File size: 20,224 Bytes
8b7c501 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 |
// Auto-generated file. Do not edit!
// Template: src/f16-dwconv2d-chw/3x3p1-neonfp16arith.c.in
// Generator: tools/xngen
//
// Copyright 2020 Google LLC
//
// This source code is licensed under the BSD-style license found in the
// LICENSE file in the root directory of this source tree.
#include <assert.h>
#include <arm_neon.h>
#include <xnnpack/dwconv.h>
#include <xnnpack/intrinsics-polyfill.h>
#include <xnnpack/math.h>
void xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_3x8(
size_t input_height,
size_t input_width,
const void* input,
const void* weights,
const void* zero,
void* output,
uint32_t padding_top,
const union xnn_f16_chw_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
assert(input_height != 0);
assert(input_width != 0);
assert(input_width % sizeof(uint16_t) == 0);
assert(padding_top == 1);
#if XNN_ARCH_ARM64
const uint16x8x2_t vminmax = vld2q_dup_u16(¶ms->neonfp16arith_stride1.min);
const float16x8_t vmin = vreinterpretq_f16_u16(vminmax.val[0]);
const float16x8_t vmax = vreinterpretq_f16_u16(vminmax.val[1]);
#else
// vld2_dup is to work around aarch32 clang bug with vld1q_dup
const uint16x4x2_t vminmax = vld2_dup_u16(¶ms->neonfp16arith_stride1.min);
const float16x8_t vmin = vreinterpretq_f16_u16(vcombine_u16(vminmax.val[0], vminmax.val[0]));
const float16x8_t vmax = vreinterpretq_f16_u16(vcombine_u16(vminmax.val[1], vminmax.val[1]));
#endif
const uint16x8_t vmask = vld1q_u16(params->neonfp16arith_stride1.mask);
const uint16_t* w = (const uint16_t*) weights;
const float16x8_t vw01234567 = vreinterpretq_f16_u16(vld1q_u16(w));
const float16x4_t vw89 = vreinterpret_f16_u32(vld1_dup_u32((const void*) (w + 8)));
const size_t input_decrement = round_up_po2(input_width, 8 * sizeof(uint16_t));
const uint16_t* i0 = zero;
const uint16_t* i1 = input;
const uint16_t* i2 = (const uint16_t*) ((uintptr_t) i1 + input_width);
const uint16_t* i3 = (const uint16_t*) ((uintptr_t) i2 + input_width);
const uint16_t* i4 = (const uint16_t*) ((uintptr_t) i3 + input_width);
uint16_t* o0 = output;
uint16_t* o1 = (uint16_t*) ((uintptr_t) o0 + input_width);
uint16_t* o2 = (uint16_t*) ((uintptr_t) o1 + input_width);
size_t output_height = input_height;
do {
if XNN_UNPREDICTABLE(output_height < 2) {
i2 = zero;
o1 = o0;
}
if XNN_UNPREDICTABLE(output_height < 3) {
i3 = zero;
o2 = o1;
}
