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/NEArithmeticAdditionKernel.h" |
| 25 | |
| 26 | #include "arm_compute/core/Error.h" |
| 27 | #include "arm_compute/core/Helpers.h" |
| 28 | #include "arm_compute/core/IAccessWindow.h" |
| 29 | #include "arm_compute/core/ITensor.h" |
Michele Di Giorgio | 81f0d15 | 2017-07-11 15:00:52 +0100 | [diff] [blame] | 30 | #include "arm_compute/core/NEON/NEFixedPoint.h" |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 31 | #include "arm_compute/core/TensorInfo.h" |
| 32 | #include "arm_compute/core/Validate.h" |
| 33 | |
| 34 | #include <algorithm> |
| 35 | #include <arm_neon.h> |
| 36 | #include <cstdint> |
| 37 | #include <map> |
| 38 | #include <string> |
| 39 | |
| 40 | using namespace arm_compute; |
| 41 | |
| 42 | namespace arm_compute |
| 43 | { |
| 44 | class Coordinates; |
| 45 | } // namespace arm_compute |
| 46 | |
| 47 | namespace |
| 48 | { |
Michele Di Giorgio | 81f0d15 | 2017-07-11 15:00:52 +0100 | [diff] [blame] | 49 | void add_wrap_QS8_QS8_QS8(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window) |
| 50 | { |
| 51 | Iterator input1(in1, window); |
| 52 | Iterator input2(in2, window); |
| 53 | Iterator output(out, window); |
| 54 | |
| 55 | execute_window_loop(window, [&](const Coordinates & id) |
| 56 | { |
| 57 | const qint8x16_t a = vld1q_qs8(reinterpret_cast<const qint8_t *>(input1.ptr())); |
| 58 | const qint8x16_t b = vld1q_qs8(reinterpret_cast<const qint8_t *>(input2.ptr())); |
| 59 | |
| 60 | vst1q_qs8(reinterpret_cast<qint8_t *>(output.ptr()), vaddq_qs8(a, b)); |
| 61 | }, |
| 62 | input1, input2, output); |
| 63 | } |
| 64 | |
| 65 | void add_saturate_QS8_QS8_QS8(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window) |
| 66 | { |
| 67 | Iterator input1(in1, window); |
| 68 | Iterator input2(in2, window); |
| 69 | Iterator output(out, window); |
| 70 | |
| 71 | execute_window_loop(window, [&](const Coordinates & id) |
| 72 | { |
| 73 | const qint8x16_t a = vld1q_qs8(reinterpret_cast<const qint8_t *>(input1.ptr())); |
| 74 | const qint8x16_t b = vld1q_qs8(reinterpret_cast<const qint8_t *>(input2.ptr())); |
| 75 | |
| 76 | vst1q_qs8(reinterpret_cast<qint8_t *>(output.ptr()), vqaddq_qs8(a, b)); |
| 77 | }, |
| 78 | input1, input2, output); |
| 79 | } |
| 80 | |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 81 | void add_wrap_U8_U8_U8(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window) |
| 82 | { |
| 83 | Iterator input1(in1, window); |
| 84 | Iterator input2(in2, window); |
| 85 | Iterator output(out, window); |
| 86 | |
| 87 | execute_window_loop(window, [&](const Coordinates & id) |
| 88 | { |
| 89 | vst1q_u8(output.ptr(), vaddq_u8(vld1q_u8(input1.ptr()), vld1q_u8(input2.ptr()))); |
| 90 | }, |
| 91 | input1, input2, output); |
| 92 | } |
| 93 | |
| 94 | void add_saturate_U8_U8_U8(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window) |
| 95 | { |
| 96 | Iterator input1(in1, window); |
| 97 | Iterator input2(in2, window); |
| 98 | Iterator output(out, window); |
| 99 | |
| 100 | execute_window_loop(window, [&](const Coordinates & id) |
| 101 | { |
| 102 | vst1q_u8(output.ptr(), vqaddq_u8(vld1q_u8(input1.ptr()), vld1q_u8(input2.ptr()))); |
| 103 | }, |
| 104 | input1, input2, output); |
| 105 | } |
| 106 | |
| 107 | inline int16x8x2_t vadd2q_s16(const int16x8x2_t &a, const int16x8x2_t &b) |
| 108 | { |
