File size: 42,417 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 |
// Copyright 2021 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 <immintrin.h>
#include <xnnpack/common.h>
#include <xnnpack/dwconv.h>
#include <xnnpack/intrinsics-polyfill.h>
#include <xnnpack/math.h>
#include <xnnpack/transpose.h>
#include <xnnpack/unaligned.h>
#include <xnnpack/vcvt.h>
#include <xnnpack/vlrelu.h>
void xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_2x4_acc2(
size_t input_height,
size_t input_width,
const float* input,
const float* weights,
const float* zero,
float* output,
uint32_t padding_top,
const union xnn_f32_chw_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
assert(input_height != 0);
assert(input_width != 0);
assert(input_width % sizeof(float) == 0);
assert(padding_top == 1);
const __m128 vmask = _mm_load_ps((const float*) params->sse_stride1.mask);
const __m128 vmax = _mm_load_ps(params->sse_stride1.max);
const __m128 vmin = _mm_load_ps(params->sse_stride1.min);
const __m128 vbias = _mm_load1_ps(weights);
const __m128 vk00 = _mm_load1_ps(weights + 1);
const __m128 vk01 = _mm_load1_ps(weights + 2);
const __m128 vk02 = _mm_load1_ps(weights + 3);
const __m128 vk10 = _mm_load1_ps(weights + 4);
const __m128 vk11 = _mm_load1_ps(weights + 5);
const __m128 vk12 = _mm_load1_ps(weights + 6);
const __m128 vk20 = _mm_load1_ps(weights + 7);
const __m128 vk21 = _mm_load1_ps(weights + 8);
const __m128 vk22 = _mm_load1_ps(weights + 9);
const size_t input_decrement = round_up_po2(input_width, 4 * sizeof(float));
const float* i0 = zero;
const float* i1 = input;
const float* i2 = (const float*) ((uintptr_t) i1 + input_width);
const float* i3 = (const float*) ((uintptr_t) i2 + input_width);
float* o0 = output;
float* o1 = (float*) ((uintptr_t) o0 + 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;
}
__m128 vi0x0123 = _mm_setzero_ps();
__m128 vi1x0123 = _mm_setzero_ps();
__m128 vi2x0123 = _mm_setzero_ps();
__m128 vi3x0123 = _mm_setzero_ps();
__m128 vi0x4567 = _mm_loadu_ps(i0);
i0 += 4;
__m128 vi1x4567 = _mm_loadu_ps(i1);
i1 += 4;
__m128 vi2x4567 = _mm_loadu_ps(i2);
i2 += 4;
__m128 vi3x4567 = _mm_loadu_ps(i3);
i3 += 4;
size_t w = input_width;
for (; w > 4 * sizeof(float); w -= 4 * sizeof(float)) {
const __m128 vi0x89AB = _mm_loadu_ps(i0);
i0 += 4;
const __m128 vi1x89AB = _mm_loadu_ps(i1);
i1 += 4;
const __m128 vi2x89AB = _mm_loadu_ps(i2);
i2 += 4;
const __m128 vi3x89AB = _mm_loadu_ps(i3);
i3 += 4;
__m128 vo0p0 = _mm_add_ps(vbias, _mm_mul_ps(vi0x4567, vk01));
__m128 vo1p0 = _mm_add_ps(vbias, _mm_mul_ps(vi1x4567, vk01));
__m128 vo0p1 = _mm_mul_ps(vi1x4567, vk11);
__m128 vo1p1 = _mm_mul_ps(vi2x4567, vk11);
vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x4567, vk21));
vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x4567, vk21));
const __m128 vi0x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi0x4567), _mm_castps_si128(vi0x0123), 12));
const __m128 vi1x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi1x4567), _mm_castps_si128(vi1x0123), 12));
const __m128 vi2x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi2x4567), _mm_castps_si128(vi2x0123), 12));
const __m128 vi3x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi3x4567), _mm_castps_si128(vi3x0123), 12));
vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi0x3456, vk00));
vo1p1 = _mm_add_ps(vo1p1, _mm_mul_ps(vi1x3456, vk00));
vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x3456, vk10));
vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x3456, vk10));
vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi2x3456, vk20));
vo1p1 = _mm_add_ps(vo1p1, _mm_mul_ps(vi3x3456, vk20));
vi0x0123 = vi0x4567;
vi1x0123 = vi1x4567;
vi2x0123 = vi2x4567;
vi3x0123 = vi3x4567;
const __m128 vi0x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi0x89AB), _mm_castps_si128(vi0x4567), 4));
const __m128 vi1x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi1x89AB), _mm_castps_si128(vi1x4567), 4));
const __m128 vi2x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi2x89AB), _mm_castps_si128(vi2x4567), 4));
const __m128 vi3x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi3x89AB), _mm_castps_si128(vi3x4567), 4));
vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi0x5678, vk02));
vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi1x5678, vk02));
vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi1x5678, vk12));
vo1p1 = _mm_add_ps(vo1p1, _mm_mul_ps(vi2x5678, vk12));
vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x5678, vk22));
vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x5678, vk22));
vi0x4567 = vi0x89AB;
vi1x4567 = vi1x89AB;
vi2x4567 = vi2x89AB;
vi3x4567 = vi3x89AB;
vo0p0 = _mm_add_ps(vo0p0, vo0p1);
vo1p0 = _mm_add_ps(vo1p0, vo1p1);
__m128 vo0 = _mm_max_ps(vo0p0, vmin);
__m128 vo1 = _mm_max_ps(vo1p0, vmin);
vo0 = _mm_min_ps(vo0, vmax);
vo1 = _mm_min_ps(vo1, vmax);
_mm_storeu_ps(o1, vo1);
o1 += 4;
_mm_storeu_ps(o0, vo0);
o0 += 4;
}
// Always process the last block of 1..4 pixels.