if XNN_UNPREDICTABLE(output_height < 4) {
i4 = zero;
}
float16x8_t vi0x01234567 = vreinterpretq_f16_u16(vmovq_n_u16(0));
float16x8_t vi1x01234567 = vreinterpretq_f16_u16(vmovq_n_u16(0));
float16x8_t vi2x01234567 = vreinterpretq_f16_u16(vmovq_n_u16(0));
float16x8_t vi3x01234567 = vreinterpretq_f16_u16(vmovq_n_u16(0));
float16x8_t vi4x01234567 = vreinterpretq_f16_u16(vmovq_n_u16(0));
float16x8_t vi0x89ABCDEF = vreinterpretq_f16_u16(vld1q_u16(i0)); i0 += 8;
float16x8_t vi1x89ABCDEF = vreinterpretq_f16_u16(vld1q_u16(i1)); i1 += 8;
float16x8_t vi2x89ABCDEF = vreinterpretq_f16_u16(vld1q_u16(i2)); i2 += 8;
float16x8_t vi3x89ABCDEF = vreinterpretq_f16_u16(vld1q_u16(i3)); i3 += 8;
float16x8_t vi4x89ABCDEF = vreinterpretq_f16_u16(vld1q_u16(i4)); i4 += 8;
size_t w = input_width;
for (; w > 8 * sizeof(uint16_t); w -= 8 * sizeof(uint16_t)) {
float16x8_t vo0p0 = vdupq_lane_f16(vget_low_f16(vw01234567), 0);
float16x8_t vo1p0 = vdupq_lane_f16(vget_low_f16(vw01234567), 0);
float16x8_t vo2p0 = vdupq_lane_f16(vget_low_f16(vw01234567), 0);
const float16x8_t vi0xGHIJKLMN = vreinterpretq_f16_u16(vld1q_u16(i0)); i0 += 8;
const float16x8_t vi1xGHIJKLMN = vreinterpretq_f16_u16(vld1q_u16(i1)); i1 += 8;
const float16x8_t vi2xGHIJKLMN = vreinterpretq_f16_u16(vld1q_u16(i2)); i2 += 8;
const float16x8_t vi3xGHIJKLMN = vreinterpretq_f16_u16(vld1q_u16(i3)); i3 += 8;
const float16x8_t vi4xGHIJKLMN = vreinterpretq_f16_u16(vld1q_u16(i4)); i4 += 8;
// Center column
#if XNN_ARCH_ARM64
vo0p0 = vfmaq_laneq_f16(vo0p0, vi0x89ABCDEF, vw01234567, 2);
#else
vo0p0 = vmlaq_lane_f16(vo0p0, vi0x89ABCDEF, vget_low_f16(vw01234567), 2);
#endif
#if XNN_ARCH_ARM64
vo1p0 = vfmaq_laneq_f16(vo1p0, vi1x89ABCDEF, vw01234567, 2);
#else
vo1p0 = vmlaq_lane_f16(vo1p0, vi1x89ABCDEF, vget_low_f16(vw01234567), 2);
#endif
#if XNN_ARCH_ARM64
vo2p0 = vfmaq_laneq_f16(vo2p0, vi2x89ABCDEF, vw01234567, 2);
#else
vo2p0 = vmlaq_lane_f16(vo2p0, vi2x89ABCDEF, vget_low_f16(vw01234567), 2);
#endif
#if XNN_ARCH_ARM64
vo0p0 = vfmaq_laneq_f16(vo0p0, vi1x89ABCDEF, vw01234567, 5);
#else
vo0p0 = vmlaq_lane_f16(vo0p0, vi1x89ABCDEF, vget_high_f16(vw01234567), 1);
#endif
#if XNN_ARCH_ARM64
vo1p0 = vfmaq_laneq_f16(vo1p0, vi2x89ABCDEF, vw01234567, 5);
#else
vo1p0 = vmlaq_lane_f16(vo1p0, vi2x89ABCDEF, vget_high_f16(vw01234567), 1);
#endif
#if XNN_ARCH_ARM64
vo2p0 = vfmaq_laneq_f16(vo2p0, vi3x89ABCDEF, vw01234567, 5);
#else
vo2p0 = vmlaq_lane_f16(vo2p0, vi3x89ABCDEF, vget_high_f16(vw01234567), 1);
