Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1 | /* |
Sang-Hoon Park | 68dd25f | 2020-10-19 16:00:11 +0100 | [diff] [blame] | 2 | * Copyright (c) 2016-2020 Arm Limited. |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 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 | */ |
Michalis Spyrou | ebcebf1 | 2020-10-21 00:04:14 +0100 | [diff] [blame] | 24 | #include "src/core/NEON/kernels/NEConvolutionKernel.h" |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 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/TensorInfo.h" |
| 31 | #include "arm_compute/core/Types.h" |
| 32 | #include "arm_compute/core/Utils.h" |
| 33 | #include "arm_compute/core/Validate.h" |
| 34 | #include "arm_compute/core/Window.h" |
Sang-Hoon Park | 68dd25f | 2020-10-19 16:00:11 +0100 | [diff] [blame] | 35 | #include "src/core/helpers/AutoConfiguration.h" |
| 36 | #include "src/core/helpers/WindowHelpers.h" |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 37 | |
| 38 | #include <algorithm> |
| 39 | #include <arm_neon.h> |
| 40 | #include <array> |
| 41 | #include <cstdint> |
| 42 | #include <cstring> |
| 43 | #include <tuple> |
| 44 | |
| 45 | namespace arm_compute |
| 46 | { |
| 47 | namespace |
| 48 | { |
| 49 | const uint16x8_t max_int16 = vdupq_n_u16(INT16_MAX); |
| 50 | |
| 51 | inline void store_results(const int32x4_t &out, const int32x4_t &out2, int16_t *output) |
| 52 | { |
| 53 | const int16x8_t s16results = vcombine_s16(vqmovn_s32(out), |
| 54 | vqmovn_s32(out2)); |
| 55 | vst1q_s16(output, s16results); |
| 56 | } |
| 57 | |
| 58 | inline void store_results(const int32x4_t &out, const int32x4_t &out2, uint8_t *output) |
| 59 | { |
| 60 | const uint8x8_t u8results = vqmovn_u16(vcombine_u16(vqmovun_s32(out), |
| 61 | vqmovun_s32(out2))); |
| 62 | vst1_u8(output, u8results); |
| 63 | } |
| 64 | |
| 65 | inline void store_results(const uint32x4_t &out, const uint32x4_t &out2, int16_t *output) |
| 66 | { |
| 67 | const uint16x8_t u16results = vcombine_u16(vqmovn_u32(out), vqmovn_u32(out2)); |
| 68 | const int16x8_t s16results = vreinterpretq_s16_u16(vminq_u16(u16results, max_int16)); |
| 69 | vst1q_s16(output, s16results); |
| 70 | } |
| 71 | |
| 72 | inline void store_results(const uint32x4_t &out, const uint32x4_t &out2, uint8_t *output) |
| 73 | { |
| 74 | const uint8x8_t u8results = vqmovn_u16(vcombine_u16(vqmovn_u32(out), |
| 75 | vqmovn_u32(out2))); |
| 76 | vst1_u8(output, u8results); |
| 77 | } |
| 78 | |
| 79 | inline void store_results(const int16x8_t &out, const int16x8_t &out2, int16_t *output) |
| 80 | { |
| 81 | vst1q_s16(output, out); |
| 82 | vst1q_s16(output + 8, out2); |
| 83 | } |
| 84 | |
| 85 | inline void store_results(const int16x8_t &out, const int16x8_t &out2, uint8_t *output) |
| 86 | { |
| 87 | const uint8x16_t u8results = vcombine_u8(vqmovun_s16(out), |
| 88 | vqmovun_s16(out2)); |
| 89 | vst1q_u8(output, u8results); |
| 90 | } |
| 91 | |
| 92 | inline void store_results(const uint16x8_t &out, const uint16x8_t &out2, uint8_t *output) |
| 93 | { |
| 94 | const uint8x16_t u8results = vcombine_u8(vqmovn_u16(out), |
| 95 | vqmovn_u16(out2)); |
| 96 | vst1q_u8(output, u8results); |
| 97 | } |
| 98 | |
| 99 | inline void store_results(const uint16x8_t &out, const uint16x8_t &out2, int16_t *output) |
| 100 | { |
| 101 | vst1q_s16(output, vreinterpretq_s16_u16(vminq_u16(out, max_int16))); |
| 102 | vst1q_s16(output + 8, vreinterpretq_s16_u16(vminq_u16(out2, max_int16))); |
| 103 | } |
| 104 | |
| 105 | inline void convolve_row3x1_unrolled(int32x4_t &out, int32x4_t &out2, const uint8x16_t &row_data, const int16x4_t &mat0, const int16x4_t &mat1, const int16x4_t &mat2) |
| 106 | { |
| 107 | // Convert to s16 and split in blocks of 4 values: |
| 108 | const int16x8_t s16_tmp0 = vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(row_data))); |
| 109 | const int16x8_t s16_tmp1 = vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(row_data))); |
| 110 | |
| 111 | const int16x4x3_t row = |
| 112 | { |
| 113 | { |
| 114 | vget_low_s16(s16_tmp0), |
| 115 | vget_high_s16(s16_tmp0), |
| 116 | vget_low_s16(s16_tmp1) |
| 117 | } |
| 118 | }; |
| 119 | |
| 120 | // Calculate row left value for pixels [0,3] |
| 121 | out = vmlal_s16(out, row.val[0], mat0); |
| 122 | // Calculate row middle value for pixels [0,3] |
| 123 | out = vmlal_s16(out, vext_s16(row.val[0], row.val[1], 1), mat1); |
| 124 | // Calculate row right value for pixels [0,3] |
| 125 | out = vmlal_s16(out, vext_s16(row.val[0], row.val[1], 2), mat2); |
| 126 | |
| 127 | // Calculate row left value for pixels [4,7] |
| 128 | out2 = vmlal_s16(out2, row.val[1], mat0); |
| 129 | // Calculate row middle value for pixels [4,7] |
| 130 | out2 = vmlal_s16(out2, vext_s16(row.val[1], row.val[2], 1), mat1); |
| 131 | // Calculate row right value for pixels [4,7] |
| 132 | out2 = vmlal_s16(out2, vext_s16(row.val[1], row.val[2], 2), mat2); |
| 133 | } |
| 134 | |
| 135 | inline void convolve_row3x1(int32x4_t &out, int32x4_t &out2, const uint8x16_t &row_data, const int16_t *convolution) |
| 136 | { |
| 137 | const int16x4_t mat0 = vld1_dup_s16(convolution); |
| 138 | const int16x4_t mat1 = vld1_dup_s16(convolution + 1); |
| 139 | const int16x4_t mat2 = vld1_dup_s16(convolution + 2); |
| 140 | |
| 141 | convolve_row3x1_unrolled(out, out2, row_data, mat0, mat1, mat2); |
| 142 | } |
| 143 | |
| 144 | inline void convolve_row5x1(int32x4_t &out, int32x4_t &out2, const uint8x16_t &row_data, const int16_t *convolution) |
| 145 | { |
| 146 | const int16x4_t mat0 = vld1_dup_s16(convolution); |
| 147 | const int16x4_t mat1 = vld1_dup_s16(convolution + 1); |
| 148 | const int16x4_t mat2 = vld1_dup_s16(convolution + 2); |
| 149 | const int16x4_t mat3 = vld1_dup_s16(convolution + 3); |
| 150 | const int16x4_t mat4 = vld1_dup_s16(convolution + 4); |
| 151 | |
| 152 | // Convert to s16 and split in blocks of 4 values: |
| 153 | const int16x8_t s16_tmp0 = vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(row_data))); |
| 154 | const int16x8_t s16_tmp1 = vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(row_data))); |
| 155 | |
| 156 | const int16x4x3_t row = |
| 157 | { |
| 158 | { |
| 159 | vget_low_s16(s16_tmp0), |
| 160 | vget_high_s16(s16_tmp0), |
| 161 | vget_low_s16(s16_tmp1) |
| 162 | } |
| 163 | }; |
| 164 | |
| 165 | // Calculate row left 2 value for pixels [0,3] |
| 166 | out = vmlal_s16(out, row.val[0], mat0); |
| 167 | // Calculate row left 1 value for pixels [0,3] |
| 168 | out = vmlal_s16(out, vext_s16(row.val[0], row.val[1], 1), mat1); |
| 169 | // Calculate row middle value for pixels [0,3] |
| 170 | out = vmlal_s16(out, vext_s16(row.val[0], row.val[1], 2), mat2); |
| 171 | // Calculate row right +1 value for pixels [0,3] |
| 172 | out = vmlal_s16(out, vext_s16(row.val[0], row.val[1], 3), mat3); |
| 173 | // Calculate row right +2 value for pixels [0,3] |
| 174 | out = vmlal_s16(out, row.val[1], mat4); |
| 175 | |
| 176 | // Calculate row left 2 value for pixels [4,7] |
| 177 | out2 = vmlal_s16(out2, row.val[1], mat0); |
| 178 | // Calculate row left 1 value for pixels [4,7] |
| 179 | out2 = vmlal_s16(out2, vext_s16(row.val[1], row.val[2], 1), mat1); |
| 180 | // Calculate row middle value for pixels [4,7] |
| 181 | out2 = vmlal_s16(out2, vext_s16(row.val[1], row.val[2], 2), mat2); |
| 182 | // Calculate row right +1 value for pixels [4,7] |
| 183 | out2 = vmlal_s16(out2, vext_s16(row.val[1], row.val[2], 3), mat3); |
| 184 | // Calculate row right +2 value for pixels [4,7] |
| 185 | out2 = vmlal_s16(out2, row.val[2], mat4); |
| 186 | } |
| 187 | |
| 188 | inline void convolve_row7x1(int32x4_t &out, int32x4_t &out2, const uint8x16_t &row_data, const int16_t *convolution) |
| 189 | { |
| 190 | const int16x4_t mat0 = vld1_dup_s16(convolution); |
| 191 | const int16x4_t mat1 = vld1_dup_s16(convolution + 1); |
| 192 | const int16x4_t mat2 = vld1_dup_s16(convolution + 2); |
| 193 | const int16x4_t mat3 = vld1_dup_s16(convolution + 3); |
| 194 | const int16x4_t mat4 = vld1_dup_s16(convolution + 4); |
| 195 | const int16x4_t mat5 = vld1_dup_s16(convolution + 5); |
| 196 | const int16x4_t mat6 = vld1_dup_s16(convolution + 6); |
| 197 | |
| 198 | // Convert to s16 and split in blocks of 4 values: |
| 199 | const int16x8_t s16_tmp0 = vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(row_data))); |
| 200 | const int16x8_t s16_tmp1 = vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(row_data))); |
| 201 | |
| 202 | const int16x4x4_t row = |
| 203 | { |
| 204 | { |
| 205 | vget_low_s16(s16_tmp0), |
| 206 | vget_high_s16(s16_tmp0), |
| 207 | vget_low_s16(s16_tmp1), |
| 208 | vget_high_s16(s16_tmp1) |
| 209 | } |
| 210 | }; |
| 211 | |
| 212 | // Calculate row left 3 value for pixels [0,3] |
| 213 | out = vmlal_s16(out, row.val[0], mat0); |
| 214 | // Calculate row left 2 value for pixels [0,3] |
| 215 | out = vmlal_s16(out, vext_s16(row.val[0], row.val[1], 1), mat1); |
| 216 | // Calculate row left 1 value for pixels [0,3] |
| 217 | out = vmlal_s16(out, vext_s16(row.val[0], row.val[1], 2), mat2); |
| 218 | // Calculate row middle value for pixels [0,3] |
| 219 | out = vmlal_s16(out, vext_s16(row.val[0], row.val[1], 3), mat3); |
| 220 | // Calculate row right +1 value for pixels [0,3] |
| 221 | out = vmlal_s16(out, row.val[1], mat4); |
| 222 | // Calculate row right +2 value for pixels [0,3] |
| 223 | out = vmlal_s16(out, vext_s16(row.val[1], row.val[2], 1), mat5); |
| 224 | // Calculate row right +3 value for pixels [0,3] |
| 225 | out = vmlal_s16(out, vext_s16(row.val[1], row.val[2], 2), mat6); |
| 226 | |
| 227 | // Calculate row left 3 value for pixels [4,7] |
| 228 | out2 = vmlal_s16(out2, row.val[1], mat0); |
| 229 | // Calculate row left 2 value for pixels [4,7] |
| 230 | out2 = vmlal_s16(out2, vext_s16(row.val[1], row.val[2], 1), mat1); |
