1 /*
2 * Copyright (c) 2017-2020 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/NERemapKernel.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/Validate.h"
31 #include "arm_compute/core/Window.h"
32 #include "src/core/AccessWindowStatic.h"
33 #include "src/core/helpers/AutoConfiguration.h"
34 #include "src/core/helpers/ScaleHelpers.h"
35 #include "src/core/helpers/WindowHelpers.h"
36
37 #include <arm_neon.h>
38 #include <cstddef>
39 #include <cstdint>
40
41 using namespace arm_compute;
42
43 namespace arm_compute
44 {
45 class Coordinates;
46 } // namespace arm_compute
47
48 namespace
49 {
offset_nearest_interpolation(const float * mapx_ptr,const float * mapy_ptr,const float32x4_t & width,const float32x4_t & height,const int32x4_t & stride)50 inline int32x4_t offset_nearest_interpolation(const float *mapx_ptr, const float *mapy_ptr, const float32x4_t &width, const float32x4_t &height, const int32x4_t &stride)
51 {
52 const float32x4_t lowerxy = vdupq_n_f32(-1.f);
53
54 float32x4_t x = vld1q_f32(mapx_ptr);
55 float32x4_t y = vld1q_f32(mapy_ptr);
56
57 // Clamp x coordinates
58 x = vmaxq_f32(lowerxy, vminq_f32(x, width));
59 y = vmaxq_f32(lowerxy, vminq_f32(y, height));
60
61 const int32x4_t x_s32 = vcvtq_s32_f32(x);
62 const int32x4_t y_s32 = vcvtq_s32_f32(y);
63
64 return vmlaq_s32(x_s32, y_s32, stride);
65 }
66
67 } // namespace
68
NERemapKernel()69 NERemapKernel::NERemapKernel()
70 : _func(nullptr), _input(nullptr), _output(nullptr), _map_x(nullptr), _map_y(nullptr)
71 {
72 }
73
border_size() const74 BorderSize NERemapKernel::border_size() const
75 {
76 return BorderSize(1);
77 }
78
configure(const ITensor * input,const ITensor * map_x,const ITensor * map_y,ITensor * output,InterpolationPolicy policy)79 void NERemapKernel::configure(const ITensor *input, const ITensor *map_x, const ITensor *map_y, ITensor *output, InterpolationPolicy policy)
80 {
81 ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8);
82 ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::U8);
83 ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(map_x, 1, DataType::F32);
84 ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(map_y, 1, DataType::F32);
85
86 _input = input;
87 _output = output;
88 _map_x = map_x;
89 _map_y = map_y;
90
91 switch(policy)
92 {
93 case InterpolationPolicy::NEAREST_NEIGHBOR:
94 {
95 _func = &NERemapKernel::remap_nearest;
96 break;
97 }
98 case InterpolationPolicy::BILINEAR:
99 {
100 _func = &NERemapKernel::remap_bilinear;
101 break;
102 }
103 default:
104 ARM_COMPUTE_ERROR("Unsupported interpolation mode");
105 break;
106 }
107
108 constexpr unsigned int num_elems_processed_per_iteration = 16;
109
110 // Configure kernel window
111 Window win = calculate_max_window(*output->info(), Steps(num_elems_processed_per_iteration));
112
113 const int total_right = ceil_to_multiple(input->info()->dimension(0), num_elems_processed_per_iteration);
114 const int access_right = total_right + (((total_right - input->info()->dimension(0)) == 0) ? border_size().right : 0);
115
116 AccessWindowStatic input_access(input->info(), -border_size().left, -border_size().top, access_right, input->info()->dimension(1) + border_size().bottom);
117
118 AccessWindowHorizontal output_access(output->info(), 0, num_elems_processed_per_iteration);
