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/NEDerivativeKernel.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 "src/core/helpers/AutoConfiguration.h"
33 #include "src/core/helpers/WindowHelpers.h"
34
35 #include <arm_neon.h>
36 #include <cstddef>
37 #include <cstdint>
38
39 using namespace arm_compute;
40
41 namespace arm_compute
42 {
43 class Coordinates;
44 } // namespace arm_compute
45
NEDerivativeKernel()46 NEDerivativeKernel::NEDerivativeKernel()
47 : _func(nullptr), _input(nullptr), _output_x(nullptr), _output_y(nullptr)
48 {
49 }
50
border_size() const51 BorderSize NEDerivativeKernel::border_size() const
52 {
53 return BorderSize(1);
54 }
55
configure(const ITensor * input,ITensor * output_x,ITensor * output_y,bool border_undefined)56 void NEDerivativeKernel::configure(const ITensor *input, ITensor *output_x, ITensor *output_y, bool border_undefined)
57 {
58 ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8);
59 ARM_COMPUTE_ERROR_ON((output_x == nullptr) && (output_y == nullptr));
60
61 const bool run_der_x = output_x != nullptr;
62 const bool run_der_y = output_y != nullptr;
63
64 if(run_der_x)
65 {
66 ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output_x, 1, DataType::S16);
67 }
68
69 if(run_der_y)
70 {
71 ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output_y, 1, DataType::S16);
72 }
73
74 _input = input;
75 _output_x = output_x;
76 _output_y = output_y;
77
78 constexpr unsigned int num_elems_processed_per_iteration = 16;
79 constexpr unsigned int num_rows_read_per_iteration = 3;
80
81 Window win = calculate_max_window(*input->info(), Steps(num_elems_processed_per_iteration), border_undefined, border_size());
82
83 AccessWindowHorizontal out_x_access(output_x == nullptr ? nullptr : output_x->info(), 0, num_elems_processed_per_iteration);
84 AccessWindowHorizontal out_y_access(output_y == nullptr ? nullptr : output_y->info(), 0, num_elems_processed_per_iteration);
85
86 // TODO(COMPMID-1503) Fix x-access input bug in NEON kernel instead of '+2'
87 AccessWindowHorizontal in_x_access(input->info(), -border_size().left, num_elems_processed_per_iteration + 2);
88 AccessWindowRectangle in_y_access(input->info(), 0, -border_size().left, num_elems_processed_per_iteration, num_rows_read_per_iteration);
89
90 // TODO(COMPMID-1503) Fix x-access input bug in NEON kernel instead of '+2'
91 AccessWindowRectangle in_xy_access(input->info(), -border_size().left, -border_size().top, num_elems_processed_per_iteration + 2, num_rows_read_per_iteration);
92
93 if(run_der_x && run_der_y)
94 {
95 _func = &NEDerivativeKernel::derivative_xy;
96 update_window_and_padding(win, in_xy_access, out_x_access, out_y_access);
97 out_y_access.set_valid_region(win, input->info()->valid_region(), border_undefined, border_size());
98 out_x_access.set_valid_region(win, input->info()->valid_region(), border_undefined, border_size());
99 }
100 else
101 {
102 if(run_der_x)
103 {
104 _func = &NEDerivativeKernel::derivative_x;
105 update_window_and_padding(win, in_x_access, out_x_access);
106 out_x_access.set_valid_region(win, input->info()->valid_region(), border_undefined, border_size());
107 }
108 else if(run_der_y)
109 {
110 _func = &NEDerivativeKernel::derivative_y;
111 update_window_and_padding(win, in_y_access, out_y_access);
112 out_y_access.set_valid_region(win, input->info()->valid_region(), border_undefined, border_size());
113 }
114 else
115 {
116 ARM_COMPUTE_ERROR("At least one output must be NOT NULL");
117 }
118 }
119
120 INEKernel::configure(win);
121 }
122
derivative_x(const Window & window)123 void NEDerivativeKernel::derivative_x(const Window &window)
124 {
125 Iterator in(_input, window);
126 Iterator out_x(_output_x, window);
127
128 /* Apply 1-D centered point discrete derivative mask ([-1 0 1]) along the X direction */
129 execute_window_loop(window, [&](const Coordinates &)
130 {
131 /* Load left and right data */
132 const uint8x16_t l_data = vld1q_u8(in.ptr() - 1);
133 const uint8x16_t r_data = vld1q_u8(in.ptr() + 1);
134
135 /* Cast to int16 and perform the subtraction between the right and left data */
136 const int16x8_t out0 = vsubq_s16(vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(r_data))),
137 vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(l_data))));
138
139 /* Cast to int16 and perform the subtraction between the right and left data */