| 231 | // Calculate row left 1 value for pixels [4,7] |
| 232 | out2 = vmlal_s16(out2, vext_s16(row.val[1], row.val[2], 2), mat2); |
| 233 | // Calculate row middle value for pixels [4,7] |
| 234 | out2 = vmlal_s16(out2, vext_s16(row.val[1], row.val[2], 3), mat3); |
| 235 | // Calculate row right +1 value for pixels [4,7] |
| 236 | out2 = vmlal_s16(out2, row.val[2], mat4); |
| 237 | // Calculate row right +2 value for pixels [4,7] |
| 238 | out2 = vmlal_s16(out2, vext_s16(row.val[2], row.val[3], 1), mat5); |
| 239 | // Calculate row right +3 value for pixels [4,7] |
| 240 | out2 = vmlal_s16(out2, vext_s16(row.val[2], row.val[3], 2), mat6); |
| 241 | } |
| 242 | |
| 243 | inline void convolve_row9x1(int32x4_t &out, int32x4_t &out2, const uint8x16_t &row_data, const int16_t *convolution) |
| 244 | { |
| 245 | const int16x4_t mat0 = vld1_dup_s16(convolution); |
| 246 | const int16x4_t mat1 = vld1_dup_s16(convolution + 1); |
| 247 | const int16x4_t mat2 = vld1_dup_s16(convolution + 2); |
| 248 | const int16x4_t mat3 = vld1_dup_s16(convolution + 3); |
| 249 | const int16x4_t mat4 = vld1_dup_s16(convolution + 4); |
| 250 | const int16x4_t mat5 = vld1_dup_s16(convolution + 5); |
| 251 | const int16x4_t mat6 = vld1_dup_s16(convolution + 6); |
| 252 | const int16x4_t mat7 = vld1_dup_s16(convolution + 7); |
| 253 | const int16x4_t mat8 = vld1_dup_s16(convolution + 8); |
| 254 | |
| 255 | // Convert to s16 and split in blocks of 4 values: |
| 256 | const int16x8_t s16_tmp0 = vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(row_data))); |
| 257 | const int16x8_t s16_tmp1 = vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(row_data))); |
| 258 | |
| 259 | const int16x4x4_t row = |
| 260 | { |
| 261 | { |
| 262 | vget_low_s16(s16_tmp0), |
| 263 | vget_high_s16(s16_tmp0), |
| 264 | vget_low_s16(s16_tmp1), |
| 265 | vget_high_s16(s16_tmp1) |
| 266 | } |
| 267 | }; |
| 268 | |
| 269 | // Calculate row left 4 value for pixels [0,3] |
| 270 | out = vmlal_s16(out, row.val[0], mat0); |
| 271 | // Calculate row left 3 value for pixels [0,3] |
| 272 | out = vmlal_s16(out, vext_s16(row.val[0], row.val[1], 1), mat1); |
| 273 | // Calculate row left 2 value for pixels [0,3] |
| 274 | out = vmlal_s16(out, vext_s16(row.val[0], row.val[1], 2), mat2); |
| 275 | // Calculate row left 1 value for pixels [0,3] |
| 276 | out = vmlal_s16(out, vext_s16(row.val[0], row.val[1], 3), mat3); |
| 277 | // Calculate row middle value for pixels [0,3] |
| 278 | out = vmlal_s16(out, row.val[1], mat4); |
| 279 | // Calculate row right +1 value for pixels [0,3] |
| 280 | out = vmlal_s16(out, vext_s16(row.val[1], row.val[2], 1), mat5); |
| 281 | // Calculate row right +2 value for pixels [0,3] |
| 282 | out = vmlal_s16(out, vext_s16(row.val[1], row.val[2], 2), mat6); |
| 283 | // Calculate row right +3 value for pixels [0,3] |
| 284 | out = vmlal_s16(out, vext_s16(row.val[1], row.val[2], 3), mat7); |
| 285 | // Calculate row right +4 value for pixels [0,3] |
| 286 | out = vmlal_s16(out, row.val[2], mat8); |
| 287 | |
| 288 | // Calculate row left 4 value for pixels [0,3] |
| 289 | out2 = vmlal_s16(out2, row.val[1], mat0); |
| 290 | // Calculate row left 3 value for pixels [0,3] |
| 291 | out2 = vmlal_s16(out2, vext_s16(row.val[1], row.val[2], 1), mat1); |
| 292 | // Calculate row left 2 value for pixels [0,3] |
| 293 | out2 = vmlal_s16(out2, vext_s16(row.val[1], row.val[2], 2), mat2); |
| 294 | // Calculate row left 1 value for pixels [0,3] |
| 295 | out2 = vmlal_s16(out2, vext_s16(row.val[1], row.val[2], 3), mat3); |
| 296 | // Calculate row middle value for pixels [0,3] |
| 297 | out2 = vmlal_s16(out2, row.val[2], mat4); |
| 298 | // Calculate row right +1 value for pixels [0,3] |
| 299 | out2 = vmlal_s16(out2, vext_s16(row.val[2], row.val[3], 1), mat5); |
| 300 | // Calculate row right +2 value for pixels [0,3] |
| 301 | out2 = vmlal_s16(out2, vext_s16(row.val[2], row.val[3], 2), mat6); |
| 302 | // Calculate row right +3 value for pixels [0,3] |
| 303 | out2 = vmlal_s16(out2, vext_s16(row.val[2], row.val[3], 3), mat7); |
| 304 | // Calculate row right +4 value for pixels [0,3] |
| 305 | out2 = vmlal_s16(out2, row.val[3], mat8); |
| 306 | } |
| 307 | } // namespace |
| 308 | |
| 309 | /****************************************************************************************\ |
| 310 | * Square Convolution * |
| 311 | \****************************************************************************************/ |
| 312 | |
| 313 | template <unsigned int matrix_size> |
| 314 | NEConvolutionKernel<matrix_size>::NEConvolutionKernel() |
| 315 | : INESimpleKernel(), _scale(0), _convolution{ {} } |
| 316 | { |
| 317 | } |
| 318 | |
| 319 | template <unsigned int matrix_size> |
| 320 | BorderSize NEConvolutionKernel<matrix_size>::border_size() const |
| 321 | { |
Michalis Spyrou | a4f378d | 2019-04-26 14:54:54 +0100 | [diff] [blame] | 322 | return BorderSize{ matrix_size / 2 }; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 323 | } |
| 324 | |
| 325 | template <unsigned int matrix_size> |
| 326 | void NEConvolutionKernel<matrix_size>::configure(const ITensor *input, ITensor *output, const int16_t *conv, uint32_t scale, bool border_undefined) |
| 327 | { |
| 328 | ARM_COMPUTE_ERROR_ON_NULLPTR(input, output, conv); |
| 329 | |
| 330 | set_shape_if_empty(*output->info(), input->info()->tensor_shape()); |
| 331 | |
| 332 | ARM_COMPUTE_ERROR_ON_MISMATCHING_SHAPES(input, output); |
| 333 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8); |
| 334 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::U8, DataType::S16); |
| 335 | |
| 336 | _input = input; |
| 337 | _output = output; |
| 338 | |
| 339 | std::copy_n(conv, _convolution.size(), _convolution.begin()); |
| 340 | |
| 341 | if(scale == 0) |
| 342 | { |
| 343 | _scale = calculate_matrix_scale(_convolution.data(), matrix_size); |
| 344 | } |
| 345 | else |
| 346 | { |
| 347 | _scale = scale; |
| 348 | } |
| 349 | |
| 350 | // Configure kernel window |
| 351 | constexpr unsigned int num_elems_processed_per_iteration = 8; |
| 352 | constexpr unsigned int num_elems_read_per_iteration = 16; |
| 353 | constexpr unsigned int num_elems_written_per_iteration = 8; |
| 354 | |
| 355 | Window win = calculate_max_window(*input->info(), Steps(num_elems_processed_per_iteration), border_undefined, border_size()); |
| 356 | AccessWindowHorizontal output_access(output->info(), 0, num_elems_written_per_iteration); |
| 357 | |
| 358 | update_window_and_padding(win, |
| 359 | AccessWindowRectangle(input->info(), -border_size().left, -border_size().top, num_elems_read_per_iteration, matrix_size), |
| 360 | output_access); |
| 361 | |
| 362 | output_access.set_valid_region(win, input->info()->valid_region(), border_undefined, border_size()); |
| 363 | |
| 364 | INEKernel::configure(win); |
| 365 | } |
| 366 | |
| 367 | template <> |
| 368 | template <typename OutputType> |
| 369 | void NEConvolutionKernel<3>::convolution(const Window &win) |
| 370 | { |
| 371 | static_assert(sizeof(OutputType) == sizeof(uint8_t) || sizeof(OutputType) == sizeof(int16_t), "The output buffer can only be u8 or s16"); |
| 372 | ARM_COMPUTE_ERROR_ON(_input->buffer() == nullptr); |
| 373 | |
| 374 | Iterator input(_input, win); |
| 375 | Iterator output(_output, win); |
| 376 | |
| 377 | // Load the matrix's coefficients into NEON registers: |
| 378 | const int16x4_t mat00 = vld1_dup_s16(_convolution.data()); |
| 379 | const int16x4_t mat01 = vld1_dup_s16(_convolution.data() + 1); |
| 380 | const int16x4_t mat02 = vld1_dup_s16(_convolution.data() + 2); |
| 381 | const int16x4_t mat10 = vld1_dup_s16(_convolution.data() + 3); |
| 382 | const int16x4_t mat11 = vld1_dup_s16(_convolution.data() + 4); |
| 383 | const int16x4_t mat12 = vld1_dup_s16(_convolution.data() + 5); |
| 384 | const int16x4_t mat20 = vld1_dup_s16(_convolution.data() + 6); |
| 385 | const int16x4_t mat21 = vld1_dup_s16(_convolution.data() + 7); |
| 386 | const int16x4_t mat22 = vld1_dup_s16(_convolution.data() + 8); |
| 387 | const float32x4_t scale_val = vdupq_n_f32(1.0f / _scale); |
| 388 | |
| 389 | const unsigned char *input_top_ptr = _input->buffer() + _input->info()->offset_element_in_bytes(Coordinates(-1, -1)); |
| 390 | const unsigned char *input_mid_ptr = _input->buffer() + _input->info()->offset_element_in_bytes(Coordinates(-1, 0)); |
| 391 | const unsigned char *input_low_ptr = _input->buffer() + _input->info()->offset_element_in_bytes(Coordinates(-1, 1)); |
| 392 | |
Michalis Spyrou | a4f378d | 2019-04-26 14:54:54 +0100 | [diff] [blame] | 393 | execute_window_loop(win, [&](const Coordinates &) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 394 | { |
| 395 | int32x4_t out = vdupq_n_s32(0); |
| 396 | int32x4_t out2 = vdupq_n_s32(0); |
| 397 | |
| 398 | // Load 16 bytes from the top row: |
| 399 | const uint8x16_t top_data = vld1q_u8(input_top_ptr + input.offset()); |
| 400 | convolve_row3x1_unrolled(out, out2, top_data, mat00, mat01, mat02); |
| 401 | |
| 402 | // Load 16 bytes from the middle row: |
| 403 | const uint8x16_t mid_data = vld1q_u8(input_mid_ptr + input.offset()); |
| 404 | convolve_row3x1_unrolled(out, out2, mid_data, mat10, mat11, mat12); |
| 405 | |
| 406 | // Load 16 bytes from the middle row: |
| 407 | const uint8x16_t low_data = vld1q_u8(input_low_ptr + input.offset()); |
| 408 | convolve_row3x1_unrolled(out, out2, low_data, mat20, mat21, mat22); |
| 409 | |
| 410 | // Apply scale |
| 411 | if(_scale != 1) |
| 412 | { |
| 413 | // Convert to F32, scale and convert back to S32 |
| 414 | out = vcvtq_s32_f32(vmulq_f32(vcvtq_f32_s32(out), scale_val)); |
| 415 | out2 = vcvtq_s32_f32(vmulq_f32(vcvtq_f32_s32(out2), scale_val)); |
| 416 | } |
| 417 | |
| 418 | // Clamp and store as U8 or S16: |
