SiCongLi | 282f324 | 2020-11-24 15:24:16 +0000 | [diff] [blame] | 1 | /* |
Manuel Bottini | 7b42786 | 2021-02-08 13:45:19 +0000 | [diff] [blame] | 2 | * Copyright (c) 2020-2021 Arm Limited. |
SiCongLi | 282f324 | 2020-11-24 15:24:16 +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 | #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 "CommonGemmExampleOptions.h" |
| 29 | #include "GemmTunerHelpers.h" |
| 30 | #include "arm_compute/core/Helpers.h" |
| 31 | #include "arm_compute/core/KernelDescriptors.h" |
| 32 | #include "arm_compute/core/KernelDescriptors.h" |
| 33 | #include "arm_compute/core/Types.h" |
| 34 | #include "arm_compute/core/utils/misc/ShapeCalculator.h" |
| 35 | #include "arm_compute/core/utils/quantization/AsymmHelpers.h" |
| 36 | #include "arm_compute/runtime/CL/CLScheduler.h" |
| 37 | #include "arm_compute/runtime/CL/CLTuner.h" |
Georgios Pinitas | 7891a73 | 2021-08-20 21:39:25 +0100 | [diff] [blame] | 38 | #include "src/gpu/cl/kernels/ClGemmLowpMatrixMultiplyReshapedOnlyRhsKernel.h" |
| 39 | #include "src/gpu/cl/kernels/ClGemmLowpReductionKernel.h" |
SiCongLi | 282f324 | 2020-11-24 15:24:16 +0000 | [diff] [blame] | 40 | #include "tests/CL/Helper.h" |
| 41 | #include "utils/Utils.h" |
| 42 | #include "utils/command_line/CommandLineOptions.h" |
| 43 | #include "utils/command_line/CommandLineParser.h" |
| 44 | |
| 45 | #include <cstdlib> |
| 46 | #include <memory> |
| 47 | |
| 48 | using namespace arm_compute; |
| 49 | using namespace utils; |
Georgios Pinitas | 4a578b9 | 2021-06-25 12:13:49 +0100 | [diff] [blame] | 50 | using namespace arm_compute::opencl::kernels; |
SiCongLi | 282f324 | 2020-11-24 15:24:16 +0000 | [diff] [blame] | 51 | using namespace arm_compute::misc::shape_calculator; |
| 52 | using namespace gemm_tuner; |
| 53 | |
| 54 | namespace |
| 55 | { |
| 56 | /** Structure holding all tunable gemm configs specific to this example/strategy */ |
| 57 | struct GemmConfigs |
| 58 | { |
| 59 | size_t m0{ 4 }; /**< Number of rows processed by the matrix multiplication */ |
| 60 | size_t n0{ 4 }; /**< Number of columns processed by the matrix multiplication */ |
| 61 | size_t k0{ 4 }; /**< Number of partial accumulations performed by the matrix multiplication */ |
| 62 | size_t h0{ 1 }; /**< Number of horizontal blocks of size (k0xn0) stored on the same output row */ |
| 63 | bool interleave_rhs{ true }; /**< Interleave rhs matrix */ |
| 64 | bool transpose_rhs{ true }; /**< Transpose rhs matrix */ |
| 65 | }; |
| 66 | |
| 67 | /** Formatted output of the GemmConfigs type |
| 68 | * |
| 69 | * @param[out] os Output stream. |
| 70 | * @param[in] configs Tunable configurations to output |
| 71 | * |
| 72 | * @return Modified output stream. |
| 73 | */ |
| 74 | ::std::ostream &operator<<(::std::ostream &os, const GemmConfigs &configs) |
| 75 | { |
| 76 | std::string false_str = std::string("false"); |
| 77 | std::string true_str = std::string("true"); |
| 78 | |
| 79 | os << "m0 : " << configs.m0 << std::endl; |
| 80 | os << "n0 : " << configs.n0 << std::endl; |
| 81 | os << "k0 : " << configs.k0 << std::endl; |
| 82 | os << "h0 : " << configs.h0 << std::endl; |
| 83 | os << "interleave_rhs : " << (configs.interleave_rhs ? true_str : false_str) << std::endl; |
