Moritz Pflanzer | 4dfc235 | 2017-08-02 14:51:36 +0100 | [diff] [blame] | 1 | /* |
Michele Di Giorgio | d9eaf61 | 2020-07-08 11:12:57 +0100 | [diff] [blame] | 2 | * Copyright (c) 2017-2020 Arm Limited. |
Moritz Pflanzer | 4dfc235 | 2017-08-02 14:51:36 +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_TEST_GEMM_FIXTURE |
| 25 | #define ARM_COMPUTE_TEST_GEMM_FIXTURE |
| 26 | |
Gian Marco Iodice | 7026b30 | 2019-06-26 17:18:11 +0100 | [diff] [blame] | 27 | #include "arm_compute/core/KernelDescriptors.h" |
Moritz Pflanzer | 4dfc235 | 2017-08-02 14:51:36 +0100 | [diff] [blame] | 28 | #include "arm_compute/core/TensorShape.h" |
| 29 | #include "arm_compute/core/Types.h" |
Moritz Pflanzer | 4dfc235 | 2017-08-02 14:51:36 +0100 | [diff] [blame] | 30 | #include "tests/AssetsLibrary.h" |
| 31 | #include "tests/Globals.h" |
| 32 | #include "tests/IAccessor.h" |
Moritz Pflanzer | a09de0c | 2017-09-01 20:41:12 +0100 | [diff] [blame] | 33 | #include "tests/framework/Asserts.h" |
| 34 | #include "tests/framework/Fixture.h" |
Moritz Pflanzer | a09de0c | 2017-09-01 20:41:12 +0100 | [diff] [blame] | 35 | #include "tests/validation/Helpers.h" |
Gian Marco Iodice | ca1f460 | 2019-07-16 15:46:48 +0100 | [diff] [blame] | 36 | #include "tests/validation/reference/ActivationLayer.h" |
Georgios Pinitas | 5a7e776 | 2017-12-01 16:27:29 +0000 | [diff] [blame] | 37 | #include "tests/validation/reference/GEMM.h" |
Moritz Pflanzer | 4dfc235 | 2017-08-02 14:51:36 +0100 | [diff] [blame] | 38 | |
| 39 | #include <random> |
| 40 | |
| 41 | namespace arm_compute |
| 42 | { |
| 43 | namespace test |
| 44 | { |
| 45 | namespace validation |
| 46 | { |
Gian Marco Iodice | f3622be | 2019-07-29 14:27:16 +0100 | [diff] [blame] | 47 | template <typename TensorType, typename AccessorType, typename FunctionType, typename T, bool disable_c = false, bool reinterpret_input_as_3d = false, bool reinterpret_output_as_3d = false> |
Gian Marco Iodice | 68a3f56 | 2018-07-26 11:44:03 +0100 | [diff] [blame] | 48 | class GEMMValidationFixture : public framework::Fixture |
Moritz Pflanzer | 4dfc235 | 2017-08-02 14:51:36 +0100 | [diff] [blame] | 49 | { |
| 50 | public: |
| 51 | template <typename...> |
Pablo Tello | 0e37b5c | 2018-10-30 11:18:37 +0000 | [diff] [blame] | 52 | void setup(TensorShape shape_a, TensorShape shape_b, TensorShape shape_c, TensorShape output_shape, float alpha, float beta, bool pretranspose, DataType data_type) |
Moritz Pflanzer | 4dfc235 | 2017-08-02 14:51:36 +0100 | [diff] [blame] | 53 | { |
Michalis Spyrou | 6bff195 | 2019-10-02 17:22:11 +0100 | [diff] [blame] | 54 | ARM_COMPUTE_UNUSED(pretranspose); |
| 55 | _target = compute_target(shape_a, shape_b, shape_c, output_shape, alpha, beta, data_type); |
| 56 | _reference = compute_reference(shape_a, shape_b, output_shape, alpha, beta, data_type); |
Moritz Pflanzer | 4dfc235 | 2017-08-02 14:51:36 +0100 | [diff] [blame] | 57 | } |
| 58 | |
| 59 | protected: |
| 60 | template <typename U> |
Pablo Tello | 0e37b5c | 2018-10-30 11:18:37 +0000 | [diff] [blame] | 61 | void fill(U &&tensor, int i, float lo = -1.f, float hi = 1.f) |
Moritz Pflanzer | 4dfc235 | 2017-08-02 14:51:36 +0100 | [diff] [blame] | 62 | { |
| 63 | switch(tensor.data_type()) |
| 64 | { |
| 65 | case DataType::F16: |
| 66 | case DataType::F32: |
| 67 | { |
Pablo Tello | 0e37b5c | 2018-10-30 11:18:37 +0000 | [diff] [blame] | 68 | std::uniform_real_distribution<> distribution(lo, hi); |
Moritz Pflanzer | 4dfc235 | 2017-08-02 14:51:36 +0100 | [diff] [blame] | 69 | library->fill(tensor, distribution, i); |
| 70 | break; |
| 71 | } |
| 72 | default: |
| 73 | library->fill_tensor_uniform(tensor, i); |
| 74 | } |
| 75 | } |
| 76 | |
| 77 | TensorType compute_target(const TensorShape &shape_a, const TensorShape &shape_b, const TensorShape &shape_c, const TensorShape &output_shape, float alpha, float beta, |
Michalis Spyrou | 6bff195 | 2019-10-02 17:22:11 +0100 | [diff] [blame] | 78 | DataType data_type) |
Moritz Pflanzer | 4dfc235 | 2017-08-02 14:51:36 +0100 | [diff] [blame] | 79 | { |
| 80 | // Create tensors |
Vidhya Sudhan Loganathan | 014333d | 2018-07-02 09:13:49 +0100 | [diff] [blame] | 81 | TensorType a = create_tensor<TensorType>(shape_a, data_type, 1); |
| 82 | TensorType b = create_tensor<TensorType>(shape_b, data_type, 1); |
| 83 | TensorType c = create_tensor<TensorType>(shape_c, data_type, 1); |
| 84 | TensorType dst = create_tensor<TensorType>(output_shape, data_type, 1); |
Moritz Pflanzer | 4dfc235 | 2017-08-02 14:51:36 +0100 | [diff] [blame] | 85 | |
| 86 | // Create and configure function |
| 87 | FunctionType gemm; |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 88 | // The GEMMinfo includes the values of the depth in case of reinterpreted 3d output. |
Gian Marco Iodice | 3139f03 | 2018-11-05 14:26:32 +0000 | [diff] [blame] | 89 | // If the output shape has the same number of dimensions of the input the method called is a 2D matrix multiplication (depth_output_reinterpreted_as_3D = 0), |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 90 | // in the other case we have to use the reinterpreted version of GEMM (depth_output_reinterpreted_as_3D = depth of the 3D output). |
Gian Marco Iodice | f3622be | 2019-07-29 14:27:16 +0100 | [diff] [blame] | 91 | gemm.configure(&a, |
| 92 | &b, |
| 93 | (disable_c) ? nullptr : &c, |
| 94 | &dst, |
| 95 | alpha, beta, |
| 96 | GEMMInfo(false, false, false, (reinterpret_output_as_3d ? output_shape[2] : 0), reinterpret_input_as_3d, false, GEMMLowpOutputStageInfo(), false, (reinterpret_input_as_3d |
| 97 | || reinterpret_output_as_3d))); |
Moritz Pflanzer | 4dfc235 | 2017-08-02 14:51:36 +0100 | [diff] [blame] | 98 | ARM_COMPUTE_EXPECT(a.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 99 | ARM_COMPUTE_EXPECT(b.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 100 | ARM_COMPUTE_EXPECT(c.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 101 | ARM_COMPUTE_EXPECT(dst.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 102 | |
| 103 | // Allocate tensors |
| 104 | a.allocator()->allocate(); |
| 105 | b.allocator()->allocate(); |
| 106 | c.allocator()->allocate(); |
| 107 | dst.allocator()->allocate(); |
| 108 | |
| 109 | ARM_COMPUTE_EXPECT(!a.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 110 | ARM_COMPUTE_EXPECT(!b.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 111 | ARM_COMPUTE_EXPECT(!c.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 112 | ARM_COMPUTE_EXPECT(!dst.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 113 | |
| 114 | // Fill tensors |
| 115 | fill(AccessorType(a), 0); |
| 116 | fill(AccessorType(b), 1); |
Pablo Tello | 0e37b5c | 2018-10-30 11:18:37 +0000 | [diff] [blame] | 117 | if(!disable_c) |
| 118 | { |
| 119 | fill(AccessorType(c), 2); |
| 120 | } |
Moritz Pflanzer | 4dfc235 | 2017-08-02 14:51:36 +0100 | [diff] [blame] | 121 | |
| 122 | // Compute GEMM function |
| 123 | gemm.run(); |
| 124 | |
| 125 | return dst; |
| 126 | } |
| 127 | |
Michalis Spyrou | 6bff195 | 2019-10-02 17:22:11 +0100 | [diff] [blame] | 128 | SimpleTensor<T> compute_reference(const TensorShape &shape_a, const TensorShape &shape_b, const TensorShape &output_shape, float alpha, float beta, |
Vidhya Sudhan Loganathan | 014333d | 2018-07-02 09:13:49 +0100 | [diff] [blame] | 129 | DataType data_type) |
Moritz Pflanzer | 4dfc235 | 2017-08-02 14:51:36 +0100 | [diff] [blame] | 130 | { |
Gian Marco Iodice | 68a3f56 | 2018-07-26 11:44:03 +0100 | [diff] [blame] | 131 | TensorShape shape_a_to_use = shape_a; |
Gian Marco Iodice | f3622be | 2019-07-29 14:27:16 +0100 | [diff] [blame] | 132 | |
Gian Marco Iodice | 68a3f56 | 2018-07-26 11:44:03 +0100 | [diff] [blame] | 133 | if(reinterpret_input_as_3d) |
| 134 | { |
| 135 | // Collapse the second and third dimension if the input is 3D |
| 136 | shape_a_to_use.collapse(2U, 1U); |
| 137 | } |
| 138 | |
Moritz Pflanzer | 4dfc235 | 2017-08-02 14:51:36 +0100 | [diff] [blame] | 139 | // Create reference |
Gian Marco Iodice | 68a3f56 | 2018-07-26 11:44:03 +0100 | [diff] [blame] | 140 | SimpleTensor<T> a{ shape_a_to_use, data_type, 1 }; |
Vidhya Sudhan Loganathan | 014333d | 2018-07-02 09:13:49 +0100 | [diff] [blame] | 141 | SimpleTensor<T> b{ shape_b, data_type, 1 }; |
Gian Marco Iodice | f3622be | 2019-07-29 14:27:16 +0100 | [diff] [blame] | 142 | SimpleTensor<T> c{ output_shape, data_type, 1 }; |
Moritz Pflanzer | 4dfc235 | 2017-08-02 14:51:36 +0100 | [diff] [blame] | 143 | |
| 144 | // Fill reference |
| 145 | fill(a, 0); |
| 146 | fill(b, 1); |
Gian Marco Iodice | f3622be | 2019-07-29 14:27:16 +0100 | [diff] [blame] | 147 | fill(c, 2); |
| 148 | |
| 149 | if(reinterpret_input_as_3d || reinterpret_output_as_3d) |
Pablo Tello | 0e37b5c | 2018-10-30 11:18:37 +0000 | [diff] [blame] | 150 | { |
Gian Marco Iodice | f3622be | 2019-07-29 14:27:16 +0100 | [diff] [blame] | 151 | const int n = shape_b[0]; |
| 152 | const int m = reinterpret_output_as_3d ? output_shape[1] * output_shape[2] : output_shape[1]; |
| 153 | const int batch_size = reinterpret_output_as_3d ? output_shape[3] : output_shape[2]; |
| 154 | |
| 155 | // In case of broadcast, we need simply copy the first into the following "M" ones |
| 156 | for(int i = 1; i < m * batch_size; i++) |
| 157 | { |
| 158 | memcpy(c.data() + i * n, c.data(), n * sizeof(T)); |
| 159 | } |
Pablo Tello | 0e37b5c | 2018-10-30 11:18:37 +0000 | [diff] [blame] | 160 | } |
Gian Marco Iodice | f3622be | 2019-07-29 14:27:16 +0100 | [diff] [blame] | 161 | |
| 162 | // Setting beta to 0 will effectively disable C for the |
| 163 | // computation of the reference: alpha * A * B + 0 * C |
| 164 | return reference::gemm<T>(a, b, c, alpha, disable_c ? 0.f : beta); |
Moritz Pflanzer | 4dfc235 | 2017-08-02 14:51:36 +0100 | [diff] [blame] | 165 | } |
| 166 | |
| 167 | TensorType _target{}; |
| 168 | SimpleTensor<T> _reference{}; |
Moritz Pflanzer | 4dfc235 | 2017-08-02 14:51:36 +0100 | [diff] [blame] | 169 | }; |
| 170 | |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 171 | template <typename TensorType, typename AccessorType, typename T, typename GEMMFunctionType> |
| 172 | class GEMMMatrixMultiplyValidationFixture : public framework::Fixture |
| 173 | { |
| 174 | public: |
| 175 | template <typename...> |
| 176 | void setup(unsigned int m, unsigned int n, unsigned int k, unsigned int batch_size, float alpha, float beta, bool broadcast_bias, bool fp16_mixed_precision, const ActivationLayerInfo &act_info, |
| 177 | DataType data_type, GPUTarget gpu_arch) |
| 178 | { |
| 179 | // Set the tensor shapes for LHS and RHS matrices |
| 180 | const TensorShape lhs_shape(k, m, batch_size); |
| 181 | const TensorShape rhs_shape(n, k, batch_size); |
| 182 | const TensorShape bias_shape(n, |
| 183 | broadcast_bias ? 1 : m, |
| 184 | broadcast_bias ? 1 : batch_size); |
| 185 | |
| 186 | _target = compute_target(lhs_shape, rhs_shape, bias_shape, data_type, alpha, beta, broadcast_bias, fp16_mixed_precision, act_info, gpu_arch); |
