26 |
|
|
27 |
%macro cglobal 1 |
%macro cglobal 1 |
28 |
%ifdef PREFIX |
%ifdef PREFIX |
29 |
|
%ifdef MARK_FUNCS |
30 |
|
global _%1:function %1.endfunc-%1 |
31 |
|
%define %1 _%1:function %1.endfunc-%1 |
32 |
|
%else |
33 |
global _%1 |
global _%1 |
34 |
%define %1 _%1 |
%define %1 _%1 |
35 |
|
%endif |
36 |
|
%else |
37 |
|
%ifdef MARK_FUNCS |
38 |
|
global %1:function %1.endfunc-%1 |
39 |
%else |
%else |
40 |
global %1 |
global %1 |
41 |
%endif |
%endif |
42 |
|
%endif |
43 |
%endmacro |
%endmacro |
44 |
|
|
45 |
;============================================================================= |
;============================================================================= |
62 |
cglobal interpolate8x8_halfpel_v_xmm |
cglobal interpolate8x8_halfpel_v_xmm |
63 |
cglobal interpolate8x8_halfpel_hv_xmm |
cglobal interpolate8x8_halfpel_hv_xmm |
64 |
|
|
65 |
|
cglobal interpolate8x8_halfpel_add_xmm |
66 |
|
cglobal interpolate8x8_halfpel_h_add_xmm |
67 |
|
cglobal interpolate8x8_halfpel_v_add_xmm |
68 |
|
cglobal interpolate8x8_halfpel_hv_add_xmm |
69 |
|
|
70 |
;=========================================================================== |
;=========================================================================== |
71 |
; |
; |
72 |
; void interpolate8x8_halfpel_h_xmm(uint8_t * const dst, |
; void interpolate8x8_halfpel_h_xmm(uint8_t * const dst, |
137 |
lea ecx,[ecx+2*edx] |
lea ecx,[ecx+2*edx] |
138 |
COPY_H_SSE_RND1 |
COPY_H_SSE_RND1 |
139 |
ret |
ret |
140 |
|
.endfunc |
141 |
|
|
142 |
;=========================================================================== |
;=========================================================================== |
143 |
; |
; |
212 |
lea ecx,[ecx+2*edx] |
lea ecx,[ecx+2*edx] |
213 |
COPY_V_SSE_RND1 |
COPY_V_SSE_RND1 |
214 |
ret |
ret |
215 |
|
.endfunc |
216 |
|
|
217 |
;=========================================================================== |
;=========================================================================== |
218 |
; |
; |
351 |
add ecx, edx |
add ecx, edx |
352 |
COPY_HV_SSE_RND1 |
COPY_HV_SSE_RND1 |
353 |
ret |
ret |
354 |
|
.endfunc |
355 |
|
|
356 |
|
;=========================================================================== |
357 |
|
; |
358 |
|
; The next functions combine both source halfpel interpolation step and the |
359 |
|
; averaging (with rouding) step to avoid wasting memory bandwidth computing |
360 |
|
; intermediate halfpel images and then averaging them. |
361 |
|
; |
362 |
|
