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/NESobel3x3Kernel.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/Types.h"
31 #include "arm_compute/core/Validate.h"
32 #include "src/core/helpers/AutoConfiguration.h"
33 #include "src/core/helpers/WindowHelpers.h"
34
35 #include <arm_neon.h>
36 #include <cstdint>
37
38 using namespace arm_compute;
39
NESobel3x3Kernel()40 NESobel3x3Kernel::NESobel3x3Kernel()
41 : _run_sobel_x(false), _run_sobel_y(false), _input(nullptr), _output_x(nullptr), _output_y(nullptr)
42 {
43 }
44
border_size() const45 BorderSize NESobel3x3Kernel::border_size() const
46 {
47 return BorderSize{ 1 };
48 }
49
configure(const ITensor * input,ITensor * output_x,ITensor * output_y,bool border_undefined)50 void NESobel3x3Kernel::configure(const ITensor *input, ITensor *output_x, ITensor *output_y, bool border_undefined)
51 {
52 ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8);
53 ARM_COMPUTE_ERROR_ON((output_x == nullptr) && (output_y == nullptr));
54
55 _run_sobel_x = output_x != nullptr;
56 _run_sobel_y = output_y != nullptr;
57
58 if(_run_sobel_x)
59 {
60 ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output_x, 1, DataType::S16);
61 }
62
63 if(_run_sobel_y)
64 {
65 ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output_y, 1, DataType::S16);
66 }
67
68 _input = input;
69 _output_x = output_x;
70 _output_y = output_y;
71
72 // Configure kernel window
73 constexpr unsigned int num_elems_processed_per_iteration = 8;
74 constexpr unsigned int num_elems_read_per_iteration = 16;
75 constexpr unsigned int num_elems_written_per_iteration = 8;
76 constexpr unsigned int num_rows_read_per_iteration = 3;
77
78 Window win = calculate_max_window(*input->info(), Steps(num_elems_processed_per_iteration), border_undefined, border_size());
79 AccessWindowHorizontal output_x_access(output_x == nullptr ? nullptr : output_x->info(), 0, num_elems_written_per_iteration);
80 AccessWindowHorizontal output_y_access(output_y == nullptr ? nullptr : output_y->info(), 0, num_elems_written_per_iteration);
81
82 update_window_and_padding(win,
83 AccessWindowRectangle(input->info(), -border_size().left, -border_size().top, num_elems_read_per_iteration, num_rows_read_per_iteration),
84 output_x_access,
85 output_y_access);
86
87 output_x_access.set_valid_region(win, input->info()->valid_region(), border_undefined, border_size());
88 output_y_access.set_valid_region(win, input->info()->valid_region(), border_undefined, border_size());
89
90 INEKernel::configure(win);
91 }
92
run(const Window & window,const ThreadInfo & info)93 void NESobel3x3Kernel::run(const Window &window, const ThreadInfo &info)
94 {
95 ARM_COMPUTE_UNUSED(info);
96 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
97 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
98
99 const unsigned char *const input_top_ptr = _input->ptr_to_element(Coordinates(-1, -1));
100 const unsigned char *const input_mid_ptr = _input->ptr_to_element(Coordinates(-1, 0));
101 const unsigned char *const input_bot_ptr = _input->ptr_to_element(Coordinates(-1, 1));
102
103 Iterator input(_input, window);
104 Iterator output_y;
105 Iterator output_x;
106
107 if(_run_sobel_y)
108 {
109 output_y = Iterator(_output_y, window);
110 }
111
112 if(_run_sobel_x)
113 {
114 output_x = Iterator(_output_x, window);
115 }
116
117 static const int16x8_t two = vdupq_n_s16(2);
118 static const int16x8_t minustwo = vdupq_n_s16(-2);
119
120 if(_run_sobel_y && _run_sobel_x)
121 {
122 execute_window_loop(window, [&](const Coordinates &)
123 {
124 const uint8x16_t top_data = vld1q_u8(input_top_ptr + input.offset());
125 const uint8x16_t mid_data = vld1q_u8(input_mid_ptr + input.offset());
126 const uint8x16_t bot_data = vld1q_u8(input_bot_ptr + input.offset());
127
128 const int16x8x2_t top_s16 =
129 {
130 {
131 vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(top_data))),
132 vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(top_data)))
133 }
134 };
135 const int16x8x2_t mid_s16 =
136 {
137 {
138 vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(mid_data))),
139 vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(mid_data)))
140 }