Michalis Spyrou | 6bff195 | 2019-10-02 17:22:11 +0100 | [diff] [blame] | 187 | _reference = compute_reference(lhs_shape, rhs_shape, data_type, alpha, beta, broadcast_bias, act_info); |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 188 | } |
| 189 | |
| 190 | protected: |
| 191 | template <typename U> |
| 192 | void fill(U &&tensor, int i) |
| 193 | { |
| 194 | std::uniform_real_distribution<> distribution(-1.0f, 1.0f); |
| 195 | library->fill(tensor, distribution, i); |
| 196 | |
| 197 | // Fill border with infinity in order to check the presence of NaN values (i.e. inf * 0) |
| 198 | std::uniform_real_distribution<> distribution_inf(std::numeric_limits<float>::infinity(), std::numeric_limits<float>::infinity()); |
| 199 | library->fill_borders_with_garbage(tensor, distribution_inf, i); |
| 200 | } |
| 201 | |
| 202 | TensorType compute_target(const TensorShape &lhs_shape, const TensorShape &rhs_shape, const TensorShape &bias_shape, DataType data_type, float alpha, float beta, bool broadcast_bias, |
| 203 | bool fp16_mixed_precision, const ActivationLayerInfo &act_info, GPUTarget gpu_arch) |
| 204 | { |
| 205 | // Create tensors |
| 206 | TensorType lhs = create_tensor<TensorType>(lhs_shape, data_type, 1); |
| 207 | TensorType rhs = create_tensor<TensorType>(rhs_shape, data_type, 1); |
| 208 | TensorType bias = create_tensor<TensorType>(bias_shape, data_type, 1); |
| 209 | TensorType dst; |
| 210 | |
| 211 | const unsigned int m = lhs_shape[1]; |
| 212 | const unsigned int n = rhs_shape[0]; |
| 213 | const unsigned int k = lhs_shape[0]; |
| 214 | GEMMReshapeInfo reshape_info(m, n, k, 1, 1, 0, false, broadcast_bias); |
| 215 | |
| 216 | // The output tensor will be auto-initialized within the function |
| 217 | |
| 218 | // Create and configure function |
| 219 | GEMMFunctionType gemm; |
| 220 | gemm.configure(gpu_arch, &lhs, &rhs, &bias, &dst, alpha, beta, false, reshape_info, fp16_mixed_precision, act_info); |
| 221 | |
| 222 | ARM_COMPUTE_EXPECT(lhs.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 223 | ARM_COMPUTE_EXPECT(rhs.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 224 | ARM_COMPUTE_EXPECT(bias.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 225 | |
| 226 | // Allocate tensors |
| 227 | lhs.allocator()->allocate(); |
| 228 | rhs.allocator()->allocate(); |
| 229 | bias.allocator()->allocate(); |
| 230 | dst.allocator()->allocate(); |
| 231 | |
| 232 | ARM_COMPUTE_EXPECT(!lhs.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 233 | ARM_COMPUTE_EXPECT(!rhs.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 234 | ARM_COMPUTE_EXPECT(!bias.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 235 | ARM_COMPUTE_EXPECT(!dst.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 236 | |
| 237 | // Fill tensors |
| 238 | fill(AccessorType(lhs), 0); |
| 239 | fill(AccessorType(rhs), 1); |
| 240 | fill(AccessorType(bias), 2); |
| 241 | |
| 242 | // Compute GEMM |
| 243 | gemm.run(); |
| 244 | |
| 245 | return dst; |
| 246 | } |
| 247 | |
Michalis Spyrou | 6bff195 | 2019-10-02 17:22:11 +0100 | [diff] [blame] | 248 | SimpleTensor<T> compute_reference(const TensorShape &lhs_shape, const TensorShape &rhs_shape, DataType data_type, float alpha, float beta, bool broadcast_bias, |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 249 | const ActivationLayerInfo &act_info) |
| 250 | { |
| 251 | TensorShape dst_shape = lhs_shape; |
| 252 | dst_shape[0] = rhs_shape[0]; |
| 253 | dst_shape[1] = lhs_shape[1]; |
| 254 | |
| 255 | // Create reference |
| 256 | SimpleTensor<T> lhs{ lhs_shape, data_type, 1 }; |
| 257 | SimpleTensor<T> rhs{ rhs_shape, data_type, 1 }; |
| 258 | SimpleTensor<T> bias{ dst_shape, data_type, 1 }; |
| 259 | |
| 260 | const int n = rhs_shape[0]; |
| 261 | const int m = lhs_shape[1]; |
| 262 | const int batch_size = lhs_shape[2]; |
| 263 | |
| 264 | // Fill reference |
| 265 | fill(lhs, 0); |
| 266 | fill(rhs, 1); |
| 267 | fill(bias, 2); |
| 268 | |
| 269 | if(broadcast_bias) |
| 270 | { |
| 271 | // In case of broadcast, we need simply copy the first into the following "M" ones |
| 272 | for(int i = 1; i < m * batch_size; i++) |
| 273 | { |
| 274 | memcpy(bias.data() + i * n, bias.data(), n * sizeof(T)); |
| 275 | } |
| 276 | } |
| 277 | |
| 278 | return reference::activation_layer(reference::gemm<T>(lhs, rhs, bias, alpha, beta), act_info); |
| 279 | } |
| 280 | |
| 281 | TensorType _target{}; |
| 282 | SimpleTensor<T> _reference{}; |
| 283 | }; |
| 284 | |
| 285 | template <typename TensorType, typename AccessorType, typename T, typename GEMMFunctionType> |
| 286 | class GEMMMatrixMultiply3DValidationFixture : public framework::Fixture |
| 287 | { |
| 288 | public: |
| 289 | template <typename...> |
| 290 | void setup(unsigned int m_w, unsigned int m_h, unsigned int n, unsigned int k, unsigned int batch_size, float alpha, float beta, bool broadcast_bias, bool fp16_mixed_precision, |
| 291 | const ActivationLayerInfo &act_info, DataType data_type, GPUTarget gpu_arch) |
| 292 | { |
Michalis Spyrou | 6bff195 | 2019-10-02 17:22:11 +0100 | [diff] [blame] | 293 | ARM_COMPUTE_UNUSED(broadcast_bias); |
| 294 | |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 295 | // In case of GEMM3D, m is the product between m_w and m_h |
| 296 | const unsigned int m = m_w * m_h; |
| 297 | |
| 298 | // Set the tensor shapes for LHS and RHS matrices |
| 299 | const TensorShape lhs_shape(k, m, batch_size); |
| 300 | const TensorShape rhs_shape(n, k, batch_size); |
| 301 | const TensorShape bias_shape(n, 1, 1); |
| 302 | |
| 303 | _target = compute_target(lhs_shape, rhs_shape, bias_shape, data_type, alpha, beta, m_h, fp16_mixed_precision, act_info, gpu_arch); |
Michalis Spyrou | 6bff195 | 2019-10-02 17:22:11 +0100 | [diff] [blame] | 304 | _reference = compute_reference(lhs_shape, rhs_shape, data_type, alpha, beta, m_h, act_info); |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 305 | } |
| 306 | |
| 307 | protected: |
| 308 | template <typename U> |
| 309 | void fill(U &&tensor, int i) |
| 310 | { |
| 311 | std::uniform_real_distribution<> distribution(-1.0f, 1.0f); |
| 312 | library->fill(tensor, distribution, i); |
| 313 | } |
| 314 | |
| 315 | TensorType compute_target(const TensorShape &lhs_shape, const TensorShape &rhs_shape, const TensorShape &bias_shape, DataType data_type, float alpha, float beta, unsigned int m_h, |
| 316 | bool fp16_mixed_precision, const ActivationLayerInfo &act_info, GPUTarget gpu_arch) |
| 317 | { |
| 318 | // Create tensors |
| 319 | TensorType lhs = create_tensor<TensorType>(lhs_shape, data_type, 1); |
| 320 | TensorType rhs = create_tensor<TensorType>(rhs_shape, data_type, 1); |
| 321 | TensorType bias = create_tensor<TensorType>(bias_shape, data_type, 1); |
| 322 | TensorType dst; |
| 323 | |
| 324 | const unsigned int m = lhs_shape[1]; |
| 325 | const unsigned int n = rhs_shape[0]; |
| 326 | const unsigned int k = lhs_shape[0]; |
| 327 | GEMMReshapeInfo reshape_info(m, n, k, 1, 1, m_h, false, true); |
| 328 | |
| 329 | // The output tensor will be auto-initialized within the function |
| 330 | |
| 331 | // Create and configure function |
| 332 | GEMMFunctionType gemm; |
| 333 | gemm.configure(gpu_arch, &lhs, &rhs, &bias, &dst, alpha, beta, false, reshape_info, fp16_mixed_precision, act_info); |
| 334 | |
| 335 | ARM_COMPUTE_EXPECT(lhs.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 336 | ARM_COMPUTE_EXPECT(rhs.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 337 | ARM_COMPUTE_EXPECT(bias.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 338 | |
| 339 | // Allocate tensors |
| 340 | lhs.allocator()->allocate(); |
| 341 | rhs.allocator()->allocate(); |
| 342 | bias.allocator()->allocate(); |
| 343 | dst.allocator()->allocate(); |
| 344 | |
| 345 | ARM_COMPUTE_EXPECT(!lhs.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 346 | ARM_COMPUTE_EXPECT(!rhs.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 347 | ARM_COMPUTE_EXPECT(!bias.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 348 | ARM_COMPUTE_EXPECT(!dst.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 349 | |
| 350 | // Fill tensors |
| 351 | fill(AccessorType(lhs), 0); |
| 352 | fill(AccessorType(rhs), 1); |
| 353 | fill(AccessorType(bias), 2); |
| 354 | |
| 355 | // Compute GEMM |
| 356 | gemm.run(); |
| 357 | |
| 358 | return dst; |
| 359 | } |
| 360 | |
Michalis Spyrou | 6bff195 | 2019-10-02 17:22:11 +0100 | [diff] [blame] | 361 | SimpleTensor<T> compute_reference(const TensorShape &lhs_shape, const TensorShape &rhs_shape, DataType data_type, float alpha, float beta, unsigned int m_h, |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 362 | const ActivationLayerInfo &act_info) |
| 363 | { |
| 364 | TensorShape dst_shape = lhs_shape; |
| 365 | dst_shape.set(0, rhs_shape[0]); |
| 366 | dst_shape.set(1, lhs_shape[1] / m_h); |
| 367 | dst_shape.set(2, m_h); |
| 368 | dst_shape.set(3, lhs_shape[2]); |
| 369 | |
| 370 | // Create reference |
| 371 | SimpleTensor<T> lhs{ lhs_shape, data_type, 1 }; |
| 372 | SimpleTensor<T> rhs{ rhs_shape, data_type, 1 }; |
| 373 | SimpleTensor<T> bias{ dst_shape, data_type, 1 }; |
| 374 | |
| 375 | const int n = rhs_shape[0]; |
| 376 | const int m = lhs_shape[1]; |
| 377 | const int batch_size = lhs_shape[2]; |
| 378 | |
| 379 | // Fill reference |
| 380 | fill(lhs, 0); |
| 381 | fill(rhs, 1); |
| 382 | fill(bias, 2); |
| 383 | |
| 384 | // In case of broadcast, we need simply copy the first into the following "M" ones |
| 385 | for(int i = 1; i < m * batch_size; i++) |
| 386 | { |
| 387 | memcpy(bias.data() + i * n, bias.data(), n * sizeof(T)); |
| 388 | } |
| 389 | |
| 390 | return reference::activation_layer(reference::gemm<T>(lhs, rhs, bias, alpha, beta), act_info); |
| 391 | } |
| 392 | |
| 393 | TensorType _target{}; |
| 394 | SimpleTensor<T> _reference{}; |
| 395 | }; |
| 396 | |
| 397 | template <typename TensorType, typename AccessorType, typename T, typename ReshapeLHSFunctionType, typename ReshapeRHSFunctionType, typename GEMMFunctionType> |
| 398 | class GEMMMatrixMultiplyInterleavedTransposedValidationFixture : public framework::Fixture |
| 399 | { |
| 400 | public: |
| 401 | template <typename...> |
| 402 | void setup(unsigned int m, unsigned int n, unsigned int k, unsigned int batch_size, float alpha, float beta, unsigned int v0, unsigned int h0, bool broadcast_bias, bool fp16_mixed_precision, |
| 403 | const ActivationLayerInfo &act_info, DataType data_type, GPUTarget gpu_arch) |
| 404 | { |
| 405 | GEMMLHSMatrixInfo lhs_info; |
| 406 | lhs_info.m0 = 4; |
| 407 | lhs_info.k0 = 4; |
| 408 | lhs_info.v0 = v0; |
| 409 | lhs_info.interleave = true; |
| 410 | lhs_info.transpose = true; |
| 411 | |
| 412 | GEMMRHSMatrixInfo rhs_info; |
| 413 | rhs_info.n0 = 16 / sizeof(T); |
| 414 | rhs_info.k0 = 1; |
| 415 | rhs_info.h0 = h0; |
| 416 | rhs_info.interleave = false; |
| 417 | rhs_info.transpose = false; |
| 418 | |
| 419 | // Set the tensor shapes for LHS and RHS matrices |
