Isabella Gottardi | 01a214a | 2018-04-09 16:00:52 +0100 | [diff] [blame] | 1 | /* |
Georgios Pinitas | 108a95e | 2019-03-27 13:55:59 +0000 | [diff] [blame] | 2 | * Copyright (c) 2017-2019 ARM Limited. |
Isabella Gottardi | 01a214a | 2018-04-09 16:00:52 +0100 | [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 | #ifndef ARM_COMPUTE_CL /* Needed by Utils.cpp to handle OpenCL exceptions properly */ |
| 25 | #error "This example needs to be built with -DARM_COMPUTE_CL" |
| 26 | #endif /* ARM_COMPUTE_CL */ |
| 27 | |
| 28 | #include "arm_compute/core/Types.h" |
| 29 | #include "arm_compute/core/utils/quantization/AsymmHelpers.h" |
| 30 | #include "arm_compute/runtime/CL/CLFunctions.h" |
| 31 | #include "arm_compute/runtime/CL/CLScheduler.h" |
| 32 | |
| 33 | #include "tests/AssetsLibrary.h" |
| 34 | #include "tests/CL/CLAccessor.h" |
| 35 | #include "tests/Globals.h" |
| 36 | #include "tests/IAccessor.h" |
| 37 | #include "tests/SimpleTensor.h" |
| 38 | #include "tests/validation/Validation.h" |
| 39 | #include "tests/validation/reference/GEMM.h" |
| 40 | #include "tests/validation/reference/GEMMLowp.h" |
| 41 | |
Georgios Pinitas | 108a95e | 2019-03-27 13:55:59 +0000 | [diff] [blame] | 42 | #include "utils/TypePrinter.h" |
Isabella Gottardi | 01a214a | 2018-04-09 16:00:52 +0100 | [diff] [blame] | 43 | #include "utils/Utils.h" |
Georgios Pinitas | 108a95e | 2019-03-27 13:55:59 +0000 | [diff] [blame] | 44 | #include "utils/command_line/CommandLineOptions.h" |
| 45 | #include "utils/command_line/CommandLineParser.h" |
| 46 | |
| 47 | #include "ValidateExample.h" |
Isabella Gottardi | 01a214a | 2018-04-09 16:00:52 +0100 | [diff] [blame] | 48 | |
| 49 | #include <cstdlib> |
| 50 | |
| 51 | using namespace arm_compute; |
| 52 | using namespace utils; |
| 53 | using namespace arm_compute::test; |
| 54 | using namespace arm_compute::test::validation; |
| 55 | |
| 56 | constexpr float abs_tolerance_f32(0.0001f); /**< F32 Absolute tolerance value for comparing reference's output against implementation's output for |
| 57 | * floating point data types in case using relative tolerance fails because of small values */ |
| 58 | RelativeTolerance<float> tolerance_f32(0.001f); /**< F32 Tolerance value for comparing reference's output against implementation's output for floating point data types */ |
| 59 | RelativeTolerance<half_float::half> tolerance_f16(half(0.2)); /**< F16 Tolerance value for comparing reference's output against implementation's output for floating point data types */ |
| 60 | constexpr float tolerance_num_f16 = 0.02f; /**< F16 Tolerance number */ |
| 61 | |
Georgios Pinitas | 108a95e | 2019-03-27 13:55:59 +0000 | [diff] [blame] | 62 | namespace arm_compute |
| 63 | { |
| 64 | DataType data_type_from_name(const std::string &name) |
| 65 | { |
| 66 | static const std::map<std::string, DataType> data_types = |
| 67 | { |
| 68 | { "f16", DataType::F16 }, |
