Gian Marco Iodice | ab18212 | 2017-10-09 15:05:40 +0100 | [diff] [blame] | 1 | /* |
Isabella Gottardi | e6630e4 | 2018-01-18 15:50:39 +0000 | [diff] [blame] | 2 | * Copyright (c) 2017-2018 ARM Limited. |
Gian Marco Iodice | ab18212 | 2017-10-09 15:05:40 +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 | #include "arm_compute/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.h" |
| 25 | |
| 26 | #include "arm_compute/core/Error.h" |
| 27 | #include "arm_compute/core/Helpers.h" |
| 28 | #include "arm_compute/core/ITensor.h" |
| 29 | #include "arm_compute/core/NEON/kernels/NEGEMMInterleave4x4Kernel.h" |
Gian Marco Iodice | ab18212 | 2017-10-09 15:05:40 +0100 | [diff] [blame] | 30 | #include "arm_compute/core/NEON/kernels/NEGEMMLowpMatrixMultiplyKernel.h" |
| 31 | #include "arm_compute/core/NEON/kernels/NEGEMMTranspose1xWKernel.h" |
Gian Marco Iodice | ab18212 | 2017-10-09 15:05:40 +0100 | [diff] [blame] | 32 | #include "arm_compute/core/TensorInfo.h" |
| 33 | #include "arm_compute/core/Types.h" |
| 34 | #include "arm_compute/core/Validate.h" |
Isabella Gottardi | e6630e4 | 2018-01-18 15:50:39 +0000 | [diff] [blame] | 35 | #include "arm_compute/core/utils/misc/ShapeCalculator.h" |
Gian Marco Iodice | ab18212 | 2017-10-09 15:05:40 +0100 | [diff] [blame] | 36 | #include "arm_compute/runtime/NEON/NEScheduler.h" |
| 37 | #include "arm_compute/runtime/TensorAllocator.h" |
| 38 | #include "support/ToolchainSupport.h" |
| 39 | |
| 40 | using namespace arm_compute; |
Isabella Gottardi | e6630e4 | 2018-01-18 15:50:39 +0000 | [diff] [blame] | 41 | using namespace arm_compute::misc::shape_calculator; |
Gian Marco Iodice | ab18212 | 2017-10-09 15:05:40 +0100 | [diff] [blame] | 42 | |
| 43 | NEGEMMLowpMatrixMultiplyCore::NEGEMMLowpMatrixMultiplyCore(std::shared_ptr<IMemoryManager> memory_manager) |
Anthony Barbier | eaefd00 | 2018-07-20 17:49:35 +0100 | [diff] [blame] | 44 | : _memory_group(memory_manager), _asm_glue(memory_manager), _mm_kernel(nullptr), _mtx_a_reshape_kernel(nullptr), _mtx_b_reshape_kernel(nullptr), _mtx_a_reduction_kernel(), _mtx_b_reduction_kernel(), |
| 45 | _offset_contribution_kernel(), _vector_sum_col(), _vector_sum_row(), _tmp_a(), _tmp_b(), _original_b(nullptr), _a_offset(0), _b_offset(0), _run_vector_matrix_multiplication(false), |
| 46 | _dot_product_path(false), _reshape_b_only_on_first_run(false), _is_prepared(false) |
Gian Marco Iodice | ab18212 | 2017-10-09 15:05:40 +0100 | [diff] [blame] | 47 | { |
| 48 | } |
| 49 | |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 50 | void NEGEMMLowpMatrixMultiplyCore::configure(const ITensor *a, const ITensor *b, const ITensor *c, ITensor *output, const GEMMInfo &gemm_info) |
Gian Marco Iodice | ab18212 | 2017-10-09 15:05:40 +0100 | [diff] [blame] | 51 | { |
Georgios Pinitas | a3b1b46 | 2017-11-16 19:24:39 +0000 | [diff] [blame] | 52 | ARM_COMPUTE_ERROR_ON_NULLPTR(a, b, output); |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 53 | ARM_COMPUTE_UNUSED(c); |
| 54 | ARM_COMPUTE_ERROR_THROW_ON(NEGEMMLowpMatrixMultiplyCore::validate(a->info(), b->info(), c != nullptr ? c->info() : nullptr, output->info(), gemm_info)); |
Gian Marco Iodice | ab18212 | 2017-10-09 15:05:40 +0100 | [diff] [blame] | 55 | |
