Moritz Pflanzer | 4dfc235 | 2017-08-02 14:51:36 +0100 | [diff] [blame] | 1 | /* |
Gian Marco Iodice | 10e88a7 | 2021-11-29 12:49:19 +0000 | [diff] [blame] | 2 | * Copyright (c) 2017-2022 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" |
SiCongLi | 1af5416 | 2021-10-06 15:25:57 +0100 | [diff] [blame] | 30 | #include "arm_compute/core/experimental/IPostOp.h" |
SiCongLi | 3177861 | 2021-11-12 17:33:45 +0000 | [diff] [blame] | 31 | #include "src/core/experimental/PostOpUtils.h" |
Moritz Pflanzer | 4dfc235 | 2017-08-02 14:51:36 +0100 | [diff] [blame] | 32 | #include "tests/AssetsLibrary.h" |
| 33 | #include "tests/Globals.h" |
| 34 | #include "tests/IAccessor.h" |
Moritz Pflanzer | a09de0c | 2017-09-01 20:41:12 +0100 | [diff] [blame] | 35 | #include "tests/framework/Asserts.h" |
| 36 | #include "tests/framework/Fixture.h" |
Moritz Pflanzer | a09de0c | 2017-09-01 20:41:12 +0100 | [diff] [blame] | 37 | #include "tests/validation/Helpers.h" |
Gian Marco Iodice | ca1f460 | 2019-07-16 15:46:48 +0100 | [diff] [blame] | 38 | #include "tests/validation/reference/ActivationLayer.h" |
SiCongLi | 1af5416 | 2021-10-06 15:25:57 +0100 | [diff] [blame] | 39 | #include "tests/validation/reference/ElementwiseOperations.h" |
Georgios Pinitas | 5a7e776 | 2017-12-01 16:27:29 +0000 | [diff] [blame] | 40 | #include "tests/validation/reference/GEMM.h" |
SiCongLi | 1af5416 | 2021-10-06 15:25:57 +0100 | [diff] [blame] | 41 | #include "tests/validation/reference/PostOps.h" |
Moritz Pflanzer | 4dfc235 | 2017-08-02 14:51:36 +0100 | [diff] [blame] | 42 | |
| 43 | #include <random> |
| 44 | |
| 45 | namespace arm_compute |
| 46 | { |
| 47 | namespace test |
| 48 | { |
| 49 | namespace validation |
| 50 | { |
Mohammed Suhail Munshi | 13a2d00 | 2022-09-05 11:57:34 +0100 | [diff] [blame] | 51 | 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, bool pretranspose_a = false, bool pretranspose_b = false, bool run_twice = false> |
Gian Marco Iodice | 68a3f56 | 2018-07-26 11:44:03 +0100 | [diff] [blame] | 52 | class GEMMValidationFixture : public framework::Fixture |
Moritz Pflanzer | 4dfc235 | 2017-08-02 14:51:36 +0100 | [diff] [blame] | 53 | { |
| 54 | public: |
| 55 | template <typename...> |
Pablo Tello | 0e37b5c | 2018-10-30 11:18:37 +0000 | [diff] [blame] | 56 | 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] | 57 | { |
Michalis Spyrou | 6bff195 | 2019-10-02 17:22:11 +0100 | [diff] [blame] | 58 | ARM_COMPUTE_UNUSED(pretranspose); |
| 59 | _target = compute_target(shape_a, shape_b, shape_c, output_shape, alpha, beta, data_type); |
| 60 | _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] | 61 | } |
| 62 | |
| 63 | protected: |
| 64 | template <typename U> |
Pablo Tello | 0e37b5c | 2018-10-30 11:18:37 +0000 | [diff] [blame] | 65 | 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] | 66 | { |
| 67 | switch(tensor.data_type()) |
| 68 | { |
| 69 | case DataType::F16: |
Giorgio Arena | 6aeb217 | 2020-12-15 15:45:43 +0000 | [diff] [blame] | 70 | { |
Giorgio Arena | a8e2aeb | 2021-01-06 11:34:57 +0000 | [diff] [blame] | 71 | arm_compute::utils::uniform_real_distribution_16bit<half> distribution{ float(lo), float(hi) }; |
Giorgio Arena | 6aeb217 | 2020-12-15 15:45:43 +0000 | [diff] [blame] | 72 | library->fill(tensor, distribution, i); |
| 73 | break; |
| 74 | } |
Moritz Pflanzer | 4dfc235 | 2017-08-02 14:51:36 +0100 | [diff] [blame] | 75 | case DataType::F32: |
| 76 | { |
Giorgio Arena | 6aeb217 | 2020-12-15 15:45:43 +0000 | [diff] [blame] | 77 | std::uniform_real_distribution<float> distribution(lo, hi); |
Moritz Pflanzer | 4dfc235 | 2017-08-02 14:51:36 +0100 | [diff] [blame] | 78 | library->fill(tensor, distribution, i); |
| 79 | break; |
| 80 | } |
| 81 | default: |
| 82 | library->fill_tensor_uniform(tensor, i); |
| 83 | } |
| 84 | } |
| 85 | |
| 86 | 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] | 87 | DataType data_type) |
Moritz Pflanzer | 4dfc235 | 2017-08-02 14:51:36 +0100 | [diff] [blame] | 88 | { |
| 89 | // Create tensors |
Vidhya Sudhan Loganathan | 014333d | 2018-07-02 09:13:49 +0100 | [diff] [blame] | 90 | TensorType a = create_tensor<TensorType>(shape_a, data_type, 1); |
| 91 | TensorType b = create_tensor<TensorType>(shape_b, data_type, 1); |
| 92 | TensorType c = create_tensor<TensorType>(shape_c, data_type, 1); |
| 93 | TensorType dst = create_tensor<TensorType>(output_shape, data_type, 1); |
Moritz Pflanzer | 4dfc235 | 2017-08-02 14:51:36 +0100 | [diff] [blame] | 94 | |
| 95 | // Create and configure function |
| 96 | FunctionType gemm; |
Isabella Gottardi | 8e74f44 | 2018-03-01 16:42:00 +0000 | [diff] [blame] | 97 | // 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] | 98 | // 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] | 99 | // 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] | 100 | gemm.configure(&a, |
| 101 | &b, |
| 102 | (disable_c) ? nullptr : &c, |
| 103 | &dst, |
| 104 | alpha, beta, |
Georgios Pinitas | 4ee8b15 | 2021-07-16 16:16:43 +0100 | [diff] [blame] | 105 | GEMMInfo(false, false, false, (reinterpret_output_as_3d ? output_shape[2] : 0), reinterpret_input_as_3d, false, GEMMLowpOutputStageInfo(), false, false, (reinterpret_input_as_3d |
Gian Marco Iodice | f3622be | 2019-07-29 14:27:16 +0100 | [diff] [blame] | 106 | || reinterpret_output_as_3d))); |
Michele Di Giorgio | 4fc10b3 | 2021-04-30 18:30:41 +0100 | [diff] [blame] | 107 | ARM_COMPUTE_ASSERT(a.info()->is_resizable()); |
| 108 | ARM_COMPUTE_ASSERT(b.info()->is_resizable()); |
| 109 | ARM_COMPUTE_ASSERT(c.info()->is_resizable()); |
| 110 | ARM_COMPUTE_ASSERT(dst.info()->is_resizable()); |
Moritz Pflanzer | 4dfc235 | 2017-08-02 14:51:36 +0100 | [diff] [blame] | 111 | |
Giorgio Arena | 63825e8 | 2021-03-25 14:54:50 +0000 | [diff] [blame] | 112 | add_padding_x({ &a, &b, &c, &dst }); |
| 113 | |
Moritz Pflanzer | 4dfc235 | 2017-08-02 14:51:36 +0100 | [diff] [blame] | 114 | // Allocate tensors |
| 115 | a.allocator()->allocate(); |
| 116 | b.allocator()->allocate(); |
| 117 | c.allocator()->allocate(); |
| 118 | dst.allocator()->allocate(); |
| 119 | |
Michele Di Giorgio | 4fc10b3 | 2021-04-30 18:30:41 +0100 | [diff] [blame] | 120 | ARM_COMPUTE_ASSERT(!a.info()->is_resizable()); |
| 121 | ARM_COMPUTE_ASSERT(!b.info()->is_resizable()); |
| 122 | ARM_COMPUTE_ASSERT(!c.info()->is_resizable()); |
| 123 | ARM_COMPUTE_ASSERT(!dst.info()->is_resizable()); |
Moritz Pflanzer | 4dfc235 | 2017-08-02 14:51:36 +0100 | [diff] [blame] | 124 | |
| 125 | // Fill tensors |
| 126 | fill(AccessorType(a), 0); |
| 127 | fill(AccessorType(b), 1); |
Pablo Tello | 0e37b5c | 2018-10-30 11:18:37 +0000 | [diff] [blame] | 128 | if(!disable_c) |
| 129 | { |
| 130 | fill(AccessorType(c), 2); |
| 131 | } |
Moritz Pflanzer | 4dfc235 | 2017-08-02 14:51:36 +0100 | [diff] [blame] | 132 | |
Mohammed Suhail Munshi | 13a2d00 | 2022-09-05 11:57:34 +0100 | [diff] [blame] | 133 | // Run with variable inputs. |
| 134 | if(run_twice) |
| 135 | { |
| 136 | gemm.run(); |
| 137 | fill(AccessorType(a), 3); // Fill tensors with new seed after run |
| 138 | fill(AccessorType(b), 4); |
| 139 | if(!disable_c) |
| 140 | { |
| 141 | fill(AccessorType(c), 5); |
| 142 | } |
| 143 | } |
| 144 | |
Moritz Pflanzer | 4dfc235 | 2017-08-02 14:51:36 +0100 | [diff] [blame] | 145 | // Compute GEMM function |
| 146 | gemm.run(); |
| 147 | |
| 148 | return dst; |
| 149 | } |
| 150 | |
Michalis Spyrou | 6bff195 | 2019-10-02 17:22:11 +0100 | [diff] [blame] | 151 | 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] | 152 | DataType data_type) |
Moritz Pflanzer | 4dfc235 | 2017-08-02 14:51:36 +0100 | [diff] [blame] | 153 | { |
Gian Marco Iodice | 68a3f56 | 2018-07-26 11:44:03 +0100 | [diff] [blame] | 154 | TensorShape shape_a_to_use = shape_a; |
| 155 | if(reinterpret_input_as_3d) |
| 156 | { |
| 157 | // Collapse the second and third dimension if the input is 3D |
| 158 | shape_a_to_use.collapse(2U, 1U); |
| 159 | } |
| 160 | |
Moritz Pflanzer | 4dfc235 | 2017-08-02 14:51:36 +0100 | [diff] [blame] | 161 | // Create reference |
Gian Marco Iodice | 68a3f56 | 2018-07-26 11:44:03 +0100 | [diff] [blame] | 162 | SimpleTensor<T> a{ shape_a_to_use, data_type, 1 }; |
Vidhya Sudhan Loganathan | 014333d | 2018-07-02 09:13:49 +0100 | [diff] [blame] | 163 | SimpleTensor<T> b{ shape_b, data_type, 1 }; |
Gian Marco Iodice | f3622be | 2019-07-29 14:27:16 +0100 | [diff] [blame] | 164 | SimpleTensor<T> c{ output_shape, data_type, 1 }; |
Moritz Pflanzer | 4dfc235 | 2017-08-02 14:51:36 +0100 | [diff] [blame] | 165 | |
| 166 | // Fill reference |
| 167 | fill(a, 0); |
| 168 | fill(b, 1); |
Gian Marco Iodice | f3622be | 2019-07-29 14:27:16 +0100 | [diff] [blame] | 169 | fill(c, 2); |
| 170 | |
| 171 | if(reinterpret_input_as_3d || reinterpret_output_as_3d) |
Pablo Tello | 0e37b5c | 2018-10-30 11:18:37 +0000 | [diff] [blame] | 172 | { |
Gian Marco Iodice | f3622be | 2019-07-29 14:27:16 +0100 | [diff] [blame] | 173 | const int n = shape_b[0]; |
| 174 | const int m = reinterpret_output_as_3d ? output_shape[1] * output_shape[2] : output_shape[1]; |
| 175 | const int batch_size = reinterpret_output_as_3d ? output_shape[3] : output_shape[2]; |
| 176 | |
Gunes Bayir | 4bfc70e | 2021-12-10 16:17:56 +0000 | [diff] [blame] | 177 | // In case of broadcast, we need to simply copy the first into the following "M" ones |
Gian Marco Iodice | f3622be | 2019-07-29 14:27:16 +0100 | [diff] [blame] | 178 | for(int i = 1; i < m * batch_size; i++) |
| 179 | { |
| 180 | memcpy(c.data() + i * n, c.data(), n * sizeof(T)); |
| 181 | } |
Pablo Tello | 0e37b5c | 2018-10-30 11:18:37 +0000 | [diff] [blame] | 182 | } |
Gunes Bayir | 4bfc70e | 2021-12-10 16:17:56 +0000 | [diff] [blame] | 183 | |
Adnan AlSinan | 3bb72b6 | 2022-05-06 12:10:11 +0100 | [diff] [blame] | 184 | /* Note: Assuming the usual batch matmul dimensions A = (B x M x K), B = (B x K x N), if pretranspose_A is set to true, then A is assumed to be (B x K x M), |
| 185 | therefore, A must be pre-transposed before passing it to the fixture. And, we transpose A again in the fixture to make it (B x M x K) |
| 186 | in order to be able to call reference implementation that works with (B x M x K) input. |
| 187 | Similarly, if pretranspose_B is set to true, then B is assumed to be (B x N x K), B must be pre-transposed before passing it to the fixture. */ |
Gunes Bayir | 4bfc70e | 2021-12-10 16:17:56 +0000 | [diff] [blame] | 188 | |
Mohammed Suhail Munshi | bc5c407 | 2022-04-27 13:49:51 +0100 | [diff] [blame] | 189 | // Define transposed shapes |
| 190 | TensorShape a_transposed_shape(a.shape().y(), a.shape().x()); |
| 191 | TensorShape b_transposed_shape(b.shape().y(), b.shape().x()); |
| 192 | |
| 193 | // Define transposed tensors |
| 194 | SimpleTensor<T> a_transposed{ a_transposed_shape, data_type }; |
| 195 | SimpleTensor<T> b_transposed{ b_transposed_shape, data_type }; |
| 196 | |
| 197 | // pretranspose a if necessary |
| 198 | if(pretranspose_a) |
| 199 | { |
| 200 | transpose_matrix<T>(a, a_transposed); |
| 201 | } |
| 202 | |
| 203 | // pretranspose b if necessary |
| 204 | if(pretranspose_b) |
| 205 | { |
| 206 | transpose_matrix<T>(b, b_transposed); |
| 207 | } |
| 208 | |
Mohammed Suhail Munshi | 13a2d00 | 2022-09-05 11:57:34 +0100 | [diff] [blame] | 209 | // Run with variable inputs. |
| 210 | if(run_twice) |
| 211 | { |
| 212 | reference::gemm<T>((pretranspose_a) ? a_transposed : a, (pretranspose_b) ? b_transposed : b, c, alpha, disable_c ? 0.f : beta); |
| 213 | fill((pretranspose_a) ? a_transposed : a, 3); |
| 214 | fill((pretranspose_b) ? b_transposed : b, 4); |
Adnan AlSinan | 26c9d1a | 2022-09-07 13:54:53 +0100 | [diff] [blame] | 215 | fill(c, 5); |
Mohammed Suhail Munshi | 13a2d00 | 2022-09-05 11:57:34 +0100 | [diff] [blame] | 216 | } |
| 217 | |
Gian Marco Iodice | f3622be | 2019-07-29 14:27:16 +0100 | [diff] [blame] | 218 | // Setting beta to 0 will effectively disable C for the |
| 219 | // computation of the reference: alpha * A * B + 0 * C |
Mohammed Suhail Munshi | bc5c407 | 2022-04-27 13:49:51 +0100 | [diff] [blame] | 220 | // Use transposed tensors if boolean enabled else use original tensors |
Adnan AlSinan | 26c9d1a | 2022-09-07 13:54:53 +0100 | [diff] [blame] | 221 | auto r = reference::gemm<T>((pretranspose_a) ? a_transposed : a, (pretranspose_b) ? b_transposed : b, c, alpha, disable_c ? 0.f : beta); |
Mohammed Suhail Munshi | 13a2d00 | 2022-09-05 11:57:34 +0100 | [diff] [blame] | 222 | return r; |
Moritz Pflanzer | 4dfc235 | 2017-08-02 14:51:36 +0100 | [diff] [blame] | 223 | } |
| 224 | |
| 225 | TensorType _target{}; |
| 226 | SimpleTensor<T> _reference{}; |
Moritz Pflanzer | 4dfc235 | 2017-08-02 14:51:36 +0100 | [diff] [blame] | 227 | }; |
| 228 | |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 229 | template <typename TensorType, typename AccessorType, typename T, typename GEMMOperatorType> |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 230 | class GEMMMatrixMultiplyValidationFixture : public framework::Fixture |
| 231 | { |
| 232 | public: |
| 233 | template <typename...> |
| 234 | 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, |
| 235 | DataType data_type, GPUTarget gpu_arch) |
| 236 | { |
| 237 | // Set the tensor shapes for LHS and RHS matrices |
| 238 | const TensorShape lhs_shape(k, m, batch_size); |
| 239 | const TensorShape rhs_shape(n, k, batch_size); |
| 240 | const TensorShape bias_shape(n, |
| 241 | broadcast_bias ? 1 : m, |
| 242 | broadcast_bias ? 1 : batch_size); |
| 243 | |
| 244 | _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] | 245 | _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] | 246 | } |
| 247 | |
| 248 | protected: |
| 249 | template <typename U> |
| 250 | void fill(U &&tensor, int i) |
| 251 | { |
Giorgio Arena | 4bdd177 | 2020-12-17 16:47:07 +0000 | [diff] [blame] | 252 | static_assert(std::is_floating_point<T>::value || std::is_same<T, half>::value, "Only floating point data types supported."); |
Giorgio Arena | 33b103b | 2021-01-08 10:37:15 +0000 | [diff] [blame] | 253 | using DistributionType = typename std::conditional<std::is_same<T, half>::value, arm_compute::utils::uniform_real_distribution_16bit<T>, std::uniform_real_distribution<T>>::type; |
Giorgio Arena | 4bdd177 | 2020-12-17 16:47:07 +0000 | [diff] [blame] | 254 | |
| 255 | DistributionType distribution{ T(-1.0f), T(1.0f) }; |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 256 | library->fill(tensor, distribution, i); |
| 257 | |
| 258 | // Fill border with infinity in order to check the presence of NaN values (i.e. inf * 0) |
Giorgio Arena | 4bdd177 | 2020-12-17 16:47:07 +0000 | [diff] [blame] | 259 | DistributionType distribution_inf{ T(std::numeric_limits<float>::infinity()), T(std::numeric_limits<float>::infinity()) }; |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 260 | library->fill_borders_with_garbage(tensor, distribution_inf, i); |
| 261 | } |
| 262 | |
| 263 | 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, |
| 264 | bool fp16_mixed_precision, const ActivationLayerInfo &act_info, GPUTarget gpu_arch) |
| 265 | { |
| 266 | // Create tensors |
| 267 | TensorType lhs = create_tensor<TensorType>(lhs_shape, data_type, 1); |
| 268 | TensorType rhs = create_tensor<TensorType>(rhs_shape, data_type, 1); |
| 269 | TensorType bias = create_tensor<TensorType>(bias_shape, data_type, 1); |
| 270 | TensorType dst; |
| 271 | |
| 272 | const unsigned int m = lhs_shape[1]; |
| 273 | const unsigned int n = rhs_shape[0]; |
| 274 | const unsigned int k = lhs_shape[0]; |
| 275 | GEMMReshapeInfo reshape_info(m, n, k, 1, 1, 0, false, broadcast_bias); |
| 276 | |
| 277 | // The output tensor will be auto-initialized within the function |
| 278 | |
| 279 | // Create and configure function |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 280 | GEMMOperatorType gemm; |
| 281 | gemm.configure(gpu_arch, lhs.info(), rhs.info(), bias.info(), dst.info(), alpha, beta, false, reshape_info, fp16_mixed_precision, act_info); |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 282 | |
Michele Di Giorgio | 4fc10b3 | 2021-04-30 18:30:41 +0100 | [diff] [blame] | 283 | ARM_COMPUTE_ASSERT(lhs.info()->is_resizable()); |
| 284 | ARM_COMPUTE_ASSERT(rhs.info()->is_resizable()); |
| 285 | ARM_COMPUTE_ASSERT(bias.info()->is_resizable()); |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 286 | |
Giorgio Arena | 63825e8 | 2021-03-25 14:54:50 +0000 | [diff] [blame] | 287 | add_padding_x({ &lhs, &rhs, &bias, &dst }); |
| 288 | |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 289 | // Allocate tensors |
| 290 | lhs.allocator()->allocate(); |
| 291 | rhs.allocator()->allocate(); |
| 292 | bias.allocator()->allocate(); |
| 293 | dst.allocator()->allocate(); |
| 294 | |
Michele Di Giorgio | 4fc10b3 | 2021-04-30 18:30:41 +0100 | [diff] [blame] | 295 | ARM_COMPUTE_ASSERT(!lhs.info()->is_resizable()); |
| 296 | ARM_COMPUTE_ASSERT(!rhs.info()->is_resizable()); |
| 297 | ARM_COMPUTE_ASSERT(!bias.info()->is_resizable()); |
| 298 | ARM_COMPUTE_ASSERT(!dst.info()->is_resizable()); |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 299 | |
| 300 | // Fill tensors |
| 301 | fill(AccessorType(lhs), 0); |
| 302 | fill(AccessorType(rhs), 1); |
| 303 | fill(AccessorType(bias), 2); |
| 304 | |
| 305 | // Compute GEMM |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 306 | ITensorPack gemm_pack({ { ACL_SRC_0, &lhs }, |
| 307 | { ACL_SRC_1, &rhs }, |
| 308 | { ACL_SRC_2, &bias }, |
| 309 | { ACL_DST, &dst } |