140 const int16x8_t out1 = vsubq_s16(vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(r_data))),
141 vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(l_data))));
142
143 /* Store result of derivative along the X direction */
144 vst1q_s16(reinterpret_cast<int16_t *>(out_x.ptr()), out0);
145 vst1q_s16(reinterpret_cast<int16_t *>(out_x.ptr()) + 8, out1);
146 },
147 in, out_x);
148 }
149
derivative_y(const Window & window)150 void NEDerivativeKernel::derivative_y(const Window &window)
151 {
152 Iterator in(_input, window);
153 Iterator out_y(_output_y, window);
154
155 const size_t stride = _input->info()->strides_in_bytes()[1];
156
157 /* Apply 1-D centered point discrete derivative mask ([-1 0 1]^T) along the Y direction */
158 execute_window_loop(window, [&](const Coordinates &)
159 {
160 /* Load top and bottom data */
161 const uint8x16_t t_data = vld1q_u8(in.ptr() - stride);
162 const uint8x16_t b_data = vld1q_u8(in.ptr() + stride);
163
164 /* Cast to int16 and perform the subtraction between the bottom and top data */
165 const int16x8_t out0 = vsubq_s16(vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(b_data))),
166 vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(t_data))));
167
168 /* Cast to int16 and perform the subtraction between the bottom and top data */
169 const int16x8_t out1 = vsubq_s16(vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(b_data))),
170 vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(t_data))));
171
172 /* Store result of derivative along the Y direction */
173 vst1q_s16(reinterpret_cast<int16_t *>(out_y.ptr()), out0);
174 vst1q_s16(reinterpret_cast<int16_t *>(out_y.ptr()) + 8, out1);
175 },
176 in, out_y);
177 }
178
derivative_xy(const Window & window)179 void NEDerivativeKernel::derivative_xy(const Window &window)
180 {
181 Iterator in(_input, window);
182 Iterator out_x(_output_x, window);
183 Iterator out_y(_output_y, window);
184
185 const size_t stride = _input->info()->strides_in_bytes()[1];
186
187 /* Apply 1-D centered point discrete derivative masks ([-1 0 1] and [-1 0 1]^T) along the X and Y directions */
188 execute_window_loop(window, [&](const Coordinates &)
189 {
190 /* Load top, bottom, left and right data */
191 const uint8x16_t t_data = vld1q_u8(in.ptr() - stride);
192 const uint8x16_t b_data = vld1q_u8(in.ptr() + stride);
193 const uint8x16_t l_data = vld1q_u8(in.ptr() - 1);
194 const uint8x16_t r_data = vld1q_u8(in.ptr() + 1);
195
196 /* Cast to int16 and perform the subtraction between the bottom and top data */
197 const int16x8_t out0 = vsubq_s16(vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(b_data))),
198 vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(t_data))));
199
200 /* Cast to int16 and perform the subtraction between the bottom and top data */
201 const int16x8_t out1 = vsubq_s16(vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(b_data))),
202 vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(t_data))));
203
204 /* Cast to int16 and perform the subtraction between the right and left data */
205 const int16x8_t out2 = vsubq_s16(vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(r_data))),
206 vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(l_data))));
207
208 /* Cast to int16 and perform the subtraction between the right and left data */
209 const int16x8_t out3 = vsubq_s16(vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(r_data))),
210 vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(l_data))));
211
212 /* Store result of derivative along the Y direction */
213 vst1q_s16(reinterpret_cast<int16_t *>(out_y.ptr()), out0);
214 vst1q_s16(reinterpret_cast<int16_t *>(out_y.ptr()) + 8, out1);
215
216 /* Store result of derivative along the X direction */
217 vst1q_s16(reinterpret_cast<int16_t *>(out_x.ptr()), out2);
218 vst1q_s16(reinterpret_cast<int16_t *>(out_x.ptr()) + 8, out3);
219 },
220 in, out_x, out_y);
221 }
222
run(const Window & window,const ThreadInfo & info)223 void NEDerivativeKernel::run(const Window &window, const ThreadInfo &info)
224 {
225 ARM_COMPUTE_UNUSED(info);
226 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
227 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
228 ARM_COMPUTE_ERROR_ON(_func == nullptr);
229
230 (this->*_func)(window);
231 }
232