Georgios Pinitas | 7221933 | 2018-06-05 14:56:06 +0100 | [diff] [blame] | 56 | // Clear state |
Anthony Barbier | 71d9b57 | 2018-07-06 17:05:59 +0100 | [diff] [blame] | 57 | _mtx_a_reshape_kernel = nullptr; |
| 58 | _mtx_b_reshape_kernel = nullptr; |
Georgios Pinitas | 7221933 | 2018-06-05 14:56:06 +0100 | [diff] [blame] | 59 | |
| 60 | // Set internal variables |
Gian Marco | c7f9b89 | 2017-11-30 14:31:13 +0000 | [diff] [blame] | 61 | _a_offset = a->info()->quantization_info().offset; |
| 62 | _b_offset = b->info()->quantization_info().offset; |
| 63 | _run_vector_matrix_multiplication = a->info()->dimension(1) < 2; |
Giorgio Arena | bb54e4e | 2018-04-05 17:20:34 +0100 | [diff] [blame] | 64 | _reshape_b_only_on_first_run = gemm_info.reshape_b_only_on_first_run(); |
Georgios Pinitas | 7221933 | 2018-06-05 14:56:06 +0100 | [diff] [blame] | 65 | _is_prepared = false; |
| 66 | _original_b = b; |
Gian Marco | e75a02b | 2017-11-08 12:24:09 +0000 | [diff] [blame] | 67 | |
Pablo Tello | eb82fd2 | 2018-02-23 13:43:50 +0000 | [diff] [blame] | 68 | #ifdef __aarch64__ |
| 69 | switch(a->info()->data_type()) |
Gian Marco Iodice | ab18212 | 2017-10-09 15:05:40 +0100 | [diff] [blame] | 70 | { |
Pablo Tello | 66c656a | 2018-03-15 10:34:58 +0000 | [diff] [blame] | 71 | case DataType::QASYMM8: |
Pablo Tello | eb82fd2 | 2018-02-23 13:43:50 +0000 | [diff] [blame] | 72 | case DataType::U8: |
Anthony Barbier | eaefd00 | 2018-07-20 17:49:35 +0100 | [diff] [blame] | 73 | case DataType::S8: |
Pablo Tello | eb82fd2 | 2018-02-23 13:43:50 +0000 | [diff] [blame] | 74 | { |
Anthony Barbier | eaefd00 | 2018-07-20 17:49:35 +0100 | [diff] [blame] | 75 | _asm_glue.configure(a, b, output, 1.f, 0.f, _reshape_b_only_on_first_run); |
| 76 | _dot_product_path = _asm_glue.is_configured(); |
Pablo Tello | eb82fd2 | 2018-02-23 13:43:50 +0000 | [diff] [blame] | 77 | break; |
| 78 | } |
| 79 | default: |
| 80 | { |
| 81 | ARM_COMPUTE_ERROR("Datatype not supported"); |
| 82 | break; |
| 83 | } |
Gian Marco Iodice | ab18212 | 2017-10-09 15:05:40 +0100 | [diff] [blame] | 84 | } |
Pablo Tello | eb82fd2 | 2018-02-23 13:43:50 +0000 | [diff] [blame] | 85 | #endif /* __aarch64__ */ |
| 86 | if(!_dot_product_path) |
Gian Marco Iodice | ab18212 | 2017-10-09 15:05:40 +0100 | [diff] [blame] | 87 | { |
Gian Marco | c7f9b89 | 2017-11-30 14:31:13 +0000 | [diff] [blame] | 88 | if(_run_vector_matrix_multiplication) |
Gian Marco Iodice | ab18212 | 2017-10-09 15:05:40 +0100 | [diff] [blame] | 89 | { |
Gian Marco | c7f9b89 | 2017-11-30 14:31:13 +0000 | [diff] [blame] | 90 | // Configure matrix multiply kernel |
| 91 | { |
| 92 | auto k = arm_compute::support::cpp14::make_unique<NEGEMMLowpMatrixMultiplyKernel>(); |
| 93 | k->configure(a, b, output); |
| 94 | _mm_kernel = std::move(k); |
| 95 | } |
Gian Marco Iodice | ab18212 | 2017-10-09 15:05:40 +0100 | [diff] [blame] | 96 | } |
Gian Marco | c7f9b89 | 2017-11-30 14:31:13 +0000 | [diff] [blame] | 97 | else |
Gian Marco Iodice | ab18212 | 2017-10-09 15:05:40 +0100 | [diff] [blame] | 98 | { |
Gian Marco | c7f9b89 | 2017-11-30 14:31:13 +0000 | [diff] [blame] | 99 | // The interleaved output matrix will have the following shape: [ a_height * 4, ceil(a_width / 4.0f) ] |
