1 |
;/************************************************************************** |
;/***************************************************************************** |
2 |
; * |
; * |
3 |
; * XVID MPEG-4 VIDEO CODEC |
; * XVID MPEG-4 VIDEO CODEC |
4 |
; * mmx quantization/dequantization |
; * mmx optimized quantization/dequantization |
5 |
|
; * |
6 |
|
; * Copyright(C) 2002 Peter Ross <pross@xvid.org> |
7 |
|
; * Copyright(C) 2002 Michael Militzer <michael@xvid.org> |
8 |
|
; * Copyright(C) 2002 Pascal Massimino <skal@planet-d.net> |
9 |
; * |
; * |
10 |
; * This program is an implementation of a part of one or more MPEG-4 |
; * This program is an implementation of a part of one or more MPEG-4 |
11 |
; * Video tools as specified in ISO/IEC 14496-2 standard. Those intending |
; * Video tools as specified in ISO/IEC 14496-2 standard. Those intending |
28 |
; * |
; * |
29 |
; * You should have received a copy of the GNU General Public License |
; * You should have received a copy of the GNU General Public License |
30 |
; * along with this program; if not, write to the Free Software |
; * along with this program; if not, write to the Free Software |
31 |
; * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. |
; * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA |
|
; * |
|
|
; *************************************************************************/ |
|
|
|
|
|
;/************************************************************************** |
|
|
; * |
|
|
; * History: |
|
|
; * |
|
|
; * 26.12.2001 minor bug fixes, dequant saturate, further optimization |
|
|
; * 19.11.2001 quant_inter_mmx now returns sum of abs. coefficient values |
|
|
; * 04.11.2001 nasm version; (c)2001 peter ross <pross@cs.rmit.edu.au> |
|
32 |
; * |
; * |
33 |
; *************************************************************************/ |
; *************************************************************************/ |
34 |
|
|
52 |
%endif |
%endif |
53 |
%endmacro |
%endmacro |
54 |
|
|
55 |
plus_one times 4 dw 1 |
align 16 |
56 |
|
|
57 |
|
plus_one times 8 dw 1 |
58 |
|
|
59 |
;=========================================================================== |
;=========================================================================== |
60 |
; |
; |
66 |
times 4 dw %1 / 2 |
times 4 dw %1 / 2 |
67 |
%endmacro |
%endmacro |
68 |
|
|
69 |
align ALIGN |
align 16 |
70 |
mmx_sub |
mmx_sub |
71 |
MMX_SUB 1 |
MMX_SUB 1 |
72 |
MMX_SUB 2 |
MMX_SUB 2 |
116 |
times 4 dw (1 << 16) / (%1 * 2) + 1 |
times 4 dw (1 << 16) / (%1 * 2) + 1 |
117 |
%endmacro |
%endmacro |
118 |
|
|
119 |
align ALIGN |
align 16 |
120 |
mmx_div |
mmx_div |
121 |
MMX_DIV 1 |
MMX_DIV 1 |
122 |
MMX_DIV 2 |
MMX_DIV 2 |
166 |
%endif |
%endif |
167 |
%endmacro |
%endmacro |
168 |
|
|
169 |
align ALIGN |
align 16 |
170 |
mmx_add |
mmx_add |
171 |
MMX_ADD 1 |
MMX_ADD 1 |
172 |
MMX_ADD 2 |
MMX_ADD 2 |
211 |
times 4 dw %1 * 2 |
times 4 dw %1 * 2 |
212 |
%endmacro |
%endmacro |
213 |
|
|
214 |
align ALIGN |
align 16 |
215 |
mmx_mul |
mmx_mul |
216 |
MMX_MUL 1 |
MMX_MUL 1 |
217 |
MMX_MUL 2 |
MMX_MUL 2 |
252 |
; |
; |
253 |
;=========================================================================== |
;=========================================================================== |
254 |
|
|
255 |
align ALIGN |
