Anthony Barbier | 3d677cc | 2018-07-23 16:42:59 +0100 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2018 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 | |
| 25 | #include "arm_compute/core/NEON/kernels/assembly/NEGEMMInterleavedMatrixMultiplyWrapper.h" |
| 26 | |
| 27 | #include "NEGEMMInterleavedStrategies.h" |
| 28 | #include "arm_compute/core/Helpers.h" |
| 29 | #include "arm_compute/core/ITensor.h" |
| 30 | #include "arm_compute/core/Utils.h" |
| 31 | #include "arm_compute/core/Validate.h" |
| 32 | #include "arm_compute/core/WindowIterator.h" |
| 33 | |
| 34 | namespace arm_compute |
| 35 | { |
| 36 | template <typename To, typename Tr, bool use_dot> |
| 37 | void NEGEMMInterleavedMatrixMultiplyWrapperTemplate<To, Tr, use_dot>::configure(const ITensor *prepared_a, const ITensor *transformed_b, ITensor *tmp_c, ITensor *c, const Window &block_walker, |
| 38 | const BlockSizes &block_sizes, const INEGEMMWrapperKernel::Params ¶ms, bool b_is_pretransposed, float alpha, float beta, unsigned int max_num_threads) |
| 39 | { |
Georgios Pinitas | ecae3a1 | 2018-10-09 15:13:12 +0100 | [diff] [blame] | 40 | using strategy = typename Kernel<To, use_dot>::strategy; |
Anthony Barbier | 3d677cc | 2018-07-23 16:42:59 +0100 | [diff] [blame] | 41 | |
| 42 | _prepared_a = prepared_a; |
| 43 | _transformed_b = transformed_b; |
| 44 | _tmp_c = tmp_c; |
| 45 | _c = c; |
| 46 | _block_walker = block_walker; |
| 47 | _block_sizes = block_sizes; |
| 48 | _params = params; |
| 49 | _b_is_pretransposed = b_is_pretransposed; |
| 50 | _alpha = alpha; |
| 51 | _beta = beta; |
| 52 | |
| 53 | auto_init_if_empty(*_tmp_c->info(), c->info()->clone()->set_tensor_shape(TensorShape{ _block_sizes.x_block * strategy::out_height(), max_num_threads })); |
| 54 | } |
| 55 | |
| 56 | template <typename To, typename Tr, bool use_dot> |
| 57 | void NEGEMMInterleavedMatrixMultiplyWrapperTemplate<To, Tr, use_dot>::transform(const MatrixMultiplyWorkload &wl, const ThreadInfo &info, const Window &batch_window, const Coordinates &start_offset, |
| 58 | const Coordinates &end_offset) |
| 59 | { |
Georgios Pinitas | ecae3a1 | 2018-10-09 15:13:12 +0100 | [diff] [blame] | 60 | using strategy = typename Kernel<To, use_dot>::strategy; |
Anthony Barbier | 3d677cc | 2018-07-23 16:42:59 +0100 | [diff] [blame] | 61 | |
| 62 | strategy strat(info.cpu_info); |
| 63 | TensorAccessor<To> prepared_a(*_prepared_a); |
| 64 | TensorAccessor<To> transformed_b(*_transformed_b); |
| 65 | TensorAccessor<Tr> c(*_c); |
| 66 | TensorAccessor<Tr> tmp_c(*_tmp_c); |
| 67 | |
| 68 | int prev_batch = -1; |
| 69 | To *a_ptr = nullptr; |
| 70 | auto window_iterator = arm_compute::create_window_iterator(batch_window, start_offset, end_offset, [&](const Coordinates & id) |
| 71 | { |
| 72 | const unsigned int y = id.x(); |
| 73 | const unsigned int batch = id.y(); |
| 74 | const unsigned int ymax = std::min(_params.M, y + strategy::out_height()); |
| 75 | |
| 76 | // If it's the first block of a new batch then reset the pointer to A. |
| 77 | if(prev_batch != static_cast<int>(batch)) |
| 78 | { |
| 79 | const unsigned int first_m = id.x(); |