Isabella Gottardi | e6630e4 | 2018-01-18 15:50:39 +0000 | [diff] [blame] | 100 | TensorInfo info_a(compute_interleaved_shape(*a->info()), 1, a->info()->data_type()); |
Gian Marco | c7f9b89 | 2017-11-30 14:31:13 +0000 | [diff] [blame] | 101 | // The transpose1xW output matrix will have the following shape: [ b_height * 16, ceil(b_width / 16.0f) ] |
Isabella Gottardi | e6630e4 | 2018-01-18 15:50:39 +0000 | [diff] [blame] | 102 | TensorInfo info_b(compute_transpose1xW_shape(*b->info()), 1, b->info()->data_type()); |
Gian Marco | c7f9b89 | 2017-11-30 14:31:13 +0000 | [diff] [blame] | 103 | _tmp_a.allocator()->init(info_a); |
| 104 | _tmp_b.allocator()->init(info_b); |
| 105 | _memory_group.manage(&_tmp_a); |
Giorgio Arena | bb54e4e | 2018-04-05 17:20:34 +0100 | [diff] [blame] | 106 | if(!_reshape_b_only_on_first_run) |
| 107 | { |
| 108 | _memory_group.manage(&_tmp_b); |
| 109 | } |
Gian Marco | c7f9b89 | 2017-11-30 14:31:13 +0000 | [diff] [blame] | 110 | |
| 111 | // Configure interleave kernel |
| 112 | { |
| 113 | auto k = arm_compute::support::cpp14::make_unique<NEGEMMInterleave4x4Kernel>(); |
| 114 | k->configure(a, &_tmp_a); |
| 115 | _mtx_a_reshape_kernel = std::move(k); |
| 116 | } |
| 117 | |
| 118 | // Configure transpose kernel |
| 119 | { |
| 120 | auto k = arm_compute::support::cpp14::make_unique<NEGEMMTranspose1xWKernel>(); |
| 121 | k->configure(b, &_tmp_b); |
| 122 | _mtx_b_reshape_kernel = std::move(k); |
| 123 | } |
| 124 | |
| 125 | // Configure matrix multiply kernel |
| 126 | { |
| 127 | auto k = arm_compute::support::cpp14::make_unique<NEGEMMLowpMatrixMultiplyKernel>(); |
| 128 | k->configure(&_tmp_a, &_tmp_b, output); |
| 129 | _mm_kernel = std::move(k); |
| 130 | } |
Gian Marco Iodice | ab18212 | 2017-10-09 15:05:40 +0100 | [diff] [blame] | 131 | } |
Gian Marco | e75a02b | 2017-11-08 12:24:09 +0000 | [diff] [blame] | 132 | } |
Gian Marco Iodice | ab18212 | 2017-10-09 15:05:40 +0100 | [diff] [blame] | 133 | |
Gian Marco | e75a02b | 2017-11-08 12:24:09 +0000 | [diff] [blame] | 134 | // Initialize matrix B reduction kernel only if _a_offset is not equal to 0 |
| 135 | if(_a_offset != 0) |
| 136 | { |
Isabella Gottardi | e6630e4 | 2018-01-18 15:50:39 +0000 | [diff] [blame] | 137 | TensorInfo info_vector_sum_col(compute_reductionA_shape(*b->info()), 1, DataType::S32); |
| 138 | |
Gian Marco | e75a02b | 2017-11-08 12:24:09 +0000 | [diff] [blame] | 139 | _vector_sum_col.allocator()->init(info_vector_sum_col); |
Giorgio Arena | bb54e4e | 2018-04-05 17:20:34 +0100 | [diff] [blame] | 140 | if(!_reshape_b_only_on_first_run) |
| 141 | { |
| 142 | _memory_group.manage(&_vector_sum_col); |
| 143 | } |
Gian Marco | e75a02b | 2017-11-08 12:24:09 +0000 | [diff] [blame] | 144 | |
| 145 | // Configure Matrix B reduction kernel |
| 146 | _mtx_b_reduction_kernel.configure(b, &_vector_sum_col, a->info()->dimension(0), false); |
| 147 | } |
| 148 | |
| 149 | // Initialize Matrix A reduction kernel only if _b_offset is not equal to 0 |
| 150 | if(_b_offset != 0) |
| 151 | { |
Isabella Gottardi | e6630e4 | 2018-01-18 15:50:39 +0000 | [diff] [blame] | 152 | TensorInfo info_vector_sum_row(compute_reductionB_shape(*a->info()), 1, DataType::S32); |
| 153 | |
