Gian Marco Iodice | ab18212 | 2017-10-09 15:05:40 +0100 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2017 ARM Limited. |
| 3 | * |
| 4 | * SPDX-License-Identifier: MIT |
| 5 | * |
| 6 | * Permission is hereby granted, free of charge, to any person obtaining a copy |
| 7 | * of this software and associated documentation files (the "Software"), to |
| 8 | * deal in the Software without restriction, including without limitation the |
| 9 | * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or |
| 10 | * sell copies of the Software, and to permit persons to whom the Software is |
| 11 | * furnished to do so, subject to the following conditions: |
| 12 | * |
| 13 | * The above copyright notice and this permission notice shall be included in all |
| 14 | * copies or substantial portions of the Software. |
| 15 | * |
| 16 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR |
| 17 | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
| 18 | * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE |
| 19 | * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER |
| 20 | * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, |
| 21 | * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE |
| 22 | * SOFTWARE. |
| 23 | */ |
| 24 | #include "arm_compute/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" |
Pablo Tello | 6ff12a0 | 2017-11-02 16:09:35 +0000 | [diff] [blame] | 29 | #include "arm_compute/core/NEON/kernels/NEGEMMAssemblyBaseKernel.h" |
Gian Marco Iodice | ab18212 | 2017-10-09 15:05:40 +0100 | [diff] [blame] | 30 | #include "arm_compute/core/NEON/kernels/NEGEMMInterleave4x4Kernel.h" |
Gian Marco Iodice | ab18212 | 2017-10-09 15:05:40 +0100 | [diff] [blame] | 31 | #include "arm_compute/core/NEON/kernels/NEGEMMLowpMatrixMultiplyKernel.h" |
| 32 | #include "arm_compute/core/NEON/kernels/NEGEMMTranspose1xWKernel.h" |
| 33 | #include "arm_compute/core/NEON/kernels/arm64/NEGEMMLowpAArch64V8P4Kernel.h" |
| 34 | #include "arm_compute/core/TensorInfo.h" |
| 35 | #include "arm_compute/core/Types.h" |
| 36 | #include "arm_compute/core/Validate.h" |
| 37 | #include "arm_compute/runtime/NEON/NEScheduler.h" |
| 38 | #include "arm_compute/runtime/TensorAllocator.h" |
| 39 | #include "support/ToolchainSupport.h" |
| 40 | |
Pablo Tello | 6ff12a0 | 2017-11-02 16:09:35 +0000 | [diff] [blame] | 41 | namespace arm_compute |
| 42 | { |
| 43 | #include "arm_compute/core/NEON/kernels/assembly/gemm_interleaved.hpp" |
Pablo Tello | 6681d24 | 2017-11-13 16:44:08 +0000 | [diff] [blame^] | 44 | #include "arm_compute/core/NEON/kernels/assembly/kernels/a64_gemm_u8_12x8.hpp" |
Pablo Tello | 6ff12a0 | 2017-11-02 16:09:35 +0000 | [diff] [blame] | 45 | } // namespace arm_compute |
| 46 | |
Gian Marco Iodice | ab18212 | 2017-10-09 15:05:40 +0100 | [diff] [blame] | 47 | using namespace arm_compute; |
| 48 | |
| 49 | NEGEMMLowpMatrixMultiplyCore::NEGEMMLowpMatrixMultiplyCore(std::shared_ptr<IMemoryManager> memory_manager) |
