Georgios Pinitas | 284cfe2 | 2018-02-13 12:15:13 +0000 | [diff] [blame] | 1 | /* |
Isabella Gottardi | 0a1090a | 2019-02-14 18:07:36 +0000 | [diff] [blame] | 2 | * Copyright (c) 2018-2019 ARM Limited. |
Georgios Pinitas | 284cfe2 | 2018-02-13 12:15:13 +0000 | [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/core/NEON/kernels/NEPermuteKernel.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/core/utils/misc/ShapeCalculator.h" |
| 33 | |
| 34 | namespace |
| 35 | { |
| 36 | #include "arm_compute/core/NEON/kernels/convolution/common/shims.hpp" |
| 37 | } // namespace |
| 38 | |
| 39 | #include <cstddef> |
| 40 | #include <cstdint> |
| 41 | |
| 42 | using namespace arm_compute; |
| 43 | |
| 44 | namespace |
| 45 | { |
Pablo Tello | 35767bc | 2018-12-05 17:36:30 +0000 | [diff] [blame] | 46 | inline bool is_permutation_supported(const PermutationVector &v) |
| 47 | { |
| 48 | static const std::array<PermutationVector, 6> permutations3 = |
| 49 | { |
Georgios Pinitas | d57891a | 2019-02-19 18:10:03 +0000 | [diff] [blame] | 50 | { |
| 51 | PermutationVector(2U, 0U, 1U), |
| 52 | PermutationVector(1U, 2U, 0U), |
| 53 | PermutationVector(0U, 1U, 2U), |
| 54 | PermutationVector(0U, 2U, 1U), |
| 55 | PermutationVector(1U, 0U, 2U), |
| 56 | PermutationVector(2U, 1U, 0U), |
| 57 | } |
Pablo Tello | 35767bc | 2018-12-05 17:36:30 +0000 | [diff] [blame] | 58 | }; |
| 59 | static const std::array<PermutationVector, 24> permutations4 = |
| 60 | { |
Georgios Pinitas | d57891a | 2019-02-19 18:10:03 +0000 | [diff] [blame] | 61 | { |
| 62 | PermutationVector(0U, 1U, 2U, 3U), |
| 63 | PermutationVector(1U, 0U, 2U, 3U), |
| 64 | PermutationVector(2U, 0U, 1U, 3U), |
| 65 | PermutationVector(0U, 2U, 1U, 3U), |
| 66 | PermutationVector(1U, 2U, 0U, 3U), |
| 67 | PermutationVector(2U, 1U, 0U, 3U), |
| 68 | PermutationVector(2U, 1U, 3U, 0U), |
| 69 | PermutationVector(1U, 2U, 3U, 0U), |
| 70 | PermutationVector(3U, 2U, 1U, 0U), |
| 71 | PermutationVector(2U, 3U, 1U, 0U), |
| 72 | PermutationVector(1U, 3U, 2U, 0U), |
| 73 | PermutationVector(3U, 1U, 2U, 0U), |
| 74 | PermutationVector(3U, 0U, 2U, 1U), |
| 75 | PermutationVector(0U, 3U, 2U, 1U), |
| 76 | PermutationVector(2U, 3U, 0U, 1U), |
| 77 | PermutationVector(3U, 2U, 0U, 1U), |
| 78 | PermutationVector(0U, 2U, 3U, 1U), |
| 79 | PermutationVector(2U, 0U, 3U, 1U), |
| 80 | PermutationVector(1U, 0U, 3U, 2U), |
| 81 | PermutationVector(0U, 1U, 3U, 2U), |
| 82 | PermutationVector(3U, 1U, 0U, 2U), |
| 83 | PermutationVector(1U, 3U, 0U, 2U), |
| 84 | PermutationVector(0U, 3U, 1U, 2U), |
| 85 | PermutationVector(3U, 0U, 1U, 2U) |
| 86 | } |
Pablo Tello | 35767bc | 2018-12-05 17:36:30 +0000 | [diff] [blame] | 87 | }; |
| 88 | |
| 89 | return (permutations3.end() != std::find(permutations3.begin(), permutations3.end(), v)) || (permutations4.end() != std::find(permutations4.begin(), permutations4.end(), v)); |
| 90 | } |
| 91 | |
Georgios Pinitas | 284cfe2 | 2018-02-13 12:15:13 +0000 | [diff] [blame] | 92 | Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const PermutationVector &perm) |
| 93 | { |
Pablo Tello | 35767bc | 2018-12-05 17:36:30 +0000 | [diff] [blame] | 94 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(!is_permutation_supported(perm), "PermutationVector not supported."); |
Georgios Pinitas | 284cfe2 | 2018-02-13 12:15:13 +0000 | [diff] [blame] | 95 | |