| 69 | { "f32", DataType::F32 }, |
| 70 | { "qasymm8", DataType::QASYMM8 }, |
| 71 | }; |
| 72 | |
| 73 | #ifndef ARM_COMPUTE_EXCEPTIONS_DISABLED |
| 74 | try |
| 75 | { |
| 76 | #endif /* ARM_COMPUTE_EXCEPTIONS_DISABLED */ |
| 77 | return data_types.at(utility::tolower(name)); |
| 78 | |
| 79 | #ifndef ARM_COMPUTE_EXCEPTIONS_DISABLED |
| 80 | } |
| 81 | catch(const std::out_of_range &) |
| 82 | { |
| 83 | throw std::invalid_argument(name); |
| 84 | } |
| 85 | #endif /* ARM_COMPUTE_EXCEPTIONS_DISABLED */ |
| 86 | } |
| 87 | |
| 88 | inline ::std::istream &operator>>(::std::istream &stream, DataType &data_type) |
| 89 | { |
| 90 | std::string value; |
| 91 | stream >> value; |
| 92 | data_type = data_type_from_name(value); |
| 93 | return stream; |
| 94 | } |
| 95 | } // namespace arm_compute |
| 96 | namespace |
| 97 | { |
| 98 | class GEMMCommandLineOptions final |
| 99 | { |
| 100 | public: |
| 101 | explicit GEMMCommandLineOptions(CommandLineParser &parser) noexcept |
| 102 | : help(parser.add_option<ToggleOption>("help")), |
| 103 | add_bias(parser.add_option<ToggleOption>("add_bias")), |
| 104 | M(parser.add_option<SimpleOption<int>>("m", 7)), |
| 105 | N(parser.add_option<SimpleOption<int>>("n", 3)), |
| 106 | K(parser.add_option<SimpleOption<int>>("k", 5)), |
| 107 | B(parser.add_option<SimpleOption<int>>("b", 1)), |
| 108 | alpha(parser.add_option<SimpleOption<float>>("alpha", 1.f)), |
| 109 | beta(parser.add_option<SimpleOption<float>>("beta", 0.f)), |
| 110 | offset_src0(parser.add_option<SimpleOption<int>>("offset_i0", 10)), |
| 111 | offset_src1(parser.add_option<SimpleOption<int>>("offset_i1", 10)), |
| 112 | offset_dst(parser.add_option<SimpleOption<int>>("offset_o", 10)), |
| 113 | scale_src0(parser.add_option<SimpleOption<float>>("scale_i0", 1.f / 255)), |
| 114 | scale_src1(parser.add_option<SimpleOption<float>>("scale_i1", 1.f / 255)), |
| 115 | scale_dst(parser.add_option<SimpleOption<float>>("scale_o", 1.f / 255)), |
| 116 | data_type() |
| 117 | { |
| 118 | // Setup data type |
| 119 | const std::set<arm_compute::DataType> supported_data_types |
| 120 | { |
| 121 | DataType::F16, |
| 122 | DataType::F32, |
| 123 | DataType::QASYMM8, |
| 124 | }; |
| 125 | data_type = parser.add_option<EnumOption<DataType>>("type", supported_data_types, DataType::F32); |
| 126 | |
| 127 | // Setup help strings |
| 128 | help->set_help("Show this help message"); |
| 129 | add_bias->set_help("Add bias to the GEMM. Used when running in QASYMM8"); |
| 130 | M->set_help("M value"); |
| 131 | N->set_help("N value"); |
| 132 | K->set_help("K value"); |
| 133 | B->set_help("B value - number of batches"); |
| 134 | alpha->set_help("Alpha value"); |
| 135 | beta->set_help("Beta value"); |
| 136 | offset_src0->set_help("Offset of first input. Used when running in QASYMM8"); |
| 137 | offset_src1->set_help("Offset of second input. Used when running in QASYMM8"); |
