Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2019 ARM Limited. |
| 3 | * |
| 4 | * SPDX-License-Identifier: MIT |
| 5 | * |
| 6 | * Permission is hereby granted, free of charge, to any person obtaining a copy |
| 7 | * of this software and associated documentation files (the "Software"), to |
| 8 | * deal in the Software without restriction, including without limitation the |
| 9 | * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or |
| 10 | * sell copies of the Software, and to permit persons to whom the Software is |
| 11 | * furnished to do so, subject to the following conditions: |
| 12 | * |
| 13 | * The above copyright notice and this permission notice shall be included in all |
| 14 | * copies or substantial portions of the Software. |
| 15 | * |
| 16 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR |
| 17 | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
| 18 | * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE |
| 19 | * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER |
| 20 | * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, |
| 21 | * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE |
| 22 | * SOFTWARE. |
| 23 | */ |
| 24 | #include "arm_compute/core/CL/kernels/CLGEMMMatrixMultiplyKernel.h" |
| 25 | #include "arm_compute/core/KernelDescriptors.h" |
| 26 | #include "arm_compute/core/Types.h" |
| 27 | #include "arm_compute/core/utils/misc/ShapeCalculator.h" |
| 28 | #include "arm_compute/runtime/CL/CLTensor.h" |
| 29 | #include "arm_compute/runtime/CL/CLTensorAllocator.h" |
| 30 | #include "tests/CL/CLAccessor.h" |
| 31 | #include "tests/CL/Helper.h" |
| 32 | #include "tests/PaddingCalculator.h" |
| 33 | #include "tests/datasets/ShapeDatasets.h" |
| 34 | #include "tests/framework/Asserts.h" |
| 35 | #include "tests/framework/Macros.h" |
| 36 | #include "tests/framework/datasets/Datasets.h" |
| 37 | #include "tests/validation/Validation.h" |
| 38 | #include "tests/validation/fixtures/GEMMFixture.h" |
| 39 | |
| 40 | namespace arm_compute |
| 41 | { |
| 42 | namespace test |
| 43 | { |
| 44 | namespace validation |
| 45 | { |
| 46 | using namespace arm_compute::misc::shape_calculator; |
| 47 | |
| 48 | // Create function for CLGEMMMatrixMultiplyKernel |
| 49 | using CLGEMMMatrixMultiplyNative = CLSynthetizeFunction<CLGEMMMatrixMultiplyKernel>; |
| 50 | |
| 51 | // Fixture for GEMMMatrixMultiplyValidationFixture |
| 52 | template <typename T> |
| 53 | using CLGEMMMatrixMultiplyNativeFixture = GEMMMatrixMultiplyValidationFixture<CLTensor, CLAccessor, T, CLGEMMMatrixMultiplyNative>; |
| 54 | |
| 55 | // Fixture for GEMMMatrixMultiply3DValidationFixture |
| 56 | template <typename T> |
| 57 | using CLGEMMMatrixMultiplyNative3DFixture = GEMMMatrixMultiply3DValidationFixture<CLTensor, CLAccessor, T, CLGEMMMatrixMultiplyNative>; |
| 58 | |
| 59 | namespace |
| 60 | { |
| 61 | // *INDENT-OFF* |
| 62 | // clang-format off |
| 63 | RelativeTolerance<float> rel_tolerance_f32(0.001f); |
| 64 | constexpr float abs_tolerance_f32(0.0001f); |
| 65 | |
| 66 | RelativeTolerance<half> rel_tolerance_f16(half(0.2)); |
| 67 | constexpr float tolerance_num_f16 = 0.02f; |
| 68 | |
| 69 | /** Alpha values to test - Precommit */ |
Gian Marco Iodice | f3622be | 2019-07-29 14:27:16 +0100 | [diff] [blame] | 70 | const auto alpha_values = framework::dataset::make("alpha", {1.0f, -0.75f} ); |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 71 | |
| 72 | /** Beta values to test - Precommit */ |
| 73 | const auto beta_values = framework::dataset::make("beta", {-0.75f, 0.0f} ); |
| 74 | |
| 75 | /** M values to test - Precommit */ |
| 76 | const auto m_values_precommit = framework::dataset::make("M", {37, 1}); |
| 77 | |
