32 |
; * |
; * |
33 |
; * History: |
; * History: |
34 |
; * |
; * |
35 |
|
; * 14.06.2002 mmx+xmm dequant_* funcs revamped -Skal- |
36 |
|
; * 24.02.2002 sse2 quant_intra / dequant_intra (have to use movdqu ???) |
37 |
|
; * 17.04.2002 sse2 quant_inter / dequant_inter |
38 |
; * 26.12.2001 minor bug fixes, dequant saturate, further optimization |
; * 26.12.2001 minor bug fixes, dequant saturate, further optimization |
39 |
; * 19.11.2001 quant_inter_mmx now returns sum of abs. coefficient values |
; * 19.11.2001 quant_inter_mmx now returns sum of abs. coefficient values |
40 |
; * 04.11.2001 nasm version; (c)2001 peter ross <pross@cs.rmit.edu.au> |
; * 04.11.2001 nasm version; (c)2001 peter ross <pross@cs.rmit.edu.au> |
61 |
%endif |
%endif |
62 |
%endmacro |
%endmacro |
63 |
|
|
64 |
plus_one times 4 dw 1 |
align 16 |
65 |
|
|
66 |
|
plus_one times 8 dw 1 |
67 |
|
|
68 |
;=========================================================================== |
;=========================================================================== |
69 |
; |
; |
75 |
times 4 dw %1 / 2 |
times 4 dw %1 / 2 |
76 |
%endmacro |
%endmacro |
77 |
|
|
78 |
align ALIGN |
align 16 |
79 |
mmx_sub |
mmx_sub |
80 |
MMX_SUB 1 |
MMX_SUB 1 |
81 |
MMX_SUB 2 |
MMX_SUB 2 |
125 |
times 4 dw (1 << 16) / (%1 * 2) + 1 |
times 4 dw (1 << 16) / (%1 * 2) + 1 |
126 |
%endmacro |
%endmacro |
127 |
|
|
128 |
align ALIGN |
align 16 |
129 |
mmx_div |
mmx_div |
130 |
MMX_DIV 1 |
MMX_DIV 1 |
131 |
MMX_DIV 2 |
MMX_DIV 2 |
175 |
%endif |
%endif |
176 |
%endmacro |
%endmacro |
177 |
|
|
178 |
align ALIGN |
align 16 |
179 |
mmx_add |
mmx_add |
180 |
MMX_ADD 1 |
MMX_ADD 1 |
181 |
MMX_ADD 2 |
MMX_ADD 2 |
220 |
times 4 dw %1 * 2 |
times 4 dw %1 * 2 |
221 |
%endmacro |
%endmacro |
222 |
|
|
223 |
align ALIGN |
align 16 |
224 |
mmx_mul |
mmx_mul |
225 |
MMX_MUL 1 |
MMX_MUL 1 |
226 |
MMX_MUL 2 |
MMX_MUL 2 |
261 |
; |
; |
262 |
;=========================================================================== |
;=========================================================================== |
263 |
|
|
264 |
align ALIGN |
align 8 |
265 |
mmx_32768_minus_2048 times 4 dw (32768-2048) |
mmx_32768_minus_2048 times 4 dw (32768-2048) |
266 |
mmx_32767_minus_2047 times 4 dw (32767-2047) |
mmx_32767_minus_2047 times 4 dw (32767-2047) |
267 |
|
|
268 |
|
align 16 |
269 |
|
mmx_2047 times 4 dw 2047 |
270 |
|
|
271 |
|
align 16 |
272 |
|
sse2_pos_2047 times 8 dw 2047 |
273 |
|
sse2_neg_2048 times 8 dw -2048 |
274 |
|
|
275 |
|
|
276 |
section .text |
section .text |
277 |
|
|
383 |
|
|
384 |
;=========================================================================== |
;=========================================================================== |
385 |
; |
; |
386 |
|
; void quant_intra_sse2(int16_t * coeff, |
387 |
|
; const int16_t const * data, |
388 |
|
; const uint32_t quant, |
389 |
|
; const uint32_t dcscalar); |
390 |
|
; |
391 |
|
;=========================================================================== |
392 |
|
|
393 |
|
align ALIGN |
394 |
|
cglobal quant_intra_sse2 |
395 |
|
quant_intra_sse2 |
396 |
|