| 420 | const TensorShape lhs_shape(k, m, batch_size); |
| 421 | const TensorShape rhs_shape(n, k, batch_size); |
| 422 | const TensorShape bias_shape(n, |
| 423 | broadcast_bias ? 1 : m, |
| 424 | broadcast_bias ? 1 : batch_size); |
| 425 | |
| 426 | _target = compute_target(lhs_shape, rhs_shape, bias_shape, lhs_info, rhs_info, data_type, alpha, beta, broadcast_bias, fp16_mixed_precision, act_info, gpu_arch); |
Michalis Spyrou | 6bff195 | 2019-10-02 17:22:11 +0100 | [diff] [blame] | 427 | _reference = compute_reference(lhs_shape, rhs_shape, data_type, alpha, beta, broadcast_bias, act_info); |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 428 | } |
| 429 | |
| 430 | protected: |
| 431 | template <typename U> |
| 432 | void fill(U &&tensor, int i) |
| 433 | { |
| 434 | std::uniform_real_distribution<> distribution(-1.0f, 1.0f); |
| 435 | library->fill(tensor, distribution, i); |
| 436 | |
| 437 | // Fill border with infinity in order to check the presence of NaN values (i.e. inf * 0) |
| 438 | std::uniform_real_distribution<> distribution_inf(std::numeric_limits<float>::infinity(), std::numeric_limits<float>::infinity()); |
| 439 | library->fill_borders_with_garbage(tensor, distribution_inf, i); |
| 440 | } |
| 441 | |
| 442 | TensorType compute_target(const TensorShape &lhs_shape, const TensorShape &rhs_shape, const TensorShape &bias_shape, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, |
| 443 | DataType data_type, float alpha, float beta, bool broadcast_bias, bool fp16_mixed_precision, const ActivationLayerInfo &act_info, GPUTarget gpu_arch) |
| 444 | { |
| 445 | // Create tensors |
| 446 | TensorType lhs = create_tensor<TensorType>(lhs_shape, data_type, 1); |
| 447 | TensorType rhs = create_tensor<TensorType>(rhs_shape, data_type, 1); |
| 448 | TensorType bias = create_tensor<TensorType>(bias_shape, data_type, 1); |
| 449 | TensorType lhs_reshaped; |
| 450 | TensorType rhs_reshaped; |
| 451 | TensorType dst; |
| 452 | |
| 453 | const unsigned int m = lhs_shape[1]; |
| 454 | const unsigned int n = rhs_shape[0]; |
| 455 | const unsigned int k = lhs_shape[0]; |
| 456 | GEMMReshapeInfo reshape_info(m, n, k, rhs_info.h0, lhs_info.v0, 0, false, broadcast_bias); |
| 457 | |
| 458 | // The output tensor will be auto-initialized within the function |
| 459 | |
| 460 | // Create and configure function |
| 461 | ReshapeLHSFunctionType reshape_lhs; |
| 462 | ReshapeRHSFunctionType reshape_rhs; |
| 463 | GEMMFunctionType gemm; |
| 464 | reshape_lhs.configure(&lhs, &lhs_reshaped, lhs_info); |
| 465 | reshape_rhs.configure(&rhs, &rhs_reshaped, rhs_info); |
| 466 | gemm.configure(gpu_arch, &lhs_reshaped, &rhs_reshaped, &bias, &dst, alpha, beta, true, reshape_info, fp16_mixed_precision, act_info); |
| 467 | |
| 468 | ARM_COMPUTE_EXPECT(lhs.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 469 | ARM_COMPUTE_EXPECT(rhs.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 470 | ARM_COMPUTE_EXPECT(bias.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 471 | |
| 472 | // Allocate tensors |
| 473 | lhs.allocator()->allocate(); |
| 474 | rhs.allocator()->allocate(); |
| 475 | lhs_reshaped.allocator()->allocate(); |
| 476 | rhs_reshaped.allocator()->allocate(); |
| 477 | bias.allocator()->allocate(); |
| 478 | dst.allocator()->allocate(); |
| 479 | |
| 480 | ARM_COMPUTE_EXPECT(!lhs.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 481 | ARM_COMPUTE_EXPECT(!rhs.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 482 | ARM_COMPUTE_EXPECT(!bias.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 483 | ARM_COMPUTE_EXPECT(!lhs_reshaped.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 484 | ARM_COMPUTE_EXPECT(!rhs_reshaped.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 485 | ARM_COMPUTE_EXPECT(!dst.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 486 | |
| 487 | // Fill tensors |
| 488 | fill(AccessorType(lhs), 0); |
| 489 | fill(AccessorType(rhs), 1); |
| 490 | fill(AccessorType(bias), 2); |
| 491 | |
| 492 | // Compute GEMM |
| 493 | reshape_lhs.run(); |
| 494 | reshape_rhs.run(); |
| 495 | gemm.run(); |
| 496 | |
| 497 | return dst; |
| 498 | } |
| 499 | |
Michalis Spyrou | 6bff195 | 2019-10-02 17:22:11 +0100 | [diff] [blame] | 500 | SimpleTensor<T> compute_reference(const TensorShape &lhs_shape, const TensorShape &rhs_shape, DataType data_type, float alpha, float beta, bool broadcast_bias, |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 501 | const ActivationLayerInfo &act_info) |
| 502 | { |
| 503 | TensorShape dst_shape = lhs_shape; |
| 504 | dst_shape[0] = rhs_shape[0]; |
| 505 | dst_shape[1] = lhs_shape[1]; |
| 506 | |
| 507 | // Create reference |
| 508 | SimpleTensor<T> lhs{ lhs_shape, data_type, 1 }; |
| 509 | SimpleTensor<T> rhs{ rhs_shape, data_type, 1 }; |
| 510 | SimpleTensor<T> bias{ dst_shape, data_type, 1 }; |
| 511 | |
| 512 | const int n = rhs_shape[0]; |
| 513 | const int m = lhs_shape[1]; |
| 514 | const int batch_size = lhs_shape[2]; |
| 515 | |
| 516 | // Fill reference |
| 517 | fill(lhs, 0); |
| 518 | fill(rhs, 1); |
| 519 | fill(bias, 2); |
| 520 | |
| 521 | if(broadcast_bias) |
| 522 | { |
| 523 | // In case of broadcast, we need simply copy the first into the following "M" ones |
| 524 | for(int i = 1; i < m * batch_size; i++) |
| 525 | { |
| 526 | memcpy(bias.data() + i * n, bias.data(), n * sizeof(T)); |
| 527 | } |
| 528 | } |
| 529 | |
| 530 | return reference::activation_layer(reference::gemm<T>(lhs, rhs, bias, alpha, beta), act_info); |
| 531 | } |
| 532 | |
| 533 | TensorType _target{}; |
| 534 | SimpleTensor<T> _reference{}; |
| 535 | }; |
| 536 | |
| 537 | template <typename TensorType, typename AccessorType, typename T, typename ReshapeLHSFunctionType, typename ReshapeRHSFunctionType, typename GEMMFunctionType> |
| 538 | class GEMMMatrixMultiplyInterleavedTransposed3DValidationFixture : public framework::Fixture |
| 539 | { |
| 540 | public: |
| 541 | template <typename...> |
| 542 | void setup(unsigned int m_w, unsigned int m_h, unsigned int n, unsigned int k, unsigned int batch_size, float alpha, float beta, unsigned int v0, unsigned int h0, bool broadcast_bias, |
| 543 | bool fp16_mixed_precision, const ActivationLayerInfo &act_info, DataType data_type, GPUTarget gpu_arch) |
| 544 | { |
Michalis Spyrou | 6bff195 | 2019-10-02 17:22:11 +0100 | [diff] [blame] | 545 | ARM_COMPUTE_UNUSED(broadcast_bias); |
| 546 | |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 547 | GEMMLHSMatrixInfo lhs_info; |
| 548 | lhs_info.m0 = 4; |
| 549 | lhs_info.k0 = 4; |
| 550 | lhs_info.v0 = v0; |
| 551 | lhs_info.interleave = true; |
| 552 | lhs_info.transpose = true; |
| 553 | |
| 554 | GEMMRHSMatrixInfo rhs_info; |
| 555 | rhs_info.n0 = 16 / sizeof(T); |
| 556 | rhs_info.k0 = 1; |
| 557 | rhs_info.h0 = h0; |
| 558 | rhs_info.interleave = false; |
| 559 | rhs_info.transpose = false; |
| 560 | |
| 561 | // In case of GEMM3D, m is the product between m_w and m_h |
| 562 | const unsigned int m = m_w * m_h; |
| 563 | |
| 564 | // Set the tensor shapes for LHS and RHS matrices |
| 565 | const TensorShape lhs_shape(k, m, batch_size); |
| 566 | const TensorShape rhs_shape(n, k, batch_size); |
| 567 | const TensorShape bias_shape(n, 1, 1); |
| 568 | |
| 569 | _target = compute_target(lhs_shape, rhs_shape, bias_shape, lhs_info, rhs_info, data_type, alpha, beta, m_h, fp16_mixed_precision, act_info, gpu_arch); |
Michalis Spyrou | 6bff195 | 2019-10-02 17:22:11 +0100 | [diff] [blame] | 570 | _reference = compute_reference(lhs_shape, rhs_shape, data_type, alpha, beta, m_h, act_info); |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 571 | } |
| 572 | |
| 573 | protected: |
| 574 | template <typename U> |
| 575 | void fill(U &&tensor, int i) |
| 576 | { |
| 577 | std::uniform_real_distribution<> distribution(-1.0f, 1.0f); |
| 578 | library->fill(tensor, distribution, i); |
| 579 | } |
| 580 | |
| 581 | TensorType compute_target(const TensorShape &lhs_shape, const TensorShape &rhs_shape, const TensorShape &bias_shape, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, |
| 582 | DataType data_type, float alpha, float beta, unsigned int m_h, bool fp16_mixed_precision, const ActivationLayerInfo &act_info, GPUTarget gpu_arch) |
| 583 | { |
| 584 | // Create tensors |
| 585 | TensorType lhs = create_tensor<TensorType>(lhs_shape, data_type, 1); |
| 586 | TensorType rhs = create_tensor<TensorType>(rhs_shape, data_type, 1); |
| 587 | TensorType bias = create_tensor<TensorType>(bias_shape, data_type, 1); |
| 588 | TensorType lhs_reshaped; |
| 589 | TensorType rhs_reshaped; |
| 590 | TensorType dst; |
| 591 | |
| 592 | const unsigned int m = lhs_shape[1]; |
| 593 | const unsigned int n = rhs_shape[0]; |
| 594 | const unsigned int k = lhs_shape[0]; |
| 595 | GEMMReshapeInfo reshape_info(m, n, k, rhs_info.h0, lhs_info.v0, m_h, false, true); |
| 596 | |
| 597 | // The output tensor will be auto-initialized within the function |
| 598 | |
| 599 | // Create and configure function |
| 600 | ReshapeLHSFunctionType reshape_lhs; |
| 601 | ReshapeRHSFunctionType reshape_rhs; |
| 602 | GEMMFunctionType gemm; |
| 603 | reshape_lhs.configure(&lhs, &lhs_reshaped, lhs_info); |
| 604 | reshape_rhs.configure(&rhs, &rhs_reshaped, rhs_info); |
| 605 | gemm.configure(gpu_arch, &lhs_reshaped, &rhs_reshaped, &bias, &dst, alpha, beta, true, reshape_info, fp16_mixed_precision, act_info); |
| 606 | |
| 607 | ARM_COMPUTE_EXPECT(lhs.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 608 | ARM_COMPUTE_EXPECT(rhs.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 609 | ARM_COMPUTE_EXPECT(bias.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 610 | |
| 611 | // Allocate tensors |
| 612 | lhs.allocator()->allocate(); |
| 613 | rhs.allocator()->allocate(); |
| 614 | lhs_reshaped.allocator()->allocate(); |
| 615 | rhs_reshaped.allocator()->allocate(); |
| 616 | bias.allocator()->allocate(); |
| 617 | dst.allocator()->allocate(); |
| 618 | |
| 619 | ARM_COMPUTE_EXPECT(!lhs.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 620 | ARM_COMPUTE_EXPECT(!rhs.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 621 | ARM_COMPUTE_EXPECT(!lhs_reshaped.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 622 | ARM_COMPUTE_EXPECT(!rhs_reshaped.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 623 | ARM_COMPUTE_EXPECT(!bias.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 624 | ARM_COMPUTE_EXPECT(!dst.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 625 | |
| 626 | // Fill tensors |
| 627 | fill(AccessorType(lhs), 0); |
| 628 | fill(AccessorType(rhs), 1); |
| 629 | fill(AccessorType(bias), 2); |
| 630 | |
| 631 | // Compute GEMM |
| 632 | reshape_lhs.run(); |
| 633 | reshape_rhs.run(); |
| 634 | gemm.run(); |
| 635 | |
| 636 | return dst; |
| 637 | } |
| 638 | |