| 310 | }); |
| 311 | gemm.run(gemm_pack); |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 312 | |
| 313 | return dst; |
| 314 | } |
| 315 | |
Michalis Spyrou | 6bff195 | 2019-10-02 17:22:11 +0100 | [diff] [blame] | 316 | 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] | 317 | const ActivationLayerInfo &act_info) |
| 318 | { |
| 319 | TensorShape dst_shape = lhs_shape; |
| 320 | dst_shape[0] = rhs_shape[0]; |
| 321 | dst_shape[1] = lhs_shape[1]; |
| 322 | |
| 323 | // Create reference |
| 324 | SimpleTensor<T> lhs{ lhs_shape, data_type, 1 }; |
| 325 | SimpleTensor<T> rhs{ rhs_shape, data_type, 1 }; |
| 326 | SimpleTensor<T> bias{ dst_shape, data_type, 1 }; |
| 327 | |
| 328 | const int n = rhs_shape[0]; |
| 329 | const int m = lhs_shape[1]; |
| 330 | const int batch_size = lhs_shape[2]; |
| 331 | |
| 332 | // Fill reference |
| 333 | fill(lhs, 0); |
| 334 | fill(rhs, 1); |
| 335 | fill(bias, 2); |
| 336 | |
| 337 | if(broadcast_bias) |
| 338 | { |
Gunes Bayir | 4bfc70e | 2021-12-10 16:17:56 +0000 | [diff] [blame] | 339 | // In case of broadcast, we need to simply copy the first into the following "M" ones |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 340 | for(int i = 1; i < m * batch_size; i++) |
| 341 | { |
| 342 | memcpy(bias.data() + i * n, bias.data(), n * sizeof(T)); |
| 343 | } |
| 344 | } |
| 345 | |
| 346 | return reference::activation_layer(reference::gemm<T>(lhs, rhs, bias, alpha, beta), act_info); |
| 347 | } |
| 348 | |
| 349 | TensorType _target{}; |
| 350 | SimpleTensor<T> _reference{}; |
| 351 | }; |
| 352 | |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 353 | template <typename TensorType, typename AccessorType, typename T, typename GEMMOperatorType> |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 354 | class GEMMMatrixMultiply3DValidationFixture : public framework::Fixture |
| 355 | { |
| 356 | public: |
| 357 | template <typename...> |
| 358 | 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, |
| 359 | const ActivationLayerInfo &act_info, DataType data_type, GPUTarget gpu_arch) |
| 360 | { |
Michalis Spyrou | 6bff195 | 2019-10-02 17:22:11 +0100 | [diff] [blame] | 361 | ARM_COMPUTE_UNUSED(broadcast_bias); |
| 362 | |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 363 | // In case of GEMM3D, m is the product between m_w and m_h |
| 364 | const unsigned int m = m_w * m_h; |
| 365 | |
| 366 | // Set the tensor shapes for LHS and RHS matrices |
| 367 | const TensorShape lhs_shape(k, m, batch_size); |
| 368 | const TensorShape rhs_shape(n, k, batch_size); |
| 369 | const TensorShape bias_shape(n, 1, 1); |
| 370 | |
| 371 | _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] | 372 | _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] | 373 | } |
| 374 | |
| 375 | protected: |
| 376 | template <typename U> |
| 377 | void fill(U &&tensor, int i) |
| 378 | { |
Giorgio Arena | 4bdd177 | 2020-12-17 16:47:07 +0000 | [diff] [blame] | 379 | static_assert(std::is_floating_point<T>::value || std::is_same<T, half>::value, "Only floating point data types supported."); |
Giorgio Arena | 33b103b | 2021-01-08 10:37:15 +0000 | [diff] [blame] | 380 | using DistributionType = typename std::conditional<std::is_same<T, half>::value, arm_compute::utils::uniform_real_distribution_16bit<T>, std::uniform_real_distribution<T>>::type; |
Giorgio Arena | 4bdd177 | 2020-12-17 16:47:07 +0000 | [diff] [blame] | 381 | |
| 382 | DistributionType distribution{ T(-1.0f), T(1.0f) }; |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 383 | library->fill(tensor, distribution, i); |
| 384 | } |
| 385 | |
| 386 | 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, |
| 387 | bool fp16_mixed_precision, const ActivationLayerInfo &act_info, GPUTarget gpu_arch) |
| 388 | { |
| 389 | // Create tensors |
| 390 | TensorType lhs = create_tensor<TensorType>(lhs_shape, data_type, 1); |
| 391 | TensorType rhs = create_tensor<TensorType>(rhs_shape, data_type, 1); |
| 392 | TensorType bias = create_tensor<TensorType>(bias_shape, data_type, 1); |
| 393 | TensorType dst; |
| 394 | |
| 395 | const unsigned int m = lhs_shape[1]; |
| 396 | const unsigned int n = rhs_shape[0]; |
| 397 | const unsigned int k = lhs_shape[0]; |
| 398 | GEMMReshapeInfo reshape_info(m, n, k, 1, 1, m_h, false, true); |
| 399 | |
| 400 | // The output tensor will be auto-initialized within the function |
| 401 | |
| 402 | // Create and configure function |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 403 | GEMMOperatorType gemm; |
| 404 | gemm.configure(gpu_arch, lhs.info(), rhs.info(), bias.info(), dst.info(), alpha, beta, false, reshape_info, fp16_mixed_precision, act_info); |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 405 | |
Michele Di Giorgio | 4fc10b3 | 2021-04-30 18:30:41 +0100 | [diff] [blame] | 406 | ARM_COMPUTE_ASSERT(lhs.info()->is_resizable()); |
| 407 | ARM_COMPUTE_ASSERT(rhs.info()->is_resizable()); |
| 408 | ARM_COMPUTE_ASSERT(bias.info()->is_resizable()); |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 409 | |
Giorgio Arena | 63825e8 | 2021-03-25 14:54:50 +0000 | [diff] [blame] | 410 | add_padding_x({ &lhs, &rhs, &bias, &dst }); |
| 411 | |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 412 | // Allocate tensors |
| 413 | lhs.allocator()->allocate(); |
| 414 | rhs.allocator()->allocate(); |
| 415 | bias.allocator()->allocate(); |
| 416 | dst.allocator()->allocate(); |
| 417 | |
Michele Di Giorgio | 4fc10b3 | 2021-04-30 18:30:41 +0100 | [diff] [blame] | 418 | ARM_COMPUTE_ASSERT(!lhs.info()->is_resizable()); |
| 419 | ARM_COMPUTE_ASSERT(!rhs.info()->is_resizable()); |
| 420 | ARM_COMPUTE_ASSERT(!bias.info()->is_resizable()); |
| 421 | ARM_COMPUTE_ASSERT(!dst.info()->is_resizable()); |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 422 | |
| 423 | // Fill tensors |
| 424 | fill(AccessorType(lhs), 0); |
| 425 | fill(AccessorType(rhs), 1); |
| 426 | fill(AccessorType(bias), 2); |
| 427 | |
| 428 | // Compute GEMM |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 429 | ITensorPack gemm_pack({ { ACL_SRC_0, &lhs }, |
| 430 | { ACL_SRC_1, &rhs }, |
| 431 | { ACL_SRC_2, &bias }, |
| 432 | { ACL_DST, &dst } |
| 433 | }); |
| 434 | gemm.run(gemm_pack); |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 435 | |
| 436 | return dst; |
| 437 | } |
| 438 | |
Michalis Spyrou | 6bff195 | 2019-10-02 17:22:11 +0100 | [diff] [blame] | 439 | 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] | 440 | const ActivationLayerInfo &act_info) |
| 441 | { |
| 442 | TensorShape dst_shape = lhs_shape; |
| 443 | dst_shape.set(0, rhs_shape[0]); |
| 444 | dst_shape.set(1, lhs_shape[1] / m_h); |
| 445 | dst_shape.set(2, m_h); |
| 446 | dst_shape.set(3, lhs_shape[2]); |
| 447 | |
| 448 | // Create reference |
| 449 | SimpleTensor<T> lhs{ lhs_shape, data_type, 1 }; |
| 450 | SimpleTensor<T> rhs{ rhs_shape, data_type, 1 }; |
| 451 | SimpleTensor<T> bias{ dst_shape, data_type, 1 }; |
| 452 | |
| 453 | const int n = rhs_shape[0]; |
| 454 | const int m = lhs_shape[1]; |
| 455 | const int batch_size = lhs_shape[2]; |
| 456 | |
| 457 | // Fill reference |
| 458 | fill(lhs, 0); |
| 459 | fill(rhs, 1); |
| 460 | fill(bias, 2); |
| 461 | |
Gunes Bayir | 4bfc70e | 2021-12-10 16:17:56 +0000 | [diff] [blame] | 462 | // In case of broadcast, we need to simply copy the first into the following "M" ones |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 463 | for(int i = 1; i < m * batch_size; i++) |
| 464 | { |
| 465 | memcpy(bias.data() + i * n, bias.data(), n * sizeof(T)); |
| 466 | } |
| 467 | |
| 468 | return reference::activation_layer(reference::gemm<T>(lhs, rhs, bias, alpha, beta), act_info); |
| 469 | } |
| 470 | |
| 471 | TensorType _target{}; |
| 472 | SimpleTensor<T> _reference{}; |
| 473 | }; |
| 474 | |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 475 | template <typename TensorType, typename AccessorType, typename T, typename ReshapeLHSOperatorType, typename ReshapeRHSOperatorType, typename GEMMOperatorType> |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 476 | class GEMMMatrixMultiplyInterleavedTransposedValidationFixture : public framework::Fixture |
| 477 | { |
| 478 | public: |
| 479 | template <typename...> |
| 480 | 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, |
| 481 | const ActivationLayerInfo &act_info, DataType data_type, GPUTarget gpu_arch) |
| 482 | { |
| 483 | GEMMLHSMatrixInfo lhs_info; |
| 484 | lhs_info.m0 = 4; |
| 485 | lhs_info.k0 = 4; |
| 486 | lhs_info.v0 = v0; |
| 487 | lhs_info.interleave = true; |
| 488 | lhs_info.transpose = true; |
| 489 | |
| 490 | GEMMRHSMatrixInfo rhs_info; |
| 491 | rhs_info.n0 = 16 / sizeof(T); |
| 492 | rhs_info.k0 = 1; |
| 493 | rhs_info.h0 = h0; |
| 494 | rhs_info.interleave = false; |
| 495 | rhs_info.transpose = false; |
| 496 | |
| 497 | // Set the tensor shapes for LHS and RHS matrices |
| 498 | const TensorShape lhs_shape(k, m, batch_size); |
| 499 | const TensorShape rhs_shape(n, k, batch_size); |
| 500 | const TensorShape bias_shape(n, |
| 501 | broadcast_bias ? 1 : m, |
| 502 | broadcast_bias ? 1 : batch_size); |
| 503 | |
| 504 | _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] | 505 | _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] | 506 | } |
| 507 | |
| 508 | protected: |
| 509 | template <typename U> |
| 510 | void fill(U &&tensor, int i) |
| 511 | { |
Giorgio Arena | 4bdd177 | 2020-12-17 16:47:07 +0000 | [diff] [blame] | 512 | static_assert(std::is_floating_point<T>::value || std::is_same<T, half>::value, "Only floating point data types supported."); |
Giorgio Arena | 33b103b | 2021-01-08 10:37:15 +0000 | [diff] [blame] | 513 | using DistributionType = typename std::conditional<std::is_same<T, half>::value, arm_compute::utils::uniform_real_distribution_16bit<T>, std::uniform_real_distribution<T>>::type; |
Giorgio Arena | 4bdd177 | 2020-12-17 16:47:07 +0000 | [diff] [blame] | 514 | |
| 515 | DistributionType distribution{ T(-1.0f), T(1.0f) }; |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 516 | library->fill(tensor, distribution, i); |
| 517 | |
| 518 | // Fill border with infinity in order to check the presence of NaN values (i.e. inf * 0) |
Giorgio Arena | 4bdd177 | 2020-12-17 16:47:07 +0000 | [diff] [blame] | 519 | DistributionType distribution_inf{ T(std::numeric_limits<float>::infinity()), T(std::numeric_limits<float>::infinity()) }; |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 520 | library->fill_borders_with_garbage(tensor, distribution_inf, i); |
| 521 | } |
| 522 | |
| 523 | TensorType compute_target(const TensorShape &lhs_shape, const TensorShape &rhs_shape, const TensorShape &bias_shape, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, |
| 524 | DataType data_type, float alpha, float beta, bool broadcast_bias, bool fp16_mixed_precision, const ActivationLayerInfo &act_info, GPUTarget gpu_arch) |
| 525 | { |
| 526 | // Create tensors |
| 527 | TensorType lhs = create_tensor<TensorType>(lhs_shape, data_type, 1); |
| 528 | TensorType rhs = create_tensor<TensorType>(rhs_shape, data_type, 1); |
| 529 | TensorType bias = create_tensor<TensorType>(bias_shape, data_type, 1); |
| 530 | TensorType lhs_reshaped; |
| 531 | TensorType rhs_reshaped; |
| 532 | TensorType dst; |
| 533 | |
| 534 | const unsigned int m = lhs_shape[1]; |
| 535 | const unsigned int n = rhs_shape[0]; |
| 536 | const unsigned int k = lhs_shape[0]; |
| 537 | GEMMReshapeInfo reshape_info(m, n, k, rhs_info.h0, lhs_info.v0, 0, false, broadcast_bias); |
| 538 | |
| 539 | // The output tensor will be auto-initialized within the function |
| 540 | |
| 541 | // Create and configure function |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 542 | ReshapeLHSOperatorType reshape_lhs; |
| 543 | ReshapeRHSOperatorType reshape_rhs; |
| 544 | GEMMOperatorType gemm; |
| 545 | reshape_lhs.configure(lhs.info(), lhs_reshaped.info(), lhs_info); |
| 546 | reshape_rhs.configure(rhs.info(), rhs_reshaped.info(), rhs_info); |
| 547 | gemm.configure(gpu_arch, lhs_reshaped.info(), rhs_reshaped.info(), bias.info(), dst.info(), alpha, beta, true, reshape_info, fp16_mixed_precision, act_info); |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 548 | |
Michele Di Giorgio | 4fc10b3 | 2021-04-30 18:30:41 +0100 | [diff] [blame] | 549 | ARM_COMPUTE_ASSERT(lhs.info()->is_resizable()); |
| 550 | ARM_COMPUTE_ASSERT(rhs.info()->is_resizable()); |
| 551 | ARM_COMPUTE_ASSERT(bias.info()->is_resizable()); |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 552 | |
Georgios Pinitas | 3dca91b | 2021-04-13 13:35:58 +0100 | [diff] [blame] | 553 | // We do not pad when using image as it needs to comply to strict pitch alignment restrictions |
Giorgio Arena | 63825e8 | 2021-03-25 14:54:50 +0000 | [diff] [blame] | 554 | if(!rhs_info.export_to_cl_image) |
| 555 | { |
| 556 | add_padding_x({ &lhs, &rhs, &lhs_reshaped, &rhs_reshaped, &bias, &dst }); |
| 557 | } |
| 558 | |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 559 | // Allocate tensors |
| 560 | lhs.allocator()->allocate(); |
| 561 | rhs.allocator()->allocate(); |
| 562 | lhs_reshaped.allocator()->allocate(); |
| 563 | rhs_reshaped.allocator()->allocate(); |
| 564 | bias.allocator()->allocate(); |
| 565 | dst.allocator()->allocate(); |
| 566 | |
Michele Di Giorgio | 4fc10b3 | 2021-04-30 18:30:41 +0100 | [diff] [blame] | 567 | ARM_COMPUTE_ASSERT(!lhs.info()->is_resizable()); |
| 568 | ARM_COMPUTE_ASSERT(!rhs.info()->is_resizable()); |
| 569 | ARM_COMPUTE_ASSERT(!bias.info()->is_resizable()); |
| 570 | ARM_COMPUTE_ASSERT(!lhs_reshaped.info()->is_resizable()); |
| 571 | ARM_COMPUTE_ASSERT(!rhs_reshaped.info()->is_resizable()); |
| 572 | ARM_COMPUTE_ASSERT(!dst.info()->is_resizable()); |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 573 | |
| 574 | // Fill tensors |
| 575 | fill(AccessorType(lhs), 0); |
| 576 | fill(AccessorType(rhs), 1); |
| 577 | fill(AccessorType(bias), 2); |
| 578 | |
| 579 | // Compute GEMM |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 580 | ITensorPack reshape_lhs_pack = { { ACL_SRC, &lhs }, { ACL_DST, &lhs_reshaped } }; |
| 581 | reshape_lhs.run(reshape_lhs_pack); |
| 582 | ITensorPack reshape_rhs_pack = { { ACL_SRC, &rhs }, { ACL_DST, &rhs_reshaped } }; |
| 583 | reshape_rhs.run(reshape_rhs_pack); |
| 584 | ITensorPack gemm_pack({ { ACL_SRC_0, &lhs_reshaped }, |
| 585 | { ACL_SRC_1, &rhs_reshaped }, |
| 586 | { ACL_SRC_2, &bias }, |
| 587 | { ACL_DST, &dst } |
| 588 | }); |
| 589 | gemm.run(gemm_pack); |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 590 | |
| 591 | return dst; |
| 592 | } |
| 593 | |
Michalis Spyrou | 6bff195 | 2019-10-02 17:22:11 +0100 | [diff] [blame] | 594 | 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] | 595 | const ActivationLayerInfo &act_info) |
| 596 | { |
| 597 | TensorShape dst_shape = lhs_shape; |
| 598 | dst_shape[0] = rhs_shape[0]; |
| 599 | dst_shape[1] = lhs_shape[1]; |
| 600 | |
| 601 | // Create reference |
| 602 | SimpleTensor<T> lhs{ lhs_shape, data_type, 1 }; |
| 603 | SimpleTensor<T> rhs{ rhs_shape, data_type, 1 }; |
| 604 | SimpleTensor<T> bias{ dst_shape, data_type, 1 }; |
| 605 | |
| 606 | const int n = rhs_shape[0]; |
| 607 | const int m = lhs_shape[1]; |
| 608 | const int batch_size = lhs_shape[2]; |
| 609 | |
| 610 | // Fill reference |
| 611 | fill(lhs, 0); |
| 612 | fill(rhs, 1); |
| 613 | fill(bias, 2); |
| 614 | |
| 615 | if(broadcast_bias) |
| 616 | { |
Gunes Bayir | 4bfc70e | 2021-12-10 16:17:56 +0000 | [diff] [blame] | 617 | // In case of broadcast, we need to simply copy the first into the following "M" ones |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 618 | for(int i = 1; i < m * batch_size; i++) |
| 619 | { |
| 620 | memcpy(bias.data() + i * n, bias.data(), n * sizeof(T)); |
| 621 | } |
| 622 | } |
| 623 | |
| 624 | return reference::activation_layer(reference::gemm<T>(lhs, rhs, bias, alpha, beta), act_info); |
| 625 | } |
| 626 | |
| 627 | TensorType _target{}; |
| 628 | SimpleTensor<T> _reference{}; |
| 629 | }; |
| 630 | |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 631 | template <typename TensorType, typename AccessorType, typename T, typename ReshapeLHSOperatorType, typename ReshapeRHSOperatorType, typename GEMMOperatorType> |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 632 | class GEMMMatrixMultiplyInterleavedTransposed3DValidationFixture : public framework::Fixture |
| 633 | { |
| 634 | public: |
| 635 | template <typename...> |
| 636 | 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, |
| 637 | bool fp16_mixed_precision, const ActivationLayerInfo &act_info, DataType data_type, GPUTarget gpu_arch) |
| 638 | { |
Michalis Spyrou | 6bff195 | 2019-10-02 17:22:11 +0100 | [diff] [blame] | 639 | ARM_COMPUTE_UNUSED(broadcast_bias); |
| 640 | |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 641 | GEMMLHSMatrixInfo lhs_info; |
| 642 | lhs_info.m0 = 4; |
| 643 | lhs_info.k0 = 4; |
| 644 | lhs_info.v0 = v0; |
| 645 | lhs_info.interleave = true; |
| 646 | lhs_info.transpose = true; |
| 647 | |
| 648 | GEMMRHSMatrixInfo rhs_info; |
| 649 | rhs_info.n0 = 16 / sizeof(T); |
| 650 | rhs_info.k0 = 1; |
| 651 | rhs_info.h0 = h0; |
| 652 | rhs_info.interleave = false; |
| 653 | rhs_info.transpose = false; |
| 654 | |
| 655 | // In case of GEMM3D, m is the product between m_w and m_h |
| 656 | const unsigned int m = m_w * m_h; |
| 657 | |
| 658 | // Set the tensor shapes for LHS and RHS matrices |