| 419 | store_results(out, out2, reinterpret_cast<OutputType *>(output.ptr())); |
| 420 | }, |
| 421 | input, output); |
| 422 | } |
| 423 | |
| 424 | template <> |
| 425 | template <typename OutputType> |
| 426 | void NEConvolutionKernel<5>::convolution(const Window &win) |
| 427 | { |
| 428 | static_assert(sizeof(OutputType) == sizeof(uint8_t) || sizeof(OutputType) == sizeof(int16_t), "The output buffer can only be u8 or s16"); |
| 429 | ARM_COMPUTE_ERROR_ON(_input->buffer() == nullptr); |
| 430 | |
| 431 | Iterator input(_input, win); |
| 432 | Iterator output(_output, win); |
| 433 | |
| 434 | const float32x4_t scale_val = vdupq_n_f32(1.0f / _scale); |
| 435 | |
| 436 | const unsigned char *input_top2_ptr = _input->buffer() + _input->info()->offset_element_in_bytes(Coordinates(-2, -2)); |
| 437 | const unsigned char *input_top1_ptr = _input->buffer() + _input->info()->offset_element_in_bytes(Coordinates(-2, -1)); |
| 438 | const unsigned char *input_mid_ptr = _input->buffer() + _input->info()->offset_element_in_bytes(Coordinates(-2, 0)); |
| 439 | const unsigned char *input_low1_ptr = _input->buffer() + _input->info()->offset_element_in_bytes(Coordinates(-2, 1)); |
| 440 | const unsigned char *input_low2_ptr = _input->buffer() + _input->info()->offset_element_in_bytes(Coordinates(-2, 2)); |
| 441 | |
Michalis Spyrou | a4f378d | 2019-04-26 14:54:54 +0100 | [diff] [blame] | 442 | execute_window_loop(win, [&](const Coordinates &) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 443 | { |
| 444 | int32x4_t out = vdupq_n_s32(0); |
| 445 | int32x4_t out2 = vdupq_n_s32(0); |
| 446 | |
| 447 | // Load 16 bytes from the top2 row: |
| 448 | const uint8x16_t data_t2 = vld1q_u8(input_top2_ptr + input.offset()); |
| 449 | convolve_row5x1(out, out2, data_t2, _convolution.data()); |
| 450 | |
| 451 | // Load 16 bytes from the top1 row: |
| 452 | const uint8x16_t data_t1 = vld1q_u8(input_top1_ptr + input.offset()); |
| 453 | convolve_row5x1(out, out2, data_t1, _convolution.data() + 5); |
| 454 | |
| 455 | // Load 16 bytes from the middle row: |
| 456 | const uint8x16_t data_m = vld1q_u8(input_mid_ptr + input.offset()); |
| 457 | convolve_row5x1(out, out2, data_m, _convolution.data() + 10); |
| 458 | |
| 459 | // Load 16 bytes from the low1 row: |
| 460 | const uint8x16_t data_b1 = vld1q_u8(input_low1_ptr + input.offset()); |
| 461 | convolve_row5x1(out, out2, data_b1, _convolution.data() + 15); |
| 462 | |
| 463 | // Load 16 bytes from the low2 row: |
| 464 | const uint8x16_t data_b2 = vld1q_u8(input_low2_ptr + input.offset()); |
| 465 | convolve_row5x1(out, out2, data_b2, _convolution.data() + 20); |
| 466 | |
| 467 | // Apply scale |
| 468 | if(_scale != 1) |
| 469 | { |
| 470 | // Convert to F32, scale and convert back to S32 |
| 471 | out = vcvtq_s32_f32(vmulq_f32(vcvtq_f32_s32(out), scale_val)); |
| 472 | out2 = vcvtq_s32_f32(vmulq_f32(vcvtq_f32_s32(out2), scale_val)); |
| 473 | } |
| 474 | |
| 475 | // Clamp and store as U8 or S16: |
| 476 | store_results(out, out2, reinterpret_cast<OutputType *>(output.ptr())); |
| 477 | }, |
| 478 | input, output); |
| 479 | } |
| 480 | |
| 481 | template <> |
| 482 | template <typename OutputType> |
| 483 | void NEConvolutionKernel<7>::convolution(const Window &win) |
| 484 | { |
| 485 | static_assert(sizeof(OutputType) == sizeof(uint8_t) || sizeof(OutputType) == sizeof(int16_t), "The output buffer can only be u8 or s16"); |
| 486 | ARM_COMPUTE_ERROR_ON(_input->buffer() == nullptr); |
| 487 | |
| 488 | Iterator input(_input, win); |
| 489 | Iterator output(_output, win); |
| 490 | |
| 491 | const float32x4_t scale_val = vdupq_n_f32(1.0f / _scale); |
| 492 | |
| 493 | const unsigned char *input_top3_ptr = _input->buffer() + _input->info()->offset_element_in_bytes(Coordinates(-3, -3)); |
| 494 | const unsigned char *input_top2_ptr = _input->buffer() + _input->info()->offset_element_in_bytes(Coordinates(-3, -2)); |
| 495 | const unsigned char *input_top1_ptr = _input->buffer() + _input->info()->offset_element_in_bytes(Coordinates(-3, -1)); |
| 496 | const unsigned char *input_mid_ptr = _input->buffer() + _input->info()->offset_element_in_bytes(Coordinates(-3, 0)); |
| 497 | const unsigned char *input_low1_ptr = _input->buffer() + _input->info()->offset_element_in_bytes(Coordinates(-3, 1)); |
| 498 | const unsigned char *input_low2_ptr = _input->buffer() + _input->info()->offset_element_in_bytes(Coordinates(-3, 2)); |
| 499 | const unsigned char *input_low3_ptr = _input->buffer() + _input->info()->offset_element_in_bytes(Coordinates(-3, 3)); |
| 500 | |
Michalis Spyrou | a4f378d | 2019-04-26 14:54:54 +0100 | [diff] [blame] | 501 | execute_window_loop(win, [&](const Coordinates &) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 502 | { |
| 503 | int32x4_t out = vdupq_n_s32(0); |
| 504 | int32x4_t out2 = vdupq_n_s32(0); |
| 505 | |
| 506 | // Load 16 bytes from the top3 row: |
| 507 | const uint8x16_t data_t3 = vld1q_u8(input_top3_ptr + input.offset()); |
| 508 | convolve_row7x1(out, out2, data_t3, _convolution.data()); |
| 509 | |
| 510 | // Load 16 bytes from the top2 row: |
| 511 | const uint8x16_t data_t2 = vld1q_u8(input_top2_ptr + input.offset()); |
| 512 | convolve_row7x1(out, out2, data_t2, _convolution.data() + 7); |
| 513 | |
| 514 | // Load 16 bytes from the top1 row: |
| 515 | const uint8x16_t data_t1 = vld1q_u8(input_top1_ptr + input.offset()); |
| 516 | convolve_row7x1(out, out2, data_t1, _convolution.data() + 14); |
| 517 | |
| 518 | // Load 16 bytes from the middle row: |
| 519 | const uint8x16_t data_m = vld1q_u8(input_mid_ptr + input.offset()); |
| 520 | convolve_row7x1(out, out2, data_m, _convolution.data() + 21); |
| 521 | |
| 522 | // Load 16 bytes from the low1 row: |
| 523 | const uint8x16_t data_b1 = vld1q_u8(input_low1_ptr + input.offset()); |
| 524 | convolve_row7x1(out, out2, data_b1, _convolution.data() + 28); |
| 525 | |
| 526 | // Load 16 bytes from the low2 row: |
| 527 | const uint8x16_t data_b2 = vld1q_u8(input_low2_ptr + input.offset()); |
| 528 | convolve_row7x1(out, out2, data_b2, _convolution.data() + 35); |
| 529 | |
| 530 | // Load 16 bytes from the low3 row: |
| 531 | const uint8x16_t data_b3 = vld1q_u8(input_low3_ptr + input.offset()); |
| 532 | convolve_row7x1(out, out2, data_b3, _convolution.data() + 42); |
| 533 | |
| 534 | // Apply scale |
| 535 | if(_scale != 1) |
| 536 | { |
| 537 | // Convert to F32, scale and convert back to S32 |
| 538 | out = vcvtq_s32_f32(vmulq_f32(vcvtq_f32_s32(out), scale_val)); |
| 539 | out2 = vcvtq_s32_f32(vmulq_f32(vcvtq_f32_s32(out2), scale_val)); |
| 540 | } |
| 541 | |
| 542 | // Clamp and store as U8 or S16: |
| 543 | store_results(out, out2, reinterpret_cast<OutputType *>(output.ptr())); |
| 544 | }, |
| 545 | input, output); |
| 546 | } |
| 547 | |
| 548 | template <> |
| 549 | template <typename OutputType> |
| 550 | void NEConvolutionKernel<9>::convolution(const Window &win) |
| 551 | { |
| 552 | static_assert(sizeof(OutputType) == sizeof(uint8_t) || sizeof(OutputType) == sizeof(int16_t), "The output buffer can only be u8 or s16"); |
| 553 | ARM_COMPUTE_ERROR_ON(_input->buffer() == nullptr); |
| 554 | |
| 555 | Iterator input(_input, win); |
| 556 | Iterator output(_output, win); |
| 557 | |
| 558 | const float32x4_t scale_val = vdupq_n_f32(1.0f / _scale); |
| 559 | |
| 560 | const unsigned char *input_top4_ptr = _input->buffer() + _input->info()->offset_element_in_bytes(Coordinates(-4, -4)); |
| 561 | const unsigned char *input_top3_ptr = _input->buffer() + _input->info()->offset_element_in_bytes(Coordinates(-4, -3)); |
| 562 | const unsigned char *input_top2_ptr = _input->buffer() + _input->info()->offset_element_in_bytes(Coordinates(-4, -2)); |
| 563 | const unsigned char *input_top1_ptr = _input->buffer() + _input->info()->offset_element_in_bytes(Coordinates(-4, -1)); |
| 564 | const unsigned char *input_mid_ptr = _input->buffer() + _input->info()->offset_element_in_bytes(Coordinates(-4, 0)); |
| 565 | const unsigned char *input_low1_ptr = _input->buffer() + _input->info()->offset_element_in_bytes(Coordinates(-4, 1)); |
| 566 | const unsigned char *input_low2_ptr = _input->buffer() + _input->info()->offset_element_in_bytes(Coordinates(-4, 2)); |
| 567 | const unsigned char *input_low3_ptr = _input->buffer() + _input->info()->offset_element_in_bytes(Coordinates(-4, 3)); |
| 568 | const unsigned char *input_low4_ptr = _input->buffer() + _input->info()->offset_element_in_bytes(Coordinates(-4, 4)); |
| 569 | |
Michalis Spyrou | a4f378d | 2019-04-26 14:54:54 +0100 | [diff] [blame] | 570 | execute_window_loop(win, [&](const Coordinates &) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 571 | { |
| 572 | int32x4_t out = vdupq_n_s32(0); |
| 573 | int32x4_t out2 = vdupq_n_s32(0); |
| 574 | |
| 575 | // Load 16 bytes from the top4 row: |
| 576 | const uint8x16_t data_t4 = vld1q_u8(input_top4_ptr + input.offset()); |
| 577 | convolve_row9x1(out, out2, data_t4, _convolution.data()); |
| 578 | |
| 579 | // Load 16 bytes from the top3 row: |
| 580 | const uint8x16_t data_t3 = vld1q_u8(input_top3_ptr + input.offset()); |
| 581 | convolve_row9x1(out, out2, data_t3, _convolution.data() + 9); |
| 582 | |
| 583 | // Load 16 bytes from the top2 row: |
| 584 | const uint8x16_t data_t2 = vld1q_u8(input_top2_ptr + input.offset()); |
| 585 | convolve_row9x1(out, out2, data_t2, _convolution.data() + 18); |
| 586 | |
| 587 | // Load 16 bytes from the top1 row: |
| 588 | const uint8x16_t data_t1 = vld1q_u8(input_top1_ptr + input.offset()); |
| 589 | convolve_row9x1(out, out2, data_t1, _convolution.data() + 27); |
| 590 | |
| 591 | // Load 16 bytes from the middle row: |
| 592 | const uint8x16_t data_m = vld1q_u8(input_mid_ptr + input.offset()); |
| 593 | convolve_row9x1(out, out2, data_m, _convolution.data() + 36); |