align 16 |
256 |
|
sse2_2047 times 8 dw 2047 |
257 |
|
|
258 |
|
align 16 |
259 |
|
mmx_2047 times 4 dw 2047 |
260 |
|
|
261 |
|
align 8 |
262 |
mmx_32768_minus_2048 times 4 dw (32768-2048) |
mmx_32768_minus_2048 times 4 dw (32768-2048) |
263 |
mmx_32767_minus_2047 times 4 dw (32767-2047) |
mmx_32767_minus_2047 times 4 dw (32767-2047) |
264 |
|
|
373 |
|
|
374 |
;=========================================================================== |
;=========================================================================== |
375 |
; |
; |
376 |
|
; void quant_intra_sse2(int16_t * coeff, |
377 |
|
; const int16_t const * data, |
378 |
|
; const uint32_t quant, |
379 |
|
; const uint32_t dcscalar); |
380 |
|
; |
381 |
|
;=========================================================================== |
382 |
|
|
383 |
|
align ALIGN |
384 |
|
cglobal quant_intra_sse2 |
385 |
|
quant_intra_sse2 |
386 |
|
|
387 |
|
push esi |
388 |
|
push edi |
389 |
|
|
390 |
|
mov edi, [esp + 8 + 4] ; coeff |
391 |
|
mov esi, [esp + 8 + 8] ; data |
392 |
|
mov eax, [esp + 8 + 12] ; quant |
393 |
|
|
394 |
|
xor ecx, ecx |
395 |
|
cmp al, 1 |
396 |
|
jz near .qas2_q1loop |
397 |
|
|
398 |
|
.qas2_not1 |
399 |
|
movq mm7, [mmx_div + eax*8 - 8] |
400 |
|
movq2dq xmm7, mm7 |
401 |
|
movlhps xmm7, xmm7 |
402 |
|
|
403 |
|
align 16 |
404 |
|
.qas2_loop |
405 |
|
movdqa xmm0, [esi + ecx*8] ; xmm0 = [1st] |
406 |
|
movdqa xmm3, [esi + ecx*8 + 16] ; xmm3 = [2nd] |
407 |
|
pxor xmm1, xmm1 |
408 |
|
pxor xmm4, xmm4 |
409 |
|
pcmpgtw xmm1, xmm0 |
410 |
|
pcmpgtw xmm4, xmm3 |
411 |
|
pxor xmm0, xmm1 |
412 |
|
pxor xmm3, xmm4 |
413 |
|
psubw xmm0, xmm1 |
414 |
|
psubw xmm3, xmm4 |
415 |
|
pmulhw xmm0, xmm7 |
416 |
|
pmulhw xmm3, xmm7 |
417 |
|
pxor xmm0, xmm1 |
418 |
|
pxor xmm3, xmm4 |
419 |
|
psubw xmm0, xmm1 |
420 |
|
psubw xmm3, xmm4 |
421 |
|
movdqa [edi + ecx*8], xmm0 |
422 |
|
movdqa [edi + ecx*8 + 16], xmm3 |
423 |
|
|
424 |
|
add ecx, 4 |
425 |
|
cmp ecx, 16 |
426 |
|
jnz .qas2_loop |
427 |
|
|
428 |
|
.qas2_done |
429 |
|
mov ecx, [esp + 8 + 16] ; dcscalar |
430 |
|
mov edx, ecx |
431 |
|
movsx eax, word [esi] |
432 |
|
shr edx, 1 |
433 |
|
cmp eax, 0 |
434 |
|
jg .qas2_gtzero |
435 |
|
|
436 |
|
sub eax, edx |
437 |
|
jmp short .qas2_mul |
438 |
|
.qas2_gtzero |
439 |
|
add eax, edx |
440 |
|
.qas2_mul |
441 |
|
cdq |
442 |
|
idiv ecx |
443 |
|
|
444 |
|
mov [edi], ax |
445 |
|
|
446 |
|
pop edi |
447 |
|
pop esi |
448 |
|
|
449 |
|
ret |
450 |
|
|
451 |
|
align 16 |
452 |
|
.qas2_q1loop |
453 |
|
movdqa xmm0, [esi + ecx*8] ; xmm0 = [1st] |
454 |
|
movdqa xmm3, [esi + ecx*8 + 16] ; xmm3 = [2nd] |
455 |
|
pxor xmm1, xmm1 |
456 |
|
pxor xmm4, xmm4 |
457 |
|
pcmpgtw xmm1, xmm0 |
458 |
|
pcmpgtw xmm4, xmm3 |
459 |
|
pxor xmm0, xmm1 |
460 |
|
pxor xmm3, xmm4 |
461 |
|
psubw xmm0, xmm1 |
462 |
|
psubw xmm3, xmm4 |
463 |
|
psrlw xmm0, 1 |
464 |
|
psrlw xmm3, 1 |
465 |
|
pxor xmm0, xmm1 |
466 |
|
pxor xmm3, xmm4 |
467 |
|
psubw xmm0, xmm1 |
468 |
|
psubw xmm3, xmm4 |
469 |
|
movdqa [edi + ecx*8], xmm0 |
470 |
|
movdqa [edi + ecx*8 + 16], xmm3 |
471 |
|
|
472 |
|
add ecx, 4 |
473 |
|
cmp ecx, 16 |
474 |
|
jnz .qas2_q1loop |
475 |
|
jmp near .qas2_done |
476 |
|
|
477 |
|
|
478 |
|
|
479 |
|
;=========================================================================== |