Michalis Spyrou | 6bff195 | 2019-10-02 17:22:11 +0100 | [diff] [blame] | 639 | SimpleTensor<T> compute_reference(const TensorShape &lhs_shape, const TensorShape &rhs_shape, DataType data_type, float alpha, float beta, unsigned int m_h, |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 640 | const ActivationLayerInfo &act_info) |
| 641 | { |
| 642 | TensorShape dst_shape = lhs_shape; |
| 643 | dst_shape.set(0, rhs_shape[0]); |
| 644 | dst_shape.set(1, lhs_shape[1] / m_h); |
| 645 | dst_shape.set(2, m_h); |
| 646 | dst_shape.set(3, lhs_shape[2]); |
| 647 | |
| 648 | // Create reference |
| 649 | SimpleTensor<T> lhs{ lhs_shape, data_type, 1 }; |
| 650 | SimpleTensor<T> rhs{ rhs_shape, data_type, 1 }; |
| 651 | SimpleTensor<T> bias{ dst_shape, data_type, 1 }; |
| 652 | |
| 653 | const int n = rhs_shape[0]; |
| 654 | const int m = lhs_shape[1]; |
| 655 | const int batch_size = lhs_shape[2]; |
| 656 | |
| 657 | // Fill reference |
| 658 | fill(lhs, 0); |
| 659 | fill(rhs, 1); |
| 660 | fill(bias, 2); |
| 661 | |
| 662 | // In case of broadcast, we need simply copy the first into the following "M" ones |
| 663 | for(int i = 1; i < m * batch_size; i++) |
| 664 | { |
| 665 | memcpy(bias.data() + i * n, bias.data(), n * sizeof(T)); |
| 666 | } |
| 667 | |
| 668 | return reference::activation_layer(reference::gemm<T>(lhs, rhs, bias, alpha, beta), act_info); |
| 669 | } |
| 670 | |
| 671 | TensorType _target{}; |
| 672 | SimpleTensor<T> _reference{}; |
| 673 | }; |
| 674 | |
Gian Marco Iodice | 0c17aa2 | 2019-09-27 09:23:15 +0100 | [diff] [blame] | 675 | template <typename TensorType, typename AccessorType, typename T, typename ReshapeLHSFunctionType, typename ReshapeRHSFunctionType, typename GEMMFunctionType, bool fp_mixed_precision = false> |
Gian Marco Iodice | bf9731e | 2018-12-12 10:18:04 +0000 | [diff] [blame] | 676 | class GEMMMatrixMultiplyReshapedValidationFixture : public framework::Fixture |
| 677 | { |
| 678 | public: |
| 679 | template <typename...> |
| 680 | void setup(unsigned int m, unsigned int n, unsigned int k, unsigned int batch_size, unsigned int m0, unsigned int n0, unsigned int k0, unsigned int v0, unsigned int h0, bool interleave_lhs, |
Gian Marco Iodice | e3a849a | 2020-06-10 17:59:30 +0100 | [diff] [blame] | 681 | bool interleave_rhs, bool export_to_cl_image, DataType data_type, float alpha, float beta, bool broadcast_bias, bool lhs_transpose, const ActivationLayerInfo &act_info) |
Gian Marco Iodice | bf9731e | 2018-12-12 10:18:04 +0000 | [diff] [blame] | 682 | { |
| 683 | GEMMLHSMatrixInfo lhs_info; |
| 684 | lhs_info.m0 = m0; |
| 685 | lhs_info.k0 = k0; |
| 686 | lhs_info.v0 = v0; |
| 687 | lhs_info.interleave = interleave_lhs; |
Giorgio Arena | ae99b6e | 2019-08-01 14:22:12 +0100 | [diff] [blame] | 688 | lhs_info.transpose = lhs_transpose; |
Gian Marco Iodice | bf9731e | 2018-12-12 10:18:04 +0000 | [diff] [blame] | 689 | |
| 690 | GEMMRHSMatrixInfo rhs_info; |
Gian Marco Iodice | e3a849a | 2020-06-10 17:59:30 +0100 | [diff] [blame] | 691 | rhs_info.n0 = n0; |
| 692 | rhs_info.k0 = k0; |
| 693 | rhs_info.h0 = h0; |
| 694 | rhs_info.interleave = interleave_rhs; |
| 695 | rhs_info.transpose = !lhs_transpose; |
| 696 | rhs_info.export_to_cl_image = export_to_cl_image; |
Gian Marco Iodice | bf9731e | 2018-12-12 10:18:04 +0000 | [diff] [blame] | 697 | |
| 698 | // Set the tensor shapes for LHS and RHS matrices |
| 699 | const TensorShape lhs_shape(k, m, batch_size); |
| 700 | const TensorShape rhs_shape(n, k, batch_size); |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 701 | const TensorShape bias_shape(n, |
| 702 | broadcast_bias ? 1 : m, |
| 703 | broadcast_bias ? 1 : batch_size); |
Gian Marco Iodice | bf9731e | 2018-12-12 10:18:04 +0000 | [diff] [blame] | 704 | |
Gian Marco Iodice | ca1f460 | 2019-07-16 15:46:48 +0100 | [diff] [blame] | 705 | _target = compute_target(lhs_shape, rhs_shape, bias_shape, lhs_info, rhs_info, data_type, alpha, beta, broadcast_bias, act_info); |
Michalis Spyrou | 6bff195 | 2019-10-02 17:22:11 +0100 | [diff] [blame] | 706 | _reference = compute_reference(lhs_shape, rhs_shape, data_type, alpha, beta, broadcast_bias, act_info); |
Gian Marco Iodice | bf9731e | 2018-12-12 10:18:04 +0000 | [diff] [blame] | 707 | } |
| 708 | |
| 709 | protected: |
| 710 | template <typename U> |
| 711 | void fill(U &&tensor, int i) |
| 712 | { |
| 713 | std::uniform_real_distribution<> distribution(-1.0f, 1.0f); |
| 714 | library->fill(tensor, distribution, i); |
Gian Marco Iodice | b87b95e | 2019-01-21 17:14:31 +0000 | [diff] [blame] | 715 | |
| 716 | // Fill border with infinity in order to check the presence of NaN values (i.e. inf * 0) |
| 717 | std::uniform_real_distribution<> distribution_inf(std::numeric_limits<float>::infinity(), std::numeric_limits<float>::infinity()); |
| 718 | library->fill_borders_with_garbage(tensor, distribution_inf, i); |
Gian Marco Iodice | bf9731e | 2018-12-12 10:18:04 +0000 | [diff] [blame] | 719 | } |
| 720 | |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 721 | TensorType compute_target(const TensorShape &lhs_shape, const TensorShape &rhs_shape, const TensorShape &bias_shape, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, |
Gian Marco Iodice | ca1f460 | 2019-07-16 15:46:48 +0100 | [diff] [blame] | 722 | DataType data_type, float alpha, float beta, bool broadcast_bias, const ActivationLayerInfo &act_info) |
Gian Marco Iodice | bf9731e | 2018-12-12 10:18:04 +0000 | [diff] [blame] | 723 | { |
| 724 | // Create tensors |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 725 | TensorType lhs = create_tensor<TensorType>(lhs_shape, data_type, 1); |
| 726 | TensorType rhs = create_tensor<TensorType>(rhs_shape, data_type, 1); |
| 727 | TensorType bias = create_tensor<TensorType>(bias_shape, data_type, 1); |
Gian Marco Iodice | bf9731e | 2018-12-12 10:18:04 +0000 | [diff] [blame] | 728 | TensorType lhs_reshaped; |
| 729 | TensorType rhs_reshaped; |
| 730 | TensorType dst; |
| 731 | |
| 732 | const unsigned int M = lhs_shape[1]; |
| 733 | const unsigned int N = rhs_shape[0]; |
| 734 | const unsigned int K = lhs_shape[0]; |
Gian Marco Iodice | 7026b30 | 2019-06-26 17:18:11 +0100 | [diff] [blame] | 735 | GEMMKernelInfo kernel_info; |
| 736 | kernel_info.m = M; |
| 737 | kernel_info.n = N; |
| 738 | kernel_info.k = K; |
| 739 | kernel_info.depth_output_gemm3d = 0; |
| 740 | kernel_info.reinterpret_input_as_3d = false; |
| 741 | kernel_info.broadcast_bias = broadcast_bias; |
Gian Marco Iodice | ca1f460 | 2019-07-16 15:46:48 +0100 | [diff] [blame] | 742 | kernel_info.activation_info = act_info; |
Gian Marco Iodice | 0c17aa2 | 2019-09-27 09:23:15 +0100 | [diff] [blame] | 743 | kernel_info.fp_mixed_precision = fp_mixed_precision; |
Gian Marco Iodice | bf9731e | 2018-12-12 10:18:04 +0000 | [diff] [blame] | 744 | |
| 745 | // The output tensor will be auto-initialized within the function |
| 746 | |
| 747 | // Create and configure function |
| 748 | ReshapeLHSFunctionType reshape_lhs; |
| 749 | ReshapeRHSFunctionType reshape_rhs; |
| 750 | GEMMFunctionType gemm; |
| 751 | reshape_lhs.configure(&lhs, &lhs_reshaped, lhs_info); |
| 752 | reshape_rhs.configure(&rhs, &rhs_reshaped, rhs_info); |
Gian Marco Iodice | 7026b30 | 2019-06-26 17:18:11 +0100 | [diff] [blame] | 753 | gemm.configure(&lhs_reshaped, &rhs_reshaped, &bias, &dst, alpha, beta, lhs_info, rhs_info, kernel_info); |
Gian Marco Iodice | bf9731e | 2018-12-12 10:18:04 +0000 | [diff] [blame] | 754 | |
| 755 | ARM_COMPUTE_EXPECT(lhs.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 756 | ARM_COMPUTE_EXPECT(rhs.info()->is_resizable(), framework::LogLevel::ERRORS); |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 757 | ARM_COMPUTE_EXPECT(bias.info()->is_resizable(), framework::LogLevel::ERRORS); |
Gian Marco Iodice | bf9731e | 2018-12-12 10:18:04 +0000 | [diff] [blame] | 758 | |
| 759 | // Allocate tensors |
| 760 | lhs.allocator()->allocate(); |
| 761 | rhs.allocator()->allocate(); |
| 762 | lhs_reshaped.allocator()->allocate(); |
| 763 | rhs_reshaped.allocator()->allocate(); |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 764 | bias.allocator()->allocate(); |
Gian Marco Iodice | bf9731e | 2018-12-12 10:18:04 +0000 | [diff] [blame] | 765 | dst.allocator()->allocate(); |
| 766 | |
| 767 | ARM_COMPUTE_EXPECT(!lhs.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 768 | ARM_COMPUTE_EXPECT(!rhs.info()->is_resizable(), framework::LogLevel::ERRORS); |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 769 | ARM_COMPUTE_EXPECT(!bias.info()->is_resizable(), framework::LogLevel::ERRORS); |
Gian Marco Iodice | bf9731e | 2018-12-12 10:18:04 +0000 | [diff] [blame] | 770 | ARM_COMPUTE_EXPECT(!lhs_reshaped.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 771 | ARM_COMPUTE_EXPECT(!rhs_reshaped.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 772 | ARM_COMPUTE_EXPECT(!dst.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 773 | |
| 774 | // Fill tensors |
| 775 | fill(AccessorType(lhs), 0); |
| 776 | fill(AccessorType(rhs), 1); |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 777 | fill(AccessorType(bias), 2); |
Gian Marco Iodice | bf9731e | 2018-12-12 10:18:04 +0000 | [diff] [blame] | 778 | |
| 779 | // Compute GEMM |
| 780 | reshape_lhs.run(); |
| 781 | reshape_rhs.run(); |
| 782 | gemm.run(); |
| 783 | |
| 784 | return dst; |
| 785 | } |
| 786 | |
Michalis Spyrou | 6bff195 | 2019-10-02 17:22:11 +0100 | [diff] [blame] | 787 | SimpleTensor<T> compute_reference(const TensorShape &lhs_shape, const TensorShape &rhs_shape, DataType data_type, float alpha, float beta, bool broadcast_bias, |
Gian Marco Iodice | ca1f460 | 2019-07-16 15:46:48 +0100 | [diff] [blame] | 788 | const ActivationLayerInfo &act_info) |
Gian Marco Iodice | bf9731e | 2018-12-12 10:18:04 +0000 | [diff] [blame] | 789 | { |
| 790 | TensorShape dst_shape = lhs_shape; |
| 791 | dst_shape[0] = rhs_shape[0]; |
| 792 | dst_shape[1] = lhs_shape[1]; |
| 793 | |
| 794 | // Create reference |
Gian Marco Iodice | 9382ab3 | 2018-12-17 15:12:07 +0000 | [diff] [blame] | 795 | SimpleTensor<T> lhs{ lhs_shape, data_type, 1 }; |
| 796 | SimpleTensor<T> rhs{ rhs_shape, data_type, 1 }; |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 797 | SimpleTensor<T> bias{ dst_shape, data_type, 1 }; |
| 798 | |
| 799 | const int n = rhs_shape[0]; |
| 800 | const int m = lhs_shape[1]; |
| 801 | const int batch_size = lhs_shape[2]; |
Gian Marco Iodice | bf9731e | 2018-12-12 10:18:04 +0000 | [diff] [blame] | 802 | |
| 803 | // Fill reference |
| 804 | fill(lhs, 0); |
| 805 | fill(rhs, 1); |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 806 | fill(bias, 2); |
Gian Marco Iodice | bf9731e | 2018-12-12 10:18:04 +0000 | [diff] [blame] | 807 | |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 808 | if(broadcast_bias) |
| 809 | { |
| 810 | // In case of broadcast, we need simply copy the first into the following "M" ones |
| 811 | for(int i = 1; i < m * batch_size; i++) |
| 812 | { |
| 813 | memcpy(bias.data() + i * n, bias.data(), n * sizeof(T)); |
| 814 | } |
| 815 | } |
| 816 | |