| 109 | const int16x8x2_t res = |
| 110 | { |
| 111 | { |
| 112 | vaddq_s16(a.val[0], b.val[0]), |
| 113 | vaddq_s16(a.val[1], b.val[1]) |
| 114 | } |
| 115 | }; |
| 116 | |
| 117 | return res; |
| 118 | } |
| 119 | |
| 120 | inline float32x4x4_t vadd4q_f32(const float32x4x4_t &a, const float32x4x4_t &b) |
| 121 | { |
| 122 | const float32x4x4_t res = |
| 123 | { |
| 124 | { |
| 125 | vaddq_f32(a.val[0], b.val[0]), |
| 126 | vaddq_f32(a.val[1], b.val[1]), |
| 127 | vaddq_f32(a.val[2], b.val[2]), |
| 128 | vaddq_f32(a.val[3], b.val[3]) |
| 129 | } |
| 130 | }; |
| 131 | |
| 132 | return res; |
| 133 | } |
| 134 | |
| 135 | inline int16x8x2_t vqadd2q_s16(const int16x8x2_t &a, const int16x8x2_t &b) |
| 136 | { |
| 137 | const int16x8x2_t res = |
| 138 | { |
| 139 | { |
| 140 | vqaddq_s16(a.val[0], b.val[0]), |
| 141 | vqaddq_s16(a.val[1], b.val[1]) |
| 142 | } |
| 143 | }; |
| 144 | |
| 145 | return res; |
| 146 | } |
| 147 | |
Ioan-Cristian Szabo | 33fd07b | 2017-10-26 15:42:24 +0100 | [diff] [blame] | 148 | #ifdef ARM_COMPUTE_AARCH64_V8_2 |
Pablo Tello | d1b0ecc | 2017-07-11 11:27:04 +0100 | [diff] [blame] | 149 | inline float16x8x2_t vadd2q_f16(const float16x8x2_t &a, const float16x8x2_t &b) |
| 150 | { |
| 151 | const float16x8x2_t res = |
| 152 | { |
| 153 | { |
| 154 | vaddq_f16(a.val[0], b.val[0]), |
| 155 | vaddq_f16(a.val[1], b.val[1]) |
| 156 | } |
| 157 | }; |
| 158 | |
| 159 | return res; |
| 160 | } |
Ioan-Cristian Szabo | 33fd07b | 2017-10-26 15:42:24 +0100 | [diff] [blame] | 161 | #endif /* ARM_COMPUTE_AARCH64_V8_2 */ |
Pablo Tello | d1b0ecc | 2017-07-11 11:27:04 +0100 | [diff] [blame] | 162 | |
| 163 | void add_F16_F16_F16(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window) |
| 164 | { |
Ioan-Cristian Szabo | 33fd07b | 2017-10-26 15:42:24 +0100 | [diff] [blame] | 165 | #ifdef ARM_COMPUTE_AARCH64_V8_2 |
Pablo Tello | d1b0ecc | 2017-07-11 11:27:04 +0100 | [diff] [blame] | 166 | Iterator input1(in1, window); |
| 167 | Iterator input2(in2, window); |
| 168 | Iterator output(out, window); |
| 169 | |
| 170 | execute_window_loop(window, [&](const Coordinates & id) |
| 171 | { |
| 172 | const float16x8x2_t a = vld2q_f16(reinterpret_cast<const float16_t *>(input1.ptr())); |
| 173 | const float16x8x2_t b = vld2q_f16(reinterpret_cast<const float16_t *>(input2.ptr())); |
| 174 | |
| 175 | vst2q_f16(reinterpret_cast<float16_t *>(output.ptr()), vadd2q_f16(a, b)); |
| 176 | }, |
| 177 | input1, input2, output); |
Ioan-Cristian Szabo | 33fd07b | 2017-10-26 15:42:24 +0100 | [diff] [blame] | 178 | #else /* ARM_COMPUTE_AARCH64_V8_2 */ |
Pablo Tello | d1b0ecc | 2017-07-11 11:27:04 +0100 | [diff] [blame] | 179 | ARM_COMPUTE_UNUSED(in1); |
| 180 | ARM_COMPUTE_UNUSED(in2); |
| 181 | ARM_COMPUTE_UNUSED(out); |
| 182 | ARM_COMPUTE_UNUSED(window); |
| 183 | ARM_COMPUTE_ERROR("Not supported, recompile the library with arch=arm64-v8.2-a"); |
Ioan-Cristian Szabo | 33fd07b | 2017-10-26 15:42:24 +0100 | [diff] [blame] | 184 | #endif /* ARM_COMPUTE_AARCH64_V8_2 */ |
Pablo Tello | d1b0ecc | 2017-07-11 11:27:04 +0100 | [diff] [blame] | 185 | } |
| 186 | |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 187 | void add_F32_F32_F32(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window) |
| 188 | { |
| 189 | Iterator input1(in1, window); |
| 190 | Iterator input2(in2, window); |
| 191 | Iterator output(out, window); |
| 192 | |
