Gitiles
Code Review Sign In
review.mlplatform.org / ml / ComputeLibrary / 6f58b37a18cfade5dbec38638926f7bd368756d9 / . / src / core / NEON / kernels / NEUpsampleLayerKernel.cpp
blob: d3d88b3bf00eab91acf44430c4610a3408e1d102 [file] [log] [blame]
Michalis Spyrou7c9541c2018-09-20 17:40:04 +0100[diff] [blame]1/*
Isabella Gottardi0a1090a2019-02-14 18:07:36 +0000[diff] [blame]2 * Copyright (c) 2018-2019 ARM Limited.
Michalis Spyrou7c9541c2018-09-20 17:40:04 +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/core/NEON/kernels/NEUpsampleLayerKernel.h"
25
Georgios Pinitas8f5802f2019-02-22 11:08:32 +0000[diff] [blame]26#include "arm_compute/core/CPP/Validate.h"
Michalis Spyrou7c9541c2018-09-20 17:40:04 +0100[diff] [blame]27#include "arm_compute/core/Error.h"
28#include "arm_compute/core/Helpers.h"
29#include "arm_compute/core/ITensor.h"
30#include "arm_compute/core/TensorInfo.h"
31#include "arm_compute/core/Validate.h"
32#include "arm_compute/core/Window.h"
33#include "arm_compute/core/utils/misc/ShapeCalculator.h"
34
35#include <arm_neon.h>
36
37namespace arm_compute
38{
39namespace
40{
41std::pair<Status, Window> validate_and_configure_window_nchw(ITensorInfo *input, ITensorInfo *output, int num_elems_processed_per_iteration_x, const Size2D &info)
42{
43 const int num_elems_processed_per_iteration_x_out = num_elems_processed_per_iteration_x * info.x();
44 Window win = calculate_max_window(*output, Steps(num_elems_processed_per_iteration_x_out));
45 AccessWindowRectangle input_access(input, 0, 0, num_elems_processed_per_iteration_x, 1, 0.5f, 0.5f);
46 AccessWindowHorizontal output_access(output, 0, num_elems_processed_per_iteration_x_out);
47 bool window_changed = update_window_and_padding(win, input_access, output_access);
48 output_access.set_valid_region(win, output->valid_region());
49
50 Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
51 return std::make_pair(err, win);
52}
53
54std::pair<Status, Window> validate_and_configure_window_nhwc(ITensorInfo *input, ITensorInfo *output, int num_elems_processed_per_iteration_x, const Size2D &info)
55{
56 ARM_COMPUTE_UNUSED(info);
57 Window win = calculate_max_window(*output, Steps(num_elems_processed_per_iteration_x));
58 AccessWindowHorizontal input_access(input, 0, num_elems_processed_per_iteration_x);
59 AccessWindowHorizontal output_access(output, 0, num_elems_processed_per_iteration_x);
60 bool window_changed = update_window_and_padding(win, input_access, output_access);
61 output_access.set_valid_region(win, output->valid_region());
62
63 Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
64 return std::make_pair(err, win);
65}
66
67std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output, int num_elems_processed_per_iteration_x, const Size2D &info)
68{
69 std::pair<Status, Window> win_config;
70 switch(input->data_layout())
71 {
72 case DataLayout::NCHW:
73 win_config = validate_and_configure_window_nchw(input, output, num_elems_processed_per_iteration_x, info);
74 break;
75 case DataLayout::NHWC:
76 win_config = validate_and_configure_window_nhwc(input, output, num_elems_processed_per_iteration_x, info);
77 break;
78 default:
79 win_config = std::make_pair(ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Unsupported data layout!"), Window{});
80 }
81
82 return win_config;
83}
84} // namespace
85NEUpsampleLayerKernel::NEUpsampleLayerKernel()
86 : _func(nullptr), _input(nullptr), _output(nullptr), _info(), _num_elems_processed_per_iteration_x()
87{
88}
89
90Status NEUpsampleLayerKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const Size2D &info, const InterpolationPolicy policy)
91{
92 ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
93 ARM_COMPUTE_UNUSED(policy);
94
95 const DataLayout data_layout = input->data_layout();
96 const int idx_width = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH);
97 const int idx_height = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT);
98
Georgios Pinitas8f5802f2019-02-22 11:08:32 +0000[diff] [blame]99 ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(input);
Michalis Spyrou7c9541c2018-09-20 17:40:04 +0100[diff] [blame]100 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F16, DataType::F32);