Gian Marco | e75a02b | 2017-11-08 12:24:09 +0000 | [diff] [blame] | 154 | _vector_sum_row.allocator()->init(info_vector_sum_row); |
| 155 | _memory_group.manage(&_vector_sum_row); |
| 156 | |
| 157 | // Configure matrix A reduction kernel |
| 158 | _mtx_a_reduction_kernel.configure(a, &_vector_sum_row, a->info()->dimension(0), false); |
| 159 | } |
| 160 | |
| 161 | // Configure offset contribution kernel |
| 162 | _offset_contribution_kernel.configure(output, _a_offset == 0 ? nullptr : &_vector_sum_col, _b_offset == 0 ? nullptr : &_vector_sum_row, a->info()->dimension(0), _a_offset, _b_offset); |
| 163 | |
| 164 | // Allocate tensors |
Gian Marco | c7f9b89 | 2017-11-30 14:31:13 +0000 | [diff] [blame] | 165 | if(!_dot_product_path && !_run_vector_matrix_multiplication) |
Gian Marco | e75a02b | 2017-11-08 12:24:09 +0000 | [diff] [blame] | 166 | { |
Pablo Tello | 6ff12a0 | 2017-11-02 16:09:35 +0000 | [diff] [blame] | 167 | _tmp_a.allocator()->allocate(); |
Georgios Pinitas | 7221933 | 2018-06-05 14:56:06 +0100 | [diff] [blame] | 168 | if(!_reshape_b_only_on_first_run) |
| 169 | { |
| 170 | _tmp_b.allocator()->allocate(); |
| 171 | } |
Pablo Tello | 6ff12a0 | 2017-11-02 16:09:35 +0000 | [diff] [blame] | 172 | } |
Gian Marco | e75a02b | 2017-11-08 12:24:09 +0000 | [diff] [blame] | 173 | |
Georgios Pinitas | 7221933 | 2018-06-05 14:56:06 +0100 | [diff] [blame] | 174 | if(_a_offset != 0 && !_reshape_b_only_on_first_run) |
Gian Marco | e75a02b | 2017-11-08 12:24:09 +0000 | [diff] [blame] | 175 | { |
| 176 | _vector_sum_col.allocator()->allocate(); |
| 177 | } |
| 178 | |
| 179 | if(_b_offset != 0) |
| 180 | { |
| 181 | _vector_sum_row.allocator()->allocate(); |
| 182 | } |
Gian Marco Iodice | ab18212 | 2017-10-09 15:05:40 +0100 | [diff] [blame] | 183 | } |
| 184 | |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 185 | Status NEGEMMLowpMatrixMultiplyCore::validate(const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c, const ITensorInfo *output, const GEMMInfo &gemm_info) |
Georgios Pinitas | a3b1b46 | 2017-11-16 19:24:39 +0000 | [diff] [blame] | 186 | { |
| 187 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(a, 1, DataType::QASYMM8); |
| 188 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::S32); |
| 189 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(a, b); |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 190 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(c != nullptr, "Bias addition not supported in NEGEMMLowpMatrixMultiplyCore"); |
Georgios Pinitas | a3b1b46 | 2017-11-16 19:24:39 +0000 | [diff] [blame] | 191 | ARM_COMPUTE_RETURN_ERROR_ON_MSG((a)->dimension(0) != (b)->dimension(1), |
| 192 | "The product AB is defined only if the number of columns in A is equal to the number of rows in B"); |
Chunosov | 5124be5 | 2017-11-22 20:42:13 +0700 | [diff] [blame] | 193 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(gemm_info.is_a_reshaped(), "Matrix A already reshaped is not supported"); |
| 194 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(gemm_info.is_b_reshaped(), "Matrix B already reshaped is not supported"); |
Georgios Pinitas | a3b1b46 | 2017-11-16 19:24:39 +0000 | [diff] [blame] | 195 | |
Georgios Pinitas | bb081ca | 2018-11-08 10:22:01 +0000 | [diff] [blame] | 196 | int32_t a_offset = a->quantization_info().offset; |