| 659 | const TensorShape lhs_shape(k, m, batch_size); |
| 660 | const TensorShape rhs_shape(n, k, batch_size); |
| 661 | const TensorShape bias_shape(n, 1, 1); |
| 662 | |
| 663 | _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] | 664 | _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] | 665 | } |
| 666 | |
| 667 | protected: |
| 668 | template <typename U> |
| 669 | void fill(U &&tensor, int i) |
| 670 | { |
Giorgio Arena | 4bdd177 | 2020-12-17 16:47:07 +0000 | [diff] [blame] | 671 | static_assert(std::is_floating_point<T>::value || std::is_same<T, half>::value, "Only floating point data types supported."); |
Giorgio Arena | 33b103b | 2021-01-08 10:37:15 +0000 | [diff] [blame] | 672 | using DistributionType = typename std::conditional<std::is_same<T, half>::value, arm_compute::utils::uniform_real_distribution_16bit<T>, std::uniform_real_distribution<T>>::type; |
Giorgio Arena | 4bdd177 | 2020-12-17 16:47:07 +0000 | [diff] [blame] | 673 | |
| 674 | DistributionType distribution{ T(-1.0f), T(1.0f) }; |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 675 | library->fill(tensor, distribution, i); |
| 676 | } |
| 677 | |
| 678 | TensorType compute_target(const TensorShape &lhs_shape, const TensorShape &rhs_shape, const TensorShape &bias_shape, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, |
| 679 | DataType data_type, float alpha, float beta, unsigned int m_h, bool fp16_mixed_precision, const ActivationLayerInfo &act_info, GPUTarget gpu_arch) |
| 680 | { |
| 681 | // Create tensors |
| 682 | TensorType lhs = create_tensor<TensorType>(lhs_shape, data_type, 1); |
| 683 | TensorType rhs = create_tensor<TensorType>(rhs_shape, data_type, 1); |
| 684 | TensorType bias = create_tensor<TensorType>(bias_shape, data_type, 1); |
| 685 | TensorType lhs_reshaped; |
| 686 | TensorType rhs_reshaped; |
| 687 | TensorType dst; |
| 688 | |
| 689 | const unsigned int m = lhs_shape[1]; |
| 690 | const unsigned int n = rhs_shape[0]; |
| 691 | const unsigned int k = lhs_shape[0]; |
| 692 | GEMMReshapeInfo reshape_info(m, n, k, rhs_info.h0, lhs_info.v0, m_h, false, true); |
| 693 | |
| 694 | // The output tensor will be auto-initialized within the function |
| 695 | |
| 696 | // Create and configure function |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 697 | ReshapeLHSOperatorType reshape_lhs; |
| 698 | ReshapeRHSOperatorType reshape_rhs; |
| 699 | GEMMOperatorType gemm; |
| 700 | reshape_lhs.configure(lhs.info(), lhs_reshaped.info(), lhs_info); |
| 701 | reshape_rhs.configure(rhs.info(), rhs_reshaped.info(), rhs_info); |
| 702 | gemm.configure(gpu_arch, lhs_reshaped.info(), rhs_reshaped.info(), bias.info(), dst.info(), alpha, beta, true, reshape_info, fp16_mixed_precision, act_info); |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 703 | |
Michele Di Giorgio | 4fc10b3 | 2021-04-30 18:30:41 +0100 | [diff] [blame] | 704 | ARM_COMPUTE_ASSERT(lhs.info()->is_resizable()); |
| 705 | ARM_COMPUTE_ASSERT(rhs.info()->is_resizable()); |
| 706 | ARM_COMPUTE_ASSERT(bias.info()->is_resizable()); |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 707 | |
Georgios Pinitas | 3dca91b | 2021-04-13 13:35:58 +0100 | [diff] [blame] | 708 | // We do not pad when using image as it needs to comply to strict pitch alignment restrictions |
Giorgio Arena | 63825e8 | 2021-03-25 14:54:50 +0000 | [diff] [blame] | 709 | if(!rhs_info.export_to_cl_image) |
| 710 | { |
| 711 | add_padding_x({ &lhs, &rhs, &lhs_reshaped, &rhs_reshaped, &bias, &dst }); |
| 712 | } |
| 713 | |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 714 | // Allocate tensors |
| 715 | lhs.allocator()->allocate(); |
| 716 | rhs.allocator()->allocate(); |
| 717 | lhs_reshaped.allocator()->allocate(); |
| 718 | rhs_reshaped.allocator()->allocate(); |
| 719 | bias.allocator()->allocate(); |
| 720 | dst.allocator()->allocate(); |
| 721 | |
Michele Di Giorgio | 4fc10b3 | 2021-04-30 18:30:41 +0100 | [diff] [blame] | 722 | ARM_COMPUTE_ASSERT(!lhs.info()->is_resizable()); |
| 723 | ARM_COMPUTE_ASSERT(!rhs.info()->is_resizable()); |
| 724 | ARM_COMPUTE_ASSERT(!lhs_reshaped.info()->is_resizable()); |
| 725 | ARM_COMPUTE_ASSERT(!rhs_reshaped.info()->is_resizable()); |
| 726 | ARM_COMPUTE_ASSERT(!bias.info()->is_resizable()); |
| 727 | ARM_COMPUTE_ASSERT(!dst.info()->is_resizable()); |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 728 | |
| 729 | // Fill tensors |
| 730 | fill(AccessorType(lhs), 0); |
| 731 | fill(AccessorType(rhs), 1); |
| 732 | fill(AccessorType(bias), 2); |
| 733 | |
| 734 | // Compute GEMM |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 735 | ITensorPack reshape_lhs_pack = { { ACL_SRC, &lhs }, { ACL_DST, &lhs_reshaped } }; |
| 736 | reshape_lhs.run(reshape_lhs_pack); |
| 737 | ITensorPack reshape_rhs_pack = { { ACL_SRC, &rhs }, { ACL_DST, &rhs_reshaped } }; |
| 738 | reshape_rhs.run(reshape_rhs_pack); |
| 739 | ITensorPack gemm_pack({ { ACL_SRC_0, &lhs_reshaped }, |
| 740 | { ACL_SRC_1, &rhs_reshaped }, |
| 741 | { ACL_SRC_2, &bias }, |
| 742 | { ACL_DST, &dst } |
| 743 | }); |
| 744 | gemm.run(gemm_pack); |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 745 | |
| 746 | return dst; |
| 747 | } |
| 748 | |
Michalis Spyrou | 6bff195 | 2019-10-02 17:22:11 +0100 | [diff] [blame] | 749 | 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] | 750 | const ActivationLayerInfo &act_info) |
| 751 | { |
| 752 | TensorShape dst_shape = lhs_shape; |
| 753 | dst_shape.set(0, rhs_shape[0]); |
| 754 | dst_shape.set(1, lhs_shape[1] / m_h); |
| 755 | dst_shape.set(2, m_h); |
| 756 | dst_shape.set(3, lhs_shape[2]); |
| 757 | |
| 758 | // Create reference |
| 759 | SimpleTensor<T> lhs{ lhs_shape, data_type, 1 }; |
| 760 | SimpleTensor<T> rhs{ rhs_shape, data_type, 1 }; |
| 761 | SimpleTensor<T> bias{ dst_shape, data_type, 1 }; |
| 762 | |
| 763 | const int n = rhs_shape[0]; |
| 764 | const int m = lhs_shape[1]; |
| 765 | const int batch_size = lhs_shape[2]; |
| 766 | |
| 767 | // Fill reference |
| 768 | fill(lhs, 0); |
| 769 | fill(rhs, 1); |
| 770 | fill(bias, 2); |
| 771 | |
Gunes Bayir | 4bfc70e | 2021-12-10 16:17:56 +0000 | [diff] [blame] | 772 | // In case of broadcast, we need to simply copy the first into the following "M" ones |
Gian Marco Iodice | d1f5476 | 2019-07-19 09:54:47 +0100 | [diff] [blame] | 773 | for(int i = 1; i < m * batch_size; i++) |
| 774 | { |
| 775 | memcpy(bias.data() + i * n, bias.data(), n * sizeof(T)); |
| 776 | } |
| 777 | |
| 778 | return reference::activation_layer(reference::gemm<T>(lhs, rhs, bias, alpha, beta), act_info); |
| 779 | } |
| 780 | |
| 781 | TensorType _target{}; |
| 782 | SimpleTensor<T> _reference{}; |
| 783 | }; |
| 784 | |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 785 | template <typename TensorType, typename AccessorType, typename T, typename ReshapeLHSOperatorType, typename ReshapeRHSOperatorType, typename GEMMOperatorType, bool fp_mixed_precision = false> |
Gian Marco Iodice | bf9731e | 2018-12-12 10:18:04 +0000 | [diff] [blame] | 786 | class GEMMMatrixMultiplyReshapedValidationFixture : public framework::Fixture |
| 787 | { |
| 788 | public: |
| 789 | template <typename...> |
| 790 | 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] | 791 | 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] | 792 | { |
| 793 | GEMMLHSMatrixInfo lhs_info; |
| 794 | lhs_info.m0 = m0; |
| 795 | lhs_info.k0 = k0; |
| 796 | lhs_info.v0 = v0; |
| 797 | lhs_info.interleave = interleave_lhs; |
Giorgio Arena | ae99b6e | 2019-08-01 14:22:12 +0100 | [diff] [blame] | 798 | lhs_info.transpose = lhs_transpose; |
Gian Marco Iodice | bf9731e | 2018-12-12 10:18:04 +0000 | [diff] [blame] | 799 | |
| 800 | GEMMRHSMatrixInfo rhs_info; |
Gian Marco Iodice | e3a849a | 2020-06-10 17:59:30 +0100 | [diff] [blame] | 801 | rhs_info.n0 = n0; |
| 802 | rhs_info.k0 = k0; |
| 803 | rhs_info.h0 = h0; |
| 804 | rhs_info.interleave = interleave_rhs; |
| 805 | rhs_info.transpose = !lhs_transpose; |
| 806 | rhs_info.export_to_cl_image = export_to_cl_image; |
Gian Marco Iodice | bf9731e | 2018-12-12 10:18:04 +0000 | [diff] [blame] | 807 | |
| 808 | // Set the tensor shapes for LHS and RHS matrices |
| 809 | const TensorShape lhs_shape(k, m, batch_size); |
| 810 | const TensorShape rhs_shape(n, k, batch_size); |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 811 | const TensorShape bias_shape(n, |
| 812 | broadcast_bias ? 1 : m, |
| 813 | broadcast_bias ? 1 : batch_size); |
Gian Marco Iodice | bf9731e | 2018-12-12 10:18:04 +0000 | [diff] [blame] | 814 | |
Sheri Zhang | cc3e53c | 2020-11-16 21:17:28 +0000 | [diff] [blame] | 815 | _target = compute_target(lhs_shape, rhs_shape, bias_shape, lhs_info, rhs_info, data_type, alpha, beta, broadcast_bias, act_info); |
| 816 | if(validate_result) |
| 817 | { |
| 818 | _reference = compute_reference(lhs_shape, rhs_shape, data_type, alpha, beta, broadcast_bias, act_info); |
| 819 | } |
Gian Marco Iodice | bf9731e | 2018-12-12 10:18:04 +0000 | [diff] [blame] | 820 | } |
| 821 | |
| 822 | protected: |
| 823 | template <typename U> |
| 824 | void fill(U &&tensor, int i) |
| 825 | { |
Giorgio Arena | 4bdd177 | 2020-12-17 16:47:07 +0000 | [diff] [blame] | 826 | static_assert(std::is_floating_point<T>::value || std::is_same<T, half>::value, "Only floating point data types supported."); |
Giorgio Arena | 33b103b | 2021-01-08 10:37:15 +0000 | [diff] [blame] | 827 | using DistributionType = typename std::conditional<std::is_same<T, half>::value, arm_compute::utils::uniform_real_distribution_16bit<T>, std::uniform_real_distribution<T>>::type; |
Giorgio Arena | 4bdd177 | 2020-12-17 16:47:07 +0000 | [diff] [blame] | 828 | |
| 829 | DistributionType distribution{ T(-1.0f), T(1.0f) }; |
Gian Marco Iodice | bf9731e | 2018-12-12 10:18:04 +0000 | [diff] [blame] | 830 | library->fill(tensor, distribution, i); |
Gian Marco Iodice | b87b95e | 2019-01-21 17:14:31 +0000 | [diff] [blame] | 831 | |
| 832 | // Fill border with infinity in order to check the presence of NaN values (i.e. inf * 0) |
Giorgio Arena | 4bdd177 | 2020-12-17 16:47:07 +0000 | [diff] [blame] | 833 | DistributionType distribution_inf{ T(std::numeric_limits<float>::infinity()), T(std::numeric_limits<float>::infinity()) }; |
Gian Marco Iodice | b87b95e | 2019-01-21 17:14:31 +0000 | [diff] [blame] | 834 | library->fill_borders_with_garbage(tensor, distribution_inf, i); |
Gian Marco Iodice | bf9731e | 2018-12-12 10:18:04 +0000 | [diff] [blame] | 835 | } |
| 836 | |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 837 | 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] | 838 | 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] | 839 | { |
| 840 | // Create tensors |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 841 | TensorType lhs = create_tensor<TensorType>(lhs_shape, data_type, 1); |
| 842 | TensorType rhs = create_tensor<TensorType>(rhs_shape, data_type, 1); |
| 843 | TensorType bias = create_tensor<TensorType>(bias_shape, data_type, 1); |
Gian Marco Iodice | bf9731e | 2018-12-12 10:18:04 +0000 | [diff] [blame] | 844 | TensorType lhs_reshaped; |
| 845 | TensorType rhs_reshaped; |
| 846 | TensorType dst; |
| 847 | |
| 848 | const unsigned int M = lhs_shape[1]; |
| 849 | const unsigned int N = rhs_shape[0]; |
| 850 | const unsigned int K = lhs_shape[0]; |
Gian Marco Iodice | 7026b30 | 2019-06-26 17:18:11 +0100 | [diff] [blame] | 851 | GEMMKernelInfo kernel_info; |
| 852 | kernel_info.m = M; |
| 853 | kernel_info.n = N; |
| 854 | kernel_info.k = K; |
| 855 | kernel_info.depth_output_gemm3d = 0; |
| 856 | kernel_info.reinterpret_input_as_3d = false; |
| 857 | kernel_info.broadcast_bias = broadcast_bias; |
Gian Marco Iodice | ca1f460 | 2019-07-16 15:46:48 +0100 | [diff] [blame] | 858 | kernel_info.activation_info = act_info; |
Gian Marco Iodice | 0c17aa2 | 2019-09-27 09:23:15 +0100 | [diff] [blame] | 859 | kernel_info.fp_mixed_precision = fp_mixed_precision; |
Gian Marco Iodice | bf9731e | 2018-12-12 10:18:04 +0000 | [diff] [blame] | 860 | |
| 861 | // The output tensor will be auto-initialized within the function |
| 862 | |
| 863 | // Create and configure function |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 864 | ReshapeLHSOperatorType reshape_lhs; |
| 865 | ReshapeRHSOperatorType reshape_rhs; |
| 866 | GEMMOperatorType gemm; |
Sheri Zhang | cc3e53c | 2020-11-16 21:17:28 +0000 | [diff] [blame] | 867 | |
| 868 | validate_result = bool(reshape_rhs.validate(rhs.info(), rhs_reshaped.info(), rhs_info)); |
| 869 | validate_result = validate_result || !rhs_info.export_to_cl_image; |
| 870 | if(!validate_result) |
| 871 | { |
| 872 | return nullptr; |
| 873 | } |
| 874 | |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 875 | reshape_lhs.configure(lhs.info(), lhs_reshaped.info(), lhs_info); |
| 876 | reshape_rhs.configure(rhs.info(), rhs_reshaped.info(), rhs_info); |
| 877 | gemm.configure(lhs_reshaped.info(), rhs_reshaped.info(), bias.info(), dst.info(), alpha, beta, lhs_info, rhs_info, kernel_info); |
Gian Marco Iodice | bf9731e | 2018-12-12 10:18:04 +0000 | [diff] [blame] | 878 | |
Michele Di Giorgio | 4fc10b3 | 2021-04-30 18:30:41 +0100 | [diff] [blame] | 879 | ARM_COMPUTE_ASSERT(lhs.info()->is_resizable()); |
| 880 | ARM_COMPUTE_ASSERT(rhs.info()->is_resizable()); |
| 881 | ARM_COMPUTE_ASSERT(bias.info()->is_resizable()); |
Gian Marco Iodice | bf9731e | 2018-12-12 10:18:04 +0000 | [diff] [blame] | 882 | |
Georgios Pinitas | 3dca91b | 2021-04-13 13:35:58 +0100 | [diff] [blame] | 883 | // We do not pad when using image as it needs to comply to strict pitch alignment restrictions |
Giorgio Arena | 63825e8 | 2021-03-25 14:54:50 +0000 | [diff] [blame] | 884 | if(!rhs_info.export_to_cl_image) |
| 885 | { |
| 886 | add_padding_x({ &lhs, &rhs, &lhs_reshaped, &rhs_reshaped, &bias, &dst }); |
| 887 | } |
| 888 | |
Gian Marco Iodice | bf9731e | 2018-12-12 10:18:04 +0000 | [diff] [blame] | 889 | // Allocate tensors |
| 890 | lhs.allocator()->allocate(); |
| 891 | rhs.allocator()->allocate(); |
| 892 | lhs_reshaped.allocator()->allocate(); |
| 893 | rhs_reshaped.allocator()->allocate(); |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 894 | bias.allocator()->allocate(); |
Gian Marco Iodice | bf9731e | 2018-12-12 10:18:04 +0000 | [diff] [blame] | 895 | dst.allocator()->allocate(); |
| 896 | |
Michele Di Giorgio | 4fc10b3 | 2021-04-30 18:30:41 +0100 | [diff] [blame] | 897 | ARM_COMPUTE_ASSERT(!lhs.info()->is_resizable()); |
| 898 | ARM_COMPUTE_ASSERT(!rhs.info()->is_resizable()); |
| 899 | ARM_COMPUTE_ASSERT(!bias.info()->is_resizable()); |
| 900 | ARM_COMPUTE_ASSERT(!lhs_reshaped.info()->is_resizable()); |
| 901 | ARM_COMPUTE_ASSERT(!rhs_reshaped.info()->is_resizable()); |
| 902 | ARM_COMPUTE_ASSERT(!dst.info()->is_resizable()); |
Gian Marco Iodice | bf9731e | 2018-12-12 10:18:04 +0000 | [diff] [blame] | 903 | |
| 904 | // Fill tensors |
| 905 | fill(AccessorType(lhs), 0); |
| 906 | fill(AccessorType(rhs), 1); |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 907 | fill(AccessorType(bias), 2); |
Gian Marco Iodice | bf9731e | 2018-12-12 10:18:04 +0000 | [diff] [blame] | 908 | |
| 909 | // Compute GEMM |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 910 | ITensorPack reshape_lhs_pack = { { ACL_SRC, &lhs }, { ACL_DST, &lhs_reshaped } }; |
| 911 | reshape_lhs.run(reshape_lhs_pack); |
| 912 | ITensorPack reshape_rhs_pack = { { ACL_SRC, &rhs }, { ACL_DST, &rhs_reshaped } }; |
| 913 | reshape_rhs.run(reshape_rhs_pack); |
| 914 | ITensorPack gemm_pack({ { ACL_SRC_0, &lhs_reshaped }, |
| 915 | { ACL_SRC_1, &rhs_reshaped }, |
| 916 | { ACL_SRC_2, &bias }, |
| 917 | { ACL_DST, &dst } |
| 918 | }); |
| 919 | gemm.run(gemm_pack); |
Gian Marco Iodice | bf9731e | 2018-12-12 10:18:04 +0000 | [diff] [blame] | 920 | |
| 921 | return dst; |
| 922 | } |
| 923 | |
Michalis Spyrou | 6bff195 | 2019-10-02 17:22:11 +0100 | [diff] [blame] | 924 | 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] | 925 | const ActivationLayerInfo &act_info) |
Gian Marco Iodice | bf9731e | 2018-12-12 10:18:04 +0000 | [diff] [blame] | 926 | { |
| 927 | TensorShape dst_shape = lhs_shape; |
| 928 | dst_shape[0] = rhs_shape[0]; |
| 929 | dst_shape[1] = lhs_shape[1]; |
| 930 | |
| 931 | // Create reference |
Gian Marco Iodice | 9382ab3 | 2018-12-17 15:12:07 +0000 | [diff] [blame] | 932 | SimpleTensor<T> lhs{ lhs_shape, data_type, 1 }; |
| 933 | SimpleTensor<T> rhs{ rhs_shape, data_type, 1 }; |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 934 | SimpleTensor<T> bias{ dst_shape, data_type, 1 }; |
| 935 | |
| 936 | const int n = rhs_shape[0]; |
| 937 | const int m = lhs_shape[1]; |
| 938 | const int batch_size = lhs_shape[2]; |
Gian Marco Iodice | bf9731e | 2018-12-12 10:18:04 +0000 | [diff] [blame] | 939 | |
| 940 | // Fill reference |
| 941 | fill(lhs, 0); |
| 942 | fill(rhs, 1); |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 943 | fill(bias, 2); |
Gian Marco Iodice | bf9731e | 2018-12-12 10:18:04 +0000 | [diff] [blame] | 944 | |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 945 | if(broadcast_bias) |
| 946 | { |
Gunes Bayir | 4bfc70e | 2021-12-10 16:17:56 +0000 | [diff] [blame] | 947 | // In case of broadcast, we need to simply copy the first into the following "M" ones |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 948 | for(int i = 1; i < m * batch_size; i++) |
| 949 | { |
| 950 | memcpy(bias.data() + i * n, bias.data(), n * sizeof(T)); |
| 951 | } |
| 952 | } |
| 953 | |