| 84 | os << "transpose_rhs : " << (configs.transpose_rhs ? true_str : false_str) << std::endl; |
| 85 | return os; |
| 86 | } |
| 87 | |
| 88 | /** Command line options for gemm configs */ |
| 89 | class GemmConfigOptions |
| 90 | { |
| 91 | public: |
| 92 | /** Constructor |
| 93 | * |
| 94 | * @param[in,out] parser A parser on which "parse()" hasn't been called yet. |
| 95 | */ |
| 96 | GemmConfigOptions(CommandLineParser &parser) |
| 97 | : m0(parser.add_positional_option<SimpleOption<size_t>>("m0", 4)), |
| 98 | n0(parser.add_positional_option<SimpleOption<size_t>>("n0", 4)), |
| 99 | k0(parser.add_positional_option<SimpleOption<size_t>>("k0", 4)), |
| 100 | h0(parser.add_positional_option<SimpleOption<size_t>>("h0", 1)), |
| 101 | interleave_rhs(parser.add_positional_option<SimpleOption<size_t>>("interleave_rhs", 1)), |
| 102 | transpose_rhs(parser.add_positional_option<SimpleOption<size_t>>("transpose_rhs", 1)) |
| 103 | { |
| 104 | m0->set_help("Number of rows processed by the matrix multiplication"); |
| 105 | n0->set_help("Number of columns processed by the matrix multiplication"); |
| 106 | k0->set_help("Number of partial accumulations performed by the matrix multiplication"); |
| 107 | h0->set_help("Number of horizontal blocks of size (k0xn0) stored on the same output row"); |
| 108 | interleave_rhs->set_help("Interleave rhs matrix (1) / Do not interleave rhs matrix (0)"); |
| 109 | transpose_rhs->set_help("Transpose rhs matrix (1) / Do not transpose rhs matrix (0)"); |
| 110 | } |
| 111 | /** Prevent instances of this class from being copied (As this class contains pointers) */ |
| 112 | GemmConfigOptions(const GemmConfigOptions &) = delete; |
| 113 | /** Prevent instances of this class from being copied (As this class contains pointers) */ |
| 114 | GemmConfigOptions &operator=(const GemmConfigOptions &) = delete; |
| 115 | /** Allow instances of this class to be moved */ |
| 116 | GemmConfigOptions(GemmConfigOptions &&) = default; |
| 117 | /** Allow instances of this class to be moved */ |
| 118 | GemmConfigOptions &operator=(GemmConfigOptions &&) = default; |
| 119 | /** Default destructor */ |
| 120 | ~GemmConfigOptions() = default; |
| 121 | |
| 122 | SimpleOption<size_t> *m0; /**< Number of rows processed by the matrix multiplication option */ |
| 123 | SimpleOption<size_t> *n0; /**< Number of columns processed by the matrix multiplication option */ |
| 124 | SimpleOption<size_t> *k0; /**< Number of partial accumulations performed by the matrix multiplication option */ |
| 125 | SimpleOption<size_t> *h0; /**< Number of horizontal blocks of size (k0xn0) stored on the same output row option */ |
| 126 | SimpleOption<size_t> *interleave_rhs; /**< Interleave rhs matrix option (1 enable; 0 disable) */ |
| 127 | SimpleOption<size_t> *transpose_rhs; /**< Transpose rhs matrix option (1 enable; 0 disable) */ |
| 128 | }; |
| 129 | |
| 130 | /** Consumes the gemm configuration options and creates a structure containing all information |
| 131 | * |
| 132 | * @param[in] options Options to consume |
| 133 | * |
| 134 | * @return Structure containing the gemm configurations |
| 135 | */ |
| 136 | GemmConfigs consume_gemm_configs(const GemmConfigOptions &options) |
| 137 | { |
| 138 | GemmConfigs configs; |
| 139 | configs.m0 = options.m0->value(); |
