Michele Di Giorgio | 1e0208a | 2021-01-22 15:42:59 +0000 | [diff] [blame] | 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/gpu/cl/kernels/ClElementwiseKernel.h" |
| 25 | |
| 26 | #include "arm_compute/core/CL/CLHelpers.h" |
| 27 | #include "arm_compute/core/CL/ICLTensor.h" |
Georgios Pinitas | 8a5146f | 2021-01-12 15:51:07 +0000 | [diff] [blame] | 28 | #include "src/common/utils/Validate.h" |
Michele Di Giorgio | 1e0208a | 2021-01-22 15:42:59 +0000 | [diff] [blame] | 29 | #include "src/core/CL/CLValidate.h" |
Michele Di Giorgio | 1e0208a | 2021-01-22 15:42:59 +0000 | [diff] [blame] | 30 | #include "src/core/helpers/AutoConfiguration.h" |
| 31 | #include "src/core/helpers/WindowHelpers.h" |
| 32 | #include "support/Cast.h" |
| 33 | #include "support/StringSupport.h" |
| 34 | #include <map> |
| 35 | |
| 36 | namespace arm_compute |
| 37 | { |
| 38 | namespace opencl |
| 39 | { |
| 40 | namespace kernels |
| 41 | { |
| 42 | namespace |
| 43 | { |
| 44 | constexpr unsigned int vector_size_byte_opencl = 16; |
| 45 | |
| 46 | std::map<ArithmeticOperation, std::string> supported_arithmetic_ops = |
| 47 | { |
| 48 | { ArithmeticOperation::ADD, "ADD" }, |
| 49 | { ArithmeticOperation::SUB, "SUB" }, |
| 50 | { ArithmeticOperation::DIV, "DIV" }, |
| 51 | { ArithmeticOperation::SQUARED_DIFF, "SQUARED_DIFF" }, |
| 52 | { ArithmeticOperation::MIN, "MIN" }, |
| 53 | { ArithmeticOperation::MAX, "MAX" }, |
| 54 | { ArithmeticOperation::POWER, "POWER" }, |
| 55 | { ArithmeticOperation::PRELU, "PRELU" }, |
| 56 | }; |
| 57 | |
| 58 | std::map<ArithmeticOperation, std::string> supported_sat_arithmetic_ops = |
| 59 | { |
| 60 | { ArithmeticOperation::ADD, "ADD" }, |
| 61 | { ArithmeticOperation::SUB, "SUB" }, |
| 62 | }; |
| 63 | |
| 64 | std::string generate_id_for_tuning_common(const std::string &kernel_name, const ITensorInfo &src1, const ITensorInfo &dst) |
| 65 | { |
| 66 | std::string config_id; |
| 67 | // Set config_id for enabling LWS tuning |
| 68 | config_id = kernel_name; |
| 69 | config_id += "_"; |
| 70 | config_id += lower_string(string_from_data_type(src1.data_type())); |
| 71 | config_id += "_"; |
| 72 | config_id += support::cpp11::to_string(dst.dimension(0)); |
| 73 | config_id += "_"; |
| 74 | config_id += support::cpp11::to_string(dst.dimension(1)); |
| 75 | return config_id; |
| 76 | } |
| 77 | |
| 78 | Status validate_arguments_with_float_only_supported_rules(const ITensorInfo &src1, const ITensorInfo &src2, const ITensorInfo &dst) |
| 79 | { |
| 80 | ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(&src1, &src2, &dst); |
| 81 | ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(&src1); |
| 82 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&src1, 1, DataType::F16, DataType::F32); |
| 83 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(&src1, &src2); |
| 84 | |
| 85 | const TensorShape out_shape = TensorShape::broadcast_shape(src1.tensor_shape(), src2.tensor_shape()); |
| 86 | |
| 87 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(out_shape.total_size() == 0, "Inputs are not broadcast compatible"); |
| 88 | |
| 89 | // Validate in case of configured dst |
| 90 | if(dst.total_size() > 0) |
| 91 | { |
| 92 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&dst, 1, DataType::F16, DataType::F32); |
| 93 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(&src1, &dst); |