| 96 | const TensorShape output_shape = misc::shape_calculator::compute_permutation_output_shape(*input, perm); |
| 97 | |
| 98 | // Validate configured output |
| 99 | if(output->total_size() != 0) |
| 100 | { |
| 101 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), output_shape); |
Isabella Gottardi | 0a1090a | 2019-02-14 18:07:36 +0000 | [diff] [blame] | 102 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(input, output); |
Georgios Pinitas | 284cfe2 | 2018-02-13 12:15:13 +0000 | [diff] [blame] | 103 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); |
Georgios Pinitas | 284cfe2 | 2018-02-13 12:15:13 +0000 | [diff] [blame] | 104 | } |
| 105 | |
| 106 | return Status{}; |
| 107 | } |
| 108 | } // namespace |
| 109 | |
| 110 | template <typename T> |
| 111 | void NEPermuteKernel::run_permute(const Window &window) |
| 112 | { |
Pablo Tello | 35767bc | 2018-12-05 17:36:30 +0000 | [diff] [blame] | 113 | const DataLayout input_layout = _input->info()->data_layout(); |
| 114 | |
Georgios Pinitas | 284cfe2 | 2018-02-13 12:15:13 +0000 | [diff] [blame] | 115 | // Input window |
| 116 | Window window_in = window; |
Pablo Tello | 35767bc | 2018-12-05 17:36:30 +0000 | [diff] [blame] | 117 | |
| 118 | // we only support these two configs in arm_compute/core/NEON/kernels/convolution/common/shims.hpp, for all others |
| 119 | // we have to fall back to C++ |
| 120 | if((input_layout == DataLayout::NCHW && _perm == PermutationVector{ 2U, 0U, 1U }) || (input_layout == DataLayout::NHWC && _perm == PermutationVector{ 1U, 2U, 0U })) |
| 121 | { |
| 122 | window_in.set(Window::DimX, Window::Dimension(window.x().start(), window.x().end(), window.x().end() - window.x().start())); |
| 123 | window_in.set(Window::DimY, Window::Dimension(window.y().start(), window.y().end(), window.y().end() - window.y().start())); |
| 124 | window_in.set(Window::DimZ, Window::Dimension(window.z().start(), window.z().end(), window.z().end() - window.z().start())); |
| 125 | window_in.set(3, Window::Dimension(window[3].start(), window[3].end(), window[3].end() - window[3].start())); |
| 126 | } |
Georgios Pinitas | 284cfe2 | 2018-02-13 12:15:13 +0000 | [diff] [blame] | 127 | |
| 128 | // Output window |
| 129 | Window window_out(window); |
| 130 | const Window::Dimension zero_window = Window::Dimension(0, 0, 0); |
| 131 | for(size_t d = 0; d <= _perm.num_dimensions(); ++d) |
| 132 | { |
| 133 | window_out.set(d, zero_window); |
| 134 | } |
| 135 | |
| 136 | // Create iterators |
| 137 | Iterator in(_input, window_in); |
| 138 | Iterator out(_output, window_out); |
| 139 | |
Pablo Tello | 35767bc | 2018-12-05 17:36:30 +0000 | [diff] [blame] | 140 | int in_row_stride = 0; |
| 141 | int in_col_stride = 0; |
| 142 | int in_channel_stride = 0; |
| 143 | int in_batch_stride = 0; |
| 144 | int n_cols = 0; |
| 145 | int n_rows = 0; |
| 146 | int n_channels = 0; |
| 147 | int n_batches = 0; |
Georgios Pinitas | 284cfe2 | 2018-02-13 12:15:13 +0000 | [diff] [blame] | 148 | |
Pablo Tello | 35767bc | 2018-12-05 17:36:30 +0000 | [diff] [blame] | 149 | switch(input_layout) |
| 150 | { |
| 151 | case DataLayout::NCHW: |
| 152 | { |
| 153 | in_row_stride = _input->info()->strides_in_bytes().y() / sizeof(T); |
| 154 | in_channel_stride = _input->info()->strides_in_bytes().z() / sizeof(T); |