Gian Marco Iodice | 0c17aa2 | 2019-09-27 09:23:15 +0100 | [diff] [blame] | 954 | if(fp_mixed_precision) |
| 955 | { |
| 956 | return reference::activation_layer(reference::gemm_mixed_precision<T>(lhs, rhs, bias, alpha, beta), act_info); |
| 957 | } |
| 958 | else |
| 959 | { |
| 960 | return reference::activation_layer(reference::gemm<T>(lhs, rhs, bias, alpha, beta), act_info); |
| 961 | } |
Gian Marco Iodice | bf9731e | 2018-12-12 10:18:04 +0000 | [diff] [blame] | 962 | } |
| 963 | |
Sheri Zhang | cc3e53c | 2020-11-16 21:17:28 +0000 | [diff] [blame] | 964 | bool validate_result = true; |
Gian Marco Iodice | 9382ab3 | 2018-12-17 15:12:07 +0000 | [diff] [blame] | 965 | TensorType _target{}; |
| 966 | SimpleTensor<T> _reference{}; |
| 967 | }; |
| 968 | |
SiCongLi | 1af5416 | 2021-10-06 15:25:57 +0100 | [diff] [blame] | 969 | /** (EXPERIMENTAL_POST_OPS)*/ |
| 970 | template <typename TensorType, typename AccessorType, typename T, typename ReshapeLHSOperatorType, typename ReshapeRHSOperatorType, typename GEMMOperatorType, bool fp_mixed_precision = false> |
| 971 | class GEMMMatrixMultiplyReshapedWithPostOpsValidationFixture : public framework::Fixture |
| 972 | { |
| 973 | public: |
| 974 | using PostOpArgBroadcast = std::tuple<bool, bool, bool>; // Instruct fixture if we need broadcasting in dimension 0, 1, 2 of each PostOp argument |
| 975 | public: |
| 976 | template <typename...> |
| 977 | 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, |
| 978 | 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, |
| 979 | const experimental::PostOpList<PostOpArgBroadcast> &post_ops) |
| 980 | { |
| 981 | GEMMLHSMatrixInfo lhs_info; |
| 982 | lhs_info.m0 = m0; |
| 983 | lhs_info.k0 = k0; |
| 984 | lhs_info.v0 = v0; |
| 985 | lhs_info.interleave = interleave_lhs; |
| 986 | lhs_info.transpose = lhs_transpose; |
| 987 | |
| 988 | GEMMRHSMatrixInfo rhs_info; |
| 989 | rhs_info.n0 = n0; |
| 990 | rhs_info.k0 = k0; |
| 991 | rhs_info.h0 = h0; |
| 992 | rhs_info.interleave = interleave_rhs; |
| 993 | rhs_info.transpose = !lhs_transpose; |
| 994 | rhs_info.export_to_cl_image = export_to_cl_image; |
| 995 | |
| 996 | // Set the tensor shapes for LHS and RHS matrices |
| 997 | const TensorShape lhs_shape(k, m, batch_size); |
| 998 | const TensorShape rhs_shape(n, k, batch_size); |
| 999 | const TensorShape bias_shape(n, |
| 1000 | broadcast_bias ? 1 : m, |
| 1001 | broadcast_bias ? 1 : batch_size); |
| 1002 | auto post_ops_with_shapes = experimental::transform_post_op_list_arguments<PostOpArgBroadcast, TensorShape>(post_ops, |
| 1003 | [ = ](auto broadcast) |
| 1004 | { |
| 1005 | return TensorShape |
| 1006 | { |
| 1007 | std::get<0>(broadcast) ? 1 : n, |
| 1008 | std::get<1>(broadcast) ? 1 : m, |
| 1009 | std::get<2>(broadcast) ? 1 : batch_size, |
| 1010 | }; |
| 1011 | }); |
| 1012 | |
| 1013 | _target = compute_target(lhs_shape, rhs_shape, bias_shape, lhs_info, rhs_info, data_type, alpha, beta, broadcast_bias, act_info, post_ops_with_shapes); |
| 1014 | if(validate_result) |
| 1015 | { |
| 1016 | _reference = compute_reference(lhs_shape, rhs_shape, data_type, alpha, beta, broadcast_bias, act_info, post_ops_with_shapes); |
| 1017 | } |
| 1018 | } |
| 1019 | |
| 1020 | protected: |
| 1021 | template <typename U> |
| 1022 | void fill(U &&tensor, int i) |
| 1023 | { |
| 1024 | static_assert(std::is_floating_point<T>::value || std::is_same<T, half>::value, "Only floating point data types supported."); |
| 1025 | using DistributionType = typename std::conditional<std::is_same<T, half>::value, arm_compute::utils::uniform_real_distribution_16bit<T>, std::uniform_real_distribution<T>>::type; |
| 1026 | |
| 1027 | DistributionType distribution{ T(-1.0f), T(1.0f) }; |
| 1028 | library->fill(tensor, distribution, i); |
| 1029 | |
| 1030 | // Fill border with infinity in order to check the presence of NaN values (i.e. inf * 0) |
| 1031 | DistributionType distribution_inf{ T(std::numeric_limits<float>::infinity()), T(std::numeric_limits<float>::infinity()) }; |
| 1032 | library->fill_borders_with_garbage(tensor, distribution_inf, i); |
| 1033 | } |
| 1034 | |
| 1035 | TensorType compute_target(const TensorShape &lhs_shape, const TensorShape &rhs_shape, const TensorShape &bias_shape, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, |
| 1036 | DataType data_type, float alpha, float beta, bool broadcast_bias, const ActivationLayerInfo &act_info, const experimental::PostOpList<TensorShape> &post_ops) |
| 1037 | { |
| 1038 | // Create tensors |
| 1039 | TensorType lhs = create_tensor<TensorType>(lhs_shape, data_type, 1); |
| 1040 | TensorType rhs = create_tensor<TensorType>(rhs_shape, data_type, 1); |
| 1041 | TensorType bias = create_tensor<TensorType>(bias_shape, data_type, 1); |
| 1042 | |
| 1043 | // Create post op tensors and populate post op with them |
| 1044 | std::vector<TensorType> post_op_tensors_holder{}; |
| 1045 | auto populated_post_ops = experimental::transform_post_op_list_arguments<TensorShape, ITensorInfo *>(post_ops, |
| 1046 | [&post_op_tensors_holder, &data_type](auto shape) |
| 1047 | { |
| 1048 | auto t = create_tensor<TensorType>(shape, data_type, 1); |
| 1049 | post_op_tensors_holder.push_back(std::move(t)); |
| 1050 | return post_op_tensors_holder.back().info(); |
| 1051 | }); |
| 1052 | TensorType lhs_reshaped; |
| 1053 | TensorType rhs_reshaped; |
| 1054 | TensorType dst; |
| 1055 | |
| 1056 | const unsigned int M = lhs_shape[1]; |
| 1057 | const unsigned int N = rhs_shape[0]; |
| 1058 | const unsigned int K = lhs_shape[0]; |
| 1059 | GEMMKernelInfo kernel_info; |
| 1060 | kernel_info.m = M; |
| 1061 | kernel_info.n = N; |
| 1062 | kernel_info.k = K; |
| 1063 | kernel_info.depth_output_gemm3d = 0; |
| 1064 | kernel_info.reinterpret_input_as_3d = false; |
| 1065 | kernel_info.broadcast_bias = broadcast_bias; |
| 1066 | kernel_info.activation_info = act_info; |
| 1067 | kernel_info.fp_mixed_precision = fp_mixed_precision; |
| 1068 | kernel_info.post_ops = populated_post_ops; |
| 1069 | |
| 1070 | // The output tensor will be auto-initialized within the function |
| 1071 | |
| 1072 | // Create and configure function |
| 1073 | ReshapeLHSOperatorType reshape_lhs; |
| 1074 | ReshapeRHSOperatorType reshape_rhs; |
| 1075 | GEMMOperatorType gemm; |
| 1076 | |
| 1077 | validate_result = bool(reshape_rhs.validate(rhs.info(), rhs_reshaped.info(), rhs_info)); |
| 1078 | validate_result = validate_result || !rhs_info.export_to_cl_image; |
| 1079 | if(!validate_result) |
| 1080 | { |
| 1081 | return nullptr; |
| 1082 | } |
| 1083 | |
| 1084 | reshape_lhs.configure(lhs.info(), lhs_reshaped.info(), lhs_info); |
| 1085 | reshape_rhs.configure(rhs.info(), rhs_reshaped.info(), rhs_info); |
| 1086 | gemm.configure(lhs_reshaped.info(), rhs_reshaped.info(), bias.info(), dst.info(), alpha, beta, lhs_info, rhs_info, kernel_info); |
| 1087 | |
| 1088 | ARM_COMPUTE_ASSERT(lhs.info()->is_resizable()); |
| 1089 | ARM_COMPUTE_ASSERT(rhs.info()->is_resizable()); |
| 1090 | ARM_COMPUTE_ASSERT(bias.info()->is_resizable()); |
| 1091 | for(const auto &tensor : post_op_tensors_holder) |
| 1092 | { |
| 1093 | ARM_COMPUTE_ASSERT(tensor.info()->is_resizable()); |
| 1094 | } |
| 1095 | |
| 1096 | // We do not pad when using image as it needs to comply to strict pitch alignment restrictions |
| 1097 | if(!rhs_info.export_to_cl_image) |
| 1098 | { |
| 1099 | add_padding_x({ &lhs, &rhs, &lhs_reshaped, &rhs_reshaped, &bias, &dst }); |
| 1100 | for(auto &tensor : post_op_tensors_holder) |
| 1101 | { |
| 1102 | add_padding_x({ &tensor }); |
| 1103 | } |
| 1104 | } |
| 1105 | |
| 1106 | // Allocate tensors |
| 1107 | lhs.allocator()->allocate(); |
| 1108 | rhs.allocator()->allocate(); |
| 1109 | lhs_reshaped.allocator()->allocate(); |
| 1110 | rhs_reshaped.allocator()->allocate(); |
| 1111 | bias.allocator()->allocate(); |
| 1112 | dst.allocator()->allocate(); |
| 1113 | for(auto &tensor : post_op_tensors_holder) |
| 1114 | { |
| 1115 | tensor.allocator()->allocate(); |
| 1116 | } |
| 1117 | |
| 1118 | ARM_COMPUTE_ASSERT(!lhs.info()->is_resizable()); |
| 1119 | ARM_COMPUTE_ASSERT(!rhs.info()->is_resizable()); |
| 1120 | ARM_COMPUTE_ASSERT(!bias.info()->is_resizable()); |
| 1121 | ARM_COMPUTE_ASSERT(!lhs_reshaped.info()->is_resizable()); |
| 1122 | ARM_COMPUTE_ASSERT(!rhs_reshaped.info()->is_resizable()); |
| 1123 | ARM_COMPUTE_ASSERT(!dst.info()->is_resizable()); |
| 1124 | for(const auto &tensor : post_op_tensors_holder) |
| 1125 | { |
| 1126 | ARM_COMPUTE_ASSERT(!tensor.info()->is_resizable()); |
| 1127 | } |
| 1128 | |
| 1129 | // Fill tensors |
| 1130 | fill(AccessorType(lhs), 0); |
| 1131 | fill(AccessorType(rhs), 1); |
| 1132 | fill(AccessorType(bias), 2); |
| 1133 | for(size_t i = 0; i < post_op_tensors_holder.size(); ++i) |
| 1134 | { |
| 1135 | fill(AccessorType(post_op_tensors_holder.at(i)), 3 + i); |
| 1136 | } |
| 1137 | |
| 1138 | // Compute GEMM |
| 1139 | ITensorPack reshape_lhs_pack = { { ACL_SRC, &lhs }, { ACL_DST, &lhs_reshaped } }; |
| 1140 | reshape_lhs.run(reshape_lhs_pack); |
| 1141 | ITensorPack reshape_rhs_pack = { { ACL_SRC, &rhs }, { ACL_DST, &rhs_reshaped } }; |
| 1142 | reshape_rhs.run(reshape_rhs_pack); |
| 1143 | ITensorPack gemm_pack({ { ACL_SRC_0, &lhs_reshaped }, |
| 1144 | { ACL_SRC_1, &rhs_reshaped }, |
| 1145 | { ACL_SRC_2, &bias }, |
| 1146 | { ACL_DST, &dst } |
| 1147 | }); |
| 1148 | for(size_t i = 0; i < post_op_tensors_holder.size(); ++i) |
| 1149 | { |
| 1150 | gemm_pack.add_tensor(experimental::get_post_op_arg_type(i), &post_op_tensors_holder.at(i)); |
| 1151 | } |
| 1152 | gemm.run(gemm_pack); |
| 1153 | |
| 1154 | return dst; |
| 1155 | } |
| 1156 | |
| 1157 | SimpleTensor<T> compute_reference(const TensorShape &lhs_shape, const TensorShape &rhs_shape, DataType data_type, float alpha, float beta, bool broadcast_bias, |
| 1158 | const ActivationLayerInfo &act_info, const experimental::PostOpList<TensorShape> &post_ops) |
| 1159 | { |
| 1160 | TensorShape dst_shape = lhs_shape; |
| 1161 | dst_shape[0] = rhs_shape[0]; |
| 1162 | dst_shape[1] = lhs_shape[1]; |
| 1163 | |
| 1164 | // Create reference |
| 1165 | SimpleTensor<T> lhs{ lhs_shape, data_type, 1 }; |
| 1166 | SimpleTensor<T> rhs{ rhs_shape, data_type, 1 }; |
| 1167 | SimpleTensor<T> bias{ dst_shape, data_type, 1 }; |
| 1168 | // Create post op tensors and populate post op with them |
| 1169 | auto populated_post_ops = experimental::transform_post_op_list_arguments<TensorShape, SimpleTensor<T>>(post_ops, [&data_type](auto shape) |
| 1170 | { |
| 1171 | return SimpleTensor<T> { shape, data_type, 1 }; |
| 1172 | }); |
| 1173 | |
| 1174 | const int n = rhs_shape[0]; |
| 1175 | const int m = lhs_shape[1]; |
| 1176 | const int batch_size = lhs_shape[2]; |
| 1177 | |
| 1178 | // Fill reference |
| 1179 | int tensor_idx = 0; |
| 1180 | fill(lhs, tensor_idx++); |
| 1181 | fill(rhs, tensor_idx++); |
| 1182 | fill(bias, tensor_idx++); |
| 1183 | for(auto &op : populated_post_ops.get_list()) |
| 1184 | { |
| 1185 | for(auto tensor : op->arguments()) |
| 1186 | { |
| 1187 | fill(*tensor, tensor_idx++); |
| 1188 | } |
| 1189 | } |
| 1190 | |
| 1191 | if(broadcast_bias) |
| 1192 | { |
Gunes Bayir | 4bfc70e | 2021-12-10 16:17:56 +0000 | [diff] [blame] | 1193 | // In case of broadcast, we need to simply copy the first into the following "M" ones |
SiCongLi | 1af5416 | 2021-10-06 15:25:57 +0100 | [diff] [blame] | 1194 | for(int i = 1; i < m * batch_size; i++) |
| 1195 | { |
| 1196 | memcpy(bias.data() + i * n, bias.data(), n * sizeof(T)); |
| 1197 | } |
| 1198 | } |
| 1199 | |
| 1200 | SimpleTensor<T> out; |
| 1201 | if(fp_mixed_precision) |
| 1202 | { |
| 1203 | out = reference::gemm_mixed_precision<T>(lhs, rhs, bias, alpha, beta); |
| 1204 | } |
| 1205 | else |
| 1206 | { |
| 1207 | out = reference::gemm<T>(lhs, rhs, bias, alpha, beta); |
| 1208 | } |
| 1209 | // Ignore activation info if post ops are used instead |
| 1210 | if(populated_post_ops.size() > 0) |
| 1211 | { |
| 1212 | out = reference::post_ops<T>(out, populated_post_ops); |
| 1213 | } |
| 1214 | else |
| 1215 | { |
| 1216 | out = reference::activation_layer(out, act_info); |
| 1217 | } |
| 1218 | return out; |
| 1219 | } |
| 1220 | |
| 1221 | bool validate_result = true; |
| 1222 | TensorType _target{}; |
| 1223 | SimpleTensor<T> _reference{}; |
| 1224 | }; |
| 1225 | |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 1226 | template <typename TensorType, typename AccessorType, typename T, typename ReshapeLHSOperatorType, typename ReshapeRHSOperatorType, typename GEMMOperatorType, bool fp_mixed_precision = false> |
Gian Marco Iodice | 9382ab3 | 2018-12-17 15:12:07 +0000 | [diff] [blame] | 1227 | class GEMMMatrixMultiplyReshaped3DValidationFixture : public framework::Fixture |
| 1228 | { |
| 1229 | public: |
| 1230 | template <typename...> |
| 1231 | 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] | 1232 | 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] | 1233 | { |
| 1234 | GEMMLHSMatrixInfo lhs_info; |
| 1235 | lhs_info.m0 = m0; |
| 1236 | lhs_info.k0 = k0; |
| 1237 | lhs_info.v0 = v0; |
| 1238 | lhs_info.interleave = interleave_lhs; |
Giorgio Arena | ae99b6e | 2019-08-01 14:22:12 +0100 | [diff] [blame] | 1239 | lhs_info.transpose = lhs_transpose; |
Gian Marco Iodice | 9382ab3 | 2018-12-17 15:12:07 +0000 | [diff] [blame] | 1240 | |
| 1241 | GEMMRHSMatrixInfo rhs_info; |
Gian Marco Iodice | e3a849a | 2020-06-10 17:59:30 +0100 | [diff] [blame] | 1242 | rhs_info.n0 = n0; |
| 1243 | rhs_info.k0 = k0; |
| 1244 | rhs_info.h0 = h0; |
| 1245 | rhs_info.interleave = interleave_rhs; |
| 1246 | rhs_info.transpose = !lhs_transpose; |
| 1247 | rhs_info.export_to_cl_image = export_to_cl_image; |
Gian Marco Iodice | 9382ab3 | 2018-12-17 15:12:07 +0000 | [diff] [blame] | 1248 | |
| 1249 | // In case of GEMM3D, m is the product between m_w and m_h |
| 1250 | const unsigned int m = m_w * m_h; |
| 1251 | |
| 1252 | // Set the tensor shapes for LHS and RHS matrices |
| 1253 | const TensorShape lhs_shape(k, m, batch_size); |
| 1254 | const TensorShape rhs_shape(n, k, batch_size); |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 1255 | const TensorShape bias_shape(n, 1, 1); |
Gian Marco Iodice | 9382ab3 | 2018-12-17 15:12:07 +0000 | [diff] [blame] | 1256 | |
Sheri Zhang | cc3e53c | 2020-11-16 21:17:28 +0000 | [diff] [blame] | 1257 | _target = compute_target(lhs_shape, rhs_shape, bias_shape, lhs_info, rhs_info, data_type, alpha, beta, m_h, act_info); |
| 1258 | if(validate_result) |
| 1259 | { |
| 1260 | _reference = compute_reference(lhs_shape, rhs_shape, data_type, alpha, beta, m_h, act_info); |
| 1261 | } |
Gian Marco Iodice | 9382ab3 | 2018-12-17 15:12:07 +0000 | [diff] [blame] | 1262 | } |
| 1263 | |
| 1264 | protected: |
| 1265 | template <typename U> |
| 1266 | void fill(U &&tensor, int i) |
| 1267 | { |
Giorgio Arena | 4bdd177 | 2020-12-17 16:47:07 +0000 | [diff] [blame] | 1268 | static_assert(std::is_floating_point<T>::value || std::is_same<T, half>::value, "Only floating point data types supported."); |
Giorgio Arena | 33b103b | 2021-01-08 10:37:15 +0000 | [diff] [blame] | 1269 | using DistributionType = typename std::conditional<std::is_same<T, half>::value, arm_compute::utils::uniform_real_distribution_16bit<T>, std::uniform_real_distribution<T>>::type; |
Giorgio Arena | 4bdd177 | 2020-12-17 16:47:07 +0000 | [diff] [blame] | 1270 | |
| 1271 | DistributionType distribution{ T(-1.0f), T(1.0f) }; |
Gian Marco Iodice | 9382ab3 | 2018-12-17 15:12:07 +0000 | [diff] [blame] | 1272 | library->fill(tensor, distribution, i); |
| 1273 | } |
| 1274 | |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 1275 | 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] | 1276 | 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] | 1277 | { |
| 1278 | // Create tensors |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 1279 | TensorType lhs = create_tensor<TensorType>(lhs_shape, data_type, 1); |
| 1280 | TensorType rhs = create_tensor<TensorType>(rhs_shape, data_type, 1); |
| 1281 | TensorType bias = create_tensor<TensorType>(bias_shape, data_type, 1); |
Gian Marco Iodice | 9382ab3 | 2018-12-17 15:12:07 +0000 | [diff] [blame] | 1282 | TensorType lhs_reshaped; |
| 1283 | TensorType rhs_reshaped; |
| 1284 | TensorType dst; |
| 1285 | |
| 1286 | const unsigned int M = lhs_shape[1]; |
| 1287 | const unsigned int N = rhs_shape[0]; |
| 1288 | const unsigned int K = lhs_shape[0]; |
Gian Marco Iodice | 7026b30 | 2019-06-26 17:18:11 +0100 | [diff] [blame] | 1289 | GEMMKernelInfo kernel_info; |
| 1290 | kernel_info.m = M; |
| 1291 | kernel_info.n = N; |
| 1292 | kernel_info.k = K; |
| 1293 | kernel_info.depth_output_gemm3d = m_h; |
| 1294 | kernel_info.reinterpret_input_as_3d = false; |
| 1295 | kernel_info.broadcast_bias = true; |
Gian Marco Iodice | ca1f460 | 2019-07-16 15:46:48 +0100 | [diff] [blame] | 1296 | kernel_info.activation_info = act_info; |
Gian Marco Iodice | 0c17aa2 | 2019-09-27 09:23:15 +0100 | [diff] [blame] | 1297 | kernel_info.fp_mixed_precision = fp_mixed_precision; |
Gian Marco Iodice | 9382ab3 | 2018-12-17 15:12:07 +0000 | [diff] [blame] | 1298 | |
| 1299 | // The output tensor will be auto-initialized within the function |
| 1300 | |
| 1301 | // Create and configure function |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 1302 | ReshapeLHSOperatorType reshape_lhs; |
| 1303 | ReshapeRHSOperatorType reshape_rhs; |
| 1304 | GEMMOperatorType gemm; |