| 94 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(detail::have_different_dimensions(out_shape, dst.tensor_shape(), 0), |
| 95 | "Wrong shape for dst"); |
| 96 | } |
| 97 | |
| 98 | return Status{}; |
| 99 | } |
| 100 | |
Giorgio Arena | 4a95bba | 2021-06-28 11:00:27 +0100 | [diff] [blame] | 101 | Status validate_arguments_divide_operation(const ITensorInfo *src1, const ITensorInfo *src2, const ITensorInfo *dst) |
Suhail Munshi | ae1c9fe | 2021-04-14 12:16:49 +0100 | [diff] [blame] | 102 | { |
| 103 | ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src1, src2, dst); |
| 104 | ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(src1); |
| 105 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src1, 1, DataType::F16, DataType::F32, DataType::S32); |
| 106 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src1, src2); |
| 107 | |
| 108 | const TensorShape out_shape = TensorShape::broadcast_shape(src1->tensor_shape(), src2->tensor_shape()); |
| 109 | |
| 110 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(out_shape.total_size() == 0, "Inputs are not broadcast compatible"); |
| 111 | |
| 112 | // Validate in case of configured dst |
| 113 | if(dst->total_size() > 0) |
| 114 | { |
| 115 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(dst, 1, DataType::F16, DataType::F32, DataType::S32); |
| 116 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src1, dst); |
| 117 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(detail::have_different_dimensions(out_shape, dst->tensor_shape(), 0), |
| 118 | "Wrong shape for dst"); |
| 119 | } |
| 120 | |
| 121 | return Status{}; |
| 122 | } |
| 123 | |
Michele Di Giorgio | 1e0208a | 2021-01-22 15:42:59 +0000 | [diff] [blame] | 124 | Status validate_arguments_with_arithmetic_rules(const ITensorInfo &src1, const ITensorInfo &src2, const ITensorInfo &dst) |
| 125 | { |
| 126 | ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(&src1); |
| 127 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&src1, 1, DataType::U8, DataType::QASYMM8, DataType::QASYMM8_SIGNED, |
| 128 | DataType::S16, DataType::QSYMM16, DataType::F16, |
| 129 | DataType::S32, DataType::F32); |
| 130 | ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(&src2); |
| 131 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&src2, 1, DataType::U8, DataType::QASYMM8, DataType::QASYMM8_SIGNED, |
| 132 | DataType::S16, DataType::QSYMM16, DataType::F16, |
| 133 | DataType::S32, DataType::F32); |
| 134 | |
| 135 | const bool is_quantized = is_data_type_quantized(src1.data_type()) || is_data_type_quantized(src2.data_type()); |
| 136 | if(is_quantized) |
| 137 | { |
| 138 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(&src1, &src2); |
| 139 | |
| 140 | if(is_data_type_quantized_symmetric(src1.data_type())) |
| 141 | { |
| 142 | const int32_t in1_offset = src1.quantization_info().uniform().offset; |
| 143 | const int32_t in2_offset = src2.quantization_info().uniform().offset; |
| 144 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(in1_offset != 0, "For quantized symmetric, offset must be zero"); |
| 145 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(in2_offset != 0, "For quantized symmetric, offset must be zero"); |
| 146 | } |
| 147 | } |
| 148 | |
| 149 | const TensorShape out_shape = TensorShape::broadcast_shape(src1.tensor_shape(), src2.tensor_shape()); |
| 150 | |
| 151 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(out_shape.total_size() == 0, "Inputs are not broadcast compatible"); |
