1 /*
2 * Copyright (c) 2018-2020 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 "src/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 #include "src/core/helpers/AutoConfiguration.h"
34 #include "src/core/helpers/WindowHelpers.h"
35
36 namespace
37 {
38 #include "src/core/NEON/kernels/convolution/common/shims.hpp"
39 } // namespace
40
41 namespace arm_compute
42 {
43 namespace
44 {
is_permutation_supported(const PermutationVector & v)45 inline bool is_permutation_supported(const PermutationVector &v)
46 {
47 static const std::array<PermutationVector, 2> permutations2 =
48 {
49 {
50 PermutationVector(0U, 1U),
51 PermutationVector(1U, 0U),
52 }
53 };
54 static const std::array<PermutationVector, 6> permutations3 =
55 {
56 {
57 PermutationVector(2U, 0U, 1U),
58 PermutationVector(1U, 2U, 0U),
59 PermutationVector(0U, 1U, 2U),
60 PermutationVector(0U, 2U, 1U),
61 PermutationVector(1U, 0U, 2U),
62 PermutationVector(2U, 1U, 0U),
63 }
64 };
65 static const std::array<PermutationVector, 24> permutations4 =
66 {
67 {
68 PermutationVector(0U, 1U, 2U, 3U),
69 PermutationVector(1U, 0U, 2U, 3U),
70 PermutationVector(2U, 0U, 1U, 3U),
71 PermutationVector(0U, 2U, 1U, 3U),
72 PermutationVector(1U, 2U, 0U, 3U),
73 PermutationVector(2U, 1U, 0U, 3U),
74 PermutationVector(2U, 1U, 3U, 0U),
75 PermutationVector(1U, 2U, 3U, 0U),
76 PermutationVector(3U, 2U, 1U, 0U),
77 PermutationVector(2U, 3U, 1U, 0U),
78 PermutationVector(1U, 3U, 2U, 0U),
79 PermutationVector(3U, 1U, 2U, 0U),
80 PermutationVector(3U, 0U, 2U, 1U),
81 PermutationVector(0U, 3U, 2U, 1U),
82 PermutationVector(2U, 3U, 0U, 1U),
83 PermutationVector(3U, 2U, 0U, 1U),
84 PermutationVector(0U, 2U, 3U, 1U),
85 PermutationVector(2U, 0U, 3U, 1U),
86 PermutationVector(1U, 0U, 3U, 2U),
87 PermutationVector(0U, 1U, 3U, 2U),
88 PermutationVector(3U, 1U, 0U, 2U),
89 PermutationVector(1U, 3U, 0U, 2U),
90 PermutationVector(0U, 3U, 1U, 2U),
91 PermutationVector(3U, 0U, 1U, 2U)
92 }
93 };
94
95 return (permutations2.end() != std::find(permutations2.begin(), permutations2.end(), v)) || (permutations3.end() != std::find(permutations3.begin(), permutations3.end(), v))
96 || (permutations4.end() != std::find(permutations4.begin(), permutations4.end(), v));
97 }
98
validate_arguments(const ITensorInfo * input,const ITensorInfo * output,const PermutationVector & perm)99 Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const PermutationVector &perm)
100 {
101 ARM_COMPUTE_RETURN_ERROR_ON(input->data_type() == DataType::UNKNOWN);
102 ARM_COMPUTE_RETURN_ERROR_ON_MSG(!is_permutation_supported(perm), "PermutationVector not supported.");
103
104 const TensorShape output_shape = misc::shape_calculator::compute_permutation_output_shape(*input, perm);
105
106 // Validate configured output
107 if(output->total_size() != 0)
108 {
109 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), output_shape);
110 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(input, output);
111 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
112 }
113
114 return Status{};
115 }
116 } // namespace
117
118 template <typename T>
run_permute(const Window & window)119 void NEPermuteKernel::run_permute(const Window &window)
120 {
121 const DataLayout input_layout = _input->info()->data_layout();
122
123 // Input window
124 Window window_in = window;
125
126 // we only support these two configs in src/core/NEON/kernels/convolution/common/shims.hpp, for all others
127 // we have to fall back to C++