#endif
#if XNN_ARCH_ARM64
vo0p0 = vfmaq_lane_f16(vo0p0, vi2x89ABCDEF, vw89, 0);
#else
vo0p0 = vmlaq_lane_f16(vo0p0, vi2x89ABCDEF, vw89, 0);
#endif
#if XNN_ARCH_ARM64
vo1p0 = vfmaq_lane_f16(vo1p0, vi3x89ABCDEF, vw89, 0);
#else
vo1p0 = vmlaq_lane_f16(vo1p0, vi3x89ABCDEF, vw89, 0);
#endif
#if XNN_ARCH_ARM64
vo2p0 = vfmaq_lane_f16(vo2p0, vi4x89ABCDEF, vw89, 0);
#else
vo2p0 = vmlaq_lane_f16(vo2p0, vi4x89ABCDEF, vw89, 0);
#endif
// Left column
const float16x8_t vi0x789ABCDE = vextq_f16(vi0x01234567, vi0x89ABCDEF, 7);
const float16x8_t vi1x789ABCDE = vextq_f16(vi1x01234567, vi1x89ABCDEF, 7);
const float16x8_t vi2x789ABCDE = vextq_f16(vi2x01234567, vi2x89ABCDEF, 7);
const float16x8_t vi3x789ABCDE = vextq_f16(vi3x01234567, vi3x89ABCDEF, 7);
const float16x8_t vi4x789ABCDE = vextq_f16(vi4x01234567, vi4x89ABCDEF, 7);
#if XNN_ARCH_ARM64
vo0p0 = vfmaq_laneq_f16(vo0p0, vi0x789ABCDE, vw01234567, 1);
#else
vo0p0 = vmlaq_lane_f16(vo0p0, vi0x789ABCDE, vget_low_f16(vw01234567), 1);
#endif
#if XNN_ARCH_ARM64
vo1p0 = vfmaq_laneq_f16(vo1p0, vi1x789ABCDE, vw01234567, 1);
#else
vo1p0 = vmlaq_lane_f16(vo1p0, vi1x789ABCDE, vget_low_f16(vw01234567), 1);
#endif
#if XNN_ARCH_ARM64
vo2p0 = vfmaq_laneq_f16(vo2p0, vi2x789ABCDE, vw01234567, 1);
#else
vo2p0 = vmlaq_lane_f16(vo2p0, vi2x789ABCDE, vget_low_f16(vw01234567), 1);
#endif
#if XNN_ARCH_ARM64
vo0p0 = vfmaq_laneq_f16(vo0p0, vi1x789ABCDE, vw01234567, 4);
#else
vo0p0 = vmlaq_lane_f16(vo0p0, vi1x789ABCDE, vget_high_f16(vw01234567), 0);
#endif
#if XNN_ARCH_ARM64
vo1p0 = vfmaq_laneq_f16(vo1p0, vi2x789ABCDE, vw01234567, 4);
#else
vo1p0 = vmlaq_lane_f16(vo1p0, vi2x789ABCDE, vget_high_f16(vw01234567), 0);
#endif
#if XNN_ARCH_ARM64
vo2p0 = vfmaq_laneq_f16(vo2p0, vi3x789ABCDE, vw01234567, 4);
#else
vo2p0 = vmlaq_lane_f16(vo2p0, vi3x789ABCDE, vget_high_f16(vw01234567), 0);
#endif
#if XNN_ARCH_ARM64
vo0p0 = vfmaq_laneq_f16(vo0p0, vi2x789ABCDE, vw01234567, 7);
#else
vo0p0 = vmlaq_lane_f16(vo0p0, vi2x789ABCDE, vget_high_f16(vw01234567), 3);
#endif
#if XNN_ARCH_ARM64
vo1p0 = vfmaq_laneq_f16(vo1p0, vi3x789ABCDE, vw01234567, 7);
#else
vo1p0 = vmlaq_lane_f16(vo1p0, vi3x789ABCDE, vget_high_f16(vw01234567), 3);
#endif
#if XNN_ARCH_ARM64
vo2p0 = vfmaq_laneq_f16(vo2p0, vi4x789ABCDE, vw01234567, 7);
#else
vo2p0 = vmlaq_lane_f16(vo2p0, vi4x789ABCDE, vget_high_f16(vw01234567), 3);
#endif
vi0x01234567 = vi0x89ABCDEF;
vi1x01234567 = vi1x89ABCDEF;
vi2x01234567 = vi2x89ABCDEF;