| 138 | offset_dst->set_help("Offset of output. Used when running in QASYMM8"); |
| 139 | scale_src0->set_help("Scale of first input. Used when running in QASYMM8"); |
| 140 | scale_src1->set_help("Scale of second input. Used when running in QASYMM8"); |
| 141 | scale_dst->set_help("Scale of output. Used when running in QASYMM8"); |
| 142 | data_type->set_help("Data type to use"); |
| 143 | } |
| 144 | /** Prevent instances of this class from being copied (As this class contains pointers) */ |
| 145 | GEMMCommandLineOptions(const GEMMCommandLineOptions &) = delete; |
| 146 | /** Prevent instances of this class from being copied (As this class contains pointers) */ |
| 147 | GEMMCommandLineOptions &operator=(const GEMMCommandLineOptions &) = delete; |
| 148 | /** Allow instances of this class to be moved */ |
| 149 | GEMMCommandLineOptions(GEMMCommandLineOptions &&) noexcept(true) = default; |
| 150 | /** Allow instances of this class to be moved */ |
| 151 | GEMMCommandLineOptions &operator=(GEMMCommandLineOptions &&) noexcept(true) = default; |
| 152 | /** Default destructor */ |
| 153 | ~GEMMCommandLineOptions() = default; |
| 154 | |
| 155 | public: |
| 156 | ToggleOption *help; |
| 157 | ToggleOption *add_bias; |
| 158 | SimpleOption<int> *M; |
| 159 | SimpleOption<int> *N; |
| 160 | SimpleOption<int> *K; |
| 161 | SimpleOption<int> *B; |
| 162 | SimpleOption<float> *alpha; |
| 163 | SimpleOption<float> *beta; |
| 164 | SimpleOption<int> *offset_src0; |
| 165 | SimpleOption<int> *offset_src1; |
| 166 | SimpleOption<int> *offset_dst; |
| 167 | SimpleOption<float> *scale_src0; |
| 168 | SimpleOption<float> *scale_src1; |
| 169 | SimpleOption<float> *scale_dst; |
| 170 | EnumOption<arm_compute::DataType> *data_type; |
| 171 | }; |
| 172 | } // namespace |
| 173 | |
Isabella Gottardi | 01a214a | 2018-04-09 16:00:52 +0100 | [diff] [blame] | 174 | class CLGEMMValidateExample : public ValidateExample |
| 175 | { |
| 176 | public: |
Anthony Barbier | e88b9bb | 2018-07-12 13:26:27 +0100 | [diff] [blame] | 177 | bool do_setup(int argc, char **argv) override |
Isabella Gottardi | 01a214a | 2018-04-09 16:00:52 +0100 | [diff] [blame] | 178 | { |
| 179 | CLScheduler::get().default_init(); |
Georgios Pinitas | 108a95e | 2019-03-27 13:55:59 +0000 | [diff] [blame] | 180 | |
| 181 | // Parse options |
| 182 | CommandLineParser parser; |
| 183 | GEMMCommandLineOptions gemm_options(parser); |
| 184 | parser.parse(argc, argv); |
| 185 | |
| 186 | // Print help |
| 187 | const bool print_help = gemm_options.help->is_set() ? gemm_options.help->value() : false; |
| 188 | if(print_help) |
Isabella Gottardi | 01a214a | 2018-04-09 16:00:52 +0100 | [diff] [blame] | 189 | { |
Georgios Pinitas | 108a95e | 2019-03-27 13:55:59 +0000 | [diff] [blame] | 190 | parser.print_help(argv[0]); |
| 191 | return false; |