| 594 | |
| 595 | // Load 16 bytes from the low1 row: |
| 596 | const uint8x16_t data_b1 = vld1q_u8(input_low1_ptr + input.offset()); |
| 597 | convolve_row9x1(out, out2, data_b1, _convolution.data() + 45); |
| 598 | |
| 599 | // Load 16 bytes from the low2 row: |
| 600 | const uint8x16_t data_b2 = vld1q_u8(input_low2_ptr + input.offset()); |
| 601 | convolve_row9x1(out, out2, data_b2, _convolution.data() + 54); |
| 602 | |
| 603 | // Load 16 bytes from the low3 row: |
| 604 | const uint8x16_t data_b3 = vld1q_u8(input_low3_ptr + input.offset()); |
| 605 | convolve_row9x1(out, out2, data_b3, _convolution.data() + 63); |
| 606 | |
| 607 | // Load 16 bytes from the low4 row: |
| 608 | const uint8x16_t data_b4 = vld1q_u8(input_low4_ptr + input.offset()); |
| 609 | convolve_row9x1(out, out2, data_b4, _convolution.data() + 72); |
| 610 | |
| 611 | // Apply scale |
| 612 | if(_scale != 1) |
| 613 | { |
| 614 | // Convert to F32, scale and convert back to S32 |
| 615 | out = vcvtq_s32_f32(vmulq_f32(vcvtq_f32_s32(out), scale_val)); |
| 616 | out2 = vcvtq_s32_f32(vmulq_f32(vcvtq_f32_s32(out2), scale_val)); |
| 617 | } |
| 618 | |
| 619 | // Clamp and store as U8 or S16: |
| 620 | store_results(out, out2, reinterpret_cast<OutputType *>(output.ptr())); |
| 621 | }, |
| 622 | input, output); |
| 623 | } |
| 624 | |
| 625 | template <unsigned int matrix_size> |
Moritz Pflanzer | c186b57 | 2017-09-07 09:48:04 +0100 | [diff] [blame] | 626 | void NEConvolutionKernel<matrix_size>::run(const Window &window, const ThreadInfo &info) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 627 | { |
Moritz Pflanzer | c186b57 | 2017-09-07 09:48:04 +0100 | [diff] [blame] | 628 | ARM_COMPUTE_UNUSED(info); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 629 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 630 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window); |
| 631 | |
Sanghoon Lee | c8a85ba | 2017-11-29 11:23:14 +0000 | [diff] [blame] | 632 | switch(_output->info()->data_type()) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 633 | { |
Sanghoon Lee | c8a85ba | 2017-11-29 11:23:14 +0000 | [diff] [blame] | 634 | case DataType::U8: |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 635 | convolution<uint8_t>(window); |
| 636 | break; |
Sanghoon Lee | c8a85ba | 2017-11-29 11:23:14 +0000 | [diff] [blame] | 637 | case DataType::S16: |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 638 | convolution<int16_t>(window); |
| 639 | break; |
| 640 | default: |
Sanghoon Lee | c8a85ba | 2017-11-29 11:23:14 +0000 | [diff] [blame] | 641 | ARM_COMPUTE_ERROR("Not supported Data type!"); |
| 642 | break; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 643 | } |
| 644 | } |
| 645 | |
| 646 | template class arm_compute::NEConvolutionKernel<3>; |
| 647 | template class arm_compute::NEConvolutionKernel<5>; |
| 648 | template class arm_compute::NEConvolutionKernel<7>; |
| 649 | template class arm_compute::NEConvolutionKernel<9>; |
| 650 | |
| 651 | /****************************************************************************************\ |
| 652 | * Separable Square Convolution * |
| 653 | \****************************************************************************************/ |
| 654 | |
| 655 | template <unsigned int matrix_size> |
| 656 | NESeparableConvolutionHorKernel<matrix_size>::NESeparableConvolutionHorKernel() |
| 657 | : _conv_row{ { 0 } }, _border_size(0) |
| 658 | { |
| 659 | } |
| 660 | |
| 661 | template <unsigned int matrix_size> |
| 662 | BorderSize NESeparableConvolutionHorKernel<matrix_size>::border_size() const |
| 663 | { |
| 664 | return _border_size; |
| 665 | } |
| 666 | |
| 667 | template <unsigned int matrix_size> |
| 668 | void NESeparableConvolutionHorKernel<matrix_size>::configure(const ITensor *input, ITensor *output, const int16_t *conv_row, bool border_undefined) |
| 669 | { |
| 670 | ARM_COMPUTE_ERROR_ON_NULLPTR(input, output, conv_row); |
| 671 | |
| 672 | set_shape_if_empty(*output->info(), input->info()->tensor_shape()); |
| 673 | |
| 674 | ARM_COMPUTE_ERROR_ON_MISMATCHING_SHAPES(input, output); |
| 675 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8); |
| 676 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::U16, DataType::S16, DataType::S32); |
| 677 | |
| 678 | _input = input; |
| 679 | _output = output; |
| 680 | std::copy_n(conv_row, _conv_row.size(), _conv_row.begin()); |
| 681 | _border_size = BorderSize(border_undefined ? 0 : matrix_size / 2, matrix_size / 2); |
| 682 | |
| 683 | // Configure kernel window |
| 684 | constexpr unsigned int num_elems_processed_per_iteration = 8; |
| 685 | constexpr unsigned int num_elems_read_per_iteration = 16; |
| 686 | constexpr unsigned int num_elems_written_per_iteration = 8; |
| 687 | |
| 688 | Window win = calculate_max_window_horizontal(*input->info(), Steps(num_elems_processed_per_iteration), border_undefined, border_size()); |
| 689 | AccessWindowHorizontal output_access(output->info(), 0, num_elems_written_per_iteration); |
| 690 | |
| 691 | update_window_and_padding(win, |
| 692 | AccessWindowHorizontal(input->info(), -border_size().left, num_elems_read_per_iteration), |
| 693 | output_access); |
| 694 | |
| 695 | output_access.set_valid_region(win, input->info()->valid_region(), border_undefined, border_size()); |
| 696 | |
| 697 | INEKernel::configure(win); |
| 698 | } |
| 699 | |
| 700 | template <unsigned int matrix_size> |
Moritz Pflanzer | c186b57 | 2017-09-07 09:48:04 +0100 | [diff] [blame] | 701 | void NESeparableConvolutionHorKernel<matrix_size>::run(const Window &window, const ThreadInfo &info) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 702 | { |
Moritz Pflanzer | c186b57 | 2017-09-07 09:48:04 +0100 | [diff] [blame] | 703 | ARM_COMPUTE_UNUSED(info); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 704 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 705 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window); |
| 706 | switch(_output->info()->data_type()) |
| 707 | { |
| 708 | case DataType::U16: |
| 709 | convolve<uint16_t>(window); |
| 710 | break; |
| 711 | case DataType::S16: |
| 712 | convolve<int16_t>(window); |
| 713 | break; |
| 714 | case DataType::S32: |
| 715 | convolve<int32_t>(window); |
| 716 | break; |
| 717 | default: |
| 718 | ARM_COMPUTE_ERROR("Unsupported intermediate data type!"); |
| 719 | break; |
| 720 | } |
| 721 | } |
| 722 | |
| 723 | template <> |
| 724 | template <> |
| 725 | inline void NESeparableConvolutionHorKernel<5>::convolve<uint16_t>(const Window &window) |
| 726 | { |
| 727 | Window win_in(window); |
| 728 | win_in.shift(Window::DimX, -2); |
| 729 | |
| 730 | Iterator input(_input, win_in); |
| 731 | Iterator output(_output, window); |
| 732 | |
Michalis Spyrou | a4f378d | 2019-04-26 14:54:54 +0100 | [diff] [blame] | 733 | execute_window_loop(window, [&](const Coordinates &) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 734 | { |
| 735 | const uint8x16_t data = vld1q_u8(input.ptr()); |
| 736 | |
| 737 | const uint16x8x2_t data_u16 = |
| 738 | { |
| 739 | { |
| 740 | vmovl_u8(vget_low_u8(data)), |
| 741 | vmovl_u8(vget_high_u8(data)) |
| 742 | } |
| 743 | }; |
| 744 | |
| 745 | uint16x8_t out = vmulq_n_u16(data_u16.val[0], _conv_row[0]); |
| 746 | out = vmlaq_n_u16(out, vextq_u16(data_u16.val[0], data_u16.val[1], 1), _conv_row[1]); |
| 747 | out = vmlaq_n_u16(out, vextq_u16(data_u16.val[0], data_u16.val[1], 2), _conv_row[2]); |
| 748 | out = vmlaq_n_u16(out, vextq_u16(data_u16.val[0], data_u16.val[1], 3), _conv_row[3]); |
| 749 | out = vmlaq_n_u16(out, vextq_u16(data_u16.val[0], data_u16.val[1], 4), _conv_row[4]); |
| 750 | |
| 751 | vst1q_u16(reinterpret_cast<uint16_t *>(output.ptr()), out); |
| 752 | }, |
| 753 | input, output); |
| 754 | } |
| 755 | |
| 756 | template <> |
| 757 | template <> |
| 758 | inline void NESeparableConvolutionHorKernel<5>::convolve<int16_t>(const Window &window) |
| 759 | { |
| 760 | Window win_in(window); |
| 761 | win_in.shift(Window::DimX, -2); |
| 762 | |
| 763 | Iterator input(_input, win_in); |
| 764 | Iterator output(_output, window); |
| 765 | |
Michalis Spyrou | a4f378d | 2019-04-26 14:54:54 +0100 | [diff] [blame] | 766 | execute_window_loop(window, [&](const Coordinates &) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 767 | { |
| 768 | const uint8x16_t data = vld1q_u8(input.ptr()); |
| 769 | |
| 770 | const int16x8x2_t data_s16 = |
| 771 | { |
| 772 | { |
| 773 | vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(data))), |
| 774 | vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(data))) |
| 775 | } |
| 776 | }; |
| 777 | |
| 778 | int16x8_t out = vmulq_n_s16(data_s16.val[0], _conv_row[0]); |
| 779 | out = vmlaq_n_s16(out, vextq_s16(data_s16.val[0], data_s16.val[1], 1), _conv_row[1]); |
| 780 | out = vmlaq_n_s16(out, vextq_s16(data_s16.val[0], data_s16.val[1], 2), _conv_row[2]); |
| 781 | out = vmlaq_n_s16(out, vextq_s16(data_s16.val[0], data_s16.val[1], 3), _conv_row[3]); |
| 782 | out = vmlaq_n_s16(out, vextq_s16(data_s16.val[0], data_s16.val[1], 4), _conv_row[4]); |
| 783 | |
| 784 | vst1q_s16(reinterpret_cast<int16_t *>(output.ptr()), out); |
| 785 | }, |
| 786 | input, output); |
| 787 | } |
| 788 | |
| 789 | template <> |
| 790 | template <> |
| 791 | void NESeparableConvolutionHorKernel<5>::convolve<int32_t>(const Window &window) |
| 792 | { |
| 793 | Window win_in(window); |
| 794 | win_in.shift(Window::DimX, -2); |
| 795 | |
| 796 | Iterator input(_input, win_in); |
| 797 | Iterator output(_output, window); |
| 798 | |
Michalis Spyrou | a4f378d | 2019-04-26 14:54:54 +0100 | [diff] [blame] | 799 | execute_window_loop(window, [&](const Coordinates &) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 800 | { |
| 801 | const uint8x16_t data = vld1q_u8(input.ptr()); |