vi3x01234567 = vi3x89ABCDEF;
vi4x01234567 = vi4x89ABCDEF;
// Right column
const float16x8_t vi0x9ABCDEFG = vextq_f16(vi0x89ABCDEF, vi0xGHIJKLMN, 1);
const float16x8_t vi1x9ABCDEFG = vextq_f16(vi1x89ABCDEF, vi1xGHIJKLMN, 1);
const float16x8_t vi2x9ABCDEFG = vextq_f16(vi2x89ABCDEF, vi2xGHIJKLMN, 1);
const float16x8_t vi3x9ABCDEFG = vextq_f16(vi3x89ABCDEF, vi3xGHIJKLMN, 1);
const float16x8_t vi4x9ABCDEFG = vextq_f16(vi4x89ABCDEF, vi4xGHIJKLMN, 1);
#if XNN_ARCH_ARM64
vo0p0 = vfmaq_laneq_f16(vo0p0, vi0x9ABCDEFG, vw01234567, 3);
#else
vo0p0 = vmlaq_lane_f16(vo0p0, vi0x9ABCDEFG, vget_low_f16(vw01234567), 3);
#endif
#if XNN_ARCH_ARM64
vo1p0 = vfmaq_laneq_f16(vo1p0, vi1x9ABCDEFG, vw01234567, 3);
#else
vo1p0 = vmlaq_lane_f16(vo1p0, vi1x9ABCDEFG, vget_low_f16(vw01234567), 3);
#endif
#if XNN_ARCH_ARM64
vo2p0 = vfmaq_laneq_f16(vo2p0, vi2x9ABCDEFG, vw01234567, 3);
#else
vo2p0 = vmlaq_lane_f16(vo2p0, vi2x9ABCDEFG, vget_low_f16(vw01234567), 3);
#endif
#if XNN_ARCH_ARM64
vo0p0 = vfmaq_laneq_f16(vo0p0, vi1x9ABCDEFG, vw01234567, 6);
#else
vo0p0 = vmlaq_lane_f16(vo0p0, vi1x9ABCDEFG, vget_high_f16(vw01234567), 2);
#endif
#if XNN_ARCH_ARM64
vo1p0 = vfmaq_laneq_f16(vo1p0, vi2x9ABCDEFG, vw01234567, 6);
#else
vo1p0 = vmlaq_lane_f16(vo1p0, vi2x9ABCDEFG, vget_high_f16(vw01234567), 2);
#endif
#if XNN_ARCH_ARM64
vo2p0 = vfmaq_laneq_f16(vo2p0, vi3x9ABCDEFG, vw01234567, 6);
#else
vo2p0 = vmlaq_lane_f16(vo2p0, vi3x9ABCDEFG, vget_high_f16(vw01234567), 2);
#endif
#if XNN_ARCH_ARM64
vo0p0 = vfmaq_lane_f16(vo0p0, vi2x9ABCDEFG, vw89, 1);
#else
vo0p0 = vmlaq_lane_f16(vo0p0, vi2x9ABCDEFG, vw89, 1);
#endif
#if XNN_ARCH_ARM64
vo1p0 = vfmaq_lane_f16(vo1p0, vi3x9ABCDEFG, vw89, 1);
#else
vo1p0 = vmlaq_lane_f16(vo1p0, vi3x9ABCDEFG, vw89, 1);
#endif
#if XNN_ARCH_ARM64
vo2p0 = vfmaq_lane_f16(vo2p0, vi4x9ABCDEFG, vw89, 1);
#else
vo2p0 = vmlaq_lane_f16(vo2p0, vi4x9ABCDEFG, vw89, 1);
#endif
vi0x89ABCDEF = vi0xGHIJKLMN;
vi1x89ABCDEF = vi1xGHIJKLMN;
vi2x89ABCDEF = vi2xGHIJKLMN;
vi3x89ABCDEF = vi3xGHIJKLMN;
vi4x89ABCDEF = vi4xGHIJKLMN;
float16x8_t vo0 = vmaxq_f16(vo0p0, vmin);
float16x8_t vo1 = vmaxq_f16(vo1p0, vmin);
float16x8_t vo2 = vmaxq_f16(vo2p0, vmin);
vo0 = vminq_f16(vo0, vmax);
vo1 = vminq_f16(vo1, vmax);
vo2 = vminq_f16(vo2, vmax);
vst1q_u16(o2, vreinterpretq_u16_f16(vo2)); o2 += 8;
vst1q_u16(o1, vreinterpretq_u16_f16(vo1)); o1 += 8;
vst1q_u16(o0, vreinterpretq_u16_f16(vo0)); o0 += 8;
}
// Always process the last block of 1..8 pixels.