| 152 | |
| 153 | // Validate in case of configured dst |
| 154 | if(dst.total_size() > 0) |
| 155 | { |
| 156 | ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(&dst); |
| 157 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&dst, 1, DataType::U8, DataType::QASYMM8, DataType::QASYMM8_SIGNED, |
| 158 | DataType::S16, DataType::QSYMM16, DataType::F16, |
| 159 | DataType::S32, DataType::F32); |
| 160 | ARM_COMPUTE_RETURN_ERROR_ON_MSG((dst.data_type() == DataType::U8) && ((src1.data_type() != DataType::U8) || (src2.data_type() != DataType::U8)), |
| 161 | "dst can only be U8 if both inputs are U8"); |
| 162 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(detail::have_different_dimensions(out_shape, dst.tensor_shape(), 0), |
| 163 | "Wrong shape for dst"); |
| 164 | |
| 165 | if(is_quantized) |
| 166 | { |
| 167 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(&src1, &dst); |
| 168 | |
| 169 | if(is_data_type_quantized_symmetric(dst.data_type())) |
| 170 | { |
| 171 | const int32_t offset = dst.quantization_info().uniform().offset; |
| 172 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(offset != 0, "For quantized symmetric, offset must be zero"); |
| 173 | } |
| 174 | } |
| 175 | } |
| 176 | return Status{}; |
| 177 | } |
| 178 | |
| 179 | CLBuildOptions generate_build_options_with_arithmetic_rules(const ITensorInfo &src1, const ITensorInfo &src2, const ITensorInfo &dst, const std::string &operation_string) |
| 180 | { |
| 181 | CLBuildOptions build_opts; |
| 182 | |
| 183 | const unsigned int num_elems_processed_per_iteration = adjust_vec_size(vector_size_byte_opencl / dst.element_size(), dst.dimension(0)); |
| 184 | |
| 185 | build_opts.add_option("-DDATA_TYPE_IN1=" + get_cl_type_from_data_type(src1.data_type())); |
| 186 | build_opts.add_option("-DDATA_TYPE_IN2=" + get_cl_type_from_data_type(src2.data_type())); |
| 187 | build_opts.add_option("-DDATA_TYPE_OUT=" + get_cl_type_from_data_type(dst.data_type())); |
| 188 | build_opts.add_option("-DVEC_SIZE_IN1=" + support::cpp11::to_string(src1.dimension(0) == 1 ? 1 : num_elems_processed_per_iteration)); |
| 189 | build_opts.add_option("-DVEC_SIZE_IN2=" + support::cpp11::to_string(src2.dimension(0) == 1 ? 1 : num_elems_processed_per_iteration)); |
| 190 | build_opts.add_option("-DVEC_SIZE_OUT=" + support::cpp11::to_string(num_elems_processed_per_iteration)); |
| 191 | build_opts.add_option("-DVEC_SIZE_LEFTOVER=" + support::cpp11::to_string(dst.dimension(0) % num_elems_processed_per_iteration)); |
| 192 | build_opts.add_option("-DOP=" + operation_string); |
| 193 | if(is_data_type_quantized(src1.data_type())) |
| 194 | { |
| 195 | const UniformQuantizationInfo iq1info = src1.quantization_info().uniform(); |
| 196 | const UniformQuantizationInfo iq2info = src2.quantization_info().uniform(); |
| 197 | const UniformQuantizationInfo oqinfo = dst.quantization_info().uniform(); |
| 198 | |
| 199 | build_opts.add_option("-DOFFSET_IN1=" + support::cpp11::to_string(iq1info.offset)); |
| 200 | build_opts.add_option("-DOFFSET_IN2=" + support::cpp11::to_string(iq2info.offset)); |
| 201 | build_opts.add_option("-DOFFSET_OUT=" + support::cpp11::to_string(oqinfo.offset)); |
| 202 | build_opts.add_option("-DSCALE_IN1=" + float_to_string_with_full_precision(iq1info.scale)); |
| 203 | build_opts.add_option("-DSCALE_IN2=" + float_to_string_with_full_precision(iq2info.scale)); |