;=========================================================================== |
363 |
|
|
364 |
|
%macro PROLOG0 0 |
365 |
|
mov ecx, [esp+ 4] ; Dst |
366 |
|
mov eax, [esp+ 8] ; Src |
367 |
|
mov edx, [esp+12] ; BpS |
368 |
|
%endmacro |
369 |
|
%macro PROLOG1 0 |
370 |
|
PROLOG0 |
371 |
|
test dword [esp+16], 1; Rounding? |
372 |
|
%endmacro |
373 |
|
%macro EPILOG 0 |
374 |
|
ret |
375 |
|
%endmacro |
376 |
|
|
377 |
|
;=========================================================================== |
378 |
|
; |
379 |
|
; void interpolate8x8_halfpel_add_xmm(uint8_t * const dst, |
380 |
|
; const uint8_t * const src, |
381 |
|
; const uint32_t stride, |
382 |
|
; const uint32_t rounding); |
383 |
|
; |
384 |
|
; |
385 |
|
;=========================================================================== |
386 |
|
|
387 |
|
%macro ADD_FF 2 |
388 |
|
movq mm0, [eax+%1] |
389 |
|
movq mm1, [eax+%2] |
390 |
|
;;--- |
391 |
|
;; movq mm2, mm0 |
392 |
|
;; movq mm3, mm1 |
393 |
|
;;--- |
394 |
|
pavgb mm0, [ecx+%1] |
395 |
|
pavgb mm1, [ecx+%2] |
396 |
|
;;-- |
397 |
|
;; por mm2, [ecx+%1] |
398 |
|
;; por mm3, [ecx+%2] |
399 |
|
;; pand mm2, [mmx_one] |
400 |
|
;; pand mm3, [mmx_one] |
401 |
|
;; psubsb mm0, mm2 |
402 |
|
;; psubsb mm1, mm3 |
403 |
|
;;-- |
404 |
|
movq [ecx+%1], mm0 |
405 |
|
movq [ecx+%2], mm1 |
406 |
|
%endmacro |
407 |
|
|
408 |
|
ALIGN 16 |
409 |
|
interpolate8x8_halfpel_add_xmm: ; 23c |
410 |
|
PROLOG1 |
411 |
|
ADD_FF 0, edx |
412 |
|
lea eax,[eax+2*edx] |
413 |
|
lea ecx,[ecx+2*edx] |
414 |
|
ADD_FF 0, edx |
415 |
|
lea eax,[eax+2*edx] |
416 |
|
lea ecx,[ecx+2*edx] |
417 |
|
ADD_FF 0, edx |
418 |
|
lea eax,[eax+2*edx] |
419 |
|
lea ecx,[ecx+2*edx] |
420 |
|
ADD_FF 0, edx |
421 |
|
EPILOG |
422 |
|
.endfunc |
423 |
|
|
424 |
|
;=========================================================================== |
425 |
|
; |
426 |
|
; void interpolate8x8_halfpel_h_add_xmm(uint8_t * const dst, |
427 |
|
; const uint8_t * const src, |
428 |
|
; const uint32_t stride, |
429 |
|
; const uint32_t rounding); |
430 |
|
; |
431 |
|
; |
432 |
|
;=========================================================================== |
433 |
|
|
434 |
|
|
435 |
|
%macro ADD_FH_RND0 2 |
436 |
|
movq mm0, [eax+%1] |
437 |
|
movq mm1, [eax+%2] |
438 |
|
pavgb mm0, [eax+%1+1] |
439 |
|
pavgb mm1, [eax+%2+1] |
440 |
|
pavgb mm0, [ecx+%1] |
441 |
|
pavgb mm1, [ecx+%2] |
442 |
|
movq [ecx+%1],mm0 |
443 |
|
movq [ecx+%2],mm1 |
444 |
|
%endmacro |
445 |
|
|
446 |
|
%macro ADD_FH_RND1 2 |
447 |
|
movq mm0, [eax+%1] |
448 |