| 140 | configs.n0 = options.n0->value(); |
| 141 | configs.k0 = options.k0->value(); |
| 142 | configs.h0 = options.h0->value(); |
| 143 | configs.interleave_rhs = options.interleave_rhs->value() != 0; |
| 144 | configs.transpose_rhs = options.transpose_rhs->value() != 0; |
| 145 | return configs; |
| 146 | } |
| 147 | |
| 148 | } // namespace |
| 149 | |
Georgios Pinitas | 4a578b9 | 2021-06-25 12:13:49 +0100 | [diff] [blame] | 150 | using ClGemmLowpMatrixMultiplyReshapedOnlyRhs = test::CLSynthetizeOperator<ClGemmLowpMatrixMultiplyReshapedOnlyRhsKernel>; |
| 151 | using ClGemmLowpMatrixAReduction = test::CLSynthetizeOperator<ClGemmLowpMatrixAReductionKernel>; |
SiCongLi | 282f324 | 2020-11-24 15:24:16 +0000 | [diff] [blame] | 152 | |
| 153 | class CLGEMMLowpMatrixMultiplyReshapedOnlyRHSFusedOutputStageFixedpointExample : public Example |
| 154 | { |
| 155 | public: |
| 156 | bool do_setup(int argc, char **argv) override |
| 157 | { |
| 158 | // Default parameters |
| 159 | CommonGemmExampleParams params; |
| 160 | GemmConfigs configs; |
| 161 | |
| 162 | // Parse command line options |
| 163 | CommandLineParser parser; |
SiCong Li | 98e33b9 | 2020-12-03 14:52:53 +0000 | [diff] [blame] | 164 | CommonGemmExampleOptions param_options(parser, DataType::QASYMM8); |
SiCongLi | 282f324 | 2020-11-24 15:24:16 +0000 | [diff] [blame] | 165 | GemmConfigOptions config_options(parser); |
| 166 | |
| 167 | parser.parse(argc, argv); |
| 168 | if(param_options.help->is_set() && param_options.help->value()) |
| 169 | { |
| 170 | parser.print_help(argv[0]); |
| 171 | return false; |
| 172 | } |
| 173 | if(!parser.validate()) |
| 174 | { |
| 175 | // Invalid arguments. Use default parameters and configs |
| 176 | std::cerr << "Invalid arguments." << std::endl; |
| 177 | parser.print_help(argv[0]); |
| 178 | std::cerr << "Falling back to default parameters and configs" << std::endl; |
| 179 | } |
| 180 | else |
| 181 | { |
| 182 | params = consume_common_gemm_example_parameters(param_options); |
| 183 | configs = consume_gemm_configs(config_options); |
| 184 | } |
| 185 | |
| 186 | std::cout << "Gemm parameters:" << std::endl; |
| 187 | std::cout << params << std::endl; |
| 188 | std::cout << "Gemm configurations:" << std::endl; |
| 189 | std::cout << configs << std::endl; |
| 190 | |
Gian Marco Iodice | ca419dd | 2021-03-03 17:25:07 +0000 | [diff] [blame] | 191 | tuner.set_tuner_mode(params.tuner_mode); |
| 192 | |
SiCongLi | 282f324 | 2020-11-24 15:24:16 +0000 | [diff] [blame] | 193 | CLScheduler::get().default_init(&tuner); |
| 194 | |
| 195 | lhs.allocator()->init(TensorInfo(TensorShape(params.K, params.M, params.B), 1, params.data_type)); |
| 196 | rhs.allocator()->init(TensorInfo(TensorShape(params.N, params.K, params.B), 1, params.data_type)); |
SiCongLi | eda87d4 | 2021-03-04 10:27:03 +0000 | [diff] [blame] | 197 | bias.allocator()->init(TensorInfo(TensorShape(params.N), 1, DataType::S32)); |
SiCongLi | 282f324 | 2020-11-24 15:24:16 +0000 | [diff] [blame] | 198 | dst.allocator()->init(TensorInfo(TensorShape(params.N, params.M, params.B), 1, params.data_type)); |
| 199 | |
| 200 | // Set arbitrary quantization information (non-zero offset to ensure offset contribution stage is included) |
| 201 | // Could be extended in the future to include a user-controlled option for offset == 0 |