Isabella Gottardi | 01a214a | 2018-04-09 16:00:52 +0100 | [diff] [blame] | 192 | } |
| 193 | |
Georgios Pinitas | 108a95e | 2019-03-27 13:55:59 +0000 | [diff] [blame] | 194 | // Consume parameters |
| 195 | consume_params(gemm_options); |
| 196 | print_parameters_internal(); |
| 197 | |
| 198 | // Calculate re-quantization parameters |
| 199 | if(data_type == DataType::QASYMM8) |
Isabella Gottardi | 01a214a | 2018-04-09 16:00:52 +0100 | [diff] [blame] | 200 | { |
Georgios Pinitas | 108a95e | 2019-03-27 13:55:59 +0000 | [diff] [blame] | 201 | float multiplier = scale_src0 * scale_src1 / scale_dst; |
| 202 | quantization::calculate_quantized_multiplier_less_than_one(multiplier, &dst_multiplier, &dst_shift); |
Isabella Gottardi | 01a214a | 2018-04-09 16:00:52 +0100 | [diff] [blame] | 203 | } |
| 204 | |
Georgios Pinitas | 108a95e | 2019-03-27 13:55:59 +0000 | [diff] [blame] | 205 | // Initialize GEMM inputs/outputs |
| 206 | src0.allocator()->init(TensorInfo(TensorShape(K, M, B), 1, data_type)); |
| 207 | src1.allocator()->init(TensorInfo(TensorShape(N, K, B), 1, data_type)); |
| 208 | src2.allocator()->init(TensorInfo(TensorShape(N, M, B), 1, data_type)); |
Isabella Gottardi | 01a214a | 2018-04-09 16:00:52 +0100 | [diff] [blame] | 209 | init_sgemm_output(dst, src0, src1, data_type); |
| 210 | |
| 211 | // Configure function |
| 212 | if(data_type == DataType::QASYMM8) |
| 213 | { |
| 214 | src0.info()->set_quantization_info(QuantizationInfo(scale_src0, offset_src0)); |
| 215 | src1.info()->set_quantization_info(QuantizationInfo(scale_src1, offset_src1)); |
| 216 | dst.info()->set_quantization_info(QuantizationInfo(scale_dst, offset_dst)); |
| 217 | biases.allocator()->init(TensorInfo(TensorShape(N), 1, DataType::S32)); |
| 218 | init_sgemm_output(tmp_dst, src0, src1, DataType::S32); |
| 219 | |
| 220 | // Configure GEMMlowp matrix multiply function |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 221 | mm_gemmlowp.configure(&src0, &src1, nullptr, &tmp_dst); |
Isabella Gottardi | 01a214a | 2018-04-09 16:00:52 +0100 | [diff] [blame] | 222 | |
| 223 | // Configure GEMMlowp output stage |
| 224 | mm_gemmlowp_output_stage.configure(&tmp_dst, add_bias ? &biases : nullptr, &dst, dst_multiplier, dst_shift, offset_dst); |
| 225 | tmp_dst.allocator()->allocate(); |
| 226 | biases.allocator()->allocate(); |
| 227 | fill(CLAccessor(biases), 3); |
| 228 | } |
| 229 | else |
| 230 | { |
| 231 | // Configure matrix multiply function |
| 232 | mm_gemm.configure(&src0, &src1, &src2, &dst, alpha, beta); |
| 233 | } |
| 234 | |
| 235 | // Allocate all the tensors |
| 236 | src0.allocator()->allocate(); |
| 237 | src1.allocator()->allocate(); |
| 238 | dst.allocator()->allocate(); |
| 239 | src2.allocator()->allocate(); |
| 240 | |
| 241 | fill(CLAccessor(src0), 0); |
| 242 | fill(CLAccessor(src1), 1); |
| 243 | fill(CLAccessor(src2), 2); |
Anthony Barbier | e88b9bb | 2018-07-12 13:26:27 +0100 | [diff] [blame] | 244 | |