|
397 |
|
push esi |
398 |
|
push edi |
399 |
|
|
400 |
|
mov edi, [esp + 8 + 4] ; coeff |
401 |
|
mov esi, [esp + 8 + 8] ; data |
402 |
|
mov eax, [esp + 8 + 12] ; quant |
403 |
|
|
404 |
|
xor ecx, ecx |
405 |
|
cmp al, 1 |
406 |
|
jz near .qas2_q1loop |
407 |
|
|
408 |
|
.qas2_not1 |
409 |
|
movq mm7, [mmx_div + eax*8 - 8] |
410 |
|
movq2dq xmm7, mm7 |
411 |
|
movlhps xmm7, xmm7 |
412 |
|
|
413 |
|
align 16 |
414 |
|
.qas2_loop |
415 |
|
movdqa xmm0, [esi + ecx*8] ; xmm0 = [1st] |
416 |
|
movdqa xmm3, [esi + ecx*8 + 16] ; xmm3 = [2nd] |
417 |
|
pxor xmm1, xmm1 |
418 |
|
pxor xmm4, xmm4 |
419 |
|
pcmpgtw xmm1, xmm0 |
420 |
|
pcmpgtw xmm4, xmm3 |
421 |
|
pxor xmm0, xmm1 |
422 |
|
pxor xmm3, xmm4 |
423 |
|
psubw xmm0, xmm1 |
424 |
|
psubw xmm3, xmm4 |
425 |
|
pmulhw xmm0, xmm7 |
426 |
|
pmulhw xmm3, xmm7 |
427 |
|
pxor xmm0, xmm1 |
428 |
|
pxor xmm3, xmm4 |
429 |
|
psubw xmm0, xmm1 |
430 |
|
psubw xmm3, xmm4 |
431 |
|
movdqa [edi + ecx*8], xmm0 |
432 |
|
movdqa [edi + ecx*8 + 16], xmm3 |
433 |
|
|
434 |
|
add ecx, 4 |
435 |
|
cmp ecx, 16 |
436 |
|
jnz .qas2_loop |
437 |
|
|
438 |
|
.qas2_done |
439 |
|
mov ecx, [esp + 8 + 16] ; dcscalar |
440 |
|
mov edx, ecx |
441 |
|
movsx eax, word [esi] |
442 |
|
shr edx, 1 |
443 |
|
cmp eax, 0 |
444 |
|
jg .qas2_gtzero |
445 |
|
|
446 |
|
sub eax, edx |
447 |
|
jmp short .qas2_mul |
448 |
|
.qas2_gtzero |
449 |
|
add eax, edx |
450 |
|
.qas2_mul |
451 |
|
cdq |
452 |
|
idiv ecx |
453 |
|
|
454 |
|
mov [edi], ax |
455 |
|
|
456 |
|
pop edi |
457 |
|
pop esi |
458 |
|
|
459 |
|
ret |
460 |
|
|
461 |
|
align 16 |
462 |
|
.qas2_q1loop |
463 |
|
movdqa xmm0, [esi + ecx*8] ; xmm0 = [1st] |
464 |
|
movdqa xmm3, [esi + ecx*8 + 16] ; xmm3 = [2nd] |
465 |
|
pxor xmm1, xmm1 |
466 |
|
pxor xmm4, xmm4 |
467 |
|
pcmpgtw xmm1, xmm0 |
468 |
|
pcmpgtw xmm4, xmm3 |
469 |
|
pxor xmm0, xmm1 |
470 |
|
pxor xmm3, xmm4 |
471 |
|
psubw xmm0, xmm1 |
472 |
|
psubw xmm3, xmm4 |
473 |
|
psrlw xmm0, 1 |
474 |
|
psrlw xmm3, 1 |
475 |
|
pxor xmm0, xmm1 |
476 |
|
pxor xmm3, xmm4 |
477 |
|
psubw xmm0, xmm1 |
478 |
|
psubw xmm3, xmm4 |
479 |
|
movdqa [edi + ecx*8], xmm0 |
480 |
|
movdqa [edi + ecx*8 + 16], xmm3 |
481 |
|
|
482 |
|
add ecx, 4 |
483 |
|
cmp ecx, 16 |
484 |
|
jnz .qas2_q1loop |
485 |
|
jmp near .qas2_done |
486 |
|
|
487 |
|
|
488 |
|
|
489 |
|
;=========================================================================== |
490 |
|
; |
491 |
; uint32_t quant_inter_mmx(int16_t * coeff, |
; uint32_t quant_inter_mmx(int16_t * coeff, |
492 |
; const int16_t const * data, |
; const int16_t const * data, |
493 |
; const uint32_t quant); |
; const uint32_t quant); |
593 |
|
|
594 |
;=========================================================================== |
;=========================================================================== |
595 |
; |
; |
596 |
|
; uint32_t quant_inter_sse2(int16_t * coeff, |
597 |
|
; const int16_t const * data, |
598 |
|
; const uint32_t quant); |
599 |
|
; |
600 |
|
;=========================================================================== |
601 |
|
|
602 |
|
align 16 |
603 |
|