480 |
|
; |
481 |
; uint32_t quant_inter_mmx(int16_t * coeff, |
; uint32_t quant_inter_mmx(int16_t * coeff, |
482 |
; const int16_t const * data, |
; const int16_t const * data, |
483 |
; const uint32_t quant); |
; const uint32_t quant); |
583 |
|
|
584 |
;=========================================================================== |
;=========================================================================== |
585 |
; |
; |
586 |
|
; uint32_t quant_inter_sse2(int16_t * coeff, |
587 |
|
; const int16_t const * data, |
588 |
|
; const uint32_t quant); |
589 |
|
; |
590 |
|
;=========================================================================== |
591 |
|
|
592 |
|
align 16 |
593 |
|
cglobal quant_inter_sse2 |
594 |
|
quant_inter_sse2 |
595 |
|
|
596 |
|
push esi |
597 |
|
push edi |
598 |
|
|
599 |
|
mov edi, [esp + 8 + 4] ; coeff |
600 |
|
mov esi, [esp + 8 + 8] ; data |
601 |
|
mov eax, [esp + 8 + 12] ; quant |
602 |
|
|
603 |
|
xor ecx, ecx |
604 |
|
|
605 |
|
pxor xmm5, xmm5 ; sum |
606 |
|
|
607 |
|
movq mm0, [mmx_sub + eax*8 - 8] ; sub |
608 |
|
movq2dq xmm6, mm0 ; load into low 8 bytes |
609 |
|
movlhps xmm6, xmm6 ; duplicate into high 8 bytes |
610 |
|
|
611 |
|
cmp al, 1 |
612 |
|
jz near .qes2_q1loop |
613 |
|
|
614 |
|
.qes2_not1 |
615 |
|
movq mm0, [mmx_div + eax*8 - 8] ; divider |
616 |
|
movq2dq xmm7, mm0 |
617 |
|
movlhps xmm7, xmm7 |
618 |
|
|
619 |
|
align 16 |
620 |
|
.qes2_loop |
621 |
|
movdqa xmm0, [esi + ecx*8] ; xmm0 = [1st] |
622 |
|
movdqa xmm3, [esi + ecx*8 + 16] ; xmm3 = [2nd] |
623 |
|
pxor xmm1, xmm1 |
624 |
|
pxor xmm4, xmm4 |
625 |
|
pcmpgtw xmm1, xmm0 |
626 |
|
pcmpgtw xmm4, xmm3 |
627 |
|
pxor xmm0, xmm1 |
628 |
|
pxor xmm3, xmm4 |
629 |
|
psubw xmm0, xmm1 |
630 |
|
psubw xmm3, xmm4 |
631 |
|
psubusw xmm0, xmm6 |
632 |
|
psubusw xmm3, xmm6 |
633 |
|
pmulhw xmm0, xmm7 |
634 |
|
pmulhw xmm3, xmm7 |
635 |
|
paddw xmm5, xmm0 |
636 |
|
pxor xmm0, xmm1 |
637 |
|
paddw xmm5, xmm3 |
638 |
|
pxor xmm3, xmm4 |
639 |
|
psubw xmm0, xmm1 |
640 |
|
psubw xmm3, xmm4 |
641 |
|
movdqa [edi + ecx*8], xmm0 |
642 |
|
movdqa [edi + ecx*8 + 16], xmm3 |
643 |
|
|
644 |
|
add ecx, 4 |
645 |
|
cmp ecx, 16 |
646 |
|
jnz .qes2_loop |
647 |
|
|
648 |
|
.qes2_done |
649 |
|
movdqu xmm6, [plus_one] |
650 |
|
pmaddwd xmm5, xmm6 |
651 |
|
movhlps xmm6, xmm5 |
652 |
|
paddd xmm5, xmm6 |
653 |
|
movdq2q mm0, xmm5 |
654 |
|
|
655 |
|
movq mm5, mm0 |
656 |
|
psrlq mm5, 32 |
657 |
|
paddd mm0, mm5 |
658 |
|
movd eax, mm0 ; return sum |
659 |
|
|
660 |
|
pop edi |
661 |
|
pop esi |
662 |
|
|
663 |
|
ret |
664 |
|
|
665 |
|
align 16 |
666 |
|
.qes2_q1loop |
667 |
|
movdqa xmm0, [esi + ecx*8] ; xmm0 = [1st] |
668 |
|
movdqa xmm3, [esi + ecx*8 + 16] ; xmm3 = [2nd] |
669 |
|
pxor xmm1, xmm1 |
670 |
|
pxor xmm4, xmm4 |
671 |
|
pcmpgtw xmm1, xmm0 |
672 |
|
pcmpgtw xmm4, xmm3 |
673 |
|
pxor xmm0, xmm1 |
674 |
|
pxor xmm3, xmm4 |
675 |
|
psubw xmm0, xmm1 |
676 |
|
psubw xmm3, xmm4 |
677 |
|
psubusw xmm0, xmm6 |
678 |
|
psubusw xmm3, xmm6 |
679 |
|
psrlw xmm0, 1 |
680 |
|
psrlw xmm3, 1 |
681 |
|
paddw xmm5, xmm0 |
682 |
|
pxor xmm0, xmm1 |
683 |
|
paddw xmm5, xmm3 |
684 |
|
pxor xmm3, xmm4 |
685 |
|
psubw xmm0, xmm1 |
686 |
|
psubw xmm3, xmm4 |
687 |
|
movdqa [edi + ecx*8], xmm0 |
688 |
|
movdqa [edi + ecx*8 + 16], xmm3 |
689 |