Gian Marco Iodice | 0c17aa2 | 2019-09-27 09:23:15 +0100 | [diff] [blame] | 817 | if(fp_mixed_precision) |
| 818 | { |
| 819 | return reference::activation_layer(reference::gemm_mixed_precision<T>(lhs, rhs, bias, alpha, beta), act_info); |
| 820 | } |
| 821 | else |
| 822 | { |
| 823 | return reference::activation_layer(reference::gemm<T>(lhs, rhs, bias, alpha, beta), act_info); |
| 824 | } |
Gian Marco Iodice | bf9731e | 2018-12-12 10:18:04 +0000 | [diff] [blame] | 825 | } |
| 826 | |
Gian Marco Iodice | 9382ab3 | 2018-12-17 15:12:07 +0000 | [diff] [blame] | 827 | TensorType _target{}; |
| 828 | SimpleTensor<T> _reference{}; |
| 829 | }; |
| 830 | |
Gian Marco Iodice | 0c17aa2 | 2019-09-27 09:23:15 +0100 | [diff] [blame] | 831 | template <typename TensorType, typename AccessorType, typename T, typename ReshapeLHSFunctionType, typename ReshapeRHSFunctionType, typename GEMMFunctionType, bool fp_mixed_precision = false> |
Gian Marco Iodice | 9382ab3 | 2018-12-17 15:12:07 +0000 | [diff] [blame] | 832 | class GEMMMatrixMultiplyReshaped3DValidationFixture : public framework::Fixture |
| 833 | { |
| 834 | public: |
| 835 | template <typename...> |
| 836 | void setup(unsigned int m_w, unsigned int m_h, unsigned int n, unsigned int k, unsigned int batch_size, unsigned int m0, unsigned int n0, unsigned int k0, unsigned int v0, unsigned int h0, |
Gian Marco Iodice | e3a849a | 2020-06-10 17:59:30 +0100 | [diff] [blame] | 837 | bool interleave_lhs, bool interleave_rhs, bool export_to_cl_image, DataType data_type, float alpha, float beta, bool lhs_transpose, const ActivationLayerInfo &act_info) |
Gian Marco Iodice | 9382ab3 | 2018-12-17 15:12:07 +0000 | [diff] [blame] | 838 | { |
| 839 | GEMMLHSMatrixInfo lhs_info; |
| 840 | lhs_info.m0 = m0; |
| 841 | lhs_info.k0 = k0; |
| 842 | lhs_info.v0 = v0; |
| 843 | lhs_info.interleave = interleave_lhs; |
Giorgio Arena | ae99b6e | 2019-08-01 14:22:12 +0100 | [diff] [blame] | 844 | lhs_info.transpose = lhs_transpose; |
Gian Marco Iodice | 9382ab3 | 2018-12-17 15:12:07 +0000 | [diff] [blame] | 845 | |
| 846 | GEMMRHSMatrixInfo rhs_info; |
Gian Marco Iodice | e3a849a | 2020-06-10 17:59:30 +0100 | [diff] [blame] | 847 | rhs_info.n0 = n0; |
| 848 | rhs_info.k0 = k0; |
| 849 | rhs_info.h0 = h0; |
| 850 | rhs_info.interleave = interleave_rhs; |
| 851 | rhs_info.transpose = !lhs_transpose; |
| 852 | rhs_info.export_to_cl_image = export_to_cl_image; |
Gian Marco Iodice | 9382ab3 | 2018-12-17 15:12:07 +0000 | [diff] [blame] | 853 | |
| 854 | // In case of GEMM3D, m is the product between m_w and m_h |
| 855 | const unsigned int m = m_w * m_h; |
| 856 | |
| 857 | // Set the tensor shapes for LHS and RHS matrices |
| 858 | const TensorShape lhs_shape(k, m, batch_size); |
| 859 | const TensorShape rhs_shape(n, k, batch_size); |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 860 | const TensorShape bias_shape(n, 1, 1); |
Gian Marco Iodice | 9382ab3 | 2018-12-17 15:12:07 +0000 | [diff] [blame] | 861 | |
Gian Marco Iodice | ca1f460 | 2019-07-16 15:46:48 +0100 | [diff] [blame] | 862 | _target = compute_target(lhs_shape, rhs_shape, bias_shape, lhs_info, rhs_info, data_type, alpha, beta, m_h, act_info); |
Michalis Spyrou | 6bff195 | 2019-10-02 17:22:11 +0100 | [diff] [blame] | 863 | _reference = compute_reference(lhs_shape, rhs_shape, data_type, alpha, beta, m_h, act_info); |
Gian Marco Iodice | 9382ab3 | 2018-12-17 15:12:07 +0000 | [diff] [blame] | 864 | } |
| 865 | |
| 866 | protected: |
| 867 | template <typename U> |
| 868 | void fill(U &&tensor, int i) |
| 869 | { |
| 870 | std::uniform_real_distribution<> distribution(-1.0f, 1.0f); |
| 871 | library->fill(tensor, distribution, i); |
| 872 | } |
| 873 | |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 874 | TensorType compute_target(const TensorShape &lhs_shape, const TensorShape &rhs_shape, const TensorShape &bias_shape, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, |
Gian Marco Iodice | ca1f460 | 2019-07-16 15:46:48 +0100 | [diff] [blame] | 875 | DataType data_type, float alpha, float beta, unsigned int m_h, const ActivationLayerInfo &act_info) |
Gian Marco Iodice | 9382ab3 | 2018-12-17 15:12:07 +0000 | [diff] [blame] | 876 | { |
| 877 | // Create tensors |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 878 | TensorType lhs = create_tensor<TensorType>(lhs_shape, data_type, 1); |
| 879 | TensorType rhs = create_tensor<TensorType>(rhs_shape, data_type, 1); |
| 880 | TensorType bias = create_tensor<TensorType>(bias_shape, data_type, 1); |
Gian Marco Iodice | 9382ab3 | 2018-12-17 15:12:07 +0000 | [diff] [blame] | 881 | TensorType lhs_reshaped; |
| 882 | TensorType rhs_reshaped; |
| 883 | TensorType dst; |
| 884 | |
| 885 | const unsigned int M = lhs_shape[1]; |
| 886 | const unsigned int N = rhs_shape[0]; |
| 887 | const unsigned int K = lhs_shape[0]; |
Gian Marco Iodice | 7026b30 | 2019-06-26 17:18:11 +0100 | [diff] [blame] | 888 | GEMMKernelInfo kernel_info; |
| 889 | kernel_info.m = M; |
| 890 | kernel_info.n = N; |
| 891 | kernel_info.k = K; |
| 892 | kernel_info.depth_output_gemm3d = m_h; |
| 893 | kernel_info.reinterpret_input_as_3d = false; |
| 894 | kernel_info.broadcast_bias = true; |
Gian Marco Iodice | ca1f460 | 2019-07-16 15:46:48 +0100 | [diff] [blame] | 895 | kernel_info.activation_info = act_info; |
Gian Marco Iodice | 0c17aa2 | 2019-09-27 09:23:15 +0100 | [diff] [blame] | 896 | kernel_info.fp_mixed_precision = fp_mixed_precision; |
Gian Marco Iodice | 9382ab3 | 2018-12-17 15:12:07 +0000 | [diff] [blame] | 897 | |
| 898 | // The output tensor will be auto-initialized within the function |
| 899 | |
| 900 | // Create and configure function |
| 901 | ReshapeLHSFunctionType reshape_lhs; |
| 902 | ReshapeRHSFunctionType reshape_rhs; |
| 903 | GEMMFunctionType gemm; |
| 904 | reshape_lhs.configure(&lhs, &lhs_reshaped, lhs_info); |
| 905 | reshape_rhs.configure(&rhs, &rhs_reshaped, rhs_info); |
Gian Marco Iodice | 7026b30 | 2019-06-26 17:18:11 +0100 | [diff] [blame] | 906 | gemm.configure(&lhs_reshaped, &rhs_reshaped, &bias, &dst, alpha, beta, lhs_info, rhs_info, kernel_info); |
Gian Marco Iodice | 9382ab3 | 2018-12-17 15:12:07 +0000 | [diff] [blame] | 907 | |
| 908 | ARM_COMPUTE_EXPECT(lhs.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 909 | ARM_COMPUTE_EXPECT(rhs.info()->is_resizable(), framework::LogLevel::ERRORS); |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 910 | ARM_COMPUTE_EXPECT(bias.info()->is_resizable(), framework::LogLevel::ERRORS); |
Gian Marco Iodice | 9382ab3 | 2018-12-17 15:12:07 +0000 | [diff] [blame] | 911 | |
| 912 | // Allocate tensors |
| 913 | lhs.allocator()->allocate(); |
| 914 | rhs.allocator()->allocate(); |
| 915 | lhs_reshaped.allocator()->allocate(); |
| 916 | rhs_reshaped.allocator()->allocate(); |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 917 | bias.allocator()->allocate(); |
Gian Marco Iodice | 9382ab3 | 2018-12-17 15:12:07 +0000 | [diff] [blame] | 918 | dst.allocator()->allocate(); |
| 919 | |
| 920 | ARM_COMPUTE_EXPECT(!lhs.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 921 | ARM_COMPUTE_EXPECT(!rhs.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 922 | ARM_COMPUTE_EXPECT(!lhs_reshaped.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 923 | ARM_COMPUTE_EXPECT(!rhs_reshaped.info()->is_resizable(), framework::LogLevel::ERRORS); |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 924 | ARM_COMPUTE_EXPECT(!bias.info()->is_resizable(), framework::LogLevel::ERRORS); |
Gian Marco Iodice | 9382ab3 | 2018-12-17 15:12:07 +0000 | [diff] [blame] | 925 | ARM_COMPUTE_EXPECT(!dst.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 926 | |
| 927 | // Fill tensors |
| 928 | fill(AccessorType(lhs), 0); |
| 929 | fill(AccessorType(rhs), 1); |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 930 | fill(AccessorType(bias), 2); |
Gian Marco Iodice | 9382ab3 | 2018-12-17 15:12:07 +0000 | [diff] [blame] | 931 | |
| 932 | // Compute GEMM |
| 933 | reshape_lhs.run(); |
| 934 | reshape_rhs.run(); |
| 935 | gemm.run(); |
| 936 | |
| 937 | return dst; |
| 938 | } |
| 939 | |
Michalis Spyrou | 6bff195 | 2019-10-02 17:22:11 +0100 | [diff] [blame] | 940 | SimpleTensor<T> compute_reference(const TensorShape &lhs_shape, const TensorShape &rhs_shape, DataType data_type, float alpha, float beta, unsigned int m_h, |
Gian Marco Iodice | ca1f460 | 2019-07-16 15:46:48 +0100 | [diff] [blame] | 941 | const ActivationLayerInfo &act_info) |
Gian Marco Iodice | 9382ab3 | 2018-12-17 15:12:07 +0000 | [diff] [blame] | 942 | { |
| 943 | TensorShape dst_shape = lhs_shape; |
| 944 | dst_shape.set(0, rhs_shape[0]); |
| 945 | dst_shape.set(1, lhs_shape[1] / m_h); |
| 946 | dst_shape.set(2, m_h); |
| 947 | dst_shape.set(3, lhs_shape[2]); |
| 948 | |
| 949 | // Create reference |
| 950 | SimpleTensor<T> lhs{ lhs_shape, data_type, 1 }; |
| 951 | SimpleTensor<T> rhs{ rhs_shape, data_type, 1 }; |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 952 | SimpleTensor<T> bias{ dst_shape, data_type, 1 }; |
| 953 | |
| 954 | const int n = rhs_shape[0]; |
| 955 | const int m = lhs_shape[1]; |
| 956 | const int batch_size = lhs_shape[2]; |
Gian Marco Iodice | 9382ab3 | 2018-12-17 15:12:07 +0000 | [diff] [blame] | 957 | |
| 958 | // Fill reference |
| 959 | fill(lhs, 0); |
| 960 | fill(rhs, 1); |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 961 | fill(bias, 2); |
Gian Marco Iodice | 9382ab3 | 2018-12-17 15:12:07 +0000 | [diff] [blame] | 962 | |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 963 | // In case of broadcast, we need simply copy the first into the following "M" ones |
| 964 | for(int i = 1; i < m * batch_size; i++) |
| 965 | { |
| 966 | memcpy(bias.data() + i * n, bias.data(), n * sizeof(T)); |
| 967 | } |
| 968 | |
Gian Marco Iodice | 0c17aa2 | 2019-09-27 09:23:15 +0100 | [diff] [blame] | 969 | if(fp_mixed_precision) |
| 970 | { |
| 971 | return reference::activation_layer(reference::gemm_mixed_precision<T>(lhs, rhs, bias, alpha, beta), act_info); |
| 972 | } |
| 973 | else |
| 974 | { |
| 975 | return reference::activation_layer(reference::gemm<T>(lhs, rhs, bias, alpha, beta), act_info); |
| 976 | } |
Gian Marco Iodice | 9382ab3 | 2018-12-17 15:12:07 +0000 | [diff] [blame] | 977 | } |
| 978 | |
| 979 | TensorType _target{}; |
| 980 | SimpleTensor<T> _reference{}; |
Gian Marco Iodice | bf9731e | 2018-12-12 10:18:04 +0000 | [diff] [blame] | 981 | }; |
Gian Marco Iodice | adc5395 | 2019-02-15 11:10:31 +0000 | [diff] [blame] | 982 | |
| 983 | template <typename TensorType, typename AccessorType, typename T, typename ReshapeRHSFunctionType, typename GEMMFunctionType> |
| 984 | class GEMMMatrixMultiplyReshapedOnlyRHSValidationFixture : public framework::Fixture |
| 985 | { |
| 986 | public: |
| 987 | template <typename...> |
| 988 | void setup(unsigned int m, unsigned int n, unsigned int k, unsigned int batch_size, unsigned int m0, unsigned int n0, unsigned int k0, unsigned int h0, |