assert(w >= 1 * sizeof(float));
assert(w <= 4 * sizeof(float));
{
vi0x4567 = _mm_and_ps(vmask, vi0x4567);
vi1x4567 = _mm_and_ps(vmask, vi1x4567);
vi2x4567 = _mm_and_ps(vmask, vi2x4567);
vi3x4567 = _mm_and_ps(vmask, vi3x4567);
__m128 vo0p0 = _mm_add_ps(vbias, _mm_mul_ps(vi0x4567, vk01));
__m128 vo1p0 = _mm_add_ps(vbias, _mm_mul_ps(vi1x4567, vk01));
__m128 vo0p1 = _mm_mul_ps(vi1x4567, vk11);
__m128 vo1p1 = _mm_mul_ps(vi2x4567, vk11);
vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x4567, vk21));
vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x4567, vk21));
const __m128 vi0x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi0x4567), _mm_castps_si128(vi0x0123), 12));
const __m128 vi1x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi1x4567), _mm_castps_si128(vi1x0123), 12));
const __m128 vi2x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi2x4567), _mm_castps_si128(vi2x0123), 12));
const __m128 vi3x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi3x4567), _mm_castps_si128(vi3x0123), 12));
vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi0x3456, vk00));
vo1p1 = _mm_add_ps(vo1p1, _mm_mul_ps(vi1x3456, vk00));
vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x3456, vk10));
vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x3456, vk10));
vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi2x3456, vk20));
vo1p1 = _mm_add_ps(vo1p1, _mm_mul_ps(vi3x3456, vk20));
const __m128i vzero = _mm_setzero_si128();
const __m128 vi0x5678 = _mm_castsi128_ps(_mm_alignr_epi8(vzero, _mm_castps_si128(vi0x4567), 4));
const __m128 vi1x5678 = _mm_castsi128_ps(_mm_alignr_epi8(vzero, _mm_castps_si128(vi1x4567), 4));
const __m128 vi2x5678 = _mm_castsi128_ps(_mm_alignr_epi8(vzero, _mm_castps_si128(vi2x4567), 4));
const __m128 vi3x5678 = _mm_castsi128_ps(_mm_alignr_epi8(vzero, _mm_castps_si128(vi3x4567), 4));
vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi0x5678, vk02));
vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi1x5678, vk02));
vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi1x5678, vk12));
vo1p1 = _mm_add_ps(vo1p1, _mm_mul_ps(vi2x5678, vk12));
vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x5678, vk22));
vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x5678, vk22));
vo0p0 = _mm_add_ps(vo0p0, vo0p1);
vo1p0 = _mm_add_ps(vo1p0, vo1p1);
__m128 vo0 = _mm_max_ps(vo0p0, vmin);
__m128 vo1 = _mm_max_ps(vo1p0, vmin);
vo0 = _mm_min_ps(vo0, vmax);
vo1 = _mm_min_ps(vo1, vmax);
if XNN_LIKELY(w == 4 * sizeof(float)) {
_mm_storeu_ps(o1, vo1);
o1 += 4;
_mm_storeu_ps(o0, vo0);
o0 += 4;
} else {
if (w & (2 * sizeof(float))) {
_mm_storel_pi((__m64*) o1, vo1);
o1 += 2;
_mm_storel_pi((__m64*) o0, vo0);
o0 += 2;
vo0 = _mm_movehl_ps(vo0, vo0);
vo1 = _mm_movehl_ps(vo1, vo1);
}
if (w & (1 * sizeof(float))) {
_mm_store_ss(o1, vo1);
o1 += 1;
_mm_store_ss(o0, vo0);
o0 += 1;
}
}
}
i0 = (const float*) ((uintptr_t) i2 - input_decrement);
i1 = (const float*) ((uintptr_t) i3 - input_decrement);
i2 = (const float*) ((uintptr_t) i1 + input_width);
i3 = (const float*) ((uintptr_t) i2 + input_width);
o0 = o1;
o1 = (float*) ((uintptr_t) o0 + input_width);
output_height = doz(output_height, 2);
} while (output_height != 0);
}
void xnn_qs16_qs8_vcvt_ukernel__ssse3_x16(
size_t batch,
const int16_t* input,
int8_t* output,
const union xnn_qs16_qs8_cvt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
assert(batch != 0);
assert(batch % sizeof(int16_t) == 0);
assert(input != NULL);
assert(output != NULL);
const __m128i vinput_bias = _mm_load_si128((const __m128i*) params->ssse3.input_bias);
const __m128i vmultiplier = _mm_load_si128((const __m128i*) params->ssse3.multiplier);
const __m128i vbias = _mm_load_si128((const __m128i*) params->ssse3.bias);
const __m128i vshuffle01 = _mm_load_si128((const __m128i*) params->ssse3.shuffle01);
const __m128i vshuffle23 = _mm_load_si128((const __m128i*) params->ssse3.shuffle23);
const __m128i vshuffle45 = _mm_load_si128((const __m128i*) params->ssse3.shuffle45);
const __m128i vshuffle67 = _mm_load_si128((const __m128i*) params->ssse3.shuffle67);
for (; batch >= 16 * sizeof(int16_t); batch -= 16 * sizeof(int16_t)) {
__m128i vx0 = _mm_loadu_si128((const __m128i*) input); input += 8;
__m128i vx2 = _mm_loadu_si128((const __m128i*) input); input += 8;
// Add 0x8000 to convert signed inputs to unsigned.
vx0 = _mm_xor_si128(vx0, vinput_bias);
vx2 = _mm_xor_si128(vx2, vinput_bias);
// Move int16 to upper part of int32
__m128i vacc0lo = _mm_shuffle_epi8(vx0, vshuffle01);
__m128i vacc0hi = _mm_shuffle_epi8(vx0, vshuffle23);
__m128i vacc1lo = _mm_shuffle_epi8(vx0, vshuffle45);
__m128i vacc1hi = _mm_shuffle_epi8(vx0, vshuffle67);
__m128i vacc2lo = _mm_shuffle_epi8(vx2, vshuffle01);
__m128i vacc2hi = _mm_shuffle_epi8(vx2, vshuffle23);
__m128i vacc3lo = _mm_shuffle_epi8(vx2, vshuffle45);
__m128i vacc3hi = _mm_shuffle_epi8(vx2, vshuffle67);
vacc0lo = _mm_mul_epu32(vacc0lo, vmultiplier);
