Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2016, 2017 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 "arm_compute/core/NEON/kernels/NEScaleKernel.h" |
| 25 | |
| 26 | #include "arm_compute/core/AccessWindowStatic.h" |
| 27 | #include "arm_compute/core/Coordinates.h" |
| 28 | #include "arm_compute/core/Error.h" |
| 29 | #include "arm_compute/core/Helpers.h" |
| 30 | #include "arm_compute/core/ITensor.h" |
| 31 | #include "arm_compute/core/TensorInfo.h" |
| 32 | #include "arm_compute/core/Validate.h" |
| 33 | #include "arm_compute/core/Window.h" |
| 34 | |
| 35 | #include <arm_neon.h> |
| 36 | #include <cstddef> |
| 37 | #include <cstdint> |
| 38 | |
| 39 | using namespace arm_compute; |
| 40 | |
| 41 | NEScaleKernel::NEScaleKernel() |
| 42 | : _func(nullptr), _offsets(nullptr), _dx(nullptr), _dy(nullptr), _input(nullptr), _output(nullptr) |
| 43 | { |
| 44 | } |
| 45 | |
| 46 | BorderSize NEScaleKernel::border_size() const |
| 47 | { |
| 48 | return BorderSize(1); |
| 49 | } |
| 50 | |
| 51 | void NEScaleKernel::configure(const ITensor *input, const ITensor *dx, const ITensor *dy, const ITensor *offsets, ITensor *output, InterpolationPolicy policy, bool border_undefined) |
| 52 | { |
Georgios Pinitas | 583137c | 2017-08-31 18:12:42 +0100 | [diff] [blame] | 53 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8, DataType::S16, DataType::F32); |
| 54 | ARM_COMPUTE_ERROR_ON_NULLPTR(output); |
| 55 | ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); |
| 56 | ARM_COMPUTE_ERROR_ON(output == input); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 57 | |
| 58 | if(policy == InterpolationPolicy::NEAREST_NEIGHBOR) |
| 59 | { |
| 60 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(offsets, 1, DataType::S32); |
| 61 | } |
| 62 | |
| 63 | if(policy == InterpolationPolicy::BILINEAR) |
| 64 | { |
| 65 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(offsets, 1, DataType::S32); |
| 66 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(dx, 1, DataType::F32); |
| 67 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(dy, 1, DataType::F32); |
| 68 | } |
| 69 | |
| 70 | ARM_COMPUTE_ERROR_ON(output->info()->dimension(0) == 0); |
| 71 | ARM_COMPUTE_ERROR_ON(output->info()->dimension(1) == 0); |
| 72 | |
| 73 | for(size_t i = 2; i < Coordinates::num_max_dimensions; ++i) |
| 74 | { |
| 75 | ARM_COMPUTE_ERROR_ON(input->info()->dimension(i) != output->info()->dimension(i)); |
| 76 | } |
| 77 | |
| 78 | _input = input; |
| 79 | _output = output; |
| 80 | _offsets = offsets; |
| 81 | _dx = dx; |
| 82 | _dy = dy; |
| 83 | |
Isabella Gottardi | 1fab09f | 2017-07-13 15:55:57 +0100 | [diff] [blame] | 84 | /* Compute the ratio between source width/height and destination width/height */ |
| 85 | const auto wr = static_cast<float>(input->info()->dimension(0)) / static_cast<float>(output->info()->dimension(0)); |
| 86 | const auto hr = static_cast<float>(input->info()->dimension(1)) / static_cast<float>(output->info()->dimension(1)); |
| 87 | |
| 88 | /* Area interpolation behaves as Nearest Neighbour in case of up-sampling */ |
| 89 | if(policy == InterpolationPolicy::AREA && wr <= 1.f && hr <= 1.f) |
| 90 | { |
| 91 | policy = InterpolationPolicy::NEAREST_NEIGHBOR; |
| 92 | } |
| 93 | |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 94 | switch(policy) |
| 95 | { |
| 96 | case InterpolationPolicy::NEAREST_NEIGHBOR: |
| 97 | { |
| 98 | _func = &NEScaleKernel::scale_nearest; |
| 99 | break; |
| 100 | } |
| 101 | case InterpolationPolicy::BILINEAR: |
| 102 | { |
| 103 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(_dx, 1, DataType::F32); |
| 104 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(_dy, 1, DataType::F32); |
| 105 | |
| 106 | _func = &NEScaleKernel::scale_bilinear; |
| 107 | break; |
| 108 | } |
| 109 | case InterpolationPolicy::AREA: |
| 110 | { |
| 111 | _func = &NEScaleKernel::scale_area; |
| 112 | break; |
| 113 | } |
| 114 | default: |
| 115 | ARM_COMPUTE_ERROR("Unsupported interpolation mode"); |
