Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +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/NEGEMM.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/TensorInfo.h" |
| 30 | #include "arm_compute/core/Types.h" |
| 31 | #include "arm_compute/core/Validate.h" |
| 32 | #include "arm_compute/runtime/NEON/NEScheduler.h" |
| 33 | #include "arm_compute/runtime/TensorAllocator.h" |
| 34 | |
| 35 | #include <cmath> |
| 36 | |
| 37 | using namespace arm_compute; |
| 38 | |
| 39 | NEGEMM::NEGEMM() |
| 40 | : _interleave_kernel(), _transpose_kernel(), _mm_kernel(), _ma_kernel(), _tmp_a(), _tmp_b(), _run_vector_matrix_multiplication(false), _run_addition(false) |
| 41 | { |
| 42 | } |
| 43 | |
| 44 | void NEGEMM::configure(const ITensor *a, const ITensor *b, const ITensor *c, ITensor *d, float alpha, float beta) |
| 45 | { |
| 46 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(a, 1, DataType::F32, DataType::F16, DataType::QS8); |
| 47 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(b, 1, DataType::F32, DataType::F16, DataType::QS8); |
| 48 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(d, 1, DataType::F32, DataType::F16, DataType::QS8); |
| 49 | |
| 50 | if(c != nullptr) |
| 51 | { |
| 52 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(c, 1, DataType::F32, DataType::F16, DataType::QS8); |
| 53 | ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(a, c); |
| 54 | ARM_COMPUTE_ERROR_ON_MSG(a->info()->dimension(1) != c->info()->dimension(1), "The C matrix must have the same number of rows as the matrix A"); |
| 55 | ARM_COMPUTE_ERROR_ON_MSG(b->info()->dimension(0) != c->info()->dimension(0), "The C matrix must have the same number of columns as the matrix B"); |
| 56 | ARM_COMPUTE_ERROR_ON_MSG(c->info()->dimension(0) != d->info()->dimension(0), "The C matrix must have the same number of rows as the output matrix"); |
| 57 | ARM_COMPUTE_ERROR_ON_MSG(c->info()->dimension(1) != d->info()->dimension(1), "The C matrix must have the same number of columns as the output matrix"); |
| 58 | } |
| 59 | |
| 60 | ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(a, b, d); |
| 61 | 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"); |
| 62 | |
| 63 | // Check if the first input tensor is a vector. If so, all the kernels for reshaping the tensors can be skipped |
| 64 | if((a->info()->dimension(1) == 1)) |
| 65 | { |
| 66 | _run_vector_matrix_multiplication = true; |
| 67 | |
| 68 | // Configure the matrix multiply kernel |
| 69 | _mm_kernel.configure(a, b, d, alpha); |
| 70 | } |
| 71 | else |
| 72 | { |
| 73 | _run_vector_matrix_multiplication = false; |
| 74 | |
| 75 | TensorShape shape_tmp_a = a->info()->tensor_shape(); |
| 76 | TensorShape shape_tmp_b = b->info()->tensor_shape(); |
| 77 | |
| 78 | shape_tmp_a.set(0, a->info()->dimension(0) * 4); |
| 79 | shape_tmp_a.set(1, std::ceil(a->info()->dimension(1) / 4.0f)); |
| 80 | |
| 81 | switch(a->info()->data_type()) |
| 82 | { |
| 83 | case DataType::F32: |
| 84 | { |
| 85 | shape_tmp_b.set(0, b->info()->dimension(1) * 4); |
| 86 | shape_tmp_b.set(1, std::ceil(b->info()->dimension(0) / 4.0f)); |
| 87 | break; |
| 88 | } |
| 89 | case DataType::F16: |
| 90 | #ifdef ARM_COMPUTE_ENABLE_FP16 |
| 91 | { |
| 92 | shape_tmp_b.set(0, b->info()->dimension(1) * 8); |
| 93 | shape_tmp_b.set(1, std::ceil(b->info()->dimension(0) / 8.0f)); |
| 94 | break; |
| 95 | } |
| 96 | #endif |
| 97 | case DataType::QS8: |
| 98 | { |
| 99 | shape_tmp_b.set(0, b->info()->dimension(1) * 16); |
| 100 | shape_tmp_b.set(1, std::ceil(b->info()->dimension(0) / 16.0f)); |
| 101 | break; |
| 102 | } |
| 103 | default: |
| 104 | { |
| 105 | ARM_COMPUTE_ERROR_ON("Data type not supported"); |
| 106 | } |
| 107 | } |
| 108 | |
| 109 | TensorInfo info_a(shape_tmp_a, 1, a->info()->data_type(), a->info()->fixed_point_position()); |
| 110 | TensorInfo info_b(shape_tmp_b, 1, b->info()->data_type(), a->info()->fixed_point_position()); |
| 111 | |
| 112 | _tmp_a.allocator()->init(info_a); |
| 113 | _tmp_b.allocator()->init(info_b); |
| 114 | |
| 115 | // Configure interleave kernel |
| 116 | _interleave_kernel.configure(a, &_tmp_a); |
| 117 | |
| 118 | // Configure transpose kernel |
| 119 | _transpose_kernel.configure(b, &_tmp_b); |
| 120 | |
| 121 | // Configure matrix multiplication kernel |
| 122 | _mm_kernel.configure(&_tmp_a, &_tmp_b, d, alpha); |
| 123 | |
| 124 | // Allocate once the all configure methods have been called |
| 125 | _tmp_a.allocator()->allocate(); |
| 126 | _tmp_b.allocator()->allocate(); |
| 127 | } |
| 128 | |
| 129 | // Configure matrix addition kernel |
| 130 | if(beta != 0 && c != nullptr) |
| 131 | { |
| 132 | _ma_kernel.configure(c, d, beta); |
| 133 | _run_addition = true; |
| 134 | } |
| 135 | } |
| 136 | |
| 137 | void NEGEMM::run() |
| 138 | { |
| 139 | if(!_run_vector_matrix_multiplication) |
| 140 | { |
| 141 | // Run interleave kernel |
| 142 | NEScheduler::get().schedule(&_interleave_kernel, Window::DimY); |
| 143 | |
| 144 | // Run transpose kernel |
| 145 | NEScheduler::get().schedule(&_transpose_kernel, Window::DimY); |
| 146 | } |
| 147 | |
| 148 | // Run matrix multiply kernel |
| 149 | NEScheduler::get().schedule(&_mm_kernel, _run_vector_matrix_multiplication ? Window::DimX : Window::DimY); |
| 150 | |
| 151 | // Run matrix addition kernel |
| 152 | if(_run_addition) |
| 153 | { |
| 154 | NEScheduler::get().schedule(&_ma_kernel, Window::DimY); |
| 155 | } |
| 156 | } |