| 245 | return true; |
Isabella Gottardi | 01a214a | 2018-04-09 16:00:52 +0100 | [diff] [blame] | 246 | } |
| 247 | |
Georgios Pinitas | 108a95e | 2019-03-27 13:55:59 +0000 | [diff] [blame] | 248 | void print_parameters_internal() |
Isabella Gottardi | 01a214a | 2018-04-09 16:00:52 +0100 | [diff] [blame] | 249 | { |
Georgios Pinitas | 108a95e | 2019-03-27 13:55:59 +0000 | [diff] [blame] | 250 | std::cout << "Datatype : " << string_from_data_type(data_type) << "\n"; |
| 251 | std::cout << "M : " << support::cpp11::to_string(M) << "\n"; |
| 252 | std::cout << "N : " << support::cpp11::to_string(N) << "\n"; |
| 253 | std::cout << "K : " << support::cpp11::to_string(K) << "\n"; |
| 254 | std::cout << "B : " << support::cpp11::to_string(B) << "\n"; |
Isabella Gottardi | 01a214a | 2018-04-09 16:00:52 +0100 | [diff] [blame] | 255 | if(data_type == DataType::QASYMM8) |
| 256 | { |
Georgios Pinitas | 108a95e | 2019-03-27 13:55:59 +0000 | [diff] [blame] | 257 | std::cout << "Scale_Src0 : " << support::cpp11::to_string(scale_src0) << "\n"; |
| 258 | std::cout << "Offset_Src0 : " << support::cpp11::to_string(offset_src0) << "\n"; |
| 259 | std::cout << "Scale_Scr1 : " << support::cpp11::to_string(scale_src1) << "\n"; |
| 260 | std::cout << "Offset_Src1 : " << support::cpp11::to_string(offset_src1) << "\n"; |
| 261 | std::cout << "Scale_Dst : " << support::cpp11::to_string(scale_dst) << "\n"; |
| 262 | std::cout << "Offset_Dst : " << support::cpp11::to_string(offset_dst) << "\n"; |
| 263 | std::cout << "Bias : " << support::cpp11::to_string(add_bias) << "\n"; |
Isabella Gottardi | 01a214a | 2018-04-09 16:00:52 +0100 | [diff] [blame] | 264 | } |
| 265 | else |
| 266 | { |
Georgios Pinitas | 108a95e | 2019-03-27 13:55:59 +0000 | [diff] [blame] | 267 | std::cout << "Alpha : " << support::cpp11::to_string(alpha) << "\n"; |
| 268 | std::cout << "Beta : " << support::cpp11::to_string(beta) << "\n"; |
Isabella Gottardi | 01a214a | 2018-04-09 16:00:52 +0100 | [diff] [blame] | 269 | } |
| 270 | } |
| 271 | |
| 272 | void do_validate() override |
| 273 | { |
| 274 | switch(data_type) |
| 275 | { |
| 276 | case DataType::F16: |
| 277 | { |
Georgios Pinitas | 108a95e | 2019-03-27 13:55:59 +0000 | [diff] [blame] | 278 | SimpleTensor<half> ref_src0 = { TensorShape(K, M, B), data_type, 1 }; |
| 279 | SimpleTensor<half> ref_src1 = { TensorShape(N, K, B), data_type, 1 }; |
| 280 | SimpleTensor<half> ref_src2 = { TensorShape(N, M, B), data_type, 1 }; |
Isabella Gottardi | 01a214a | 2018-04-09 16:00:52 +0100 | [diff] [blame] | 281 | |
| 282 | fill(ref_src0, 0); |
| 283 | fill(ref_src1, 1); |
| 284 | fill(ref_src2, 2); |
| 285 | |
| 286 | SimpleTensor<half> ref_dst = reference::gemm<half>(ref_src0, ref_src1, ref_src2, alpha, beta); |
| 287 | validate(CLAccessor(dst), ref_dst, tolerance_f16, tolerance_num_f16); |
| 288 | break; |
| 289 | } |