| 116 | } |
| 117 | |
| 118 | constexpr unsigned int num_elems_processed_per_iteration = 16; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 119 | |
| 120 | // Configure kernel window |
| 121 | Window win = calculate_max_window(*output->info(), Steps(num_elems_processed_per_iteration)); |
| 122 | |
Isabella Gottardi | 1fab09f | 2017-07-13 15:55:57 +0100 | [diff] [blame] | 123 | const ValidRegion &input_valid_region = input->info()->valid_region(); |
| 124 | |
| 125 | // Reads can occur within the valid region of the input |
| 126 | AccessWindowStatic input_access(input->info(), |
| 127 | input_valid_region.anchor[0] - border_size().left, input_valid_region.anchor[1] - border_size().top, |
| 128 | input_valid_region.anchor[0] + input_valid_region.shape[0] + border_size().right, |
| 129 | input_valid_region.anchor[1] + input_valid_region.shape[1] + border_size().bottom); |
| 130 | AccessWindowHorizontal offsets_access(offsets == nullptr ? nullptr : offsets->info(), 0, num_elems_processed_per_iteration); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 131 | AccessWindowHorizontal dx_access(dx == nullptr ? nullptr : dx->info(), 0, num_elems_processed_per_iteration); |
| 132 | AccessWindowHorizontal dy_access(dy == nullptr ? nullptr : dy->info(), 0, num_elems_processed_per_iteration); |
| 133 | AccessWindowHorizontal output_access(output->info(), 0, num_elems_processed_per_iteration); |
| 134 | |
| 135 | update_window_and_padding(win, |
| 136 | input_access, |
| 137 | offsets_access, |
| 138 | dx_access, |
| 139 | dy_access, |
| 140 | output_access); |
| 141 | |
Isabella Gottardi | 1fab09f | 2017-07-13 15:55:57 +0100 | [diff] [blame] | 142 | output_access.set_valid_region(win, calculate_valid_region_scale(*(input->info()), output->info()->tensor_shape(), policy, border_size(), border_undefined)); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 143 | INEKernel::configure(win); |
| 144 | } |
| 145 | |
| 146 | void NEScaleKernel::scale_nearest(const Window &window) |
| 147 | { |
| 148 | const size_t input_stride = _input->info()->strides_in_bytes()[1]; |
| 149 | |
| 150 | // Compute the ratio between source height and destination height |
| 151 | const auto hr = static_cast<float>(_input->info()->dimension(1)) / static_cast<float>(_output->info()->dimension(1)); |
| 152 | |
| 153 | // Don't increment in X and Y direction for the input tensor |
| 154 | // A pointer to the start of this plane is needed as base for the precomputed offsets |
| 155 | Window win_in(window); |
| 156 | win_in.set(Window::DimX, Window::Dimension(0, 0, 0)); |
| 157 | win_in.set(Window::DimY, Window::Dimension(0, 0, 0)); |
| 158 | |
| 159 | Window win_off; |
| 160 | win_off.set(Window::DimX, window[Window::DimX]); |
| 161 | win_off.set(Window::DimY, window[Window::DimY]); |
| 162 | |
| 163 | for(size_t d = Window::DimZ; d < _offsets->info()->num_dimensions(); ++d) |
| 164 | { |
| 165 | win_off.set(d, Window::Dimension(0, 0, 0)); |
| 166 | } |
| 167 | |
| 168 | Iterator in(_input, win_in); |
| 169 | Iterator out(_output, window); |
| 170 | Iterator offsets(_offsets, win_off); |
| 171 | |
| 172 | switch(_input->info()->data_type()) |
| 173 | { |
| 174 | case DataType::U8: |
| 175 | { |
| 176 | uint8x16_t tmp = vdupq_n_u8(0); |
| 177 | |
| 178 | execute_window_loop(window, [&](const Coordinates & id) |
| 179 | { |
| 180 | const auto offsets_ptr = reinterpret_cast<const int32_t *>(offsets.ptr()); |
| 181 | const uint8_t *const in_ptr = in.ptr(); |
| 182 | |
Pablo Tello | f5f34bb | 2017-08-22 13:34:13 +0100 | [diff] [blame] | 183 | const int in_yi = std::floor((id.y() + 0.5f) * hr); |
| 184 | const int in_yi_clamped = std::min(static_cast<int>(_input->info()->dimension(1)), std::max(in_yi, -1)); |
| 185 | ARM_COMPUTE_ERROR_ON(in_yi_clamped < -1 || in_yi_clamped > static_cast<int>(_input->info()->dimension(1))); |
| 186 | const int offset_row = in_yi_clamped * input_stride; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 187 | |