|
|
690 |
|
add ecx,4 |
691 |
|
cmp ecx,16 |
692 |
|
jnz .qes2_q1loop |
693 |
|
jmp .qes2_done |
694 |
|
|
695 |
|
|
696 |
|
;=========================================================================== |
697 |
|
; |
698 |
; void dequant_intra_mmx(int16_t *data, |
; void dequant_intra_mmx(int16_t *data, |
699 |
; const int16_t const *coeff, |
; const int16_t const *coeff, |
700 |
; const uint32_t quant, |
; const uint32_t quant, |
702 |
; |
; |
703 |
;=========================================================================== |
;=========================================================================== |
704 |
|
|
705 |
|
; note: we only saturate to +2047 *before* restoring the sign. |
706 |
|
; Hence, final clamp really is [-2048,2047] |
707 |
|
|
708 |
align ALIGN |
align ALIGN |
709 |
cglobal dequant_intra_mmx |
cglobal dequant_intra_mmx |
710 |
dequant_intra_mmx |
dequant_intra_mmx: |
|
|
|
|
push esi |
|
|
push edi |
|
711 |
|
|
712 |
mov edi, [esp + 8 + 4] ; data |
mov edx, [esp+ 4] ; data |
713 |
mov esi, [esp + 8 + 8] ; coeff |
mov ecx, [esp+ 8] ; coeff |
714 |
mov eax, [esp + 8 + 12] ; quant |
mov eax, [esp+12] ; quant |
715 |
|
movq mm6, [mmx_add + eax*8 - 8] ; quant or quant-1 |
716 |
movq mm6, [mmx_add + eax * 8 - 8] |
movq mm7, [mmx_mul + eax*8 - 8] ; 2*quant |
717 |
movq mm7, [mmx_mul + eax * 8 - 8] |
mov eax, -16 |
|
xor eax, eax |
|
718 |
|
|
719 |
align ALIGN |
align ALIGN |
720 |
.loop |
.loop |
721 |
movq mm0, [esi + 8*eax] ; mm0 = [coeff] |
movq mm0, [ecx+8*eax+8*16] ; c = coeff[i] |
722 |
movq mm3, [esi + 8*eax + 8] ; |
movq mm3, [ecx+8*eax+8*16 + 8] ; c' = coeff[i+1] |
723 |
pxor mm1, mm1 ; mm1 = 0 |
pxor mm1, mm1 |
724 |
pxor mm4, mm4 ; |
pxor mm4, mm4 |
725 |
pcmpgtw mm1, mm0 ; mm1 = (0 > mm0) |
pcmpgtw mm1, mm0 ; sign(c) |
726 |
pcmpgtw mm4, mm3 ; |
pcmpgtw mm4, mm3 ; sign(c') |
727 |
pxor mm2, mm2 ; mm2 = 0 |
pxor mm2, mm2 |
728 |
pxor mm5, mm5 ; |
pxor mm5, mm5 |
729 |
pcmpeqw mm2, mm0 ; mm2 = (0 == mm0) |
pcmpeqw mm2, mm0 ; c is zero |
730 |
pcmpeqw mm5, mm3 ; |
pcmpeqw mm5, mm3 ; c' is zero |
731 |
pandn mm2, mm6 ; mm2 = (iszero ? 0 : add) |
pandn mm2, mm6 ; offset = isZero ? 0 : quant_add |
732 |
pandn mm5, mm6 ; |
pandn mm5, mm6 |
733 |
pxor mm0, mm1 ; mm0 = |mm0| |
pxor mm0, mm1 ; negate if negative |
734 |
pxor mm3, mm4 ; |
pxor mm3, mm4 ; negate if negative |
735 |
psubw mm0, mm1 ; displace |
psubw mm0, mm1 |
|
psubw mm3, mm4 ; |
|
|
pmullw mm0, mm7 ; mm0 *= 2Q |
|
|
pmullw mm3, mm7 ; |
|
|
paddw mm0, mm2 ; mm0 += mm2 (add) |
|
|
paddw mm3, mm5 ; |
|
|
pxor mm0, mm1 ; mm0 *= sign(mm0) |
|
|
pxor mm3, mm4 ; |
|
|
psubw mm0, mm1 ; undisplace |
|
736 |
psubw mm3, mm4 |
psubw mm3, mm4 |
737 |
|
pmullw mm0, mm7 ; *= 2Q |
738 |
|
pmullw mm3, mm7 ; *= 2Q |
739 |
|
paddw mm0, mm2 ; + offset |
740 |
|
paddw mm3, mm5 ; + offset |
741 |
|
paddw mm0, mm1 ; negate back |
742 |
|
paddw mm3, mm4 ; negate back |
743 |
|
|
744 |
%ifdef SATURATE |
; saturates to +2047 |
745 |
movq mm2, [mmx_32767_minus_2047] |
movq mm2, [mmx_32767_minus_2047] |
746 |
movq mm4, [mmx_32768_minus_2048] |
add eax, 2 |
747 |
paddsw mm0, mm2 |
paddsw mm0, mm2 |
748 |
paddsw mm3, mm2 |
paddsw mm3, mm2 |
749 |
psubsw mm0, mm2 |
psubsw mm0, mm2 |
750 |
psubsw mm3, mm2 |
psubsw mm3, mm2 |
|
psubsw mm0, mm4 |
|
|
psubsw mm3, mm4 |