| 204 | build_opts.add_option("-DSCALE_OUT=" + float_to_string_with_full_precision(oqinfo.scale)); |
| 205 | } |
Suhail Munshi | ae1c9fe | 2021-04-14 12:16:49 +0100 | [diff] [blame] | 206 | build_opts.add_option_if(src1.data_type() == DataType::S32, "-DS32"); |
| 207 | |
Michele Di Giorgio | 1e0208a | 2021-01-22 15:42:59 +0000 | [diff] [blame] | 208 | return build_opts; |
| 209 | } |
| 210 | |
| 211 | std::pair<Status, Window> configure_window_arithmetic_common(ITensorInfo &dst) |
| 212 | { |
| 213 | const unsigned int num_elems_processed_per_iteration = adjust_vec_size(vector_size_byte_opencl / dst.element_size(), dst.dimension(0)); |
| 214 | Window win = calculate_max_window(dst, Steps(num_elems_processed_per_iteration)); |
| 215 | return std::make_pair(Status{}, win); |
| 216 | } |
| 217 | |
| 218 | std::pair<Status, Window> validate_and_configure_window_for_arithmetic_operators(ITensorInfo &src1, ITensorInfo &src2, ITensorInfo &dst) |
| 219 | { |
| 220 | const std::pair<TensorShape, ValidRegion> broadcast_pair = ITensorInfo::broadcast_shape_and_valid_region(src1, src2); |
| 221 | const TensorShape &out_shape = broadcast_pair.first; |
| 222 | |
| 223 | set_shape_if_empty(dst, out_shape); |
| 224 | |
| 225 | if(src1.data_type() == DataType::S16 || src2.data_type() == DataType::S16) |
| 226 | { |
| 227 | set_format_if_unknown(dst, Format::S16); |
| 228 | } |
| 229 | else if(src1.data_type() == DataType::F16 || src2.data_type() == DataType::F16) |
| 230 | { |
| 231 | set_format_if_unknown(dst, Format::F16); |
| 232 | } |
| 233 | else if(src1.data_type() == DataType::F32 || src2.data_type() == DataType::F32) |
| 234 | { |
| 235 | set_format_if_unknown(dst, Format::F32); |
| 236 | } |
| 237 | else if(src1.data_type() == DataType::QASYMM8 || src2.data_type() == DataType::QASYMM8) |
| 238 | { |
| 239 | set_data_type_if_unknown(dst, DataType::QASYMM8); |
| 240 | } |
| 241 | else if(src1.data_type() == DataType::QASYMM8_SIGNED || src2.data_type() == DataType::QASYMM8_SIGNED) |
| 242 | { |
| 243 | set_data_type_if_unknown(dst, DataType::QASYMM8_SIGNED); |
| 244 | } |
| 245 | else if(src1.data_type() == DataType::QSYMM16 || src2.data_type() == DataType::QSYMM16) |
| 246 | { |
| 247 | set_data_type_if_unknown(dst, DataType::QSYMM16); |
| 248 | } |
| 249 | |
| 250 | return configure_window_arithmetic_common(dst); |
| 251 | } |
| 252 | |
| 253 | std::pair<Status, Window> validate_and_configure_window_for_logical_binary_operators(ITensorInfo &src1, ITensorInfo &src2, ITensorInfo &dst) |
| 254 | { |
| 255 | const std::pair<TensorShape, ValidRegion> broadcast_pair = ITensorInfo::broadcast_shape_and_valid_region(src1, src2); |
| 256 | const TensorShape &out_shape = broadcast_pair.first; |
| 257 | |
| 258 | set_shape_if_empty(dst, out_shape); |
| 259 | set_data_type_if_unknown(dst, DataType::U8); |
| 260 | |
| 261 | // The arithmetic utility functions can be share |
| 262 | return configure_window_arithmetic_common(dst); |
| 263 | } |
| 264 | |
| 265 | std::pair<Status, Window> validate_and_configure_window_for_division(ITensorInfo &src1, ITensorInfo &src2, ITensorInfo &dst) |
| 266 | { |
| 267 | const std::pair<TensorShape, ValidRegion> broadcast_pair = ITensorInfo::broadcast_shape_and_valid_region(src1, src2); |
| 268 | const TensorShape &out_shape = broadcast_pair.first; |
| 269 | auto_init_if_empty(dst, out_shape, 1, src1.data_type()); |
| 270 | return configure_window_arithmetic_common(dst); |