| 78 | /** N values to test - Precommit */ |
| 79 | const auto n_values_precommit = framework::dataset::make("N", 51); |
| 80 | |
| 81 | /** K values to test - Precommit */ |
| 82 | const auto k_values_precommit = framework::dataset::make("K", 23); |
| 83 | |
| 84 | /** M values to test - Nightly */ |
| 85 | const auto m_values_nightly = framework::dataset::make("M", {421, 1}); |
| 86 | |
| 87 | /** N values to test - Nightly */ |
| 88 | const auto n_values_nightly = framework::dataset::make("N", {323, 1103}); |
| 89 | |
| 90 | /** K values to test - Nightly */ |
| 91 | const auto k_values_nightly = framework::dataset::make("K", 207); |
| 92 | |
| 93 | /** M_W values to test - Precommit */ |
| 94 | const auto m_w_values_precommit = framework::dataset::make("M_W", 5); |
| 95 | |
| 96 | /** M_H values to test - Precommit */ |
| 97 | const auto m_h_values_precommit = framework::dataset::make("M_H", 7); |
| 98 | |
| 99 | /** M_W values to test - Nightly */ |
| 100 | const auto m_w_values_nightly = framework::dataset::make("M_W", 13); |
| 101 | |
| 102 | /** M_H values to test - Nightly */ |
| 103 | const auto m_h_values_nightly = framework::dataset::make("M_H", 27); |
| 104 | |
| 105 | /** Batch size values to test */ |
| 106 | const auto b_values = framework::dataset::make("batch_size", 1, 3); |
| 107 | |
| 108 | /** Activation values to test */ |
| 109 | const auto act_values = framework::dataset::make("Activation", |
| 110 | { |
| 111 | ActivationLayerInfo(), |
| 112 | ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 8.f, 2.f), |
| 113 | }); |
| 114 | |
| 115 | /** Broadcast bias from vector to matrix */ |
| 116 | const auto broadcast_bias_values = framework::dataset::make("broadcast_bias", {false, true} ); |
| 117 | |
| 118 | /** GPU architectures values to test */ |
| 119 | const auto gpu_arch_values = framework::dataset::make("GPUArch", |
| 120 | { |
| 121 | GPUTarget::MIDGARD, |
| 122 | GPUTarget::BIFROST |
| 123 | }); |
| 124 | |
| 125 | /** Data types values to test in the configuration */ |
| 126 | const auto data_type_values = framework::dataset::make("DataType", |
| 127 | { |
| 128 | DataType::F32, |
| 129 | DataType::F16 |
| 130 | }); |
| 131 | |
| 132 | /** M values to test */ |
| 133 | const auto fp16_mixed_precision_values = framework::dataset::make("fp16_mixed_precision", {true, false}); |
| 134 | |
| 135 | /** Configuration test */ |
| 136 | void validate_configuration(unsigned int m_value, unsigned int n_value, unsigned int k_value, unsigned int b_value, bool broadcast_bias, bool fp16_mixed_precision, const ActivationLayerInfo &act_info, DataType data_type, GPUTarget gpu_arch_value) |
| 137 | { |
| 138 | GEMMReshapeInfo reshape_info(m_value, n_value, k_value, 1, 1, 0, false, broadcast_bias); |
| 139 | |
| 140 | const TensorShape lhs_shape(k_value, m_value, b_value); |
| 141 | const TensorShape rhs_shape(n_value, k_value, b_value); |
| 142 | |
| 143 | const TensorShape dst_shape = compute_mm_shape(TensorInfo(lhs_shape, 1, data_type), |
| 144 | TensorInfo(rhs_shape, 1, data_type), |
| 145 | reshape_info); |
| 146 | |
| 147 | const TensorShape bias_shape(n_value, |
| 148 | broadcast_bias? 1 : m_value, |
| 149 | broadcast_bias? 1 : b_value); |
| 150 | |
| 151 | // Create tensors |
| 152 | CLTensor lhs = create_tensor<CLTensor>(lhs_shape, data_type); |
| 153 | CLTensor rhs = create_tensor<CLTensor>(rhs_shape, data_type); |
| 154 | CLTensor bias = create_tensor<CLTensor>(bias_shape, data_type); |
| 155 | CLTensor dst = create_tensor<CLTensor>(dst_shape, data_type); |
| 156 | |
| 157 | ARM_COMPUTE_EXPECT(lhs.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 158 | ARM_COMPUTE_EXPECT(rhs.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 159 | ARM_COMPUTE_EXPECT(bias.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 160 | ARM_COMPUTE_EXPECT(dst.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 161 | |