| 188 | tmp = vsetq_lane_u8(in_ptr[offsets_ptr[0] + offset_row], tmp, 0); |
| 189 | tmp = vsetq_lane_u8(in_ptr[offsets_ptr[1] + offset_row], tmp, 1); |
| 190 | tmp = vsetq_lane_u8(in_ptr[offsets_ptr[2] + offset_row], tmp, 2); |
| 191 | tmp = vsetq_lane_u8(in_ptr[offsets_ptr[3] + offset_row], tmp, 3); |
| 192 | tmp = vsetq_lane_u8(in_ptr[offsets_ptr[4] + offset_row], tmp, 4); |
| 193 | tmp = vsetq_lane_u8(in_ptr[offsets_ptr[5] + offset_row], tmp, 5); |
| 194 | tmp = vsetq_lane_u8(in_ptr[offsets_ptr[6] + offset_row], tmp, 6); |
| 195 | tmp = vsetq_lane_u8(in_ptr[offsets_ptr[7] + offset_row], tmp, 7); |
| 196 | tmp = vsetq_lane_u8(in_ptr[offsets_ptr[8] + offset_row], tmp, 8); |
| 197 | tmp = vsetq_lane_u8(in_ptr[offsets_ptr[9] + offset_row], tmp, 9); |
| 198 | tmp = vsetq_lane_u8(in_ptr[offsets_ptr[10] + offset_row], tmp, 10); |
| 199 | tmp = vsetq_lane_u8(in_ptr[offsets_ptr[11] + offset_row], tmp, 11); |
| 200 | tmp = vsetq_lane_u8(in_ptr[offsets_ptr[12] + offset_row], tmp, 12); |
| 201 | tmp = vsetq_lane_u8(in_ptr[offsets_ptr[13] + offset_row], tmp, 13); |
| 202 | tmp = vsetq_lane_u8(in_ptr[offsets_ptr[14] + offset_row], tmp, 14); |
| 203 | tmp = vsetq_lane_u8(in_ptr[offsets_ptr[15] + offset_row], tmp, 15); |
| 204 | |
| 205 | vst1q_u8(out.ptr(), tmp); |
| 206 | }, |
| 207 | in, offsets, out); |
| 208 | break; |
| 209 | } |
| 210 | case DataType::S16: |
| 211 | { |
| 212 | int16x8x2_t tmp = |
| 213 | { |
| 214 | { |
| 215 | vdupq_n_s16(0), |
| 216 | vdupq_n_s16(0) |
| 217 | } |
| 218 | }; |
| 219 | |
| 220 | execute_window_loop(window, [&](const Coordinates & id) |
| 221 | { |
| 222 | const auto offsets_ptr = reinterpret_cast<const int32_t *>(offsets.ptr()); |
| 223 | |
Isabella Gottardi | 1fab09f | 2017-07-13 15:55:57 +0100 | [diff] [blame] | 224 | const int in_yi = (id.y() + 0.5f) * hr; |
| 225 | const int offset_row = in_yi * input_stride; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 226 | |
| 227 | tmp.val[0] = vsetq_lane_s16(*reinterpret_cast<const int16_t *>(in.ptr() + offsets_ptr[0] + offset_row), tmp.val[0], 0); |
| 228 | tmp.val[0] = vsetq_lane_s16(*reinterpret_cast<const int16_t *>(in.ptr() + offsets_ptr[2] + offset_row), tmp.val[0], 1); |
| 229 | tmp.val[0] = vsetq_lane_s16(*reinterpret_cast<const int16_t *>(in.ptr() + offsets_ptr[4] + offset_row), tmp.val[0], 2); |
| 230 | tmp.val[0] = vsetq_lane_s16(*reinterpret_cast<const int16_t *>(in.ptr() + offsets_ptr[6] + offset_row), tmp.val[0], 3); |
| 231 | tmp.val[0] = vsetq_lane_s16(*reinterpret_cast<const int16_t *>(in.ptr() + offsets_ptr[8] + offset_row), tmp.val[0], 4); |
| 232 | tmp.val[0] = vsetq_lane_s16(*reinterpret_cast<const int16_t *>(in.ptr() + offsets_ptr[10] + offset_row), tmp.val[0], 5); |
| 233 | tmp.val[0] = vsetq_lane_s16(*reinterpret_cast<const int16_t *>(in.ptr() + offsets_ptr[12] + offset_row), tmp.val[0], 6); |
| 234 | tmp.val[0] = vsetq_lane_s16(*reinterpret_cast<const int16_t *>(in.ptr() + offsets_ptr[14] + offset_row), tmp.val[0], 7); |
| 235 | |
| 236 | tmp.val[1] = vsetq_lane_s16(*reinterpret_cast<const int16_t *>(in.ptr() + offsets_ptr[1] + offset_row), tmp.val[1], 0); |
| 237 | tmp.val[1] = vsetq_lane_s16(*reinterpret_cast<const int16_t *>(in.ptr() + offsets_ptr[3] + offset_row), tmp.val[1], 1); |
| 238 | tmp.val[1] = vsetq_lane_s16(*reinterpret_cast<const int16_t *>(in.ptr() + offsets_ptr[5] + offset_row), tmp.val[1], 2); |
| 239 | tmp.val[1] = vsetq_lane_s16(*reinterpret_cast<const int16_t *>(in.ptr() + offsets_ptr[7] + offset_row), tmp.val[1], 3); |
| 240 | tmp.val[1] = vsetq_lane_s16(*reinterpret_cast<const int16_t *>(in.ptr() + offsets_ptr[9] + offset_row), tmp.val[1], 4); |
| 241 | tmp.val[1] = vsetq_lane_s16(*reinterpret_cast<const int16_t *>(in.ptr() + offsets_ptr[11] + offset_row), tmp.val[1], 5); |
| 242 | tmp.val[1] = vsetq_lane_s16(*reinterpret_cast<const int16_t *>(in.ptr() + offsets_ptr[13] + offset_row), tmp.val[1], 6); |
| 243 | tmp.val[1] = vsetq_lane_s16(*reinterpret_cast<const int16_t *>(in.ptr() + offsets_ptr[15] + offset_row), tmp.val[1], 7); |