| 155 | in_batch_stride = _input->info()->strides_in_bytes()[3] / sizeof(T); |
| 156 | n_cols = _input->info()->tensor_shape().x(); |
| 157 | n_rows = window_in.y().step(); |
| 158 | n_channels = _input->info()->tensor_shape().z(); |
| 159 | n_batches = _input->info()->tensor_shape()[3]; |
| 160 | break; |
| 161 | } |
| 162 | case DataLayout::NHWC: |
| 163 | { |
| 164 | in_col_stride = _input->info()->strides_in_bytes().y() / sizeof(T); |
| 165 | in_row_stride = _input->info()->strides_in_bytes().z() / sizeof(T); |
| 166 | in_batch_stride = _input->info()->strides_in_bytes()[3] / sizeof(T); |
| 167 | n_channels = _input->info()->tensor_shape().x(); |
| 168 | n_cols = window_in.y().step(); |
| 169 | n_rows = _input->info()->tensor_shape().z(); |
| 170 | n_batches = _input->info()->tensor_shape()[3]; |
| 171 | break; |
| 172 | } |
| 173 | default: |
| 174 | { |
| 175 | ARM_COMPUTE_ERROR("Invalid input data layout."); |
| 176 | break; |
| 177 | } |
| 178 | } |
| 179 | |
| 180 | // CHW -> HWC |
| 181 | if(input_layout == DataLayout::NCHW && _perm == PermutationVector{ 2U, 0U, 1U }) |
| 182 | { |
Georgios Pinitas | 284cfe2 | 2018-02-13 12:15:13 +0000 | [diff] [blame] | 183 | const int out_channel_stride = _output->info()->strides_in_bytes().x() / sizeof(T); |
| 184 | const int out_col_stride = _output->info()->strides_in_bytes().y() / sizeof(T); |
| 185 | const int out_row_stride = _output->info()->strides_in_bytes().z() / sizeof(T); |
| 186 | const int out_batch_stride = _output->info()->strides_in_bytes()[3] / sizeof(T); |
Georgios Pinitas | 284cfe2 | 2018-02-13 12:15:13 +0000 | [diff] [blame] | 187 | execute_window_loop(window_in, [&](const Coordinates & id) |
| 188 | { |
| 189 | const int idx = id[0] * out_col_stride + id[1] * out_row_stride + id[2] * out_channel_stride; |
| 190 | reorder::nchw_to_nhwc(reinterpret_cast<const T *>(in.ptr()), reinterpret_cast<T *>(out.ptr()) + idx, |
| 191 | n_batches, n_channels, n_rows, n_cols, |
| 192 | in_batch_stride, in_channel_stride, in_row_stride, |
| 193 | out_batch_stride, out_row_stride, out_col_stride); |
| 194 | }, |
| 195 | in, out); |
| 196 | } |
| 197 | // HWC -> CHW |
Pablo Tello | 35767bc | 2018-12-05 17:36:30 +0000 | [diff] [blame] | 198 | else if(input_layout == DataLayout::NHWC && _perm == PermutationVector{ 1U, 2U, 0U }) |
Georgios Pinitas | 284cfe2 | 2018-02-13 12:15:13 +0000 | [diff] [blame] | 199 | { |
Georgios Pinitas | 284cfe2 | 2018-02-13 12:15:13 +0000 | [diff] [blame] | 200 | const int out_col_stride = _output->info()->strides_in_bytes().x() / sizeof(T); |
| 201 | const int out_row_stride = _output->info()->strides_in_bytes().y() / sizeof(T); |
| 202 | const int out_channel_stride = _output->info()->strides_in_bytes().z() / sizeof(T); |
| 203 | const int out_batch_stride = _output->info()->strides_in_bytes()[3] / sizeof(T); |
Georgios Pinitas | 284cfe2 | 2018-02-13 12:15:13 +0000 | [diff] [blame] | 204 | execute_window_loop(window_in, [&](const Coordinates & id) |
| 205 | { |
| 206 | const int idx = id[0] * out_channel_stride + id[1] * out_col_stride + id[2] * out_row_stride; |
| 207 | reorder::nhwc_to_nchw(reinterpret_cast<const T *>(in.ptr()), reinterpret_cast<T *>(out.ptr()) + idx, |
| 208 | n_batches, n_rows, n_cols, n_channels, |
| 209 | in_batch_stride, in_row_stride, in_col_stride, |
| 210 | out_batch_stride, out_channel_stride, out_row_stride); |