assert(w >= 1 * sizeof(uint16_t));
assert(w <= 8 * sizeof(uint16_t));
{
float16x8_t vo0p0 = vdupq_lane_f16(vget_low_f16(vw01234567), 0);
float16x8_t vo1p0 = vdupq_lane_f16(vget_low_f16(vw01234567), 0);
float16x8_t vo2p0 = vdupq_lane_f16(vget_low_f16(vw01234567), 0);
vi0x89ABCDEF = vreinterpretq_f16_u16(vandq_u16(vmask, vreinterpretq_u16_f16(vi0x89ABCDEF)));
vi1x89ABCDEF = vreinterpretq_f16_u16(vandq_u16(vmask, vreinterpretq_u16_f16(vi1x89ABCDEF)));
vi2x89ABCDEF = vreinterpretq_f16_u16(vandq_u16(vmask, vreinterpretq_u16_f16(vi2x89ABCDEF)));
vi3x89ABCDEF = vreinterpretq_f16_u16(vandq_u16(vmask, vreinterpretq_u16_f16(vi3x89ABCDEF)));
vi4x89ABCDEF = vreinterpretq_f16_u16(vandq_u16(vmask, vreinterpretq_u16_f16(vi4x89ABCDEF)));
// Center column
#if XNN_ARCH_ARM64
vo0p0 = vfmaq_laneq_f16(vo0p0, vi0x89ABCDEF, vw01234567, 2);
#else
vo0p0 = vmlaq_lane_f16(vo0p0, vi0x89ABCDEF, vget_low_f16(vw01234567), 2);
#endif
#if XNN_ARCH_ARM64
vo1p0 = vfmaq_laneq_f16(vo1p0, vi1x89ABCDEF, vw01234567, 2);
#else
vo1p0 = vmlaq_lane_f16(vo1p0, vi1x89ABCDEF, vget_low_f16(vw01234567), 2);
#endif
#if XNN_ARCH_ARM64
vo2p0 = vfmaq_laneq_f16(vo2p0, vi2x89ABCDEF, vw01234567, 2);
#else
vo2p0 = vmlaq_lane_f16(vo2p0, vi2x89ABCDEF, vget_low_f16(vw01234567), 2);
#endif
#if XNN_ARCH_ARM64
vo0p0 = vfmaq_laneq_f16(vo0p0, vi1x89ABCDEF, vw01234567, 5);
#else
vo0p0 = vmlaq_lane_f16(vo0p0, vi1x89ABCDEF, vget_high_f16(vw01234567), 1);
#endif
#if XNN_ARCH_ARM64
vo1p0 = vfmaq_laneq_f16(vo1p0, vi2x89ABCDEF, vw01234567, 5);
#else
vo1p0 = vmlaq_lane_f16(vo1p0, vi2x89ABCDEF, vget_high_f16(vw01234567), 1);
#endif
#if XNN_ARCH_ARM64
vo2p0 = vfmaq_laneq_f16(vo2p0, vi3x89ABCDEF, vw01234567, 5);
#else
vo2p0 = vmlaq_lane_f16(vo2p0, vi3x89ABCDEF, vget_high_f16(vw01234567), 1);
#endif
#if XNN_ARCH_ARM64
vo0p0 = vfmaq_lane_f16(vo0p0, vi2x89ABCDEF, vw89, 0);
#else
vo0p0 = vmlaq_lane_f16(vo0p0, vi2x89ABCDEF, vw89, 0);
#endif
#if XNN_ARCH_ARM64
vo1p0 = vfmaq_lane_f16(vo1p0, vi3x89ABCDEF, vw89, 0);
#else
vo1p0 = vmlaq_lane_f16(vo1p0, vi3x89ABCDEF, vw89, 0);
#endif
#if XNN_ARCH_ARM64
vo2p0 = vfmaq_lane_f16(vo2p0, vi4x89ABCDEF, vw89, 0);
#else
vo2p0 = vmlaq_lane_f16(vo2p0, vi4x89ABCDEF, vw89, 0);
#endif
// Left column
const float16x8_t vi0x789ABCDE = vextq_f16(vi0x01234567, vi0x89ABCDEF, 7);
const float16x8_t vi1x789ABCDE = vextq_f16(vi1x01234567, vi1x89ABCDEF, 7);
const float16x8_t vi2x789ABCDE = vextq_f16(vi2x01234567, vi2x89ABCDEF, 7);
const float16x8_t vi3x789ABCDE = vextq_f16(vi3x01234567, vi3x89ABCDEF, 7);
const float16x8_t vi4x789ABCDE = vextq_f16(vi4x01234567, vi4x89ABCDEF, 7);
#if XNN_ARCH_ARM64