| 244 | |
| 245 | vst2q_s16(reinterpret_cast<int16_t *>(out.ptr()), tmp); |
| 246 | }, |
| 247 | in, offsets, out); |
| 248 | break; |
| 249 | } |
Georgios Pinitas | 583137c | 2017-08-31 18:12:42 +0100 | [diff] [blame] | 250 | case DataType::F32: |
| 251 | { |
| 252 | float32x4x4_t tmp = |
| 253 | { |
| 254 | { |
| 255 | vdupq_n_f32(0), |
| 256 | vdupq_n_f32(0), |
| 257 | vdupq_n_f32(0), |
| 258 | vdupq_n_f32(0) |
| 259 | } |
| 260 | }; |
| 261 | |
| 262 | execute_window_loop(window, [&](const Coordinates & id) |
| 263 | { |
| 264 | const auto offsets_ptr = reinterpret_cast<const int32_t *>(offsets.ptr()); |
| 265 | |
| 266 | const int in_yi = (id.y() + 0.5f) * hr; |
| 267 | const int offset_row = in_yi * input_stride; |
| 268 | |
| 269 | tmp.val[0] = vsetq_lane_f32(*reinterpret_cast<const float *>(in.ptr() + offsets_ptr[0] + offset_row), tmp.val[0], 0); |
| 270 | tmp.val[0] = vsetq_lane_f32(*reinterpret_cast<const float *>(in.ptr() + offsets_ptr[4] + offset_row), tmp.val[0], 1); |
| 271 | tmp.val[0] = vsetq_lane_f32(*reinterpret_cast<const float *>(in.ptr() + offsets_ptr[8] + offset_row), tmp.val[0], 2); |
| 272 | tmp.val[0] = vsetq_lane_f32(*reinterpret_cast<const float *>(in.ptr() + offsets_ptr[12] + offset_row), tmp.val[0], 3); |
| 273 | |
| 274 | tmp.val[1] = vsetq_lane_f32(*reinterpret_cast<const float *>(in.ptr() + offsets_ptr[1] + offset_row), tmp.val[1], 0); |
| 275 | tmp.val[1] = vsetq_lane_f32(*reinterpret_cast<const float *>(in.ptr() + offsets_ptr[5] + offset_row), tmp.val[1], 1); |
| 276 | tmp.val[1] = vsetq_lane_f32(*reinterpret_cast<const float *>(in.ptr() + offsets_ptr[9] + offset_row), tmp.val[1], 2); |
| 277 | tmp.val[1] = vsetq_lane_f32(*reinterpret_cast<const float *>(in.ptr() + offsets_ptr[13] + offset_row), tmp.val[1], 3); |
| 278 | |
| 279 | tmp.val[2] = vsetq_lane_f32(*reinterpret_cast<const float *>(in.ptr() + offsets_ptr[2] + offset_row), tmp.val[2], 0); |
| 280 | tmp.val[2] = vsetq_lane_f32(*reinterpret_cast<const float *>(in.ptr() + offsets_ptr[6] + offset_row), tmp.val[2], 1); |
| 281 | tmp.val[2] = vsetq_lane_f32(*reinterpret_cast<const float *>(in.ptr() + offsets_ptr[10] + offset_row), tmp.val[2], 2); |
| 282 | tmp.val[2] = vsetq_lane_f32(*reinterpret_cast<const float *>(in.ptr() + offsets_ptr[14] + offset_row), tmp.val[2], 3); |
| 283 | |
| 284 | tmp.val[3] = vsetq_lane_f32(*reinterpret_cast<const float *>(in.ptr() + offsets_ptr[3] + offset_row), tmp.val[3], 0); |
| 285 | tmp.val[3] = vsetq_lane_f32(*reinterpret_cast<const float *>(in.ptr() + offsets_ptr[7] + offset_row), tmp.val[3], 1); |
| 286 | tmp.val[3] = vsetq_lane_f32(*reinterpret_cast<const float *>(in.ptr() + offsets_ptr[11] + offset_row), tmp.val[3], 2); |
| 287 | tmp.val[3] = vsetq_lane_f32(*reinterpret_cast<const float *>(in.ptr() + offsets_ptr[15] + offset_row), tmp.val[3], 3); |
| 288 | |
| 289 | vst4q_f32(reinterpret_cast<float *>(out.ptr()), tmp); |
| 290 | }, |
| 291 | in, offsets, out); |
| 292 | break; |
| 293 | } |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 294 | default: |
| 295 | ARM_COMPUTE_ERROR("Not supported"); |
| 296 | break; |
| 297 | } |
| 298 | } |
| 299 | |
| 300 | void NEScaleKernel::scale_bilinear(const Window &window) |
| 301 | { |
Georgios Pinitas | 583137c | 2017-08-31 18:12:42 +0100 | [diff] [blame] | 302 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(_input, 1, DataType::U8, DataType::S16, DataType::F32); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 303 | |
| 304 | // Compute the ratio between source height and destination height |
| 305 | const auto hr = static_cast<float>(_input->info()->dimension(1)) / static_cast<float>(_output->info()->dimension(1)); |
| 306 | |
| 307 | // Don't increment in X and Y direction for the input tensor |
| 308 | // A pointer to the start of this plane is needed as base for the precomputed offsets |
| 309 | Window win_in(window); |
| 310 | win_in.set(Window::DimX, Window::Dimension(0, 0, 0)); |
| 311 | win_in.set(Window::DimY, Window::Dimension(0, 0, 0)); |
