1 /*
2 * Copyright (c) 2016-2020 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/NEGaussianPyramidKernel.h"
25
26 #include "arm_compute/core/Coordinates.h"
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/Types.h"
32 #include "arm_compute/core/Validate.h"
33 #include "arm_compute/core/Window.h"
34 #include "src/core/NEON/INEKernel.h"
35 #include "src/core/helpers/AutoConfiguration.h"
36 #include "src/core/helpers/WindowHelpers.h"
37
38 #include <arm_neon.h>
39 #include <cstddef>
40 #include <cstdint>
41 #include <tuple>
42
43 using namespace arm_compute;
44
NEGaussianPyramidHorKernel()45 NEGaussianPyramidHorKernel::NEGaussianPyramidHorKernel()
46 : _l2_load_offset(0)
47 {
48 }
49
border_size() const50 BorderSize NEGaussianPyramidHorKernel::border_size() const
51 {
52 return BorderSize{ 0, 2 };
53 }
54
configure(const ITensor * input,ITensor * output)55 void NEGaussianPyramidHorKernel::configure(const ITensor *input, ITensor *output)
56 {
57 ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8);
58 ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::S16);
59 ARM_COMPUTE_ERROR_ON(input->info()->dimension(1) != output->info()->dimension(1));
60
61 for(size_t i = 2; i < Coordinates::num_max_dimensions; ++i)
62 {
63 ARM_COMPUTE_ERROR_ON(input->info()->dimension(i) != output->info()->dimension(i));
64 }
65
66 _input = input;
67 _output = output;
68
69 // Configure kernel window
70 constexpr unsigned int num_elems_processed_per_iteration = 16;
71 constexpr unsigned int num_elems_read_per_iteration = 32;
72 constexpr unsigned int num_elems_written_per_iteration = 8;
73 const float scale_x = static_cast<float>(output->info()->dimension(0)) / input->info()->dimension(0);
74
75 Window win = calculate_max_window_horizontal(*input->info(), Steps(num_elems_processed_per_iteration));
76 AccessWindowHorizontal output_access(output->info(), 0, num_elems_written_per_iteration, scale_x);
77
78 // Sub sampling selects odd pixels (1, 3, 5, ...) for images with even
79 // width and even pixels (0, 2, 4, ...) for images with odd width. (Whether
80 // a pixel is even or odd is determined based on the tensor shape not the
81 // valid region!)
82 // Thus the offset from which the first pixel (L2) for the convolution is
83 // loaded depends on the anchor and shape of the valid region.
84 // In the case of an even shape (= even image width) we need to load L2
85 // from -2 if the anchor is odd and from -1 if the anchor is even. That
86 // makes sure that L2 is always loaded from an odd pixel.
87 // On the other hand, for an odd shape (= odd image width) we need to load
88 // L2 from -1 if the anchor is odd and from -2 if the anchor is even to
89 // achieve the opposite effect.
90 // The condition can be simplified to checking whether anchor + shape is
91 // odd (-2) or even (-1) as only adding an odd and an even number will have
92 // an odd result.