Gian Marco | e75a02b | 2017-11-08 12:24:09 +0000 | [diff] [blame] | 50 | : _memory_group(std::move(memory_manager)), _mm_kernel(nullptr), _mtx_a_reshape_kernel(nullptr), _mtx_b_reshape_kernel(nullptr), _mtx_a_reduction_kernel(), _mtx_b_reduction_kernel(), |
| 51 | _offset_contribution_kernel(), _vector_sum_col(), _vector_sum_row(), _tmp_a(), _tmp_b(), _workspace(), _a_offset(0), _b_offset(0) |
Gian Marco Iodice | ab18212 | 2017-10-09 15:05:40 +0100 | [diff] [blame] | 52 | { |
| 53 | } |
| 54 | |
| 55 | void NEGEMMLowpMatrixMultiplyCore::configure(const ITensor *a, const ITensor *b, ITensor *output) |
| 56 | { |
Gian Marco | e75a02b | 2017-11-08 12:24:09 +0000 | [diff] [blame] | 57 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(a, 1, DataType::QASYMM8); |
Gian Marco Iodice | ab18212 | 2017-10-09 15:05:40 +0100 | [diff] [blame] | 58 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::S32); |
| 59 | ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(a, b); |
| 60 | ARM_COMPUTE_ERROR_ON_MSG((a)->info()->dimension(0) != (b)->info()->dimension(1), "The product AB is defined only if the number of columns in A is equal to the number of rows in B"); |
| 61 | ARM_COMPUTE_ERROR_ON_MSG((a)->info()->dimension(1) != (output)->info()->dimension(1), "The output matrix must have the same number of rows as the matrix A"); |
| 62 | ARM_COMPUTE_ERROR_ON_MSG((b)->info()->dimension(0) != (output)->info()->dimension(0), "The output matrix must have the same number of columns as the matrix B"); |
| 63 | |
Gian Marco | e75a02b | 2017-11-08 12:24:09 +0000 | [diff] [blame] | 64 | bool dot_product_path = false; |
| 65 | |
| 66 | _a_offset = a->info()->quantization_info().offset; |
| 67 | _b_offset = b->info()->quantization_info().offset; |
| 68 | |
Gian Marco Iodice | ab18212 | 2017-10-09 15:05:40 +0100 | [diff] [blame] | 69 | #ifdef ARM_COMPUTE_AARCH64_V8_2 |
| 70 | // Check for DOT product instruction |
| 71 | const struct CPUInfo ci = NEScheduler::get().cpu_info(); |
| 72 | const int cpu_has_dotprod = static_cast<int>(ci.CPU) & static_cast<int>(CPUTarget::DOT); |
| 73 | |
| 74 | if(cpu_has_dotprod != 0) |
| 75 | { |
Gian Marco | e75a02b | 2017-11-08 12:24:09 +0000 | [diff] [blame] | 76 | dot_product_path = true; |
| 77 | |
Gian Marco Iodice | ab18212 | 2017-10-09 15:05:40 +0100 | [diff] [blame] | 78 | // Configure matrix multiply kernel |
Pablo Tello | 6ff12a0 | 2017-11-02 16:09:35 +0000 | [diff] [blame] | 79 | struct CPUInfo ci = NEScheduler::get().cpu_info(); |
| 80 | const int M = output->info()->tensor_shape().y(); |
| 81 | const int N = output->info()->tensor_shape().x(); |
| 82 | const int K = a->info()->tensor_shape().x(); |
| 83 | |
Pablo Tello | 6681d24 | 2017-11-13 16:44:08 +0000 | [diff] [blame^] | 84 | const size_t workbench_size = GemmInterleaved<gemm_u8_12x8, gemm_u8_12x8::operand_type, gemm_u8_12x8::result_type>(&ci, M, N, K, false, false).get_working_size(); |
| 85 | constexpr size_t alignment = 4096; |
| 86 | _workspace.allocator()->init(TensorInfo(TensorShape{ (workbench_size + alignment - 1) * NEScheduler::get().num_threads() }, 1, DataType::U8)); |