|
|
paddsw mm0, mm4 |
|
|
paddsw mm3, mm4 |
|
|
%endif |
|
751 |
|
|
752 |
movq [edi + 8*eax], mm0 ; [data] = mm0 |
pxor mm0, mm1 |
753 |
movq [edi + 8*eax + 8], mm3 |
pxor mm3, mm4 |
754 |
|
movq [edx + 8*eax + 8*16 - 2*8], mm0 |
755 |
|
movq [edx + 8*eax + 8*16+8 - 2*8], mm3 |
756 |
|
jnz near .loop |
757 |
|
|
758 |
|
; deal with DC |
759 |
|
|
760 |
|
movd mm0, [ecx] |
761 |
|
pmullw mm0, [esp+16] ; dcscalar |
762 |
|
movq mm2, [mmx_32767_minus_2047] |
763 |
|
paddsw mm0, mm2 |
764 |
|
psubsw mm0, mm2 |
765 |
|
movq mm3, [mmx_32768_minus_2048] |
766 |
|
psubsw mm0, mm3 |
767 |
|
paddsw mm0, mm3 |
768 |
|
movd eax, mm0 |
769 |
|
mov [edx], ax |
770 |
|
|
771 |
|
ret |
772 |
|
|
773 |
|
;=========================================================================== |
774 |
|
; |
775 |
|
; void dequant_intra_xmm(int16_t *data, |
776 |
|
; const int16_t const *coeff, |
777 |
|
; const uint32_t quant, |
778 |
|
; const uint32_t dcscalar); |
779 |
|
; |
780 |
|
;=========================================================================== |
781 |
|
|
782 |
|
; this is the same as dequant_inter_mmx, except that we're |
783 |
|
; saturating using 'pminsw' (saves 2 cycles/loop => ~5% faster) |
784 |
|
|
785 |
|
align ALIGN |
786 |
|
cglobal dequant_intra_xmm |
787 |
|
dequant_intra_xmm: |
788 |
|
|
789 |
|
mov edx, [esp+ 4] ; data |
790 |
|
mov ecx, [esp+ 8] ; coeff |
791 |
|
mov eax, [esp+12] ; quant |
792 |
|
movq mm6, [mmx_add + eax*8 - 8] ; quant or quant-1 |
793 |
|
movq mm7, [mmx_mul + eax*8 - 8] ; 2*quant |
794 |
|
mov eax, -16 |
795 |
|
|
796 |
|
align ALIGN |
797 |
|
.loop |
798 |
|
movq mm0, [ecx+8*eax+8*16] ; c = coeff[i] |
799 |
|
movq mm3, [ecx+8*eax+8*16 + 8] ; c' = coeff[i+1] |
800 |
|
pxor mm1, mm1 |
801 |
|
pxor mm4, mm4 |
802 |
|
pcmpgtw mm1, mm0 ; sign(c) |
803 |
|
pcmpgtw mm4, mm3 ; sign(c') |
804 |
|
pxor mm2, mm2 |
805 |
|
pxor mm5, mm5 |
806 |
|
pcmpeqw mm2, mm0 ; c is zero |
807 |
|
pcmpeqw mm5, mm3 ; c' is zero |
808 |
|
pandn mm2, mm6 ; offset = isZero ? 0 : quant_add |
809 |
|
pandn mm5, mm6 |
810 |
|
pxor mm0, mm1 ; negate if negative |
811 |
|
pxor mm3, mm4 ; negate if negative |
812 |
|
psubw mm0, mm1 |
813 |
|
psubw mm3, mm4 |
814 |
|
pmullw mm0, mm7 ; *= 2Q |
815 |
|
pmullw mm3, mm7 ; *= 2Q |
816 |
|
paddw mm0, mm2 ; + offset |
817 |
|
paddw mm3, mm5 ; + offset |
818 |
|
paddw mm0, mm1 ; negate back |
819 |
|
paddw mm3, mm4 ; negate back |
820 |
|
|
821 |
|
; saturates to +2047 |
822 |
|
movq mm2, [mmx_2047] |
823 |
|
pminsw mm0, mm2 |
824 |
add eax, 2 |
add eax, 2 |
825 |
cmp eax, 16 |
pminsw mm3, mm2 |
826 |
|
|
827 |
|
pxor mm0, mm1 |
828 |
|
pxor mm3, mm4 |
829 |
|
movq [edx + 8*eax + 8*16 - 2*8], mm0 |
830 |
|
movq [edx + 8*eax + 8*16+8 - 2*8], mm3 |
831 |
jnz near .loop |
jnz near .loop |
832 |
|
|
833 |
mov ax, [esi] ; ax = data[0] |
; deal with DC |
|
imul ax, [esp + 8 + 16] ; eax = data[0] * dcscalar |
|
834 |
|
|
835 |
%ifdef SATURATE |
movd mm0, [ecx] |
836 |
cmp ax, -2048 |
pmullw mm0, [esp+16] ; dcscalar |
837 |
jl .set_n2048 |
movq mm2, [mmx_32767_minus_2047] |
838 |
cmp ax, 2047 |
paddsw mm0, mm2 |
839 |
jg .set_2047 |
psubsw mm0, mm2 |
840 |
%endif |
movq mm2, [mmx_32768_minus_2048] |
841 |
mov [edi], ax |
psubsw mm0, mm2 |
842 |
|
paddsw mm0, mm2 |
843 |
|
movd eax, mm0 |
844 |
|