| 193 | execute_window_loop(window, [&](const Coordinates & id) |
| 194 | { |
| 195 | const float32x4x4_t a = vld4q_f32(reinterpret_cast<const float *>(input1.ptr())); |
| 196 | const float32x4x4_t b = vld4q_f32(reinterpret_cast<const float *>(input2.ptr())); |
| 197 | |
| 198 | vst4q_f32(reinterpret_cast<float *>(output.ptr()), vadd4q_f32(a, b)); |
| 199 | }, |
| 200 | input1, input2, output); |
| 201 | } |
| 202 | |
| 203 | void add_wrap_S16_S16_S16(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window) |
| 204 | { |
| 205 | Iterator input1(in1, window); |
| 206 | Iterator input2(in2, window); |
| 207 | Iterator output(out, window); |
| 208 | |
| 209 | execute_window_loop(window, [&](const Coordinates & id) |
| 210 | { |
| 211 | const int16x8x2_t a = vld2q_s16(reinterpret_cast<const int16_t *>(input1.ptr())); |
| 212 | const int16x8x2_t b = vld2q_s16(reinterpret_cast<const int16_t *>(input2.ptr())); |
| 213 | |
| 214 | vst2q_s16(reinterpret_cast<int16_t *>(output.ptr()), vadd2q_s16(a, b)); |
| 215 | }, |
| 216 | input1, input2, output); |
| 217 | } |
| 218 | |
| 219 | void add_saturate_S16_S16_S16(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window) |
| 220 | { |
| 221 | Iterator input1(in1, window); |
| 222 | Iterator input2(in2, window); |
| 223 | Iterator output(out, window); |
| 224 | |
| 225 | execute_window_loop(window, [&](const Coordinates & id) |
| 226 | { |
| 227 | const int16x8x2_t a = vld2q_s16(reinterpret_cast<const int16_t *>(input1.ptr())); |
| 228 | const int16x8x2_t b = vld2q_s16(reinterpret_cast<const int16_t *>(input2.ptr())); |
| 229 | |
| 230 | vst2q_s16(reinterpret_cast<int16_t *>(output.ptr()), vqadd2q_s16(a, b)); |
| 231 | }, |
| 232 | input1, input2, output); |
| 233 | } |
| 234 | |
| 235 | void add_wrap_S16_U8_S16(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window) |
| 236 | { |
| 237 | Iterator input1(in1, window); |
| 238 | Iterator input2(in2, window); |
| 239 | Iterator output(out, window); |
| 240 | |
| 241 | execute_window_loop(window, [&](const Coordinates & id) |
| 242 | { |
| 243 | const int16x8x2_t a = |
| 244 | { |
| 245 | { |
| 246 | vld1q_s16(reinterpret_cast<const int16_t *>(input1.ptr())), |
| 247 | vld1q_s16(reinterpret_cast<const int16_t *>(input1.ptr()) + 8) |
| 248 | } |
| 249 | }; |
| 250 | const uint8x16_t b = vld1q_u8(input2.ptr()); |
| 251 | |
| 252 | vst1q_s16(reinterpret_cast<int16_t *>(output.ptr()), vaddq_s16(a.val[0], vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(b))))); |
| 253 | vst1q_s16(reinterpret_cast<int16_t *>(output.ptr()) + 8, vaddq_s16(a.val[1], vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(b))))); |
| 254 | }, |
| 255 | input1, input2, output); |
| 256 | } |
| 257 | |
| 258 | void add_saturate_S16_U8_S16(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window) |
| 259 | { |
| 260 | Iterator input1(in1, window); |
| 261 | Iterator input2(in2, window); |
| 262 | Iterator output(out, window); |
| 263 | |
| 264 | execute_window_loop(window, [&](const Coordinates & id) |
| 265 | { |
| 266 | const int16x8x2_t a = |
| 267 | { |
| 268 | { |
| 269 | vld1q_s16(reinterpret_cast<const int16_t *>(input1.ptr())), |
| 270 | vld1q_s16(reinterpret_cast<const int16_t *>(input1.ptr()) + 8) |
| 271 | } |
| 272 | }; |
| 273 | const uint8x16_t b = vld1q_u8(input2.ptr()); |
| 274 | |