| 197 | int32_t b_offset = b->quantization_info().offset; |
| 198 | const bool reshape_b_only_on_first_run = gemm_info.reshape_b_only_on_first_run(); |
Georgios Pinitas | a3b1b46 | 2017-11-16 19:24:39 +0000 | [diff] [blame] | 199 | |
Georgios Pinitas | bb081ca | 2018-11-08 10:22:01 +0000 | [diff] [blame] | 200 | // Check if we need to run the optimized assembly kernel |
| 201 | const bool run_optimised = bool(NEGEMMAssemblyDispatch::validate(a, b, output, 1.f, 0.f, reshape_b_only_on_first_run)); |
| 202 | |
| 203 | if(run_optimised) |
Georgios Pinitas | a3b1b46 | 2017-11-16 19:24:39 +0000 | [diff] [blame] | 204 | { |
Georgios Pinitas | bb081ca | 2018-11-08 10:22:01 +0000 | [diff] [blame] | 205 | if(output->total_size() != 0) |
| 206 | { |
| 207 | ARM_COMPUTE_RETURN_ERROR_ON(b->dimension(0) != output->dimension(0)); |
| 208 | if(gemm_info.depth_output_gemm3d() != 0) |
| 209 | { |
| 210 | if(gemm_info.reinterpret_input_as_3d()) |
| 211 | { |
| 212 | ARM_COMPUTE_RETURN_ERROR_ON(a->dimension(1) != output->dimension(1)); |
| 213 | ARM_COMPUTE_RETURN_ERROR_ON(a->dimension(2) != output->dimension(2)); |
| 214 | } |
| 215 | else |
| 216 | { |
| 217 | ARM_COMPUTE_RETURN_ERROR_ON(a->dimension(1) != output->dimension(1) * output->dimension(2)); |
| 218 | } |
| 219 | } |
| 220 | else |
| 221 | { |
| 222 | ARM_COMPUTE_RETURN_ERROR_ON(a->dimension(1) != output->dimension(1)); |
| 223 | } |
| 224 | } |
Georgios Pinitas | a3b1b46 | 2017-11-16 19:24:39 +0000 | [diff] [blame] | 225 | } |
| 226 | else |
Georgios Pinitas | a3b1b46 | 2017-11-16 19:24:39 +0000 | [diff] [blame] | 227 | { |
Georgios Pinitas | bb081ca | 2018-11-08 10:22:01 +0000 | [diff] [blame] | 228 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(gemm_info.reinterpret_input_as_3d(), "NEGEMM cannot reinterpret the input tensor as 3D"); |
| 229 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(gemm_info.depth_output_gemm3d() != 0, "NEGEMM cannot reinterpret the output tensor as 3D"); |
| 230 | |
| 231 | const bool run_vector_matrix_multiplication = a->dimension(1) < 2; |
| 232 | if(!run_vector_matrix_multiplication) |
| 233 | { |
| 234 | // The interleaved output matrix will have the following shape: [ a_height * 4, ceil(a_width / 4.0f) ] |
| 235 | TensorShape shape_tmp_a = a->tensor_shape(); |
| 236 | shape_tmp_a.set(0, a->dimension(0) * 4); |
| 237 | shape_tmp_a.set(1, std::ceil(a->dimension(1) / 4.f)); |
| 238 | |
| 239 | // The transpose1xW output matrix will have the following shape: [ b_height * 16, ceil(b_width / 16.0f) ] |
| 240 | TensorShape shape_tmp_b = b->tensor_shape(); |
| 241 | shape_tmp_b.set(0, b->dimension(1) * 16); |
| 242 | shape_tmp_b.set(1, std::ceil(b->dimension(0) / 16.f)); |
| 243 | |
| 244 | TensorInfo info_a(shape_tmp_a, 1, a->data_type()); |
| 245 | TensorInfo info_b(shape_tmp_b, 1, b->data_type()); |
| 246 | |
| 247 | ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMInterleave4x4Kernel::validate(a, &info_a)); |
| 248 | ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMTranspose1xWKernel::validate(b, &info_b)); |
| 249 | ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMLowpMatrixMultiplyKernel::validate(&info_a, &info_b, output)); |