vacc0hi = _mm_mul_epu32(vacc0hi, vmultiplier);
vacc1lo = _mm_mul_epu32(vacc1lo, vmultiplier);
vacc1hi = _mm_mul_epu32(vacc1hi, vmultiplier);
vacc2lo = _mm_mul_epu32(vacc2lo, vmultiplier);
vacc2hi = _mm_mul_epu32(vacc2hi, vmultiplier);
vacc3lo = _mm_mul_epu32(vacc3lo, vmultiplier);
vacc3hi = _mm_mul_epu32(vacc3hi, vmultiplier);
vacc0lo = _mm_add_epi64(vacc0lo, vbias);
vacc0hi = _mm_add_epi64(vacc0hi, vbias);
vacc1lo = _mm_add_epi64(vacc1lo, vbias);
vacc1hi = _mm_add_epi64(vacc1hi, vbias);
vacc2lo = _mm_add_epi64(vacc2lo, vbias);
vacc2hi = _mm_add_epi64(vacc2hi, vbias);
vacc3lo = _mm_add_epi64(vacc3lo, vbias);
vacc3hi = _mm_add_epi64(vacc3hi, vbias);
__m128i vacc0 = _mm_castps_si128(_mm_shuffle_ps(_mm_castsi128_ps(vacc0lo), _mm_castsi128_ps(vacc0hi), _MM_SHUFFLE(3, 1, 3, 1)));
__m128i vacc1 = _mm_castps_si128(_mm_shuffle_ps(_mm_castsi128_ps(vacc1lo), _mm_castsi128_ps(vacc1hi), _MM_SHUFFLE(3, 1, 3, 1)));
__m128i vacc2 = _mm_castps_si128(_mm_shuffle_ps(_mm_castsi128_ps(vacc2lo), _mm_castsi128_ps(vacc2hi), _MM_SHUFFLE(3, 1, 3, 1)));
__m128i vacc3 = _mm_castps_si128(_mm_shuffle_ps(_mm_castsi128_ps(vacc3lo), _mm_castsi128_ps(vacc3hi), _MM_SHUFFLE(3, 1, 3, 1)));
// Pack 8 ints into 8 shorts
vacc0 = _mm_packs_epi32(vacc0, vacc1);
vacc2 = _mm_packs_epi32(vacc2, vacc3);
// Pack 16 shorts into 16 bytes
const __m128i vy0 = _mm_packs_epi16(vacc0, vacc2);
_mm_storeu_si128((__m128i*) output, vy0); output += 16;
}
for (; batch >= 4 * sizeof(int16_t); batch -= 4 * sizeof(int16_t)) {
__m128i vx = _mm_loadu_si128((const __m128i*) input); input += 4;
vx = _mm_xor_si128(vx, vinput_bias);
__m128i vacclo = _mm_shuffle_epi8(vx, vshuffle01);
__m128i vacchi = _mm_shuffle_epi8(vx, vshuffle23);
vacclo = _mm_mul_epu32(vacclo, vmultiplier);
vacchi = _mm_mul_epu32(vacchi, vmultiplier);
vacclo = _mm_add_epi64(vacclo, vbias);
vacchi = _mm_add_epi64(vacchi, vbias);
__m128i vacc = _mm_castps_si128(_mm_shuffle_ps(_mm_castsi128_ps(vacclo), _mm_castsi128_ps(vacchi), _MM_SHUFFLE(3, 1, 3, 1)));
vacc = _mm_packs_epi32(vacc, vacc);
const __m128i vy = _mm_packs_epi16(vacc, vacc);
_mm_storeu_si32(output, vy);
output += 4;
}
if XNN_UNLIKELY(batch != 0) {
assert(batch >= 1 * sizeof(int16_t));
assert(batch <= 3 * sizeof(int16_t));
__m128i vx = _mm_loadu_si128((const __m128i*) input);
vx = _mm_xor_si128(vx, vinput_bias);
__m128i vacclo = _mm_shuffle_epi8(vx, vshuffle01);
__m128i vacchi = _mm_shuffle_epi8(vx, vshuffle23);
vacclo = _mm_mul_epu32(vacclo, vmultiplier);
vacchi = _mm_mul_epu32(vacchi, vmultiplier);
vacclo = _mm_add_epi64(vacclo, vbias);
vacchi = _mm_add_epi64(vacchi, vbias);
__m128i vacc = _mm_castps_si128(_mm_shuffle_ps(_mm_castsi128_ps(vacclo), _mm_castsi128_ps(vacchi), _MM_SHUFFLE(3, 1, 3, 1)));
vacc = _mm_packs_epi32(vacc, vacc);
const __m128i vy = _mm_packs_epi16(vacc, vacc);
uint32_t vy_lo = (uint32_t) _mm_cvtsi128_si32(vy);
if (batch & (2 * sizeof(int16_t))) {
unaligned_store_u16(output, (uint16_t) vy_lo);
vy_lo >>= 16;
output += 2;
}
if (batch & (1 * sizeof(int16_t))) {
*output = (int8_t) vy_lo;
}
}
}
void xnn_qs8_vcvt_ukernel__ssse3_x32(
size_t batch,
const int8_t* input,
int8_t* output,
const union xnn_qs8_cvt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
assert(batch != 0);
assert(batch % sizeof(int8_t) == 0);
assert(input != NULL);
assert(output != NULL);
const __m128i vinput_zero_point = _mm_load_si128((const __m128i*) params->ssse3.input_zero_point);
const __m128i vmultiplier = _mm_load_si128((const __m128i*) params->ssse3.multiplier);
const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->ssse3.output_zero_point);
for (; batch >= 32 * sizeof(int8_t); batch -= 32 * sizeof(int8_t)) {
const __m128i vx0 = _mm_loadu_si128((const __m128i*) input);
const __m128i vx1 = _mm_loadu_si128((const __m128i*) (input + 16));
input += 32;
const __m128i vm0 = _mm_cmpgt_epi8(_mm_setzero_si128(), vx0);
__m128i vacc0 = _mm_unpacklo_epi8(vx0, vm0);
__m128i vacc1 = _mm_unpackhi_epi8(vx0, vm0);
const __m128i vm1 = _mm_cmpgt_epi8(_mm_setzero_si128(), vx1);
__m128i vacc2 = _mm_unpacklo_epi8(vx1, vm1);
__m128i vacc3 = _mm_unpackhi_epi8(vx1, vm1);
vacc0 = _mm_sub_epi16(vinput_zero_point, vacc0);
vacc1 = _mm_sub_epi16(vinput_zero_point, vacc1);
vacc2 = _mm_sub_epi16(vinput_zero_point, vacc2);
vacc3 = _mm_sub_epi16(vinput_zero_point, vacc3);
vacc0 = _mm_slli_epi16(vacc0, 7);
vacc1 = _mm_slli_epi16(vacc1, 7);
vacc2 = _mm_slli_epi16(vacc2, 7);
vacc3 = _mm_slli_epi16(vacc3, 7);
vacc0 = _mm_mulhrs_epi16(vacc0, vmultiplier);
vacc1 = _mm_mulhrs_epi16(vacc1, vmultiplier);
vacc2 = _mm_mulhrs_epi16(vacc2, vmultiplier);
vacc3 = _mm_mulhrs_epi16(vacc3, vmultiplier);
vacc0 = _mm_adds_epi16(vacc0, voutput_zero_point);