mov [edx], ax |
845 |
|
|
|
pop edi |
|
|
pop esi |
|
846 |
ret |
ret |
847 |
|
|
848 |
%ifdef SATURATE |
|
849 |
|
;=========================================================================== |
850 |
|
; |
851 |
|
; void dequant_intra_sse2(int16_t *data, |
852 |
|
; const int16_t const *coeff, |
853 |
|
; const uint32_t quant, |
854 |
|
; const uint32_t dcscalar); |
855 |
|
; |
856 |
|
;=========================================================================== |
857 |
align ALIGN |
align ALIGN |
858 |
.set_n2048 |
cglobal dequant_intra_sse2 |
859 |
mov word [edi], -2048 |
dequant_intra_sse2: |
860 |
pop edi |
mov edx, [esp+ 4] ; data |
861 |
pop esi |
mov ecx, [esp+ 8] ; coeff |
862 |
ret |
mov eax, [esp+12] ; quant |
863 |
|
movq mm6, [mmx_add + eax * 8 - 8] |
864 |
|
movq mm7, [mmx_mul + eax * 8 - 8] |
865 |
|
movq2dq xmm6, mm6 |
866 |
|
movq2dq xmm7, mm7 |
867 |
|
movlhps xmm6, xmm6 |
868 |
|
movlhps xmm7, xmm7 |
869 |
|
mov eax, -16 |
870 |
|
|
871 |
align ALIGN |
align ALIGN |
872 |
.set_2047 |
.loop |
873 |
mov word [edi], 2047 |
movdqa xmm0, [ecx + 8*16 + 8*eax] ; c = coeff[i] |
874 |
pop edi |
movdqa xmm3, [ecx + 8*16 + 8*eax+ 16] |
875 |
pop esi |
pxor xmm1, xmm1 |
876 |
|
pxor xmm4, xmm4 |
877 |
|
pcmpgtw xmm1, xmm0 ; sign(c) |
878 |
|
pcmpgtw xmm4, xmm3 |
879 |
|
pxor xmm2, xmm2 |
880 |
|
pxor xmm5, xmm5 |
881 |
|
pcmpeqw xmm2, xmm0 ; c is zero |
882 |
|
pcmpeqw xmm5, xmm3 |
883 |
|
pandn xmm2, xmm6 ; offset = isZero ? 0 : quant_add |
884 |
|
pandn xmm5, xmm6 |
885 |
|
pxor xmm0, xmm1 ; negate if negative |
886 |
|
pxor xmm3, xmm4 |
887 |
|
psubw xmm0, xmm1 |
888 |
|
psubw xmm3, xmm4 |
889 |
|
pmullw xmm0, xmm7 ; *= 2Q |
890 |
|
pmullw xmm3, xmm7 |
891 |
|
paddw xmm0, xmm2 ; + offset |
892 |
|
paddw xmm3, xmm5 |
893 |
|
paddw xmm0, xmm1 ; negate back |
894 |
|
paddw xmm3, xmm4 |
895 |
|
|
896 |
|
; saturates to +2047 |
897 |
|
movdqa xmm2, [sse2_2047] |
898 |
|
pminsw xmm0, xmm2 |
899 |
|
add eax, 4 |
900 |
|
pminsw xmm3, xmm2 |
901 |
|
|
902 |
|
pxor xmm0, xmm1 |
903 |
|
pxor xmm3, xmm4 |
904 |
|
movdqa [edx + 8*16 - 8*4 + 8*eax], xmm0 |
905 |
|
movdqa [edx + 8*16 - 8*4 + 8*eax + 16], xmm3 |
906 |
|
jnz near .loop |
907 |
|
|
908 |
|
; deal with DC |
909 |
|
movd mm0, [ecx] |
910 |
|
pmullw mm0, [esp+16] ; dcscalar |
911 |
|
movq mm2, [mmx_32767_minus_2047] |
912 |
|
paddsw mm0, mm2 |
913 |
|
psubsw mm0, mm2 |
914 |
|
movq mm2, [mmx_32768_minus_2048] |
915 |
|
psubsw mm0, mm2 |
916 |
|
paddsw mm0, mm2 |
917 |
|
movd eax, mm0 |
918 |
|
mov [edx], ax |
919 |
|
|
920 |
ret |
ret |
921 |
%endif |
|
922 |
|
|
923 |
|
|
924 |
|
|
925 |
;=========================================================================== |
;=========================================================================== |
926 |
; |
; |
932 |
|
|
933 |
align ALIGN |
align ALIGN |
934 |
cglobal dequant_inter_mmx |
cglobal dequant_inter_mmx |
935 |
dequant_inter_mmx |
dequant_inter_mmx: |
|
|
|
|
push esi |
|
|
push edi |
|
|
|
|
|
mov edi, [esp + 8 + 4] ; data |
|
|
mov esi, [esp + 8 + 8] ; coeff |
|
|
mov eax, [esp + 8 + 12] ; quant |
|
|
movq mm6, [mmx_add + eax * 8 - 8] |
|
|
movq mm7, [mmx_mul + eax * 8 - 8] |
|
936 |
|
|
937 |
xor eax, eax |
mov edx, [esp+ 4] ; data |
938 |
|
mov ecx, [esp+ 8] ; coeff |
939 |
|
mov eax, [esp+12] ; quant |
940 |