Pablo Tello | 6ff12a0 | 2017-11-02 16:09:35 +0000 | [diff] [blame] | 87 | _memory_group.manage(&_workspace); |
Gian Marco | e75a02b | 2017-11-08 12:24:09 +0000 | [diff] [blame] | 88 | |
Pablo Tello | 6ff12a0 | 2017-11-02 16:09:35 +0000 | [diff] [blame] | 89 | // Configure matrix multiplication kernel |
| 90 | auto k = arm_compute::support::cpp14::make_unique<NEGEMMLowpAArch64V8P4Kernel>(); |
| 91 | k->configure(a, b, output, &_workspace, 1.f, 1.f); |
| 92 | _mm_kernel = std::move(k); |
Gian Marco Iodice | ab18212 | 2017-10-09 15:05:40 +0100 | [diff] [blame] | 93 | } |
| 94 | else |
| 95 | #endif /* ARM_COMPUTE_AARCH64_V8_2 */ |
| 96 | { |
| 97 | // The interleaved output matrix will have the following shape: [ a_height * 4, ceil(a_width / 4.0f) ] |
| 98 | TensorShape shape_tmp_a = a->info()->tensor_shape(); |
| 99 | shape_tmp_a.set(0, a->info()->dimension(0) * 4); |
| 100 | shape_tmp_a.set(1, std::ceil(a->info()->dimension(1) / 4.f)); |
| 101 | |
| 102 | // The transpose1xW output matrix will have the following shape: [ b_height * 16, ceil(b_width / 16.0f) ] |
| 103 | TensorShape shape_tmp_b = b->info()->tensor_shape(); |
| 104 | shape_tmp_b.set(0, b->info()->dimension(1) * 16); |
| 105 | shape_tmp_b.set(1, std::ceil(b->info()->dimension(0) / 16.f)); |
| 106 | |
| 107 | TensorInfo info_a(shape_tmp_a, 1, a->info()->data_type()); |
| 108 | TensorInfo info_b(shape_tmp_b, 1, b->info()->data_type()); |
| 109 | _tmp_a.allocator()->init(info_a); |
| 110 | _tmp_b.allocator()->init(info_b); |
| 111 | _memory_group.manage(&_tmp_a); |
| 112 | _memory_group.manage(&_tmp_b); |
| 113 | |
| 114 | // Configure interleave kernel |
| 115 | { |
| 116 | auto k = arm_compute::support::cpp14::make_unique<NEGEMMInterleave4x4Kernel>(); |
| 117 | k->configure(a, &_tmp_a); |
| 118 | _mtx_a_reshape_kernel = std::move(k); |
| 119 | } |
| 120 | |
| 121 | // Configure transpose kernel |
| 122 | { |
| 123 | auto k = arm_compute::support::cpp14::make_unique<NEGEMMTranspose1xWKernel>(); |
| 124 | k->configure(b, &_tmp_b); |
| 125 | _mtx_b_reshape_kernel = std::move(k); |
| 126 | } |
| 127 | |
| 128 | // Configure matrix multiply kernel |
| 129 | { |
| 130 | auto k = arm_compute::support::cpp14::make_unique<NEGEMMLowpMatrixMultiplyKernel>(); |
| 131 | k->configure(&_tmp_a, &_tmp_b, output); |
| 132 | _mm_kernel = std::move(k); |
| 133 | } |
Gian Marco | e75a02b | 2017-11-08 12:24:09 +0000 | [diff] [blame] | 134 | } |
Gian Marco Iodice | ab18212 | 2017-10-09 15:05:40 +0100 | [diff] [blame] | 135 | |
Gian Marco | e75a02b | 2017-11-08 12:24:09 +0000 | [diff] [blame] | 136 | // Initialize matrix B reduction kernel only if _a_offset is not equal to 0 |
| 137 | if(_a_offset != 0) |
| 138 | { |
| 139 | TensorShape shape_vector_sum_col = b->info()->tensor_shape(); |
| 140 | shape_vector_sum_col.remove_dimension(1); |
| 141 | TensorInfo info_vector_sum_col(shape_vector_sum_col, 1, DataType::S32); |
| 142 | _vector_sum_col.allocator()->init(info_vector_sum_col); |
| 143 | _memory_group.manage(&_vector_sum_col); |
| 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 | { |