Sheri Zhang | cc3e53c | 2020-11-16 21:17:28 +0000 | [diff] [blame] | 1305 | |
| 1306 | validate_result = bool(reshape_rhs.validate(rhs.info(), rhs_reshaped.info(), rhs_info)); |
| 1307 | validate_result = validate_result || !rhs_info.export_to_cl_image; |
| 1308 | if(!validate_result) |
| 1309 | { |
| 1310 | return nullptr; |
| 1311 | } |
| 1312 | |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 1313 | reshape_lhs.configure(lhs.info(), lhs_reshaped.info(), lhs_info); |
| 1314 | reshape_rhs.configure(rhs.info(), rhs_reshaped.info(), rhs_info); |
| 1315 | gemm.configure(lhs_reshaped.info(), rhs_reshaped.info(), bias.info(), dst.info(), alpha, beta, lhs_info, rhs_info, kernel_info); |
Gian Marco Iodice | 9382ab3 | 2018-12-17 15:12:07 +0000 | [diff] [blame] | 1316 | |
Michele Di Giorgio | 4fc10b3 | 2021-04-30 18:30:41 +0100 | [diff] [blame] | 1317 | ARM_COMPUTE_ASSERT(lhs.info()->is_resizable()); |
| 1318 | ARM_COMPUTE_ASSERT(rhs.info()->is_resizable()); |
| 1319 | ARM_COMPUTE_ASSERT(bias.info()->is_resizable()); |
Gian Marco Iodice | 9382ab3 | 2018-12-17 15:12:07 +0000 | [diff] [blame] | 1320 | |
Georgios Pinitas | 3dca91b | 2021-04-13 13:35:58 +0100 | [diff] [blame] | 1321 | // We do not pad when using image as it needs to comply to strict pitch alignment restrictions |
Giorgio Arena | 63825e8 | 2021-03-25 14:54:50 +0000 | [diff] [blame] | 1322 | if(!rhs_info.export_to_cl_image) |
| 1323 | { |
| 1324 | add_padding_x({ &lhs, &rhs, &lhs_reshaped, &rhs_reshaped, &bias, &dst }); |
| 1325 | } |
| 1326 | |
Gian Marco Iodice | 9382ab3 | 2018-12-17 15:12:07 +0000 | [diff] [blame] | 1327 | // Allocate tensors |
| 1328 | lhs.allocator()->allocate(); |
| 1329 | rhs.allocator()->allocate(); |
| 1330 | lhs_reshaped.allocator()->allocate(); |
| 1331 | rhs_reshaped.allocator()->allocate(); |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 1332 | bias.allocator()->allocate(); |
Gian Marco Iodice | 9382ab3 | 2018-12-17 15:12:07 +0000 | [diff] [blame] | 1333 | dst.allocator()->allocate(); |
| 1334 | |
Michele Di Giorgio | 4fc10b3 | 2021-04-30 18:30:41 +0100 | [diff] [blame] | 1335 | ARM_COMPUTE_ASSERT(!lhs.info()->is_resizable()); |
| 1336 | ARM_COMPUTE_ASSERT(!rhs.info()->is_resizable()); |
| 1337 | ARM_COMPUTE_ASSERT(!lhs_reshaped.info()->is_resizable()); |
| 1338 | ARM_COMPUTE_ASSERT(!rhs_reshaped.info()->is_resizable()); |
| 1339 | ARM_COMPUTE_ASSERT(!bias.info()->is_resizable()); |
| 1340 | ARM_COMPUTE_ASSERT(!dst.info()->is_resizable()); |
Gian Marco Iodice | 9382ab3 | 2018-12-17 15:12:07 +0000 | [diff] [blame] | 1341 | |
| 1342 | // Fill tensors |
| 1343 | fill(AccessorType(lhs), 0); |
| 1344 | fill(AccessorType(rhs), 1); |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 1345 | fill(AccessorType(bias), 2); |
Gian Marco Iodice | 9382ab3 | 2018-12-17 15:12:07 +0000 | [diff] [blame] | 1346 | |
| 1347 | // Compute GEMM |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 1348 | ITensorPack reshape_lhs_pack = { { ACL_SRC, &lhs }, { ACL_DST, &lhs_reshaped } }; |
| 1349 | reshape_lhs.run(reshape_lhs_pack); |
| 1350 | ITensorPack reshape_rhs_pack = { { ACL_SRC, &rhs }, { ACL_DST, &rhs_reshaped } }; |
| 1351 | reshape_rhs.run(reshape_rhs_pack); |
| 1352 | ITensorPack gemm_pack({ { ACL_SRC_0, &lhs_reshaped }, |
| 1353 | { ACL_SRC_1, &rhs_reshaped }, |
| 1354 | { ACL_SRC_2, &bias }, |
| 1355 | { ACL_DST, &dst } |
| 1356 | }); |
| 1357 | gemm.run(gemm_pack); |
Gian Marco Iodice | 9382ab3 | 2018-12-17 15:12:07 +0000 | [diff] [blame] | 1358 | |
| 1359 | return dst; |
| 1360 | } |
| 1361 | |
Michalis Spyrou | 6bff195 | 2019-10-02 17:22:11 +0100 | [diff] [blame] | 1362 | 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] | 1363 | const ActivationLayerInfo &act_info) |
Gian Marco Iodice | 9382ab3 | 2018-12-17 15:12:07 +0000 | [diff] [blame] | 1364 | { |
| 1365 | TensorShape dst_shape = lhs_shape; |
| 1366 | dst_shape.set(0, rhs_shape[0]); |
| 1367 | dst_shape.set(1, lhs_shape[1] / m_h); |
| 1368 | dst_shape.set(2, m_h); |
| 1369 | dst_shape.set(3, lhs_shape[2]); |
| 1370 | |
| 1371 | // Create reference |
| 1372 | SimpleTensor<T> lhs{ lhs_shape, data_type, 1 }; |
| 1373 | SimpleTensor<T> rhs{ rhs_shape, data_type, 1 }; |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 1374 | SimpleTensor<T> bias{ dst_shape, data_type, 1 }; |
| 1375 | |
| 1376 | const int n = rhs_shape[0]; |
| 1377 | const int m = lhs_shape[1]; |
| 1378 | const int batch_size = lhs_shape[2]; |
Gian Marco Iodice | 9382ab3 | 2018-12-17 15:12:07 +0000 | [diff] [blame] | 1379 | |
| 1380 | // Fill reference |
| 1381 | fill(lhs, 0); |
| 1382 | fill(rhs, 1); |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 1383 | fill(bias, 2); |
Gian Marco Iodice | 9382ab3 | 2018-12-17 15:12:07 +0000 | [diff] [blame] | 1384 | |
Gunes Bayir | 4bfc70e | 2021-12-10 16:17:56 +0000 | [diff] [blame] | 1385 | // In case of broadcast, we need to simply copy the first into the following "M" ones |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 1386 | for(int i = 1; i < m * batch_size; i++) |
| 1387 | { |
| 1388 | memcpy(bias.data() + i * n, bias.data(), n * sizeof(T)); |
| 1389 | } |
| 1390 | |
Gian Marco Iodice | 0c17aa2 | 2019-09-27 09:23:15 +0100 | [diff] [blame] | 1391 | if(fp_mixed_precision) |
| 1392 | { |
| 1393 | return reference::activation_layer(reference::gemm_mixed_precision<T>(lhs, rhs, bias, alpha, beta), act_info); |
| 1394 | } |
| 1395 | else |
| 1396 | { |
| 1397 | return reference::activation_layer(reference::gemm<T>(lhs, rhs, bias, alpha, beta), act_info); |
| 1398 | } |
Gian Marco Iodice | 9382ab3 | 2018-12-17 15:12:07 +0000 | [diff] [blame] | 1399 | } |
| 1400 | |
Sheri Zhang | cc3e53c | 2020-11-16 21:17:28 +0000 | [diff] [blame] | 1401 | bool validate_result = true; |
Gian Marco Iodice | 9382ab3 | 2018-12-17 15:12:07 +0000 | [diff] [blame] | 1402 | TensorType _target{}; |
| 1403 | SimpleTensor<T> _reference{}; |
Gian Marco Iodice | bf9731e | 2018-12-12 10:18:04 +0000 | [diff] [blame] | 1404 | }; |
Gian Marco Iodice | adc5395 | 2019-02-15 11:10:31 +0000 | [diff] [blame] | 1405 | |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 1406 | template <typename TensorType, typename AccessorType, typename T, typename ReshapeRHSOperatorType, typename GEMMOperatorType> |
Gian Marco Iodice | adc5395 | 2019-02-15 11:10:31 +0000 | [diff] [blame] | 1407 | class GEMMMatrixMultiplyReshapedOnlyRHSValidationFixture : public framework::Fixture |
| 1408 | { |
| 1409 | public: |
| 1410 | template <typename...> |
| 1411 | 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] | 1412 | 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] | 1413 | { |
| 1414 | GEMMLHSMatrixInfo lhs_info; |
| 1415 | lhs_info.m0 = m0; |
| 1416 | lhs_info.k0 = k0; |
| 1417 | |
| 1418 | GEMMRHSMatrixInfo rhs_info; |
Gian Marco Iodice | 781cba7 | 2020-06-19 16:56:57 +0100 | [diff] [blame] | 1419 | rhs_info.n0 = n0; |
| 1420 | rhs_info.k0 = k0; |
| 1421 | rhs_info.h0 = h0; |
| 1422 | rhs_info.interleave = interleave_rhs; |
| 1423 | rhs_info.transpose = transpose_rhs; |
| 1424 | rhs_info.export_to_cl_image = export_to_cl_image; |
Gian Marco Iodice | adc5395 | 2019-02-15 11:10:31 +0000 | [diff] [blame] | 1425 | |
| 1426 | // Set the tensor shapes for LHS and RHS matrices |
| 1427 | const TensorShape lhs_shape(k, m, batch_size); |
| 1428 | const TensorShape rhs_shape(n, k, batch_size); |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 1429 | const TensorShape bias_shape(n, |
| 1430 | broadcast_bias ? 1 : m, |
| 1431 | broadcast_bias ? 1 : batch_size); |
Georgios Pinitas | b0f342e | 2019-05-21 13:32:43 +0100 | [diff] [blame] | 1432 | |
Sheri Zhang | cc3e53c | 2020-11-16 21:17:28 +0000 | [diff] [blame] | 1433 | _target = compute_target(lhs_shape, rhs_shape, bias_shape, lhs_info, rhs_info, data_type, alpha, beta, broadcast_bias, act_info); |
| 1434 | if(validate_result) |
| 1435 | { |
| 1436 | _reference = compute_reference(lhs_shape, rhs_shape, data_type, alpha, beta, broadcast_bias, act_info); |
| 1437 | } |
Gian Marco Iodice | adc5395 | 2019-02-15 11:10:31 +0000 | [diff] [blame] | 1438 | } |
| 1439 | |
| 1440 | protected: |
| 1441 | template <typename U> |
| 1442 | void fill(U &&tensor, int i) |
| 1443 | { |
Giorgio Arena | 4bdd177 | 2020-12-17 16:47:07 +0000 | [diff] [blame] | 1444 | static_assert(std::is_floating_point<T>::value || std::is_same<T, half>::value, "Only floating point data types supported."); |
Giorgio Arena | 33b103b | 2021-01-08 10:37:15 +0000 | [diff] [blame] | 1445 | using DistributionType = typename std::conditional<std::is_same<T, half>::value, arm_compute::utils::uniform_real_distribution_16bit<T>, std::uniform_real_distribution<T>>::type; |
Giorgio Arena | 4bdd177 | 2020-12-17 16:47:07 +0000 | [diff] [blame] | 1446 | |
| 1447 | DistributionType distribution{ T(-1.0f), T(1.0f) }; |
Gian Marco Iodice | adc5395 | 2019-02-15 11:10:31 +0000 | [diff] [blame] | 1448 | library->fill(tensor, distribution, i); |
| 1449 | |
| 1450 | // Fill border with infinity in order to check the presence of NaN values (i.e. inf * 0) |
Giorgio Arena | 4bdd177 | 2020-12-17 16:47:07 +0000 | [diff] [blame] | 1451 | DistributionType distribution_inf{ T(std::numeric_limits<float>::infinity()), T(std::numeric_limits<float>::infinity()) }; |
Gian Marco Iodice | adc5395 | 2019-02-15 11:10:31 +0000 | [diff] [blame] | 1452 | library->fill_borders_with_garbage(tensor, distribution_inf, i); |
| 1453 | } |
| 1454 | |
Georgios Pinitas | b0f342e | 2019-05-21 13:32:43 +0100 | [diff] [blame] | 1455 | 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] | 1456 | 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] | 1457 | { |
| 1458 | // Create tensors |
Georgios Pinitas | b0f342e | 2019-05-21 13:32:43 +0100 | [diff] [blame] | 1459 | TensorType lhs = create_tensor<TensorType>(lhs_shape, data_type, 1); |
| 1460 | TensorType rhs = create_tensor<TensorType>(rhs_shape, data_type, 1); |
| 1461 | TensorType bias = create_tensor<TensorType>(bias_shape, data_type, 1); |
Gian Marco Iodice | adc5395 | 2019-02-15 11:10:31 +0000 | [diff] [blame] | 1462 | TensorType rhs_reshaped; |
| 1463 | TensorType dst; |
| 1464 | |
| 1465 | const unsigned int M = lhs_shape[1]; |
| 1466 | const unsigned int N = rhs_shape[0]; |
| 1467 | const unsigned int K = lhs_shape[0]; |
Gian Marco Iodice | 7026b30 | 2019-06-26 17:18:11 +0100 | [diff] [blame] | 1468 | GEMMKernelInfo kernel_info; |
| 1469 | kernel_info.m = M; |
| 1470 | kernel_info.n = N; |
| 1471 | kernel_info.k = K; |
| 1472 | kernel_info.depth_output_gemm3d = 0; |
| 1473 | kernel_info.reinterpret_input_as_3d = false; |
| 1474 | kernel_info.broadcast_bias = broadcast_bias; |
Gian Marco Iodice | ca1f460 | 2019-07-16 15:46:48 +0100 | [diff] [blame] | 1475 | kernel_info.activation_info = act_info; |
Gian Marco Iodice | adc5395 | 2019-02-15 11:10:31 +0000 | [diff] [blame] | 1476 | |
| 1477 | // The output tensor will be auto-initialized within the function |
| 1478 | |
| 1479 | // Create and configure function |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 1480 | ReshapeRHSOperatorType reshape_rhs; |
| 1481 | GEMMOperatorType gemm; |
Sheri Zhang | cc3e53c | 2020-11-16 21:17:28 +0000 | [diff] [blame] | 1482 | |
| 1483 | validate_result = bool(reshape_rhs.validate(rhs.info(), rhs_reshaped.info(), rhs_info)); |
| 1484 | validate_result = validate_result || !rhs_info.export_to_cl_image; |
| 1485 | if(!validate_result) |
| 1486 | { |
| 1487 | return nullptr; |
| 1488 | } |
| 1489 | |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 1490 | reshape_rhs.configure(rhs.info(), rhs_reshaped.info(), rhs_info); |
| 1491 | gemm.configure(lhs.info(), rhs_reshaped.info(), bias.info(), dst.info(), alpha, beta, lhs_info, rhs_info, kernel_info); |
Gian Marco Iodice | adc5395 | 2019-02-15 11:10:31 +0000 | [diff] [blame] | 1492 | |
Michele Di Giorgio | 4fc10b3 | 2021-04-30 18:30:41 +0100 | [diff] [blame] | 1493 | ARM_COMPUTE_ASSERT(lhs.info()->is_resizable()); |
| 1494 | ARM_COMPUTE_ASSERT(rhs.info()->is_resizable()); |
| 1495 | ARM_COMPUTE_ASSERT(bias.info()->is_resizable()); |
Gian Marco Iodice | adc5395 | 2019-02-15 11:10:31 +0000 | [diff] [blame] | 1496 | |
Georgios Pinitas | 3dca91b | 2021-04-13 13:35:58 +0100 | [diff] [blame] | 1497 | // We do not pad when using image as it needs to comply to strict pitch alignment restrictions |
Giorgio Arena | 63825e8 | 2021-03-25 14:54:50 +0000 | [diff] [blame] | 1498 | if(!rhs_info.export_to_cl_image) |
| 1499 | { |
| 1500 | add_padding_x({ &lhs, &rhs, &rhs_reshaped, &bias, &dst }); |
| 1501 | } |
| 1502 | |
Gian Marco Iodice | adc5395 | 2019-02-15 11:10:31 +0000 | [diff] [blame] | 1503 | // Allocate tensors |
| 1504 | lhs.allocator()->allocate(); |
| 1505 | rhs.allocator()->allocate(); |
| 1506 | rhs_reshaped.allocator()->allocate(); |
Georgios Pinitas | b0f342e | 2019-05-21 13:32:43 +0100 | [diff] [blame] | 1507 | bias.allocator()->allocate(); |
Gian Marco Iodice | adc5395 | 2019-02-15 11:10:31 +0000 | [diff] [blame] | 1508 | dst.allocator()->allocate(); |
| 1509 | |
Michele Di Giorgio | 4fc10b3 | 2021-04-30 18:30:41 +0100 | [diff] [blame] | 1510 | ARM_COMPUTE_ASSERT(!lhs.info()->is_resizable()); |
| 1511 | ARM_COMPUTE_ASSERT(!rhs.info()->is_resizable()); |
| 1512 | ARM_COMPUTE_ASSERT(!rhs_reshaped.info()->is_resizable()); |
| 1513 | ARM_COMPUTE_ASSERT(!bias.info()->is_resizable()); |
| 1514 | ARM_COMPUTE_ASSERT(!dst.info()->is_resizable()); |
Gian Marco Iodice | adc5395 | 2019-02-15 11:10:31 +0000 | [diff] [blame] | 1515 | |
| 1516 | // Fill tensors |
| 1517 | fill(AccessorType(lhs), 0); |
| 1518 | fill(AccessorType(rhs), 1); |
Georgios Pinitas | b0f342e | 2019-05-21 13:32:43 +0100 | [diff] [blame] | 1519 | fill(AccessorType(bias), 2); |
Gian Marco Iodice | adc5395 | 2019-02-15 11:10:31 +0000 | [diff] [blame] | 1520 | |
| 1521 | // Compute GEMM |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 1522 | ITensorPack reshape_rhs_pack = { { ACL_SRC, &rhs }, { ACL_DST, &rhs_reshaped } }; |
| 1523 | reshape_rhs.run(reshape_rhs_pack); |
| 1524 | ITensorPack gemm_pack({ { ACL_SRC_0, &lhs }, |
| 1525 | { ACL_SRC_1, &rhs_reshaped }, |
| 1526 | { ACL_SRC_2, &bias }, |
| 1527 | { ACL_DST, &dst } |
| 1528 | }); |
| 1529 | gemm.run(gemm_pack); |
Gian Marco Iodice | adc5395 | 2019-02-15 11:10:31 +0000 | [diff] [blame] | 1530 | |
| 1531 | return dst; |
| 1532 | } |
| 1533 | |
Michalis Spyrou | 6bff195 | 2019-10-02 17:22:11 +0100 | [diff] [blame] | 1534 | 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] | 1535 | const ActivationLayerInfo &act_info) |
Gian Marco Iodice | adc5395 | 2019-02-15 11:10:31 +0000 | [diff] [blame] | 1536 | { |
| 1537 | TensorShape dst_shape = lhs_shape; |
| 1538 | dst_shape[0] = rhs_shape[0]; |
| 1539 | dst_shape[1] = lhs_shape[1]; |
| 1540 | |
| 1541 | // Create reference |
| 1542 | SimpleTensor<T> lhs{ lhs_shape, data_type, 1 }; |
| 1543 | SimpleTensor<T> rhs{ rhs_shape, data_type, 1 }; |
Georgios Pinitas | b0f342e | 2019-05-21 13:32:43 +0100 | [diff] [blame] | 1544 | SimpleTensor<T> bias{ dst_shape, data_type, 1 }; |
| 1545 | |
| 1546 | const int n = rhs_shape[0]; |
| 1547 | const int m = lhs_shape[1]; |
| 1548 | const int batch_size = lhs_shape[2]; |
Gian Marco Iodice | adc5395 | 2019-02-15 11:10:31 +0000 | [diff] [blame] | 1549 | |
| 1550 | // Fill reference |
| 1551 | fill(lhs, 0); |
| 1552 | fill(rhs, 1); |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 1553 | fill(bias, 2); |
Gian Marco Iodice | adc5395 | 2019-02-15 11:10:31 +0000 | [diff] [blame] | 1554 | |
Georgios Pinitas | b0f342e | 2019-05-21 13:32:43 +0100 | [diff] [blame] | 1555 | if(broadcast_bias) |
| 1556 | { |
Gunes Bayir | 4bfc70e | 2021-12-10 16:17:56 +0000 | [diff] [blame] | 1557 | // In case of broadcast, we need to simply copy the first into the following "M" ones |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 1558 | for(int i = 1; i < m * batch_size; i++) |
Georgios Pinitas | b0f342e | 2019-05-21 13:32:43 +0100 | [diff] [blame] | 1559 | { |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 1560 | memcpy(bias.data() + i * n, bias.data(), n * sizeof(T)); |
Georgios Pinitas | b0f342e | 2019-05-21 13:32:43 +0100 | [diff] [blame] | 1561 | } |
| 1562 | } |
Georgios Pinitas | b0f342e | 2019-05-21 13:32:43 +0100 | [diff] [blame] | 1563 | |
Gian Marco Iodice | ca1f460 | 2019-07-16 15:46:48 +0100 | [diff] [blame] | 1564 | 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] | 1565 | } |
| 1566 | |
Sheri Zhang | cc3e53c | 2020-11-16 21:17:28 +0000 | [diff] [blame] | 1567 | bool validate_result = true; |
Gian Marco Iodice | adc5395 | 2019-02-15 11:10:31 +0000 | [diff] [blame] | 1568 | TensorType _target{}; |
| 1569 | SimpleTensor<T> _reference{}; |
| 1570 | }; |
| 1571 | |
SiCongLi | afa1972 | 2021-10-24 19:12:33 +0100 | [diff] [blame] | 1572 | /** (EXPERIMENTAL_POST_OPS)*/ |
| 1573 | template <typename TensorType, typename AccessorType, typename T, typename ReshapeRHSOperatorType, typename GEMMOperatorType> |
| 1574 | class GEMMMatrixMultiplyReshapedOnlyRHSWithPostOpsValidationFixture : public framework::Fixture |
| 1575 | { |
| 1576 | public: |
| 1577 | using PostOpArgBroadcast = std::tuple<bool, bool, bool>; // Instruct fixture if we need broadcasting in dimension 0, 1, 2 of each PostOp argument |
| 1578 | template <typename...> |
| 1579 | 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, |
| 1580 | 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, |
| 1581 | const experimental::PostOpList<PostOpArgBroadcast> &post_ops) |
| 1582 | { |