| 275 | vst1q_s16(reinterpret_cast<int16_t *>(output.ptr()), vqaddq_s16(a.val[0], vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(b))))); |
| 276 | vst1q_s16(reinterpret_cast<int16_t *>(output.ptr()) + 8, vqaddq_s16(a.val[1], vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(b))))); |
| 277 | }, |
| 278 | input1, input2, output); |
| 279 | } |
| 280 | |
| 281 | inline void add_wrap_U8_S16_S16(const ITensor *input1, const ITensor *input2, ITensor *output, const Window &window) |
| 282 | { |
| 283 | //Simply swap the two input buffers: |
| 284 | add_wrap_S16_U8_S16(input2, input1, output, window); |
| 285 | } |
| 286 | |
| 287 | inline void add_saturate_U8_S16_S16(const ITensor *input1, const ITensor *input2, ITensor *output, const Window &window) |
| 288 | { |
| 289 | //Simply swap the two input buffers: |
| 290 | add_saturate_S16_U8_S16(input2, input1, output, window); |
| 291 | } |
| 292 | |
| 293 | void add_wrap_U8_U8_S16(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window) |
| 294 | { |
| 295 | Iterator input1(in1, window); |
| 296 | Iterator input2(in2, window); |
| 297 | Iterator output(out, window); |
| 298 | |
| 299 | execute_window_loop(window, [&](const Coordinates & id) |
| 300 | { |
| 301 | const uint8x16_t a = vld1q_u8(input1.ptr()); |
| 302 | const uint8x16_t b = vld1q_u8(input2.ptr()); |
| 303 | |
| 304 | const int16x8x2_t a_s16 = |
| 305 | { |
| 306 | { |
| 307 | vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(a))), |
| 308 | vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(a))) |
| 309 | } |
| 310 | }; |
| 311 | |
| 312 | const int16x8x2_t b_s16 = |
| 313 | { |
| 314 | { |
| 315 | vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(b))), |
| 316 | vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(b))) |
| 317 | } |
| 318 | }; |
| 319 | |
| 320 | vst1q_s16(reinterpret_cast<int16_t *>(output.ptr()), vaddq_s16(a_s16.val[0], b_s16.val[0])); |
| 321 | vst1q_s16(reinterpret_cast<int16_t *>(output.ptr()) + 8, vaddq_s16(a_s16.val[1], b_s16.val[1])); |
| 322 | }, |
| 323 | input1, input2, output); |
| 324 | } |
| 325 | |
| 326 | void add_saturate_U8_U8_S16(const ITensor *in1, const ITensor *in2, ITensor *out, const Window &window) |
| 327 | { |
| 328 | Iterator input1(in1, window); |
| 329 | Iterator input2(in2, window); |
| 330 | Iterator output(out, window); |
| 331 | |
| 332 | execute_window_loop(window, [&](const Coordinates & id) |
| 333 | { |
| 334 | const uint8x16_t a = vld1q_u8(input1.ptr()); |
| 335 | const uint8x16_t b = vld1q_u8(input2.ptr()); |
| 336 | |
| 337 | const int16x8x2_t a_s16 = |
| 338 | { |
| 339 | { |
| 340 | vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(a))), |
| 341 | vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(a))) |
| 342 | } |
| 343 | }; |
| 344 | |
| 345 | const int16x8x2_t b_s16 = |
| 346 | { |
| 347 | { |
| 348 | vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(b))), |
| 349 | vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(b))) |
| 350 | } |
| 351 | }; |
| 352 | |
| 353 | vst1q_s16(reinterpret_cast<int16_t *>(output.ptr()), vqaddq_s16(a_s16.val[0], b_s16.val[0])); |
| 354 | vst1q_s16(reinterpret_cast<int16_t *>(output.ptr()) + 8, vqaddq_s16(a_s16.val[1], b_s16.val[1])); |
| 355 | }, |
| 356 | input1, input2, output); |
| 357 | } |
| 358 | } // namespace |
| 359 | |
| 360 | NEArithmeticAdditionKernel::NEArithmeticAdditionKernel() |
| 361 | : _func(nullptr), _input1(nullptr), _input2(nullptr), _output(nullptr) |