vacc1 = _mm_adds_epi16(vacc1, voutput_zero_point);
vacc2 = _mm_adds_epi16(vacc2, voutput_zero_point);
vacc3 = _mm_adds_epi16(vacc3, voutput_zero_point);
const __m128i vy0 = _mm_packs_epi16(vacc0, vacc1);
const __m128i vy1 = _mm_packs_epi16(vacc2, vacc3);
_mm_storeu_si128((__m128i*) output, vy0);
_mm_storeu_si128((__m128i*) (output + 16), vy1);
output += 32;
}
for (; batch >= 16 * sizeof(int8_t); batch -= 16 * sizeof(int8_t)) {
const __m128i vx = _mm_loadu_si128((const __m128i*) input);
input += 16;
const __m128i vm = _mm_cmpgt_epi8(_mm_setzero_si128(), vx);
__m128i vacc_lo = _mm_unpacklo_epi8(vx, vm);
__m128i vacc_hi = _mm_unpackhi_epi8(vx, vm);
vacc_lo = _mm_sub_epi16(vinput_zero_point, vacc_lo);
vacc_hi = _mm_sub_epi16(vinput_zero_point, vacc_hi);
vacc_lo = _mm_slli_epi16(vacc_lo, 7);
vacc_hi = _mm_slli_epi16(vacc_hi, 7);
vacc_lo = _mm_mulhrs_epi16(vacc_lo, vmultiplier);
vacc_hi = _mm_mulhrs_epi16(vacc_hi, vmultiplier);
vacc_lo = _mm_adds_epi16(vacc_lo, voutput_zero_point);
vacc_hi = _mm_adds_epi16(vacc_hi, voutput_zero_point);
const __m128i vy = _mm_packs_epi16(vacc_lo, vacc_hi);
_mm_storeu_si128((__m128i*) output, vy);
output += 16;
}
if XNN_UNLIKELY(batch != 0) {
assert(batch >= 1 * sizeof(int8_t));
assert(batch <= 15 * sizeof(int8_t));
const __m128i vx = _mm_loadu_si128((const __m128i*) input);
const __m128i vm = _mm_cmpgt_epi8(_mm_setzero_si128(), vx);
__m128i vacc_lo = _mm_unpacklo_epi8(vx, vm);
__m128i vacc_hi = _mm_unpackhi_epi8(vx, vm);
vacc_lo = _mm_sub_epi16(vinput_zero_point, vacc_lo);
vacc_hi = _mm_sub_epi16(vinput_zero_point, vacc_hi);
vacc_lo = _mm_slli_epi16(vacc_lo, 7);
vacc_hi = _mm_slli_epi16(vacc_hi, 7);
vacc_lo = _mm_mulhrs_epi16(vacc_lo, vmultiplier);
vacc_hi = _mm_mulhrs_epi16(vacc_hi, vmultiplier);
vacc_lo = _mm_adds_epi16(vacc_lo, voutput_zero_point);
vacc_hi = _mm_adds_epi16(vacc_hi, voutput_zero_point);
__m128i vy = _mm_packs_epi16(vacc_lo, vacc_hi);
if (batch & (8 * sizeof(int8_t))) {
_mm_storel_epi64((__m128i*) output, vy);
vy = _mm_unpackhi_epi64(vy, vy);
output += 8;
}
if (batch & (4 * sizeof(int8_t))) {
unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vy));
vy = _mm_srli_epi64(vy, 32);
output += 4;
}
uint32_t vy_lo = (uint32_t) _mm_cvtsi128_si32(vy);
if (batch & (2 * sizeof(int8_t))) {
unaligned_store_u16(output, (uint16_t) vy_lo);
vy_lo >>= 16;
output += 2;
}
if (batch & (1 * sizeof(int8_t))) {
*output = (int8_t) vy_lo;
}
}
}
void xnn_qs8_vlrelu_ukernel__ssse3_x32(
size_t batch,
const int8_t* input,
int8_t* output,
const union xnn_qs8_lrelu_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
assert(batch != 0);
assert(batch % sizeof(int8_t) == 0);
assert(input != NULL);
assert(output != NULL);
const __m128i vinput_zero_point = _mm_load_si128((const __m128i*) params->sse2.input_zero_point);
const __m128i vmultiplier_diff = _mm_load_si128((const __m128i*) params->sse2.multiplier_diff);
const __m128i vmultiplier_base = _mm_load_si128((const __m128i*) params->sse2.multiplier_base);
const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
for (; batch >= 32 * sizeof(int8_t); batch -= 32 * sizeof(int8_t)) {
const __m128i vx0 = _mm_loadu_si128((const __m128i*) input);
const __m128i vx1 = _mm_loadu_si128((const __m128i*) (input + 16));
input += 32;
const __m128i vm0 = _mm_cmpgt_epi8(_mm_setzero_si128(), vx0);
__m128i vacc0 = _mm_unpacklo_epi8(vx0, vm0);
__m128i vacc1 = _mm_unpackhi_epi8(vx0, vm0);
const __m128i vm1 = _mm_cmpgt_epi8(_mm_setzero_si128(), vx1);
__m128i vacc2 = _mm_unpacklo_epi8(vx1, vm1);
__m128i vacc3 = _mm_unpackhi_epi8(vx1, vm1);
__m128i vmultiplier0 = _mm_cmpgt_epi16(vacc0, vinput_zero_point);
vacc0 = _mm_sub_epi16(vinput_zero_point, vacc0);
__m128i vmultiplier1 = _mm_cmpgt_epi16(vacc1, vinput_zero_point);
vacc1 = _mm_sub_epi16(vinput_zero_point, vacc1);
__m128i vmultiplier2 = _mm_cmpgt_epi16(vacc2, vinput_zero_point);
vacc2 = _mm_sub_epi16(vinput_zero_point, vacc2);
__m128i vmultiplier3 = _mm_cmpgt_epi16(vacc3, vinput_zero_point);
vacc3 = _mm_sub_epi16(vinput_zero_point, vacc3);
vmultiplier0 = _mm_and_si128(vmultiplier0, vmultiplier_diff);
vacc0 = _mm_slli_epi16(vacc0, 7);
vmultiplier0 = _mm_xor_si128(vmultiplier0, vmultiplier_base);
vmultiplier1 = _mm_and_si128(vmultiplier1, vmultiplier_diff);
vacc1 = _mm_slli_epi16(vacc1, 7);
vmultiplier1 = _mm_xor_si128(vmultiplier1, vmultiplier_base);
vmultiplier2 = _mm_and_si128(vmultiplier2, vmultiplier_diff);
vacc2 = _mm_slli_epi16(vacc2, 7);
vmultiplier2 = _mm_xor_si128(vmultiplier2, vmultiplier_base);
vmultiplier3 = _mm_and_si128(vmultiplier3, vmultiplier_diff);
vacc3 = _mm_slli_epi16(vacc3, 7);
vmultiplier3 = _mm_xor_si128(vmultiplier3, vmultiplier_base);
vacc0 = _mm_mulhrs_epi16(vacc0, vmultiplier0);
vacc1 = _mm_mulhrs_epi16(vacc1, vmultiplier1);