| 271 | } |
| 272 | } // namespace |
| 273 | |
Giorgio Arena | 4a95bba | 2021-06-28 11:00:27 +0100 | [diff] [blame] | 274 | ClElementwiseKernel::ClElementwiseKernel() |
| 275 | { |
| 276 | _type = CLKernelType::ELEMENTWISE; |
| 277 | } |
| 278 | |
Michele Di Giorgio | 1e0208a | 2021-01-22 15:42:59 +0000 | [diff] [blame] | 279 | void ClElementwiseKernel::configure_common(const ClCompileContext &compile_context, ITensorInfo *src1, ITensorInfo *src2, ITensorInfo *dst) |
| 280 | { |
| 281 | // Configure kernel window |
| 282 | auto win_config = validate_and_configure_window(*src1, *src2, *dst); |
| 283 | ARM_COMPUTE_ERROR_THROW_ON(win_config.first); |
| 284 | |
Michele Di Giorgio | 1e0208a | 2021-01-22 15:42:59 +0000 | [diff] [blame] | 285 | std::string kernel_name = "elementwise_operation_" + name(); |
| 286 | if(is_data_type_quantized(src1->data_type())) |
| 287 | { |
| 288 | kernel_name += "_quantized"; |
| 289 | } |
| 290 | |
| 291 | // Set kernel build options |
| 292 | CLBuildOptions build_opts = generate_build_options(*src1, *src2, *dst); |
| 293 | if(_act_info.enabled()) |
| 294 | { |
| 295 | build_opts.add_option("-DACTIVATION_TYPE=" + lower_string(string_from_activation_func(_act_info.activation()))); |
| 296 | build_opts.add_option("-DA_VAL=" + float_to_string_with_full_precision(_act_info.a())); |
| 297 | build_opts.add_option("-DB_VAL=" + float_to_string_with_full_precision(_act_info.b())); |
| 298 | } |
| 299 | |
| 300 | // Create kernel |
| 301 | _kernel = create_kernel(compile_context, kernel_name, build_opts.options()); |
| 302 | |
| 303 | ICLKernel::configure_internal(win_config.second); |
| 304 | |
| 305 | _config_id = generate_id_for_tuning(kernel_name, *src1, *dst); |
| 306 | } |
| 307 | |
| 308 | void ClElementwiseKernel::run_op(ITensorPack &tensors, const Window &window, ::cl::CommandQueue &queue) |
| 309 | { |
| 310 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 311 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window); |
| 312 | |
| 313 | const auto src_0 = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_0)); |
| 314 | const auto src_1 = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_1)); |
| 315 | auto dst = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST)); |
| 316 | |
| 317 | const TensorShape &in_shape1 = src_0->info()->tensor_shape(); |
| 318 | const TensorShape &in_shape2 = src_1->info()->tensor_shape(); |
| 319 | const TensorShape &out_shape = dst->info()->tensor_shape(); |
| 320 | |
| 321 | bool can_collapse = true; |
| 322 | const bool is_vector = in_shape1.num_dimensions() == 1 || in_shape2.num_dimensions() == 1; |
| 323 | if(std::min(in_shape1.total_size(), in_shape2.total_size()) > 1 && !is_vector) |
| 324 | { |
| 325 | can_collapse = (std::min(in_shape1.num_dimensions(), in_shape2.num_dimensions()) > Window::DimZ); |
| 326 | for(size_t d = Window::DimZ; can_collapse && (d < out_shape.num_dimensions()); d++) |
| 327 | { |
| 328 | can_collapse = (in_shape1[d] == in_shape2[d]); |
| 329 | } |
| 330 | } |
| 331 | |
| 332 | bool has_collapsed = false; |
| 333 | Window collapsed = can_collapse ? window.collapse_if_possible(ICLKernel::window(), Window::DimZ, &has_collapsed) : window; |
| 334 | |
| 335 | const TensorShape &in_shape1_collapsed = has_collapsed ? in_shape1.collapsed_from(Window::DimZ) : in_shape1; |
| 336 | const TensorShape &in_shape2_collapsed = has_collapsed ? in_shape2.collapsed_from(Window::DimZ) : in_shape2; |