Gian Marco Iodice | 781cba7 | 2020-06-19 16:56:57 +0100 | [diff] [blame] | 989 | bool interleave_rhs, bool transpose_rhs, bool export_to_cl_image, DataType data_type, float alpha, float beta, bool broadcast_bias, const ActivationLayerInfo &act_info) |
Gian Marco Iodice | adc5395 | 2019-02-15 11:10:31 +0000 | [diff] [blame] | 990 | { |
| 991 | GEMMLHSMatrixInfo lhs_info; |
| 992 | lhs_info.m0 = m0; |
| 993 | lhs_info.k0 = k0; |
| 994 | |
| 995 | GEMMRHSMatrixInfo rhs_info; |
Gian Marco Iodice | 781cba7 | 2020-06-19 16:56:57 +0100 | [diff] [blame] | 996 | rhs_info.n0 = n0; |
| 997 | rhs_info.k0 = k0; |
| 998 | rhs_info.h0 = h0; |
| 999 | rhs_info.interleave = interleave_rhs; |
| 1000 | rhs_info.transpose = transpose_rhs; |
| 1001 | rhs_info.export_to_cl_image = export_to_cl_image; |
Gian Marco Iodice | adc5395 | 2019-02-15 11:10:31 +0000 | [diff] [blame] | 1002 | |
| 1003 | // Set the tensor shapes for LHS and RHS matrices |
| 1004 | const TensorShape lhs_shape(k, m, batch_size); |
| 1005 | const TensorShape rhs_shape(n, k, batch_size); |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 1006 | const TensorShape bias_shape(n, |
| 1007 | broadcast_bias ? 1 : m, |
| 1008 | broadcast_bias ? 1 : batch_size); |
Georgios Pinitas | b0f342e | 2019-05-21 13:32:43 +0100 | [diff] [blame] | 1009 | |
Gian Marco Iodice | ca1f460 | 2019-07-16 15:46:48 +0100 | [diff] [blame] | 1010 | _target = compute_target(lhs_shape, rhs_shape, bias_shape, lhs_info, rhs_info, data_type, alpha, beta, broadcast_bias, act_info); |
Michalis Spyrou | 6bff195 | 2019-10-02 17:22:11 +0100 | [diff] [blame] | 1011 | _reference = compute_reference(lhs_shape, rhs_shape, data_type, alpha, beta, broadcast_bias, act_info); |
Gian Marco Iodice | adc5395 | 2019-02-15 11:10:31 +0000 | [diff] [blame] | 1012 | } |
| 1013 | |
| 1014 | protected: |
| 1015 | template <typename U> |
| 1016 | void fill(U &&tensor, int i) |
| 1017 | { |
| 1018 | std::uniform_real_distribution<> distribution(-1.0f, 1.0f); |
| 1019 | library->fill(tensor, distribution, i); |
| 1020 | |
| 1021 | // Fill border with infinity in order to check the presence of NaN values (i.e. inf * 0) |
| 1022 | std::uniform_real_distribution<> distribution_inf(std::numeric_limits<float>::infinity(), std::numeric_limits<float>::infinity()); |
| 1023 | library->fill_borders_with_garbage(tensor, distribution_inf, i); |
| 1024 | } |
| 1025 | |
Georgios Pinitas | b0f342e | 2019-05-21 13:32:43 +0100 | [diff] [blame] | 1026 | TensorType compute_target(const TensorShape &lhs_shape, const TensorShape &rhs_shape, const TensorShape &bias_shape, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, |
Gian Marco Iodice | ca1f460 | 2019-07-16 15:46:48 +0100 | [diff] [blame] | 1027 | DataType data_type, float alpha, float beta, bool broadcast_bias, const ActivationLayerInfo &act_info) |
Gian Marco Iodice | adc5395 | 2019-02-15 11:10:31 +0000 | [diff] [blame] | 1028 | { |
| 1029 | // Create tensors |
Georgios Pinitas | b0f342e | 2019-05-21 13:32:43 +0100 | [diff] [blame] | 1030 | TensorType lhs = create_tensor<TensorType>(lhs_shape, data_type, 1); |
| 1031 | TensorType rhs = create_tensor<TensorType>(rhs_shape, data_type, 1); |
| 1032 | TensorType bias = create_tensor<TensorType>(bias_shape, data_type, 1); |
Gian Marco Iodice | adc5395 | 2019-02-15 11:10:31 +0000 | [diff] [blame] | 1033 | TensorType rhs_reshaped; |
| 1034 | TensorType dst; |
| 1035 | |
| 1036 | const unsigned int M = lhs_shape[1]; |
| 1037 | const unsigned int N = rhs_shape[0]; |
| 1038 | const unsigned int K = lhs_shape[0]; |
Gian Marco Iodice | 7026b30 | 2019-06-26 17:18:11 +0100 | [diff] [blame] | 1039 | GEMMKernelInfo kernel_info; |
| 1040 | kernel_info.m = M; |
| 1041 | kernel_info.n = N; |
| 1042 | kernel_info.k = K; |
| 1043 | kernel_info.depth_output_gemm3d = 0; |
| 1044 | kernel_info.reinterpret_input_as_3d = false; |
| 1045 | kernel_info.broadcast_bias = broadcast_bias; |
Gian Marco Iodice | ca1f460 | 2019-07-16 15:46:48 +0100 | [diff] [blame] | 1046 | kernel_info.activation_info = act_info; |
Gian Marco Iodice | adc5395 | 2019-02-15 11:10:31 +0000 | [diff] [blame] | 1047 | |
| 1048 | // The output tensor will be auto-initialized within the function |
| 1049 | |
| 1050 | // Create and configure function |
| 1051 | ReshapeRHSFunctionType reshape_rhs; |
| 1052 | GEMMFunctionType gemm; |
| 1053 | reshape_rhs.configure(&rhs, &rhs_reshaped, rhs_info); |
Gian Marco Iodice | 7026b30 | 2019-06-26 17:18:11 +0100 | [diff] [blame] | 1054 | gemm.configure(&lhs, &rhs_reshaped, &bias, &dst, alpha, beta, lhs_info, rhs_info, kernel_info); |
Gian Marco Iodice | adc5395 | 2019-02-15 11:10:31 +0000 | [diff] [blame] | 1055 | |
| 1056 | ARM_COMPUTE_EXPECT(lhs.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 1057 | ARM_COMPUTE_EXPECT(rhs.info()->is_resizable(), framework::LogLevel::ERRORS); |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 1058 | ARM_COMPUTE_EXPECT(bias.info()->is_resizable(), framework::LogLevel::ERRORS); |
Gian Marco Iodice | adc5395 | 2019-02-15 11:10:31 +0000 | [diff] [blame] | 1059 | |
| 1060 | // Allocate tensors |
| 1061 | lhs.allocator()->allocate(); |
| 1062 | rhs.allocator()->allocate(); |
| 1063 | rhs_reshaped.allocator()->allocate(); |
Georgios Pinitas | b0f342e | 2019-05-21 13:32:43 +0100 | [diff] [blame] | 1064 | bias.allocator()->allocate(); |
Gian Marco Iodice | adc5395 | 2019-02-15 11:10:31 +0000 | [diff] [blame] | 1065 | dst.allocator()->allocate(); |
| 1066 | |
| 1067 | ARM_COMPUTE_EXPECT(!lhs.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 1068 | ARM_COMPUTE_EXPECT(!rhs.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 1069 | ARM_COMPUTE_EXPECT(!rhs_reshaped.info()->is_resizable(), framework::LogLevel::ERRORS); |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 1070 | ARM_COMPUTE_EXPECT(!bias.info()->is_resizable(), framework::LogLevel::ERRORS); |
Gian Marco Iodice | adc5395 | 2019-02-15 11:10:31 +0000 | [diff] [blame] | 1071 | ARM_COMPUTE_EXPECT(!dst.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 1072 | |
| 1073 | // Fill tensors |
| 1074 | fill(AccessorType(lhs), 0); |
| 1075 | fill(AccessorType(rhs), 1); |
Georgios Pinitas | b0f342e | 2019-05-21 13:32:43 +0100 | [diff] [blame] | 1076 | fill(AccessorType(bias), 2); |
Gian Marco Iodice | adc5395 | 2019-02-15 11:10:31 +0000 | [diff] [blame] | 1077 | |
| 1078 | // Compute GEMM |
| 1079 | reshape_rhs.run(); |
| 1080 | gemm.run(); |
| 1081 | |
| 1082 | return dst; |
| 1083 | } |
| 1084 | |
Michalis Spyrou | 6bff195 | 2019-10-02 17:22:11 +0100 | [diff] [blame] | 1085 | SimpleTensor<T> compute_reference(const TensorShape &lhs_shape, const TensorShape &rhs_shape, DataType data_type, float alpha, float beta, bool broadcast_bias, |
Gian Marco Iodice | ca1f460 | 2019-07-16 15:46:48 +0100 | [diff] [blame] | 1086 | const ActivationLayerInfo &act_info) |
Gian Marco Iodice | adc5395 | 2019-02-15 11:10:31 +0000 | [diff] [blame] | 1087 | { |
| 1088 | TensorShape dst_shape = lhs_shape; |
| 1089 | dst_shape[0] = rhs_shape[0]; |
| 1090 | dst_shape[1] = lhs_shape[1]; |
| 1091 | |
| 1092 | // Create reference |
| 1093 | SimpleTensor<T> lhs{ lhs_shape, data_type, 1 }; |
| 1094 | SimpleTensor<T> rhs{ rhs_shape, data_type, 1 }; |
Georgios Pinitas | b0f342e | 2019-05-21 13:32:43 +0100 | [diff] [blame] | 1095 | SimpleTensor<T> bias{ dst_shape, data_type, 1 }; |
| 1096 | |
| 1097 | const int n = rhs_shape[0]; |
| 1098 | const int m = lhs_shape[1]; |
| 1099 | const int batch_size = lhs_shape[2]; |
Gian Marco Iodice | adc5395 | 2019-02-15 11:10:31 +0000 | [diff] [blame] | 1100 | |
| 1101 | // Fill reference |
| 1102 | fill(lhs, 0); |
| 1103 | fill(rhs, 1); |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 1104 | fill(bias, 2); |
Gian Marco Iodice | adc5395 | 2019-02-15 11:10:31 +0000 | [diff] [blame] | 1105 | |
Georgios Pinitas | b0f342e | 2019-05-21 13:32:43 +0100 | [diff] [blame] | 1106 | if(broadcast_bias) |
| 1107 | { |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 1108 | // In case of broadcast, we need simply copy the first into the following "M" ones |
| 1109 | for(int i = 1; i < m * batch_size; i++) |
Georgios Pinitas | b0f342e | 2019-05-21 13:32:43 +0100 | [diff] [blame] | 1110 | { |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 1111 | memcpy(bias.data() + i * n, bias.data(), n * sizeof(T)); |
Georgios Pinitas | b0f342e | 2019-05-21 13:32:43 +0100 | [diff] [blame] | 1112 | } |
| 1113 | } |
Georgios Pinitas | b0f342e | 2019-05-21 13:32:43 +0100 | [diff] [blame] | 1114 | |
Gian Marco Iodice | ca1f460 | 2019-07-16 15:46:48 +0100 | [diff] [blame] | 1115 | return reference::activation_layer(reference::gemm<T>(lhs, rhs, bias, alpha, beta), act_info); |
Gian Marco Iodice | adc5395 | 2019-02-15 11:10:31 +0000 | [diff] [blame] | 1116 | } |
| 1117 | |
| 1118 | TensorType _target{}; |
| 1119 | SimpleTensor<T> _reference{}; |
| 1120 | }; |
| 1121 | |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 1122 | template <typename TensorType, typename AccessorType, typename T, typename ReshapeRHSFunctionType, typename GEMMFunctionType> |
| 1123 | class GEMMMatrixMultiplyReshapedOnlyRHS3DValidationFixture : public framework::Fixture |
| 1124 | { |
| 1125 | public: |
| 1126 | template <typename...> |
| 1127 | void setup(unsigned int m_w, unsigned int m_h, unsigned int n, unsigned int k, unsigned int batch_size, unsigned int m0, unsigned int n0, unsigned int k0, unsigned int h0, |
Gian Marco Iodice | 9ae06d4 | 2020-10-22 16:37:12 +0100 | [diff] [blame^] | 1128 | bool interleave_rhs, bool transpose_rhs, bool export_to_cl_image, bool has_pad_y, DataType data_type, float alpha, float beta, const ActivationLayerInfo &act_info) |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 1129 | { |
| 1130 | GEMMLHSMatrixInfo lhs_info; |
| 1131 | lhs_info.m0 = m0; |
| 1132 | lhs_info.k0 = k0; |
| 1133 | |
| 1134 | GEMMRHSMatrixInfo rhs_info; |
Gian Marco Iodice | 781cba7 | 2020-06-19 16:56:57 +0100 | [diff] [blame] | 1135 | rhs_info.n0 = n0; |
| 1136 | rhs_info.k0 = k0; |
| 1137 | rhs_info.h0 = h0; |
| 1138 | rhs_info.interleave = interleave_rhs; |
| 1139 | rhs_info.transpose = transpose_rhs; |
| 1140 | rhs_info.export_to_cl_image = export_to_cl_image; |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 1141 | |
| 1142 | // In case of GEMM3D, m is the product between m_w and m_h |
| 1143 | const unsigned int m = m_w * m_h; |
| 1144 | |
| 1145 | // Set the tensor shapes for LHS and RHS matrices |
| 1146 | const TensorShape lhs_shape(k, m, batch_size); |
| 1147 | const TensorShape rhs_shape(n, k, batch_size); |
| 1148 | const TensorShape bias_shape(n, 1, 1); |
| 1149 | |
Gian Marco Iodice | 9ae06d4 | 2020-10-22 16:37:12 +0100 | [diff] [blame^] | 1150 | _target = compute_target(lhs_shape, rhs_shape, bias_shape, lhs_info, rhs_info, data_type, alpha, beta, m_h, act_info, has_pad_y); |
Michalis Spyrou | 6bff195 | 2019-10-02 17:22:11 +0100 | [diff] [blame] | 1151 | _reference = compute_reference(lhs_shape, rhs_shape, data_type, alpha, beta, m_h, act_info); |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 1152 | } |
