Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [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/CL/functions/CLGEMMLowpMatrixMultiplyCore.h" |
| 25 | |
| 26 | #include "arm_compute/core/CL/ICLTensor.h" |
| 27 | #include "arm_compute/core/Error.h" |
| 28 | #include "arm_compute/core/Helpers.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/CL/CLScheduler.h" |
| 33 | |
| 34 | using namespace arm_compute; |
| 35 | |
| 36 | CLGEMMLowpMatrixMultiplyCore::CLGEMMLowpMatrixMultiplyCore(std::shared_ptr<IMemoryManager> memory_manager) |
| 37 | : _memory_group(std::move(memory_manager)), _mm_kernel(), _mtx_a_reshape_kernel(), _mtx_b_reshape_kernel(), _mtx_a_reduction_kernel(), _mtx_b_reduction_kernel(), _offset_contribution_kernel(), |
| 38 | _vector_sum_col(), _vector_sum_row(), _tmp_a(), _tmp_b(), _a_offset(0), _b_offset(0), _is_interleaved_transposed(true) |
| 39 | { |
| 40 | } |
| 41 | |
| 42 | void CLGEMMLowpMatrixMultiplyCore::configure(const ICLTensor *a, const ICLTensor *b, ICLTensor *output) |
| 43 | { |
| 44 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(a, 1, DataType::QASYMM8); |
| 45 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::S32); |
| 46 | ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(a, b); |
| 47 | 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"); |
| 48 | 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"); |
| 49 | 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"); |
| 50 | |
| 51 | _a_offset = a->info()->quantization_info().offset; |
| 52 | _b_offset = b->info()->quantization_info().offset; |
| 53 | |
| 54 | // If the input tensor has less than 16 rows, we run a special version of GEMMLowp without reshaping the input tensors |
| 55 | _is_interleaved_transposed = a->info()->dimension(1) > 16; |
| 56 | |
| 57 | const ICLTensor *matrix_a = a; |
| 58 | const ICLTensor *matrix_b = b; |
| 59 | |
| 60 | if(_is_interleaved_transposed) |
| 61 | { |
| 62 | matrix_a = &_tmp_a; |
| 63 | matrix_b = &_tmp_b; |
| 64 | |
| 65 | // The interleaved output matrix will have the following shape: [ a_height * 4, ceil(a_width / 4.0f) ] |
| 66 | TensorShape shape_tmp_a = a->info()->tensor_shape(); |
| 67 | shape_tmp_a.set(0, a->info()->dimension(0) * 4); |
| 68 | shape_tmp_a.set(1, std::ceil(a->info()->dimension(1) / 4.f)); |
| 69 | |
| 70 | // The transpose1xW output matrix will have the following shape: [ b_height * 16, ceil(b_width / 16.0f) ] |
| 71 | TensorShape shape_tmp_b = b->info()->tensor_shape(); |
| 72 | shape_tmp_b.set(0, b->info()->dimension(1) * 16); |
| 73 | shape_tmp_b.set(1, std::ceil(b->info()->dimension(0) / 16.f)); |
| 74 | |
| 75 | TensorInfo info_a(shape_tmp_a, 1, a->info()->data_type()); |
| 76 | TensorInfo info_b(shape_tmp_b, 1, b->info()->data_type()); |
| 77 | _tmp_a.allocator()->init(info_a); |
| 78 | _tmp_b.allocator()->init(info_b); |
| 79 | _memory_group.manage(&_tmp_a); |
| 80 | _memory_group.manage(&_tmp_b); |
| 81 | |
| 82 | // Configure interleave kernel |
| 83 | _mtx_a_reshape_kernel.configure(a, &_tmp_a); |
| 84 | |
| 85 | // Configure transpose kernel |
| 86 | _mtx_b_reshape_kernel.configure(b, &_tmp_b); |
| 87 | } |
| 88 | |