119 AccessWindowHorizontal mapx_access(map_x->info(), 0, num_elems_processed_per_iteration);
120 AccessWindowHorizontal mapy_access(map_y->info(), 0, num_elems_processed_per_iteration);
121
122 update_window_and_padding(win, input_access, mapx_access, mapy_access, output_access);
123
124 output_access.set_valid_region(win, ValidRegion(Coordinates(), output->info()->tensor_shape()));
125
126 INEKernel::configure(win);
127 }
128
remap_nearest(const Window & window)129 void NERemapKernel::remap_nearest(const Window &window)
130 {
131 // Don't increment in X and Y direction for the input tensor
132 // A pointer to the start of this plane is needed as base for the precomputed offsets
133 Window win_in(window);
134 win_in.set(Window::DimX, Window::Dimension(0, 0, 0));
135 win_in.set(Window::DimY, Window::Dimension(0, 0, 0));
136
137 Iterator in(_input, win_in);
138 Iterator out(_output, window);
139 Iterator mapx(_map_x, window);
140 Iterator mapy(_map_y, window);
141
142 const float32x4_t width = vdupq_n_f32(static_cast<float>(_input->info()->dimension(0)));
143 const float32x4_t height = vdupq_n_f32(static_cast<float>(_input->info()->dimension(1)));
144 const int32x4_t in_stride = vdupq_n_s32(static_cast<int32_t>(_input->info()->strides_in_bytes()[1]));
145
146 execute_window_loop(window, [&](const Coordinates &)
147 {
148 const auto mapx_ptr = reinterpret_cast<const float *>(mapx.ptr());
149 const auto mapy_ptr = reinterpret_cast<const float *>(mapy.ptr());
150 const uint8_t *in_ptr = in.ptr();
151
152 const int32x4_t offset0 = offset_nearest_interpolation(mapx_ptr + 0, mapy_ptr + 0, width, height, in_stride);
153 const int32x4_t offset1 = offset_nearest_interpolation(mapx_ptr + 4, mapy_ptr + 4, width, height, in_stride);
154 const int32x4_t offset2 = offset_nearest_interpolation(mapx_ptr + 8, mapy_ptr + 8, width, height, in_stride);
155 const int32x4_t offset3 = offset_nearest_interpolation(mapx_ptr + 12, mapy_ptr + 12, width, height, in_stride);
156
157 uint8x16_t tmp = vdupq_n_u8(0);
158 tmp = vsetq_lane_u8(in_ptr[vgetq_lane_s32(offset0, 0)], tmp, 0);
159 tmp = vsetq_lane_u8(in_ptr[vgetq_lane_s32(offset0, 1)], tmp, 1);
160 tmp = vsetq_lane_u8(in_ptr[vgetq_lane_s32(offset0, 2)], tmp, 2);
161 tmp = vsetq_lane_u8(in_ptr[vgetq_lane_s32(offset0, 3)], tmp, 3);
162 tmp = vsetq_lane_u8(in_ptr[vgetq_lane_s32(offset1, 0)], tmp, 4);
163 tmp = vsetq_lane_u8(in_ptr[vgetq_lane_s32(offset1, 1)], tmp, 5);
164 tmp = vsetq_lane_u8(in_ptr[vgetq_lane_s32(offset1, 2)], tmp, 6);
165 tmp = vsetq_lane_u8(in_ptr[vgetq_lane_s32(offset1, 3)], tmp, 7);
166 tmp = vsetq_lane_u8(in_ptr[vgetq_lane_s32(offset2, 0)], tmp, 8);
167 tmp = vsetq_lane_u8(in_ptr[vgetq_lane_s32(offset2, 1)], tmp, 9);
168 tmp = vsetq_lane_u8(in_ptr[vgetq_lane_s32(offset2, 2)], tmp, 10);
169 tmp = vsetq_lane_u8(in_ptr[vgetq_lane_s32(offset2, 3)], tmp, 11);
170 tmp = vsetq_lane_u8(in_ptr[vgetq_lane_s32(offset3, 0)], tmp, 12);
171 tmp = vsetq_lane_u8(in_ptr[vgetq_lane_s32(offset3, 1)], tmp, 13);
172 tmp = vsetq_lane_u8(in_ptr[vgetq_lane_s32(offset3, 2)], tmp, 14);
173 tmp = vsetq_lane_u8(in_ptr[vgetq_lane_s32(offset3, 3)], tmp, 15);
174 vst1q_u8(out.ptr(), tmp);
175 },
176 in, out, mapx, mapy);
177 }
178
remap_bilinear(const Window & window)179 void NERemapKernel::remap_bilinear(const Window &window)
180 {
181 using namespace scale_helpers;