cglobal quant_inter_sse2 |
604 |
|
quant_inter_sse2 |
605 |
|
|
606 |
|
push esi |
607 |
|
push edi |
608 |
|
|
609 |
|
mov edi, [esp + 8 + 4] ; coeff |
610 |
|
mov esi, [esp + 8 + 8] ; data |
611 |
|
mov eax, [esp + 8 + 12] ; quant |
612 |
|
|
613 |
|
xor ecx, ecx |
614 |
|
|
615 |
|
pxor xmm5, xmm5 ; sum |
616 |
|
|
617 |
|
movq mm0, [mmx_sub + eax*8 - 8] ; sub |
618 |
|
movq2dq xmm6, mm0 ; load into low 8 bytes |
619 |
|
movlhps xmm6, xmm6 ; duplicate into high 8 bytes |
620 |
|
|
621 |
|
cmp al, 1 |
622 |
|
jz near .qes2_q1loop |
623 |
|
|
624 |
|
.qes2_not1 |
625 |
|
movq mm0, [mmx_div + eax*8 - 8] ; divider |
626 |
|
movq2dq xmm7, mm0 |
627 |
|
movlhps xmm7, xmm7 |
628 |
|
|
629 |
|
align 16 |
630 |
|
.qes2_loop |
631 |
|
movdqa xmm0, [esi + ecx*8] ; xmm0 = [1st] |
632 |
|
movdqa xmm3, [esi + ecx*8 + 16] ; xmm3 = [2nd] |
633 |
|
pxor xmm1, xmm1 |
634 |
|
pxor xmm4, xmm4 |
635 |
|
pcmpgtw xmm1, xmm0 |
636 |
|
pcmpgtw xmm4, xmm3 |
637 |
|
pxor xmm0, xmm1 |
638 |
|
pxor xmm3, xmm4 |
639 |
|
psubw xmm0, xmm1 |
640 |
|
psubw xmm3, xmm4 |
641 |
|
psubusw xmm0, xmm6 |
642 |
|
psubusw xmm3, xmm6 |
643 |
|
pmulhw xmm0, xmm7 |
644 |
|
pmulhw xmm3, xmm7 |
645 |
|
paddw xmm5, xmm0 |
646 |
|
pxor xmm0, xmm1 |
647 |
|
paddw xmm5, xmm3 |
648 |
|
pxor xmm3, xmm4 |
649 |
|
psubw xmm0, xmm1 |
650 |
|
psubw xmm3, xmm4 |
651 |
|
movdqa [edi + ecx*8], xmm0 |
652 |
|
movdqa [edi + ecx*8 + 16], xmm3 |
653 |
|
|
654 |
|
add ecx, 4 |
655 |
|
cmp ecx, 16 |
656 |
|
jnz .qes2_loop |
657 |
|
|
658 |
|
.qes2_done |
659 |
|
movdqu xmm6, [plus_one] |
660 |
|
pmaddwd xmm5, xmm6 |
661 |
|
movhlps xmm6, xmm5 |
662 |
|
paddd xmm5, xmm6 |
663 |
|
movdq2q mm0, xmm5 |
664 |
|
|
665 |
|
movq mm5, mm0 |
666 |
|
psrlq mm5, 32 |
667 |
|
paddd mm0, mm5 |
668 |
|
movd eax, mm0 ; return sum |
669 |
|
|
670 |
|
pop edi |
671 |
|
pop esi |
672 |
|
|
673 |
|
ret |
674 |
|
|
675 |
|
align 16 |
676 |
|
.qes2_q1loop |
677 |
|
movdqa xmm0, [esi + ecx*8] ; xmm0 = [1st] |
678 |
|
movdqa xmm3, [esi + ecx*8 + 16] ; xmm3 = [2nd] |
679 |
|
pxor xmm1, xmm1 |
680 |
|
pxor xmm4, xmm4 |
681 |
|
pcmpgtw xmm1, xmm0 |
682 |
|
pcmpgtw xmm4, xmm3 |
683 |
|
pxor xmm0, xmm1 |
684 |
|
pxor xmm3, xmm4 |
685 |
|
psubw xmm0, xmm1 |
686 |
|
psubw xmm3, xmm4 |
687 |
|
psubusw xmm0, xmm6 |
688 |
|
psubusw xmm3, xmm6 |
689 |
|
psrlw xmm0, 1 |
690 |
|
psrlw xmm3, 1 |
691 |
|
paddw xmm5, xmm0 |
692 |
|
pxor xmm0, xmm1 |
693 |
|
paddw xmm5, xmm3 |
694 |
|
pxor xmm3, xmm4 |
695 |
|
psubw xmm0, xmm1 |
696 |
|
psubw xmm3, xmm4 |
697 |
|
movdqa [edi + ecx*8], xmm0 |
698 |
|
movdqa [edi + ecx*8 + 16], xmm3 |
699 |
|
|
700 |
|
add ecx,4 |
701 |
|
cmp ecx,16 |
702 |
|
jnz .qes2_q1loop |
703 |
|
jmp .qes2_done |
704 |
|
|
705 |
|
|
706 |
|
;=========================================================================== |
707 |
|
; |
708 |
; void dequant_intra_mmx(int16_t *data, |
; void dequant_intra_mmx(int16_t *data, |
709 |
; const int16_t const *coeff, |
; const int16_t const *coeff, |
710 |
; const uint32_t quant, |
; const uint32_t quant, |
712 |
; |
; |
713 |
;=========================================================================== |