| 211 | }, |
| 212 | in, out); |
| 213 | } |
| 214 | else |
| 215 | { |
Pablo Tello | 35767bc | 2018-12-05 17:36:30 +0000 | [diff] [blame] | 216 | // All other cases fall back to C++ |
| 217 | // Permute strides |
| 218 | Strides strides = _output->info()->strides_in_bytes(); |
| 219 | Strides perm_strides = strides; |
| 220 | permute_strides(perm_strides, _perm); |
| 221 | const int perm_stride_3 = _input->info()->num_dimensions() >= 4 ? perm_strides[3] : 0; |
| 222 | execute_window_loop(window, [&](const Coordinates & id) |
| 223 | { |
| 224 | const int idx = id[0] * perm_strides[0] + id[1] * perm_strides[1] + id[2] * perm_strides[2] + id[3] * perm_stride_3; |
| 225 | *(reinterpret_cast<T *>(out.ptr() + idx)) = *(reinterpret_cast<const T *>(in.ptr())); |
| 226 | }, |
| 227 | in, out); |
Georgios Pinitas | 284cfe2 | 2018-02-13 12:15:13 +0000 | [diff] [blame] | 228 | } |
| 229 | } |
| 230 | |
| 231 | NEPermuteKernel::NEPermuteKernel() |
| 232 | : _func(), _input(nullptr), _output(nullptr), _perm() |
| 233 | { |
| 234 | } |
| 235 | |
| 236 | void NEPermuteKernel::configure(const ITensor *input, ITensor *output, const PermutationVector &perm) |
| 237 | { |
| 238 | ARM_COMPUTE_ERROR_ON_NULLPTR(input, output); |
| 239 | const TensorShape output_shape = misc::shape_calculator::compute_permutation_output_shape(*input->info(), perm); |
| 240 | // Output auto inizialitation if not yet initialized |
| 241 | auto_init_if_empty(*output->info(), input->info()->clone()->set_tensor_shape(output_shape)); |
| 242 | |
| 243 | // Perform validation step |
| 244 | ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), perm)); |
| 245 | |
| 246 | _input = input; |
| 247 | _output = output; |
| 248 | _perm = perm; |
| 249 | |
| 250 | switch(input->info()->element_size()) |
| 251 | { |
| 252 | case 1: |
| 253 | _func = &NEPermuteKernel::run_permute<uint8_t>; |
| 254 | break; |
| 255 | case 2: |
| 256 | _func = &NEPermuteKernel::run_permute<uint16_t>; |
| 257 | break; |
| 258 | case 4: |
| 259 | _func = &NEPermuteKernel::run_permute<uint32_t>; |
| 260 | break; |
| 261 | default: |
| 262 | ARM_COMPUTE_ERROR("Element size not supported"); |
| 263 | break; |
| 264 | } |
| 265 | |
| 266 | // Configure kernel window |
| 267 | Window win = calculate_max_window(*input->info(), Steps()); |
| 268 | |
| 269 | // The NEPermute doesn't need padding so update_window_and_padding() can be skipped |
| 270 | Coordinates coord; |
| 271 | coord.set_num_dimensions(output->info()->num_dimensions()); |
| 272 | output->info()->set_valid_region(ValidRegion(coord, output->info()->tensor_shape())); |
| 273 | |
| 274 | ICPPKernel::configure(win); |
| 275 | } |
| 276 | |
| 277 | Status NEPermuteKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const PermutationVector &perm) |
| 278 | { |
| 279 | ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, perm)); |
| 280 | return Status{}; |
| 281 | } |
| 282 | |
| 283 | void NEPermuteKernel::run(const Window &window, const ThreadInfo &info) |
| 284 | { |
| 285 | ARM_COMPUTE_UNUSED(info); |
| 286 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 287 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICPPKernel::window(), window); |
| 288 | |
| 289 | if(_func != nullptr) |
| 290 | { |
| 291 | (this->*_func)(window); |
| 292 | } |
| 293 | } |