| 337 | |
| 338 | Window slice = collapsed.first_slice_window_3D(); |
| 339 | Window slice_src1 = slice.broadcast_if_dimension_le_one(in_shape1_collapsed); |
| 340 | Window slice_src2 = slice.broadcast_if_dimension_le_one(in_shape2_collapsed); |
| 341 | do |
| 342 | { |
| 343 | unsigned int idx = 0; |
| 344 | add_3D_tensor_argument(idx, src_0, slice_src1); |
| 345 | add_3D_tensor_argument(idx, src_1, slice_src2); |
| 346 | add_3D_tensor_argument(idx, dst, slice); |
| 347 | |
| 348 | enqueue(queue, *this, slice, lws_hint()); |
| 349 | ARM_COMPUTE_UNUSED(collapsed.slide_window_slice_3D(slice_src1)); |
| 350 | ARM_COMPUTE_UNUSED(collapsed.slide_window_slice_3D(slice_src2)); |
| 351 | } |
| 352 | while(collapsed.slide_window_slice_3D(slice)); |
| 353 | } |
| 354 | |
| 355 | /** Logical binary */ |
| 356 | |
| 357 | void ClLogicalBinaryKernel::configure(const ClCompileContext &compile_context, LogicalOperation op, ITensorInfo *src1, ITensorInfo *src2, ITensorInfo *dst) |
| 358 | { |
| 359 | ARM_COMPUTE_ERROR_ON_NULLPTR(src1, src2, dst); |
| 360 | ARM_COMPUTE_ERROR_THROW_ON(ClLogicalBinaryKernel::validate(op, src1, src2, dst)); |
| 361 | _op = op; |
| 362 | configure_common(compile_context, src1, src2, dst); |
| 363 | } |
| 364 | |
| 365 | Status ClLogicalBinaryKernel::validate(LogicalOperation op, const ITensorInfo *src1, const ITensorInfo *src2, const ITensorInfo *dst) |
| 366 | { |
| 367 | ARM_COMPUTE_UNUSED(op); |
| 368 | ARM_COMPUTE_ASSERT(op != LogicalOperation::Unknown && op != LogicalOperation::Not); |
| 369 | ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src1, src2, dst); |
| 370 | |
| 371 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src1, 1, DataType::U8); |
| 372 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src1, src2); |
| 373 | |
| 374 | ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments_with_arithmetic_rules(*src1, *src2, *dst)); |
| 375 | ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window_for_logical_binary_operators(*src1->clone(), *src2->clone(), *dst->clone()).first); |
| 376 | |
| 377 | return Status{}; |
| 378 | } |
| 379 | |
| 380 | std::string ClLogicalBinaryKernel::name() |
| 381 | { |
| 382 | switch(_op) |
| 383 | { |
| 384 | case LogicalOperation::And: |
| 385 | return "AND"; |
| 386 | case LogicalOperation::Or: |
| 387 | return "OR"; |
| 388 | case LogicalOperation::Not: |
| 389 | /* fall through */ |
| 390 | default: |
| 391 | ARM_COMPUTE_ASSERT(true); |
| 392 | } |
| 393 | return ""; |
| 394 | } |
| 395 | |
| 396 | std::pair<Status, Window> ClLogicalBinaryKernel::validate_and_configure_window(ITensorInfo &src1, ITensorInfo &src2, ITensorInfo &dst) |
| 397 | { |
| 398 | return validate_and_configure_window_for_logical_binary_operators(src1, src2, dst); |
| 399 | } |
| 400 | |
| 401 | CLBuildOptions ClLogicalBinaryKernel::generate_build_options(const ITensorInfo &src1, const ITensorInfo &src2, const ITensorInfo &dst) |
| 402 | { |
| 403 | // The arithmetic utility functions can be share |
| 404 | return generate_build_options_with_arithmetic_rules(src1, src2, dst, name()); |
| 405 | } |
| 406 | |
| 407 | std::string ClLogicalBinaryKernel::generate_id_for_tuning(const std::string &kernel_name, const ITensorInfo &src1, const ITensorInfo &dst) |
| 408 | { |
| 409 | return generate_id_for_tuning_common(kernel_name, src1, dst); |
| 410 | } |
| 411 | |