;=========================================================================== |
714 |
|
|
715 |
|
; note: we only saturate to +2047 *before* restoring the sign. |
716 |
|
; Hence, final clamp really is [-2048,2047] |
717 |
|
|
718 |
align ALIGN |
align ALIGN |
719 |
cglobal dequant_intra_mmx |
cglobal dequant_intra_mmx |
720 |
dequant_intra_mmx |
dequant_intra_mmx: |
721 |
|
|
722 |
|
mov edx, [esp+ 4] ; data |
723 |
|
mov ecx, [esp+ 8] ; coeff |
724 |
|
mov eax, [esp+12] ; quant |
725 |
|
movq mm6, [mmx_add + eax*8 - 8] ; quant or quant-1 |
726 |
|
movq mm7, [mmx_mul + eax*8 - 8] ; 2*quant |
727 |
|
mov eax, -16 |
728 |
|
|
729 |
|
align ALIGN |
730 |
|
.loop |
731 |
|
movq mm0, [ecx+8*eax+8*16] ; c = coeff[i] |
732 |
|
movq mm3, [ecx+8*eax+8*16 + 8] ; c' = coeff[i+1] |
733 |
|
pxor mm1, mm1 |
734 |
|
pxor mm4, mm4 |
735 |
|
pcmpgtw mm1, mm0 ; sign(c) |
736 |
|
pcmpgtw mm4, mm3 ; sign(c') |
737 |
|
pxor mm2, mm2 |
738 |
|
pxor mm5, mm5 |
739 |
|
pcmpeqw mm2, mm0 ; c is zero |
740 |
|
pcmpeqw mm5, mm3 ; c' is zero |
741 |
|
pandn mm2, mm6 ; offset = isZero ? 0 : quant_add |
742 |
|
pandn mm5, mm6 |
743 |
|
pxor mm0, mm1 ; negate if negative |
744 |
|
pxor mm3, mm4 ; negate if negative |
745 |
|
psubw mm0, mm1 |
746 |
|
psubw mm3, mm4 |
747 |
|
pmullw mm0, mm7 ; *= 2Q |
748 |
|
pmullw mm3, mm7 ; *= 2Q |
749 |
|
paddw mm0, mm2 ; + offset |
750 |
|
paddw mm3, mm5 ; + offset |
751 |
|
paddw mm0, mm1 ; negate back |
752 |
|
paddw mm3, mm4 ; negate back |
753 |
|
|
754 |
|
; saturates to +2047 |
755 |
|
movq mm2, [mmx_32767_minus_2047] |
756 |
|
add eax, 2 |
757 |
|
paddsw mm0, mm2 |
758 |
|
paddsw mm3, mm2 |
759 |
|
psubsw mm0, mm2 |
760 |
|
psubsw mm3, mm2 |
761 |
|
|
762 |
|
pxor mm0, mm1 |
763 |
|
pxor mm3, mm4 |
764 |
|
movq [edx + 8*eax + 8*16 - 2*8], mm0 |
765 |
|
movq [edx + 8*eax + 8*16+8 - 2*8], mm3 |
766 |
|
jnz near .loop |
767 |
|
|
768 |
|
; deal with DC |
769 |
|
|
770 |
|
movd mm0, [ecx] |
771 |
|
pmullw mm0, [esp+16] ; dcscalar |
772 |
|
movq mm2, [mmx_32767_minus_2047] |
773 |
|
paddsw mm0, mm2 |
774 |
|
psubsw mm0, mm2 |
775 |
|
movq mm3, [mmx_32768_minus_2048] |
776 |
|
psubsw mm0, mm3 |
777 |
|
paddsw mm0, mm3 |
778 |
|
movd eax, mm0 |
779 |
|
mov [edx], ax |
780 |
|
|
781 |
|
ret |
782 |
|
|
783 |
|
;=========================================================================== |
784 |
|
; |
785 |
|
; void dequant_intra_xmm(int16_t *data, |
786 |
|
; const int16_t const *coeff, |
787 |
|
; const uint32_t quant, |
788 |
|
; const uint32_t dcscalar); |
789 |
|
; |
790 |
|
;=========================================================================== |
791 |
|
|
792 |
|
; this is the same as dequant_inter_mmx, except that we're |
793 |
|
; saturating using 'pminsw' (saves 2 cycles/loop => ~5% faster) |
794 |
|
|
795 |
|
align ALIGN |
796 |
|
cglobal dequant_intra_xmm |
797 |
|
dequant_intra_xmm: |
798 |
|
|
799 |
|
mov edx, [esp+ 4] ; data |
800 |
|
mov ecx, [esp+ 8] ; coeff |
801 |
|
mov eax, [esp+12] ; quant |
802 |
|
movq mm6, [mmx_add + eax*8 - 8] ; quant or quant-1 |
803 |
|
movq mm7, [mmx_mul + eax*8 - 8] ; 2*quant |