| 312 | |
| 313 | Window win_off; |
| 314 | win_off.set(Window::DimX, window.x()); |
| 315 | win_off.set(Window::DimY, window.y()); |
| 316 | |
| 317 | for(size_t d = Window::DimZ; d < _offsets->info()->num_dimensions(); ++d) |
| 318 | { |
| 319 | win_off.set(d, Window::Dimension(0, 0, 0)); |
| 320 | } |
| 321 | |
| 322 | Iterator in(_input, win_in); |
| 323 | Iterator out(_output, window); |
| 324 | Iterator offsets(_offsets, win_off); |
| 325 | Iterator dx(_dx, win_off); |
| 326 | Iterator dy(_dy, win_off); |
| 327 | |
| 328 | /* Input image stride */ |
Georgios Pinitas | 583137c | 2017-08-31 18:12:42 +0100 | [diff] [blame] | 329 | const size_t in_stide_in_bytes = _input->info()->strides_in_bytes()[1]; |
| 330 | const size_t in_stride = in_stide_in_bytes / _input->info()->element_size(); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 331 | |
Georgios Pinitas | 583137c | 2017-08-31 18:12:42 +0100 | [diff] [blame] | 332 | switch(_input->info()->data_type()) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 333 | { |
Georgios Pinitas | 583137c | 2017-08-31 18:12:42 +0100 | [diff] [blame] | 334 | case DataType::U8: |
| 335 | { |
| 336 | execute_window_loop(window, [&](const Coordinates & id) |
| 337 | { |
| 338 | const auto offsets_ptr = reinterpret_cast<const int32_t *>(offsets.ptr()); |
| 339 | const auto dx_ptr = reinterpret_cast<const float *>(dx.ptr()); |
| 340 | const auto dy_ptr = reinterpret_cast<const float *>(dy.ptr()); |
| 341 | const auto in_ptr = reinterpret_cast<const uint8_t *>(in.ptr()); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 342 | |
Georgios Pinitas | 583137c | 2017-08-31 18:12:42 +0100 | [diff] [blame] | 343 | const int in_yi = std::floor((id.y() + 0.5f) * hr - 0.5f); |
| 344 | const int offset_row = in_yi * in_stide_in_bytes; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 345 | |
Georgios Pinitas | 583137c | 2017-08-31 18:12:42 +0100 | [diff] [blame] | 346 | uint8x8_t tmp0 = vdup_n_u8(0); |
| 347 | tmp0 = vset_lane_u8(delta_bilinear_c1(&in_ptr[offsets_ptr[0] + offset_row], in_stride, dx_ptr[0], dy_ptr[0]), tmp0, 0); |
| 348 | tmp0 = vset_lane_u8(delta_bilinear_c1(&in_ptr[offsets_ptr[1] + offset_row], in_stride, dx_ptr[1], dy_ptr[1]), tmp0, 1); |
| 349 | tmp0 = vset_lane_u8(delta_bilinear_c1(&in_ptr[offsets_ptr[2] + offset_row], in_stride, dx_ptr[2], dy_ptr[2]), tmp0, 2); |
| 350 | tmp0 = vset_lane_u8(delta_bilinear_c1(&in_ptr[offsets_ptr[3] + offset_row], in_stride, dx_ptr[3], dy_ptr[3]), tmp0, 3); |
| 351 | tmp0 = vset_lane_u8(delta_bilinear_c1(&in_ptr[offsets_ptr[4] + offset_row], in_stride, dx_ptr[4], dy_ptr[4]), tmp0, 4); |
| 352 | tmp0 = vset_lane_u8(delta_bilinear_c1(&in_ptr[offsets_ptr[5] + offset_row], in_stride, dx_ptr[5], dy_ptr[5]), tmp0, 5); |
| 353 | tmp0 = vset_lane_u8(delta_bilinear_c1(&in_ptr[offsets_ptr[6] + offset_row], in_stride, dx_ptr[6], dy_ptr[6]), tmp0, 6); |
| 354 | tmp0 = vset_lane_u8(delta_bilinear_c1(&in_ptr[offsets_ptr[7] + offset_row], in_stride, dx_ptr[7], dy_ptr[7]), tmp0, 7); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 355 | |
Georgios Pinitas | 583137c | 2017-08-31 18:12:42 +0100 | [diff] [blame] | 356 | uint8x8_t tmp1 = vdup_n_u8(0); |
| 357 | tmp1 = vset_lane_u8(delta_bilinear_c1(&in_ptr[offsets_ptr[8] + offset_row], in_stride, dx_ptr[8], dy_ptr[8]), tmp1, 0); |
| 358 | tmp1 = vset_lane_u8(delta_bilinear_c1(&in_ptr[offsets_ptr[9] + offset_row], in_stride, dx_ptr[9], dy_ptr[9]), tmp1, 1); |
| 359 | tmp1 = vset_lane_u8(delta_bilinear_c1(&in_ptr[offsets_ptr[10] + offset_row], in_stride, dx_ptr[10], dy_ptr[10]), tmp1, 2); |
| 360 | tmp1 = vset_lane_u8(delta_bilinear_c1(&in_ptr[offsets_ptr[11] + offset_row], in_stride, dx_ptr[11], dy_ptr[11]), tmp1, 3); |
| 361 | tmp1 = vset_lane_u8(delta_bilinear_c1(&in_ptr[offsets_ptr[12] + offset_row], in_stride, dx_ptr[12], dy_ptr[12]), tmp1, 4); |
| 362 | tmp1 = vset_lane_u8(delta_bilinear_c1(&in_ptr[offsets_ptr[13] + offset_row], in_stride, dx_ptr[13], dy_ptr[13]), tmp1, 5); |