| 1583 | GEMMLHSMatrixInfo lhs_info; |
| 1584 | lhs_info.m0 = m0; |
| 1585 | lhs_info.k0 = k0; |
| 1586 | |
| 1587 | GEMMRHSMatrixInfo rhs_info; |
| 1588 | rhs_info.n0 = n0; |
| 1589 | rhs_info.k0 = k0; |
| 1590 | rhs_info.h0 = h0; |
| 1591 | rhs_info.interleave = interleave_rhs; |
| 1592 | rhs_info.transpose = transpose_rhs; |
| 1593 | rhs_info.export_to_cl_image = export_to_cl_image; |
| 1594 | |
| 1595 | // Set the tensor shapes for LHS and RHS matrices |
| 1596 | const TensorShape lhs_shape(k, m, batch_size); |
| 1597 | const TensorShape rhs_shape(n, k, batch_size); |
| 1598 | const TensorShape bias_shape(n, |
| 1599 | broadcast_bias ? 1 : m, |
| 1600 | broadcast_bias ? 1 : batch_size); |
| 1601 | auto post_ops_with_shapes = experimental::transform_post_op_list_arguments<PostOpArgBroadcast, TensorShape>(post_ops, |
| 1602 | [ = ](auto broadcast) |
| 1603 | { |
| 1604 | return TensorShape |
| 1605 | { |
| 1606 | std::get<0>(broadcast) ? 1 : n, |
| 1607 | std::get<1>(broadcast) ? 1 : m, |
| 1608 | std::get<2>(broadcast) ? 1 : batch_size, |
| 1609 | }; |
| 1610 | }); |
| 1611 | |
| 1612 | _target = compute_target(lhs_shape, rhs_shape, bias_shape, lhs_info, rhs_info, data_type, alpha, beta, broadcast_bias, act_info, post_ops_with_shapes); |
| 1613 | if(validate_result) |
| 1614 | { |
| 1615 | _reference = compute_reference(lhs_shape, rhs_shape, data_type, alpha, beta, broadcast_bias, act_info, post_ops_with_shapes); |
| 1616 | } |
| 1617 | } |
| 1618 | |
| 1619 | protected: |
| 1620 | template <typename U> |
| 1621 | void fill(U &&tensor, int i) |
| 1622 | { |
| 1623 | static_assert(std::is_floating_point<T>::value || std::is_same<T, half>::value, "Only floating point data types supported."); |
| 1624 | using DistributionType = typename std::conditional<std::is_same<T, half>::value, arm_compute::utils::uniform_real_distribution_16bit<T>, std::uniform_real_distribution<T>>::type; |
| 1625 | |
| 1626 | DistributionType distribution{ T(-1.0f), T(1.0f) }; |
| 1627 | library->fill(tensor, distribution, i); |
| 1628 | |
| 1629 | // Fill border with infinity in order to check the presence of NaN values (i.e. inf * 0) |
| 1630 | DistributionType distribution_inf{ T(std::numeric_limits<float>::infinity()), T(std::numeric_limits<float>::infinity()) }; |
| 1631 | library->fill_borders_with_garbage(tensor, distribution_inf, i); |
| 1632 | } |
| 1633 | |
| 1634 | TensorType compute_target(const TensorShape &lhs_shape, const TensorShape &rhs_shape, const TensorShape &bias_shape, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, |
| 1635 | DataType data_type, float alpha, float beta, bool broadcast_bias, const ActivationLayerInfo &act_info, const experimental::PostOpList<TensorShape> &post_ops) |
| 1636 | { |
| 1637 | // Create tensors |
| 1638 | TensorType lhs = create_tensor<TensorType>(lhs_shape, data_type, 1); |
| 1639 | TensorType rhs = create_tensor<TensorType>(rhs_shape, data_type, 1); |
| 1640 | TensorType bias = create_tensor<TensorType>(bias_shape, data_type, 1); |
| 1641 | TensorType rhs_reshaped; |
| 1642 | TensorType dst; |
| 1643 | // Create post op tensors and populate post op with them |
| 1644 | std::vector<TensorType> post_op_tensors_holder{}; |
| 1645 | auto populated_post_ops = experimental::transform_post_op_list_arguments<TensorShape, ITensorInfo *>(post_ops, |
| 1646 | [&post_op_tensors_holder, &data_type](auto shape) |
| 1647 | { |
| 1648 | auto t = create_tensor<TensorType>(shape, data_type, 1); |
| 1649 | post_op_tensors_holder.push_back(std::move(t)); |
| 1650 | return post_op_tensors_holder.back().info(); |
| 1651 | }); |
| 1652 | |
| 1653 | const unsigned int M = lhs_shape[1]; |
| 1654 | const unsigned int N = rhs_shape[0]; |
| 1655 | const unsigned int K = lhs_shape[0]; |
| 1656 | GEMMKernelInfo kernel_info; |
| 1657 | kernel_info.m = M; |
| 1658 | kernel_info.n = N; |
| 1659 | kernel_info.k = K; |
| 1660 | kernel_info.depth_output_gemm3d = 0; |
| 1661 | kernel_info.reinterpret_input_as_3d = false; |
| 1662 | kernel_info.broadcast_bias = broadcast_bias; |
| 1663 | kernel_info.activation_info = act_info; |
| 1664 | kernel_info.post_ops = populated_post_ops; |
| 1665 | |
| 1666 | // The output tensor will be auto-initialized within the function |
| 1667 | |
| 1668 | // Create and configure function |
| 1669 | ReshapeRHSOperatorType reshape_rhs; |
| 1670 | GEMMOperatorType gemm; |
| 1671 | |
| 1672 | validate_result = bool(reshape_rhs.validate(rhs.info(), rhs_reshaped.info(), rhs_info)); |
| 1673 | validate_result = validate_result || !rhs_info.export_to_cl_image; |
| 1674 | if(!validate_result) |
| 1675 | { |
| 1676 | return nullptr; |
| 1677 | } |
| 1678 | |
| 1679 | reshape_rhs.configure(rhs.info(), rhs_reshaped.info(), rhs_info); |
| 1680 | gemm.configure(lhs.info(), rhs_reshaped.info(), bias.info(), dst.info(), alpha, beta, lhs_info, rhs_info, kernel_info); |
| 1681 | |
| 1682 | ARM_COMPUTE_ASSERT(lhs.info()->is_resizable()); |
| 1683 | ARM_COMPUTE_ASSERT(rhs.info()->is_resizable()); |
| 1684 | ARM_COMPUTE_ASSERT(bias.info()->is_resizable()); |
| 1685 | for(const auto &tensor : post_op_tensors_holder) |
| 1686 | { |
| 1687 | ARM_COMPUTE_ASSERT(tensor.info()->is_resizable()); |
| 1688 | } |
| 1689 | |
| 1690 | // We do not pad when using image as it needs to comply to strict pitch alignment restrictions |
| 1691 | if(!rhs_info.export_to_cl_image) |
| 1692 | { |
| 1693 | add_padding_x({ &lhs, &rhs, &rhs_reshaped, &bias, &dst }); |
| 1694 | for(auto &tensor : post_op_tensors_holder) |
| 1695 | { |
| 1696 | add_padding_x({ &tensor }); |
| 1697 | } |
| 1698 | } |
| 1699 | |
| 1700 | // Allocate tensors |
| 1701 | lhs.allocator()->allocate(); |
| 1702 | rhs.allocator()->allocate(); |
| 1703 | rhs_reshaped.allocator()->allocate(); |
| 1704 | bias.allocator()->allocate(); |
| 1705 | dst.allocator()->allocate(); |
| 1706 | for(auto &tensor : post_op_tensors_holder) |
| 1707 | { |
| 1708 | tensor.allocator()->allocate(); |
| 1709 | } |
| 1710 | |
| 1711 | ARM_COMPUTE_ASSERT(!lhs.info()->is_resizable()); |
| 1712 | ARM_COMPUTE_ASSERT(!rhs.info()->is_resizable()); |
| 1713 | ARM_COMPUTE_ASSERT(!rhs_reshaped.info()->is_resizable()); |
| 1714 | ARM_COMPUTE_ASSERT(!bias.info()->is_resizable()); |
| 1715 | ARM_COMPUTE_ASSERT(!dst.info()->is_resizable()); |
| 1716 | for(const auto &tensor : post_op_tensors_holder) |
| 1717 | { |
| 1718 | ARM_COMPUTE_ASSERT(!tensor.info()->is_resizable()); |
| 1719 | } |
| 1720 | |
| 1721 | // Fill tensors |
| 1722 | fill(AccessorType(lhs), 0); |
| 1723 | fill(AccessorType(rhs), 1); |
| 1724 | fill(AccessorType(bias), 2); |
| 1725 | for(size_t i = 0; i < post_op_tensors_holder.size(); ++i) |
| 1726 | { |
| 1727 | fill(AccessorType(post_op_tensors_holder.at(i)), 3 + i); |
| 1728 | } |
| 1729 | |
| 1730 | // Compute GEMM |
| 1731 | ITensorPack reshape_rhs_pack = { { ACL_SRC, &rhs }, { ACL_DST, &rhs_reshaped } }; |
| 1732 | reshape_rhs.run(reshape_rhs_pack); |
| 1733 | ITensorPack gemm_pack({ { ACL_SRC_0, &lhs }, |
| 1734 | { ACL_SRC_1, &rhs_reshaped }, |
| 1735 | { ACL_SRC_2, &bias }, |
| 1736 | { ACL_DST, &dst } |
| 1737 | }); |
| 1738 | for(size_t i = 0; i < post_op_tensors_holder.size(); ++i) |
| 1739 | { |
| 1740 | gemm_pack.add_tensor(experimental::get_post_op_arg_type(i), &post_op_tensors_holder.at(i)); |
| 1741 | } |
| 1742 | gemm.run(gemm_pack); |
| 1743 | |
| 1744 | return dst; |
| 1745 | } |
| 1746 | |
| 1747 | SimpleTensor<T> compute_reference(const TensorShape &lhs_shape, const TensorShape &rhs_shape, DataType data_type, float alpha, float beta, bool broadcast_bias, |
| 1748 | const ActivationLayerInfo &act_info, const experimental::PostOpList<TensorShape> &post_ops) |
| 1749 | { |
| 1750 | TensorShape dst_shape = lhs_shape; |
| 1751 | dst_shape[0] = rhs_shape[0]; |
| 1752 | dst_shape[1] = lhs_shape[1]; |
| 1753 | |
| 1754 | // Create reference |
| 1755 | SimpleTensor<T> lhs{ lhs_shape, data_type, 1 }; |
| 1756 | SimpleTensor<T> rhs{ rhs_shape, data_type, 1 }; |
| 1757 | SimpleTensor<T> bias{ dst_shape, data_type, 1 }; |
| 1758 | // Create post op tensors and populate post op with them |
| 1759 | auto populated_post_ops = experimental::transform_post_op_list_arguments<TensorShape, SimpleTensor<T>>(post_ops, [&data_type](auto shape) |
| 1760 | { |
| 1761 | return SimpleTensor<T> { shape, data_type, 1 }; |
| 1762 | }); |
| 1763 | |
| 1764 | const int n = rhs_shape[0]; |
| 1765 | const int m = lhs_shape[1]; |
| 1766 | const int batch_size = lhs_shape[2]; |
| 1767 | |
| 1768 | // Fill reference |
| 1769 | int tensor_idx = 0; |
| 1770 | fill(lhs, tensor_idx++); |
| 1771 | fill(rhs, tensor_idx++); |
| 1772 | fill(bias, tensor_idx++); |
| 1773 | for(auto &op : populated_post_ops.get_list()) |
| 1774 | { |
| 1775 | for(auto tensor : op->arguments()) |
| 1776 | { |
| 1777 | fill(*tensor, tensor_idx++); |
| 1778 | } |
| 1779 | } |
| 1780 | |
| 1781 | if(broadcast_bias) |
| 1782 | { |
Gunes Bayir | 4bfc70e | 2021-12-10 16:17:56 +0000 | [diff] [blame] | 1783 | // In case of broadcast, we need to simply copy the first into the following "M" ones |
SiCongLi | afa1972 | 2021-10-24 19:12:33 +0100 | [diff] [blame] | 1784 | for(int i = 1; i < m * batch_size; i++) |
| 1785 | { |
| 1786 | memcpy(bias.data() + i * n, bias.data(), n * sizeof(T)); |
| 1787 | } |
| 1788 | } |
| 1789 | |
| 1790 | SimpleTensor<T> out; |
| 1791 | out = reference::gemm<T>(lhs, rhs, bias, alpha, beta); |
| 1792 | // Ignore activation info if post ops are used instead |
| 1793 | if(populated_post_ops.size() > 0) |
| 1794 | { |
| 1795 | out = reference::post_ops<T>(out, populated_post_ops); |
| 1796 | } |
| 1797 | else |
| 1798 | { |
| 1799 | out = reference::activation_layer(out, act_info); |
| 1800 | } |
| 1801 | return out; |
| 1802 | } |
| 1803 | |
| 1804 | bool validate_result = true; |
| 1805 | TensorType _target{}; |
| 1806 | SimpleTensor<T> _reference{}; |
| 1807 | }; |
| 1808 | |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 1809 | template <typename TensorType, typename AccessorType, typename T, typename ReshapeRHSOperatorType, typename GEMMOperatorType> |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 1810 | class GEMMMatrixMultiplyReshapedOnlyRHS3DValidationFixture : public framework::Fixture |
| 1811 | { |
| 1812 | public: |
| 1813 | template <typename...> |
| 1814 | 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] | 1815 | 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] | 1816 | { |
| 1817 | GEMMLHSMatrixInfo lhs_info; |
| 1818 | lhs_info.m0 = m0; |
| 1819 | lhs_info.k0 = k0; |
| 1820 | |
| 1821 | GEMMRHSMatrixInfo rhs_info; |
Gian Marco Iodice | 781cba7 | 2020-06-19 16:56:57 +0100 | [diff] [blame] | 1822 | rhs_info.n0 = n0; |
| 1823 | rhs_info.k0 = k0; |
| 1824 | rhs_info.h0 = h0; |
| 1825 | rhs_info.interleave = interleave_rhs; |
| 1826 | rhs_info.transpose = transpose_rhs; |
| 1827 | rhs_info.export_to_cl_image = export_to_cl_image; |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 1828 | |
| 1829 | // In case of GEMM3D, m is the product between m_w and m_h |
| 1830 | const unsigned int m = m_w * m_h; |
| 1831 | |
| 1832 | // Set the tensor shapes for LHS and RHS matrices |
| 1833 | const TensorShape lhs_shape(k, m, batch_size); |
| 1834 | const TensorShape rhs_shape(n, k, batch_size); |
| 1835 | const TensorShape bias_shape(n, 1, 1); |
| 1836 | |
Sheri Zhang | cc3e53c | 2020-11-16 21:17:28 +0000 | [diff] [blame] | 1837 | _target = compute_target(lhs_shape, rhs_shape, bias_shape, lhs_info, rhs_info, data_type, alpha, beta, m_h, act_info, has_pad_y); |
| 1838 | if(validate_result) |
| 1839 | { |
| 1840 | _reference = compute_reference(lhs_shape, rhs_shape, data_type, alpha, beta, m_h, act_info); |
| 1841 | } |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 1842 | } |
| 1843 | |
| 1844 | protected: |
| 1845 | template <typename U> |
| 1846 | void fill(U &&tensor, int i) |
| 1847 | { |
Giorgio Arena | 4bdd177 | 2020-12-17 16:47:07 +0000 | [diff] [blame] | 1848 | static_assert(std::is_floating_point<T>::value || std::is_same<T, half>::value, "Only floating point data types supported."); |
Giorgio Arena | 33b103b | 2021-01-08 10:37:15 +0000 | [diff] [blame] | 1849 | using DistributionType = typename std::conditional<std::is_same<T, half>::value, arm_compute::utils::uniform_real_distribution_16bit<T>, std::uniform_real_distribution<T>>::type; |
Giorgio Arena | 4bdd177 | 2020-12-17 16:47:07 +0000 | [diff] [blame] | 1850 | |
| 1851 | DistributionType distribution{ T(-1.0f), T(1.0f) }; |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 1852 | library->fill(tensor, distribution, i); |
| 1853 | } |
| 1854 | |
| 1855 | TensorType compute_target(const TensorShape &lhs_shape, const TensorShape &rhs_shape, const TensorShape &bias_shape, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, |
| 1856 | DataType data_type, float alpha, float beta, |
Gian Marco Iodice | 9ae06d4 | 2020-10-22 16:37:12 +0100 | [diff] [blame] | 1857 | 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] | 1858 | { |
| 1859 | // Create tensors |
| 1860 | TensorType lhs = create_tensor<TensorType>(lhs_shape, data_type, 1); |
| 1861 | TensorType rhs = create_tensor<TensorType>(rhs_shape, data_type, 1); |
| 1862 | TensorType bias = create_tensor<TensorType>(bias_shape, data_type, 1); |
| 1863 | TensorType rhs_reshaped; |
| 1864 | TensorType dst; |
| 1865 | |
| 1866 | const unsigned int M = lhs_shape[1]; |
| 1867 | const unsigned int N = rhs_shape[0]; |
| 1868 | const unsigned int K = lhs_shape[0]; |
Gian Marco Iodice | 7026b30 | 2019-06-26 17:18:11 +0100 | [diff] [blame] | 1869 | GEMMKernelInfo kernel_info; |
| 1870 | kernel_info.m = M; |
| 1871 | kernel_info.n = N; |
| 1872 | kernel_info.k = K; |
| 1873 | kernel_info.depth_output_gemm3d = m_h; |
| 1874 | kernel_info.reinterpret_input_as_3d = false; |
| 1875 | kernel_info.broadcast_bias = true; |
Gian Marco Iodice | ca1f460 | 2019-07-16 15:46:48 +0100 | [diff] [blame] | 1876 | kernel_info.activation_info = act_info; |
Gian Marco Iodice | 9ae06d4 | 2020-10-22 16:37:12 +0100 | [diff] [blame] | 1877 | kernel_info.has_pad_y = has_pad_y; |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 1878 | |
| 1879 | // The output tensor will be auto-initialized within the function |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 1880 | // Create and configure function |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 1881 | ReshapeRHSOperatorType reshape_rhs; |
| 1882 | GEMMOperatorType gemm; |
Sheri Zhang | cc3e53c | 2020-11-16 21:17:28 +0000 | [diff] [blame] | 1883 | |
| 1884 | validate_result = bool(reshape_rhs.validate(rhs.info(), rhs_reshaped.info(), rhs_info)); |
| 1885 | validate_result = validate_result || !rhs_info.export_to_cl_image; |
| 1886 | if(!validate_result) |
| 1887 | { |
| 1888 | return nullptr; |
| 1889 | } |
| 1890 | |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 1891 | reshape_rhs.configure(rhs.info(), rhs_reshaped.info(), rhs_info); |
| 1892 | gemm.configure(lhs.info(), rhs_reshaped.info(), bias.info(), dst.info(), alpha, beta, lhs_info, rhs_info, kernel_info); |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 1893 | |
Gian Marco Iodice | 9ae06d4 | 2020-10-22 16:37:12 +0100 | [diff] [blame] | 1894 | if(has_pad_y) |
| 1895 | { |
| 1896 | // Add dummy padding into lhs to validate has_pad_y path |
| 1897 | lhs.info()->extend_padding(PaddingSize(2, 0, 2, 0)); |
| 1898 | dst.info()->extend_padding(PaddingSize(2, 0, 1, 0)); |
| 1899 | } |
| 1900 | |
Michele Di Giorgio | 4fc10b3 | 2021-04-30 18:30:41 +0100 | [diff] [blame] | 1901 | ARM_COMPUTE_ASSERT(lhs.info()->is_resizable()); |
| 1902 | ARM_COMPUTE_ASSERT(rhs.info()->is_resizable()); |
| 1903 | ARM_COMPUTE_ASSERT(bias.info()->is_resizable()); |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 1904 | |
Georgios Pinitas | 3dca91b | 2021-04-13 13:35:58 +0100 | [diff] [blame] | 1905 | // We do not pad when using image as it needs to comply to strict pitch alignment restrictions |
Giorgio Arena | 63825e8 | 2021-03-25 14:54:50 +0000 | [diff] [blame] | 1906 | if(!rhs_info.export_to_cl_image) |
| 1907 | { |
| 1908 | add_padding_x({ &lhs, &rhs, &rhs_reshaped, &bias, &dst }); |
| 1909 | } |
| 1910 | |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 1911 | // Allocate tensors |
| 1912 | lhs.allocator()->allocate(); |
| 1913 | rhs.allocator()->allocate(); |
| 1914 | rhs_reshaped.allocator()->allocate(); |
| 1915 | bias.allocator()->allocate(); |
| 1916 | dst.allocator()->allocate(); |
| 1917 | |
Michele Di Giorgio | 4fc10b3 | 2021-04-30 18:30:41 +0100 | [diff] [blame] | 1918 | ARM_COMPUTE_ASSERT(!lhs.info()->is_resizable()); |
| 1919 | ARM_COMPUTE_ASSERT(!rhs.info()->is_resizable()); |
| 1920 | ARM_COMPUTE_ASSERT(!rhs_reshaped.info()->is_resizable()); |
| 1921 | ARM_COMPUTE_ASSERT(!bias.info()->is_resizable()); |
| 1922 | ARM_COMPUTE_ASSERT(!dst.info()->is_resizable()); |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 1923 | |
| 1924 | // Fill tensors |
| 1925 | fill(AccessorType(lhs), 0); |
| 1926 | fill(AccessorType(rhs), 1); |
| 1927 | fill(AccessorType(bias), 2); |
| 1928 | |
| 1929 | // Compute GEMM |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 1930 | ITensorPack reshape_rhs_pack = { { ACL_SRC, &rhs }, { ACL_DST, &rhs_reshaped } }; |
| 1931 | reshape_rhs.run(reshape_rhs_pack); |
| 1932 | ITensorPack gemm_pack({ { ACL_SRC_0, &lhs }, |
| 1933 | { ACL_SRC_1, &rhs_reshaped }, |
| 1934 | { ACL_SRC_2, &bias }, |