804 |
|
mov eax, -16 |
805 |
|
|
806 |
|
align ALIGN |
807 |
|
.loop |
808 |
|
movq mm0, [ecx+8*eax+8*16] ; c = coeff[i] |
809 |
|
movq mm3, [ecx+8*eax+8*16 + 8] ; c' = coeff[i+1] |
810 |
|
pxor mm1, mm1 |
811 |
|
pxor mm4, mm4 |
812 |
|
pcmpgtw mm1, mm0 ; sign(c) |
813 |
|
pcmpgtw mm4, mm3 ; sign(c') |
814 |
|
pxor mm2, mm2 |
815 |
|
pxor mm5, mm5 |
816 |
|
pcmpeqw mm2, mm0 ; c is zero |
817 |
|
pcmpeqw mm5, mm3 ; c' is zero |
818 |
|
pandn mm2, mm6 ; offset = isZero ? 0 : quant_add |
819 |
|
pandn mm5, mm6 |
820 |
|
pxor mm0, mm1 ; negate if negative |
821 |
|
pxor mm3, mm4 ; negate if negative |
822 |
|
psubw mm0, mm1 |
823 |
|
psubw mm3, mm4 |
824 |
|
pmullw mm0, mm7 ; *= 2Q |
825 |
|
pmullw mm3, mm7 ; *= 2Q |
826 |
|
paddw mm0, mm2 ; + offset |
827 |
|
paddw mm3, mm5 ; + offset |
828 |
|
paddw mm0, mm1 ; negate back |
829 |
|
paddw mm3, mm4 ; negate back |
830 |
|
|
831 |
|
; saturates to +2047 |
832 |
|
movq mm2, [mmx_2047] |
833 |
|
pminsw mm0, mm2 |
834 |
|
add eax, 2 |
835 |
|
pminsw mm3, mm2 |
836 |
|
|
837 |
|
pxor mm0, mm1 |
838 |
|
pxor mm3, mm4 |
839 |
|
movq [edx + 8*eax + 8*16 - 2*8], mm0 |
840 |
|
movq [edx + 8*eax + 8*16+8 - 2*8], mm3 |
841 |
|
jnz near .loop |
842 |
|
|
843 |
|
; deal with DC |
844 |
|
|
845 |
|
movd mm0, [ecx] |
846 |
|
pmullw mm0, [esp+16] ; dcscalar |
847 |
|
movq mm2, [mmx_32767_minus_2047] |
848 |
|
paddsw mm0, mm2 |
849 |
|
psubsw mm0, mm2 |
850 |
|
movq mm2, [mmx_32768_minus_2048] |
851 |
|
psubsw mm0, mm2 |
852 |
|
paddsw mm0, mm2 |
853 |
|
movd eax, mm0 |
854 |
|
mov [edx], ax |
855 |
|
|
856 |
|
ret |
857 |
|
|
858 |
|
;=========================================================================== |
859 |
|
; |
860 |
|
; void dequant_intra_sse2(int16_t *data, |
861 |
|
; const int16_t const *coeff, |
862 |
|
; const uint32_t quant, |
863 |
|
; const uint32_t dcscalar); |
864 |
|
; |
865 |
|
;=========================================================================== |
866 |
|
|
867 |
|
align 16 |
868 |
|
cglobal dequant_intra_sse2 |
869 |
|
dequant_intra_sse2: |
870 |
|
|
871 |
push esi |
push esi |
872 |
push edi |
push edi |
877 |
|
|
878 |
movq mm6, [mmx_add + eax * 8 - 8] |
movq mm6, [mmx_add + eax * 8 - 8] |
879 |
movq mm7, [mmx_mul + eax * 8 - 8] |
movq mm7, [mmx_mul + eax * 8 - 8] |
880 |
|
movq2dq xmm6, mm6 |
881 |
|
movq2dq xmm7, mm7 |
882 |
|
movlhps xmm6, xmm6 |
883 |
|
movlhps xmm7, xmm7 |
884 |
|
|
885 |
xor eax, eax |
xor eax, eax |
886 |
|
|
887 |
align ALIGN |
align 16 |
888 |
.loop |
.das2_loop |
889 |
movq mm0, [esi + 8*eax] ; mm0 = [coeff] |
movdqa xmm0, [esi + eax*8] |
890 |
movq mm3, [esi + 8*eax + 8] ; |
movdqa xmm3, [esi + eax*8 + 16] |
891 |
pxor mm1, mm1 ; mm1 = 0 |
pxor xmm1, xmm1 |
892 |
pxor mm4, mm4 ; |
pxor xmm4, xmm4 |
893 |
pcmpgtw mm1, mm0 ; mm1 = (0 > mm0) |
pcmpgtw xmm1, xmm0 |
894 |
pcmpgtw mm4, mm3 ; |
pcmpgtw xmm4, xmm3 |
895 |
pxor mm2, mm2 ; mm2 = 0 |
pxor xmm2, xmm2 |
896 |
pxor mm5, mm5 ; |
pxor xmm5, xmm5 |
897 |
pcmpeqw mm2, mm0 ; mm2 = (0 == mm0) |
pcmpeqw xmm2, xmm0 |
898 |
pcmpeqw mm5, mm3 ; |
pcmpeqw xmm5, xmm3 |
899 |