182
183 // Don't increment in X and Y direction for the input tensor
184 // A pointer to the start of this plane is needed as base for the precomputed offsets
185 Window win_in(window);
186 win_in.set(Window::DimX, Window::Dimension(0, 0, 0));
187 win_in.set(Window::DimY, Window::Dimension(0, 0, 0));
188
189 Iterator in(_input, win_in);
190 Iterator out(_output, window);
191 Iterator mapx(_map_x, window);
192 Iterator mapy(_map_y, window);
193
194 const size_t width = _input->info()->dimension(0);
195 const size_t height = _input->info()->dimension(1);
196 const size_t in_stride = _input->info()->strides_in_bytes()[1];
197
198 execute_window_loop(window, [&](const Coordinates &)
199 {
200 const auto mapx_ptr = reinterpret_cast<float *>(mapx.ptr());
201 const auto mapy_ptr = reinterpret_cast<float *>(mapy.ptr());
202 const uint8_t *in_ptr = in.ptr();
203
204 uint8x8_t tmp0 = vdup_n_u8(0);
205 tmp0 = vset_lane_u8(pixel_bilinear_c1_clamp(in_ptr, in_stride, width, height, mapx_ptr[0], mapy_ptr[0]), tmp0, 0);
206 tmp0 = vset_lane_u8(pixel_bilinear_c1_clamp(in_ptr, in_stride, width, height, mapx_ptr[1], mapy_ptr[1]), tmp0, 1);
207 tmp0 = vset_lane_u8(pixel_bilinear_c1_clamp(in_ptr, in_stride, width, height, mapx_ptr[2], mapy_ptr[2]), tmp0, 2);
208 tmp0 = vset_lane_u8(pixel_bilinear_c1_clamp(in_ptr, in_stride, width, height, mapx_ptr[3], mapy_ptr[3]), tmp0, 3);
209 tmp0 = vset_lane_u8(pixel_bilinear_c1_clamp(in_ptr, in_stride, width, height, mapx_ptr[4], mapy_ptr[4]), tmp0, 4);
210 tmp0 = vset_lane_u8(pixel_bilinear_c1_clamp(in_ptr, in_stride, width, height, mapx_ptr[5], mapy_ptr[5]), tmp0, 5);
211 tmp0 = vset_lane_u8(pixel_bilinear_c1_clamp(in_ptr, in_stride, width, height, mapx_ptr[6], mapy_ptr[6]), tmp0, 6);
212 tmp0 = vset_lane_u8(pixel_bilinear_c1_clamp(in_ptr, in_stride, width, height, mapx_ptr[7], mapy_ptr[7]), tmp0, 7);
213
214 uint8x8_t tmp1 = vdup_n_u8(0);
215 tmp1 = vset_lane_u8(pixel_bilinear_c1_clamp(in_ptr, in_stride, width, height, mapx_ptr[8], mapy_ptr[8]), tmp1, 0);
216 tmp1 = vset_lane_u8(pixel_bilinear_c1_clamp(in_ptr, in_stride, width, height, mapx_ptr[9], mapy_ptr[9]), tmp1, 1);
217 tmp1 = vset_lane_u8(pixel_bilinear_c1_clamp(in_ptr, in_stride, width, height, mapx_ptr[10], mapy_ptr[10]), tmp1, 2);
218 tmp1 = vset_lane_u8(pixel_bilinear_c1_clamp(in_ptr, in_stride, width, height, mapx_ptr[11], mapy_ptr[11]), tmp1, 3);
219 tmp1 = vset_lane_u8(pixel_bilinear_c1_clamp(in_ptr, in_stride, width, height, mapx_ptr[12], mapy_ptr[12]), tmp1, 4);
220 tmp1 = vset_lane_u8(pixel_bilinear_c1_clamp(in_ptr, in_stride, width, height, mapx_ptr[13], mapy_ptr[13]), tmp1, 5);
221 tmp1 = vset_lane_u8(pixel_bilinear_c1_clamp(in_ptr, in_stride, width, height, mapx_ptr[14], mapy_ptr[14]), tmp1, 6);
222 tmp1 = vset_lane_u8(pixel_bilinear_c1_clamp(in_ptr, in_stride, width, height, mapx_ptr[15], mapy_ptr[15]), tmp1, 7);
223
224 vst1q_u8(out.ptr(), vcombine_u8(tmp0, tmp1));
225 },
226 in, out, mapx, mapy);
227 }
228
run(const Window & window,const ThreadInfo & info)229 void NERemapKernel::run(const Window &window, const ThreadInfo &info)
230 {
231 ARM_COMPUTE_UNUSED(info);
232 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
233 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
234 ARM_COMPUTE_ERROR_ON(_func == nullptr);
235
236 (this->*_func)(window);
237 }
238