| 1935 | { ACL_DST, &dst } |
| 1936 | }); |
| 1937 | gemm.run(gemm_pack); |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 1938 | |
| 1939 | return dst; |
| 1940 | } |
| 1941 | |
Michalis Spyrou | 6bff195 | 2019-10-02 17:22:11 +0100 | [diff] [blame] | 1942 | 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] | 1943 | const ActivationLayerInfo &act_info) |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 1944 | { |
| 1945 | TensorShape dst_shape = lhs_shape; |
| 1946 | dst_shape.set(0, rhs_shape[0]); |
| 1947 | dst_shape.set(1, lhs_shape[1] / m_h); |
| 1948 | dst_shape.set(2, m_h); |
| 1949 | dst_shape.set(3, lhs_shape[2]); |
| 1950 | |
| 1951 | // Create reference |
| 1952 | SimpleTensor<T> lhs{ lhs_shape, data_type, 1 }; |
| 1953 | SimpleTensor<T> rhs{ rhs_shape, data_type, 1 }; |
| 1954 | SimpleTensor<T> bias{ dst_shape, data_type, 1 }; |
| 1955 | |
| 1956 | const int n = rhs_shape[0]; |
| 1957 | const int m = lhs_shape[1]; |
| 1958 | const int batch_size = lhs_shape[2]; |
| 1959 | |
| 1960 | // Fill reference |
| 1961 | fill(lhs, 0); |
| 1962 | fill(rhs, 1); |
| 1963 | fill(bias, 2); |
| 1964 | |
Gunes Bayir | 4bfc70e | 2021-12-10 16:17:56 +0000 | [diff] [blame] | 1965 | // In case of broadcast, we need to simply copy the first into the following "M" ones |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 1966 | for(int i = 1; i < m * batch_size; i++) |
| 1967 | { |
| 1968 | memcpy(bias.data() + i * n, bias.data(), n * sizeof(T)); |
| 1969 | } |
| 1970 | |
Gian Marco Iodice | ca1f460 | 2019-07-16 15:46:48 +0100 | [diff] [blame] | 1971 | 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] | 1972 | } |
| 1973 | |
Sheri Zhang | cc3e53c | 2020-11-16 21:17:28 +0000 | [diff] [blame] | 1974 | bool validate_result = true; |
Gian Marco Iodice | e16c890 | 2019-06-14 16:11:10 +0100 | [diff] [blame] | 1975 | TensorType _target{}; |
| 1976 | SimpleTensor<T> _reference{}; |
| 1977 | }; |
| 1978 | |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 1979 | template <typename TensorType, typename AccessorType, typename T, typename GEMMOperatorType> |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 1980 | class GEMMMatrixMultiplyNativeValidationFixture : public framework::Fixture |
| 1981 | { |
| 1982 | public: |
| 1983 | template <typename...> |
Gian Marco Iodice | ca1f460 | 2019-07-16 15:46:48 +0100 | [diff] [blame] | 1984 | 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, |
| 1985 | const ActivationLayerInfo &act_info) |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 1986 | { |
| 1987 | GEMMLHSMatrixInfo lhs_info; |
| 1988 | lhs_info.m0 = m0; |
| 1989 | lhs_info.k0 = k0; |
| 1990 | |
| 1991 | GEMMRHSMatrixInfo rhs_info; |
| 1992 | rhs_info.n0 = n0; |
| 1993 | rhs_info.k0 = k0; |
| 1994 | |
| 1995 | // Set the tensor shapes for LHS and RHS matrices |
| 1996 | const TensorShape lhs_shape(k, m, batch_size); |
| 1997 | const TensorShape rhs_shape(n, k, batch_size); |
Gian Marco Iodice | 944170e | 2019-06-24 14:40:30 +0100 | [diff] [blame] | 1998 | const TensorShape bias_shape(n, |
| 1999 | broadcast_bias ? 1 : m, |
| 2000 | broadcast_bias ? 1 : batch_size); |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 2001 | |
Gian Marco Iodice | ca1f460 | 2019-07-16 15:46:48 +0100 | [diff] [blame] | 2002 | _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] | 2003 | _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] | 2004 | } |
| 2005 | |
| 2006 | protected: |
| 2007 | template <typename U> |
| 2008 | void fill(U &&tensor, int i) |
| 2009 | { |
Giorgio Arena | 4bdd177 | 2020-12-17 16:47:07 +0000 | [diff] [blame] | 2010 | static_assert(std::is_floating_point<T>::value || std::is_same<T, half>::value, "Only floating point data types supported."); |
Giorgio Arena | 33b103b | 2021-01-08 10:37:15 +0000 | [diff] [blame] | 2011 | using DistributionType = typename std::conditional<std::is_same<T, half>::value, arm_compute::utils::uniform_real_distribution_16bit<T>, std::uniform_real_distribution<T>>::type; |
Giorgio Arena | 4bdd177 | 2020-12-17 16:47:07 +0000 | [diff] [blame] | 2012 | |
| 2013 | DistributionType distribution{ T(-1.0f), T(1.0f) }; |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 2014 | library->fill(tensor, distribution, i); |
| 2015 | |
| 2016 | // Fill border with infinity in order to check the presence of NaN values (i.e. inf * 0) |
Giorgio Arena | 4bdd177 | 2020-12-17 16:47:07 +0000 | [diff] [blame] | 2017 | DistributionType distribution_inf{ T(std::numeric_limits<float>::infinity()), T(std::numeric_limits<float>::infinity()) }; |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 2018 | library->fill_borders_with_garbage(tensor, distribution_inf, i); |
| 2019 | } |
| 2020 | |
Gian Marco Iodice | 944170e | 2019-06-24 14:40:30 +0100 | [diff] [blame] | 2021 | 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] | 2022 | 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] | 2023 | { |
| 2024 | // Create tensors |
Gian Marco Iodice | 944170e | 2019-06-24 14:40:30 +0100 | [diff] [blame] | 2025 | TensorType lhs = create_tensor<TensorType>(lhs_shape, data_type, 1); |
| 2026 | TensorType rhs = create_tensor<TensorType>(rhs_shape, data_type, 1); |
| 2027 | TensorType bias = create_tensor<TensorType>(bias_shape, data_type, 1); |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 2028 | TensorType dst; |
| 2029 | |
| 2030 | const unsigned int M = lhs_shape[1]; |
| 2031 | const unsigned int N = rhs_shape[0]; |
| 2032 | const unsigned int K = lhs_shape[0]; |
Gian Marco Iodice | 7026b30 | 2019-06-26 17:18:11 +0100 | [diff] [blame] | 2033 | GEMMKernelInfo kernel_info; |
| 2034 | kernel_info.m = M; |
| 2035 | kernel_info.n = N; |
| 2036 | kernel_info.k = K; |
| 2037 | kernel_info.depth_output_gemm3d = 0; |
| 2038 | kernel_info.reinterpret_input_as_3d = false; |
| 2039 | kernel_info.broadcast_bias = broadcast_bias; |
Gian Marco Iodice | ca1f460 | 2019-07-16 15:46:48 +0100 | [diff] [blame] | 2040 | kernel_info.activation_info = act_info; |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 2041 | |
| 2042 | // Create and configure function |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 2043 | GEMMOperatorType gemm; |
| 2044 | gemm.configure(lhs.info(), rhs.info(), bias.info(), dst.info(), alpha, beta, lhs_info, rhs_info, kernel_info); |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 2045 | |
Michele Di Giorgio | 4fc10b3 | 2021-04-30 18:30:41 +0100 | [diff] [blame] | 2046 | ARM_COMPUTE_ASSERT(lhs.info()->is_resizable()); |
| 2047 | ARM_COMPUTE_ASSERT(rhs.info()->is_resizable()); |
| 2048 | ARM_COMPUTE_ASSERT(bias.info()->is_resizable()); |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 2049 | |
Giorgio Arena | 63825e8 | 2021-03-25 14:54:50 +0000 | [diff] [blame] | 2050 | add_padding_x({ &lhs, &rhs, &bias, &dst }); |
| 2051 | |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 2052 | // Allocate tensors |
| 2053 | lhs.allocator()->allocate(); |
| 2054 | rhs.allocator()->allocate(); |
Gian Marco Iodice | 944170e | 2019-06-24 14:40:30 +0100 | [diff] [blame] | 2055 | bias.allocator()->allocate(); |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 2056 | dst.allocator()->allocate(); |
| 2057 | |
Michele Di Giorgio | 4fc10b3 | 2021-04-30 18:30:41 +0100 | [diff] [blame] | 2058 | ARM_COMPUTE_ASSERT(!lhs.info()->is_resizable()); |
| 2059 | ARM_COMPUTE_ASSERT(!rhs.info()->is_resizable()); |
| 2060 | ARM_COMPUTE_ASSERT(!bias.info()->is_resizable()); |
| 2061 | ARM_COMPUTE_ASSERT(!dst.info()->is_resizable()); |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 2062 | |
| 2063 | // Fill tensors |
| 2064 | fill(AccessorType(lhs), 0); |
| 2065 | fill(AccessorType(rhs), 1); |
Gian Marco Iodice | 944170e | 2019-06-24 14:40:30 +0100 | [diff] [blame] | 2066 | fill(AccessorType(bias), 2); |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 2067 | |
| 2068 | // Compute GEMM |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 2069 | ITensorPack gemm_pack({ { ACL_SRC_0, &lhs }, |
| 2070 | { ACL_SRC_1, &rhs }, |
| 2071 | { ACL_SRC_2, &bias }, |
| 2072 | { ACL_DST, &dst } |
| 2073 | }); |
| 2074 | gemm.run(gemm_pack); |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 2075 | |
| 2076 | return dst; |
| 2077 | } |
| 2078 | |
Michalis Spyrou | 6bff195 | 2019-10-02 17:22:11 +0100 | [diff] [blame] | 2079 | 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] | 2080 | const ActivationLayerInfo &act_info) |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 2081 | { |
| 2082 | TensorShape dst_shape = lhs_shape; |
| 2083 | dst_shape[0] = rhs_shape[0]; |
| 2084 | dst_shape[1] = lhs_shape[1]; |
| 2085 | |
| 2086 | // Create reference |
| 2087 | SimpleTensor<T> lhs{ lhs_shape, data_type, 1 }; |
| 2088 | SimpleTensor<T> rhs{ rhs_shape, data_type, 1 }; |
Gian Marco Iodice | 944170e | 2019-06-24 14:40:30 +0100 | [diff] [blame] | 2089 | SimpleTensor<T> bias{ dst_shape, data_type, 1 }; |
| 2090 | |
| 2091 | const int n = rhs_shape[0]; |
| 2092 | const int m = lhs_shape[1]; |
| 2093 | const int batch_size = lhs_shape[2]; |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 2094 | |
| 2095 | // Fill reference |
| 2096 | fill(lhs, 0); |
| 2097 | fill(rhs, 1); |
Gian Marco Iodice | 944170e | 2019-06-24 14:40:30 +0100 | [diff] [blame] | 2098 | fill(bias, 2); |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 2099 | |
Gian Marco Iodice | 944170e | 2019-06-24 14:40:30 +0100 | [diff] [blame] | 2100 | if(broadcast_bias) |
| 2101 | { |
Gunes Bayir | 4bfc70e | 2021-12-10 16:17:56 +0000 | [diff] [blame] | 2102 | // In case of broadcast, we need to simply copy the first into the following "M" ones |
Gian Marco Iodice | 944170e | 2019-06-24 14:40:30 +0100 | [diff] [blame] | 2103 | for(int i = 1; i < m * batch_size; i++) |
| 2104 | { |
| 2105 | memcpy(bias.data() + i * n, bias.data(), n * sizeof(T)); |
| 2106 | } |
| 2107 | } |
| 2108 | |
Gian Marco Iodice | ca1f460 | 2019-07-16 15:46:48 +0100 | [diff] [blame] | 2109 | 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] | 2110 | } |
| 2111 | |
| 2112 | TensorType _target{}; |
| 2113 | SimpleTensor<T> _reference{}; |
| 2114 | }; |
| 2115 | |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 2116 | template <typename TensorType, typename AccessorType, typename T, typename GEMMOperatorType> |
SiCongLi | afa1972 | 2021-10-24 19:12:33 +0100 | [diff] [blame] | 2117 | class GEMMMatrixMultiplyNativeWithPostOpsValidationFixture : public framework::Fixture |
| 2118 | { |
| 2119 | public: |
| 2120 | using PostOpArgBroadcast = std::tuple<bool, bool, bool>; // Instruct fixture if we need broadcasting in dimension 0, 1, 2 of each PostOp argument |
| 2121 | public: |
| 2122 | template <typename...> |
| 2123 | 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, |
| 2124 | const ActivationLayerInfo &act_info, const experimental::PostOpList<PostOpArgBroadcast> &post_ops) |
| 2125 | { |
| 2126 | GEMMLHSMatrixInfo lhs_info; |
| 2127 | lhs_info.m0 = m0; |
| 2128 | lhs_info.k0 = k0; |
| 2129 | |
| 2130 | GEMMRHSMatrixInfo rhs_info; |
| 2131 | rhs_info.n0 = n0; |
| 2132 | rhs_info.k0 = k0; |
| 2133 | |
| 2134 | // Set the tensor shapes for LHS and RHS matrices |
| 2135 | const TensorShape lhs_shape(k, m, batch_size); |
| 2136 | const TensorShape rhs_shape(n, k, batch_size); |
| 2137 | const TensorShape bias_shape(n, |
| 2138 | broadcast_bias ? 1 : m, |
| 2139 | broadcast_bias ? 1 : batch_size); |
| 2140 | const auto post_ops_with_shapes = experimental::transform_post_op_list_arguments<PostOpArgBroadcast, TensorShape>(post_ops, |
| 2141 | [ = ](auto broadcast) |
| 2142 | { |
| 2143 | return TensorShape |
| 2144 | { |
| 2145 | std::get<0>(broadcast) ? 1 : n, |
| 2146 | std::get<1>(broadcast) ? 1 : m, |
| 2147 | std::get<2>(broadcast) ? 1 : batch_size, |
| 2148 | }; |
| 2149 | }); |
| 2150 | |
| 2151 | _target = compute_target(lhs_shape, rhs_shape, bias_shape, lhs_info, rhs_info, data_type, alpha, beta, broadcast_bias, act_info, post_ops_with_shapes); |
| 2152 | _reference = compute_reference(lhs_shape, rhs_shape, data_type, alpha, beta, broadcast_bias, act_info, post_ops_with_shapes); |
| 2153 | } |
| 2154 | |
| 2155 | protected: |
| 2156 | template <typename U> |
| 2157 | void fill(U &&tensor, int i) |
| 2158 | { |
| 2159 | static_assert(std::is_floating_point<T>::value || std::is_same<T, half>::value, "Only floating point data types supported."); |
| 2160 | using DistributionType = typename std::conditional<std::is_same<T, half>::value, arm_compute::utils::uniform_real_distribution_16bit<T>, std::uniform_real_distribution<T>>::type; |
| 2161 | |
| 2162 | DistributionType distribution{ T(-1.0f), T(1.0f) }; |
| 2163 | library->fill(tensor, distribution, i); |
| 2164 | |
| 2165 | // Fill border with infinity in order to check the presence of NaN values (i.e. inf * 0) |
| 2166 | DistributionType distribution_inf{ T(std::numeric_limits<float>::infinity()), T(std::numeric_limits<float>::infinity()) }; |
| 2167 | library->fill_borders_with_garbage(tensor, distribution_inf, i); |
| 2168 | } |
| 2169 | |
| 2170 | TensorType compute_target(const TensorShape &lhs_shape, const TensorShape &rhs_shape, const TensorShape &bias_shape, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, |
| 2171 | DataType data_type, float alpha, float beta, bool broadcast_bias, const ActivationLayerInfo &act_info, const experimental::PostOpList<TensorShape> &post_ops) |
| 2172 | { |
| 2173 | // Create tensors |
| 2174 | TensorType lhs = create_tensor<TensorType>(lhs_shape, data_type, 1); |
| 2175 | TensorType rhs = create_tensor<TensorType>(rhs_shape, data_type, 1); |
| 2176 | TensorType bias = create_tensor<TensorType>(bias_shape, data_type, 1); |
| 2177 | TensorType dst; |
| 2178 | // Create post op tensors and populate post op with them |
| 2179 | std::vector<TensorType> post_op_tensors_holder{}; |
| 2180 | auto populated_post_ops = experimental::transform_post_op_list_arguments<TensorShape, ITensorInfo *>(post_ops, |
| 2181 | [&post_op_tensors_holder, &data_type](auto shape) |
| 2182 | { |
| 2183 | auto t = create_tensor<TensorType>(shape, data_type, 1); |
| 2184 | post_op_tensors_holder.push_back(std::move(t)); |
| 2185 | return post_op_tensors_holder.back().info(); |
| 2186 | }); |
| 2187 | |
| 2188 | const unsigned int M = lhs_shape[1]; |
| 2189 | const unsigned int N = rhs_shape[0]; |
| 2190 | const unsigned int K = lhs_shape[0]; |
| 2191 | GEMMKernelInfo kernel_info; |
| 2192 | kernel_info.m = M; |
| 2193 | kernel_info.n = N; |
| 2194 | kernel_info.k = K; |
| 2195 | kernel_info.depth_output_gemm3d = 0; |
| 2196 | kernel_info.reinterpret_input_as_3d = false; |
| 2197 | kernel_info.broadcast_bias = broadcast_bias; |
| 2198 | kernel_info.activation_info = act_info; |
| 2199 | kernel_info.post_ops = populated_post_ops; |
| 2200 | |
| 2201 | // Create and configure function |
| 2202 | GEMMOperatorType gemm; |
| 2203 | gemm.configure(lhs.info(), rhs.info(), bias.info(), dst.info(), alpha, beta, lhs_info, rhs_info, kernel_info); |
| 2204 | |
| 2205 | ARM_COMPUTE_ASSERT(lhs.info()->is_resizable()); |
| 2206 | ARM_COMPUTE_ASSERT(rhs.info()->is_resizable()); |
| 2207 | ARM_COMPUTE_ASSERT(bias.info()->is_resizable()); |
| 2208 | for(const auto &tensor : post_op_tensors_holder) |
| 2209 | { |
| 2210 | ARM_COMPUTE_ASSERT(tensor.info()->is_resizable()); |
| 2211 | } |
| 2212 | |
| 2213 | add_padding_x({ &lhs, &rhs, &bias, &dst }); |
| 2214 | for(auto &tensor : post_op_tensors_holder) |
| 2215 | { |
| 2216 | add_padding_x({ &tensor }); |
| 2217 | } |
| 2218 | |
| 2219 | // Allocate tensors |
| 2220 | lhs.allocator()->allocate(); |
| 2221 | rhs.allocator()->allocate(); |
| 2222 | bias.allocator()->allocate(); |
| 2223 | dst.allocator()->allocate(); |
| 2224 | for(auto &tensor : post_op_tensors_holder) |
| 2225 | { |
| 2226 | tensor.allocator()->allocate(); |
| 2227 | } |
| 2228 | |
| 2229 | ARM_COMPUTE_ASSERT(!lhs.info()->is_resizable()); |
| 2230 | ARM_COMPUTE_ASSERT(!rhs.info()->is_resizable()); |
| 2231 | ARM_COMPUTE_ASSERT(!bias.info()->is_resizable()); |
| 2232 | ARM_COMPUTE_ASSERT(!dst.info()->is_resizable()); |
| 2233 | for(const auto &tensor : post_op_tensors_holder) |
| 2234 | { |
| 2235 | ARM_COMPUTE_ASSERT(!tensor.info()->is_resizable()); |
| 2236 | } |
| 2237 | |
| 2238 | // Fill tensors |
| 2239 | fill(AccessorType(lhs), 0); |
| 2240 | fill(AccessorType(rhs), 1); |
| 2241 | fill(AccessorType(bias), 2); |
| 2242 | for(size_t i = 0; i < post_op_tensors_holder.size(); ++i) |
| 2243 | { |
| 2244 | fill(AccessorType(post_op_tensors_holder.at(i)), 3 + i); |
| 2245 | } |
| 2246 | |
| 2247 | // Compute GEMM |
| 2248 | ITensorPack gemm_pack({ { ACL_SRC_0, &lhs }, |
| 2249 | { ACL_SRC_1, &rhs }, |
| 2250 | { ACL_SRC_2, &bias }, |
| 2251 | { ACL_DST, &dst } |
| 2252 | }); |
| 2253 | for(size_t i = 0; i < post_op_tensors_holder.size(); ++i) |
| 2254 | { |
| 2255 | gemm_pack.add_tensor(experimental::get_post_op_arg_type(i), &post_op_tensors_holder.at(i)); |
| 2256 | } |
| 2257 | gemm.run(gemm_pack); |
| 2258 | |
| 2259 | return dst; |
| 2260 | } |
| 2261 | |
| 2262 | SimpleTensor<T> compute_reference(const TensorShape &lhs_shape, const TensorShape &rhs_shape, DataType data_type, float alpha, float beta, bool broadcast_bias, |
| 2263 | const ActivationLayerInfo &act_info, const experimental::PostOpList<TensorShape> &post_ops) |
| 2264 | { |
| 2265 | TensorShape dst_shape = lhs_shape; |
| 2266 | dst_shape[0] = rhs_shape[0]; |
| 2267 | dst_shape[1] = lhs_shape[1]; |