|
movq mm1, [eax+%2] |
449 |
|
movq mm4, mm0 |
450 |
|
movq mm5, mm1 |
451 |
|
movq mm2, [eax+%1+1] |
452 |
|
movq mm3, [eax+%2+1] |
453 |
|
pavgb mm0, mm2 |
454 |
|
; lea ?? |
455 |
|
pxor mm2, mm4 |
456 |
|
pavgb mm1, mm3 |
457 |
|
pxor mm3, mm5 |
458 |
|
pand mm2, [mmx_one] |
459 |
|
pand mm3, [mmx_one] |
460 |
|
psubb mm0, mm2 |
461 |
|
psubb mm1, mm3 |
462 |
|
pavgb mm0, [ecx+%1] |
463 |
|
pavgb mm1, [ecx+%2] |
464 |
|
movq [ecx+%1],mm0 |
465 |
|
movq [ecx+%2],mm1 |
466 |
|
%endmacro |
467 |
|
|
468 |
|
ALIGN 16 |
469 |
|
interpolate8x8_halfpel_h_add_xmm: ; 32c |
470 |
|
PROLOG1 |
471 |
|
jnz near .Loop1 |
472 |
|
ADD_FH_RND0 0, edx |
473 |
|
lea eax,[eax+2*edx] |
474 |
|
lea ecx,[ecx+2*edx] |
475 |
|
ADD_FH_RND0 0, edx |
476 |
|
lea eax,[eax+2*edx] |
477 |
|
lea ecx,[ecx+2*edx] |
478 |
|
ADD_FH_RND0 0, edx |
479 |
|
lea eax,[eax+2*edx] |
480 |
|
lea ecx,[ecx+2*edx] |
481 |
|
ADD_FH_RND0 0, edx |
482 |
|
EPILOG |
483 |
|
|
484 |
|
.Loop1 |
485 |
|
; we use: (i+j)/2 = ( i+j+1 )/2 - (i^j)&1 |
486 |
|
; movq mm7, [mmx_one] |
487 |
|
ADD_FH_RND1 0, edx |
488 |
|
lea eax,[eax+2*edx] |
489 |
|
lea ecx,[ecx+2*edx] |
490 |
|
ADD_FH_RND1 0, edx |
491 |
|
lea eax,[eax+2*edx] |
492 |
|
lea ecx,[ecx+2*edx] |
493 |
|
ADD_FH_RND1 0, edx |
494 |
|
lea eax,[eax+2*edx] |
495 |
|
lea ecx,[ecx+2*edx] |
496 |
|
ADD_FH_RND1 0, edx |
497 |
|
EPILOG |
498 |
|
.endfunc |
499 |
|
|
500 |
|
|
501 |
|
;=========================================================================== |
502 |
|
; |
503 |
|
; void interpolate8x8_halfpel_v_add_xmm(uint8_t * const dst, |
504 |
|
; const uint8_t * const src, |
505 |
|
; const uint32_t stride, |
506 |
|
; const uint32_t rounding); |
507 |
|
; |
508 |
|
; |
509 |
|
;=========================================================================== |
510 |
|
|
511 |
|
%macro ADD_8_HF_RND0 0 |
512 |
|
movq mm0, [eax] |
513 |
|
movq mm1, [eax+edx] |
514 |
|
pavgb mm0, mm1 |
515 |
|
pavgb mm1, [eax+2*edx] |
516 |
|
lea eax,[eax+2*edx] |
517 |
|
pavgb mm0, [ecx] |
518 |
|
pavgb mm1, [ecx+edx] |
519 |
|
movq [ecx],mm0 |
520 |
|
movq [ecx+edx],mm1 |
521 |
|
%endmacro |
522 |
|
|
523 |
|
%macro ADD_8_HF_RND1 0 |
524 |
|
movq mm1, [eax+edx] |
525 |
|
movq mm2, [eax+2*edx] |
526 |
|
lea eax,[eax+2*edx] |
527 |
|
movq mm4, mm0 |
528 |
|
movq mm5, mm1 |
529 |
|
pavgb mm0, mm1 |
530 |
|
pxor mm4, mm1 |
531 |
|
pavgb mm1, mm2 |
532 |
|
pxor mm5, mm2 |
533 |
|
pand mm4, mm7 ; lsb's of (i^j)... |
534 |
|
pand mm5, mm7 ; lsb's of (i^j)... |
535 |
|