| 362 | { |
| 363 | } |
| 364 | |
| 365 | void NEArithmeticAdditionKernel::configure(const ITensor *input1, const ITensor *input2, ITensor *output, ConvertPolicy policy) |
| 366 | { |
| 367 | ARM_COMPUTE_ERROR_ON_NULLPTR(input1, input2, output); |
| 368 | |
Georgios Pinitas | f0dea70 | 2017-07-03 18:17:28 +0100 | [diff] [blame] | 369 | // Auto initialize output if not initialized |
| 370 | { |
| 371 | set_shape_if_empty(*output->info(), input1->info()->tensor_shape()); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 372 | |
Georgios Pinitas | f0dea70 | 2017-07-03 18:17:28 +0100 | [diff] [blame] | 373 | if(input1->info()->data_type() == DataType::S16 || input2->info()->data_type() == DataType::S16) |
| 374 | { |
| 375 | set_format_if_unknown(*output->info(), Format::S16); |
| 376 | } |
Pablo Tello | d1b0ecc | 2017-07-11 11:27:04 +0100 | [diff] [blame] | 377 | else if(input1->info()->data_type() == DataType::F16 || input2->info()->data_type() == DataType::F16) |
| 378 | { |
| 379 | set_format_if_unknown(*output->info(), Format::F16); |
| 380 | } |
Georgios Pinitas | f0dea70 | 2017-07-03 18:17:28 +0100 | [diff] [blame] | 381 | else if(input1->info()->data_type() == DataType::F32 || input2->info()->data_type() == DataType::F32) |
| 382 | { |
| 383 | set_format_if_unknown(*output->info(), Format::F32); |
| 384 | } |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 385 | } |
| 386 | |
| 387 | ARM_COMPUTE_ERROR_ON_MISMATCHING_SHAPES(input1, input2, output); |
Michele Di Giorgio | 81f0d15 | 2017-07-11 15:00:52 +0100 | [diff] [blame] | 388 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input1, 1, DataType::QS8, DataType::U8, DataType::QS16, DataType::S16, DataType::F16, DataType::F32); |
| 389 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input2, 1, DataType::QS8, DataType::U8, DataType::QS16, DataType::S16, DataType::F16, DataType::F32); |
| 390 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::QS8, DataType::U8, DataType::QS16, DataType::S16, DataType::F16, DataType::F32); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 391 | ARM_COMPUTE_ERROR_ON_MSG(output->info()->data_type() == DataType::U8 && (input1->info()->data_type() != DataType::U8 || input2->info()->data_type() != DataType::U8), |
| 392 | "Output can only be U8 if both inputs are U8"); |
Michele Di Giorgio | 81f0d15 | 2017-07-11 15:00:52 +0100 | [diff] [blame] | 393 | if(is_data_type_fixed_point(input1->info()->data_type()) || is_data_type_fixed_point(input2->info()->data_type()) || is_data_type_fixed_point(output->info()->data_type())) |
| 394 | { |
| 395 | // Check that all data types are the same and all fixed-point positions are the same |
| 396 | ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input1, input2, output); |
| 397 | } |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 398 | |
| 399 | static std::map<std::string, AddFunction *> map_function = |
| 400 | { |
Michele Di Giorgio | 81f0d15 | 2017-07-11 15:00:52 +0100 | [diff] [blame] | 401 | { "add_wrap_QS8_QS8_QS8", &add_wrap_QS8_QS8_QS8 }, |
| 402 | { "add_saturate_QS8_QS8_QS8", &add_saturate_QS8_QS8_QS8 }, |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 403 | { "add_wrap_U8_U8_U8", &add_wrap_U8_U8_U8 }, |
| 404 | { "add_saturate_U8_U8_U8", &add_saturate_U8_U8_U8 }, |
| 405 | { "add_wrap_S16_U8_S16", &add_wrap_S16_U8_S16 }, |