SiCong Li | 98e33b9 | 2020-12-03 14:52:53 +0000 | [diff] [blame] | 202 | const QuantizationInfo q_info |
| 203 | { |
| 204 | 0.012, 3 |
| 205 | }; |
| 206 | lhs.info()->set_quantization_info(q_info); |
| 207 | rhs.info()->set_quantization_info(q_info); |
| 208 | bias.info()->set_quantization_info(q_info); |
| 209 | dst.info()->set_quantization_info(q_info); |
SiCongLi | 282f324 | 2020-11-24 15:24:16 +0000 | [diff] [blame] | 210 | |
| 211 | GEMMLHSMatrixInfo lhs_info; |
| 212 | lhs_info.m0 = configs.m0; |
| 213 | lhs_info.k0 = configs.k0; |
| 214 | |
| 215 | GEMMRHSMatrixInfo rhs_info; |
| 216 | rhs_info.n0 = configs.n0; |
| 217 | rhs_info.k0 = configs.k0; |
| 218 | rhs_info.h0 = configs.h0; |
| 219 | rhs_info.interleave = configs.interleave_rhs; |
| 220 | rhs_info.transpose = configs.transpose_rhs; |
| 221 | rhs_info.export_to_cl_image = false; // CL image not supported for quantized cases yet |
| 222 | |
| 223 | rhs_reshaped.allocator()->init(TensorInfo(compute_rhs_reshaped_shape(*rhs.info(), rhs_info), 1, params.data_type)); |
SiCong Li | 98e33b9 | 2020-12-03 14:52:53 +0000 | [diff] [blame] | 224 | rhs_reshaped.info()->set_quantization_info(q_info); |
SiCongLi | 282f324 | 2020-11-24 15:24:16 +0000 | [diff] [blame] | 225 | if(rhs_info.export_to_cl_image) |
| 226 | { |
Manuel Bottini | 7b42786 | 2021-02-08 13:45:19 +0000 | [diff] [blame] | 227 | if(!examples::gemm_tuner_helpers::update_padding_for_cl_image(rhs_reshaped.info())) |
| 228 | { |
| 229 | std::cerr << "cl_image is not supported on the device, disable export_to_cl_image" << std::endl; |
| 230 | return false; |
| 231 | } |
SiCongLi | 282f324 | 2020-11-24 15:24:16 +0000 | [diff] [blame] | 232 | } |
| 233 | |
| 234 | // Configure output stage for quantized case |
| 235 | GEMMLowpOutputStageInfo gemmlowp_output_stage; |
| 236 | gemmlowp_output_stage.type = GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT; |
| 237 | gemmlowp_output_stage.output_data_type = dst.info()->data_type(); |
| 238 | gemmlowp_output_stage.gemmlowp_offset = 0; |
| 239 | { |
SiCongLi | 282f324 | 2020-11-24 15:24:16 +0000 | [diff] [blame] | 240 | gemmlowp_output_stage.is_quantized_per_channel = false; |
| 241 | // Num_filters is 1 unless quantized type is of per_channel type. Could be extended in the future to support per-channel quantization. |
| 242 | const unsigned int num_filters = 1; |
| 243 | |
| 244 | dst_multipliers.allocator()->init(TensorInfo(TensorShape(num_filters), 1, DataType::S32)); |
| 245 | dst_shifts.allocator()->init(TensorInfo(TensorShape(num_filters), 1, DataType::S32)); |
| 246 | |
| 247 | gemmlowp_output_stage.gemmlowp_multipliers.resize(num_filters); |
| 248 | gemmlowp_output_stage.gemmlowp_shifts.resize(num_filters); |
| 249 | quantization::compute_quantized_multipliers_and_shifts(lhs.info(), |
| 250 | rhs.info(), |
| 251 | dst.info(), |
SiCongLi | 282f324 | 2020-11-24 15:24:16 +0000 | [diff] [blame] | 252 | gemmlowp_output_stage.gemmlowp_multipliers.data(), |
| 253 | gemmlowp_output_stage.gemmlowp_shifts.data()); |
| 254 | gemmlowp_output_stage.gemmlowp_multiplier = gemmlowp_output_stage.gemmlowp_multipliers[0]; |
| 255 | gemmlowp_output_stage.gemmlowp_shift = gemmlowp_output_stage.gemmlowp_shifts[0]; |
| 256 | |