|
movq mm6, [mmx_add + eax*8 - 8] ; quant or quant-1 |
941 |
|
movq mm7, [mmx_mul + eax*8 - 8] ; 2*quant |
942 |
|
mov eax, -16 |
943 |
|
|
944 |
align ALIGN |
align ALIGN |
945 |
.loop |
.loop |
946 |
movq mm0, [esi + 8*eax] ; mm0 = [coeff] |
movq mm0, [ecx+8*eax+8*16] ; c = coeff[i] |
947 |
movq mm3, [esi + 8*eax + 8] ; |
movq mm3, [ecx+8*eax+8*16 + 8] ; c' = coeff[i+1] |
948 |
pxor mm1, mm1 ; mm1 = 0 |
pxor mm1, mm1 |
949 |
pxor mm4, mm4 ; |
pxor mm4, mm4 |
950 |
pcmpgtw mm1, mm0 ; mm1 = (0 > mm0) |
pcmpgtw mm1, mm0 ; sign(c) |
951 |
pcmpgtw mm4, mm3 ; |
pcmpgtw mm4, mm3 ; sign(c') |
952 |
pxor mm2, mm2 ; mm2 = 0 |
pxor mm2, mm2 |
953 |
pxor mm5, mm5 ; |
pxor mm5, mm5 |
954 |
pcmpeqw mm2, mm0 ; mm2 = (0 == mm0) |
pcmpeqw mm2, mm0 ; c is zero |
955 |
pcmpeqw mm5, mm3 ; |
pcmpeqw mm5, mm3 ; c' is zero |
956 |
pandn mm2, mm6 ; mm2 = (iszero ? 0 : add) |
pandn mm2, mm6 ; offset = isZero ? 0 : quant_add |
957 |
pandn mm5, mm6 ; |
pandn mm5, mm6 |
958 |
pxor mm0, mm1 ; mm0 = |mm0| |
pxor mm0, mm1 ; negate if negative |
959 |
pxor mm3, mm4 ; |
pxor mm3, mm4 ; negate if negative |
960 |
psubw mm0, mm1 ; displace |
psubw mm0, mm1 |
|
psubw mm3, mm4 ; |
|
|
pmullw mm0, mm7 ; mm0 *= 2Q |
|
|
pmullw mm3, mm7 ; |
|
|
paddw mm0, mm2 ; mm0 += mm2 (add) |
|
|
paddw mm3, mm5 ; |
|
|
pxor mm0, mm1 ; mm0 *= sign(mm0) |
|
|
pxor mm3, mm4 ; |
|
|
psubw mm0, mm1 ; undisplace |
|
961 |
psubw mm3, mm4 |
psubw mm3, mm4 |
962 |
|
pmullw mm0, mm7 ; *= 2Q |
963 |
|
pmullw mm3, mm7 ; *= 2Q |
964 |
|
paddw mm0, mm2 ; + offset |
965 |
|
paddw mm3, mm5 ; + offset |
966 |
|
paddw mm0, mm1 ; negate back |
967 |
|
paddw mm3, mm4 ; negate back |
968 |
|
|
969 |
%ifdef SATURATE |
; saturates to +2047 |
970 |
movq mm2, [mmx_32767_minus_2047] |
movq mm2, [mmx_32767_minus_2047] |
971 |
movq mm4, [mmx_32768_minus_2048] |
add eax, 2 |
972 |
paddsw mm0, mm2 |
paddsw mm0, mm2 |
973 |
paddsw mm3, mm2 |
paddsw mm3, mm2 |
974 |
psubsw mm0, mm2 |
psubsw mm0, mm2 |
975 |
psubsw mm3, mm2 |
psubsw mm3, mm2 |
|
psubsw mm0, mm4 |
|
|
psubsw mm3, mm4 |
|
|
paddsw mm0, mm4 |
|
|
paddsw mm3, mm4 |
|
|
%endif |
|
976 |
|
|
977 |
movq [edi + 8*eax], mm0 |
pxor mm0, mm1 |
978 |
movq [edi + 8*eax + 8], mm3 |
pxor mm3, mm4 |
979 |
|
movq [edx + 8*eax + 8*16 - 2*8], mm0 |
980 |
|
movq [edx + 8*eax + 8*16+8 - 2*8], mm3 |
981 |
|
jnz near .loop |
982 |
|
|
983 |
|
ret |
984 |
|
|
985 |
|
;=========================================================================== |
986 |
|
; |
987 |
|
; void dequant_inter_xmm(int16_t * data, |
988 |
|
; const int16_t * const coeff, |
989 |
|
; const uint32_t quant); |
990 |
|
; |
991 |
|
;=========================================================================== |
992 |
|
|
993 |
|
; this is the same as dequant_inter_mmx, |
994 |
|
; except that we're saturating using 'pminsw' (saves 2 cycles/loop) |
995 |
|
|
996 |
|
align ALIGN |
997 |
|
cglobal dequant_inter_xmm |
998 |
|
dequant_inter_xmm: |
999 |
|
|
1000 |
|
mov edx, [esp+ 4] ; data |
1001 |
|
mov ecx, [esp+ 8] ; coeff |
1002 |
|
mov eax, [esp+12] ; quant |
1003 |
|
movq mm6, [mmx_add + eax*8 - 8] ; quant or quant-1 |
1004 |
|
movq mm7, [mmx_mul + eax*8 - 8] ; 2*quant |
1005 |
|
mov eax, -16 |
1006 |
|
|
1007 |
|
align ALIGN |
1008 |
|
.loop |
1009 |
|
movq mm0, [ecx+8*eax+8*16] ; c = coeff[i] |