vacc2 = _mm_mulhrs_epi16(vacc2, vmultiplier2);
vacc3 = _mm_mulhrs_epi16(vacc3, vmultiplier3);
vacc0 = _mm_adds_epi16(vacc0, voutput_zero_point);
vacc1 = _mm_adds_epi16(vacc1, voutput_zero_point);
vacc2 = _mm_adds_epi16(vacc2, voutput_zero_point);
vacc3 = _mm_adds_epi16(vacc3, voutput_zero_point);
const __m128i vy0 = _mm_packs_epi16(vacc0, vacc1);
const __m128i vy1 = _mm_packs_epi16(vacc2, vacc3);
_mm_storeu_si128((__m128i*) output, vy0);
_mm_storeu_si128((__m128i*) (output + 16), vy1);
output += 32;
}
for (; batch >= 16 * sizeof(int8_t); batch -= 16 * sizeof(int8_t)) {
const __m128i vx = _mm_loadu_si128((const __m128i*) input);
input += 16;
const __m128i vm = _mm_cmpgt_epi8(_mm_setzero_si128(), vx);
__m128i vacc_lo = _mm_unpacklo_epi8(vx, vm);
__m128i vacc_hi = _mm_unpackhi_epi8(vx, vm);
__m128i vmultiplier_lo = _mm_cmpgt_epi16(vacc_lo, vinput_zero_point);
__m128i vmultiplier_hi = _mm_cmpgt_epi16(vacc_hi, vinput_zero_point);
vacc_lo = _mm_sub_epi16(vinput_zero_point, vacc_lo);
vacc_hi = _mm_sub_epi16(vinput_zero_point, vacc_hi);
vmultiplier_lo = _mm_and_si128(vmultiplier_lo, vmultiplier_diff);
vmultiplier_hi = _mm_and_si128(vmultiplier_hi, vmultiplier_diff);
vacc_lo = _mm_slli_epi16(vacc_lo, 7);
vacc_hi = _mm_slli_epi16(vacc_hi, 7);
vmultiplier_lo = _mm_xor_si128(vmultiplier_lo, vmultiplier_base);
vmultiplier_hi = _mm_xor_si128(vmultiplier_hi, vmultiplier_base);
vacc_lo = _mm_mulhrs_epi16(vacc_lo, vmultiplier_lo);
vacc_hi = _mm_mulhrs_epi16(vacc_hi, vmultiplier_hi);
vacc_lo = _mm_adds_epi16(vacc_lo, voutput_zero_point);
vacc_hi = _mm_adds_epi16(vacc_hi, voutput_zero_point);
const __m128i vy = _mm_packs_epi16(vacc_lo, vacc_hi);
_mm_storeu_si128((__m128i*) output, vy);
output += 16;
}
if XNN_UNLIKELY(batch != 0) {
assert(batch >= 1 * sizeof(int8_t));
assert(batch <= 15 * sizeof(int8_t));
const __m128i vx = _mm_loadu_si128((const __m128i*) input);
const __m128i vm = _mm_cmpgt_epi8(_mm_setzero_si128(), vx);
__m128i vacc_lo = _mm_unpacklo_epi8(vx, vm);
__m128i vacc_hi = _mm_unpackhi_epi8(vx, vm);
__m128i vmultiplier_lo = _mm_cmpgt_epi16(vacc_lo, vinput_zero_point);
__m128i vmultiplier_hi = _mm_cmpgt_epi16(vacc_hi, vinput_zero_point);
vacc_lo = _mm_sub_epi16(vinput_zero_point, vacc_lo);
vacc_hi = _mm_sub_epi16(vinput_zero_point, vacc_hi);
vmultiplier_lo = _mm_and_si128(vmultiplier_lo, vmultiplier_diff);
vmultiplier_hi = _mm_and_si128(vmultiplier_hi, vmultiplier_diff);
vacc_lo = _mm_slli_epi16(vacc_lo, 7);
vacc_hi = _mm_slli_epi16(vacc_hi, 7);
vmultiplier_lo = _mm_xor_si128(vmultiplier_lo, vmultiplier_base);
vmultiplier_hi = _mm_xor_si128(vmultiplier_hi, vmultiplier_base);
vacc_lo = _mm_mulhrs_epi16(vacc_lo, vmultiplier_lo);
vacc_hi = _mm_mulhrs_epi16(vacc_hi, vmultiplier_hi);
vacc_lo = _mm_adds_epi16(vacc_lo, voutput_zero_point);
vacc_hi = _mm_adds_epi16(vacc_hi, voutput_zero_point);
__m128i vy = _mm_packs_epi16(vacc_lo, vacc_hi);
if (batch & (8 * sizeof(int8_t))) {
_mm_storel_epi64((__m128i*) output, vy);
vy = _mm_unpackhi_epi64(vy, vy);
output += 8;
}
if (batch & (4 * sizeof(int8_t))) {
unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vy));
vy = _mm_srli_epi64(vy, 32);
output += 4;
}
uint32_t vy_lo = (uint32_t) _mm_cvtsi128_si32(vy);
if (batch & (2 * sizeof(int8_t))) {
unaligned_store_u16(output, (uint16_t) vy_lo);
vy_lo >>= 16;
output += 2;
}
if (batch & (1 * sizeof(int8_t))) {
*output = (int8_t) vy_lo;
}
}
}
void xnn_qu8_vcvt_ukernel__ssse3_x32(
size_t batch,
const uint8_t* input,
uint8_t* output,
const union xnn_qu8_cvt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
assert(batch != 0);
assert(batch % sizeof(uint8_t) == 0);
assert(input != NULL);
assert(output != NULL);
const __m128i vinput_zero_point = _mm_load_si128((const __m128i*) params->ssse3.input_zero_point);
const __m128i vmultiplier = _mm_load_si128((const __m128i*) params->ssse3.multiplier);
const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->ssse3.output_zero_point);
const __m128i vzero = _mm_setzero_si128();
for (; batch >= 32 * sizeof(uint8_t); batch -= 32 * sizeof(uint8_t)) {
const __m128i vx0 = _mm_loadu_si128((const __m128i*) input);
const __m128i vx1 = _mm_loadu_si128((const __m128i*) (input + 16));
input += 32;
__m128i vacc0 = _mm_unpacklo_epi8(vx0, vzero);
__m128i vacc1 = _mm_unpackhi_epi8(vx0, vzero);
__m128i vacc2 = _mm_unpacklo_epi8(vx1, vzero);
__m128i vacc3 = _mm_unpackhi_epi8(vx1, vzero);
vacc0 = _mm_sub_epi16(vinput_zero_point, vacc0);
vacc1 = _mm_sub_epi16(vinput_zero_point, vacc1);
vacc2 = _mm_sub_epi16(vinput_zero_point, vacc2);
vacc3 = _mm_sub_epi16(vinput_zero_point, vacc3);
vacc0 = _mm_slli_epi16(vacc0, 7);
vacc1 = _mm_slli_epi16(vacc1, 7);
vacc2 = _mm_slli_epi16(vacc2, 7);
vacc3 = _mm_slli_epi16(vacc3, 7);
vacc0 = _mm_mulhrs_epi16(vacc0, vmultiplier);
vacc1 = _mm_mulhrs_epi16(vacc1, vmultiplier);