101 ARM_COMPUTE_RETURN_ERROR_ON_MSG(info.x() != 2 || info.y() != 2, "Only stride 2 is supported");
102 ARM_COMPUTE_RETURN_ERROR_ON_MSG(policy != InterpolationPolicy::NEAREST_NEIGHBOR, "Only nearest neighbor policy supported");
103
104 // Check output if configured
105 if(output->total_size() != 0)
106 {
107 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
108 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(input, output);
109 ARM_COMPUTE_RETURN_ERROR_ON(output->dimension(idx_width) != info.x() * input->dimension(idx_width));
110 ARM_COMPUTE_RETURN_ERROR_ON(output->dimension(idx_height) != info.y() * input->dimension(idx_height));
Isabella Gottardi0a1090a2019-02-14 18:07:36 +0000[diff] [blame]111 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(input, output);
Michalis Spyrou7c9541c2018-09-20 17:40:04 +0100[diff] [blame]112 }
113
114 const int num_elems_processed_per_iteration_x = 16 / input->element_size();
115 ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(),
116 output->clone().get(), num_elems_processed_per_iteration_x, info)
117 .first);
118 return Status{};
119}
120
121void NEUpsampleLayerKernel::upsample_f32_nchw(const arm_compute::Window &window)
122{
123 Window window_in(window);
124 window_in.set(Window::DimX, Window::Dimension(0, _input->info()->dimension(0), _num_elems_processed_per_iteration_x));
125
126 Window window_out(window);
127 window_out.set(Window::DimY, Window::Dimension(0, _output->info()->dimension(1), _info.y()));
128
129 Iterator input(_input, window_in);
130 Iterator output(_output, window_out);
131 const int offset_y_out = _output->info()->strides_in_bytes().y() / sizeof(float);
132
Michalis Spyroua4f378d2019-04-26 14:54:54 +0100[diff] [blame]133 execute_window_loop(window_out, [&](const Coordinates &)
Michalis Spyrou7c9541c2018-09-20 17:40:04 +0100[diff] [blame]134 {
135 const float32x4_t data = vld1q_f32(reinterpret_cast<const float *>(input.ptr()));
136 const float32x4_t data_out1 = { vgetq_lane_f32(data, 0), vgetq_lane_f32(data, 0), vgetq_lane_f32(data, 1), vgetq_lane_f32(data, 1) };
137 const float32x4_t data_out2 = { vgetq_lane_f32(data, 2), vgetq_lane_f32(data, 2), vgetq_lane_f32(data, 3), vgetq_lane_f32(data, 3) };
138 auto out = reinterpret_cast<float *>(output.ptr());
139
140 vst1q_f32(out, data_out1);
141 vst1q_f32(out + 4, data_out2);
142 vst1q_f32(out + offset_y_out, data_out1);
143 vst1q_f32(out + offset_y_out + 4, data_out2);
144 },
145 input, output);
146}
147
148void NEUpsampleLayerKernel::upsample_f32_nhwc(const arm_compute::Window &window)
149{
150 Window window_out(window);
151 window_out.set(Window::DimY, Window::Dimension(0, _output->info()->dimension(1), _info.x()));
152 window_out.set(Window::DimZ, Window::Dimension(0, _output->info()->dimension(2), _info.y()));
153
154 Iterator input(_input, window);
155 Iterator output(_output, window_out);
156
157 const int offset_y_out = _output->info()->strides_in_bytes().y() / sizeof(float);
158 const int offset_z_out = _output->info()->strides_in_bytes().z() / sizeof(float);
159
Michalis Spyroua4f378d2019-04-26 14:54:54 +0100[diff] [blame]160 execute_window_loop(window_out, [&](const Coordinates &)
Michalis Spyrou7c9541c2018-09-20 17:40:04 +0100[diff] [blame]161 {
162 const float32x4_t data = vld1q_f32(reinterpret_cast<const float *>(input.ptr()));
163 auto out = reinterpret_cast<float *>(output.ptr());
164
165 vst1q_f32(out, data);
166 vst1q_f32(out + offset_y_out, data);
167 vst1q_f32(out + offset_z_out, data);
168 vst1q_f32(out + offset_y_out + offset_z_out, data);
169 },
170 input, output);
171}
172
173void NEUpsampleLayerKernel::upsample_qasymm8_nchw(const arm_compute::Window &window)
174{
175 Window window_in(window);
176 window_in.set(Window::DimX, Window::Dimension(0, _input->info()->dimension(0), _num_elems_processed_per_iteration_x));
177
178 Window window_out(window);
179 window_out.set(Window::DimY, Window::Dimension(0, _output->info()->dimension(1), _info.y()));
180
181 Iterator input(_input, window_in);
182 Iterator output(_output, window_out);