| 802 | |
| 803 | const int16x8x2_t data_s16 = |
| 804 | { |
| 805 | { |
| 806 | vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(data))), |
| 807 | vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(data))) |
| 808 | } |
| 809 | }; |
| 810 | |
| 811 | const int16x8_t data_s16_l1 = vextq_s16(data_s16.val[0], data_s16.val[1], 1); |
| 812 | const int16x8_t data_s16_m = vextq_s16(data_s16.val[0], data_s16.val[1], 2); |
| 813 | const int16x8_t data_s16_r1 = vextq_s16(data_s16.val[0], data_s16.val[1], 3); |
| 814 | const int16x8_t data_s16_r2 = vextq_s16(data_s16.val[0], data_s16.val[1], 4); |
| 815 | |
| 816 | int32x4_t out_low = vmull_n_s16(vget_low_s16(data_s16.val[0]), _conv_row[0]); |
| 817 | out_low = vmlal_n_s16(out_low, vget_low_s16(data_s16_l1), _conv_row[1]); |
| 818 | out_low = vmlal_n_s16(out_low, vget_low_s16(data_s16_m), _conv_row[2]); |
| 819 | out_low = vmlal_n_s16(out_low, vget_low_s16(data_s16_r1), _conv_row[3]); |
| 820 | out_low = vmlal_n_s16(out_low, vget_low_s16(data_s16_r2), _conv_row[4]); |
| 821 | |
| 822 | vst1q_s32(reinterpret_cast<int32_t *>(output.ptr()), out_low); |
| 823 | |
| 824 | int32x4_t out_high = vmull_n_s16(vget_high_s16(data_s16.val[0]), _conv_row[0]); |
| 825 | out_high = vmlal_n_s16(out_high, vget_high_s16(data_s16_l1), _conv_row[1]); |
| 826 | out_high = vmlal_n_s16(out_high, vget_high_s16(data_s16_m), _conv_row[2]); |
| 827 | out_high = vmlal_n_s16(out_high, vget_high_s16(data_s16_r1), _conv_row[3]); |
| 828 | out_high = vmlal_n_s16(out_high, vget_high_s16(data_s16_r2), _conv_row[4]); |
| 829 | |
| 830 | vst1q_s32(reinterpret_cast<int32_t *>(output.ptr()) + 4, out_high); |
| 831 | }, |
| 832 | input, output); |
| 833 | } |
| 834 | |
| 835 | template <> |
| 836 | template <> |
| 837 | inline void NESeparableConvolutionHorKernel<7>::convolve<uint16_t>(const Window &window) |
| 838 | { |
| 839 | Window win_in(window); |
| 840 | win_in.shift(Window::DimX, -3); |
| 841 | |
| 842 | Iterator input(_input, win_in); |
| 843 | Iterator output(_output, window); |
| 844 | |
Michalis Spyrou | a4f378d | 2019-04-26 14:54:54 +0100 | [diff] [blame] | 845 | execute_window_loop(window, [&](const Coordinates &) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 846 | { |
| 847 | const uint8x16_t data = vld1q_u8(input.ptr()); |
| 848 | |
| 849 | const uint16x8x2_t data_u16 = |
| 850 | { |
| 851 | { |
| 852 | vmovl_u8(vget_low_u8(data)), |
| 853 | vmovl_u8(vget_high_u8(data)) |
| 854 | } |
| 855 | }; |
| 856 | |
| 857 | uint16x8_t out = vmulq_n_u16(data_u16.val[0], _conv_row[0]); |
| 858 | out = vmlaq_n_u16(out, vextq_u16(data_u16.val[0], data_u16.val[1], 1), _conv_row[1]); |
| 859 | out = vmlaq_n_u16(out, vextq_u16(data_u16.val[0], data_u16.val[1], 2), _conv_row[2]); |
| 860 | out = vmlaq_n_u16(out, vextq_u16(data_u16.val[0], data_u16.val[1], 3), _conv_row[3]); |
| 861 | out = vmlaq_n_u16(out, vextq_u16(data_u16.val[0], data_u16.val[1], 4), _conv_row[4]); |
| 862 | out = vmlaq_n_u16(out, vextq_u16(data_u16.val[0], data_u16.val[1], 5), _conv_row[5]); |
| 863 | out = vmlaq_n_u16(out, vextq_u16(data_u16.val[0], data_u16.val[1], 6), _conv_row[6]); |
| 864 | |
| 865 | vst1q_u16(reinterpret_cast<uint16_t *>(output.ptr()), out); |
| 866 | }, |
| 867 | input, output); |
| 868 | } |
| 869 | |
| 870 | template <> |
| 871 | template <> |
| 872 | inline void NESeparableConvolutionHorKernel<7>::convolve<int16_t>(const Window &window) |
| 873 | { |
| 874 | Window win_in(window); |
| 875 | win_in.shift(Window::DimX, -3); |
| 876 | |
| 877 | Iterator input(_input, win_in); |
| 878 | Iterator output(_output, window); |
| 879 | |
Michalis Spyrou | a4f378d | 2019-04-26 14:54:54 +0100 | [diff] [blame] | 880 | execute_window_loop(window, [&](const Coordinates &) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 881 | { |
| 882 | const uint8x16_t data = vld1q_u8(input.ptr()); |
| 883 | |
| 884 | const int16x8x2_t data_s16 = |
| 885 | { |
| 886 | { |
| 887 | vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(data))), |
| 888 | vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(data))) |
| 889 | } |
| 890 | }; |
| 891 | |
| 892 | int16x8_t out = vmulq_n_s16(data_s16.val[0], _conv_row[0]); |
| 893 | out = vmlaq_n_s16(out, vextq_s16(data_s16.val[0], data_s16.val[1], 1), _conv_row[1]); |
| 894 | out = vmlaq_n_s16(out, vextq_s16(data_s16.val[0], data_s16.val[1], 2), _conv_row[2]); |
| 895 | out = vmlaq_n_s16(out, vextq_s16(data_s16.val[0], data_s16.val[1], 3), _conv_row[3]); |
| 896 | out = vmlaq_n_s16(out, vextq_s16(data_s16.val[0], data_s16.val[1], 4), _conv_row[4]); |
| 897 | out = vmlaq_n_s16(out, vextq_s16(data_s16.val[0], data_s16.val[1], 5), _conv_row[5]); |
| 898 | out = vmlaq_n_s16(out, vextq_s16(data_s16.val[0], data_s16.val[1], 6), _conv_row[6]); |
| 899 | |
| 900 | vst1q_s16(reinterpret_cast<int16_t *>(output.ptr()), out); |
| 901 | }, |
| 902 | input, output); |
| 903 | } |
| 904 | |
| 905 | template <> |
| 906 | template <> |
| 907 | void NESeparableConvolutionHorKernel<7>::convolve<int32_t>(const Window &window) |
| 908 | { |
| 909 | Window win_in(window); |
| 910 | win_in.shift(Window::DimX, -3); |
| 911 | |
| 912 | Iterator input(_input, win_in); |
| 913 | Iterator output(_output, window); |
| 914 | |
Michalis Spyrou | a4f378d | 2019-04-26 14:54:54 +0100 | [diff] [blame] | 915 | execute_window_loop(window, [&](const Coordinates &) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 916 | { |
| 917 | const uint8x16_t data = vld1q_u8(input.ptr()); |
| 918 | |
| 919 | const int16x8x2_t data_s16 = |
| 920 | { |
| 921 | { |
| 922 | vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(data))), |
| 923 | vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(data))) |
| 924 | } |
| 925 | }; |
| 926 | |
| 927 | const int16x8_t data_s16_l2 = vextq_s16(data_s16.val[0], data_s16.val[1], 1); |
| 928 | const int16x8_t data_s16_l1 = vextq_s16(data_s16.val[0], data_s16.val[1], 2); |
| 929 | const int16x8_t data_s16_m = vextq_s16(data_s16.val[0], data_s16.val[1], 3); |
| 930 | const int16x8_t data_s16_r1 = vextq_s16(data_s16.val[0], data_s16.val[1], 4); |
| 931 | const int16x8_t data_s16_r2 = vextq_s16(data_s16.val[0], data_s16.val[1], 5); |
| 932 | const int16x8_t data_s16_r3 = vextq_s16(data_s16.val[0], data_s16.val[1], 6); |
| 933 | |
| 934 | int32x4_t out_low = vmull_n_s16(vget_low_s16(data_s16.val[0]), _conv_row[0]); |
| 935 | out_low = vmlal_n_s16(out_low, vget_low_s16(data_s16_l2), _conv_row[1]); |
| 936 | out_low = vmlal_n_s16(out_low, vget_low_s16(data_s16_l1), _conv_row[2]); |
| 937 | out_low = vmlal_n_s16(out_low, vget_low_s16(data_s16_m), _conv_row[3]); |
| 938 | out_low = vmlal_n_s16(out_low, vget_low_s16(data_s16_r1), _conv_row[4]); |
| 939 | out_low = vmlal_n_s16(out_low, vget_low_s16(data_s16_r2), _conv_row[5]); |
| 940 | out_low = vmlal_n_s16(out_low, vget_low_s16(data_s16_r3), _conv_row[6]); |
| 941 | |
| 942 | vst1q_s32(reinterpret_cast<int32_t *>(output.ptr()), out_low); |
| 943 | |
| 944 | int32x4_t out_high = vmull_n_s16(vget_high_s16(data_s16.val[0]), _conv_row[0]); |
| 945 | out_high = vmlal_n_s16(out_high, vget_high_s16(data_s16_l2), _conv_row[1]); |
| 946 | out_high = vmlal_n_s16(out_high, vget_high_s16(data_s16_l1), _conv_row[2]); |
| 947 | out_high = vmlal_n_s16(out_high, vget_high_s16(data_s16_m), _conv_row[3]); |
| 948 | out_high = vmlal_n_s16(out_high, vget_high_s16(data_s16_r1), _conv_row[4]); |
| 949 | out_high = vmlal_n_s16(out_high, vget_high_s16(data_s16_r2), _conv_row[5]); |
| 950 | out_high = vmlal_n_s16(out_high, vget_high_s16(data_s16_r3), _conv_row[6]); |
| 951 | |
| 952 | vst1q_s32(reinterpret_cast<int32_t *>(output.ptr()) + 4, out_high); |
| 953 | }, |
| 954 | input, output); |
| 955 | } |
| 956 | |
| 957 | template <> |
| 958 | template <> |
| 959 | inline void NESeparableConvolutionHorKernel<9>::convolve<uint16_t>(const Window &window) |
| 960 | { |
| 961 | Window win_in(window); |
| 962 | win_in.shift(Window::DimX, -4); |
| 963 | |
| 964 | Iterator input(_input, win_in); |
| 965 | Iterator output(_output, window); |
| 966 | |
Michalis Spyrou | a4f378d | 2019-04-26 14:54:54 +0100 | [diff] [blame] | 967 | execute_window_loop(window, [&](const Coordinates &) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 968 | { |
| 969 | const uint8x16_t data = vld1q_u8(input.ptr()); |
| 970 | |
| 971 | const uint16x8x2_t data_u16 = |
| 972 | { |
| 973 | { |
| 974 | vmovl_u8(vget_low_u8(data)), |
| 975 | vmovl_u8(vget_high_u8(data)) |
| 976 | } |
| 977 | }; |
| 978 | |
| 979 | uint16x8_t out = vmulq_n_u16(data_u16.val[0], _conv_row[0]); |
| 980 | out = vmlaq_n_u16(out, vextq_u16(data_u16.val[0], data_u16.val[1], 1), _conv_row[1]); |
| 981 | out = vmlaq_n_u16(out, vextq_u16(data_u16.val[0], data_u16.val[1], 2), _conv_row[2]); |
| 982 | out = vmlaq_n_u16(out, vextq_u16(data_u16.val[0], data_u16.val[1], 3), _conv_row[3]); |
| 983 | out = vmlaq_n_u16(out, vextq_u16(data_u16.val[0], data_u16.val[1], 4), _conv_row[4]); |
| 984 | out = vmlaq_n_u16(out, vextq_u16(data_u16.val[0], data_u16.val[1], 5), _conv_row[5]); |
| 985 | out = vmlaq_n_u16(out, vextq_u16(data_u16.val[0], data_u16.val[1], 6), _conv_row[6]); |
| 986 | out = vmlaq_n_u16(out, vextq_u16(data_u16.val[0], data_u16.val[1], 7), _conv_row[7]); |
| 987 | out = vmlaq_n_u16(out, data_u16.val[1], _conv_row[8]); |
| 988 | |
| 989 | vst1q_u16(reinterpret_cast<uint16_t *>(output.ptr()), out); |
| 990 | }, |
| 991 | input, output); |
| 992 | } |
| 993 | |
| 994 | template <> |
| 995 | template <> |
| 996 | inline void NESeparableConvolutionHorKernel<9>::convolve<int16_t>(const Window &window) |
| 997 | { |
| 998 | Window win_in(window); |
| 999 | win_in.shift(Window::DimX, -4); |
| 1000 | |
| 1001 | Iterator input(_input, win_in); |
| 1002 | Iterator output(_output, window); |
| 1003 | |