vacc2 = _mm_mulhrs_epi16(vacc2, vmultiplier);
vacc3 = _mm_mulhrs_epi16(vacc3, vmultiplier);
vacc0 = _mm_adds_epi16(vacc0, voutput_zero_point);
vacc1 = _mm_adds_epi16(vacc1, voutput_zero_point);
vacc2 = _mm_adds_epi16(vacc2, voutput_zero_point);
vacc3 = _mm_adds_epi16(vacc3, voutput_zero_point);
const __m128i vy0 = _mm_packus_epi16(vacc0, vacc1);
const __m128i vy1 = _mm_packus_epi16(vacc2, vacc3);
_mm_storeu_si128((__m128i*) output, vy0);
_mm_storeu_si128((__m128i*) (output + 16), vy1);
output += 32;
}
for (; batch >= 16 * sizeof(uint8_t); batch -= 16 * sizeof(uint8_t)) {
const __m128i vx = _mm_loadu_si128((const __m128i*) input);
input += 16;
__m128i vacc_lo = _mm_unpacklo_epi8(vx, vzero);
__m128i vacc_hi = _mm_unpackhi_epi8(vx, vzero);
vacc_lo = _mm_sub_epi16(vinput_zero_point, vacc_lo);
vacc_hi = _mm_sub_epi16(vinput_zero_point, vacc_hi);
vacc_lo = _mm_slli_epi16(vacc_lo, 7);
vacc_hi = _mm_slli_epi16(vacc_hi, 7);
vacc_lo = _mm_mulhrs_epi16(vacc_lo, vmultiplier);
vacc_hi = _mm_mulhrs_epi16(vacc_hi, vmultiplier);
vacc_lo = _mm_adds_epi16(vacc_lo, voutput_zero_point);
vacc_hi = _mm_adds_epi16(vacc_hi, voutput_zero_point);
const __m128i vy = _mm_packus_epi16(vacc_lo, vacc_hi);
_mm_storeu_si128((__m128i*) output, vy);
output += 16;
}
if XNN_UNLIKELY(batch != 0) {
assert(batch >= 1 * sizeof(uint8_t));
assert(batch <= 15 * sizeof(uint8_t));
const __m128i vx = _mm_loadu_si128((const __m128i*) input);
__m128i vacc_lo = _mm_unpacklo_epi8(vx, vzero);
__m128i vacc_hi = _mm_unpackhi_epi8(vx, vzero);
vacc_lo = _mm_sub_epi16(vinput_zero_point, vacc_lo);
vacc_hi = _mm_sub_epi16(vinput_zero_point, vacc_hi);
vacc_lo = _mm_slli_epi16(vacc_lo, 7);
vacc_hi = _mm_slli_epi16(vacc_hi, 7);
vacc_lo = _mm_mulhrs_epi16(vacc_lo, vmultiplier);
vacc_hi = _mm_mulhrs_epi16(vacc_hi, vmultiplier);
vacc_lo = _mm_adds_epi16(vacc_lo, voutput_zero_point);
vacc_hi = _mm_adds_epi16(vacc_hi, voutput_zero_point);
__m128i vy = _mm_packus_epi16(vacc_lo, vacc_hi);
if (batch & (8 * sizeof(uint8_t))) {
_mm_storel_epi64((__m128i*) output, vy);
vy = _mm_unpackhi_epi64(vy, vy);
output += 8;
}
if (batch & (4 * sizeof(uint8_t))) {
unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vy));
vy = _mm_srli_epi64(vy, 32);
output += 4;
}
uint32_t vy_lo = (uint32_t) _mm_cvtsi128_si32(vy);
if (batch & (2 * sizeof(uint8_t))) {
unaligned_store_u16(output, (uint16_t) vy_lo);
vy_lo >>= 16;
output += 2;
}
if (batch & (1 * sizeof(uint8_t))) {
*output = (uint8_t) vy_lo;
}
}
}
void xnn_qu8_vlrelu_ukernel__ssse3_x32(
size_t batch,
const uint8_t* input,
uint8_t* output,
const union xnn_qu8_lrelu_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
assert(batch != 0);
assert(batch % sizeof(uint8_t) == 0);
assert(input != NULL);
assert(output != NULL);
const __m128i vinput_zero_point = _mm_load_si128((const __m128i*) params->sse2.input_zero_point);
const __m128i vmultiplier_diff = _mm_load_si128((const __m128i*) params->sse2.multiplier_diff);
const __m128i vmultiplier_base = _mm_load_si128((const __m128i*) params->sse2.multiplier_base);
const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
const __m128i vzero = _mm_setzero_si128();
for (; batch >= 32 * sizeof(uint8_t); batch -= 32 * sizeof(uint8_t)) {
const __m128i vx0 = _mm_loadu_si128((const __m128i*) input);
const __m128i vx1 = _mm_loadu_si128((const __m128i*) (input + 16));
input += 32;
__m128i vacc0 = _mm_unpacklo_epi8(vx0, vzero);
__m128i vacc1 = _mm_unpackhi_epi8(vx0, vzero);
__m128i vacc2 = _mm_unpacklo_epi8(vx1, vzero);
__m128i vacc3 = _mm_unpackhi_epi8(vx1, vzero);
__m128i vmultiplier0 = _mm_cmpgt_epi16(vacc0, vinput_zero_point);
vacc0 = _mm_sub_epi16(vinput_zero_point, vacc0);
__m128i vmultiplier1 = _mm_cmpgt_epi16(vacc1, vinput_zero_point);
vacc1 = _mm_sub_epi16(vinput_zero_point, vacc1);
__m128i vmultiplier2 = _mm_cmpgt_epi16(vacc2, vinput_zero_point);
vacc2 = _mm_sub_epi16(vinput_zero_point, vacc2);
__m128i vmultiplier3 = _mm_cmpgt_epi16(vacc3, vinput_zero_point);
vacc3 = _mm_sub_epi16(vinput_zero_point, vacc3);
vmultiplier0 = _mm_and_si128(vmultiplier0, vmultiplier_diff);
vacc0 = _mm_slli_epi16(vacc0, 7);
vmultiplier0 = _mm_xor_si128(vmultiplier0, vmultiplier_base);
vmultiplier1 = _mm_and_si128(vmultiplier1, vmultiplier_diff);
vacc1 = _mm_slli_epi16(vacc1, 7);
vmultiplier1 = _mm_xor_si128(vmultiplier1, vmultiplier_base);
vmultiplier2 = _mm_and_si128(vmultiplier2, vmultiplier_diff);
vacc2 = _mm_slli_epi16(vacc2, 7);
vmultiplier2 = _mm_xor_si128(vmultiplier2, vmultiplier_base);
vmultiplier3 = _mm_and_si128(vmultiplier3, vmultiplier_diff);
vacc3 = _mm_slli_epi16(vacc3, 7);
vmultiplier3 = _mm_xor_si128(vmultiplier3, vmultiplier_base);
vacc0 = _mm_mulhrs_epi16(vacc0, vmultiplier0);
vacc1 = _mm_mulhrs_epi16(vacc1, vmultiplier1);
vacc2 = _mm_mulhrs_epi16(vacc2, vmultiplier2);