| 1153 | |
| 1154 | protected: |
| 1155 | template <typename U> |
| 1156 | void fill(U &&tensor, int i) |
| 1157 | { |
| 1158 | std::uniform_real_distribution<> distribution(-1.0f, 1.0f); |
| 1159 | library->fill(tensor, distribution, i); |
| 1160 | } |
| 1161 | |
| 1162 | TensorType compute_target(const TensorShape &lhs_shape, const TensorShape &rhs_shape, const TensorShape &bias_shape, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, |
| 1163 | DataType data_type, float alpha, float beta, |
Gian Marco Iodice | 9ae06d4 | 2020-10-22 16:37:12 +0100 | [diff] [blame^] | 1164 | unsigned int m_h, const ActivationLayerInfo &act_info, bool has_pad_y) |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 1165 | { |
| 1166 | // Create tensors |
| 1167 | TensorType lhs = create_tensor<TensorType>(lhs_shape, data_type, 1); |
| 1168 | TensorType rhs = create_tensor<TensorType>(rhs_shape, data_type, 1); |
| 1169 | TensorType bias = create_tensor<TensorType>(bias_shape, data_type, 1); |
| 1170 | TensorType rhs_reshaped; |
| 1171 | TensorType dst; |
| 1172 | |
| 1173 | const unsigned int M = lhs_shape[1]; |
| 1174 | const unsigned int N = rhs_shape[0]; |
| 1175 | const unsigned int K = lhs_shape[0]; |
Gian Marco Iodice | 7026b30 | 2019-06-26 17:18:11 +0100 | [diff] [blame] | 1176 | GEMMKernelInfo kernel_info; |
| 1177 | kernel_info.m = M; |
| 1178 | kernel_info.n = N; |
| 1179 | kernel_info.k = K; |
| 1180 | kernel_info.depth_output_gemm3d = m_h; |
| 1181 | kernel_info.reinterpret_input_as_3d = false; |
| 1182 | kernel_info.broadcast_bias = true; |
Gian Marco Iodice | ca1f460 | 2019-07-16 15:46:48 +0100 | [diff] [blame] | 1183 | kernel_info.activation_info = act_info; |
Gian Marco Iodice | 9ae06d4 | 2020-10-22 16:37:12 +0100 | [diff] [blame^] | 1184 | kernel_info.has_pad_y = has_pad_y; |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 1185 | |
| 1186 | // The output tensor will be auto-initialized within the function |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 1187 | // Create and configure function |
| 1188 | ReshapeRHSFunctionType reshape_rhs; |
| 1189 | GEMMFunctionType gemm; |
| 1190 | reshape_rhs.configure(&rhs, &rhs_reshaped, rhs_info); |
Gian Marco Iodice | 7026b30 | 2019-06-26 17:18:11 +0100 | [diff] [blame] | 1191 | gemm.configure(&lhs, &rhs_reshaped, &bias, &dst, alpha, beta, lhs_info, rhs_info, kernel_info); |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 1192 | |
Gian Marco Iodice | 9ae06d4 | 2020-10-22 16:37:12 +0100 | [diff] [blame^] | 1193 | if(has_pad_y) |
| 1194 | { |
| 1195 | // Add dummy padding into lhs to validate has_pad_y path |
| 1196 | lhs.info()->extend_padding(PaddingSize(2, 0, 2, 0)); |
| 1197 | dst.info()->extend_padding(PaddingSize(2, 0, 1, 0)); |
| 1198 | } |
| 1199 | |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 1200 | ARM_COMPUTE_EXPECT(lhs.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 1201 | ARM_COMPUTE_EXPECT(rhs.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 1202 | ARM_COMPUTE_EXPECT(bias.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 1203 | |
| 1204 | // Allocate tensors |
| 1205 | lhs.allocator()->allocate(); |
| 1206 | rhs.allocator()->allocate(); |
| 1207 | rhs_reshaped.allocator()->allocate(); |
| 1208 | bias.allocator()->allocate(); |
| 1209 | dst.allocator()->allocate(); |
| 1210 | |
| 1211 | ARM_COMPUTE_EXPECT(!lhs.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 1212 | ARM_COMPUTE_EXPECT(!rhs.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 1213 | ARM_COMPUTE_EXPECT(!rhs_reshaped.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 1214 | ARM_COMPUTE_EXPECT(!bias.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 1215 | ARM_COMPUTE_EXPECT(!dst.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 1216 | |
| 1217 | // Fill tensors |
| 1218 | fill(AccessorType(lhs), 0); |
| 1219 | fill(AccessorType(rhs), 1); |
| 1220 | fill(AccessorType(bias), 2); |
| 1221 | |
| 1222 | // Compute GEMM |
| 1223 | reshape_rhs.run(); |
| 1224 | gemm.run(); |
| 1225 | |
| 1226 | return dst; |
| 1227 | } |
| 1228 | |
Michalis Spyrou | 6bff195 | 2019-10-02 17:22:11 +0100 | [diff] [blame] | 1229 | SimpleTensor<T> compute_reference(const TensorShape &lhs_shape, const TensorShape &rhs_shape, DataType data_type, float alpha, float beta, unsigned int m_h, |
Gian Marco Iodice | ca1f460 | 2019-07-16 15:46:48 +0100 | [diff] [blame] | 1230 | const ActivationLayerInfo &act_info) |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 1231 | { |
| 1232 | TensorShape dst_shape = lhs_shape; |
| 1233 | dst_shape.set(0, rhs_shape[0]); |
| 1234 | dst_shape.set(1, lhs_shape[1] / m_h); |
| 1235 | dst_shape.set(2, m_h); |
| 1236 | dst_shape.set(3, lhs_shape[2]); |
| 1237 | |
| 1238 | // Create reference |
| 1239 | SimpleTensor<T> lhs{ lhs_shape, data_type, 1 }; |
| 1240 | SimpleTensor<T> rhs{ rhs_shape, data_type, 1 }; |
| 1241 | SimpleTensor<T> bias{ dst_shape, data_type, 1 }; |
| 1242 | |
| 1243 | const int n = rhs_shape[0]; |
| 1244 | const int m = lhs_shape[1]; |
| 1245 | const int batch_size = lhs_shape[2]; |
| 1246 | |
| 1247 | // Fill reference |
| 1248 | fill(lhs, 0); |
| 1249 | fill(rhs, 1); |
| 1250 | fill(bias, 2); |
| 1251 | |
| 1252 | // In case of broadcast, we need simply copy the first into the following "M" ones |
| 1253 | for(int i = 1; i < m * batch_size; i++) |
| 1254 | { |
| 1255 | memcpy(bias.data() + i * n, bias.data(), n * sizeof(T)); |
| 1256 | } |
| 1257 | |
Gian Marco Iodice | ca1f460 | 2019-07-16 15:46:48 +0100 | [diff] [blame] | 1258 | return reference::activation_layer(reference::gemm<T>(lhs, rhs, bias, alpha, beta), act_info); |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 1259 | } |
| 1260 | |
| 1261 | TensorType _target{}; |
| 1262 | SimpleTensor<T> _reference{}; |
| 1263 | }; |
| 1264 | |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 1265 | template <typename TensorType, typename AccessorType, typename T, typename GEMMFunctionType> |
| 1266 | class GEMMMatrixMultiplyNativeValidationFixture : public framework::Fixture |
| 1267 | { |
| 1268 | public: |
| 1269 | template <typename...> |
Gian Marco Iodice | ca1f460 | 2019-07-16 15:46:48 +0100 | [diff] [blame] | 1270 | void setup(unsigned int m, unsigned int n, unsigned int k, unsigned int batch_size, unsigned int m0, unsigned int n0, unsigned int k0, DataType data_type, float alpha, float beta, bool broadcast_bias, |
| 1271 | const ActivationLayerInfo &act_info) |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 1272 | { |
| 1273 | GEMMLHSMatrixInfo lhs_info; |
| 1274 | lhs_info.m0 = m0; |
| 1275 | lhs_info.k0 = k0; |
| 1276 | |
| 1277 | GEMMRHSMatrixInfo rhs_info; |
| 1278 | rhs_info.n0 = n0; |
| 1279 | rhs_info.k0 = k0; |
| 1280 | |
| 1281 | // Set the tensor shapes for LHS and RHS matrices |
| 1282 | const TensorShape lhs_shape(k, m, batch_size); |
| 1283 | const TensorShape rhs_shape(n, k, batch_size); |
Gian Marco Iodice | 944170e | 2019-06-24 14:40:30 +0100 | [diff] [blame] | 1284 | const TensorShape bias_shape(n, |
| 1285 | broadcast_bias ? 1 : m, |
| 1286 | broadcast_bias ? 1 : batch_size); |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 1287 | |
Gian Marco Iodice | ca1f460 | 2019-07-16 15:46:48 +0100 | [diff] [blame] | 1288 | _target = compute_target(lhs_shape, rhs_shape, bias_shape, lhs_info, rhs_info, data_type, alpha, beta, broadcast_bias, act_info); |
Michalis Spyrou | 6bff195 | 2019-10-02 17:22:11 +0100 | [diff] [blame] | 1289 | _reference = compute_reference(lhs_shape, rhs_shape, data_type, alpha, beta, broadcast_bias, act_info); |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 1290 | } |
| 1291 | |
| 1292 | protected: |
| 1293 | template <typename U> |
| 1294 | void fill(U &&tensor, int i) |
| 1295 | { |
| 1296 | std::uniform_real_distribution<> distribution(-1.0f, 1.0f); |
| 1297 | library->fill(tensor, distribution, i); |
| 1298 | |
| 1299 | // Fill border with infinity in order to check the presence of NaN values (i.e. inf * 0) |
| 1300 | std::uniform_real_distribution<> distribution_inf(std::numeric_limits<float>::infinity(), std::numeric_limits<float>::infinity()); |
| 1301 | library->fill_borders_with_garbage(tensor, distribution_inf, i); |
| 1302 | } |
| 1303 | |
Gian Marco Iodice | 944170e | 2019-06-24 14:40:30 +0100 | [diff] [blame] | 1304 | TensorType compute_target(const TensorShape &lhs_shape, const TensorShape &rhs_shape, const TensorShape &bias_shape, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, |
Gian Marco Iodice | ca1f460 | 2019-07-16 15:46:48 +0100 | [diff] [blame] | 1305 | DataType data_type, float alpha, float beta, bool broadcast_bias, const ActivationLayerInfo &act_info) |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 1306 | { |
| 1307 | // Create tensors |
Gian Marco Iodice | 944170e | 2019-06-24 14:40:30 +0100 | [diff] [blame] | 1308 | TensorType lhs = create_tensor<TensorType>(lhs_shape, data_type, 1); |
| 1309 | TensorType rhs = create_tensor<TensorType>(rhs_shape, data_type, 1); |
| 1310 | TensorType bias = create_tensor<TensorType>(bias_shape, data_type, 1); |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 1311 | TensorType dst; |
| 1312 | |
| 1313 | const unsigned int M = lhs_shape[1]; |
| 1314 | const unsigned int N = rhs_shape[0]; |
| 1315 | const unsigned int K = lhs_shape[0]; |
Gian Marco Iodice | 7026b30 | 2019-06-26 17:18:11 +0100 | [diff] [blame] | 1316 | GEMMKernelInfo kernel_info; |
| 1317 | kernel_info.m = M; |
| 1318 | kernel_info.n = N; |
| 1319 | kernel_info.k = K; |
| 1320 | kernel_info.depth_output_gemm3d = 0; |
| 1321 | kernel_info.reinterpret_input_as_3d = false; |
| 1322 | kernel_info.broadcast_bias = broadcast_bias; |
Gian Marco Iodice | ca1f460 | 2019-07-16 15:46:48 +0100 | [diff] [blame] | 1323 | kernel_info.activation_info = act_info; |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 1324 | |
| 1325 | // Create and configure function |
| 1326 | GEMMFunctionType gemm; |
Gian Marco Iodice | 7026b30 | 2019-06-26 17:18:11 +0100 | [diff] [blame] | 1327 | gemm.configure(&lhs, &rhs, &bias, &dst, alpha, beta, lhs_info, rhs_info, kernel_info); |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 1328 | |
| 1329 | ARM_COMPUTE_EXPECT(lhs.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 1330 | ARM_COMPUTE_EXPECT(rhs.info()->is_resizable(), framework::LogLevel::ERRORS); |
Gian Marco Iodice | 944170e | 2019-06-24 14:40:30 +0100 | [diff] [blame] | 1331 | ARM_COMPUTE_EXPECT(bias.info()->is_resizable(), framework::LogLevel::ERRORS); |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 1332 | |
| 1333 | // Allocate tensors |
| 1334 | lhs.allocator()->allocate(); |
| 1335 | rhs.allocator()->allocate(); |
Gian Marco Iodice | 944170e | 2019-06-24 14:40:30 +0100 | [diff] [blame] | 1336 | bias.allocator()->allocate(); |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 1337 | dst.allocator()->allocate(); |