pandn mm2, mm6 ; mm2 = (iszero ? 0 : add) |
pandn xmm2, xmm6 |
900 |
pandn mm5, mm6 ; |
pandn xmm5, xmm6 |
901 |
pxor mm0, mm1 ; mm0 = |mm0| |
pxor xmm0, xmm1 |
902 |
pxor mm3, mm4 ; |
pxor xmm3, xmm4 |
903 |
psubw mm0, mm1 ; displace |
psubw xmm0, xmm1 |
904 |
psubw mm3, mm4 ; |
psubw xmm3, xmm4 |
905 |
pmullw mm0, mm7 ; mm0 *= 2Q |
pmullw xmm0, xmm7 |
906 |
pmullw mm3, mm7 ; |
pmullw xmm3, xmm7 |
907 |
paddw mm0, mm2 ; mm0 += mm2 (add) |
paddw xmm0, xmm2 |
908 |
paddw mm3, mm5 ; |
paddw xmm3, xmm5 |
909 |
pxor mm0, mm1 ; mm0 *= sign(mm0) |
pxor xmm0, xmm1 |
910 |
pxor mm3, mm4 ; |
pxor xmm3, xmm4 |
911 |
psubw mm0, mm1 ; undisplace |
psubw xmm0, xmm1 |
912 |
psubw mm3, mm4 |
psubw xmm3, xmm4 |
913 |
|
|
914 |
%ifdef SATURATE |
%ifdef SATURATE |
915 |
movq mm2, [mmx_32767_minus_2047] |
movdqu xmm2, [sse2_pos_2047] |
916 |
movq mm4, [mmx_32768_minus_2048] |
movdqu xmm4, [sse2_neg_2048] |
917 |
paddsw mm0, mm2 |
pminsw xmm0, xmm2 |
918 |
paddsw mm3, mm2 |
pminsw xmm3, xmm2 |
919 |
psubsw mm0, mm2 |
pmaxsw xmm0, xmm4 |
920 |
psubsw mm3, mm2 |
pmaxsw xmm3, xmm4 |
|
psubsw mm0, mm4 |
|
|
psubsw mm3, mm4 |
|
|
paddsw mm0, mm4 |
|
|
paddsw mm3, mm4 |
|
921 |
%endif |
%endif |
922 |
|
|
923 |
movq [edi + 8*eax], mm0 ; [data] = mm0 |
movdqa [edi + eax*8], xmm0 |
924 |
movq [edi + 8*eax + 8], mm3 |
movdqa [edi + eax*8 + 16], xmm3 |
925 |
|
|
926 |
add eax, 2 |
add eax, 4 |
927 |
cmp eax, 16 |
cmp eax, 16 |
928 |
jnz near .loop |
jnz near .das2_loop |
929 |
|
|
930 |
mov ax, [esi] ; ax = data[0] |
mov ax, [esi] ; ax = data[0] |
931 |
imul ax, [esp + 8 + 16] ; eax = data[0] * dcscalar |
imul ax, [esp + 8 + 16] ; eax = data[0] * dcscalar |
932 |
|
|
933 |
%ifdef SATURATE |
%ifdef SATURATE |
934 |
cmp ax, -2048 |
cmp ax, -2048 |
935 |
jl .set_n2048 |
jl .das2_set_n2048 |
936 |
cmp ax, 2047 |
cmp ax, 2047 |
937 |
jg .set_2047 |
jg .das2_set_2047 |
938 |
%endif |
%endif |
939 |
mov [edi], ax |
mov [edi], ax |
940 |
|
|
943 |
ret |
ret |
944 |
|
|
945 |
%ifdef SATURATE |
%ifdef SATURATE |
946 |
align ALIGN |
align 16 |
947 |
.set_n2048 |
.das2_set_n2048 |
948 |
mov word [edi], -2048 |
mov word [edi], -2048 |
949 |
pop edi |
pop edi |
950 |
pop esi |
pop esi |
951 |
ret |
ret |
952 |
|
|
953 |
align ALIGN |
align 16 |
954 |
.set_2047 |
.das2_set_2047 |
955 |
mov word [edi], 2047 |
mov word [edi], 2047 |
956 |
pop edi |
pop edi |
957 |
pop esi |
pop esi |
968 |
|
|
969 |
align ALIGN |
align ALIGN |
970 |
cglobal dequant_inter_mmx |
cglobal dequant_inter_mmx |
971 |
dequant_inter_mmx |
dequant_inter_mmx: |
972 |
|
|
973 |
|
mov edx, [esp+ 4] ; data |
974 |
|
mov ecx, [esp+ 8] ; coeff |
975 |
|
mov eax, [esp+12] ; quant |
976 |
|
movq mm6, [mmx_add + eax*8 - 8] ; quant or quant-1 |
977 |
|
movq mm7, [mmx_mul + eax*8 - 8] ; 2*quant |
978 |
|
mov eax, -16 |
979 |
|
|
980 |
|
align ALIGN |
981 |
|
.loop |
982 |
|
movq mm0, [ecx+8*eax+8*16] ; c = coeff[i] |
983 |
|
movq mm3, [ecx+8*eax+8*16 + 8] ; c' = coeff[i+1] |
984 |
|
pxor mm1, mm1 |
985 |
|
pxor mm4, mm4 |
986 |
|
pcmpgtw mm1, mm0 ; sign(c) |
987 |
|
pcmpgtw mm4, mm3 ; sign(c') |
988 |
|
pxor mm2, mm2 |
989 |