vacc3 = _mm_mulhrs_epi16(vacc3, vmultiplier3);
vacc0 = _mm_adds_epi16(vacc0, voutput_zero_point);
vacc1 = _mm_adds_epi16(vacc1, voutput_zero_point);
vacc2 = _mm_adds_epi16(vacc2, voutput_zero_point);
vacc3 = _mm_adds_epi16(vacc3, voutput_zero_point);
const __m128i vy0 = _mm_packus_epi16(vacc0, vacc1);
const __m128i vy1 = _mm_packus_epi16(vacc2, vacc3);
_mm_storeu_si128((__m128i*) output, vy0);
_mm_storeu_si128((__m128i*) (output + 16), vy1);
output += 32;
}
for (; batch >= 16 * sizeof(uint8_t); batch -= 16 * sizeof(uint8_t)) {
const __m128i vx = _mm_loadu_si128((const __m128i*) input);
input += 16;
__m128i vacc_lo = _mm_unpacklo_epi8(vx, vzero);
__m128i vacc_hi = _mm_unpackhi_epi8(vx, vzero);
__m128i vmultiplier_lo = _mm_cmpgt_epi16(vacc_lo, vinput_zero_point);
__m128i vmultiplier_hi = _mm_cmpgt_epi16(vacc_hi, vinput_zero_point);
vacc_lo = _mm_sub_epi16(vinput_zero_point, vacc_lo);
vacc_hi = _mm_sub_epi16(vinput_zero_point, vacc_hi);
vmultiplier_lo = _mm_and_si128(vmultiplier_lo, vmultiplier_diff);
vmultiplier_hi = _mm_and_si128(vmultiplier_hi, vmultiplier_diff);
vacc_lo = _mm_slli_epi16(vacc_lo, 7);
vacc_hi = _mm_slli_epi16(vacc_hi, 7);
vmultiplier_lo = _mm_xor_si128(vmultiplier_lo, vmultiplier_base);
vmultiplier_hi = _mm_xor_si128(vmultiplier_hi, vmultiplier_base);
vacc_lo = _mm_mulhrs_epi16(vacc_lo, vmultiplier_lo);
vacc_hi = _mm_mulhrs_epi16(vacc_hi, vmultiplier_hi);
vacc_lo = _mm_adds_epi16(vacc_lo, voutput_zero_point);
vacc_hi = _mm_adds_epi16(vacc_hi, voutput_zero_point);
const __m128i vy = _mm_packus_epi16(vacc_lo, vacc_hi);
_mm_storeu_si128((__m128i*) output, vy);
output += 16;
}
if XNN_UNLIKELY(batch != 0) {
assert(batch >= 1 * sizeof(uint8_t));
assert(batch <= 15 * sizeof(uint8_t));
const __m128i vx = _mm_loadu_si128((const __m128i*) input);
__m128i vacc_lo = _mm_unpacklo_epi8(vx, vzero);
__m128i vacc_hi = _mm_unpackhi_epi8(vx, vzero);
__m128i vmultiplier_lo = _mm_cmpgt_epi16(vacc_lo, vinput_zero_point);
__m128i vmultiplier_hi = _mm_cmpgt_epi16(vacc_hi, vinput_zero_point);
vacc_lo = _mm_sub_epi16(vinput_zero_point, vacc_lo);
vacc_hi = _mm_sub_epi16(vinput_zero_point, vacc_hi);
vmultiplier_lo = _mm_and_si128(vmultiplier_lo, vmultiplier_diff);
vmultiplier_hi = _mm_and_si128(vmultiplier_hi, vmultiplier_diff);
vacc_lo = _mm_slli_epi16(vacc_lo, 7);
vacc_hi = _mm_slli_epi16(vacc_hi, 7);
vmultiplier_lo = _mm_xor_si128(vmultiplier_lo, vmultiplier_base);
vmultiplier_hi = _mm_xor_si128(vmultiplier_hi, vmultiplier_base);
vacc_lo = _mm_mulhrs_epi16(vacc_lo, vmultiplier_lo);
vacc_hi = _mm_mulhrs_epi16(vacc_hi, vmultiplier_hi);
vacc_lo = _mm_adds_epi16(vacc_lo, voutput_zero_point);
vacc_hi = _mm_adds_epi16(vacc_hi, voutput_zero_point);
__m128i vy = _mm_packus_epi16(vacc_lo, vacc_hi);
if (batch & (8 * sizeof(uint8_t))) {
_mm_storel_epi64((__m128i*) output, vy);
vy = _mm_unpackhi_epi64(vy, vy);
output += 8;
}
if (batch & (4 * sizeof(uint8_t))) {
unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vy));
vy = _mm_srli_epi64(vy, 32);
output += 4;
}
uint32_t vy_lo = (uint32_t) _mm_cvtsi128_si32(vy);
if (batch & (2 * sizeof(uint8_t))) {
unaligned_store_u16(output, (uint16_t) vy_lo);
vy_lo >>= 16;
output += 2;
}
if (batch & (1 * sizeof(uint8_t))) {
*output = (uint8_t) vy_lo;
}
}
}
void xnn_x24_transposec_ukernel__4x4_ssse3(
const void *input,
void * output,
size_t input_stride,
size_t output_stride,
size_t block_width,
size_t block_height,
const union xnn_x24_transpose_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
assert(output_stride >= block_height * 3);
assert(input_stride >= block_width * 3);
assert(output_stride >= block_height * 3);
assert(input_stride >= block_width * 3);
const size_t tile_height = 4;
const size_t tile_width = 4;
const size_t tile_wbytes = tile_width * 3;
const size_t input_reset = tile_wbytes - round_down_po2(block_height, tile_height) * input_stride;
const size_t output_reset = tile_width * output_stride - block_height * 3;
const size_t tile_stride = tile_height * input_stride;
const uint8_t* i0 = (const uint8_t*) input;
const uint8_t* i1 = (const uint8_t*) ((uintptr_t) i0 + input_stride);
const uint8_t* i2 = (const uint8_t*) ((uintptr_t) i1 + input_stride);
const uint8_t* i3 = (const uint8_t*) ((uintptr_t) i2 + input_stride);
uint8_t* o0 = (uint8_t*) output;
uint8_t* o1 = (uint8_t*) ((uintptr_t) o0 + output_stride);
uint8_t* o2 = (uint8_t*) ((uintptr_t) o1 + output_stride);
uint8_t* o3 = (uint8_t*) ((uintptr_t) o2 + output_stride);
const __m128i vperm0 = _mm_loadu_si128((const __m128i*) params->ssse3.pos0);
const __m128i vperm1 = _mm_loadu_si128((const __m128i*) params->ssse3.pos1);
const __m128i vperm2 = _mm_loadu_si128((const __m128i*) params->ssse3.pos2);
const __m128i vperm3 = _mm_loadu_si128((const __m128i*) params->ssse3.pos3);
