_op == ArithmeticOperation::DIV || _op == ArithmeticOperation::POWER) |
giuros01 | 49f7c02 | 2018-12-03 19:25:22 +0000 | [diff] [blame] | 382 | { |
Usama Arif | 52c54f6 | 2019-05-14 10:22:36 +0100 | [diff] [blame] | 383 | // Division and Power operators don't support integer arithmetic |
| 384 | return validate_arguments_with_float_only_supported_rules(input1, input2, output); |
giuros01 | 49f7c02 | 2018-12-03 19:25:22 +0000 | [diff] [blame] | 385 | } |
| 386 | else |
| 387 | { |
| 388 | return validate_arguments_with_arithmetic_rules(input1, input2, output); |
| 389 | } |
giuros01 | 164a272 | 2018-11-20 18:34:46 +0000 | [diff] [blame] | 390 | } |
| 391 | |
| 392 | CLBuildOptions CLArithmeticOperationKernel::generate_build_options(const ITensorInfo &input1, const ITensorInfo &input2, const ITensorInfo &output) |
| 393 | { |
| 394 | return generate_build_options_with_arithmetic_rules(input1, input2, output, name()); |
| 395 | } |
| 396 | std::string CLArithmeticOperationKernel::generate_id_for_tuning(const std::string &kernel_name, const ITensorInfo &input1, const ITensorInfo &output) |
| 397 | { |
| 398 | return generate_id_for_tuning_common(kernel_name, input1, output); |
| 399 | } |
| 400 | |
| 401 | std::string CLArithmeticOperationKernel::name() |
| 402 | { |
| 403 | return supported_arithmetic_ops[_op]; |
| 404 | } |
| 405 | } // namespace arm_compute |