93 _l2_load_offset = -border_size().left;
94
95 if((_input->info()->valid_region().anchor[0] + _input->info()->valid_region().shape[0]) % 2 == 0)
96 {
97 _l2_load_offset += 1;
98 }
99
100 // Replace input access with static window
101 update_window_and_padding(win,
102 AccessWindowHorizontal(input->info(), _l2_load_offset, num_elems_read_per_iteration),
103 output_access);
104
105 output->info()->set_valid_region(ValidRegion(Coordinates(), output->info()->tensor_shape()));
106
107 INEKernel::configure(win);
108 }
109
run(const Window & window,const ThreadInfo & info)110 void NEGaussianPyramidHorKernel::run(const Window &window, const ThreadInfo &info)
111 {
112 ARM_COMPUTE_UNUSED(info);
113 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
114 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
115 ARM_COMPUTE_ERROR_ON(window.x().step() % 2);
116
117 static const int16x8_t six = vdupq_n_s16(6);
118 static const int16x8_t four = vdupq_n_s16(4);
119
120 Window win_in(window);
121 win_in.shift(Window::DimX, _l2_load_offset);
122
123 Iterator in(_input, win_in);
124
125 // The output is half the width of the input
126 Window win_out(window);
127 win_out.scale(Window::DimX, 0.5f);
128
129 Iterator out(_output, win_out);
130
131 execute_window_loop(window, [&](const Coordinates &)
132 {
133 const uint8x16x2_t data_2q = vld2q_u8(in.ptr());
134 const uint8x16_t &data_even = data_2q.val[0];
135 const uint8x16_t &data_odd = data_2q.val[1];
136
137 const int16x8_t data_l2 = vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(data_even)));
138 const int16x8_t data_l1 = vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(data_odd)));
139 const int16x8_t data_m = vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(vextq_u8(data_even, data_even, 1))));
140 const int16x8_t data_r1 = vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(vextq_u8(data_odd, data_odd, 1))));
141 const int16x8_t data_r2 = vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(vextq_u8(data_even, data_even, 2))));
142
143 int16x8_t out_val = vaddq_s16(data_l2, data_r2);
144 out_val = vmlaq_s16(out_val, data_l1, four);
145 out_val = vmlaq_s16(out_val, data_m, six);
146 out_val = vmlaq_s16(out_val, data_r1, four);
147
148 vst1q_s16(reinterpret_cast<int16_t *>(out.ptr()), out_val);
149 },
150 in, out);
151 }
152
NEGaussianPyramidVertKernel()153 NEGaussianPyramidVertKernel::NEGaussianPyramidVertKernel()
154 : _t2_load_offset(0)
155 {
156 }
157
border_size() const158 BorderSize NEGaussianPyramidVertKernel::border_size() const
159 {
160 return BorderSize{ 2, 0 };
161 }
162
configure(const ITensor * input,ITensor * output)163 void NEGaussianPyramidVertKernel::configure(const ITensor *input, ITensor *output)
164 {
165 ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::S16);
166 ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::U8);
167 ARM_COMPUTE_ERROR_ON(input->info()->dimension(0) != output->info()->dimension(0));
168
169 for(size_t i = 2; i < Coordinates::num_max_dimensions; ++i)
170 {
171 ARM_COMPUTE_ERROR_ON(input->info()->dimension(i) != output->info()->dimension(i));
172 }
173
174 _input = input;
175 _output = output;
176
177 // Configure kernel window
178 constexpr unsigned int num_elems_processed_per_iteration = 16;
179 constexpr unsigned int num_rows_processed_per_iteration = 2;
180
181 constexpr unsigned int num_elems_written_per_iteration = 16;
182 constexpr unsigned int num_rows_written_per_iteration = 1;
183
184 constexpr unsigned int num_elems_read_per_iteration = 16;
185 constexpr unsigned int num_rows_read_per_iteration = 5;
186
187 const float scale_y = static_cast<float>(output->info()->dimension(1)) / input->info()->dimension(1);
188
189 Window win = calculate_max_window(*input->info(), Steps(num_elems_processed_per_iteration, num_rows_processed_per_iteration));
190 AccessWindowRectangle output_access(output->info(), 0, 0, num_elems_written_per_iteration, num_rows_written_per_iteration, 1.f, scale_y);
191
192 // Determine whether we need to load even or odd rows. See above for a
193 // detailed explanation.