const __m128i vperm4 = _mm_loadu_si128((const __m128i*) params->ssse3.pos4);
const __m128i vperm5 = _mm_loadu_si128((const __m128i*) params->ssse3.pos5);
do {
if XNN_UNPREDICTABLE(block_width < 2) {
o1 = o0;
}
if XNN_UNPREDICTABLE(block_width <= 2) {
o2 = o0;
}
if XNN_UNPREDICTABLE(block_width < 4) {
o3 = o0;
}
size_t bh = block_height;
for (; bh >= 4; bh -= 4) {
const __m128i v0 = _mm_loadu_si128((const __m128i*) i0);
const __m128i v1 = _mm_loadu_si128((const __m128i*) i1);
const __m128i v2 = _mm_loadu_si128((const __m128i*) i2);
const __m128i v3 = _mm_loadu_si128((const __m128i*) i3);
i0 = (const uint8_t*) ((uintptr_t) i0 + tile_stride);
i1 = (const uint8_t*) ((uintptr_t) i1 + tile_stride);
i2 = (const uint8_t*) ((uintptr_t) i2 + tile_stride);
i3 = (const uint8_t*) ((uintptr_t) i3 + tile_stride);
const __m128i v1_0 = _mm_unpacklo_epi8(v0, v1);
const __m128i v1_1 = _mm_unpackhi_epi8(v0, v1);
const __m128i v1_2 = _mm_unpacklo_epi8(v2, v3);
const __m128i v1_3 = _mm_unpackhi_epi8(v2, v3);
const __m128i v3_0 = _mm_unpacklo_epi8(v1_0, v1_2);
const __m128i v3_1 = _mm_unpackhi_epi8(v1_0, v1_2);
const __m128i v3_2 = _mm_unpacklo_epi8(v1_1, v1_3);
__m128i v4_0 = _mm_shuffle_epi8(v3_0, vperm0);
__m128i v4_1 = _mm_or_si128(_mm_shuffle_epi8(v3_0, vperm2), _mm_shuffle_epi8(v3_1, vperm3));
__m128i v4_2 = _mm_or_si128(_mm_shuffle_epi8(v3_1, vperm4), _mm_shuffle_epi8(v3_2, vperm5));
__m128i v4_3 = _mm_shuffle_epi8(v3_2, vperm1);
_mm_storel_epi64((__m128i*) o3, v4_3);
_mm_storel_epi64((__m128i*) o2, v4_2);
_mm_storel_epi64((__m128i*) o1, v4_1);
_mm_storel_epi64((__m128i*) o0, v4_0);
o3 += 8;
o2 += 8;
o1 += 8;
o0 += 8;
v4_3 = _mm_unpackhi_epi64(v4_3, v4_3);
unaligned_store_u32(o3, (uint32_t) _mm_cvtsi128_si32(v4_3));
v4_2 = _mm_unpackhi_epi64(v4_2, v4_2);
unaligned_store_u32(o2, (uint32_t) _mm_cvtsi128_si32(v4_2));
v4_1 = _mm_unpackhi_epi64(v4_1, v4_1);
unaligned_store_u32(o1, (uint32_t) _mm_cvtsi128_si32(v4_1));
v4_0 = _mm_unpackhi_epi64(v4_0, v4_0);
unaligned_store_u32(o0, (uint32_t) _mm_cvtsi128_si32(v4_0));
o3 += 4;
o2 += 4;
o1 += 4;
o0 += 4;
}
if (bh != 0) {
if XNN_UNPREDICTABLE(bh <= 2) {
i2 = i0;
}
if XNN_UNPREDICTABLE(bh < 2) {
i1 = i0;
}
const __m128i v0 = _mm_loadu_si128((const __m128i*) i0);
const __m128i v1 = _mm_loadu_si128((const __m128i*) i1);
const __m128i v2 = _mm_loadu_si128((const __m128i*) i2);
const __m128i v1_0 = _mm_unpacklo_epi8(v0, v1);
const __m128i v1_1 = _mm_unpackhi_epi8(v0, v1);
const __m128i v1_2 = _mm_unpacklo_epi8(v2, v2);
const __m128i v1_3 = _mm_unpackhi_epi8(v2, v2);
const __m128i v3_0 = _mm_unpacklo_epi8(v1_0, v1_2);
const __m128i v3_1 = _mm_unpackhi_epi8(v1_0, v1_2);
const __m128i v3_2 = _mm_unpacklo_epi8(v1_1, v1_3);
__m128i v4_0 = _mm_shuffle_epi8(v3_0, vperm0);
__m128i v4_1 = _mm_or_si128(_mm_shuffle_epi8(v3_0, vperm2), _mm_shuffle_epi8(v3_1, vperm3));
__m128i v4_2 = _mm_or_si128(_mm_shuffle_epi8(v3_1, vperm4), _mm_shuffle_epi8(v3_2, vperm5));
__m128i v4_3 = _mm_shuffle_epi8(v3_2, vperm1);
if (bh & 2) {
unaligned_store_u32(o3, (uint32_t) _mm_cvtsi128_si32(v4_3));
unaligned_store_u32(o2, (uint32_t) _mm_cvtsi128_si32(v4_2));
unaligned_store_u32(o1, (uint32_t) _mm_cvtsi128_si32(v4_1));
unaligned_store_u32(o0, (uint32_t) _mm_cvtsi128_si32(v4_0));
o3 += 4;
o2 += 4;
o1 += 4;
o0 += 4;
unaligned_store_u16(o3, (uint16_t) _mm_extract_epi16(v4_3, 2));
unaligned_store_u16(o2, (uint16_t) _mm_extract_epi16(v4_2, 2));
unaligned_store_u16(o1, (uint16_t) _mm_extract_epi16(v4_1, 2));
unaligned_store_u16(o0, (uint16_t) _mm_extract_epi16(v4_0, 2));
o3 += 2;
o2 += 2;
o1 += 2;
o0 += 2;
v4_3 = _mm_bsrli_si128(v4_3, 6);
v4_2 = _mm_bsrli_si128(v4_2, 6);
v4_1 = _mm_bsrli_si128(v4_1, 6);
v4_0 = _mm_bsrli_si128(v4_0, 6);
}
if (bh & 1) {
unaligned_store_u16(o3, (uint16_t) _mm_cvtsi128_si32(v4_3));
unaligned_store_u16(o2, (uint16_t) _mm_cvtsi128_si32(v4_2));
unaligned_store_u16(o1, (uint16_t) _mm_cvtsi128_si32(v4_1));
unaligned_store_u16(o0, (uint16_t) _mm_cvtsi128_si32(v4_0));
o3 += 2;
o2 += 2;
o1 += 2;
o0 += 2;
*((uint8_t*) o3) = (uint8_t) _mm_cvtsi128_si32(_mm_bsrli_si128(v4_3, 2));
*((uint8_t*) o2) = (uint8_t) _mm_cvtsi128_si32(_mm_bsrli_si128(v4_2, 2));
*((uint8_t*) o1) = (uint8_t) _mm_cvtsi128_si32(_mm_bsrli_si128(v4_1, 2));
*((uint8_t*) o0) = (uint8_t) _mm_cvtsi128_si32(_mm_bsrli_si128(v4_0, 2));
o3 += 1;
o2 += 1;
o1 += 1;
o0 += 1;
}
}
i0 = (const uint8_t*) ((uintptr_t) i0 + input_reset);
i1 = (const uint8_t*) ((uintptr_t) i0 + input_stride);
i2 = (const uint8_t*) ((uintptr_t) i1 + input_stride);
i3 = (const uint8_t*) ((uintptr_t) i2 + input_stride);
o0 = (uint8_t*) ((uintptr_t) o0 + output_reset);
o1 = (uint8_t*) ((uintptr_t) o1 + output_reset);
o2 = (uint8_t*) ((uintptr_t) o2 + output_reset);
o3 = (uint8_t*) ((uintptr_t) o3 + output_reset);
block_width = doz(block_width, tile_width);
} while (block_width != 0);
}
|