| 2268 | |
| 2269 | // Create reference |
| 2270 | SimpleTensor<T> lhs{ lhs_shape, data_type, 1 }; |
| 2271 | SimpleTensor<T> rhs{ rhs_shape, data_type, 1 }; |
| 2272 | SimpleTensor<T> bias{ dst_shape, data_type, 1 }; |
| 2273 | // Create post op tensors and populate post op with them |
| 2274 | auto populated_post_ops = experimental::transform_post_op_list_arguments<TensorShape, SimpleTensor<T>>(post_ops, [&data_type](auto shape) |
| 2275 | { |
| 2276 | return SimpleTensor<T> { shape, data_type, 1 }; |
| 2277 | }); |
| 2278 | |
| 2279 | const int n = rhs_shape[0]; |
| 2280 | const int m = lhs_shape[1]; |
| 2281 | const int batch_size = lhs_shape[2]; |
| 2282 | |
| 2283 | // Fill reference |
| 2284 | int tensor_idx = 0; |
| 2285 | fill(lhs, tensor_idx++); |
| 2286 | fill(rhs, tensor_idx++); |
| 2287 | fill(bias, tensor_idx++); |
| 2288 | for(auto &op : populated_post_ops.get_list()) |
| 2289 | { |
| 2290 | for(auto tensor : op->arguments()) |
| 2291 | { |
| 2292 | fill(*tensor, tensor_idx++); |
| 2293 | } |
| 2294 | } |
| 2295 | |
| 2296 | if(broadcast_bias) |
| 2297 | { |
Gunes Bayir | 4bfc70e | 2021-12-10 16:17:56 +0000 | [diff] [blame] | 2298 | // In case of broadcast, we need to simply copy the first into the following "M" ones |
SiCongLi | afa1972 | 2021-10-24 19:12:33 +0100 | [diff] [blame] | 2299 | for(int i = 1; i < m * batch_size; i++) |
| 2300 | { |
| 2301 | memcpy(bias.data() + i * n, bias.data(), n * sizeof(T)); |
| 2302 | } |
| 2303 | } |
| 2304 | |
| 2305 | SimpleTensor<T> out; |
| 2306 | out = reference::gemm<T>(lhs, rhs, bias, alpha, beta); |
| 2307 | // Ignore activation info if post ops are used instead |
| 2308 | if(populated_post_ops.size() > 0) |
| 2309 | { |
| 2310 | out = reference::post_ops<T>(out, populated_post_ops); |
| 2311 | } |
| 2312 | else |
| 2313 | { |
| 2314 | out = reference::activation_layer(out, act_info); |
| 2315 | } |
| 2316 | return out; |
| 2317 | } |
| 2318 | |
| 2319 | TensorType _target{}; |
| 2320 | SimpleTensor<T> _reference{}; |
| 2321 | }; |
| 2322 | |
| 2323 | template <typename TensorType, typename AccessorType, typename T, typename GEMMOperatorType> |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 2324 | class GEMMMatrixMultiplyNative3DValidationFixture : public framework::Fixture |
| 2325 | { |
| 2326 | public: |
| 2327 | template <typename...> |
Gian Marco Iodice | ca1f460 | 2019-07-16 15:46:48 +0100 | [diff] [blame] | 2328 | 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, |
| 2329 | const ActivationLayerInfo &act_info) |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 2330 | { |
| 2331 | GEMMLHSMatrixInfo lhs_info; |
| 2332 | lhs_info.m0 = m0; |
| 2333 | lhs_info.k0 = k0; |
| 2334 | |
| 2335 | GEMMRHSMatrixInfo rhs_info; |
| 2336 | rhs_info.n0 = n0; |
| 2337 | rhs_info.k0 = k0; |
| 2338 | |
| 2339 | // In case of GEMM3D, m is the product between m_w and m_h |
| 2340 | const unsigned int m = m_w * m_h; |
| 2341 | |
| 2342 | // Set the tensor shapes for LHS and RHS matrices |
| 2343 | const TensorShape lhs_shape(k, m, batch_size); |
| 2344 | const TensorShape rhs_shape(n, k, batch_size); |
Gian Marco Iodice | 944170e | 2019-06-24 14:40:30 +0100 | [diff] [blame] | 2345 | const TensorShape bias_shape(n, 1, 1); |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 2346 | |
Gian Marco Iodice | ca1f460 | 2019-07-16 15:46:48 +0100 | [diff] [blame] | 2347 | _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] | 2348 | _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] | 2349 | } |
| 2350 | |
| 2351 | protected: |
| 2352 | template <typename U> |
| 2353 | void fill(U &&tensor, int i) |
| 2354 | { |
Giorgio Arena | 4bdd177 | 2020-12-17 16:47:07 +0000 | [diff] [blame] | 2355 | static_assert(std::is_floating_point<T>::value || std::is_same<T, half>::value, "Only floating point data types supported."); |
Giorgio Arena | 33b103b | 2021-01-08 10:37:15 +0000 | [diff] [blame] | 2356 | using DistributionType = typename std::conditional<std::is_same<T, half>::value, arm_compute::utils::uniform_real_distribution_16bit<T>, std::uniform_real_distribution<T>>::type; |
Giorgio Arena | 4bdd177 | 2020-12-17 16:47:07 +0000 | [diff] [blame] | 2357 | |
| 2358 | DistributionType distribution{ T(-1.0f), T(1.0f) }; |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 2359 | library->fill(tensor, distribution, i); |
| 2360 | } |
| 2361 | |
Gian Marco Iodice | 944170e | 2019-06-24 14:40:30 +0100 | [diff] [blame] | 2362 | 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] | 2363 | 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] | 2364 | { |
| 2365 | // Create tensors |
Gian Marco Iodice | 944170e | 2019-06-24 14:40:30 +0100 | [diff] [blame] | 2366 | TensorType lhs = create_tensor<TensorType>(lhs_shape, data_type, 1); |
| 2367 | TensorType rhs = create_tensor<TensorType>(rhs_shape, data_type, 1); |
| 2368 | TensorType bias = create_tensor<TensorType>(bias_shape, data_type, 1); |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 2369 | TensorType dst; |
| 2370 | |
| 2371 | const unsigned int M = lhs_shape[1]; |
| 2372 | const unsigned int N = rhs_shape[0]; |
| 2373 | const unsigned int K = lhs_shape[0]; |
Gian Marco Iodice | 7026b30 | 2019-06-26 17:18:11 +0100 | [diff] [blame] | 2374 | GEMMKernelInfo kernel_info; |
| 2375 | kernel_info.m = M; |
| 2376 | kernel_info.n = N; |
| 2377 | kernel_info.k = K; |
| 2378 | kernel_info.depth_output_gemm3d = m_h; |
| 2379 | kernel_info.reinterpret_input_as_3d = false; |
| 2380 | kernel_info.broadcast_bias = true; |
Gian Marco Iodice | ca1f460 | 2019-07-16 15:46:48 +0100 | [diff] [blame] | 2381 | kernel_info.activation_info = act_info; |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 2382 | |
| 2383 | // The output tensor will be auto-initialized within the function |
| 2384 | |
| 2385 | // Create and configure function |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 2386 | GEMMOperatorType gemm; |
| 2387 | gemm.configure(lhs.info(), rhs.info(), bias.info(), dst.info(), alpha, beta, lhs_info, rhs_info, kernel_info); |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 2388 | |
Michele Di Giorgio | 4fc10b3 | 2021-04-30 18:30:41 +0100 | [diff] [blame] | 2389 | ARM_COMPUTE_ASSERT(lhs.info()->is_resizable()); |
| 2390 | ARM_COMPUTE_ASSERT(rhs.info()->is_resizable()); |
| 2391 | ARM_COMPUTE_ASSERT(bias.info()->is_resizable()); |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 2392 | |
Giorgio Arena | 63825e8 | 2021-03-25 14:54:50 +0000 | [diff] [blame] | 2393 | add_padding_x({ &lhs, &rhs, &bias, &dst }); |
| 2394 | |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 2395 | // Allocate tensors |
| 2396 | lhs.allocator()->allocate(); |
| 2397 | rhs.allocator()->allocate(); |
Gian Marco Iodice | 944170e | 2019-06-24 14:40:30 +0100 | [diff] [blame] | 2398 | bias.allocator()->allocate(); |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 2399 | dst.allocator()->allocate(); |
| 2400 | |
Michele Di Giorgio | 4fc10b3 | 2021-04-30 18:30:41 +0100 | [diff] [blame] | 2401 | ARM_COMPUTE_ASSERT(!lhs.info()->is_resizable()); |
| 2402 | ARM_COMPUTE_ASSERT(!rhs.info()->is_resizable()); |
| 2403 | ARM_COMPUTE_ASSERT(!bias.info()->is_resizable()); |
| 2404 | ARM_COMPUTE_ASSERT(!dst.info()->is_resizable()); |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 2405 | |
| 2406 | // Fill tensors |
| 2407 | fill(AccessorType(lhs), 0); |
| 2408 | fill(AccessorType(rhs), 1); |
Gian Marco Iodice | 944170e | 2019-06-24 14:40:30 +0100 | [diff] [blame] | 2409 | fill(AccessorType(bias), 2); |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 2410 | |
| 2411 | // Compute GEMM |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 2412 | ITensorPack gemm_pack({ { ACL_SRC_0, &lhs }, |
| 2413 | { ACL_SRC_1, &rhs }, |
| 2414 | { ACL_SRC_2, &bias }, |
| 2415 | { ACL_DST, &dst } |
| 2416 | }); |
| 2417 | gemm.run(gemm_pack); |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 2418 | |
| 2419 | return dst; |
| 2420 | } |
| 2421 | |
Michalis Spyrou | 6bff195 | 2019-10-02 17:22:11 +0100 | [diff] [blame] | 2422 | 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] | 2423 | const ActivationLayerInfo &act_info) |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 2424 | { |
| 2425 | TensorShape dst_shape = lhs_shape; |
| 2426 | dst_shape.set(0, rhs_shape[0]); |
| 2427 | dst_shape.set(1, lhs_shape[1] / m_h); |
| 2428 | dst_shape.set(2, m_h); |
| 2429 | dst_shape.set(3, lhs_shape[2]); |
| 2430 | |
| 2431 | // Create reference |
| 2432 | SimpleTensor<T> lhs{ lhs_shape, data_type, 1 }; |
| 2433 | SimpleTensor<T> rhs{ rhs_shape, data_type, 1 }; |
Gian Marco Iodice | 944170e | 2019-06-24 14:40:30 +0100 | [diff] [blame] | 2434 | SimpleTensor<T> bias{ dst_shape, data_type, 1 }; |
| 2435 | |
| 2436 | const int n = rhs_shape[0]; |
| 2437 | const int m = lhs_shape[1]; |
| 2438 | const int batch_size = lhs_shape[2]; |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 2439 | |
| 2440 | // Fill reference |
| 2441 | fill(lhs, 0); |
| 2442 | fill(rhs, 1); |
Gian Marco Iodice | 944170e | 2019-06-24 14:40:30 +0100 | [diff] [blame] | 2443 | fill(bias, 2); |
giuros01 | b3204e7 | 2019-04-01 13:50:22 +0100 | [diff] [blame] | 2444 | |
Gunes Bayir | 4bfc70e | 2021-12-10 16:17:56 +0000 | [diff] [blame] | 2445 | // In case of broadcast, we need to simply copy the first into the following "M" ones |
Gian Marco Iodice | 944170e | 2019-06-24 14:40:30 +0100 | [diff] [blame] | 2446 | for(int i = 1; i < m * batch_size; i++) |
| 2447 | { |
| 2448 | memcpy(bias.data() + i * n, bias.data(), n * sizeof(T)); |
| 2449 | } |
| 2450 | |
Gian Marco Iodice | ca1f460 | 2019-07-16 15:46:48 +0100 | [diff] [blame] | 2451 | 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] | 2452 | } |
| 2453 | |
| 2454 | TensorType _target{}; |
| 2455 | SimpleTensor<T> _reference{}; |
| 2456 | }; |
| 2457 | |
Gunes Bayir | 4bfc70e | 2021-12-10 16:17:56 +0000 | [diff] [blame] | 2458 | template <typename TensorType, typename AccessorType, typename T, typename ReshapeRHSOperatorType, typename GEMMOperatorType> |
| 2459 | class GEMMMatrixMultiplyReshapedOnlyRhsMMULValidationFixture : public framework::Fixture |
| 2460 | { |
| 2461 | public: |
| 2462 | template <typename...> |
| 2463 | void setup(unsigned int m, unsigned int n, unsigned int k, unsigned int batch_size, unsigned int m0, unsigned int n0, unsigned int k0, bool export_to_cl_image, DataType data_type, float alpha, |
| 2464 | float beta, bool broadcast_bias, |
| 2465 | const ActivationLayerInfo &act_info) |
| 2466 | { |
| 2467 | GEMMLHSMatrixInfo lhs_info; |
| 2468 | lhs_info.m0 = m0; |
| 2469 | lhs_info.k0 = k0; |
| 2470 | |
| 2471 | GEMMRHSMatrixInfo rhs_info; |
| 2472 | rhs_info.n0 = n0; |
| 2473 | rhs_info.k0 = k0; |
| 2474 | rhs_info.interleave = true; |
| 2475 | rhs_info.transpose = false; |
| 2476 | rhs_info.h0 = 4; |
| 2477 | rhs_info.export_to_cl_image = export_to_cl_image; |
| 2478 | |
| 2479 | // Set the tensor shapes for LHS and RHS matrices |
| 2480 | const TensorShape lhs_shape(k, m, batch_size); |
| 2481 | const TensorShape rhs_shape(n, k, batch_size); |
| 2482 | const TensorShape bias_shape(n, |
| 2483 | broadcast_bias ? 1 : m, |
| 2484 | broadcast_bias ? 1 : batch_size); |
| 2485 | |
| 2486 | _target = compute_target(lhs_shape, rhs_shape, bias_shape, lhs_info, rhs_info, data_type, alpha, beta, broadcast_bias, act_info); |
| 2487 | _reference = compute_reference(lhs_shape, rhs_shape, data_type, alpha, beta, broadcast_bias, act_info); |
| 2488 | } |
| 2489 | |
| 2490 | protected: |
| 2491 | template <typename U> |
| 2492 | void fill(U &&tensor, int i) |
| 2493 | { |
| 2494 | static_assert(std::is_floating_point<T>::value || std::is_same<T, half>::value, "Only floating point data types supported."); |
| 2495 | using DistributionType = typename std::conditional<std::is_same<T, half>::value, arm_compute::utils::uniform_real_distribution_16bit<T>, std::uniform_real_distribution<T>>::type; |
| 2496 | |
| 2497 | DistributionType distribution{ T(-1.0f), T(1.0f) }; |
| 2498 | library->fill(tensor, distribution, i); |
| 2499 | |
| 2500 | // Fill border with infinity in order to check the presence of NaN values (i.e. inf * 0) |
| 2501 | DistributionType distribution_inf{ T(std::numeric_limits<float>::infinity()), T(std::numeric_limits<float>::infinity()) }; |
| 2502 | library->fill_borders_with_garbage(tensor, distribution_inf, i); |
| 2503 | } |
| 2504 | |
| 2505 | TensorType compute_target(const TensorShape &lhs_shape, const TensorShape &rhs_shape, const TensorShape &bias_shape, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, |
| 2506 | DataType data_type, float alpha, float beta, bool broadcast_bias, const ActivationLayerInfo &act_info) |
| 2507 | { |
| 2508 | // Create tensors |
| 2509 | TensorType lhs = create_tensor<TensorType>(lhs_shape, data_type, 1); |
| 2510 | TensorType rhs = create_tensor<TensorType>(rhs_shape, data_type, 1); |
| 2511 | TensorType bias = create_tensor<TensorType>(bias_shape, data_type, 1); |
| 2512 | TensorType rhs_reshaped; |
| 2513 | TensorType dst; |
| 2514 | |
| 2515 | const unsigned int M = lhs_shape[1]; |
| 2516 | const unsigned int N = rhs_shape[0]; |
| 2517 | const unsigned int K = lhs_shape[0]; |
| 2518 | GEMMKernelInfo kernel_info; |
| 2519 | kernel_info.m = M; |
| 2520 | kernel_info.n = N; |
| 2521 | kernel_info.k = K; |
| 2522 | kernel_info.depth_output_gemm3d = 0; |
| 2523 | kernel_info.reinterpret_input_as_3d = false; |
| 2524 | kernel_info.broadcast_bias = broadcast_bias; |
| 2525 | kernel_info.activation_info = act_info; |
| 2526 | |
| 2527 | // Create and configure function |
| 2528 | ReshapeRHSOperatorType reshape_rhs; |
| 2529 | GEMMOperatorType gemm; |
| 2530 | |
| 2531 | validate_result = bool(reshape_rhs.validate(rhs.info(), rhs_reshaped.info(), rhs_info)); |
| 2532 | if(!validate_result) |
| 2533 | { |
| 2534 | return nullptr; |
| 2535 | } |
| 2536 | |
| 2537 | reshape_rhs.configure(rhs.info(), rhs_reshaped.info(), rhs_info); |
| 2538 | |
| 2539 | validate_result = bool(gemm.validate(lhs.info(), rhs_reshaped.info(), bias.info(), dst.info(), alpha, beta, lhs_info, rhs_info, kernel_info)); |
| 2540 | if(!validate_result) |
| 2541 | { |
| 2542 | return nullptr; |
| 2543 | } |
| 2544 | |
| 2545 | gemm.configure(lhs.info(), rhs_reshaped.info(), bias.info(), dst.info(), alpha, beta, lhs_info, rhs_info, kernel_info); |
| 2546 | |
| 2547 | ARM_COMPUTE_ASSERT(lhs.info()->is_resizable()); |
| 2548 | ARM_COMPUTE_ASSERT(rhs.info()->is_resizable()); |
| 2549 | ARM_COMPUTE_ASSERT(bias.info()->is_resizable()); |
| 2550 | |
| 2551 | // Allocate tensors |
| 2552 | lhs.allocator()->allocate(); |
| 2553 | rhs.allocator()->allocate(); |
| 2554 | rhs_reshaped.allocator()->allocate(); |
| 2555 | bias.allocator()->allocate(); |
| 2556 | dst.allocator()->allocate(); |
| 2557 | |
| 2558 | ARM_COMPUTE_ASSERT(!lhs.info()->is_resizable()); |
| 2559 | ARM_COMPUTE_ASSERT(!rhs.info()->is_resizable()); |
| 2560 | ARM_COMPUTE_ASSERT(!rhs_reshaped.info()->is_resizable()); |
| 2561 | ARM_COMPUTE_ASSERT(!bias.info()->is_resizable()); |
| 2562 | ARM_COMPUTE_ASSERT(!dst.info()->is_resizable()); |
| 2563 | |
| 2564 | // Fill tensors |
| 2565 | fill(AccessorType(lhs), 0); |
| 2566 | fill(AccessorType(rhs), 1); |
| 2567 | fill(AccessorType(bias), 2); |
| 2568 | |
| 2569 | // Compute GEMM |
| 2570 | ITensorPack reshape_rhs_pack = { { ACL_SRC, &rhs }, { ACL_DST, &rhs_reshaped } }; |
| 2571 | reshape_rhs.run(reshape_rhs_pack); |
| 2572 | ITensorPack gemm_pack({ { ACL_SRC_0, &lhs }, |
| 2573 | { ACL_SRC_1, &rhs_reshaped }, |
| 2574 | { ACL_SRC_2, &bias }, |
| 2575 | { ACL_DST, &dst } |
| 2576 | }); |
| 2577 | gemm.run(gemm_pack); |
| 2578 | |
| 2579 | return dst; |
| 2580 | } |
| 2581 | |
| 2582 | SimpleTensor<T> compute_reference(const TensorShape &lhs_shape, const TensorShape &rhs_shape, DataType data_type, float alpha, float beta, bool broadcast_bias, |
| 2583 | const ActivationLayerInfo &act_info) |
| 2584 | { |
| 2585 | if(!validate_result) |
| 2586 | return SimpleTensor<T>(); |
| 2587 | |
| 2588 | TensorShape dst_shape = lhs_shape; |
| 2589 | dst_shape[0] = rhs_shape[0]; |
| 2590 | dst_shape[1] = lhs_shape[1]; |
| 2591 | |
| 2592 | // Create reference |
| 2593 | SimpleTensor<T> lhs{ lhs_shape, data_type, 1 }; |
| 2594 | SimpleTensor<T> rhs{ rhs_shape, data_type, 1 }; |
| 2595 | SimpleTensor<T> bias{ dst_shape, data_type, 1 }; |
| 2596 | |
| 2597 | const int n = rhs_shape[0]; |
| 2598 | const int m = lhs_shape[1]; |
| 2599 | const int batch_size = lhs_shape[2]; |
| 2600 | |
| 2601 | // Fill reference |
| 2602 | fill(lhs, 0); |
| 2603 | fill(rhs, 1); |
| 2604 | fill(bias, 2); |
| 2605 | |
| 2606 | if(broadcast_bias) |
| 2607 | { |
| 2608 | // In case of broadcast, we need to simply copy the first into the following "M" ones |
| 2609 | for(int i = 1; i < m * batch_size; i++) |
| 2610 | { |
| 2611 | memcpy(bias.data() + i * n, bias.data(), n * sizeof(T)); |
| 2612 | } |
| 2613 | } |
| 2614 | |
| 2615 | return reference::activation_layer(reference::gemm<T>(lhs, rhs, bias, alpha, beta), act_info); |
| 2616 | } |
| 2617 | |
| 2618 | bool validate_result = true; |
| 2619 | TensorType _target{}; |
| 2620 | SimpleTensor<T> _reference{}; |
| 2621 | }; |
| 2622 | |
Moritz Pflanzer | 4dfc235 | 2017-08-02 14:51:36 +0100 | [diff] [blame] | 2623 | } // namespace validation |
| 2624 | } // namespace test |
| 2625 | } // namespace arm_compute |
| 2626 | #endif /* ARM_COMPUTE_TEST_GEMM_FIXTURE */ |