128 if((input_layout == DataLayout::NCHW && _perm == PermutationVector{ 2U, 0U, 1U }) || (input_layout == DataLayout::NHWC && _perm == PermutationVector{ 1U, 2U, 0U }))
129 {
130 window_in.set(Window::DimX, Window::Dimension(window.x().start(), window.x().end(), window.x().end() - window.x().start()));
131 window_in.set(Window::DimY, Window::Dimension(window.y().start(), window.y().end(), window.y().end() - window.y().start()));
132 window_in.set(Window::DimZ, Window::Dimension(window.z().start(), window.z().end(), window.z().end() - window.z().start()));
133 window_in.set(3, Window::Dimension(window[3].start(), window[3].end(), window[3].end() - window[3].start()));
134 }
135
136 // Output window
137 Window window_out(window);
138 const Window::Dimension zero_window = Window::Dimension(0, 0, 0);
139 for(size_t d = 0; d <= _output->info()->num_dimensions(); ++d)
140 {
141 window_out.set(d, zero_window);
142 }
143
144 // Create iterators
145 Iterator in(_input, window_in);
146 Iterator out(_output, window_out);
147
148 int in_row_stride = 0;
149 int in_col_stride = 0;
150 int in_channel_stride = 0;
151 int in_batch_stride = 0;
152 int n_cols = 0;
153 int n_rows = 0;
154 int n_channels = 0;
155 int n_batches = 0;
156
157 switch(input_layout)
158 {
159 case DataLayout::NCHW:
160 {
161 in_row_stride = _input->info()->strides_in_bytes().y() / sizeof(T);
162 in_channel_stride = _input->info()->strides_in_bytes().z() / sizeof(T);
163 in_batch_stride = _input->info()->strides_in_bytes()[3] / sizeof(T);
164 n_cols = _input->info()->tensor_shape().x();
165 n_rows = window_in.y().step();
166 n_channels = _input->info()->tensor_shape().z();
167 n_batches = _input->info()->tensor_shape()[3];
168 break;
169 }
170 case DataLayout::NHWC:
171 {
172 in_col_stride = _input->info()->strides_in_bytes().y() / sizeof(T);
173 in_row_stride = _input->info()->strides_in_bytes().z() / sizeof(T);
174 in_batch_stride = _input->info()->strides_in_bytes()[3] / sizeof(T);
175 n_channels = _input->info()->tensor_shape().x();
176 n_cols = window_in.y().step();
177 n_rows = _input->info()->tensor_shape().z();
178 n_batches = _input->info()->tensor_shape()[3];
179 break;
180 }
181 default:
182 {
183 ARM_COMPUTE_ERROR("Invalid input data layout.");
184 break;
185 }
186 }
187
188 // CHW -> HWC
189 if(input_layout == DataLayout::NCHW && _perm == PermutationVector{ 2U, 0U, 1U })
190 {
191 const int out_channel_stride = _output->info()->strides_in_bytes().x() / sizeof(T);
192 const int out_col_stride = _output->info()->strides_in_bytes().y() / sizeof(T);
193 const int out_row_stride = _output->info()->strides_in_bytes().z() / sizeof(T);
194 const int out_batch_stride = _output->info()->strides_in_bytes()[3] / sizeof(T);
195 execute_window_loop(window_in, [&](const Coordinates & id)
196 {
197 const int idx = id[0] * out_col_stride + id[1] * out_row_stride + id[2] * out_channel_stride;
198 reorder::nchw_to_nhwc(reinterpret_cast<const T *>(in.ptr()), reinterpret_cast<T *>(out.ptr()) + idx,
199 n_batches, n_channels, n_rows, n_cols,
200 in_batch_stride, in_channel_stride, in_row_stride,
201 out_batch_stride, out_row_stride, out_col_stride);
202 },
203 in, out);
204 }
205 // HWC -> CHW
206 else if(input_layout == DataLayout::NHWC && _perm == PermutationVector{ 1U, 2U, 0U })
207 {
208 const int out_col_stride = _output->info()->strides_in_bytes().x() / sizeof(T);
209 const int out_row_stride = _output->info()->strides_in_bytes().y() / sizeof(T);
210 const int out_channel_stride = _output->info()->strides_in_bytes().z() / sizeof(T);