psubb mm0, mm4 ; ...are substracted from result of pavgb |
536 |
|
pavgb mm0, [ecx] |
537 |
|
movq [ecx], mm0 |
538 |
|
psubb mm1, mm5 ; ...are substracted from result of pavgb |
539 |
|
pavgb mm1, [ecx+edx] |
540 |
|
movq [ecx+edx], mm1 |
541 |
|
%endmacro |
542 |
|
|
543 |
|
ALIGN 16 |
544 |
|
interpolate8x8_halfpel_v_add_xmm: |
545 |
|
PROLOG1 |
546 |
|
|
547 |
|
jnz near .Loop1 |
548 |
|
pxor mm7, mm7 ; this is a NOP |
549 |
|
|
550 |
|
ADD_8_HF_RND0 |
551 |
|
lea ecx,[ecx+2*edx] |
552 |
|
ADD_8_HF_RND0 |
553 |
|
lea ecx,[ecx+2*edx] |
554 |
|
ADD_8_HF_RND0 |
555 |
|
lea ecx,[ecx+2*edx] |
556 |
|
ADD_8_HF_RND0 |
557 |
|
EPILOG |
558 |
|
|
559 |
|
.Loop1 |
560 |
|
movq mm0, [eax] ; loop invariant |
561 |
|
movq mm7, [mmx_one] |
562 |
|
|
563 |
|
ADD_8_HF_RND1 |
564 |
|
movq mm0, mm2 |
565 |
|
lea ecx,[ecx+2*edx] |
566 |
|
ADD_8_HF_RND1 |
567 |
|
movq mm0, mm2 |
568 |
|
lea ecx,[ecx+2*edx] |
569 |
|
ADD_8_HF_RND1 |
570 |
|
movq mm0, mm2 |
571 |
|
lea ecx,[ecx+2*edx] |
572 |
|
ADD_8_HF_RND1 |
573 |
|
EPILOG |
574 |
|
.endfunc |
575 |
|
|
576 |
|
; The trick is to correct the result of 'pavgb' with some combination of the |
577 |
|
; lsb's of the 4 input values i,j,k,l, and their intermediate 'pavgb' (s and t). |
578 |
|
; The boolean relations are: |
579 |
|
; (i+j+k+l+3)/4 = (s+t+1)/2 - (ij&kl)&st |
580 |
|
; (i+j+k+l+2)/4 = (s+t+1)/2 - (ij|kl)&st |
581 |
|
; (i+j+k+l+1)/4 = (s+t+1)/2 - (ij&kl)|st |
582 |
|
; (i+j+k+l+0)/4 = (s+t+1)/2 - (ij|kl)|st |
583 |
|
; with s=(i+j+1)/2, t=(k+l+1)/2, ij = i^j, kl = k^l, st = s^t. |
584 |
|
|
585 |
|
; Moreover, we process 2 lines at a times, for better overlapping (~15% faster). |
586 |
|
|
587 |
|
;=========================================================================== |
588 |
|
; |
589 |
|
; void interpolate8x8_halfpel_hv_add_xmm(uint8_t * const dst, |
590 |
|
; const uint8_t * const src, |
591 |
|
; const uint32_t stride, |
592 |
|
; const uint32_t rounding); |
593 |
|
; |
594 |
|
; |
595 |
|
;=========================================================================== |
596 |
|
|
597 |
|
%macro ADD_HH_RND0 0 |
598 |
|
lea eax,[eax+edx] |
599 |
|
|
600 |
|
movq mm0, [eax] |
601 |
|
movq mm1, [eax+1] |
602 |
|
|
603 |
|
movq mm6, mm0 |
604 |
|
pavgb mm0, mm1 ; mm0=(j+k+1)/2. preserved for next step |
605 |
|
lea eax,[eax+edx] |
606 |
|
pxor mm1, mm6 ; mm1=(j^k). preserved for next step |
607 |
|
|
608 |
|
por mm3, mm1 ; ij |= jk |
609 |
|
movq mm6, mm2 |
610 |
|
pxor mm6, mm0 ; mm6 = s^t |
611 |
|
pand mm3, mm6 ; (ij|jk) &= st |
612 |
|