183 const int offset_y_out = _output->info()->strides_in_bytes().y() / sizeof(uint8_t);
184
Michalis Spyroua4f378d2019-04-26 14:54:54 +0100[diff] [blame]185 execute_window_loop(window_out, [&](const Coordinates &)
Michalis Spyrou7c9541c2018-09-20 17:40:04 +0100[diff] [blame]186 {
187 const uint8x16_t data = vld1q_u8(reinterpret_cast<const uint8_t *>(input.ptr()));
188 const uint8x16_t data_out1 = { vgetq_lane_u8(data, 0), vgetq_lane_u8(data, 0), vgetq_lane_u8(data, 1), vgetq_lane_u8(data, 1),
189 vgetq_lane_u8(data, 2), vgetq_lane_u8(data, 2), vgetq_lane_u8(data, 3), vgetq_lane_u8(data, 3),
190 vgetq_lane_u8(data, 4), vgetq_lane_u8(data, 4), vgetq_lane_u8(data, 5), vgetq_lane_u8(data, 5),
191 vgetq_lane_u8(data, 6), vgetq_lane_u8(data, 6), vgetq_lane_u8(data, 7), vgetq_lane_u8(data, 7)
192 };
193 const uint8x16_t data_out2 =
194 {
195 vgetq_lane_u8(data, 8), vgetq_lane_u8(data, 8), vgetq_lane_u8(data, 9), vgetq_lane_u8(data, 9),
196 vgetq_lane_u8(data, 10), vgetq_lane_u8(data, 10), vgetq_lane_u8(data, 11), vgetq_lane_u8(data, 11),
197 vgetq_lane_u8(data, 12), vgetq_lane_u8(data, 12), vgetq_lane_u8(data, 13), vgetq_lane_u8(data, 13),
198 vgetq_lane_u8(data, 14), vgetq_lane_u8(data, 14), vgetq_lane_u8(data, 15), vgetq_lane_u8(data, 15)
199 };
200 auto out = reinterpret_cast<uint8_t *>(output.ptr());
201
202 vst1q_u8(out, data_out1);
203 vst1q_u8(out + 16, data_out2);
204 vst1q_u8(out + offset_y_out, data_out1);
205 vst1q_u8(out + offset_y_out + 16, data_out2);
206 },
207 input, output);
208}
209
210void NEUpsampleLayerKernel::upsample_qasymm8_nhwc(const arm_compute::Window &window)
211{
212 Window window_out(window);
213 window_out.set(Window::DimY, Window::Dimension(0, _output->info()->dimension(1), _info.x()));
214 window_out.set(Window::DimZ, Window::Dimension(0, _output->info()->dimension(2), _info.y()));
215
216 Iterator input(_input, window);
217 Iterator output(_output, window_out);
218
219 const int offset_y_out = _output->info()->strides_in_bytes().y() / sizeof(uint8_t);
220 const int offset_z_out = _output->info()->strides_in_bytes().z() / sizeof(uint8_t);
Michalis Spyroua4f378d2019-04-26 14:54:54 +0100[diff] [blame]221 execute_window_loop(window_out, [&](const Coordinates &)
Michalis Spyrou7c9541c2018-09-20 17:40:04 +0100[diff] [blame]222 {
223 const uint8x16_t data = vld1q_u8(reinterpret_cast<const uint8_t *>(input.ptr()));
224 auto out = reinterpret_cast<uint8_t *>(output.ptr());
225
226 vst1q_u8(out, data);
227 vst1q_u8(out + offset_y_out, data);
228 vst1q_u8(out + offset_z_out, data);
229 vst1q_u8(out + offset_y_out + offset_z_out, data);
230 },
231 input, output);
232}
233
234void NEUpsampleLayerKernel::upsample_f16_nchw(const arm_compute::Window &window)
235{
236 ARM_COMPUTE_UNUSED(window);
237#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
238 Window window_in(window);
239 window_in.set(Window::DimX, Window::Dimension(0, _input->info()->dimension(0), _num_elems_processed_per_iteration_x));
240
241 Window window_out(window);
242 window_out.set(Window::DimY, Window::Dimension(0, _output->info()->dimension(1), _info.y()));
243
244 Iterator input(_input, window_in);
245 Iterator output(_output, window_out);
246 const int offset_y_out = _output->info()->strides_in_bytes().y() / sizeof(float16_t);
247
Michalis Spyroua4f378d2019-04-26 14:54:54 +0100[diff] [blame]248 execute_window_loop(window_out, [&](const Coordinates &)
Michalis Spyrou7c9541c2018-09-20 17:40:04 +0100[diff] [blame]249 {
250 const float16x8_t data = vld1q_f16(reinterpret_cast<const float16_t *>(input.ptr()));
251 const float16x8_t data_out1 = { vgetq_lane_f16(data, 0), vgetq_lane_f16(data, 0), vgetq_lane_f16(data, 1), vgetq_lane_f16(data, 1),
252 vgetq_lane_f16(data, 2), vgetq_lane_f16(data, 2), vgetq_lane_f16(data, 3), vgetq_lane_f16(data, 3)
253 };
254 const float16x8_t data_out2 = { vgetq_lane_f16(data, 4), vgetq_lane_f16(data, 4), vgetq_lane_f16(data, 5), vgetq_lane_f16(data, 5),
255 vgetq_lane_f16(data, 6), vgetq_lane_f16(data, 6), vgetq_lane_f16(data, 7), vgetq_lane_f16(data, 7)
256 };
257 auto out = reinterpret_cast<float16_t *>(output.ptr());
258
259 vst1q_f16(out, data_out1);
260 vst1q_f16(out + 8, data_out2);