| 162 | // Create and configure function |
| 163 | CLGEMMMatrixMultiplyNative gemm; |
| 164 | gemm.configure(gpu_arch_value, &lhs, &rhs, &bias, &dst, 1.0f, 2.0f, false, reshape_info, fp16_mixed_precision, act_info); |
| 165 | } |
| 166 | } // namespace |
| 167 | |
| 168 | TEST_SUITE(CL) |
| 169 | TEST_SUITE(GEMMMatrixMultiply) |
| 170 | TEST_SUITE(Float) |
| 171 | TEST_SUITE(FP32) |
| 172 | DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(combine(combine( |
| 173 | m_values_precommit, |
| 174 | n_values_precommit), |
| 175 | k_values_precommit), |
| 176 | framework::dataset::make("batch_size", 1)), |
| 177 | broadcast_bias_values), |
| 178 | framework::dataset::make("fp16_mixed_precision", false)), |
| 179 | act_values), |
| 180 | data_type_values), |
| 181 | gpu_arch_values), |
| 182 | m_value, n_value, k_value, b_value, broadcast_bias, fp16_mixed_precision_value, act_value, data_type_value, gpu_arch_value) |
| 183 | { |
| 184 | validate_configuration(m_value, n_value, k_value, b_value, broadcast_bias, fp16_mixed_precision_value, act_value, data_type_value, gpu_arch_value); |
| 185 | } |
| 186 | |
| 187 | FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMMatrixMultiplyNativeFixture<float>, framework::DatasetMode::ALL, |
| 188 | combine(combine(combine(combine(combine(combine(combine(combine(combine(combine( |
| 189 | m_values_precommit, |
| 190 | n_values_precommit), |
| 191 | k_values_precommit), |
| 192 | b_values), |
| 193 | alpha_values), |
| 194 | beta_values), |
| 195 | broadcast_bias_values), |
| 196 | framework::dataset::make("fp16_mixed_precision", false)), |
| 197 | act_values), |
| 198 | framework::dataset::make("DataType", DataType::F32)), |
| 199 | gpu_arch_values)) |
| 200 | { |
| 201 | // Validate output |
| 202 | validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32); |
| 203 | } |
| 204 | |
| 205 | FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMMatrixMultiplyNativeFixture<float>, framework::DatasetMode::NIGHTLY, |
| 206 | combine(combine(combine(combine(combine(combine(combine(combine(combine(combine( |
| 207 | m_values_nightly, |
| 208 | n_values_nightly), |
| 209 | k_values_nightly), |
| 210 | b_values), |
| 211 | alpha_values), |
| 212 | beta_values), |
| 213 | broadcast_bias_values), |
| 214 | framework::dataset::make("fp16_mixed_precision", false)), |
| 215 | act_values), |
| 216 | framework::dataset::make("DataType", DataType::F32)), |
| 217 | gpu_arch_values)) |
| 218 | { |
| 219 | // Validate output |
| 220 | validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32); |
| 221 | } |
| 222 | |
| 223 | FIXTURE_DATA_TEST_CASE(RunSmall3D, CLGEMMMatrixMultiplyNative3DFixture<float>, framework::DatasetMode::ALL, |
| 224 | combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine( |
| 225 | m_w_values_precommit, |
| 226 | m_h_values_precommit), |
| 227 | n_values_precommit), |
| 228 | k_values_precommit), |
| 229 | b_values), |
| 230 | alpha_values), |
| 231 | beta_values), |
| 232 | broadcast_bias_values), |
| 233 | framework::dataset::make("fp16_mixed_precision", false)), |
| 234 | act_values), |
| 235 | framework::dataset::make("DataType", DataType::F32)), |
| 236 | gpu_arch_values)) |
| 237 | { |
| 238 | // Validate output |
| 239 | validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32); |
| 240 | } |
| 241 | |
| 242 | FIXTURE_DATA_TEST_CASE(RunLarge3D, CLGEMMMatrixMultiplyNative3DFixture<float>, framework::DatasetMode::NIGHTLY, |
| 243 | combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine( |
| 244 | m_w_values_nightly, |