| 257 | // No fused activation |
| 258 | PixelValue min_val{}; |
| 259 | PixelValue max_val{}; |
| 260 | std::tie(min_val, max_val) = get_min_max(dst.info()->data_type()); |
| 261 | |
| 262 | auto min_activation = min_val.get<int32_t>(); |
| 263 | auto max_activation = max_val.get<int32_t>(); |
| 264 | |
| 265 | // Set the GEMMLowp output stage info |
| 266 | gemmlowp_output_stage.gemmlowp_offset = dst.info()->quantization_info().uniform().offset; |
| 267 | gemmlowp_output_stage.gemmlowp_min_bound = min_activation; |
| 268 | gemmlowp_output_stage.gemmlowp_max_bound = max_activation; |
| 269 | } |
| 270 | |
| 271 | GEMMKernelInfo gemm_info; |
| 272 | gemm_info.m = params.M; |
| 273 | gemm_info.n = params.N; |
| 274 | gemm_info.k = params.K; |
| 275 | gemm_info.depth_output_gemm3d = 0; |
| 276 | gemm_info.reinterpret_input_as_3d = false; |
| 277 | gemm_info.broadcast_bias = true; |
| 278 | gemm_info.fp_mixed_precision = false; |
| 279 | gemm_info.has_pad_y = false; |
| 280 | gemm_info.mult_transpose1xW_width = configs.h0; |
| 281 | gemm_info.lhs_info = lhs_info; |
| 282 | gemm_info.rhs_info = rhs_info; |
| 283 | gemm_info.a_offset = lhs.info()->quantization_info().uniform().offset; |
| 284 | gemm_info.b_offset = rhs.info()->quantization_info().uniform().offset; |
| 285 | gemm_info.output_stage = gemmlowp_output_stage; |
| 286 | |
| 287 | // Initialize Matrix A reduction kernel only if _b_offset is not equal to 0 |
| 288 | if(gemm_info.b_offset != 0) |
| 289 | { |
| 290 | const TensorInfo info_vector_sum_row(compute_reductionB_shape(*lhs.info()), 1, DataType::S32); |
| 291 | vector_sum_row.allocator()->init(info_vector_sum_row); |
| 292 | |
Georgios Pinitas | 4a578b9 | 2021-06-25 12:13:49 +0100 | [diff] [blame] | 293 | mtx_a_reduction = std::make_unique<ClGemmLowpMatrixAReduction>(); |
SiCongLi | 282f324 | 2020-11-24 15:24:16 +0000 | [diff] [blame] | 294 | |
| 295 | if(!mtx_a_reduction->validate(lhs.info(), vector_sum_row.info(), GEMMLowpReductionKernelInfo{})) |
| 296 | { |
| 297 | std::cerr << "Invalid arguments for CLGEMMLowpMatrixAReductionKernel." << std::endl; |
| 298 | return false; |
| 299 | } |
| 300 | |
Georgios Pinitas | 4a578b9 | 2021-06-25 12:13:49 +0100 | [diff] [blame] | 301 | mtx_a_reduction->configure(lhs.info(), vector_sum_row.info(), GEMMLowpReductionKernelInfo{}); |
SiCongLi | 282f324 | 2020-11-24 15:24:16 +0000 | [diff] [blame] | 302 | } |
| 303 | // Initialize matrix B reduction kernel only if _a_offset is not equal to 0 |
| 304 | if(gemm_info.a_offset != 0) |
| 305 | { |
| 306 | const TensorInfo info_vector_sum_col(compute_reductionA_shape(*rhs.info()), 1, DataType::S32); |
| 307 | vector_sum_col.allocator()->init(info_vector_sum_col); |
| 308 | // There's no need for a Matrix B reduction kernel as this is assumed to be run only once in the prepare stage |
| 309 | } |
| 310 | |
| 311 | // Validate argments |
| 312 | if(!gemm.validate(lhs.info(), rhs_reshaped.info(), dst.info(), gemm_info, gemm_info.a_offset == 0 ? nullptr : vector_sum_col.info(), |
| 313 | gemm_info.b_offset == 0 ? nullptr : vector_sum_row.info(), bias.info(), dst_multipliers.info(), dst_shifts.info())) |
| 314 | { |
Georgios Pinitas | 4a578b9 | 2021-06-25 12:13:49 +0100 | [diff] [blame] | 315 | std::cerr << "Invalid arguments for ClGemmLowpMatrixMultiplyReshapedOnlyRhsKernel." << std::endl; |
