Lines Matching refs:in1
488 * felem_mul sets |out| = |in1| * |in2|
490 * in1[i] < 2^64
493 * out[i] < 17 * max(in1[i]) * max(in2[i])
495 static void felem_mul(largefelem out, const felem in1, const felem in2)
500 out[0] = ((uint128_t) in1[0]) * in2[0];
502 out[1] = ((uint128_t) in1[0]) * in2[1] + ((uint128_t) in1[1]) * in2[0];
504 out[2] = ((uint128_t) in1[0]) * in2[2] +
505 ((uint128_t) in1[1]) * in2[1] + ((uint128_t) in1[2]) * in2[0];
507 out[3] = ((uint128_t) in1[0]) * in2[3] +
508 ((uint128_t) in1[1]) * in2[2] +
509 ((uint128_t) in1[2]) * in2[1] + ((uint128_t) in1[3]) * in2[0];
511 out[4] = ((uint128_t) in1[0]) * in2[4] +
512 ((uint128_t) in1[1]) * in2[3] +
513 ((uint128_t) in1[2]) * in2[2] +
514 ((uint128_t) in1[3]) * in2[1] + ((uint128_t) in1[4]) * in2[0];
516 out[5] = ((uint128_t) in1[0]) * in2[5] +
517 ((uint128_t) in1[1]) * in2[4] +
518 ((uint128_t) in1[2]) * in2[3] +
519 ((uint128_t) in1[3]) * in2[2] +
520 ((uint128_t) in1[4]) * in2[1] + ((uint128_t) in1[5]) * in2[0];
522 out[6] = ((uint128_t) in1[0]) * in2[6] +
523 ((uint128_t) in1[1]) * in2[5] +
524 ((uint128_t) in1[2]) * in2[4] +
525 ((uint128_t) in1[3]) * in2[3] +
526 ((uint128_t) in1[4]) * in2[2] +
527 ((uint128_t) in1[5]) * in2[1] + ((uint128_t) in1[6]) * in2[0];
529 out[7] = ((uint128_t) in1[0]) * in2[7] +
530 ((uint128_t) in1[1]) * in2[6] +
531 ((uint128_t) in1[2]) * in2[5] +
532 ((uint128_t) in1[3]) * in2[4] +
533 ((uint128_t) in1[4]) * in2[3] +
534 ((uint128_t) in1[5]) * in2[2] +
535 ((uint128_t) in1[6]) * in2[1] + ((uint128_t) in1[7]) * in2[0];
537 out[8] = ((uint128_t) in1[0]) * in2[8] +
538 ((uint128_t) in1[1]) * in2[7] +
539 ((uint128_t) in1[2]) * in2[6] +
540 ((uint128_t) in1[3]) * in2[5] +
541 ((uint128_t) in1[4]) * in2[4] +
542 ((uint128_t) in1[5]) * in2[3] +
543 ((uint128_t) in1[6]) * in2[2] +
544 ((uint128_t) in1[7]) * in2[1] + ((uint128_t) in1[8]) * in2[0];
548 out[0] += ((uint128_t) in1[1]) * in2x2[8] +
549 ((uint128_t) in1[2]) * in2x2[7] +
550 ((uint128_t) in1[3]) * in2x2[6] +
551 ((uint128_t) in1[4]) * in2x2[5] +
552 ((uint128_t) in1[5]) * in2x2[4] +
553 ((uint128_t) in1[6]) * in2x2[3] +
554 ((uint128_t) in1[7]) * in2x2[2] + ((uint128_t) in1[8]) * in2x2[1];
556 out[1] += ((uint128_t) in1[2]) * in2x2[8] +
557 ((uint128_t) in1[3]) * in2x2[7] +
558 ((uint128_t) in1[4]) * in2x2[6] +
559 ((uint128_t) in1[5]) * in2x2[5] +
560 ((uint128_t) in1[6]) * in2x2[4] +
561 ((uint128_t) in1[7]) * in2x2[3] + ((uint128_t) in1[8]) * in2x2[2];
563 out[2] += ((uint128_t) in1[3]) * in2x2[8] +
564 ((uint128_t) in1[4]) * in2x2[7] +
565 ((uint128_t) in1[5]) * in2x2[6] +
566 ((uint128_t) in1[6]) * in2x2[5] +
567 ((uint128_t) in1[7]) * in2x2[4] + ((uint128_t) in1[8]) * in2x2[3];
569 out[3] += ((uint128_t) in1[4]) * in2x2[8] +
570 ((uint128_t) in1[5]) * in2x2[7] +
571 ((uint128_t) in1[6]) * in2x2[6] +
572 ((uint128_t) in1[7]) * in2x2[5] + ((uint128_t) in1[8]) * in2x2[4];
574 out[4] += ((uint128_t) in1[5]) * in2x2[8] +
575 ((uint128_t) in1[6]) * in2x2[7] +
576 ((uint128_t) in1[7]) * in2x2[6] + ((uint128_t) in1[8]) * in2x2[5];
578 out[5] += ((uint128_t) in1[6]) * in2x2[8] +
579 ((uint128_t) in1[7]) * in2x2[7] + ((uint128_t) in1[8]) * in2x2[6];
581 out[6] += ((uint128_t) in1[7]) * in2x2[8] +
582 ((uint128_t) in1[8]) * in2x2[7];
584 out[7] += ((uint128_t) in1[8]) * in2x2[8];
678 static void felem_mul_reduce(felem out, const felem in1, const felem in2)
681 felem_mul(tmp, in1, in2);