| 412 | /** Arithmetic operations with saturation*/ |
| 413 | void ClSaturatedArithmeticKernel::configure(const ClCompileContext &compile_context, ArithmeticOperation op, ITensorInfo *input1, ITensorInfo *input2, ITensorInfo *output, |
| 414 | const ConvertPolicy &policy, |
| 415 | const ActivationLayerInfo &act_info) |
| 416 | { |
| 417 | ARM_COMPUTE_ERROR_ON_NULLPTR(input1, input2, output); |
| 418 | ARM_COMPUTE_ERROR_THROW_ON(ClSaturatedArithmeticKernel::validate(op, input1, input2, output, policy, act_info)); |
| 419 | auto padding_info = get_padding_info({ input1, input2, output }); |
| 420 | |
| 421 | _policy = policy; |
| 422 | _op = op; |
| 423 | _act_info = act_info; |
| 424 | configure_common(compile_context, input1, input2, output); |
| 425 | ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info)); |
| 426 | } |
| 427 | |
| 428 | Status ClSaturatedArithmeticKernel::validate(ArithmeticOperation op, const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, const ConvertPolicy &policy, |
| 429 | const ActivationLayerInfo &act_info) |
| 430 | { |
| 431 | ARM_COMPUTE_UNUSED(op, policy); |
| 432 | ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input1, input2, output); |
| 433 | ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments_with_arithmetic_rules(*input1, *input2, *output)); |
| 434 | ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window_for_arithmetic_operators(*input1->clone(), *input2->clone(), *output->clone()).first); |
| 435 | ARM_COMPUTE_RETURN_ERROR_ON(act_info.enabled() && !is_data_type_float(output->data_type())); |
| 436 | |
| 437 | return Status{}; |
| 438 | } |
| 439 | |
| 440 | std::pair<Status, Window> ClSaturatedArithmeticKernel::validate_and_configure_window(ITensorInfo &input1, ITensorInfo &input2, ITensorInfo &output) |
| 441 | { |
| 442 | return validate_and_configure_window_for_arithmetic_operators(input1, input2, output); |
| 443 | } |
| 444 | |
| 445 | CLBuildOptions ClSaturatedArithmeticKernel::generate_build_options(const ITensorInfo &input1, const ITensorInfo &input2, const ITensorInfo &output) |
| 446 | { |
| 447 | const bool has_float_out = is_data_type_float(output.data_type()); |
| 448 | auto build_options = generate_build_options_with_arithmetic_rules(input1, input2, output, name()); |
| 449 | build_options.add_option((_policy == ConvertPolicy::WRAP || has_float_out) ? "-DWRAP" : "-DSATURATE"); |
| 450 | return build_options; |
| 451 | } |
| 452 | |
| 453 | std::string ClSaturatedArithmeticKernel::generate_id_for_tuning(const std::string &kernel_name, const ITensorInfo &input1, const ITensorInfo &output) |
| 454 | { |
| 455 | auto config_id = generate_id_for_tuning_common(kernel_name, input1, output); |
| 456 | config_id += (_policy == ConvertPolicy::WRAP) ? "_wrap_" : "_saturate_"; |
| 457 | config_id += lower_string(string_from_data_layout(input1.data_layout())); |
| 458 | return config_id; |
| 459 | } |
| 460 | |
| 461 | std::string ClSaturatedArithmeticKernel::name() |
| 462 | { |
| 463 | return supported_sat_arithmetic_ops[_op]; |
| 464 | } |
| 465 | |
| 466 | /** Arithmetic operations*/ |
| 467 | void ClArithmeticKernel::configure(const ClCompileContext &compile_context, ArithmeticOperation op, ITensorInfo *src1, ITensorInfo *src2, ITensorInfo *dst, |
| 468 | const ActivationLayerInfo &act_info) |
| 469 | { |
| 470 | ARM_COMPUTE_ERROR_ON_NULLPTR(src1, src2, dst); |