| 80 | a_ptr = prepared_a(0, first_m, batch); |
| 81 | prev_batch = batch; |
| 82 | } |
| 83 | |
| 84 | // Call matrix multiply assembly routine to process the block: |
| 85 | strat.kernel(a_ptr, transformed_b(wl._offset_transformed_b), tmp_c(0, info.thread_id), 1, wl._bblocks, wl._kern_k); |
| 86 | a_ptr += strategy::out_height() * wl._kern_k; |
| 87 | |
| 88 | // Merge the result with the other blocks' results: |
| 89 | strat.transforms.Merge(c(0, 0, batch, wl._multi), tmp_c(0, info.thread_id), c.stride(1), y, ymax, wl._x0, wl._xmax, _alpha, (wl._k0 == 0 ? _beta : static_cast<Tr>(1))); |
| 90 | }); |
| 91 | auto on_new_row_size = [&](unsigned int start, unsigned int end) |
| 92 | { |
| 93 | //Nothing to do |
| 94 | }; |
| 95 | window_iterator.iterate_2D(on_new_row_size); |
| 96 | } |
| 97 | |
| 98 | template <typename To, typename Tr, bool use_dot> |
| 99 | void NEGEMMInterleavedMatrixMultiplyWrapperTemplate<To, Tr, use_dot>::create_workloads(std::vector<MatrixMultiplyWorkload> &workloads) |
| 100 | { |
Georgios Pinitas | ecae3a1 | 2018-10-09 15:13:12 +0100 | [diff] [blame] | 101 | using strategy = typename Kernel<To, use_dot>::strategy; |
Anthony Barbier | 3d677cc | 2018-07-23 16:42:59 +0100 | [diff] [blame] | 102 | |
| 103 | unsigned int offset_transformed_b = 0; |
| 104 | execute_window_loop(_block_walker, [&](const Coordinates & id) |
| 105 | { |
| 106 | const unsigned int x0 = id.x(); |
| 107 | const unsigned int k0 = id.y(); |
| 108 | const unsigned int multi = id.z(); |
| 109 | |
| 110 | const unsigned int xmax = std::min(x0 + _block_walker.x().step(), _params.N); |
| 111 | const unsigned int kmax = std::min(k0 + _block_walker.y().step(), _params.K); |
| 112 | |
| 113 | // Figure out how many "K" the kernel will actually process. |
| 114 | const int kern_k = ceil_to_multiple(kmax - k0, strategy::k_unroll()); |
| 115 | const int bblocks = DIV_CEIL(xmax - x0, strategy::out_width()); |
| 116 | |
| 117 | workloads.push_back(MatrixMultiplyWorkload(offset_transformed_b, x0, xmax, k0, kmax, multi, kern_k, bblocks)); |
| 118 | |
| 119 | if(_b_is_pretransposed) |
| 120 | { |
| 121 | offset_transformed_b += bblocks * strategy::out_width() * kern_k; |
| 122 | } |
| 123 | else |
| 124 | { |
| 125 | ARM_COMPUTE_ERROR("Not supported"); |
| 126 | } |
| 127 | }); |
| 128 | } |
| 129 | |
| 130 | //TODO: regroup somewhere ? |
| 131 | template class NEGEMMInterleavedMatrixMultiplyWrapperTemplate<float, float>; |
| 132 | #ifdef __aarch64__ |
| 133 | template class NEGEMMInterleavedMatrixMultiplyWrapperTemplate<uint8_t, uint32_t>; |
| 134 | template class NEGEMMInterleavedMatrixMultiplyWrapperTemplate<int8_t, int32_t>; |
| 135 | template class NEGEMMInterleavedMatrixMultiplyWrapperTemplate<uint8_t, uint32_t, true>; |
| 136 | template class NEGEMMInterleavedMatrixMultiplyWrapperTemplate<int8_t, int32_t, true>; |
| 137 | #endif /* __aarch64__ */ |
| 138 | |
| 139 | #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC |
| 140 | template class NEGEMMInterleavedMatrixMultiplyWrapperTemplate<float16_t, float16_t>; |
| 141 | #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ |
| 142 | } // namespace arm_compute |