| 250 | } |
| 251 | else |
| 252 | { |
| 253 | ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMLowpMatrixMultiplyKernel::validate(a, b, output)); |
| 254 | } |
Georgios Pinitas | a3b1b46 | 2017-11-16 19:24:39 +0000 | [diff] [blame] | 255 | } |
| 256 | |
| 257 | TensorInfo info_vector_sum_col, info_vector_sum_row; |
| 258 | |
| 259 | // Validate matrix B reduction kernel only if _a_offset is not equal to 0 |
| 260 | if(a_offset != 0) |
| 261 | { |
Isabella Gottardi | e6630e4 | 2018-01-18 15:50:39 +0000 | [diff] [blame] | 262 | info_vector_sum_col = TensorInfo(compute_reductionA_shape(*b), 1, DataType::S32); |
Georgios Pinitas | a3b1b46 | 2017-11-16 19:24:39 +0000 | [diff] [blame] | 263 | |
| 264 | // Configure Matrix B reduction kernel |
| 265 | ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMLowpMatrixBReductionKernel::validate(b, &info_vector_sum_col, a->dimension(0), false)); |
| 266 | } |
| 267 | |
| 268 | // Validate Matrix A reduction kernel only if _b_offset is not equal to 0 |
| 269 | if(b_offset != 0) |
| 270 | { |
Isabella Gottardi | e6630e4 | 2018-01-18 15:50:39 +0000 | [diff] [blame] | 271 | info_vector_sum_row = TensorInfo(compute_reductionB_shape(*a), 1, DataType::S32); |
Georgios Pinitas | a3b1b46 | 2017-11-16 19:24:39 +0000 | [diff] [blame] | 272 | |
| 273 | // Configure matrix A reduction kernel |
| 274 | ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMLowpMatrixAReductionKernel::validate(a, &info_vector_sum_row, a->dimension(0), false)); |
| 275 | } |
| 276 | |
| 277 | // Validate offset contribution kernel |
| 278 | ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMLowpOffsetContributionKernel::validate(output, |
| 279 | a_offset == 0 ? nullptr : &info_vector_sum_col, |
| 280 | b_offset == 0 ? nullptr : &info_vector_sum_row, |
| 281 | a_offset, b_offset)); |
| 282 | |
Georgios Pinitas | 631c41a | 2017-12-06 11:53:03 +0000 | [diff] [blame] | 283 | return Status{}; |
Georgios Pinitas | a3b1b46 | 2017-11-16 19:24:39 +0000 | [diff] [blame] | 284 | } |
| 285 | |
Gian Marco Iodice | ab18212 | 2017-10-09 15:05:40 +0100 | [diff] [blame] | 286 | void NEGEMMLowpMatrixMultiplyCore::run() |
| 287 | { |
Georgios Pinitas | 7221933 | 2018-06-05 14:56:06 +0100 | [diff] [blame] | 288 | prepare(); |
| 289 | |
Gian Marco Iodice | ab18212 | 2017-10-09 15:05:40 +0100 | [diff] [blame] | 290 | _memory_group.acquire(); |
| 291 | |
Georgios Pinitas | 7221933 | 2018-06-05 14:56:06 +0100 | [diff] [blame] | 292 | // Reshape inputs |
| 293 | if(_mtx_a_reshape_kernel) |
Pablo Tello | 6ff12a0 | 2017-11-02 16:09:35 +0000 | [diff] [blame] | 294 | { |
Georgios Pinitas | 7221933 | 2018-06-05 14:56:06 +0100 | [diff] [blame] | 295 | NEScheduler::get().schedule(_mtx_a_reshape_kernel.get(), Window::DimY); |
| 296 | } |
| 297 | if(_mtx_b_reshape_kernel && !_reshape_b_only_on_first_run) |
| 298 | { |
| 299 | NEScheduler::get().schedule(_mtx_b_reshape_kernel.get(), Window::DimY); |
Pablo Tello | 6ff12a0 | 2017-11-02 16:09:35 +0000 | [diff] [blame] | 300 | } |
Gian Marco Iodice | ab18212 | 2017-10-09 15:05:40 +0100 | [diff] [blame] | 301 | |
Georgios Pinitas | 7221933 | 2018-06-05 14:56:06 +0100 | [diff] [blame] | 302 | // Run GEMM |