1010 |
|
movq mm3, [ecx+8*eax+8*16 + 8] ; c' = coeff[i+1] |
1011 |
|
pxor mm1, mm1 |
1012 |
|
pxor mm4, mm4 |
1013 |
|
pcmpgtw mm1, mm0 ; sign(c) |
1014 |
|
pcmpgtw mm4, mm3 ; sign(c') |
1015 |
|
pxor mm2, mm2 |
1016 |
|
pxor mm5, mm5 |
1017 |
|
pcmpeqw mm2, mm0 ; c is zero |
1018 |
|
pcmpeqw mm5, mm3 ; c' is zero |
1019 |
|
pandn mm2, mm6 ; offset = isZero ? 0 : quant_add |
1020 |
|
pandn mm5, mm6 |
1021 |
|
pxor mm0, mm1 ; negate if negative |
1022 |
|
pxor mm3, mm4 ; negate if negative |
1023 |
|
psubw mm0, mm1 |
1024 |
|
psubw mm3, mm4 |
1025 |
|
pmullw mm0, mm7 ; *= 2Q |
1026 |
|
pmullw mm3, mm7 ; *= 2Q |
1027 |
|
paddw mm0, mm2 ; + offset |
1028 |
|
paddw mm3, mm5 ; + offset |
1029 |
|
paddw mm0, mm1 ; start restoring sign |
1030 |
|
paddw mm3, mm4 ; start restoring sign |
1031 |
|
|
1032 |
|
; saturates to +2047 |
1033 |
|
movq mm2, [mmx_2047] |
1034 |
|
pminsw mm0, mm2 |
1035 |
add eax, 2 |
add eax, 2 |
1036 |
cmp eax, 16 |
pminsw mm3, mm2 |
1037 |
|
|
1038 |
|
pxor mm0, mm1 ; finish restoring sign |
1039 |
|
pxor mm3, mm4 ; finish restoring sign |
1040 |
|
movq [edx + 8*eax + 8*16 - 2*8], mm0 |
1041 |
|
movq [edx + 8*eax + 8*16+8 - 2*8], mm3 |
1042 |
jnz near .loop |
jnz near .loop |
1043 |
|
|
1044 |
pop edi |
ret |
1045 |
pop esi |
|
1046 |
|
;=========================================================================== |
1047 |
|
; |
1048 |
|
; void dequant_inter_sse2(int16_t * data, |
1049 |
|
; const int16_t * const coeff, |
1050 |
|
; const uint32_t quant); |
1051 |
|
; |
1052 |
|
;=========================================================================== |
1053 |
|
align ALIGN |
1054 |
|
cglobal dequant_inter_sse2 |
1055 |
|
dequant_inter_sse2 |
1056 |
|
mov edx, [esp + 4] ; data |
1057 |
|
mov ecx, [esp + 8] ; coeff |
1058 |
|
mov eax, [esp + 12] ; quant |
1059 |
|
movq mm6, [mmx_add + eax * 8 - 8] |
1060 |
|
movq mm7, [mmx_mul + eax * 8 - 8] |
1061 |
|
movq2dq xmm6, mm6 |
1062 |
|
movq2dq xmm7, mm7 |
1063 |
|
movlhps xmm6, xmm6 |
1064 |
|
movlhps xmm7, xmm7 |
1065 |
|
mov eax, -16 |
1066 |
|
|
1067 |
|
align ALIGN |
1068 |
|
.loop |
1069 |
|
movdqa xmm0, [ecx + 8*16 + 8*eax] ; c = coeff[i] |
1070 |
|
movdqa xmm3, [ecx + 8*16 + 8*eax + 16] |
1071 |
|
|
1072 |
|
pxor xmm1, xmm1 |
1073 |
|
pxor xmm4, xmm4 |
1074 |
|
pcmpgtw xmm1, xmm0 ; sign(c) |
1075 |
|
pcmpgtw xmm4, xmm3 |
1076 |
|
pxor xmm2, xmm2 |
1077 |
|
pxor xmm5, xmm5 |
1078 |
|
pcmpeqw xmm2, xmm0 ; c is zero |
1079 |
|
pcmpeqw xmm5, xmm3 |
1080 |
|
pandn xmm2, xmm6 |
1081 |
|
pandn xmm5, xmm6 |
1082 |
|
pxor xmm0, xmm1 ; negate if negative |
1083 |
|
pxor xmm3, xmm4 |
1084 |
|
psubw xmm0, xmm1 |
1085 |
|
psubw xmm3, xmm4 |
1086 |
|
pmullw xmm0, xmm7 ; *= 2Q |
1087 |
|
pmullw xmm3, xmm7 |
1088 |
|
paddw xmm0, xmm2 ; + offset |
1089 |
|
paddw xmm3, xmm5 |
1090 |
|
|
1091 |
|
paddw xmm0, xmm1 ; start restoring sign |
1092 |
|
paddw xmm3, xmm4 |
1093 |
|
|
1094 |
|
; saturates to +2047 |
1095 |
|
movdqa xmm2, [sse2_2047] |
1096 |
|
pminsw xmm0, xmm2 |
1097 |
|
add eax, 4 |
1098 |
|
pminsw xmm3, xmm2 |
1099 |
|
|
1100 |
|
pxor xmm0, xmm1 ; finish restoring sign |
1101 |
|
pxor xmm3, xmm4 |
1102 |
|
movdqa [edx + 8*16 - 8*4 + 8*eax], xmm0 |
1103 |
|
movdqa [edx + 8*16 - 8*4 + 8*eax + 16], xmm3 |
1104 |
|
jnz near .loop |
1105 |
|
|
1106 |
ret |
ret |