Michalis Spyrou | a4f378d | 2019-04-26 14:54:54 +0100 | [diff] [blame] | 1004 | execute_window_loop(window, [&](const Coordinates &) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1005 | { |
| 1006 | const uint8x16_t data = vld1q_u8(input.ptr()); |
| 1007 | |
| 1008 | const int16x8x2_t data_s16 = |
| 1009 | { |
| 1010 | { |
| 1011 | vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(data))), |
| 1012 | vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(data))) |
| 1013 | } |
| 1014 | }; |
| 1015 | |
| 1016 | int16x8_t out = vmulq_n_s16(data_s16.val[0], _conv_row[0]); |
| 1017 | out = vmlaq_n_s16(out, vextq_s16(data_s16.val[0], data_s16.val[1], 1), _conv_row[1]); |
| 1018 | out = vmlaq_n_s16(out, vextq_s16(data_s16.val[0], data_s16.val[1], 2), _conv_row[2]); |
| 1019 | out = vmlaq_n_s16(out, vextq_s16(data_s16.val[0], data_s16.val[1], 3), _conv_row[3]); |
| 1020 | out = vmlaq_n_s16(out, vextq_s16(data_s16.val[0], data_s16.val[1], 4), _conv_row[4]); |
| 1021 | out = vmlaq_n_s16(out, vextq_s16(data_s16.val[0], data_s16.val[1], 5), _conv_row[5]); |
| 1022 | out = vmlaq_n_s16(out, vextq_s16(data_s16.val[0], data_s16.val[1], 6), _conv_row[6]); |
| 1023 | out = vmlaq_n_s16(out, vextq_s16(data_s16.val[0], data_s16.val[1], 7), _conv_row[7]); |
| 1024 | out = vmlaq_n_s16(out, data_s16.val[1], _conv_row[8]); |
| 1025 | |
| 1026 | vst1q_s16(reinterpret_cast<int16_t *>(output.ptr()), out); |
| 1027 | }, |
| 1028 | input, output); |
| 1029 | } |
| 1030 | |
| 1031 | template <> |
| 1032 | template <> |
| 1033 | void NESeparableConvolutionHorKernel<9>::convolve<int32_t>(const Window &window) |
| 1034 | { |
| 1035 | Window win_in(window); |
| 1036 | win_in.shift(Window::DimX, -4); |
| 1037 | |
| 1038 | Iterator input(_input, win_in); |
| 1039 | Iterator output(_output, window); |
| 1040 | |
Michalis Spyrou | a4f378d | 2019-04-26 14:54:54 +0100 | [diff] [blame] | 1041 | execute_window_loop(window, [&](const Coordinates &) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1042 | { |
| 1043 | const uint8x16_t data = vld1q_u8(input.ptr()); |
| 1044 | |
| 1045 | const int16x8x2_t data_s16 = |
| 1046 | { |
| 1047 | { |
| 1048 | vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(data))), |
| 1049 | vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(data))) |
| 1050 | } |
| 1051 | }; |
| 1052 | |
| 1053 | const int16x8_t data_s16_l3 = vextq_s16(data_s16.val[0], data_s16.val[1], 1); |
| 1054 | const int16x8_t data_s16_l2 = vextq_s16(data_s16.val[0], data_s16.val[1], 2); |
| 1055 | const int16x8_t data_s16_l1 = vextq_s16(data_s16.val[0], data_s16.val[1], 3); |
| 1056 | const int16x8_t data_s16_m = vextq_s16(data_s16.val[0], data_s16.val[1], 4); |
| 1057 | const int16x8_t data_s16_r1 = vextq_s16(data_s16.val[0], data_s16.val[1], 5); |
| 1058 | const int16x8_t data_s16_r2 = vextq_s16(data_s16.val[0], data_s16.val[1], 6); |
| 1059 | const int16x8_t data_s16_r3 = vextq_s16(data_s16.val[0], data_s16.val[1], 7); |
| 1060 | |
| 1061 | int32x4_t out_low = vmull_n_s16(vget_low_s16(data_s16.val[0]), _conv_row[0]); |
| 1062 | out_low = vmlal_n_s16(out_low, vget_low_s16(data_s16_l3), _conv_row[1]); |
| 1063 | out_low = vmlal_n_s16(out_low, vget_low_s16(data_s16_l2), _conv_row[2]); |
| 1064 | out_low = vmlal_n_s16(out_low, vget_low_s16(data_s16_l1), _conv_row[3]); |
| 1065 | out_low = vmlal_n_s16(out_low, vget_low_s16(data_s16_m), _conv_row[4]); |
| 1066 | out_low = vmlal_n_s16(out_low, vget_low_s16(data_s16_r1), _conv_row[5]); |
| 1067 | out_low = vmlal_n_s16(out_low, vget_low_s16(data_s16_r2), _conv_row[6]); |
| 1068 | out_low = vmlal_n_s16(out_low, vget_low_s16(data_s16_r3), _conv_row[7]); |
| 1069 | out_low = vmlal_n_s16(out_low, vget_low_s16(data_s16.val[1]), _conv_row[8]); |
| 1070 | |
| 1071 | vst1q_s32(reinterpret_cast<int32_t *>(output.ptr()), out_low); |
| 1072 | |
| 1073 | int32x4_t out_high = vmull_n_s16(vget_high_s16(data_s16.val[0]), _conv_row[0]); |
| 1074 | out_high = vmlal_n_s16(out_high, vget_high_s16(data_s16_l3), _conv_row[1]); |
| 1075 | out_high = vmlal_n_s16(out_high, vget_high_s16(data_s16_l2), _conv_row[2]); |
| 1076 | out_high = vmlal_n_s16(out_high, vget_high_s16(data_s16_l1), _conv_row[3]); |
| 1077 | out_high = vmlal_n_s16(out_high, vget_high_s16(data_s16_m), _conv_row[4]); |
| 1078 | out_high = vmlal_n_s16(out_high, vget_high_s16(data_s16_r1), _conv_row[5]); |
| 1079 | out_high = vmlal_n_s16(out_high, vget_high_s16(data_s16_r2), _conv_row[6]); |
| 1080 | out_high = vmlal_n_s16(out_high, vget_high_s16(data_s16_r3), _conv_row[7]); |
| 1081 | out_high = vmlal_n_s16(out_high, vget_high_s16(data_s16.val[1]), _conv_row[8]); |
| 1082 | |
| 1083 | vst1q_s32(reinterpret_cast<int32_t *>(output.ptr()) + 4, out_high); |
| 1084 | }, |
| 1085 | input, output); |
| 1086 | } |
| 1087 | |
| 1088 | template class arm_compute::NESeparableConvolutionHorKernel<5>; |
| 1089 | template class arm_compute::NESeparableConvolutionHorKernel<7>; |
| 1090 | template class arm_compute::NESeparableConvolutionHorKernel<9>; |
| 1091 | |
| 1092 | template <unsigned int matrix_size> |
| 1093 | NESeparableConvolutionVertKernel<matrix_size>::NESeparableConvolutionVertKernel() |
| 1094 | : _conv_col{ { 0 } }, _scale(0) |
| 1095 | { |
| 1096 | } |
| 1097 | |
| 1098 | template <unsigned int matrix_size> |
| 1099 | BorderSize NESeparableConvolutionVertKernel<matrix_size>::border_size() const |
| 1100 | { |
Michalis Spyrou | a4f378d | 2019-04-26 14:54:54 +0100 | [diff] [blame] | 1101 | return BorderSize{ matrix_size / 2, 0 }; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1102 | } |
| 1103 | |
| 1104 | template <unsigned int matrix_size> |
| 1105 | void NESeparableConvolutionVertKernel<matrix_size>::configure(const ITensor *input, ITensor *output, const int16_t *conv_col, uint32_t scale, bool border_undefined) |
| 1106 | { |
| 1107 | ARM_COMPUTE_ERROR_ON_NULLPTR(input, output, conv_col); |
| 1108 | |
| 1109 | set_shape_if_empty(*output->info(), input->info()->tensor_shape()); |
| 1110 | |
| 1111 | ARM_COMPUTE_ERROR_ON_MISMATCHING_SHAPES(input, output); |
| 1112 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U16, DataType::S16, DataType::S32); |
| 1113 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::U8, DataType::S16); |
| 1114 | ARM_COMPUTE_ERROR_ON(scale == 0); |
| 1115 | |
| 1116 | _input = input; |
| 1117 | _output = output; |
| 1118 | std::copy_n(conv_col, _conv_col.size(), _conv_col.begin()); |
| 1119 | _scale = scale; |
| 1120 | |
| 1121 | // Configure kernel window |
| 1122 | constexpr unsigned int num_elems_processed_per_iteration = 16; |
| 1123 | constexpr unsigned int num_elems_read_per_iteration = 16; |
| 1124 | constexpr unsigned int num_elems_written_per_iteration = 16; |
| 1125 | |
| 1126 | Window win = calculate_max_window(*input->info(), Steps(num_elems_processed_per_iteration), border_undefined, border_size()); |
| 1127 | AccessWindowHorizontal output_access(output->info(), 0, num_elems_written_per_iteration); |
| 1128 | |
| 1129 | update_window_and_padding(win, |
| 1130 | AccessWindowRectangle(input->info(), 0, -border_size().top, num_elems_read_per_iteration, matrix_size), |
| 1131 | output_access); |
| 1132 | |
| 1133 | output_access.set_valid_region(win, input->info()->valid_region(), border_undefined, border_size()); |
| 1134 | |
| 1135 | INEKernel::configure(win); |
| 1136 | } |
| 1137 | |
| 1138 | template <unsigned int matrix_size> |
Moritz Pflanzer | c186b57 | 2017-09-07 09:48:04 +0100 | [diff] [blame] | 1139 | void NESeparableConvolutionVertKernel<matrix_size>::run(const Window &window, const ThreadInfo &info) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1140 | { |
Moritz Pflanzer | c186b57 | 2017-09-07 09:48:04 +0100 | [diff] [blame] | 1141 | ARM_COMPUTE_UNUSED(info); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1142 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 1143 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window); |
| 1144 | |
| 1145 | switch(_input->info()->data_type()) |
| 1146 | { |
| 1147 | case DataType::U16: |
| 1148 | switch(_output->info()->data_type()) |
| 1149 | { |
| 1150 | case DataType::U8: |
| 1151 | convolution_u16<uint8_t>(window); |
| 1152 | break; |
| 1153 | case DataType::S16: |
| 1154 | convolution_u16<int16_t>(window); |
| 1155 | break; |
| 1156 | default: |
| 1157 | ARM_COMPUTE_ERROR("Not supported"); |
| 1158 | } |
| 1159 | break; |
| 1160 | case DataType::S16: |
| 1161 | switch(_output->info()->data_type()) |
| 1162 | { |
| 1163 | case DataType::U8: |
| 1164 | convolution_s16<uint8_t>(window); |
| 1165 | break; |
| 1166 | case DataType::S16: |
| 1167 | convolution_s16<int16_t>(window); |
| 1168 | break; |
| 1169 | default: |
| 1170 | ARM_COMPUTE_ERROR("Not supported"); |
| 1171 | } |
| 1172 | break; |
| 1173 | case DataType::S32: |
| 1174 | switch(_output->info()->data_type()) |
| 1175 | { |
| 1176 | case DataType::U8: |
| 1177 | convolution_s32<uint8_t>(window); |
| 1178 | break; |
| 1179 | case DataType::S16: |
| 1180 | convolution_s32<int16_t>(window); |
| 1181 | break; |
| 1182 | default: |
| 1183 | ARM_COMPUTE_ERROR("Not supported"); |
| 1184 | } |
| 1185 | break; |
| 1186 | default: |
| 1187 | ARM_COMPUTE_ERROR("Unsupported intermediate data type!"); |
| 1188 | break; |
| 1189 | } |
| 1190 | } |
| 1191 | |
| 1192 | template <unsigned int matrix_size> |
| 1193 | template <typename OutputType> |
| 1194 | void NESeparableConvolutionVertKernel<matrix_size>::convolution_u16(const Window &win) |
| 1195 | { |
| 1196 | static_assert(sizeof(OutputType) == sizeof(uint8_t) || sizeof(OutputType) == sizeof(int16_t), "The output buffer can only be u8 or s16"); |
| 1197 | |
| 1198 | Window win_in(win); |
| 1199 | win_in.set_dimension_step(Window::DimX, 8); |
| 1200 | |
| 1201 | Iterator in(_input, win_in); |
| 1202 | Iterator out(_output, win); |