|
pxor mm5, mm5 |
990 |
|
pcmpeqw mm2, mm0 ; c is zero |
991 |
|
pcmpeqw mm5, mm3 ; c' is zero |
992 |
|
pandn mm2, mm6 ; offset = isZero ? 0 : quant_add |
993 |
|
pandn mm5, mm6 |
994 |
|
pxor mm0, mm1 ; negate if negative |
995 |
|
pxor mm3, mm4 ; negate if negative |
996 |
|
psubw mm0, mm1 |
997 |
|
psubw mm3, mm4 |
998 |
|
pmullw mm0, mm7 ; *= 2Q |
999 |
|
pmullw mm3, mm7 ; *= 2Q |
1000 |
|
paddw mm0, mm2 ; + offset |
1001 |
|
paddw mm3, mm5 ; + offset |
1002 |
|
paddw mm0, mm1 ; negate back |
1003 |
|
paddw mm3, mm4 ; negate back |
1004 |
|
|
1005 |
|
; saturates to +2047 |
1006 |
|
movq mm2, [mmx_32767_minus_2047] |
1007 |
|
add eax, 2 |
1008 |
|
paddsw mm0, mm2 |
1009 |
|
paddsw mm3, mm2 |
1010 |
|
psubsw mm0, mm2 |
1011 |
|
psubsw mm3, mm2 |
1012 |
|
|
1013 |
|
pxor mm0, mm1 |
1014 |
|
pxor mm3, mm4 |
1015 |
|
movq [edx + 8*eax + 8*16 - 2*8], mm0 |
1016 |
|
movq [edx + 8*eax + 8*16+8 - 2*8], mm3 |
1017 |
|
jnz near .loop |
1018 |
|
|
1019 |
|
ret |
1020 |
|
|
1021 |
|
;=========================================================================== |
1022 |
|
; |
1023 |
|
; void dequant_inter_xmm(int16_t * data, |
1024 |
|
; const int16_t * const coeff, |
1025 |
|
; const uint32_t quant); |
1026 |
|
; |
1027 |
|
;=========================================================================== |
1028 |
|
|
1029 |
|
; this is the same as dequant_inter_mmx, |
1030 |
|
; except that we're saturating using 'pminsw' (saves 2 cycles/loop) |
1031 |
|
|
1032 |
|
align ALIGN |
1033 |
|
cglobal dequant_inter_xmm |
1034 |
|
dequant_inter_xmm: |
1035 |
|
|
1036 |
|
mov edx, [esp+ 4] ; data |
1037 |
|
mov ecx, [esp+ 8] ; coeff |
1038 |
|
mov eax, [esp+12] ; quant |
1039 |
|
movq mm6, [mmx_add + eax*8 - 8] ; quant or quant-1 |
1040 |
|
movq mm7, [mmx_mul + eax*8 - 8] ; 2*quant |
1041 |
|
mov eax, -16 |
1042 |
|
|
1043 |
|
align ALIGN |
1044 |
|
.loop |
1045 |
|
movq mm0, [ecx+8*eax+8*16] ; c = coeff[i] |
1046 |
|
movq mm3, [ecx+8*eax+8*16 + 8] ; c' = coeff[i+1] |
1047 |
|
pxor mm1, mm1 |
1048 |
|
pxor mm4, mm4 |
1049 |
|
pcmpgtw mm1, mm0 ; sign(c) |
1050 |
|
pcmpgtw mm4, mm3 ; sign(c') |
1051 |
|
pxor mm2, mm2 |
1052 |
|
pxor mm5, mm5 |
1053 |
|
pcmpeqw mm2, mm0 ; c is zero |
1054 |
|
pcmpeqw mm5, mm3 ; c' is zero |
1055 |
|
pandn mm2, mm6 ; offset = isZero ? 0 : quant_add |
1056 |
|
pandn mm5, mm6 |
1057 |
|
pxor mm0, mm1 ; negate if negative |
1058 |
|
pxor mm3, mm4 ; negate if negative |
1059 |
|
psubw mm0, mm1 |
1060 |
|
psubw mm3, mm4 |
1061 |
|
pmullw mm0, mm7 ; *= 2Q |
1062 |
|
pmullw mm3, mm7 ; *= 2Q |
1063 |
|
paddw mm0, mm2 ; + offset |
1064 |
|
paddw mm3, mm5 ; + offset |
1065 |
|
paddw mm0, mm1 ; start restoring sign |
1066 |
|
paddw mm3, mm4 ; start restoring sign |
1067 |
|
|
1068 |
|
; saturates to +2047 |
1069 |
|
movq mm2, [mmx_2047] |
1070 |
|
pminsw mm0, mm2 |
1071 |
|
add eax, 2 |
1072 |
|
pminsw mm3, mm2 |
1073 |
|
|
1074 |
|
pxor mm0, mm1 ; finish restoring sign |
1075 |
|
pxor mm3, mm4 ; finish restoring sign |
1076 |
|
movq [edx + 8*eax + 8*16 - 2*8], mm0 |
1077 |
|
movq [edx + 8*eax + 8*16+8 - 2*8], mm3 |
1078 |
|
jnz near .loop |