SiCongLi | 282f324 | 2020-11-24 15:24:16 +0000 | [diff] [blame] | 316 | return false; |
| 317 | } |
| 318 | |
| 319 | // Configure function |
Georgios Pinitas | 4a578b9 | 2021-06-25 12:13:49 +0100 | [diff] [blame] | 320 | gemm.configure(lhs.info(), rhs_reshaped.info(), dst.info(), gemm_info, |
| 321 | gemm_info.a_offset == 0 ? nullptr : vector_sum_col.info(), gemm_info.b_offset == 0 ? nullptr : vector_sum_row.info(), |
| 322 | bias.info(), dst_multipliers.info(), dst_shifts.info()); |
SiCongLi | 282f324 | 2020-11-24 15:24:16 +0000 | [diff] [blame] | 323 | |
| 324 | // Allocate tensors |
| 325 | lhs.allocator()->allocate(); |
| 326 | rhs.allocator()->allocate(); |
| 327 | rhs_reshaped.allocator()->allocate(); |
| 328 | bias.allocator()->allocate(); |
| 329 | dst.allocator()->allocate(); |
| 330 | vector_sum_col.allocator()->allocate(); |
| 331 | vector_sum_row.allocator()->allocate(); |
| 332 | dst_multipliers.allocator()->allocate(); |
| 333 | dst_shifts.allocator()->allocate(); |
| 334 | |
| 335 | return true; |
| 336 | } |
| 337 | void do_run() override |
| 338 | { |
| 339 | if(mtx_a_reduction != nullptr) |
| 340 | { |
Georgios Pinitas | 4a578b9 | 2021-06-25 12:13:49 +0100 | [diff] [blame] | 341 | ITensorPack red_pack({ { ACL_SRC, &lhs }, { ACL_DST, &dst } }); |
| 342 | mtx_a_reduction->run(red_pack); |
SiCongLi | 282f324 | 2020-11-24 15:24:16 +0000 | [diff] [blame] | 343 | } |
Georgios Pinitas | 4a578b9 | 2021-06-25 12:13:49 +0100 | [diff] [blame] | 344 | |
| 345 | ITensorPack gemm_pack({ { ACL_SRC_0, &lhs }, { ACL_SRC_1, &rhs }, { ACL_BIAS, &bias }, { ACL_VEC_COL_SUM, &vector_sum_col }, { ACL_VEC_ROW_SUM, &vector_sum_row }, { ACL_SHIFTS, &dst_shifts }, { ACL_MULTIPLIERS, &dst_multipliers }, { ACL_DST, &dst } }); |
| 346 | gemm.run(gemm_pack); |
SiCongLi | 282f324 | 2020-11-24 15:24:16 +0000 | [diff] [blame] | 347 | |
| 348 | // Make sure all the OpenCL jobs are done executing: |
| 349 | CLScheduler::get().sync(); |
| 350 | } |
| 351 | |
| 352 | void do_teardown() override |
| 353 | { |
| 354 | } |
| 355 | |
| 356 | private: |
| 357 | CLTensor lhs{}; |
| 358 | CLTensor rhs{}; |
| 359 | CLTensor rhs_reshaped{}; |
| 360 | CLTensor bias{}; |
| 361 | CLTensor dst{}; |
| 362 | CLTensor vector_sum_col{}; |
| 363 | CLTensor vector_sum_row{}; |
| 364 | CLTensor dst_multipliers{}; |
| 365 | CLTensor dst_shifts{}; |
| 366 | CLTuner tuner{}; |
Georgios Pinitas | 4a578b9 | 2021-06-25 12:13:49 +0100 | [diff] [blame] | 367 | ClGemmLowpMatrixMultiplyReshapedOnlyRhs gemm{}; |
| 368 | std::unique_ptr<ClGemmLowpMatrixAReduction> mtx_a_reduction{ nullptr }; |
SiCongLi | 282f324 | 2020-11-24 15:24:16 +0000 | [diff] [blame] | 369 | }; |
| 370 | |
| 371 | /** Main test program for gemmlowp reshaped rhs only with fused output stage fixedpoint |
| 372 | * |
| 373 | * @param[in] argc Number of arguments |
| 374 | * @param[in] argv Arguments ( [optional] M, [optional] N, [optional] K, [optional] B, [optional] m0, [optional] n0, [optional] k0, [optional] h0, [optional] interleave_rhs, [optional] transpose_rhs ) |
| 375 | */ |
| 376 | int main(int argc, char **argv) |
| 377 | { |
| 378 | return run_example<CLGEMMLowpMatrixMultiplyReshapedOnlyRHSFusedOutputStageFixedpointExample>(argc, argv); |
| 379 | } |