| 290 | case DataType::F32: |
| 291 | { |
Georgios Pinitas | 108a95e | 2019-03-27 13:55:59 +0000 | [diff] [blame] | 292 | SimpleTensor<float> ref_src0 = { TensorShape(K, M, B), data_type, 1 }; |
| 293 | SimpleTensor<float> ref_src1 = { TensorShape(N, K, B), data_type, 1 }; |
| 294 | SimpleTensor<float> ref_src2 = { TensorShape(N, M, B), data_type, 1 }; |
Isabella Gottardi | 01a214a | 2018-04-09 16:00:52 +0100 | [diff] [blame] | 295 | |
| 296 | fill(ref_src0, 0); |
| 297 | fill(ref_src1, 1); |
| 298 | fill(ref_src2, 2); |
| 299 | |
| 300 | SimpleTensor<float> ref_dst = reference::gemm<float>(ref_src0, ref_src1, ref_src2, alpha, beta); |
| 301 | validate(CLAccessor(dst), ref_dst, tolerance_f32, 0.f, abs_tolerance_f32); |
| 302 | break; |
| 303 | } |
| 304 | case DataType::QASYMM8: |
| 305 | { |
Georgios Pinitas | 108a95e | 2019-03-27 13:55:59 +0000 | [diff] [blame] | 306 | SimpleTensor<uint8_t> ref_src0{ TensorShape(K, M, B), data_type, 1 }; |
| 307 | SimpleTensor<uint8_t> ref_src1{ TensorShape(N, K, B), data_type, 1 }; |
Isabella Gottardi | 01a214a | 2018-04-09 16:00:52 +0100 | [diff] [blame] | 308 | SimpleTensor<uint8_t> ref_dst; |
| 309 | |
| 310 | // Fill reference |
| 311 | fill(ref_src0, 0); |
| 312 | fill(ref_src1, 1); |
| 313 | |
Georgios Pinitas | 108a95e | 2019-03-27 13:55:59 +0000 | [diff] [blame] | 314 | SimpleTensor<int32_t> ref_tmp_dst = reference::gemmlowp_matrix_multiply_core<int32_t, uint8_t>(ref_src0, ref_src1, TensorShape(N, M, B), offset_src0, offset_src1); |
Isabella Gottardi | 01a214a | 2018-04-09 16:00:52 +0100 | [diff] [blame] | 315 | |
| 316 | if(add_bias) |
| 317 | { |
| 318 | SimpleTensor<int32_t> biases{ TensorShape(N), DataType::S32, 1 }; |
| 319 | // Fill bias |
| 320 | fill(biases, 3); |
| 321 | ref_dst = reference::gemmlowp_quantize_down_int32_to_uint8_scale_by_fixedpoint<int32_t>(ref_tmp_dst, biases, dst_multiplier, dst_shift, offset_dst); |
| 322 | } |
| 323 | else |
| 324 | { |
| 325 | ref_dst = reference::gemmlowp_quantize_down_int32_to_uint8_scale_by_fixedpoint<int32_t>(ref_tmp_dst, dst_multiplier, dst_shift, offset_dst); |
| 326 | } |
| 327 | validate(CLAccessor(dst), ref_dst); |
| 328 | break; |
| 329 | } |
| 330 | default: |
| 331 | break; |
| 332 | } |
| 333 | } |
| 334 | void do_run() override |
| 335 | { |
| 336 | // Execute the function |
| 337 | if(data_type == DataType::QASYMM8) |
| 338 | { |
| 339 | // Run gemmlowp |
| 340 | mm_gemmlowp.run(); |
| 341 | // Run output stage |
| 342 | mm_gemmlowp_output_stage.run(); |
| 343 | } |
| 344 | else |
| 345 | { |
| 346 | // Run gemm |
| 347 | mm_gemm.run(); |
| 348 | } |
| 349 | |
| 350 | // Make sure all the OpenCL jobs are done executing: |
| 351 | CLScheduler::get().sync(); |
| 352 | } |
| 353 | |
| 354 | private: |
| 355 | template <typename U> |
| 356 | void fill(U &&tensor, int i) |
| 357 | { |
| 358 | switch(tensor.data_type()) |
| 359 | { |
| 360 | case DataType::F16: |