pavgb mm2, mm0 ; mm2 = (s+t+1)/2 |
613 |
|
pand mm3, mm7 ; mask lsb |
614 |
|
psubb mm2, mm3 ; apply. |
615 |
|
|
616 |
|
pavgb mm2, [ecx] |
617 |
|
movq [ecx], mm2 |
618 |
|
|
619 |
|
movq mm2, [eax] |
620 |
|
movq mm3, [eax+1] |
621 |
|
movq mm6, mm2 |
622 |
|
pavgb mm2, mm3 ; preserved for next iteration |
623 |
|
lea ecx,[ecx+edx] |
624 |
|
pxor mm3, mm6 ; preserved for next iteration |
625 |
|
|
626 |
|
por mm1, mm3 |
627 |
|
movq mm6, mm0 |
628 |
|
pxor mm6, mm2 |
629 |
|
pand mm1, mm6 |
630 |
|
pavgb mm0, mm2 |
631 |
|
|
632 |
|
pand mm1, mm7 |
633 |
|
psubb mm0, mm1 |
634 |
|
|
635 |
|
pavgb mm0, [ecx] |
636 |
|
movq [ecx], mm0 |
637 |
|
%endmacro |
638 |
|
|
639 |
|
%macro ADD_HH_RND1 0 |
640 |
|
lea eax,[eax+edx] |
641 |
|
|
642 |
|
movq mm0, [eax] |
643 |
|
movq mm1, [eax+1] |
644 |
|
|
645 |
|
movq mm6, mm0 |
646 |
|
pavgb mm0, mm1 ; mm0=(j+k+1)/2. preserved for next step |
647 |
|
lea eax,[eax+edx] |
648 |
|
pxor mm1, mm6 ; mm1=(j^k). preserved for next step |
649 |
|
|
650 |
|
pand mm3, mm1 |
651 |
|
movq mm6, mm2 |
652 |
|
pxor mm6, mm0 |
653 |
|
por mm3, mm6 |
654 |
|
pavgb mm2, mm0 |
655 |
|
pand mm3, mm7 |
656 |
|
psubb mm2, mm3 |
657 |
|
|
658 |
|
pavgb mm2, [ecx] |
659 |
|
movq [ecx], mm2 |
660 |
|
|
661 |
|
movq mm2, [eax] |
662 |
|
movq mm3, [eax+1] |
663 |
|
movq mm6, mm2 |
664 |
|
pavgb mm2, mm3 ; preserved for next iteration |
665 |
|
lea ecx,[ecx+edx] |
666 |
|
pxor mm3, mm6 ; preserved for next iteration |
667 |
|
|
668 |
|
pand mm1, mm3 |
669 |
|
movq mm6, mm0 |
670 |
|
pxor mm6, mm2 |
671 |
|
por mm1, mm6 |
672 |
|
pavgb mm0, mm2 |
673 |
|
pand mm1, mm7 |
674 |
|
psubb mm0, mm1 |
675 |
|
|
676 |
|
pavgb mm0, [ecx] |
677 |
|
movq [ecx], mm0 |
678 |
|
%endmacro |
679 |
|
|
680 |
|
ALIGN 16 |
681 |
|
interpolate8x8_halfpel_hv_add_xmm: |
682 |
|
PROLOG1 |
683 |
|
|
684 |
|
movq mm7, [mmx_one] |
685 |
|
|
686 |
|
; loop invariants: mm2=(i+j+1)/2 and mm3= i^j |
687 |
|
movq mm2, [eax] |
688 |
|
movq mm3, [eax+1] |
689 |
|
movq mm6, mm2 |
690 |
|
pavgb mm2, mm3 |
691 |
|
pxor mm3, mm6 ; mm2/mm3 ready |
692 |
|
|
693 |
|
jnz near .Loop1 |
694 |
|
|
695 |
|
ADD_HH_RND0 |
696 |
|
add ecx, edx |
697 |
|
ADD_HH_RND0 |
698 |
|
add ecx, edx |
699 |
|
ADD_HH_RND0 |
700 |
|
add ecx, edx |
701 |
|
ADD_HH_RND0 |
702 |
|
EPILOG |
703 |
|
|
704 |
|
.Loop1 |
705 |
|
ADD_HH_RND1 |
706 |
|
add ecx, edx |
707 |
|
ADD_HH_RND1 |
708 |
|
add ecx, edx |
709 |
|
ADD_HH_RND1 |
710 |
|
add ecx, edx |
711 |
|
ADD_HH_RND1 |
712 |
|
|
713 |
|
EPILOG |
714 |
|
.endfunc |
715 |
|
|