| 152 | TensorShape shape_vector_sum_row = a->info()->tensor_shape(); |
| 153 | shape_vector_sum_row.set(Window::DimX, a->info()->dimension(1)); |
| 154 | shape_vector_sum_row.remove_dimension(1); |
| 155 | TensorInfo info_vector_sum_row(shape_vector_sum_row, 1, DataType::S32); |
| 156 | _vector_sum_row.allocator()->init(info_vector_sum_row); |
| 157 | _memory_group.manage(&_vector_sum_row); |
| 158 | |
| 159 | // Configure matrix A reduction kernel |
| 160 | _mtx_a_reduction_kernel.configure(a, &_vector_sum_row, a->info()->dimension(0), false); |
| 161 | } |
| 162 | |
| 163 | // Configure offset contribution kernel |
| 164 | _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); |
| 165 | |
| 166 | // Allocate tensors |
| 167 | if(!dot_product_path) |
| 168 | { |
Pablo Tello | 6ff12a0 | 2017-11-02 16:09:35 +0000 | [diff] [blame] | 169 | _tmp_a.allocator()->allocate(); |
| 170 | _tmp_b.allocator()->allocate(); |
| 171 | } |
Gian Marco | e75a02b | 2017-11-08 12:24:09 +0000 | [diff] [blame] | 172 | else |
| 173 | { |
| 174 | _workspace.allocator()->allocate(); |
| 175 | } |
| 176 | |
| 177 | if(_a_offset != 0) |
| 178 | { |
| 179 | _vector_sum_col.allocator()->allocate(); |
| 180 | } |
| 181 | |
| 182 | if(_b_offset != 0) |
| 183 | { |
| 184 | _vector_sum_row.allocator()->allocate(); |
| 185 | } |
Gian Marco Iodice | ab18212 | 2017-10-09 15:05:40 +0100 | [diff] [blame] | 186 | } |
| 187 | |
| 188 | void NEGEMMLowpMatrixMultiplyCore::run() |
| 189 | { |
| 190 | _memory_group.acquire(); |
| 191 | |
Pablo Tello | 6ff12a0 | 2017-11-02 16:09:35 +0000 | [diff] [blame] | 192 | if(_mtx_a_reshape_kernel) |
| 193 | { |
| 194 | NEScheduler::get().schedule(_mtx_a_reshape_kernel.get(), Window::DimY); |
| 195 | } |
Gian Marco Iodice | ab18212 | 2017-10-09 15:05:40 +0100 | [diff] [blame] | 196 | |
Pablo Tello | 6ff12a0 | 2017-11-02 16:09:35 +0000 | [diff] [blame] | 197 | if(_mtx_b_reshape_kernel) |
| 198 | { |
| 199 | NEScheduler::get().schedule(_mtx_b_reshape_kernel.get(), Window::DimY); |
| 200 | } |
Gian Marco Iodice | ab18212 | 2017-10-09 15:05:40 +0100 | [diff] [blame] | 201 | |
Gian Marco Iodice | ab18212 | 2017-10-09 15:05:40 +0100 | [diff] [blame] | 202 | NEScheduler::get().schedule(_mm_kernel.get(), Window::DimY); |
| 203 | |
Gian Marco | e75a02b | 2017-11-08 12:24:09 +0000 | [diff] [blame] | 204 | // Run matrix A reduction kernel only if _b_offset is not equal to 0 |
| 205 | if(_b_offset != 0) |
| 206 | { |
| 207 | NEScheduler::get().schedule(&_mtx_a_reduction_kernel, Window::DimX); |
| 208 | } |
| 209 | |
| 210 | // Run matrix B reduction kernel only if _a_offset is not equal to 0 |
| 211 | if(_a_offset != 0) |
| 212 | { |
| 213 | NEScheduler::get().schedule(&_mtx_b_reduction_kernel, Window::DimX); |
| 214 | } |
| 215 | |
| 216 | // Run offset contribution kernel |
| 217 | NEScheduler::get().schedule(&_offset_contribution_kernel, Window::DimY); |
| 218 | |
Gian Marco Iodice | ab18212 | 2017-10-09 15:05:40 +0100 | [diff] [blame] | 219 | _memory_group.release(); |
Pablo Tello | 6ff12a0 | 2017-11-02 16:09:35 +0000 | [diff] [blame] | 220 | } |