2 * Copyright (c) 2017 ARM Limited.
4 * SPDX-License-Identifier: MIT
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:
13 * The above copyright notice and this permission notice shall be included in all
14 * copies or substantial portions of the Software.
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
24 #include "arm_compute/core/NEON/kernels/NERemapKernel.h"
26 #include "arm_compute/core/AccessWindowStatic.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/Validate.h"
32 #include "arm_compute/core/Window.h"
38 using namespace arm_compute;
43 } // namespace arm_compute
47 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)
49 static const float32x4_t lowerxy = vdupq_n_f32(-1.0f);
51 float32x4_t x = vld1q_f32(mapx_ptr);
52 float32x4_t y = vld1q_f32(mapy_ptr);
54 // Clamp x coordinates
55 x = vmaxq_f32(lowerxy, vminq_f32(x, width));
56 y = vmaxq_f32(lowerxy, vminq_f32(y, height));
58 const int32x4_t x_s32 = vcvtq_s32_f32(x);
59 const int32x4_t y_s32 = vcvtq_s32_f32(y);
61 return vmlaq_s32(x_s32, y_s32, stride);
66 NERemapKernel::NERemapKernel()
67 : _func(nullptr), _input(nullptr), _output(nullptr), _map_x(nullptr), _map_y(nullptr)
71 void NERemapKernel::configure(const ITensor *input, const ITensor *map_x, const ITensor *map_y, ITensor *output, InterpolationPolicy policy)
73 ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8);
74 ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::U8);
75 ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(map_x, 1, DataType::F32);
76 ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(map_y, 1, DataType::F32);
85 case InterpolationPolicy::NEAREST_NEIGHBOR:
87 _func = &NERemapKernel::remap_nearest;
90 case InterpolationPolicy::BILINEAR:
92 _func = &NERemapKernel::remap_bilinear;
96 ARM_COMPUTE_ERROR("Unsupported interpolation mode");
100 const unsigned int num_elems_processed_per_iteration = 16;
102 // Configure kernel window
103 Window win = calculate_max_window(*output->info(), Steps(num_elems_processed_per_iteration));
105 AccessWindowStatic output_access(output->info(), 0, 0, output->info()->dimension(0), output->info()->dimension(1));
107 update_window_and_padding(win,
108 AccessWindowRectangle(input->info(), 0, 0, num_elems_processed_per_iteration, 1),
109 AccessWindowRectangle(map_x->info(), 0, 0, num_elems_processed_per_iteration, 1),
110 AccessWindowRectangle(map_y->info(), 0, 0, num_elems_processed_per_iteration, 1),
113 output_access.set_valid_region(win, ValidRegion(Coordinates(0, 0), output->info()->tensor_shape()));
115 INEKernel::configure(win);
118 void NERemapKernel::remap_nearest(const Window &window)
120 // Don't increment in X and Y direction for the input tensor
121 // A pointer to the start of this plane is needed as base for the precomputed offsets
122 Window win_in(window);
123 win_in.set(Window::DimX, Window::Dimension(0, 0, 0));
124 win_in.set(Window::DimY, Window::Dimension(0, 0, 0));
126 Iterator in(_input, win_in);
127 Iterator out(_output, window);
128 Iterator mapx(_map_x, window);
129 Iterator mapy(_map_y, window);
131 const float32x4_t width = vdupq_n_f32(static_cast<float>(_input->info()->dimension(0)));
132 const float32x4_t height = vdupq_n_f32(static_cast<float>(_input->info()->dimension(1)));
133 const int32x4_t in_stride = vdupq_n_s32(static_cast<int32_t>(_input->info()->strides_in_bytes()[1]));
135 execute_window_loop(window, [&](const Coordinates & id)
137 const auto mapx_ptr = reinterpret_cast<const float *>(mapx.ptr());
138 const auto mapy_ptr = reinterpret_cast<const float *>(mapy.ptr());
139 const uint8_t *in_ptr = in.ptr();
141 const int32x4_t offset0 = offset_nearest_interpolation(mapx_ptr + 0, mapy_ptr + 0, width, height, in_stride);
142 const int32x4_t offset1 = offset_nearest_interpolation(mapx_ptr + 4, mapy_ptr + 4, width, height, in_stride);
143 const int32x4_t offset2 = offset_nearest_interpolation(mapx_ptr + 8, mapy_ptr + 8, width, height, in_stride);
144 const int32x4_t offset3 = offset_nearest_interpolation(mapx_ptr + 12, mapy_ptr + 12, width, height, in_stride);
146 uint8x8_t tmp0 = vdup_n_u8(0);
147 tmp0 = vset_lane_u8(in_ptr[vgetq_lane_s32(offset0, 0)], tmp0, 0);
148 tmp0 = vset_lane_u8(in_ptr[vgetq_lane_s32(offset0, 1)], tmp0, 1);
149 tmp0 = vset_lane_u8(in_ptr[vgetq_lane_s32(offset0, 2)], tmp0, 2);