194 _t2_load_offset = -border_size().top;
195
196 if((_input->info()->valid_region().anchor[1] + _input->info()->valid_region().shape[1]) % 2 == 0)
197 {
198 _t2_load_offset += 1;
199 }
200
201 update_window_and_padding(win,
202 AccessWindowRectangle(input->info(), 0, _t2_load_offset, num_elems_read_per_iteration, num_rows_read_per_iteration),
203 output_access);
204
205 output->info()->set_valid_region(ValidRegion(Coordinates(), output->info()->tensor_shape()));
206
207 INEKernel::configure(win);
208 }
209
run(const Window & window,const ThreadInfo & info)210 void NEGaussianPyramidVertKernel::run(const Window &window, const ThreadInfo &info)
211 {
212 ARM_COMPUTE_UNUSED(info);
213 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
214 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
215 ARM_COMPUTE_ERROR_ON(window.x().step() != 16);
216 ARM_COMPUTE_ERROR_ON(window.y().step() % 2);
217 ARM_COMPUTE_ERROR_ON(_input->buffer() == nullptr);
218
219 static const uint16x8_t six = vdupq_n_u16(6);
220 static const uint16x8_t four = vdupq_n_u16(4);
221
222 Window win_in(window);
223 // Need to load two times 8 values instead of 16 values once
224 win_in.set_dimension_step(Window::DimX, 8);
225 win_in.shift(Window::DimY, _t2_load_offset);
226
227 Iterator in(_input, win_in);
228
229 // Output's height is half of input's
230 Window win_out(window);
231 win_out.scale(Window::DimY, 0.5f);
232
233 Iterator out(_output, win_out);
234
235 const uint8_t *input_top2_ptr = _input->buffer() + _input->info()->offset_element_in_bytes(Coordinates(0, 0));
236 const uint8_t *input_top_ptr = _input->buffer() + _input->info()->offset_element_in_bytes(Coordinates(0, 1));
237 const uint8_t *input_mid_ptr = _input->buffer() + _input->info()->offset_element_in_bytes(Coordinates(0, 2));
238 const uint8_t *input_low_ptr = _input->buffer() + _input->info()->offset_element_in_bytes(Coordinates(0, 3));
239 const uint8_t *input_low2_ptr = _input->buffer() + _input->info()->offset_element_in_bytes(Coordinates(0, 4));
240
241 execute_window_loop(window, [&](const Coordinates &)
242 {
243 // Low data
244 const uint16x8_t data_low_t2 = vreinterpretq_u16_s16(vld1q_s16(reinterpret_cast<const int16_t *>(input_top2_ptr + in.offset())));
245 const uint16x8_t data_low_t1 = vreinterpretq_u16_s16(vld1q_s16(reinterpret_cast<const int16_t *>(input_top_ptr + in.offset())));
246 const uint16x8_t data_low_m = vreinterpretq_u16_s16(vld1q_s16(reinterpret_cast<const int16_t *>(input_mid_ptr + in.offset())));
247 const uint16x8_t data_low_b1 = vreinterpretq_u16_s16(vld1q_s16(reinterpret_cast<const int16_t *>(input_low_ptr + in.offset())));
248 const uint16x8_t data_low_b2 = vreinterpretq_u16_s16(vld1q_s16(reinterpret_cast<const int16_t *>(input_low2_ptr + in.offset())));
249
250 uint16x8_t out_low = vaddq_u16(data_low_t2, data_low_b2);
251 out_low = vmlaq_u16(out_low, data_low_t1, four);
252 out_low = vmlaq_u16(out_low, data_low_m, six);
253 out_low = vmlaq_u16(out_low, data_low_b1, four);
254
255 in.increment(Window::DimX);
256
257 // High data
258 const uint16x8_t data_high_t2 = vreinterpretq_u16_s16(vld1q_s16(reinterpret_cast<const int16_t *>(input_top2_ptr + in.offset())));
259 const uint16x8_t data_high_t1 = vreinterpretq_u16_s16(vld1q_s16(reinterpret_cast<const int16_t *>(input_top_ptr + in.offset())));
260 const uint16x8_t data_high_m = vreinterpretq_u16_s16(vld1q_s16(reinterpret_cast<const int16_t *>(input_mid_ptr + in.offset())));
261 const uint16x8_t data_high_b1 = vreinterpretq_u16_s16(vld1q_s16(reinterpret_cast<const int16_t *>(input_low_ptr + in.offset())));
262 const uint16x8_t data_high_b2 = vreinterpretq_u16_s16(vld1q_s16(reinterpret_cast<const int16_t *>(input_low2_ptr + in.offset())));
263
264 uint16x8_t out_high = vaddq_u16(data_high_t2, data_high_b2);
265 out_high = vmlaq_u16(out_high, data_high_t1, four);
266 out_high = vmlaq_u16(out_high, data_high_m, six);
267 out_high = vmlaq_u16(out_high, data_high_b1, four);
268
269 vst1q_u8(out.ptr(), vcombine_u8(vqshrn_n_u16(out_low, 8), vqshrn_n_u16(out_high, 8)));
270 },
271 in, out);
272 }
273