| 1338 | |
| 1339 | ARM_COMPUTE_EXPECT(!lhs.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 1340 | ARM_COMPUTE_EXPECT(!rhs.info()->is_resizable(), framework::LogLevel::ERRORS); |
Gian Marco Iodice | 944170e | 2019-06-24 14:40:30 +0100 | [diff] [blame] | 1341 | ARM_COMPUTE_EXPECT(!bias.info()->is_resizable(), framework::LogLevel::ERRORS); |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 1342 | ARM_COMPUTE_EXPECT(!dst.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 1343 | |
| 1344 | // Fill tensors |
| 1345 | fill(AccessorType(lhs), 0); |
| 1346 | fill(AccessorType(rhs), 1); |
Gian Marco Iodice | 944170e | 2019-06-24 14:40:30 +0100 | [diff] [blame] | 1347 | fill(AccessorType(bias), 2); |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 1348 | |
| 1349 | // Compute GEMM |
| 1350 | gemm.run(); |
| 1351 | |
| 1352 | return dst; |
| 1353 | } |
| 1354 | |
Michalis Spyrou | 6bff195 | 2019-10-02 17:22:11 +0100 | [diff] [blame] | 1355 | SimpleTensor<T> compute_reference(const TensorShape &lhs_shape, const TensorShape &rhs_shape, DataType data_type, float alpha, float beta, bool broadcast_bias, |
Gian Marco Iodice | ca1f460 | 2019-07-16 15:46:48 +0100 | [diff] [blame] | 1356 | const ActivationLayerInfo &act_info) |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 1357 | { |
| 1358 | TensorShape dst_shape = lhs_shape; |
| 1359 | dst_shape[0] = rhs_shape[0]; |
| 1360 | dst_shape[1] = lhs_shape[1]; |
| 1361 | |
| 1362 | // Create reference |
| 1363 | SimpleTensor<T> lhs{ lhs_shape, data_type, 1 }; |
| 1364 | SimpleTensor<T> rhs{ rhs_shape, data_type, 1 }; |
Gian Marco Iodice | 944170e | 2019-06-24 14:40:30 +0100 | [diff] [blame] | 1365 | SimpleTensor<T> bias{ dst_shape, data_type, 1 }; |
| 1366 | |
| 1367 | const int n = rhs_shape[0]; |
| 1368 | const int m = lhs_shape[1]; |
| 1369 | const int batch_size = lhs_shape[2]; |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 1370 | |
| 1371 | // Fill reference |
| 1372 | fill(lhs, 0); |
| 1373 | fill(rhs, 1); |
Gian Marco Iodice | 944170e | 2019-06-24 14:40:30 +0100 | [diff] [blame] | 1374 | fill(bias, 2); |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 1375 | |
Gian Marco Iodice | 944170e | 2019-06-24 14:40:30 +0100 | [diff] [blame] | 1376 | if(broadcast_bias) |
| 1377 | { |
| 1378 | // In case of broadcast, we need simply copy the first into the following "M" ones |
| 1379 | for(int i = 1; i < m * batch_size; i++) |
| 1380 | { |
| 1381 | memcpy(bias.data() + i * n, bias.data(), n * sizeof(T)); |
| 1382 | } |
| 1383 | } |
| 1384 | |
Gian Marco Iodice | ca1f460 | 2019-07-16 15:46:48 +0100 | [diff] [blame] | 1385 | return reference::activation_layer(reference::gemm<T>(lhs, rhs, bias, alpha, beta), act_info); |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 1386 | } |
| 1387 | |
| 1388 | TensorType _target{}; |
| 1389 | SimpleTensor<T> _reference{}; |
| 1390 | }; |
| 1391 | |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 1392 | template <typename TensorType, typename AccessorType, typename T, typename GEMMFunctionType> |
| 1393 | class GEMMMatrixMultiplyNative3DValidationFixture : public framework::Fixture |
| 1394 | { |
| 1395 | public: |
| 1396 | template <typename...> |
Gian Marco Iodice | ca1f460 | 2019-07-16 15:46:48 +0100 | [diff] [blame] | 1397 | void setup(unsigned int m_w, unsigned int m_h, unsigned int n, unsigned int k, unsigned int batch_size, unsigned int m0, unsigned int n0, unsigned int k0, DataType data_type, float alpha, float beta, |
| 1398 | const ActivationLayerInfo &act_info) |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 1399 | { |
| 1400 | GEMMLHSMatrixInfo lhs_info; |
| 1401 | lhs_info.m0 = m0; |
| 1402 | lhs_info.k0 = k0; |
| 1403 | |
| 1404 | GEMMRHSMatrixInfo rhs_info; |
| 1405 | rhs_info.n0 = n0; |
| 1406 | rhs_info.k0 = k0; |
| 1407 | |
| 1408 | // In case of GEMM3D, m is the product between m_w and m_h |
| 1409 | const unsigned int m = m_w * m_h; |
| 1410 | |
| 1411 | // Set the tensor shapes for LHS and RHS matrices |
| 1412 | const TensorShape lhs_shape(k, m, batch_size); |
| 1413 | const TensorShape rhs_shape(n, k, batch_size); |
Gian Marco Iodice | 944170e | 2019-06-24 14:40:30 +0100 | [diff] [blame] | 1414 | const TensorShape bias_shape(n, 1, 1); |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 1415 | |
Gian Marco Iodice | ca1f460 | 2019-07-16 15:46:48 +0100 | [diff] [blame] | 1416 | _target = compute_target(lhs_shape, rhs_shape, bias_shape, lhs_info, rhs_info, data_type, alpha, beta, m_h, act_info); |
Michalis Spyrou | 6bff195 | 2019-10-02 17:22:11 +0100 | [diff] [blame] | 1417 | _reference = compute_reference(lhs_shape, rhs_shape, data_type, alpha, beta, m_h, act_info); |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 1418 | } |
| 1419 | |
| 1420 | protected: |
| 1421 | template <typename U> |
| 1422 | void fill(U &&tensor, int i) |
| 1423 | { |
| 1424 | std::uniform_real_distribution<> distribution(-1.0f, 1.0f); |
| 1425 | library->fill(tensor, distribution, i); |
| 1426 | } |
| 1427 | |
Gian Marco Iodice | 944170e | 2019-06-24 14:40:30 +0100 | [diff] [blame] | 1428 | TensorType compute_target(const TensorShape &lhs_shape, const TensorShape &rhs_shape, const TensorShape &bias_shape, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, |
Gian Marco Iodice | ca1f460 | 2019-07-16 15:46:48 +0100 | [diff] [blame] | 1429 | DataType data_type, float alpha, float beta, unsigned int m_h, const ActivationLayerInfo &act_info) |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 1430 | { |
| 1431 | // Create tensors |
Gian Marco Iodice | 944170e | 2019-06-24 14:40:30 +0100 | [diff] [blame] | 1432 | TensorType lhs = create_tensor<TensorType>(lhs_shape, data_type, 1); |
| 1433 | TensorType rhs = create_tensor<TensorType>(rhs_shape, data_type, 1); |
| 1434 | TensorType bias = create_tensor<TensorType>(bias_shape, data_type, 1); |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 1435 | TensorType dst; |
| 1436 | |
| 1437 | const unsigned int M = lhs_shape[1]; |
| 1438 | const unsigned int N = rhs_shape[0]; |
| 1439 | const unsigned int K = lhs_shape[0]; |
Gian Marco Iodice | 7026b30 | 2019-06-26 17:18:11 +0100 | [diff] [blame] | 1440 | GEMMKernelInfo kernel_info; |
| 1441 | kernel_info.m = M; |
| 1442 | kernel_info.n = N; |
| 1443 | kernel_info.k = K; |
| 1444 | kernel_info.depth_output_gemm3d = m_h; |
| 1445 | kernel_info.reinterpret_input_as_3d = false; |
| 1446 | kernel_info.broadcast_bias = true; |
Gian Marco Iodice | ca1f460 | 2019-07-16 15:46:48 +0100 | [diff] [blame] | 1447 | kernel_info.activation_info = act_info; |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 1448 | |
| 1449 | // The output tensor will be auto-initialized within the function |
| 1450 | |
| 1451 | // Create and configure function |
| 1452 | GEMMFunctionType gemm; |
Gian Marco Iodice | 7026b30 | 2019-06-26 17:18:11 +0100 | [diff] [blame] | 1453 | gemm.configure(&lhs, &rhs, &bias, &dst, alpha, beta, lhs_info, rhs_info, kernel_info); |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 1454 | |
| 1455 | ARM_COMPUTE_EXPECT(lhs.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 1456 | ARM_COMPUTE_EXPECT(rhs.info()->is_resizable(), framework::LogLevel::ERRORS); |
Gian Marco Iodice | 944170e | 2019-06-24 14:40:30 +0100 | [diff] [blame] | 1457 | ARM_COMPUTE_EXPECT(bias.info()->is_resizable(), framework::LogLevel::ERRORS); |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 1458 | |
| 1459 | // Allocate tensors |
| 1460 | lhs.allocator()->allocate(); |
| 1461 | rhs.allocator()->allocate(); |
Gian Marco Iodice | 944170e | 2019-06-24 14:40:30 +0100 | [diff] [blame] | 1462 | bias.allocator()->allocate(); |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 1463 | dst.allocator()->allocate(); |
| 1464 | |
| 1465 | ARM_COMPUTE_EXPECT(!lhs.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 1466 | ARM_COMPUTE_EXPECT(!rhs.info()->is_resizable(), framework::LogLevel::ERRORS); |
Gian Marco Iodice | 944170e | 2019-06-24 14:40:30 +0100 | [diff] [blame] | 1467 | ARM_COMPUTE_EXPECT(!bias.info()->is_resizable(), framework::LogLevel::ERRORS); |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 1468 | ARM_COMPUTE_EXPECT(!dst.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 1469 | |
| 1470 | // Fill tensors |
| 1471 | fill(AccessorType(lhs), 0); |
| 1472 | fill(AccessorType(rhs), 1); |
Gian Marco Iodice | 944170e | 2019-06-24 14:40:30 +0100 | [diff] [blame] | 1473 | fill(AccessorType(bias), 2); |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 1474 | |
| 1475 | // Compute GEMM |
| 1476 | gemm.run(); |
| 1477 | |
| 1478 | return dst; |
| 1479 | } |
| 1480 | |
Michalis Spyrou | 6bff195 | 2019-10-02 17:22:11 +0100 | [diff] [blame] | 1481 | SimpleTensor<T> compute_reference(const TensorShape &lhs_shape, const TensorShape &rhs_shape, DataType data_type, float alpha, float beta, unsigned int m_h, |
Gian Marco Iodice | ca1f460 | 2019-07-16 15:46:48 +0100 | [diff] [blame] | 1482 | const ActivationLayerInfo &act_info) |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 1483 | { |
| 1484 | TensorShape dst_shape = lhs_shape; |
| 1485 | dst_shape.set(0, rhs_shape[0]); |
| 1486 | dst_shape.set(1, lhs_shape[1] / m_h); |
| 1487 | dst_shape.set(2, m_h); |
| 1488 | dst_shape.set(3, lhs_shape[2]); |
| 1489 | |
| 1490 | // Create reference |
| 1491 | SimpleTensor<T> lhs{ lhs_shape, data_type, 1 }; |
| 1492 | SimpleTensor<T> rhs{ rhs_shape, data_type, 1 }; |
Gian Marco Iodice | 944170e | 2019-06-24 14:40:30 +0100 | [diff] [blame] | 1493 | SimpleTensor<T> bias{ dst_shape, data_type, 1 }; |
| 1494 | |
| 1495 | const int n = rhs_shape[0]; |
| 1496 | const int m = lhs_shape[1]; |
| 1497 | const int batch_size = lhs_shape[2]; |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 1498 | |
| 1499 | // Fill reference |
| 1500 | fill(lhs, 0); |
| 1501 | fill(rhs, 1); |
Gian Marco Iodice | 944170e | 2019-06-24 14:40:30 +0100 | [diff] [blame] | 1502 | fill(bias, 2); |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 1503 | |
Gian Marco Iodice | 944170e | 2019-06-24 14:40:30 +0100 | [diff] [blame] | 1504 | // In case of broadcast, we need simply copy the first into the following "M" ones |
| 1505 | for(int i = 1; i < m * batch_size; i++) |
| 1506 | { |
| 1507 | memcpy(bias.data() + i * n, bias.data(), n * sizeof(T)); |
| 1508 | } |
| 1509 | |
Gian Marco Iodice | ca1f460 | 2019-07-16 15:46:48 +0100 | [diff] [blame] | 1510 | return reference::activation_layer(reference::gemm<T>(lhs, rhs, bias, alpha, beta), act_info); |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 1511 | } |
| 1512 | |
| 1513 | TensorType _target{}; |
| 1514 | SimpleTensor<T> _reference{}; |
| 1515 | }; |
| 1516 | |
Moritz Pflanzer | 4dfc235 | 2017-08-02 14:51:36 +0100 | [diff] [blame] | 1517 | } // namespace validation |
| 1518 | } // namespace test |
| 1519 | } // namespace arm_compute |
| 1520 | #endif /* ARM_COMPUTE_TEST_GEMM_FIXTURE */ |