141 };
142 const int16x8x2_t bot_s16 =
143 {
144 {
145 vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(bot_data))),
146 vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(bot_data)))
147 }
148 };
149
150 //SOBEL Y
151 //top left
152 int16x8_t out_y = vnegq_s16(top_s16.val[0]);
153 //top mid
154 out_y = vmlaq_s16(out_y, vextq_s16(top_s16.val[0], top_s16.val[1], 1), minustwo);
155 //top right
156 out_y = vsubq_s16(out_y, vextq_s16(top_s16.val[0], top_s16.val[1], 2));
157 //bot left
158 out_y = vaddq_s16(out_y, bot_s16.val[0]);
159 //bot mid
160 out_y = vmlaq_s16(out_y, vextq_s16(bot_s16.val[0], bot_s16.val[1], 1), two);
161 //bot right
162 out_y = vaddq_s16(out_y, vextq_s16(bot_s16.val[0], bot_s16.val[1], 2));
163
164 vst1q_s16(reinterpret_cast<int16_t *>(output_y.ptr()), out_y);
165
166 //SOBEL X
167 //top left
168 int16x8_t out_x = vnegq_s16(top_s16.val[0]);
169 //top right
170 out_x = vaddq_s16(out_x, vextq_s16(top_s16.val[0], top_s16.val[1], 2));
171 //mid left
172 out_x = vmlaq_s16(out_x, mid_s16.val[0], minustwo);
173 //mid right
174 out_x = vmlaq_s16(out_x, vextq_s16(mid_s16.val[0], mid_s16.val[1], 2), two);
175 //bot left
176 out_x = vsubq_s16(out_x, bot_s16.val[0]);
177 //bot right
178 out_x = vaddq_s16(out_x, vextq_s16(bot_s16.val[0], bot_s16.val[1], 2));
179
180 vst1q_s16(reinterpret_cast<int16_t *>(output_x.ptr()), out_x);
181 },
182 input, output_x, output_y);
183 }
184 else if(_run_sobel_x)
185 {
186 execute_window_loop(window, [&](const Coordinates &)
187 {
188 const uint8x16_t top_data = vld1q_u8(input_top_ptr + input.offset());
189 const uint8x16_t mid_data = vld1q_u8(input_mid_ptr + input.offset());
190 const uint8x16_t bot_data = vld1q_u8(input_bot_ptr + input.offset());
191
192 const int16x8x2_t top_s16 =
193 {
194 {
195 vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(top_data))),
196 vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(top_data)))
197 }
198 };
199 const int16x8x2_t mid_s16 =
200 {
201 {
202 vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(mid_data))),
203 vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(mid_data)))
204 }
205 };
206 const int16x8x2_t bot_s16 =
207 {
208 {
209 vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(bot_data))),
210 vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(bot_data)))
211 }
212 };
213
214 //SOBEL X
215 //top left
216 int16x8_t out = vnegq_s16(top_s16.val[0]);
217 //top right
218 out = vaddq_s16(out, vextq_s16(top_s16.val[0], top_s16.val[1], 2));
219 //mid left
220 out = vmlaq_s16(out, mid_s16.val[0], minustwo);
221 //mid right
222 out = vmlaq_s16(out, vextq_s16(mid_s16.val[0], mid_s16.val[1], 2), two);
223 //bot left
224 out = vsubq_s16(out, bot_s16.val[0]);
225 //bot right
226 out = vaddq_s16(out, vextq_s16(bot_s16.val[0], bot_s16.val[1], 2));
227
228 vst1q_s16(reinterpret_cast<int16_t *>(output_x.ptr()), out);
229 },
230 input, output_x);
231 }
232 else if(_run_sobel_y)
233 {
234 execute_window_loop(window, [&](const Coordinates &)
235 {
236 const uint8x16_t top_data = vld1q_u8(input_top_ptr + input.offset());
237 const uint8x16_t bot_data = vld1q_u8(input_bot_ptr + input.offset());
238
239 const int16x8x2_t top_s16 =
240 {
241 {
242 vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(top_data))),
243 vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(top_data)))
244 }
245 };
246 const int16x8x2_t bot_s16 =
247 {
248 {
249 vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(bot_data))),
250 vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(bot_data)))
251 }
252 };
253
254 //SOBEL Y
255 //top left
256 int16x8_t out = vnegq_s16(top_s16.val[0]);
257 //top mid
258 out = vmlaq_s16(out, vextq_s16(top_s16.val[0], top_s16.val[1], 1), minustwo);
259 //top right
260 out = vsubq_s16(out, vextq_s16(top_s16.val[0], top_s16.val[1], 2));
261 //bot left
262 out = vaddq_s16(out, bot_s16.val[0]);
263 //bot mid
264 out = vmlaq_s16(out, vextq_s16(bot_s16.val[0], bot_s16.val[1], 1), two);
265 //bot right
266 out = vaddq_s16(out, vextq_s16(bot_s16.val[0], bot_s16.val[1], 2));
267
268 vst1q_s16(reinterpret_cast<int16_t *>(output_y.ptr()), out);
269 },
270 input, output_y);
271 }
272 }
273