| 471 | ARM_COMPUTE_ERROR_THROW_ON(ClArithmeticKernel::validate(op, src1, src2, dst, act_info)); |
| 472 | auto padding_info = get_padding_info({ src1, src2, dst }); |
| 473 | |
| 474 | _op = op; |
| 475 | _act_info = act_info; |
| 476 | configure_common(compile_context, src1, src2, dst); |
| 477 | ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info)); |
| 478 | } |
| 479 | |
| 480 | Status ClArithmeticKernel::validate(ArithmeticOperation op, const ITensorInfo *src1, const ITensorInfo *src2, const ITensorInfo *dst, const ActivationLayerInfo &act_info) |
| 481 | { |
| 482 | ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src1, src2, dst); |
Suhail Munshi | ae1c9fe | 2021-04-14 12:16:49 +0100 | [diff] [blame] | 483 | if(op == ArithmeticOperation::DIV) |
Michele Di Giorgio | 1e0208a | 2021-01-22 15:42:59 +0000 | [diff] [blame] | 484 | { |
Suhail Munshi | ae1c9fe | 2021-04-14 12:16:49 +0100 | [diff] [blame] | 485 | // Partial integer support S32/F32/F16 |
| 486 | ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments_divide_operation(src1, src2, dst)); |
| 487 | ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window_for_division(*src1->clone(), *src2->clone(), *dst->clone()).first); |
| 488 | } |
| 489 | else if(op == ArithmeticOperation::POWER) |
| 490 | { |
| 491 | // Power operators doesn't support integer arithmetic |
Michele Di Giorgio | 1e0208a | 2021-01-22 15:42:59 +0000 | [diff] [blame] | 492 | ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments_with_float_only_supported_rules(*src1, *src2, *dst)); |
| 493 | ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window_for_division(*src1->clone(), *src2->clone(), *dst->clone()).first); |
| 494 | } |
| 495 | else |
| 496 | { |
| 497 | ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments_with_arithmetic_rules(*src1, *src2, *dst)); |
| 498 | ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window_for_arithmetic_operators(*src1->clone(), *src2->clone(), *dst->clone()).first); |
| 499 | } |
| 500 | ARM_COMPUTE_RETURN_ERROR_ON(act_info.enabled() && !is_data_type_float(dst->data_type())); |
| 501 | |
| 502 | return Status{}; |
| 503 | } |
| 504 | std::pair<Status, Window> ClArithmeticKernel::validate_and_configure_window(ITensorInfo &src1, ITensorInfo &src2, ITensorInfo &dst) |
| 505 | { |
| 506 | if(_op == ArithmeticOperation::DIV || _op == ArithmeticOperation::POWER) |
| 507 | { |
| 508 | // Division and Power operators don't support integer arithmetic |
| 509 | return validate_and_configure_window_for_division(src1, src2, dst); |
| 510 | } |
| 511 | else |
| 512 | { |
| 513 | return validate_and_configure_window_for_arithmetic_operators(src1, src2, dst); |
| 514 | } |
| 515 | } |
| 516 | |
| 517 | CLBuildOptions ClArithmeticKernel::generate_build_options(const ITensorInfo &src1, const ITensorInfo &src2, const ITensorInfo &dst) |
| 518 | { |
| 519 | return generate_build_options_with_arithmetic_rules(src1, src2, dst, name()); |
| 520 | } |
| 521 | std::string ClArithmeticKernel::generate_id_for_tuning(const std::string &kernel_name, const ITensorInfo &src1, const ITensorInfo &dst) |
| 522 | { |
| 523 | return generate_id_for_tuning_common(kernel_name, src1, dst); |
| 524 | } |
| 525 | |
| 526 | std::string ClArithmeticKernel::name() |
| 527 | { |
| 528 | return supported_arithmetic_ops[_op]; |
| 529 | } |
| 530 | } // namespace kernels |
| 531 | } // namespace opencl |
| 532 | } // namespace arm_compute |