vo0p0 = vfmaq_laneq_f16(vo0p0, vi0x789ABCDE, vw01234567, 1);
#else
vo0p0 = vmlaq_lane_f16(vo0p0, vi0x789ABCDE, vget_low_f16(vw01234567), 1);
#endif
#if XNN_ARCH_ARM64
vo1p0 = vfmaq_laneq_f16(vo1p0, vi1x789ABCDE, vw01234567, 1);
#else
vo1p0 = vmlaq_lane_f16(vo1p0, vi1x789ABCDE, vget_low_f16(vw01234567), 1);
#endif
#if XNN_ARCH_ARM64
vo2p0 = vfmaq_laneq_f16(vo2p0, vi2x789ABCDE, vw01234567, 1);
#else
vo2p0 = vmlaq_lane_f16(vo2p0, vi2x789ABCDE, vget_low_f16(vw01234567), 1);
#endif
#if XNN_ARCH_ARM64
vo0p0 = vfmaq_laneq_f16(vo0p0, vi1x789ABCDE, vw01234567, 4);
#else
vo0p0 = vmlaq_lane_f16(vo0p0, vi1x789ABCDE, vget_high_f16(vw01234567), 0);
#endif
#if XNN_ARCH_ARM64
vo1p0 = vfmaq_laneq_f16(vo1p0, vi2x789ABCDE, vw01234567, 4);
#else
vo1p0 = vmlaq_lane_f16(vo1p0, vi2x789ABCDE, vget_high_f16(vw01234567), 0);
#endif
#if XNN_ARCH_ARM64
vo2p0 = vfmaq_laneq_f16(vo2p0, vi3x789ABCDE, vw01234567, 4);
#else
vo2p0 = vmlaq_lane_f16(vo2p0, vi3x789ABCDE, vget_high_f16(vw01234567), 0);
#endif
#if XNN_ARCH_ARM64
vo0p0 = vfmaq_laneq_f16(vo0p0, vi2x789ABCDE, vw01234567, 7);
#else
vo0p0 = vmlaq_lane_f16(vo0p0, vi2x789ABCDE, vget_high_f16(vw01234567), 3);
#endif
#if XNN_ARCH_ARM64
vo1p0 = vfmaq_laneq_f16(vo1p0, vi3x789ABCDE, vw01234567, 7);
#else
vo1p0 = vmlaq_lane_f16(vo1p0, vi3x789ABCDE, vget_high_f16(vw01234567), 3);
#endif
#if XNN_ARCH_ARM64
vo2p0 = vfmaq_laneq_f16(vo2p0, vi4x789ABCDE, vw01234567, 7);
#else
vo2p0 = vmlaq_lane_f16(vo2p0, vi4x789ABCDE, vget_high_f16(vw01234567), 3);
#endif
// Right column
const float16x8_t vzero = vreinterpretq_f16_u16(vmovq_n_u16(0));
const float16x8_t vi0x9ABCDEFG = vextq_f16(vi0x89ABCDEF, vzero, 1);
const float16x8_t vi1x9ABCDEFG = vextq_f16(vi1x89ABCDEF, vzero, 1);
const float16x8_t vi2x9ABCDEFG = vextq_f16(vi2x89ABCDEF, vzero, 1);
const float16x8_t vi3x9ABCDEFG = vextq_f16(vi3x89ABCDEF, vzero, 1);
const float16x8_t vi4x9ABCDEFG = vextq_f16(vi4x89ABCDEF, vzero, 1);
#if XNN_ARCH_ARM64
vo0p0 = vfmaq_laneq_f16(vo0p0, vi0x9ABCDEFG, vw01234567, 3);
#else
vo0p0 = vmlaq_lane_f16(vo0p0, vi0x9ABCDEFG, vget_low_f16(vw01234567), 3);
#endif
#if XNN_ARCH_ARM64
vo1p0 = vfmaq_laneq_f16(vo1p0, vi1x9ABCDEFG, vw01234567, 3);
#else
vo1p0 = vmlaq_lane_f16(vo1p0, vi1x9ABCDEFG, vget_low_f16(vw01234567), 3);
#endif
#if XNN_ARCH_ARM64
vo2p0 = vfmaq_laneq_f16(vo2p0, vi2x9ABCDEFG, vw01234567, 3);
#else
vo2p0 = vmlaq_lane_f16(vo2p0, vi2x9ABCDEFG, vget_low_f16(vw01234567), 3);
#endif
#if XNN_ARCH_ARM64
vo0p0 = vfmaq_laneq_f16(vo0p0, vi1x9ABCDEFG, vw01234567, 6);
#else
vo0p0 = vmlaq_lane_f16(vo0p0, vi1x9ABCDEFG, vget_high_f16(vw01234567), 2);
#endif
#if XNN_ARCH_ARM64