150 tmp0 = vset_lane_u8(in_ptr[vgetq_lane_s32(offset0, 3)], tmp0, 3);
151 tmp0 = vset_lane_u8(in_ptr[vgetq_lane_s32(offset1, 0)], tmp0, 4);
152 tmp0 = vset_lane_u8(in_ptr[vgetq_lane_s32(offset1, 1)], tmp0, 5);
153 tmp0 = vset_lane_u8(in_ptr[vgetq_lane_s32(offset1, 2)], tmp0, 6);
154 tmp0 = vset_lane_u8(in_ptr[vgetq_lane_s32(offset1, 3)], tmp0, 7);
156 uint8x8_t tmp1 = vdup_n_u8(0);
157 tmp1 = vset_lane_u8(in_ptr[vgetq_lane_s32(offset2, 0)], tmp1, 0);
158 tmp1 = vset_lane_u8(in_ptr[vgetq_lane_s32(offset2, 1)], tmp1, 1);
159 tmp1 = vset_lane_u8(in_ptr[vgetq_lane_s32(offset2, 2)], tmp1, 2);
160 tmp1 = vset_lane_u8(in_ptr[vgetq_lane_s32(offset2, 3)], tmp1, 3);
161 tmp1 = vset_lane_u8(in_ptr[vgetq_lane_s32(offset3, 0)], tmp1, 4);
162 tmp1 = vset_lane_u8(in_ptr[vgetq_lane_s32(offset3, 1)], tmp1, 5);
163 tmp1 = vset_lane_u8(in_ptr[vgetq_lane_s32(offset3, 2)], tmp1, 6);
164 tmp1 = vset_lane_u8(in_ptr[vgetq_lane_s32(offset3, 3)], tmp1, 7);
166 vst1q_u8(out.ptr(), vcombine_u8(tmp0, tmp1));
168 in, out, mapx, mapy);
171 void NERemapKernel::remap_bilinear(const Window &window)
173 // Don't increment in X and Y direction for the input tensor
174 // A pointer to the start of this plane is needed as base for the precomputed offsets
175 Window win_in(window);
176 win_in.set(Window::DimX, Window::Dimension(0, 0, 0));
177 win_in.set(Window::DimY, Window::Dimension(0, 0, 0));
179 Iterator in(_input, win_in);
180 Iterator out(_output, window);
181 Iterator mapx(_map_x, window);
182 Iterator mapy(_map_y, window);
184 const size_t width = _input->info()->dimension(0);
185 const size_t height = _input->info()->dimension(1);
186 const size_t in_stride = _input->info()->strides_in_bytes()[1];
188 execute_window_loop(window, [&](const Coordinates & id)
190 const auto mapx_ptr = reinterpret_cast<float *>(mapx.ptr());
191 const auto mapy_ptr = reinterpret_cast<float *>(mapy.ptr());
192 const uint8_t *in_ptr = in.ptr();
194 uint8x8_t tmp0 = vdup_n_u8(0);
195 tmp0 = vset_lane_u8(pixel_bilinear_c1u8_clamp(in_ptr, in_stride, width, height, mapx_ptr[0], mapy_ptr[0]), tmp0, 0);
196 tmp0 = vset_lane_u8(pixel_bilinear_c1u8_clamp(in_ptr, in_stride, width, height, mapx_ptr[1], mapy_ptr[1]), tmp0, 1);
197 tmp0 = vset_lane_u8(pixel_bilinear_c1u8_clamp(in_ptr, in_stride, width, height, mapx_ptr[2], mapy_ptr[2]), tmp0, 2);
198 tmp0 = vset_lane_u8(pixel_bilinear_c1u8_clamp(in_ptr, in_stride, width, height, mapx_ptr[3], mapy_ptr[3]), tmp0, 3);
199 tmp0 = vset_lane_u8(pixel_bilinear_c1u8_clamp(in_ptr, in_stride, width, height, mapx_ptr[4], mapy_ptr[4]), tmp0, 4);
200 tmp0 = vset_lane_u8(pixel_bilinear_c1u8_clamp(in_ptr, in_stride, width, height, mapx_ptr[5], mapy_ptr[5]), tmp0, 5);
201 tmp0 = vset_lane_u8(pixel_bilinear_c1u8_clamp(in_ptr, in_stride, width, height, mapx_ptr[6], mapy_ptr[6]), tmp0, 6);
202 tmp0 = vset_lane_u8(pixel_bilinear_c1u8_clamp(in_ptr, in_stride, width, height, mapx_ptr[7], mapy_ptr[7]), tmp0, 7);
204 uint8x8_t tmp1 = vdup_n_u8(0);
205 tmp1 = vset_lane_u8(pixel_bilinear_c1u8_clamp(in_ptr, in_stride, width, height, mapx_ptr[8], mapy_ptr[8]), tmp1, 0);
206 tmp1 = vset_lane_u8(pixel_bilinear_c1u8_clamp(in_ptr, in_stride, width, height, mapx_ptr[9], mapy_ptr[9]), tmp1, 1);
207 tmp1 = vset_lane_u8(pixel_bilinear_c1u8_clamp(in_ptr, in_stride, width, height, mapx_ptr[10], mapy_ptr[10]), tmp1, 2);
208 tmp1 = vset_lane_u8(pixel_bilinear_c1u8_clamp(in_ptr, in_stride, width, height, mapx_ptr[11], mapy_ptr[11]), tmp1, 3);
209 tmp1 = vset_lane_u8(pixel_bilinear_c1u8_clamp(in_ptr, in_stride, width, height, mapx_ptr[12], mapy_ptr[12]), tmp1, 4);
210 tmp1 = vset_lane_u8(pixel_bilinear_c1u8_clamp(in_ptr, in_stride, width, height, mapx_ptr[13], mapy_ptr[13]), tmp1, 5);
211 tmp1 = vset_lane_u8(pixel_bilinear_c1u8_clamp(in_ptr, in_stride, width, height, mapx_ptr[14], mapy_ptr[14]), tmp1, 6);
212 tmp1 = vset_lane_u8(pixel_bilinear_c1u8_clamp(in_ptr, in_stride, width, height, mapx_ptr[15], mapy_ptr[15]), tmp1, 7);
214 vst1q_u8(out.ptr(), vcombine_u8(tmp0, tmp1));
216 in, out, mapx, mapy);
219 void NERemapKernel::run(const Window &window)
221 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
222 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
223 ARM_COMPUTE_ERROR_ON(_func == nullptr);
225 (this->*_func)(window);