Anthony Barbier | eaefd00 | 2018-07-20 17:49:35 +0100 | [diff] [blame] | 303 | if(_asm_glue.is_configured()) |
Pablo Tello | eb82fd2 | 2018-02-23 13:43:50 +0000 | [diff] [blame] | 304 | { |
Anthony Barbier | eaefd00 | 2018-07-20 17:49:35 +0100 | [diff] [blame] | 305 | _asm_glue.run(); |
Pablo Tello | eb82fd2 | 2018-02-23 13:43:50 +0000 | [diff] [blame] | 306 | } |
| 307 | else |
| 308 | { |
| 309 | NEScheduler::get().schedule(_mm_kernel.get(), Window::DimY); |
| 310 | } |
Gian Marco Iodice | ab18212 | 2017-10-09 15:05:40 +0100 | [diff] [blame] | 311 | |
Gian Marco | e75a02b | 2017-11-08 12:24:09 +0000 | [diff] [blame] | 312 | // Run matrix A reduction kernel only if _b_offset is not equal to 0 |
| 313 | if(_b_offset != 0) |
| 314 | { |
| 315 | NEScheduler::get().schedule(&_mtx_a_reduction_kernel, Window::DimX); |
| 316 | } |
| 317 | |
| 318 | // Run matrix B reduction kernel only if _a_offset is not equal to 0 |
Georgios Pinitas | 7221933 | 2018-06-05 14:56:06 +0100 | [diff] [blame] | 319 | if(_a_offset != 0 && !_reshape_b_only_on_first_run) |
Gian Marco | e75a02b | 2017-11-08 12:24:09 +0000 | [diff] [blame] | 320 | { |
| 321 | NEScheduler::get().schedule(&_mtx_b_reduction_kernel, Window::DimX); |
| 322 | } |
| 323 | |
| 324 | // Run offset contribution kernel |
| 325 | NEScheduler::get().schedule(&_offset_contribution_kernel, Window::DimY); |
| 326 | |
Gian Marco Iodice | ab18212 | 2017-10-09 15:05:40 +0100 | [diff] [blame] | 327 | _memory_group.release(); |
Georgios Pinitas | 7221933 | 2018-06-05 14:56:06 +0100 | [diff] [blame] | 328 | } |
Giorgio Arena | bb54e4e | 2018-04-05 17:20:34 +0100 | [diff] [blame] | 329 | |
Georgios Pinitas | 7221933 | 2018-06-05 14:56:06 +0100 | [diff] [blame] | 330 | void NEGEMMLowpMatrixMultiplyCore::prepare() |
| 331 | { |
| 332 | if(!_is_prepared) |
| 333 | { |
| 334 | // Run assembly reshape |
Anthony Barbier | eaefd00 | 2018-07-20 17:49:35 +0100 | [diff] [blame] | 335 | if(_asm_glue.is_configured() && _reshape_b_only_on_first_run) |
Georgios Pinitas | 7221933 | 2018-06-05 14:56:06 +0100 | [diff] [blame] | 336 | { |
| 337 | ARM_COMPUTE_ERROR_ON(!_original_b->is_used()); |
| 338 | |
Anthony Barbier | eaefd00 | 2018-07-20 17:49:35 +0100 | [diff] [blame] | 339 | _asm_glue.prepare(); |
Georgios Pinitas | 7221933 | 2018-06-05 14:56:06 +0100 | [diff] [blame] | 340 | _original_b->mark_as_unused(); |
| 341 | } |
| 342 | // Run non-assembly reshape |
| 343 | else if(_mtx_b_reshape_kernel && _reshape_b_only_on_first_run) |
| 344 | { |
| 345 | ARM_COMPUTE_ERROR_ON(!_original_b->is_used()); |
| 346 | |
| 347 | // Run reshape kernel and mark original weights tensor as unused |
| 348 | _tmp_b.allocator()->allocate(); |
| 349 | NEScheduler::get().schedule(_mtx_b_reshape_kernel.get(), Window::DimY); |
| 350 | _original_b->mark_as_unused(); |
| 351 | } |
| 352 | |
| 353 | // Run matrix B reduction kernel only if _a_offset is not equal to 0 |
| 354 | if(_a_offset != 0 && _reshape_b_only_on_first_run) |
| 355 | { |
| 356 | _vector_sum_col.allocator()->allocate(); |
| 357 | NEScheduler::get().schedule(&_mtx_b_reduction_kernel, Window::DimX); |
| 358 | } |
| 359 | |
| 360 | _is_prepared = true; |
| 361 | } |
Pablo Tello | 6ff12a0 | 2017-11-02 16:09:35 +0000 | [diff] [blame] | 362 | } |