| 1203 | |
| 1204 | std::array<unsigned char *, matrix_size> input_ptrs{ {} }; |
| 1205 | const float32x4_t oneoverscale = vdupq_n_f32(1.0f / _scale); |
| 1206 | const int k_half = matrix_size / 2; |
| 1207 | |
| 1208 | // Set row pointers |
| 1209 | for(int i = -k_half; i <= k_half; ++i) |
| 1210 | { |
| 1211 | input_ptrs[k_half + i] = _input->ptr_to_element(Coordinates(0, i)); |
| 1212 | } |
| 1213 | |
Michalis Spyrou | a4f378d | 2019-04-26 14:54:54 +0100 | [diff] [blame] | 1214 | execute_window_loop(win, [&](const Coordinates &) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1215 | { |
| 1216 | uint16x8_t out0 = vdupq_n_u16(0); |
| 1217 | uint16x8_t out1 = vdupq_n_u16(0); |
| 1218 | |
| 1219 | // First half |
| 1220 | for(unsigned int r = 0; r < matrix_size; ++r) |
| 1221 | { |
| 1222 | const uint16x8_t data = vld1q_u16(reinterpret_cast<const uint16_t *>(input_ptrs[r] + in.offset())); |
| 1223 | out0 = vmlaq_n_u16(out0, data, _conv_col[r]); |
| 1224 | } |
| 1225 | |
| 1226 | in.increment(Window::DimX); |
| 1227 | |
| 1228 | // Second half |
| 1229 | for(unsigned int r = 0; r < matrix_size; ++r) |
| 1230 | { |
| 1231 | const uint16x8_t data = vld1q_u16(reinterpret_cast<const uint16_t *>(input_ptrs[r] + in.offset())); |
| 1232 | out1 = vmlaq_n_u16(out1, data, _conv_col[r]); |
| 1233 | } |
| 1234 | |
| 1235 | //scale the result if needed |
| 1236 | if(_scale != 1) |
| 1237 | { |
| 1238 | float32x4_t out0_f32_high = vcvtq_f32_u32(vmovl_u16(vget_high_u16(out0))); |
| 1239 | float32x4_t out0_f32_low = vcvtq_f32_u32(vmovl_u16(vget_low_u16(out0))); |
| 1240 | out0_f32_high = vmulq_f32(out0_f32_high, oneoverscale); |
| 1241 | out0_f32_low = vmulq_f32(out0_f32_low, oneoverscale); |
| 1242 | store_results(vcvtq_u32_f32(out0_f32_low), vcvtq_u32_f32(out0_f32_high), reinterpret_cast<OutputType *>(out.ptr())); |
| 1243 | |
| 1244 | float32x4_t out1_f32_high = vcvtq_f32_u32(vmovl_u16(vget_high_u16(out1))); |
| 1245 | float32x4_t out1_f32_low = vcvtq_f32_u32(vmovl_u16(vget_low_u16(out1))); |
| 1246 | out1_f32_high = vmulq_f32(out1_f32_high, oneoverscale); |
| 1247 | out1_f32_low = vmulq_f32(out1_f32_low, oneoverscale); |
| 1248 | store_results(vcvtq_u32_f32(out1_f32_low), vcvtq_u32_f32(out1_f32_high), reinterpret_cast<OutputType *>(out.ptr()) + 8); |
| 1249 | } |
| 1250 | else |
| 1251 | { |
| 1252 | store_results(out0, out1, reinterpret_cast<OutputType *>(out.ptr())); |
| 1253 | } |
| 1254 | }, |
| 1255 | in, out); |
| 1256 | } |
| 1257 | |
| 1258 | template <unsigned int matrix_size> |
| 1259 | template <typename OutputType> |
| 1260 | void NESeparableConvolutionVertKernel<matrix_size>::convolution_s16(const Window &win) |
| 1261 | { |
| 1262 | static_assert(sizeof(OutputType) == sizeof(uint8_t) || sizeof(OutputType) == sizeof(int16_t), "The output buffer can only be u8 or s16"); |
| 1263 | |
| 1264 | Window win_in(win); |
| 1265 | win_in.set_dimension_step(Window::DimX, 8); |
| 1266 | |
| 1267 | Iterator in(_input, win_in); |
| 1268 | Iterator out(_output, win); |
| 1269 | |
| 1270 | std::array<unsigned char *, matrix_size> input_ptrs{ {} }; |
| 1271 | const float32x4_t oneoverscale = vdupq_n_f32(1.0f / _scale); |
| 1272 | const int k_half = matrix_size / 2; |
| 1273 | |
| 1274 | // Set row pointers |
| 1275 | for(int i = -k_half; i <= k_half; ++i) |
| 1276 | { |
| 1277 | input_ptrs[k_half + i] = _input->ptr_to_element(Coordinates(0, i)); |
| 1278 | } |
| 1279 | |
Michalis Spyrou | a4f378d | 2019-04-26 14:54:54 +0100 | [diff] [blame] | 1280 | execute_window_loop(win, [&](const Coordinates &) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1281 | { |
| 1282 | int16x8_t out0 = vdupq_n_s16(0); |
| 1283 | int16x8_t out1 = vdupq_n_s16(0); |
| 1284 | |
| 1285 | // First half |
| 1286 | for(unsigned int r = 0; r < matrix_size; ++r) |
| 1287 | { |
| 1288 | const int16x8_t data = vld1q_s16(reinterpret_cast<const int16_t *>(input_ptrs[r] + in.offset())); |
| 1289 | out0 = vmlaq_n_s16(out0, data, _conv_col[r]); |
| 1290 | } |
| 1291 | |
| 1292 | in.increment(Window::DimX); |
| 1293 | |
| 1294 | // Second half |
| 1295 | for(unsigned int r = 0; r < matrix_size; ++r) |
| 1296 | { |
| 1297 | const int16x8_t data = vld1q_s16(reinterpret_cast<const int16_t *>(input_ptrs[r] + in.offset())); |
| 1298 | out1 = vmlaq_n_s16(out1, data, _conv_col[r]); |
| 1299 | } |
| 1300 | |
| 1301 | //scale the result if needed |
| 1302 | if(_scale != 1) |
| 1303 | { |
| 1304 | float32x4_t out0_f32_high = vcvtq_f32_s32(vmovl_s16(vget_high_s16(out0))); |
| 1305 | float32x4_t out0_f32_low = vcvtq_f32_s32(vmovl_s16(vget_low_s16(out0))); |
| 1306 | out0_f32_high = vmulq_f32(out0_f32_high, oneoverscale); |
| 1307 | out0_f32_low = vmulq_f32(out0_f32_low, oneoverscale); |
| 1308 | store_results(vcvtq_s32_f32(out0_f32_low), vcvtq_s32_f32(out0_f32_high), reinterpret_cast<OutputType *>(out.ptr())); |
| 1309 | |
| 1310 | float32x4_t out1_f32_high = vcvtq_f32_s32(vmovl_s16(vget_high_s16(out1))); |
| 1311 | float32x4_t out1_f32_low = vcvtq_f32_s32(vmovl_s16(vget_low_s16(out1))); |
| 1312 | out1_f32_high = vmulq_f32(out1_f32_high, oneoverscale); |
| 1313 | out1_f32_low = vmulq_f32(out1_f32_low, oneoverscale); |
| 1314 | store_results(vcvtq_s32_f32(out1_f32_low), vcvtq_s32_f32(out1_f32_high), reinterpret_cast<OutputType *>(out.ptr()) + 8); |
| 1315 | } |
| 1316 | else |
| 1317 | { |
| 1318 | store_results(out0, out1, reinterpret_cast<OutputType *>(out.ptr())); |
| 1319 | } |
| 1320 | }, |
| 1321 | in, out); |
| 1322 | } |
| 1323 | |
| 1324 | template <unsigned int matrix_size> |
| 1325 | template <typename OutputType> |
| 1326 | void NESeparableConvolutionVertKernel<matrix_size>::convolution_s32(const Window &win) |
| 1327 | { |
| 1328 | static_assert(sizeof(OutputType) == sizeof(uint8_t) || sizeof(OutputType) == sizeof(int16_t), "The output buffer can only be u8 or s16"); |
| 1329 | |
| 1330 | Window win_in(win); |
| 1331 | win_in.set_dimension_step(Window::DimX, 8); |
| 1332 | |
| 1333 | Iterator in(_input, win_in); |
| 1334 | Iterator out(_output, win); |
| 1335 | |
| 1336 | std::array<unsigned char *, matrix_size> input_ptrs{ {} }; |
| 1337 | const float32x4_t oneoverscale = vdupq_n_f32(1.0f / _scale); |
| 1338 | const int k_half = matrix_size / 2; |
| 1339 | |
| 1340 | // Set row pointers |
| 1341 | for(int i = -k_half; i <= k_half; ++i) |
| 1342 | { |
| 1343 | input_ptrs[k_half + i] = _input->ptr_to_element(Coordinates(0, i)); |
| 1344 | } |
| 1345 | |
| 1346 | const int32x4_t zero = vdupq_n_s32(0); |
| 1347 | |
Michalis Spyrou | a4f378d | 2019-04-26 14:54:54 +0100 | [diff] [blame] | 1348 | execute_window_loop(win, [&](const Coordinates &) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1349 | { |
| 1350 | int32x4x2_t out0 = |
| 1351 | { |
| 1352 | { |
| 1353 | zero, |
| 1354 | zero |
| 1355 | } |
| 1356 | }; |
| 1357 | |
| 1358 | int32x4x2_t out1 = |
| 1359 | { |
| 1360 | { |
| 1361 | zero, |
| 1362 | zero |
| 1363 | } |
| 1364 | }; |
| 1365 | |
| 1366 | // First half |
| 1367 | for(unsigned int r = 0; r < matrix_size; ++r) |
| 1368 | { |
| 1369 | const int32x4x2_t data = vld2q_s32(reinterpret_cast<const int32_t *>(input_ptrs[r] + in.offset())); |
| 1370 | out0.val[0] = vmlaq_n_s32(out0.val[0], data.val[0], _conv_col[r]); |
| 1371 | out0.val[1] = vmlaq_n_s32(out0.val[1], data.val[1], _conv_col[r]); |
| 1372 | } |
| 1373 | |
| 1374 | in.increment(Window::DimX); |
| 1375 | |
| 1376 | // Second half |
| 1377 | for(unsigned int r = 0; r < matrix_size; ++r) |
| 1378 | { |
| 1379 | const int32x4x2_t data = vld2q_s32(reinterpret_cast<const int32_t *>(input_ptrs[r] + in.offset())); |
| 1380 | out1.val[0] = vmlaq_n_s32(out1.val[0], data.val[0], _conv_col[r]); |
| 1381 | out1.val[1] = vmlaq_n_s32(out1.val[1], data.val[1], _conv_col[r]); |
| 1382 | } |
| 1383 | |
| 1384 | //scale the result if needed |
| 1385 | if(_scale != 1) |
| 1386 | { |
| 1387 | float32x4_t out0_f32_odd = vcvtq_f32_s32(out0.val[0]); |
| 1388 | float32x4_t out0_f32_even = vcvtq_f32_s32(out0.val[1]); |
| 1389 | out0_f32_odd = vmulq_f32(out0_f32_odd, oneoverscale); |
| 1390 | out0_f32_even = vmulq_f32(out0_f32_even, oneoverscale); |
| 1391 | out0.val[0] = vcvtq_s32_f32(out0_f32_odd); |
| 1392 | out0.val[1] = vcvtq_s32_f32(out0_f32_even); |
| 1393 | |
| 1394 | float32x4_t out1_f32_odd = vcvtq_f32_s32(out1.val[0]); |
| 1395 | float32x4_t out1_f32_even = vcvtq_f32_s32(out1.val[1]); |
| 1396 | out1_f32_odd = vmulq_f32(out1_f32_odd, oneoverscale); |
| 1397 | out1_f32_even = vmulq_f32(out1_f32_even, oneoverscale); |
| 1398 | out1.val[0] = vcvtq_s32_f32(out1_f32_odd); |
| 1399 | out1.val[1] = vcvtq_s32_f32(out1_f32_even); |
| 1400 | } |
| 1401 | |
| 1402 | const int32x4x2_t out0_s32 = vzipq_s32(out0.val[0], out0.val[1]); |
| 1403 | store_results(out0_s32.val[0], out0_s32.val[1], reinterpret_cast<OutputType *>(out.ptr())); |
| 1404 | |
| 1405 | const int32x4x2_t out1_s32 = vzipq_s32(out1.val[0], out1.val[1]); |
| 1406 | store_results(out1_s32.val[0], out1_s32.val[1], reinterpret_cast<OutputType *>(out.ptr()) + 8); |
| 1407 | }, |
| 1408 | in, out); |
| 1409 | } |
| 1410 | |
| 1411 | template class arm_compute::NESeparableConvolutionVertKernel<5>; |
| 1412 | template class arm_compute::NESeparableConvolutionVertKernel<7>; |
| 1413 | template class arm_compute::NESeparableConvolutionVertKernel<9>; |
| 1414 | |
| 1415 | /****************************************************************************************\ |
| 1416 | * Rectangle Convolution * |
| 1417 | \****************************************************************************************/ |
| 1418 | |
| 1419 | NEConvolutionRectangleKernel::NEConvolutionRectangleKernel() |
| 1420 | : _input(nullptr), _output(nullptr), _scale(0), _convolution(), _border_size(), _func_idx(0) |
| 1421 | { |
| 1422 | } |
| 1423 | |