1079 |
|
|
1080 |
|
ret |
1081 |
|
|
1082 |
|
;=========================================================================== |
1083 |
|
; |
1084 |
|
; void dequant_inter_sse2(int16_t * data, |
1085 |
|
; const int16_t * const coeff, |
1086 |
|
; const uint32_t quant); |
1087 |
|
; |
1088 |
|
;=========================================================================== |
1089 |
|
|
1090 |
|
align 16 |
1091 |
|
cglobal dequant_inter_sse2 |
1092 |
|
dequant_inter_sse2 |
1093 |
|
|
1094 |
push esi |
push esi |
1095 |
push edi |
push edi |
1100 |
movq mm6, [mmx_add + eax * 8 - 8] |
movq mm6, [mmx_add + eax * 8 - 8] |
1101 |
movq mm7, [mmx_mul + eax * 8 - 8] |
movq mm7, [mmx_mul + eax * 8 - 8] |
1102 |
|
|
1103 |
|
movq2dq xmm6, mm6 |
1104 |
|
movq2dq xmm7, mm7 |
1105 |
|
movlhps xmm6, xmm6 |
1106 |
|
movlhps xmm7, xmm7 |
1107 |
|
|
1108 |
xor eax, eax |
xor eax, eax |
1109 |
|
|
1110 |
align ALIGN |
align 16 |
1111 |
.loop |
.des2_loop |
1112 |
movq mm0, [esi + 8*eax] ; mm0 = [coeff] |
movdqa xmm0, [esi + eax*8] ; xmm0 = [coeff] |
1113 |
movq mm3, [esi + 8*eax + 8] ; |
movdqa xmm3, [esi + eax*8 + 16] |
1114 |
pxor mm1, mm1 ; mm1 = 0 |
pxor xmm1, xmm1 |
1115 |
pxor mm4, mm4 ; |
pxor xmm4, xmm4 |
1116 |
pcmpgtw mm1, mm0 ; mm1 = (0 > mm0) |
pcmpgtw xmm1, xmm0 |
1117 |
pcmpgtw mm4, mm3 ; |
pcmpgtw xmm4, xmm3 |
1118 |
pxor mm2, mm2 ; mm2 = 0 |
pxor xmm2, xmm2 |
1119 |
pxor mm5, mm5 ; |
pxor xmm5, xmm5 |
1120 |
pcmpeqw mm2, mm0 ; mm2 = (0 == mm0) |
pcmpeqw xmm2, xmm0 |
1121 |
pcmpeqw mm5, mm3 ; |
pcmpeqw xmm5, xmm3 |
1122 |
pandn mm2, mm6 ; mm2 = (iszero ? 0 : add) |
pandn xmm2, xmm6 |
1123 |
pandn mm5, mm6 ; |
pandn xmm5, xmm6 |
1124 |
pxor mm0, mm1 ; mm0 = |mm0| |
pxor xmm0, xmm1 |
1125 |
pxor mm3, mm4 ; |
pxor xmm3, xmm4 |
1126 |
psubw mm0, mm1 ; displace |
psubw xmm0, xmm1 |
1127 |
psubw mm3, mm4 ; |
psubw xmm3, xmm4 |
1128 |
pmullw mm0, mm7 ; mm0 *= 2Q |
pmullw xmm0, xmm7 |
1129 |
pmullw mm3, mm7 ; |
pmullw xmm3, xmm7 |
1130 |
paddw mm0, mm2 ; mm0 += mm2 (add) |
paddw xmm0, xmm2 |
1131 |
paddw mm3, mm5 ; |
paddw xmm3, xmm5 |
1132 |
pxor mm0, mm1 ; mm0 *= sign(mm0) |
pxor xmm0, xmm1 |
1133 |
pxor mm3, mm4 ; |
pxor xmm3, xmm4 |
1134 |
psubw mm0, mm1 ; undisplace |
psubw xmm0, xmm1 |
1135 |
psubw mm3, mm4 |
psubw xmm3, xmm4 |
1136 |
|
|
1137 |
%ifdef SATURATE |
%ifdef SATURATE |
1138 |
movq mm2, [mmx_32767_minus_2047] |
movdqu xmm2, [sse2_pos_2047] |
1139 |
movq mm4, [mmx_32768_minus_2048] |
movdqu xmm4, [sse2_neg_2048] |
1140 |
paddsw mm0, mm2 |
pminsw xmm0, xmm2 |
1141 |
paddsw mm3, mm2 |
pminsw xmm3, xmm2 |
1142 |
psubsw mm0, mm2 |
pmaxsw xmm0, xmm4 |
1143 |
psubsw mm3, mm2 |
pmaxsw xmm3, xmm4 |
|
psubsw mm0, mm4 |
|
|
psubsw mm3, mm4 |
|
|
paddsw mm0, mm4 |
|
|
paddsw mm3, mm4 |
|
1144 |
%endif |
%endif |
1145 |
|
|
1146 |
movq [edi + 8*eax], mm0 |
movdqa [edi + eax*8], xmm0 |
1147 |
movq [edi + 8*eax + 8], mm3 |
movdqa [edi + eax*8 + 16], xmm3 |
1148 |
|
|
1149 |
add eax, 2 |
add eax, 4 |
1150 |
cmp eax, 16 |
cmp eax, 16 |
1151 |
jnz near .loop |
jnz near .des2_loop |
1152 |
|
|
1153 |
pop edi |
pop edi |
1154 |
pop esi |
pop esi |