| 89 | // Configure matrix multiply kernel |
| 90 | _mm_kernel.configure(matrix_a, matrix_b, output, _is_interleaved_transposed); |
| 91 | |
| 92 | // Initialize matrix B reduction kernel only if _a_offset is not equal to 0 |
| 93 | if(_a_offset != 0) |
| 94 | { |
| 95 | TensorShape shape_vector_sum_col = b->info()->tensor_shape(); |
| 96 | if(b->info()->num_dimensions() > 1) |
| 97 | { |
| 98 | shape_vector_sum_col.remove_dimension(1); |
| 99 | } |
| 100 | TensorInfo info_vector_sum_col(shape_vector_sum_col, 1, DataType::S32); |
| 101 | _vector_sum_col.allocator()->init(info_vector_sum_col); |
| 102 | _memory_group.manage(&_vector_sum_col); |
| 103 | |
| 104 | // Configure Matrix B reduction kernel |
| 105 | _mtx_b_reduction_kernel.configure(b, &_vector_sum_col); |
| 106 | } |
| 107 | |
| 108 | // Initialize Matrix A reduction kernel only if _b_offset is not equal to 0 |
| 109 | if(_b_offset != 0) |
| 110 | { |
| 111 | TensorShape shape_vector_sum_row = a->info()->tensor_shape(); |
| 112 | shape_vector_sum_row.set(Window::DimX, a->info()->dimension(1)); |
| 113 | if(a->info()->num_dimensions() > 1) |
| 114 | { |
| 115 | shape_vector_sum_row.remove_dimension(1); |
| 116 | } |
| 117 | TensorInfo info_vector_sum_row(shape_vector_sum_row, 1, DataType::S32); |
| 118 | _vector_sum_row.allocator()->init(info_vector_sum_row); |
| 119 | _memory_group.manage(&_vector_sum_row); |
| 120 | |
| 121 | // Configure matrix A reduction kernel |
| 122 | _mtx_a_reduction_kernel.configure(a, &_vector_sum_row); |
| 123 | } |
| 124 | |
| 125 | // Configure offset contribution kernel |
| 126 | _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); |
| 127 | |
| 128 | // Allocate tensors |
| 129 | if(_is_interleaved_transposed) |
| 130 | { |
| 131 | _tmp_a.allocator()->allocate(); |
| 132 | _tmp_b.allocator()->allocate(); |
| 133 | } |
| 134 | |
| 135 | if(_a_offset != 0) |
| 136 | { |
| 137 | _vector_sum_col.allocator()->allocate(); |
| 138 | } |
| 139 | |
| 140 | if(_b_offset != 0) |
| 141 | { |
| 142 | _vector_sum_row.allocator()->allocate(); |
| 143 | } |
| 144 | } |
| 145 | |
| 146 | void CLGEMMLowpMatrixMultiplyCore::run() |
| 147 | { |
| 148 | _memory_group.acquire(); |
| 149 | |
| 150 | if(_is_interleaved_transposed) |
| 151 | { |
| 152 | // Run reshape matrix A |
| 153 | CLScheduler::get().enqueue(_mtx_a_reshape_kernel, false); |
| 154 | |
| 155 | // Run reshape matrix B |
| 156 | CLScheduler::get().enqueue(_mtx_b_reshape_kernel, false); |
| 157 | } |
| 158 | |
| 159 | // Run matrix multiply |
| 160 | CLScheduler::get().enqueue(_mm_kernel, false); |
| 161 | |
| 162 | // Run matrix A reduction kernel only if _b_offset is not equal to 0 |
| 163 | if(_b_offset != 0) |
| 164 | { |
| 165 | CLScheduler::get().enqueue(_mtx_a_reduction_kernel, false); |
| 166 | } |
| 167 | |
| 168 | // Run matrix B reduction kernel only if _a_offset is not equal to 0 |
| 169 | if(_a_offset != 0) |
| 170 | { |
| 171 | CLScheduler::get().enqueue(_mtx_b_reduction_kernel, false); |
| 172 | } |
| 173 | |
| 174 | // Run offset contribution kernel |
| 175 | CLScheduler::get().enqueue(_offset_contribution_kernel, true); |
| 176 | |
| 177 | _memory_group.release(); |
| 178 | } |