| 363 | tmp1 = vset_lane_u8(delta_bilinear_c1(&in_ptr[offsets_ptr[14] + offset_row], in_stride, dx_ptr[14], dy_ptr[14]), tmp1, 6); |
| 364 | tmp1 = vset_lane_u8(delta_bilinear_c1(&in_ptr[offsets_ptr[15] + offset_row], in_stride, dx_ptr[15], dy_ptr[15]), tmp1, 7); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 365 | |
Georgios Pinitas | 583137c | 2017-08-31 18:12:42 +0100 | [diff] [blame] | 366 | vst1q_u8(out.ptr(), vcombine_u8(tmp0, tmp1)); |
| 367 | }, |
| 368 | in, offsets, dx, dy, out); |
| 369 | break; |
| 370 | } |
| 371 | case DataType::S16: |
| 372 | { |
| 373 | execute_window_loop(window, [&](const Coordinates & id) |
| 374 | { |
| 375 | const auto offsets_ptr = reinterpret_cast<const int32_t *>(offsets.ptr()); |
| 376 | const auto dx_ptr = reinterpret_cast<const float *>(dx.ptr()); |
| 377 | const auto dy_ptr = reinterpret_cast<const float *>(dy.ptr()); |
| 378 | |
| 379 | const int in_yi = std::floor((id.y() + 0.5f) * hr - 0.5f); |
| 380 | const int offset_row = in_yi * in_stide_in_bytes; |
| 381 | |
| 382 | int16x8x2_t tmp = |
| 383 | { |
| 384 | { |
| 385 | vdupq_n_s16(0), |
| 386 | vdupq_n_s16(0) |
| 387 | } |
| 388 | }; |
| 389 | |
| 390 | tmp.val[0] = vsetq_lane_s16(delta_bilinear_c1(reinterpret_cast<const int16_t *>(in.ptr() + offsets_ptr[0] + offset_row), in_stride, dx_ptr[0], dy_ptr[0]), tmp.val[0], 0); |
| 391 | tmp.val[0] = vsetq_lane_s16(delta_bilinear_c1(reinterpret_cast<const int16_t *>(in.ptr() + offsets_ptr[2] + offset_row), in_stride, dx_ptr[2], dy_ptr[2]), tmp.val[0], 1); |
| 392 | tmp.val[0] = vsetq_lane_s16(delta_bilinear_c1(reinterpret_cast<const int16_t *>(in.ptr() + offsets_ptr[4] + offset_row), in_stride, dx_ptr[4], dy_ptr[4]), tmp.val[0], 2); |
| 393 | tmp.val[0] = vsetq_lane_s16(delta_bilinear_c1(reinterpret_cast<const int16_t *>(in.ptr() + offsets_ptr[6] + offset_row), in_stride, dx_ptr[6], dy_ptr[6]), tmp.val[0], 3); |
| 394 | tmp.val[0] = vsetq_lane_s16(delta_bilinear_c1(reinterpret_cast<const int16_t *>(in.ptr() + offsets_ptr[8] + offset_row), in_stride, dx_ptr[8], dy_ptr[8]), tmp.val[0], 4); |
| 395 | tmp.val[0] = vsetq_lane_s16(delta_bilinear_c1(reinterpret_cast<const int16_t *>(in.ptr() + offsets_ptr[10] + offset_row), in_stride, dx_ptr[10], dy_ptr[10]), tmp.val[0], 5); |
| 396 | tmp.val[0] = vsetq_lane_s16(delta_bilinear_c1(reinterpret_cast<const int16_t *>(in.ptr() + offsets_ptr[12] + offset_row), in_stride, dx_ptr[12], dy_ptr[12]), tmp.val[0], 6); |
| 397 | tmp.val[0] = vsetq_lane_s16(delta_bilinear_c1(reinterpret_cast<const int16_t *>(in.ptr() + offsets_ptr[14] + offset_row), in_stride, dx_ptr[14], dy_ptr[14]), tmp.val[0], 7); |
| 398 | |
| 399 | tmp.val[1] = vsetq_lane_s16(delta_bilinear_c1(reinterpret_cast<const int16_t *>(in.ptr() + offsets_ptr[1] + offset_row), in_stride, dx_ptr[1], dy_ptr[1]), tmp.val[1], 0); |
| 400 | tmp.val[1] = vsetq_lane_s16(delta_bilinear_c1(reinterpret_cast<const int16_t *>(in.ptr() + offsets_ptr[3] + offset_row), in_stride, dx_ptr[3], dy_ptr[3]), tmp.val[1], 1); |
| 401 | tmp.val[1] = vsetq_lane_s16(delta_bilinear_c1(reinterpret_cast<const int16_t *>(in.ptr() + offsets_ptr[5] + offset_row), in_stride, dx_ptr[5], dy_ptr[5]), tmp.val[1], 2); |
| 402 | tmp.val[1] = vsetq_lane_s16(delta_bilinear_c1(reinterpret_cast<const int16_t *>(in.ptr() + offsets_ptr[7] + offset_row), in_stride, dx_ptr[7], dy_ptr[7]), tmp.val[1], 3); |
| 403 | tmp.val[1] = vsetq_lane_s16(delta_bilinear_c1(reinterpret_cast<const int16_t *>(in.ptr() + offsets_ptr[9] + offset_row), in_stride, dx_ptr[9], dy_ptr[9]), tmp.val[1], 4); |
| 404 | tmp.val[1] = vsetq_lane_s16(delta_bilinear_c1(reinterpret_cast<const int16_t *>(in.ptr() + offsets_ptr[11] + offset_row), in_stride, dx_ptr[11], dy_ptr[11]), tmp.val[1], 5); |
| 405 | tmp.val[1] = vsetq_lane_s16(delta_bilinear_c1(reinterpret_cast<const int16_t *>(in.ptr() + offsets_ptr[13] + offset_row), in_stride, dx_ptr[13], dy_ptr[13]), tmp.val[1], 6); |
| 406 | tmp.val[1] = vsetq_lane_s16(delta_bilinear_c1(reinterpret_cast<const int16_t *>(in.ptr() + offsets_ptr[15] + offset_row), in_stride, dx_ptr[15], dy_ptr[15]), tmp.val[1], 7); |