vo1p0 = vfmaq_laneq_f16(vo1p0, vi2x9ABCDEFG, vw01234567, 6);
#else
vo1p0 = vmlaq_lane_f16(vo1p0, vi2x9ABCDEFG, vget_high_f16(vw01234567), 2);
#endif
#if XNN_ARCH_ARM64
vo2p0 = vfmaq_laneq_f16(vo2p0, vi3x9ABCDEFG, vw01234567, 6);
#else
vo2p0 = vmlaq_lane_f16(vo2p0, vi3x9ABCDEFG, vget_high_f16(vw01234567), 2);
#endif
#if XNN_ARCH_ARM64
vo0p0 = vfmaq_lane_f16(vo0p0, vi2x9ABCDEFG, vw89, 1);
#else
vo0p0 = vmlaq_lane_f16(vo0p0, vi2x9ABCDEFG, vw89, 1);
#endif
#if XNN_ARCH_ARM64
vo1p0 = vfmaq_lane_f16(vo1p0, vi3x9ABCDEFG, vw89, 1);
#else
vo1p0 = vmlaq_lane_f16(vo1p0, vi3x9ABCDEFG, vw89, 1);
#endif
#if XNN_ARCH_ARM64
vo2p0 = vfmaq_lane_f16(vo2p0, vi4x9ABCDEFG, vw89, 1);
#else
vo2p0 = vmlaq_lane_f16(vo2p0, vi4x9ABCDEFG, vw89, 1);
#endif
float16x8_t vo0 = vmaxq_f16(vo0p0, vmin);
float16x8_t vo1 = vmaxq_f16(vo1p0, vmin);
float16x8_t vo2 = vmaxq_f16(vo2p0, vmin);
vo0 = vminq_f16(vo0, vmax);
vo1 = vminq_f16(vo1, vmax);
vo2 = vminq_f16(vo2, vmax);
if XNN_LIKELY(w == 8 * sizeof(uint16_t)) {
vst1q_u16(o2, vreinterpretq_u16_f16(vo2)); o2 += 8;
vst1q_u16(o1, vreinterpretq_u16_f16(vo1)); o1 += 8;
vst1q_u16(o0, vreinterpretq_u16_f16(vo0)); o0 += 8;
} else {
float16x4_t vo2_lo = vget_low_f16(vo2);
float16x4_t vo1_lo = vget_low_f16(vo1);
float16x4_t vo0_lo = vget_low_f16(vo0);
if (w & (4 * sizeof(uint16_t))) {
vst1_u16(o2, vreinterpret_u16_f16(vo2_lo)); o2 += 4;
vst1_u16(o1, vreinterpret_u16_f16(vo1_lo)); o1 += 4;
vst1_u16(o0, vreinterpret_u16_f16(vo0_lo)); o0 += 4;
vo2_lo = vget_high_f16(vo2);
vo1_lo = vget_high_f16(vo1);
vo0_lo = vget_high_f16(vo0);
}
if (w & (2 * sizeof(uint16_t))) {
vst1_lane_u32((void*) o2, vreinterpret_u32_f16(vo2_lo), 0); o2 += 2;
vst1_lane_u32((void*) o1, vreinterpret_u32_f16(vo1_lo), 0); o1 += 2;
vst1_lane_u32((void*) o0, vreinterpret_u32_f16(vo0_lo), 0); o0 += 2;
vo0_lo = vext_f16(vo0_lo, vo0_lo, 2);
vo1_lo = vext_f16(vo1_lo, vo1_lo, 2);
vo2_lo = vext_f16(vo2_lo, vo2_lo, 2);
}
if (w & (1 * sizeof(uint16_t))) {
vst1_lane_u16(o2, vreinterpret_u16_f16(vo2_lo), 0); o2 += 1;
vst1_lane_u16(o1, vreinterpret_u16_f16(vo1_lo), 0); o1 += 1;
vst1_lane_u16(o0, vreinterpret_u16_f16(vo0_lo), 0); o0 += 1;
}
}
}
i0 = (const uint16_t*) ((uintptr_t) i3 - input_decrement);
i1 = (const uint16_t*) ((uintptr_t) i4 - input_decrement);
i2 = (const uint16_t*) ((uintptr_t) i1 + input_width);
i3 = (const uint16_t*) ((uintptr_t) i2 + input_width);
i4 = (const uint16_t*) ((uintptr_t) i3 + input_width);
o0 = o2;
o1 = (uint16_t*) ((uintptr_t) o0 + input_width);
o2 = (uint16_t*) ((uintptr_t) o1 + input_width);
output_height = doz(output_height, 3);
} while (output_height != 0);
}
|