| 245 | m_h_values_nightly), |
| 246 | n_values_nightly), |
| 247 | k_values_nightly), |
| 248 | b_values), |
| 249 | alpha_values), |
| 250 | beta_values), |
| 251 | broadcast_bias_values), |
| 252 | framework::dataset::make("fp16_mixed_precision", false)), |
| 253 | act_values), |
| 254 | framework::dataset::make("DataType", DataType::F32)), |
| 255 | gpu_arch_values)) |
| 256 | { |
| 257 | // Validate output |
| 258 | validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32); |
| 259 | } |
| 260 | |
| 261 | TEST_SUITE_END() // FP32 |
| 262 | |
| 263 | TEST_SUITE(FP16) |
| 264 | FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMMatrixMultiplyNativeFixture<half>, framework::DatasetMode::ALL, |
| 265 | combine(combine(combine(combine(combine(combine(combine(combine(combine(combine( |
| 266 | m_values_precommit, |
| 267 | n_values_precommit), |
| 268 | k_values_precommit), |
| 269 | b_values), |
| 270 | alpha_values), |
| 271 | beta_values), |
| 272 | broadcast_bias_values), |
| 273 | fp16_mixed_precision_values), |
| 274 | act_values), |
| 275 | framework::dataset::make("DataType", DataType::F16)), |
| 276 | gpu_arch_values)) |
| 277 | { |
| 278 | // Validate output |
| 279 | validate(CLAccessor(_target), _reference, rel_tolerance_f16, tolerance_num_f16); |
| 280 | } |
| 281 | |
| 282 | FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMMatrixMultiplyNativeFixture<half>, framework::DatasetMode::NIGHTLY, |
| 283 | combine(combine(combine(combine(combine(combine(combine(combine(combine(combine( |
| 284 | m_values_nightly, |
| 285 | n_values_nightly), |
| 286 | k_values_nightly), |
| 287 | b_values), |
| 288 | alpha_values), |
| 289 | beta_values), |
| 290 | broadcast_bias_values), |
| 291 | fp16_mixed_precision_values), |
| 292 | act_values), |
| 293 | framework::dataset::make("DataType", DataType::F16)), |
| 294 | gpu_arch_values)) |
| 295 | { |
| 296 | // Validate output |
| 297 | validate(CLAccessor(_target), _reference, rel_tolerance_f16, tolerance_num_f16); |
| 298 | } |
| 299 | |
| 300 | FIXTURE_DATA_TEST_CASE(RunSmall3D, CLGEMMMatrixMultiplyNative3DFixture<half>, framework::DatasetMode::ALL, |
| 301 | combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine( |
| 302 | m_w_values_precommit, |
| 303 | m_h_values_precommit), |
| 304 | n_values_precommit), |
| 305 | k_values_precommit), |
| 306 | b_values), |
| 307 | alpha_values), |
| 308 | beta_values), |
| 309 | broadcast_bias_values), |
| 310 | fp16_mixed_precision_values), |
| 311 | act_values), |
| 312 | framework::dataset::make("DataType", DataType::F16)), |
| 313 | gpu_arch_values)) |
| 314 | { |
| 315 | // Validate output |
| 316 | validate(CLAccessor(_target), _reference, rel_tolerance_f16, tolerance_num_f16); |
| 317 | } |
| 318 | |
| 319 | FIXTURE_DATA_TEST_CASE(RunLarge3D, CLGEMMMatrixMultiplyNative3DFixture<half>, framework::DatasetMode::NIGHTLY, |
| 320 | combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine( |
| 321 | m_w_values_nightly, |
| 322 | m_h_values_nightly), |
| 323 | n_values_nightly), |
| 324 | k_values_nightly), |
| 325 | b_values), |
| 326 | alpha_values), |
| 327 | beta_values), |
| 328 | broadcast_bias_values), |
| 329 | fp16_mixed_precision_values), |
| 330 | act_values), |
| 331 | framework::dataset::make("DataType", DataType::F16)), |
| 332 | gpu_arch_values)) |
| 333 | { |
| 334 | // Validate output |
| 335 | validate(CLAccessor(_target), _reference, rel_tolerance_f16, tolerance_num_f16); |
| 336 | } |
| 337 | |
| 338 | TEST_SUITE_END() // FP16 |
| 339 | TEST_SUITE_END() // Float |
| 340 | TEST_SUITE_END() // GEMMMatrixMuliplty |
| 341 | TEST_SUITE_END() // CL |
| 342 | } // namespace validation |
| 343 | } // namespace test |
| 344 | } // namespace arm_compute |