| 407 | |
| 408 | vst2q_s16(reinterpret_cast<int16_t *>(out.ptr()), tmp); |
| 409 | }, |
| 410 | in, offsets, dx, dy, out); |
| 411 | break; |
| 412 | } |
| 413 | case DataType::F32: |
| 414 | { |
| 415 | execute_window_loop(window, [&](const Coordinates & id) |
| 416 | { |
| 417 | const auto offsets_ptr = reinterpret_cast<const int32_t *>(offsets.ptr()); |
| 418 | const auto dx_ptr = reinterpret_cast<const float *>(dx.ptr()); |
| 419 | const auto dy_ptr = reinterpret_cast<const float *>(dy.ptr()); |
| 420 | |
| 421 | const int in_yi = std::floor((id.y() + 0.5f) * hr - 0.5f); |
| 422 | const int offset_row = in_yi * in_stide_in_bytes; |
| 423 | |
| 424 | float32x4x4_t tmp = |
| 425 | { |
| 426 | { |
| 427 | vdupq_n_f32(0), |
| 428 | vdupq_n_f32(0), |
| 429 | vdupq_n_f32(0), |
| 430 | vdupq_n_f32(0) |
| 431 | } |
| 432 | }; |
| 433 | |
| 434 | tmp.val[0] = vsetq_lane_f32(delta_bilinear_c1(reinterpret_cast<const float *>(in.ptr() + offsets_ptr[0] + offset_row), in_stride, dx_ptr[0], dy_ptr[0]), tmp.val[0], 0); |
| 435 | tmp.val[0] = vsetq_lane_f32(delta_bilinear_c1(reinterpret_cast<const float *>(in.ptr() + offsets_ptr[4] + offset_row), in_stride, dx_ptr[4], dy_ptr[4]), tmp.val[0], 1); |
| 436 | tmp.val[0] = vsetq_lane_f32(delta_bilinear_c1(reinterpret_cast<const float *>(in.ptr() + offsets_ptr[8] + offset_row), in_stride, dx_ptr[8], dy_ptr[8]), tmp.val[0], 2); |
| 437 | tmp.val[0] = vsetq_lane_f32(delta_bilinear_c1(reinterpret_cast<const float *>(in.ptr() + offsets_ptr[12] + offset_row), in_stride, dx_ptr[12], dy_ptr[12]), tmp.val[0], 3); |
| 438 | |
| 439 | tmp.val[1] = vsetq_lane_f32(delta_bilinear_c1(reinterpret_cast<const float *>(in.ptr() + offsets_ptr[1] + offset_row), in_stride, dx_ptr[1], dy_ptr[1]), tmp.val[1], 0); |
| 440 | tmp.val[1] = vsetq_lane_f32(delta_bilinear_c1(reinterpret_cast<const float *>(in.ptr() + offsets_ptr[5] + offset_row), in_stride, dx_ptr[5], dy_ptr[5]), tmp.val[1], 1); |
| 441 | tmp.val[1] = vsetq_lane_f32(delta_bilinear_c1(reinterpret_cast<const float *>(in.ptr() + offsets_ptr[9] + offset_row), in_stride, dx_ptr[9], dy_ptr[9]), tmp.val[1], 2); |
| 442 | tmp.val[1] = vsetq_lane_f32(delta_bilinear_c1(reinterpret_cast<const float *>(in.ptr() + offsets_ptr[13] + offset_row), in_stride, dx_ptr[13], dy_ptr[13]), tmp.val[1], 3); |
| 443 | |
| 444 | tmp.val[2] = vsetq_lane_f32(delta_bilinear_c1(reinterpret_cast<const float *>(in.ptr() + offsets_ptr[2] + offset_row), in_stride, dx_ptr[2], dy_ptr[2]), tmp.val[2], 0); |
| 445 | tmp.val[2] = vsetq_lane_f32(delta_bilinear_c1(reinterpret_cast<const float *>(in.ptr() + offsets_ptr[6] + offset_row), in_stride, dx_ptr[6], dy_ptr[6]), tmp.val[2], 1); |
| 446 | tmp.val[2] = vsetq_lane_f32(delta_bilinear_c1(reinterpret_cast<const float *>(in.ptr() + offsets_ptr[10] + offset_row), in_stride, dx_ptr[10], dy_ptr[10]), tmp.val[2], 2); |
| 447 | tmp.val[2] = vsetq_lane_f32(delta_bilinear_c1(reinterpret_cast<const float *>(in.ptr() + offsets_ptr[14] + offset_row), in_stride, dx_ptr[14], dy_ptr[14]), tmp.val[2], 3); |
| 448 | |
| 449 | tmp.val[3] = vsetq_lane_f32(delta_bilinear_c1(reinterpret_cast<const float *>(in.ptr() + offsets_ptr[3] + offset_row), in_stride, dx_ptr[3], dy_ptr[3]), tmp.val[3], 0); |
| 450 | tmp.val[3] = vsetq_lane_f32(delta_bilinear_c1(reinterpret_cast<const float *>(in.ptr() + offsets_ptr[7] + offset_row), in_stride, dx_ptr[7], dy_ptr[7]), tmp.val[3], 1); |
| 451 | tmp.val[3] = vsetq_lane_f32(delta_bilinear_c1(reinterpret_cast<const float *>(in.ptr() + offsets_ptr[11] + offset_row), in_stride, dx_ptr[11], dy_ptr[11]), tmp.val[3], 2); |
| 452 | tmp.val[3] = vsetq_lane_f32(delta_bilinear_c1(reinterpret_cast<const float *>(in.ptr() + offsets_ptr[15] + offset_row), in_stride, dx_ptr[15], dy_ptr[15]), tmp.val[3], 3); |
| 453 | |
| 454 | vst4q_f32(reinterpret_cast<float *>(out.ptr()), tmp); |
| 455 | }, |
| 456 | in, offsets, dx, dy, out); |
| 457 | break; |
| 458 | } |
| 459 | default: |
| 460 | ARM_COMPUTE_ERROR("Not supported"); |
