Lines Matching refs:_mm_mul_epu32
247 p->S21.v = _mm_mul_epu32(p->R21.v, FIVE);
248 p->S22.v = _mm_mul_epu32(p->R22.v, FIVE);
249 p->S23.v = _mm_mul_epu32(p->R23.v, FIVE);
250 p->S24.v = _mm_mul_epu32(p->R24.v, FIVE);
301 T0 = _mm_mul_epu32(H0, p->R20.v);
302 T1 = _mm_mul_epu32(H0, p->R21.v);
303 T2 = _mm_mul_epu32(H0, p->R22.v);
304 T3 = _mm_mul_epu32(H0, p->R23.v);
305 T4 = _mm_mul_epu32(H0, p->R24.v);
306 T5 = _mm_mul_epu32(H1, p->S24.v);
307 T6 = _mm_mul_epu32(H1, p->R20.v);
310 T5 = _mm_mul_epu32(H2, p->S23.v);
311 T6 = _mm_mul_epu32(H2, p->S24.v);
314 T5 = _mm_mul_epu32(H3, p->S22.v);
315 T6 = _mm_mul_epu32(H3, p->S23.v);
318 T5 = _mm_mul_epu32(H4, p->S21.v);
319 T6 = _mm_mul_epu32(H4, p->S22.v);
322 T5 = _mm_mul_epu32(H1, p->R21.v);
323 T6 = _mm_mul_epu32(H1, p->R22.v);
326 T5 = _mm_mul_epu32(H2, p->R20.v);
327 T6 = _mm_mul_epu32(H2, p->R21.v);
330 T5 = _mm_mul_epu32(H3, p->S24.v);
331 T6 = _mm_mul_epu32(H3, p->R20.v);
334 T5 = _mm_mul_epu32(H4, p->S23.v);
335 T6 = _mm_mul_epu32(H4, p->S24.v);
338 T5 = _mm_mul_epu32(H1, p->R23.v);
340 T5 = _mm_mul_epu32(H2, p->R22.v);
342 T5 = _mm_mul_epu32(H3, p->R21.v);
344 T5 = _mm_mul_epu32(H4, p->R20.v);
360 T5 = _mm_mul_epu32(M0, p->R20.v);
361 T6 = _mm_mul_epu32(M0, p->R21.v);
364 T5 = _mm_mul_epu32(M1, p->S24.v);
365 T6 = _mm_mul_epu32(M1, p->R20.v);
368 T5 = _mm_mul_epu32(M2, p->S23.v);
369 T6 = _mm_mul_epu32(M2, p->S24.v);
372 T5 = _mm_mul_epu32(M3, p->S22.v);
373 T6 = _mm_mul_epu32(M3, p->S23.v);
376 T5 = _mm_mul_epu32(M4, p->S21.v);
377 T6 = _mm_mul_epu32(M4, p->S22.v);
380 T5 = _mm_mul_epu32(M0, p->R22.v);
381 T6 = _mm_mul_epu32(M0, p->R23.v);
384 T5 = _mm_mul_epu32(M1, p->R21.v);
385 T6 = _mm_mul_epu32(M1, p->R22.v);
388 T5 = _mm_mul_epu32(M2, p->R20.v);
389 T6 = _mm_mul_epu32(M2, p->R21.v);
392 T5 = _mm_mul_epu32(M3, p->S24.v);
393 T6 = _mm_mul_epu32(M3, p->R20.v);
396 T5 = _mm_mul_epu32(M4, p->S23.v);
397 T6 = _mm_mul_epu32(M4, p->S24.v);
400 T5 = _mm_mul_epu32(M0, p->R24.v);
402 T5 = _mm_mul_epu32(M1, p->R23.v);
404 T5 = _mm_mul_epu32(M2, p->R22.v);
406 T5 = _mm_mul_epu32(M3, p->R21.v);
408 T5 = _mm_mul_epu32(M4, p->R20.v);
441 T0 = _mm_add_epi64(T0, _mm_mul_epu32(C2, FIVE));
498 T0 = _mm_mul_epu32(H0, p->R20.v);
499 T1 = _mm_mul_epu32(H0, p->R21.v);
500 T2 = _mm_mul_epu32(H0, p->R22.v);
501 T3 = _mm_mul_epu32(H0, p->R23.v);
502 T4 = _mm_mul_epu32(H0, p->R24.v);
503 T5 = _mm_mul_epu32(H1, p->S24.v);
504 T6 = _mm_mul_epu32(H1, p->R20.v);
507 T5 = _mm_mul_epu32(H2, p->S23.v);
508 T6 = _mm_mul_epu32(H2, p->S24.v);
511 T5 = _mm_mul_epu32(H3, p->S22.v);
512 T6 = _mm_mul_epu32(H3, p->S23.v);
515 T5 = _mm_mul_epu32(H4, p->S21.v);
516 T6 = _mm_mul_epu32(H4, p->S22.v);
519 T5 = _mm_mul_epu32(H1, p->R21.v);
520 T6 = _mm_mul_epu32(H1, p->R22.v);
523 T5 = _mm_mul_epu32(H2, p->R20.v);
524 T6 = _mm_mul_epu32(H2, p->R21.v);
527 T5 = _mm_mul_epu32(H3, p->S24.v);
528 T6 = _mm_mul_epu32(H3, p->R20.v);
531 T5 = _mm_mul_epu32(H4, p->S23.v);
532 T6 = _mm_mul_epu32(H4, p->S24.v);
535 T5 = _mm_mul_epu32(H1, p->R23.v);
537 T5 = _mm_mul_epu32(H2, p->R22.v);
539 T5 = _mm_mul_epu32(H3, p->R21.v);
541 T5 = _mm_mul_epu32(H4, p->R20.v);
574 T0 = _mm_add_epi64(T0, _mm_mul_epu32(C2, FIVE));
611 T0 = _mm_mul_epu32(H0, p->R20.v);
612 T1 = _mm_mul_epu32(H0, p->R21.v);
613 T2 = _mm_mul_epu32(H0, p->R22.v);
614 T3 = _mm_mul_epu32(H0, p->R23.v);
615 T4 = _mm_mul_epu32(H0, p->R24.v);
616 T5 = _mm_mul_epu32(H1, p->S24.v);
617 T6 = _mm_mul_epu32(H1, p->R20.v);
620 T5 = _mm_mul_epu32(H2, p->S23.v);
621 T6 = _mm_mul_epu32(H2, p->S24.v);
624 T5 = _mm_mul_epu32(H3, p->S22.v);
625 T6 = _mm_mul_epu32(H3, p->S23.v);
628 T5 = _mm_mul_epu32(H4, p->S21.v);
629 T6 = _mm_mul_epu32(H4, p->S22.v);
632 T5 = _mm_mul_epu32(H1, p->R21.v);
633 T6 = _mm_mul_epu32(H1, p->R22.v);
636 T5 = _mm_mul_epu32(H2, p->R20.v);
637 T6 = _mm_mul_epu32(H2, p->R21.v);
640 T5 = _mm_mul_epu32(H3, p->S24.v);
641 T6 = _mm_mul_epu32(H3, p->R20.v);
644 T5 = _mm_mul_epu32(H4, p->S23.v);
645 T6 = _mm_mul_epu32(H4, p->S24.v);
648 T5 = _mm_mul_epu32(H1, p->R23.v);
650 T5 = _mm_mul_epu32(H2, p->R22.v);
652 T5 = _mm_mul_epu32(H3, p->R21.v);
654 T5 = _mm_mul_epu32(H4, p->R20.v);
668 T0 = _mm_add_epi64(T0, _mm_mul_epu32(C2, FIVE));