| 406 | { "add_saturate_S16_U8_S16", &add_saturate_S16_U8_S16 }, |
| 407 | { "add_wrap_U8_S16_S16", &add_wrap_U8_S16_S16 }, |
| 408 | { "add_saturate_U8_S16_S16", &add_saturate_U8_S16_S16 }, |
| 409 | { "add_wrap_U8_U8_S16", &add_wrap_U8_U8_S16 }, |
| 410 | { "add_saturate_U8_U8_S16", &add_saturate_U8_U8_S16 }, |
Michele Di Giorgio | 81f0d15 | 2017-07-11 15:00:52 +0100 | [diff] [blame] | 411 | { "add_wrap_QS16_QS16_QS16", &add_wrap_S16_S16_S16 }, |
| 412 | { "add_saturate_QS16_QS16_QS16", &add_saturate_S16_S16_S16 }, |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 413 | { "add_wrap_S16_S16_S16", &add_wrap_S16_S16_S16 }, |
| 414 | { "add_saturate_S16_S16_S16", &add_saturate_S16_S16_S16 }, |
| 415 | { "add_wrap_F32_F32_F32", &add_F32_F32_F32 }, |
| 416 | { "add_saturate_F32_F32_F32", &add_F32_F32_F32 }, |
Pablo Tello | d1b0ecc | 2017-07-11 11:27:04 +0100 | [diff] [blame] | 417 | { "add_wrap_F16_F16_F16", &add_F16_F16_F16 }, |
| 418 | { "add_saturate_F16_F16_F16", &add_F16_F16_F16 }, |
| 419 | |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 420 | }; |
| 421 | |
| 422 | _input1 = input1; |
| 423 | _input2 = input2; |
| 424 | _output = output; |
| 425 | |
| 426 | std::string function_to_call("add_"); |
| 427 | function_to_call += policy == ConvertPolicy::WRAP ? "wrap_" : "saturate_"; |
| 428 | function_to_call += string_from_data_type(input1->info()->data_type()) + "_"; |
| 429 | function_to_call += string_from_data_type(input2->info()->data_type()) + "_"; |
| 430 | function_to_call += string_from_data_type(output->info()->data_type()); |
| 431 | |
| 432 | auto it = map_function.find(function_to_call); |
| 433 | |
| 434 | if(it != map_function.end()) |
| 435 | { |
| 436 | _func = it->second; |
| 437 | } |
| 438 | else |
| 439 | { |
| 440 | ARM_COMPUTE_ERROR("You called arithmetic addition with the wrong tensor data type"); |
| 441 | } |
| 442 | |
| 443 | constexpr unsigned int num_elems_processed_per_iteration = 16; |
| 444 | |
| 445 | // Configure kernel window |
| 446 | Window win = calculate_max_window(*input1->info(), Steps(num_elems_processed_per_iteration)); |
| 447 | AccessWindowHorizontal output_access(output->info(), 0, num_elems_processed_per_iteration); |
| 448 | |
| 449 | update_window_and_padding(win, |
| 450 | AccessWindowHorizontal(input1->info(), 0, num_elems_processed_per_iteration), |
| 451 | AccessWindowHorizontal(input2->info(), 0, num_elems_processed_per_iteration), |
| 452 | output_access); |
| 453 | |
| 454 | ValidRegion valid_region = intersect_valid_regions(input1->info()->valid_region(), |
| 455 | input2->info()->valid_region()); |
| 456 | |
| 457 | output_access.set_valid_region(win, valid_region); |
| 458 | |
| 459 | INEKernel::configure(win); |
| 460 | } |
| 461 | |
Moritz Pflanzer | c186b57 | 2017-09-07 09:48:04 +0100 | [diff] [blame] | 462 | void NEArithmeticAdditionKernel::run(const Window &window, const ThreadInfo &info) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 463 | { |
Moritz Pflanzer | c186b57 | 2017-09-07 09:48:04 +0100 | [diff] [blame] | 464 | ARM_COMPUTE_UNUSED(info); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 465 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 466 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window); |
| 467 | ARM_COMPUTE_ERROR_ON(_func == nullptr); |
| 468 | |
| 469 | (*_func)(_input1, _input2, _output, window); |
| 470 | } |