| 361 | case DataType::F32: |
| 362 | { |
| 363 | std::uniform_real_distribution<> distribution(-1.0f, 1.0f); |
| 364 | library->fill(tensor, distribution, i); |
| 365 | break; |
| 366 | } |
| 367 | case DataType::S32: |
| 368 | case DataType::QASYMM8: |
| 369 | { |
| 370 | std::uniform_int_distribution<> distribution(-6000, 6000); |
| 371 | library->fill(tensor, distribution, i); |
| 372 | break; |
| 373 | } |
| 374 | default: |
| 375 | library->fill_tensor_uniform(tensor, i); |
| 376 | } |
| 377 | } |
| 378 | |
Georgios Pinitas | 108a95e | 2019-03-27 13:55:59 +0000 | [diff] [blame] | 379 | void consume_params(const GEMMCommandLineOptions &opts) |
| 380 | { |
| 381 | ARM_COMPUTE_ERROR_ON(opts.M->value() <= 0); |
| 382 | ARM_COMPUTE_ERROR_ON(opts.N->value() <= 0); |
| 383 | ARM_COMPUTE_ERROR_ON(opts.K->value() <= 0); |
| 384 | ARM_COMPUTE_ERROR_ON(opts.B->value() <= 0); |
| 385 | M = opts.M->value(); |
| 386 | N = opts.N->value(); |
| 387 | K = opts.K->value(); |
| 388 | B = opts.B->value(); |
| 389 | alpha = opts.alpha->value(); |
| 390 | beta = opts.beta->value(); |
| 391 | offset_src0 = opts.offset_src0->value(); |
| 392 | offset_src1 = opts.offset_src1->value(); |
| 393 | offset_dst = opts.offset_dst->value(); |
| 394 | scale_src0 = opts.scale_src0->value(); |
| 395 | scale_src1 = opts.scale_src1->value(); |
| 396 | scale_dst = opts.scale_dst->value(); |
| 397 | add_bias = opts.add_bias->is_set() ? opts.add_bias->value() : true; |
| 398 | data_type = opts.data_type->value(); |
| 399 | } |
| 400 | |
Isabella Gottardi | 01a214a | 2018-04-09 16:00:52 +0100 | [diff] [blame] | 401 | CLTensor src0{}, src1{}, src2{}, dst{}; |
| 402 | CLTensor tmp_dst{}, biases{}; |
| 403 | |
| 404 | CLGEMM mm_gemm{}; |
| 405 | CLGEMMLowpMatrixMultiplyCore mm_gemmlowp{}; |
| 406 | CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint mm_gemmlowp_output_stage{}; |
| 407 | |
Georgios Pinitas | 108a95e | 2019-03-27 13:55:59 +0000 | [diff] [blame] | 408 | size_t M{ 7 }, N{ 3 }, K{ 5 }, B{ 1 }; |
Isabella Gottardi | 01a214a | 2018-04-09 16:00:52 +0100 | [diff] [blame] | 409 | DataType data_type{ DataType::F32 }; |
| 410 | float alpha{ 1.0 }, beta{ 0.0 }; |
| 411 | int offset_src0{ 10 }, offset_src1{ 10 }, offset_dst{ 10 }; |
| 412 | float scale_src0{ 1.0f / 255 }, scale_src1{ 1.0f / 255 }, scale_dst{ 1.0f / 255 }; |
| 413 | int32_t dst_multiplier{ 0 }, dst_shift{ 0 }; |
| 414 | bool add_bias{ true }; |
| 415 | }; |
| 416 | |
| 417 | /** Main program for gemm test |
| 418 | * |
| 419 | * @param[in] argc Number of arguments |
Georgios Pinitas | 108a95e | 2019-03-27 13:55:59 +0000 | [diff] [blame] | 420 | * @param[in] argv Arguments |
Isabella Gottardi | 01a214a | 2018-04-09 16:00:52 +0100 | [diff] [blame] | 421 | * |
| 422 | */ |
| 423 | int main(int argc, char **argv) |
| 424 | { |
| 425 | return utils::run_example<CLGEMMValidateExample>(argc, argv); |
| 426 | } |