| 461 | break; |
| 462 | } |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 463 | } |
| 464 | |
| 465 | void NEScaleKernel::scale_area(const Window &window) |
| 466 | { |
| 467 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(_input, 1, DataType::U8); |
| 468 | |
| 469 | // Don't increment in X and Y direction for the input tensor |
| 470 | // A pointer to the start of this plane is needed as base for the precomputed offsets |
| 471 | Window win_in(window); |
| 472 | win_in.set(Window::DimX, Window::Dimension(0, 0, 0)); |
| 473 | win_in.set(Window::DimY, Window::Dimension(0, 0, 0)); |
| 474 | |
| 475 | Iterator in(_input, win_in); |
| 476 | Iterator out(_output, window); |
| 477 | |
| 478 | const auto wr = static_cast<float>(_input->info()->dimension(0)) / static_cast<float>(_output->info()->dimension(0)); |
| 479 | const auto hr = static_cast<float>(_input->info()->dimension(1)) / static_cast<float>(_output->info()->dimension(1)); |
| 480 | const auto w = _input->info()->dimension(0); |
| 481 | const auto h = _input->info()->dimension(1); |
| 482 | const size_t in_stride = _input->info()->strides_in_bytes()[1]; |
| 483 | |
| 484 | execute_window_loop(window, [&](const Coordinates & id) |
| 485 | { |
| 486 | const auto in_ptr = reinterpret_cast<const uint8_t *>(in.ptr()); |
| 487 | |
| 488 | uint8x8_t tmp0 = vdup_n_u8(0); |
| 489 | tmp0 = vset_lane_u8(pixel_area_c1u8_clamp(in_ptr, in_stride, w, h, wr, hr, id.x(), id.y()), tmp0, 0); |
| 490 | tmp0 = vset_lane_u8(pixel_area_c1u8_clamp(in_ptr, in_stride, w, h, wr, hr, id.x() + 1, id.y()), tmp0, 1); |
| 491 | tmp0 = vset_lane_u8(pixel_area_c1u8_clamp(in_ptr, in_stride, w, h, wr, hr, id.x() + 2, id.y()), tmp0, 2); |
| 492 | tmp0 = vset_lane_u8(pixel_area_c1u8_clamp(in_ptr, in_stride, w, h, wr, hr, id.x() + 3, id.y()), tmp0, 3); |
| 493 | tmp0 = vset_lane_u8(pixel_area_c1u8_clamp(in_ptr, in_stride, w, h, wr, hr, id.x() + 4, id.y()), tmp0, 4); |
| 494 | tmp0 = vset_lane_u8(pixel_area_c1u8_clamp(in_ptr, in_stride, w, h, wr, hr, id.x() + 5, id.y()), tmp0, 5); |
| 495 | tmp0 = vset_lane_u8(pixel_area_c1u8_clamp(in_ptr, in_stride, w, h, wr, hr, id.x() + 6, id.y()), tmp0, 6); |
| 496 | tmp0 = vset_lane_u8(pixel_area_c1u8_clamp(in_ptr, in_stride, w, h, wr, hr, id.x() + 7, id.y()), tmp0, 7); |
| 497 | |
| 498 | uint8x8_t tmp1 = vdup_n_u8(0); |
| 499 | tmp1 = vset_lane_u8(pixel_area_c1u8_clamp(in_ptr, in_stride, w, h, wr, hr, id.x() + 8, id.y()), tmp1, 0); |
| 500 | tmp1 = vset_lane_u8(pixel_area_c1u8_clamp(in_ptr, in_stride, w, h, wr, hr, id.x() + 9, id.y()), tmp1, 1); |
| 501 | tmp1 = vset_lane_u8(pixel_area_c1u8_clamp(in_ptr, in_stride, w, h, wr, hr, id.x() + 10, id.y()), tmp1, 2); |
| 502 | tmp1 = vset_lane_u8(pixel_area_c1u8_clamp(in_ptr, in_stride, w, h, wr, hr, id.x() + 11, id.y()), tmp1, 3); |
| 503 | tmp1 = vset_lane_u8(pixel_area_c1u8_clamp(in_ptr, in_stride, w, h, wr, hr, id.x() + 12, id.y()), tmp1, 4); |
| 504 | tmp1 = vset_lane_u8(pixel_area_c1u8_clamp(in_ptr, in_stride, w, h, wr, hr, id.x() + 13, id.y()), tmp1, 5); |
| 505 | tmp1 = vset_lane_u8(pixel_area_c1u8_clamp(in_ptr, in_stride, w, h, wr, hr, id.x() + 14, id.y()), tmp1, 6); |
| 506 | tmp1 = vset_lane_u8(pixel_area_c1u8_clamp(in_ptr, in_stride, w, h, wr, hr, id.x() + 15, id.y()), tmp1, 7); |
| 507 | |
| 508 | vst1q_u8(out.ptr(), vcombine_u8(tmp0, tmp1)); |
| 509 | }, |
| 510 | in, out); |
| 511 | } |
| 512 | |
Moritz Pflanzer | c186b57 | 2017-09-07 09:48:04 +0100 | [diff] [blame] | 513 | void NEScaleKernel::run(const Window &window, const ThreadInfo &info) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 514 | { |
Moritz Pflanzer | c186b57 | 2017-09-07 09:48:04 +0100 | [diff] [blame] | 515 | ARM_COMPUTE_UNUSED(info); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 516 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 517 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window); |
| 518 | ARM_COMPUTE_ERROR_ON(_func == nullptr); |
| 519 | |
| 520 | (this->*_func)(window); |
| 521 | } |