261 vst1q_f16(out + offset_y_out, data_out1);
262 vst1q_f16(out + offset_y_out + 8, data_out2);
263 },
264 input, output);
265#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
266}
267
268void NEUpsampleLayerKernel::upsample_f16_nhwc(const arm_compute::Window &window)
269{
270 ARM_COMPUTE_UNUSED(window);
271#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
272 Window window_out(window);
273 window_out.set(Window::DimY, Window::Dimension(0, _output->info()->dimension(1), _info.x()));
274 window_out.set(Window::DimZ, Window::Dimension(0, _output->info()->dimension(2), _info.y()));
275
276 Iterator input(_input, window);
277 Iterator output(_output, window_out);
278 const int offset_y_out = _output->info()->strides_in_bytes().y() / sizeof(float16_t);
279 const int offset_z_out = _output->info()->strides_in_bytes().z() / sizeof(float16_t);
280
Michalis Spyroua4f378d2019-04-26 14:54:54 +0100[diff] [blame]281 execute_window_loop(window_out, [&](const Coordinates &)
Michalis Spyrou7c9541c2018-09-20 17:40:04 +0100[diff] [blame]282 {
283 const float16x8_t data = vld1q_f16(reinterpret_cast<const float16_t *>(input.ptr()));
284 auto out = reinterpret_cast<float16_t *>(output.ptr());
285
286 vst1q_f16(out, data);
287 vst1q_f16(out + offset_y_out, data);
288 vst1q_f16(out + offset_z_out, data);
289 vst1q_f16(out + offset_y_out + offset_z_out, data);
290 },
291 input, output);
292#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
293}
294
295void NEUpsampleLayerKernel::configure(const ITensor *input, ITensor *output, const Size2D &info, const InterpolationPolicy policy)
296{
297 ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
298 ARM_COMPUTE_UNUSED(policy);
299
300 _input = input;
301 _output = output;
302 _info = info;
303
304 const DataLayout data_layout = input->info()->data_layout();
305
306 TensorShape output_shape = misc::shape_calculator::compute_upsample_shape(*input->info(), info);
307 auto_init_if_empty(*output->info(), output_shape, 1, input->info()->data_type());
308 output->info()->set_data_layout(data_layout);
309
310 // Perform validation step
311 ARM_COMPUTE_ERROR_THROW_ON(NEUpsampleLayerKernel::validate(input->info(), output->info(), info, policy));
312
313 _num_elems_processed_per_iteration_x = 16 / output->info()->element_size();
314
315 switch(data_layout)
316 {
317 case DataLayout::NCHW:
318 {
319 switch(input->info()->data_type())
320 {
321 case DataType::QASYMM8:
322 _func = &NEUpsampleLayerKernel::upsample_qasymm8_nchw;
323 break;
324 case DataType::F32:
325 _func = &NEUpsampleLayerKernel::upsample_f32_nchw;
326 break;
327#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
328 case DataType::F16:
329 _func = &NEUpsampleLayerKernel::upsample_f16_nchw;
330 break;
331#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
332 default:
333 ARM_COMPUTE_ERROR("Not implemented");
334 }
335 break;
336 }
337 case DataLayout::NHWC:
338 {
339 switch(input->info()->data_type())
340 {
341 case DataType::QASYMM8:
342 _func = &NEUpsampleLayerKernel::upsample_qasymm8_nhwc;
343 break;
344 case DataType::F32:
345 _func = &NEUpsampleLayerKernel::upsample_f32_nhwc;
346 break;
347#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
348 case DataType::F16:
349 _func = &NEUpsampleLayerKernel::upsample_f16_nhwc;
350 break;
351#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
352 default:
353 ARM_COMPUTE_ERROR("Not implemented");
354 }
355 break;
356 }
357 default:
358 ARM_COMPUTE_ERROR("Not implemented");
359 }
360
361 // Configure window
362 std::pair<Status, Window> win_config = validate_and_configure_window(input->info(),
363 output->info(),
364 _num_elems_processed_per_iteration_x,
365 info);
366 ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
367 INEKernel::configure(win_config.second);
368}
369
370void NEUpsampleLayerKernel::run(const Window &window, const ThreadInfo &info)
371{
372 ARM_COMPUTE_UNUSED(info);
373 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
374 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
375 ARM_COMPUTE_ERROR_ON(_func == nullptr);
376
377 (this->*_func)(window);
378}
379} // namespace arm_compute