211 const int out_batch_stride = _output->info()->strides_in_bytes()[3] / sizeof(T);
212 execute_window_loop(window_in, [&](const Coordinates & id)
213 {
214 const int idx = id[0] * out_channel_stride + id[1] * out_col_stride + id[2] * out_row_stride;
215 reorder::nhwc_to_nchw(reinterpret_cast<const T *>(in.ptr()), reinterpret_cast<T *>(out.ptr()) + idx,
216 n_batches, n_rows, n_cols, n_channels,
217 in_batch_stride, in_row_stride, in_col_stride,
218 out_batch_stride, out_channel_stride, out_row_stride);
219 },
220 in, out);
221 }
222 else
223 {
224 // All other cases fall back to C++
225 // Permute strides
226 Strides strides = _output->info()->strides_in_bytes();
227 Strides perm_strides = strides;
228 permute_strides(perm_strides, _perm);
229 const int perm_stride_3 = _input->info()->num_dimensions() >= 4 ? perm_strides[3] : 0;
230 execute_window_loop(window, [&](const Coordinates & id)
231 {
232 const int idx = id[0] * perm_strides[0] + id[1] * perm_strides[1] + id[2] * perm_strides[2] + id[3] * perm_stride_3;
233 *(reinterpret_cast<T *>(out.ptr() + idx)) = *(reinterpret_cast<const T *>(in.ptr()));
234 },
235 in, out);
236 }
237 }
238
NEPermuteKernel()239 NEPermuteKernel::NEPermuteKernel()
240 : _func(), _input(nullptr), _output(nullptr), _perm()
241 {
242 }
243
configure(const ITensor * input,ITensor * output,const PermutationVector & perm)244 void NEPermuteKernel::configure(const ITensor *input, ITensor *output, const PermutationVector &perm)
245 {
246 ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
247 const TensorShape output_shape = misc::shape_calculator::compute_permutation_output_shape(*input->info(), perm);
248 // Output auto inizialitation if not yet initialized
249 auto_init_if_empty(*output->info(), input->info()->clone()->set_tensor_shape(output_shape));
250
251 // Perform validation step
252 ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), perm));
253
254 _input = input;
255 _output = output;
256 _perm = perm;
257
258 switch(input->info()->element_size())
259 {
260 case 1:
261 _func = &NEPermuteKernel::run_permute<uint8_t>;
262 break;
263 case 2:
264 _func = &NEPermuteKernel::run_permute<uint16_t>;
265 break;
266 case 4:
267 _func = &NEPermuteKernel::run_permute<uint32_t>;
268 break;
269 default:
270 ARM_COMPUTE_ERROR("Element size not supported");
271 break;
272 }
273
274 // Configure kernel window
275 Window win = calculate_max_window(*input->info(), Steps());
276
277 // The NEPermute doesn't need padding so update_window_and_padding() can be skipped
278 Coordinates coord;
279 coord.set_num_dimensions(output->info()->num_dimensions());
280 output->info()->set_valid_region(ValidRegion(coord, output->info()->tensor_shape()));
281
282 ICPPKernel::configure(win);
283 }
284
validate(const ITensorInfo * input,const ITensorInfo * output,const PermutationVector & perm)285 Status NEPermuteKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const PermutationVector &perm)
286 {
287 ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, perm));
288 return Status{};
289 }
290
run(const Window & window,const ThreadInfo & info)291 void NEPermuteKernel::run(const Window &window, const ThreadInfo &info)
292 {
293 ARM_COMPUTE_UNUSED(info);
294 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
295 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICPPKernel::window(), window);
296
297 if(_func != nullptr)
298 {
299 (this->*_func)(window);
300 }
301 }
302 } // namespace arm_compute
303