| 1424 | BorderSize NEConvolutionRectangleKernel::border_size() const |
| 1425 | { |
| 1426 | return _border_size; |
| 1427 | } |
| 1428 | |
| 1429 | void NEConvolutionRectangleKernel::configure(const ITensor *input, ITensor *output, const int16_t *conv, uint32_t width, uint32_t height, uint32_t scale, bool border_undefined) |
| 1430 | { |
| 1431 | ARM_COMPUTE_ERROR_ON_NULLPTR(input, output, conv); |
| 1432 | |
| 1433 | set_shape_if_empty(*output->info(), input->info()->tensor_shape()); |
| 1434 | |
| 1435 | ARM_COMPUTE_ERROR_ON_MISMATCHING_SHAPES(input, output); |
| 1436 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8); |
| 1437 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::U8, DataType::S16); |
| 1438 | ARM_COMPUTE_ERROR_ON(width != 3 && width != 5 && width != 7 && width != 9); |
| 1439 | ARM_COMPUTE_ERROR_ON(height != 3 && height != 5 && height != 7 && height != 9); |
| 1440 | ARM_COMPUTE_ERROR_ON(0 == scale); |
| 1441 | |
| 1442 | _input = input; |
| 1443 | _output = output; |
| 1444 | _scale = scale; |
| 1445 | _border_size = BorderSize(height / 2, width / 2); |
| 1446 | |
| 1447 | // Setup the convolution matrix |
| 1448 | const uint32_t nr_elements = width * height; |
| 1449 | _convolution.resize(nr_elements); |
| 1450 | std::copy_n(conv, nr_elements, _convolution.begin()); |
| 1451 | |
| 1452 | // Set function index to help choose appropriate function in run() |
| 1453 | _func_idx = get_index(height) * 4 + get_index(width); |
| 1454 | ARM_COMPUTE_ERROR_ON(_func_idx > (_nr_supported_sizes * _nr_supported_sizes)); |
| 1455 | |
| 1456 | // Configure kernel window |
| 1457 | constexpr unsigned int num_elems_processed_per_iteration = 8; |
| 1458 | constexpr unsigned int num_elems_read_per_iteration = 16; |
| 1459 | constexpr unsigned int num_elems_written_per_iteration = 8; |
| 1460 | |
Diego Lopez Recas | 0021d75 | 2017-12-18 14:42:56 +0000 | [diff] [blame] | 1461 | Window win = calculate_max_window(*input->info(), Steps(num_elems_processed_per_iteration), border_undefined, _border_size); |
| 1462 | AccessWindowHorizontal output_access(output->info(), 0, num_elems_written_per_iteration); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1463 | |
| 1464 | update_window_and_padding(win, |
| 1465 | AccessWindowRectangle(input->info(), -_border_size.left, -_border_size.top, num_elems_read_per_iteration, height), |
| 1466 | output_access); |
| 1467 | |
| 1468 | output_access.set_valid_region(win, input->info()->valid_region(), border_undefined, _border_size); |
| 1469 | |
| 1470 | INEKernel::configure(win); |
| 1471 | } |
| 1472 | |
Moritz Pflanzer | c186b57 | 2017-09-07 09:48:04 +0100 | [diff] [blame] | 1473 | void NEConvolutionRectangleKernel::run(const Window &window, const ThreadInfo &info) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1474 | { |
Moritz Pflanzer | c186b57 | 2017-09-07 09:48:04 +0100 | [diff] [blame] | 1475 | ARM_COMPUTE_UNUSED(info); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1476 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 1477 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window); |
| 1478 | |
| 1479 | using ConvolutionRectangleFunction = void (NEConvolutionRectangleKernel::*)(const Window & window); |
| 1480 | |
| 1481 | // uint8_t function table |
| 1482 | static const std::array<ConvolutionRectangleFunction, 16> func_table_u8 = |
| 1483 | { |
| 1484 | { |
| 1485 | &NEConvolutionRectangleKernel::convolution<uint8_t, 3, 3>, |
| 1486 | &NEConvolutionRectangleKernel::convolution<uint8_t, 3, 5>, |
| 1487 | &NEConvolutionRectangleKernel::convolution<uint8_t, 3, 7>, |
| 1488 | &NEConvolutionRectangleKernel::convolution<uint8_t, 3, 9>, |
| 1489 | &NEConvolutionRectangleKernel::convolution<uint8_t, 5, 3>, |
| 1490 | &NEConvolutionRectangleKernel::convolution<uint8_t, 5, 5>, |
| 1491 | &NEConvolutionRectangleKernel::convolution<uint8_t, 5, 7>, |
| 1492 | &NEConvolutionRectangleKernel::convolution<uint8_t, 5, 9>, |
| 1493 | &NEConvolutionRectangleKernel::convolution<uint8_t, 7, 3>, |
| 1494 | &NEConvolutionRectangleKernel::convolution<uint8_t, 7, 5>, |
| 1495 | &NEConvolutionRectangleKernel::convolution<uint8_t, 7, 7>, |
| 1496 | &NEConvolutionRectangleKernel::convolution<uint8_t, 7, 9>, |
| 1497 | &NEConvolutionRectangleKernel::convolution<uint8_t, 9, 3>, |
| 1498 | &NEConvolutionRectangleKernel::convolution<uint8_t, 9, 5>, |
| 1499 | &NEConvolutionRectangleKernel::convolution<uint8_t, 9, 7>, |
| 1500 | &NEConvolutionRectangleKernel::convolution<uint8_t, 9, 9> |
| 1501 | } |
| 1502 | }; |
| 1503 | // int16_t function table |
| 1504 | static const std::array<ConvolutionRectangleFunction, 16> func_table_s16 = |
| 1505 | { |
| 1506 | { |
| 1507 | &NEConvolutionRectangleKernel::convolution<int16_t, 3, 3>, |
| 1508 | &NEConvolutionRectangleKernel::convolution<int16_t, 3, 5>, |
| 1509 | &NEConvolutionRectangleKernel::convolution<int16_t, 3, 7>, |
| 1510 | &NEConvolutionRectangleKernel::convolution<int16_t, 3, 9>, |
| 1511 | &NEConvolutionRectangleKernel::convolution<int16_t, 5, 3>, |
| 1512 | &NEConvolutionRectangleKernel::convolution<int16_t, 5, 5>, |
| 1513 | &NEConvolutionRectangleKernel::convolution<int16_t, 5, 7>, |
| 1514 | &NEConvolutionRectangleKernel::convolution<int16_t, 5, 9>, |
| 1515 | &NEConvolutionRectangleKernel::convolution<int16_t, 7, 3>, |
| 1516 | &NEConvolutionRectangleKernel::convolution<int16_t, 7, 5>, |
| 1517 | &NEConvolutionRectangleKernel::convolution<int16_t, 7, 7>, |
| 1518 | &NEConvolutionRectangleKernel::convolution<int16_t, 7, 9>, |
| 1519 | &NEConvolutionRectangleKernel::convolution<int16_t, 9, 3>, |
| 1520 | &NEConvolutionRectangleKernel::convolution<int16_t, 9, 5>, |
| 1521 | &NEConvolutionRectangleKernel::convolution<int16_t, 9, 7>, |
| 1522 | &NEConvolutionRectangleKernel::convolution<int16_t, 9, 9> |
| 1523 | } |
| 1524 | }; |
| 1525 | |
| 1526 | // Run appropriate function |
Sanghoon Lee | d7ba539 | 2017-12-13 11:28:50 +0000 | [diff] [blame] | 1527 | switch(_output->info()->data_type()) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1528 | { |
Sanghoon Lee | d7ba539 | 2017-12-13 11:28:50 +0000 | [diff] [blame] | 1529 | case DataType::U8: |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1530 | ARM_COMPUTE_ERROR_ON(_func_idx >= func_table_u8.size()); |
| 1531 | (this->*func_table_u8[_func_idx])(window); |
| 1532 | break; |
Sanghoon Lee | d7ba539 | 2017-12-13 11:28:50 +0000 | [diff] [blame] | 1533 | case DataType::S16: |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1534 | ARM_COMPUTE_ERROR_ON(_func_idx >= func_table_s16.size()); |
| 1535 | (this->*func_table_s16[_func_idx])(window); |
| 1536 | break; |
| 1537 | default: |
| 1538 | ARM_COMPUTE_ERROR("Not supported"); |
| 1539 | } |
| 1540 | } |
| 1541 | |
| 1542 | unsigned int NEConvolutionRectangleKernel::get_index(uint32_t val) |
| 1543 | { |
| 1544 | switch(val) |
| 1545 | { |
| 1546 | case 3: |
| 1547 | return 0; |
| 1548 | case 5: |
| 1549 | return 1; |
| 1550 | case 7: |
| 1551 | return 2; |
| 1552 | case 9: |
| 1553 | return 3; |
| 1554 | default: |
| 1555 | ARM_COMPUTE_ERROR("Not supported dimension size"); |
| 1556 | return 0; |
| 1557 | } |
| 1558 | } |
| 1559 | |
| 1560 | template <typename OutputType, unsigned int rows, unsigned int cols> |
| 1561 | void NEConvolutionRectangleKernel::convolution(const Window &win) |
| 1562 | { |
| 1563 | static_assert(sizeof(OutputType) == sizeof(uint8_t) || sizeof(OutputType) == sizeof(int16_t), "The output buffer can only be u8 or s16"); |
| 1564 | ARM_COMPUTE_ERROR_ON(_input->buffer() == nullptr); |
| 1565 | |
| 1566 | Iterator input(_input, win); |
| 1567 | Iterator output(_output, win); |
| 1568 | |
| 1569 | std::array<unsigned char *, rows> input_ptrs{ {} }; |
| 1570 | const int16_t *conv = _convolution.data(); |
| 1571 | const float32x4_t scale_val = vdupq_n_f32(1.0f / _scale); |
| 1572 | const int k_row_half = rows / 2; |
| 1573 | const int k_col_half = cols / 2; |
| 1574 | |
| 1575 | // Set row pointers |
| 1576 | for(int i = -k_row_half; i <= k_row_half; ++i) |
| 1577 | { |
| 1578 | input_ptrs[k_row_half + i] = _input->buffer() + _input->info()->offset_element_in_bytes(Coordinates(-k_col_half, i)); |
| 1579 | } |
| 1580 | |
Michalis Spyrou | a4f378d | 2019-04-26 14:54:54 +0100 | [diff] [blame] | 1581 | execute_window_loop(win, [&](const Coordinates &) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1582 | { |
| 1583 | int32x4_t out = vdupq_n_s32(0); |
| 1584 | int32x4_t out2 = vdupq_n_s32(0); |
| 1585 | |
| 1586 | // Perform appropriate convolution |
| 1587 | for(unsigned int r = 0; r < rows; ++r) |
| 1588 | { |
| 1589 | const uint8x16_t data = vld1q_u8(input_ptrs[r] + input.offset()); |
| 1590 | if(3 == cols) |
| 1591 | { |
| 1592 | convolve_row3x1(out, out2, data, conv + r * cols); |
| 1593 | } |
| 1594 | else if(5 == cols) |
| 1595 | { |
| 1596 | convolve_row5x1(out, out2, data, conv + r * cols); |
| 1597 | } |
| 1598 | else if(7 == cols) |
| 1599 | { |
| 1600 | convolve_row7x1(out, out2, data, conv + r * cols); |
| 1601 | } |
| 1602 | else if(9 == cols) |
| 1603 | { |
| 1604 | convolve_row9x1(out, out2, data, conv + r * cols); |
| 1605 | } |
| 1606 | else |
| 1607 | { |
| 1608 | ARM_COMPUTE_ERROR("Unsupported number of columns"); |
| 1609 | } |
| 1610 | } |
| 1611 | |
| 1612 | // Apply scale |
| 1613 | if(_scale != 1) |
| 1614 | { |
| 1615 | // Convert to F32, scale and convert back to S32 |
| 1616 | out = vcvtq_s32_f32(vmulq_f32(vcvtq_f32_s32(out), scale_val)); |
| 1617 | out2 = vcvtq_s32_f32(vmulq_f32(vcvtq_f32_s32(out2), scale_val)); |
| 1618 | } |
| 1619 | |
| 1620 | // Clamp and store as U8 or S16: |
| 1621 | store_results(out, out2, reinterpret_cast<OutputType *>(output.ptr())); |
| 1622 | }, |
| 1623 | input, output); |
| 1624 | } |
| 1625 | } // namespace arm_compute |