Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 1 | /* |
Gian Marco | 1aede36 | 2018-02-05 15:07:36 +0000 | [diff] [blame^] | 2 | * Copyright (c) 2017-2018 ARM Limited. |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +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/GLES_COMPUTE/functions/GCGEMM.h" |
| 25 | |
| 26 | #include "arm_compute/core/Error.h" |
| 27 | #include "arm_compute/core/GLES_COMPUTE/GCHelpers.h" |
| 28 | #include "arm_compute/core/GLES_COMPUTE/IGCTensor.h" |
| 29 | #include "arm_compute/core/GLES_COMPUTE/kernels/GCGEMMInterleave4x4Kernel.h" |
| 30 | #include "arm_compute/core/GLES_COMPUTE/kernels/GCGEMMMatrixAdditionKernel.h" |
| 31 | #include "arm_compute/core/GLES_COMPUTE/kernels/GCGEMMMatrixMultiplyKernel.h" |
| 32 | #include "arm_compute/core/GLES_COMPUTE/kernels/GCGEMMTranspose1xWKernel.h" |
| 33 | #include "arm_compute/core/Helpers.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/GLES_COMPUTE/GCScheduler.h" |
| 38 | #include "arm_compute/runtime/ITensorAllocator.h" |
| 39 | |
| 40 | using namespace arm_compute; |
Joel Liang | 1c5ffd6 | 2017-12-28 10:09:51 +0800 | [diff] [blame] | 41 | using namespace arm_compute::gles_compute; |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 42 | |
| 43 | GCGEMM::GCGEMM() |
| 44 | : _interleave_kernel(), _transpose_kernel(), _mm_kernel(), _ma_kernel(), _tmp_a(), _tmp_b(), _is_interleaved_transposed(false), _run_addition(false) |
| 45 | { |
| 46 | } |
| 47 | |
Gian Marco | 1aede36 | 2018-02-05 15:07:36 +0000 | [diff] [blame^] | 48 | void GCGEMM::configure(const IGCTensor *a, const IGCTensor *b, const IGCTensor *c, IGCTensor *output, float alpha, float beta, const GEMMInfo &gemm_info) |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 49 | { |
| 50 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(a, 1, DataType::F32); |
| 51 | ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(a, b, output); |
Gian Marco | 1aede36 | 2018-02-05 15:07:36 +0000 | [diff] [blame^] | 52 | ARM_COMPUTE_ERROR_ON_MSG(gemm_info.is_a_reshaped(), "Matrix A already reshaped is not supported"); |
| 53 | ARM_COMPUTE_ERROR_ON_MSG(gemm_info.is_b_reshaped(), "Matrix B already reshaped is not supported"); |
| 54 | ARM_COMPUTE_ERROR_ON_MSG(gemm_info.reshape_b_only_on_first_run(), "Reshape matrix B only on first run is not supported"); |
| 55 | ARM_COMPUTE_UNUSED(gemm_info); |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 56 | |
| 57 | if(c != nullptr) |
| 58 | { |
| 59 | ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(a, c); |
| 60 | 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"); |
| 61 | 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 C"); |
| 62 | ARM_COMPUTE_ERROR_ON_MSG(c->info()->dimension(0) != output->info()->dimension(0), "The C matrix must have the same number of rows as the output matrix"); |
| 63 | ARM_COMPUTE_ERROR_ON_MSG(c->info()->dimension(1) != output->info()->dimension(1), "The C matrix must have the same number of columns as the output matrix"); |
| 64 | } |
| 65 | |
| 66 | 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"); |
| 67 | |
| 68 | // If the input tensor has less than 16 rows, we run a special version of GEMM without reshaping the input tensors |
| 69 | _is_interleaved_transposed = a->info()->dimension(1) > 16; |
| 70 | |
| 71 | const IGCTensor *matrix_a = a; |
| 72 | const IGCTensor *matrix_b = b; |
| 73 | |
| 74 | if(_is_interleaved_transposed) |
| 75 | { |
| 76 | matrix_a = &_tmp_a; |
| 77 | matrix_b = &_tmp_b; |
| 78 | |
| 79 | TensorShape shape_tmp_a = a->info()->tensor_shape(); |
| 80 | TensorShape shape_tmp_b = b->info()->tensor_shape(); |
| 81 | |
| 82 | shape_tmp_a.set(0, a->info()->dimension(0) * 4); |
| 83 | shape_tmp_a.set(1, std::ceil(a->info()->dimension(1) / 4.0f)); |
| 84 | |
| 85 | const unsigned int transpose_w = max_gc_vector_width / data_size_from_type(b->info()->data_type()); |
| 86 | shape_tmp_b.set(0, b->info()->dimension(1) * transpose_w); |
| 87 | shape_tmp_b.set(1, std::ceil(b->info()->dimension(0) / static_cast<float>(transpose_w))); |
| 88 | |
| 89 | TensorInfo info_a(shape_tmp_a, 1, a->info()->data_type(), a->info()->fixed_point_position()); |
| 90 | _tmp_a.allocator()->init(info_a); |
| 91 | |
| 92 | TensorInfo info_b(shape_tmp_b, 1, b->info()->data_type(), b->info()->fixed_point_position()); |
| 93 | _tmp_b.allocator()->init(info_b); |
| 94 | |
| 95 | // Configure interleave kernel |
| 96 | _interleave_kernel.configure(a, &_tmp_a); |
| 97 | |
| 98 | // Configure transpose kernel |
| 99 | _transpose_kernel.configure(b, &_tmp_b); |
| 100 | } |
| 101 | |
| 102 | _mm_kernel.configure(matrix_a, matrix_b, output, alpha, _is_interleaved_transposed); |
| 103 | |
| 104 | if(_is_interleaved_transposed) |
| 105 | { |
| 106 | // Allocate intermediate tensors |
| 107 | _tmp_a.allocator()->allocate(); |
| 108 | _tmp_b.allocator()->allocate(); |
| 109 | } |
| 110 | |
| 111 | // Configure matrix addition kernel |
| 112 | if(beta != 0 && c != nullptr) |
| 113 | { |
| 114 | _ma_kernel.configure(c, output, beta); |
| 115 | _run_addition = true; |
| 116 | } |
| 117 | } |
| 118 | |
| 119 | void GCGEMM::run() |
| 120 | { |
| 121 | if(_is_interleaved_transposed) |
| 122 | { |
| 123 | // Run interleave kernel |
Joel Liang | eb8f71e | 2017-12-27 13:16:00 +0800 | [diff] [blame] | 124 | GCScheduler::get().dispatch(_interleave_kernel, false); |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 125 | |
| 126 | // Run transpose kernel |
Joel Liang | eb8f71e | 2017-12-27 13:16:00 +0800 | [diff] [blame] | 127 | GCScheduler::get().dispatch(_transpose_kernel, false); |
| 128 | GCScheduler::get().memory_barrier(); |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 129 | } |
| 130 | |
| 131 | // Run matrix multiply kernel |
Joel Liang | eb8f71e | 2017-12-27 13:16:00 +0800 | [diff] [blame] | 132 | GCScheduler::get().dispatch(_mm_kernel, !_run_addition); |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 133 | |
| 134 | // Run matrix addition kernel |
| 135 | if(_run_addition) |
| 136 | { |
Joel Liang | eb8f71e | 2017-12-27 13:16:00 +0800 | [diff] [blame] | 137 | GCScheduler::get().memory_barrier(); |
| 138 | GCScheduler::get().dispatch(_ma_kernel); |
Anthony Barbier | 7068f99 | 2017-10-26 15:23:08 +0100 | [diff] [blame] | 139 | } |
| 140 | } |