32 |
; * |
; * |
33 |
; * History: |
; * History: |
34 |
; * |
; * |
35 |
|
; * 24.02.2002 sse2 quant_intra / dequant_intra (have to use movdqu ???) |
36 |
|
; * 17.04.2002 sse2 quant_inter / dequant_inter |
37 |
; * 26.12.2001 minor bug fixes, dequant saturate, further optimization |
; * 26.12.2001 minor bug fixes, dequant saturate, further optimization |
38 |
; * 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 |
39 |
; * 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> |
64 |
|
|
65 |
plus_one times 8 dw 1 |
plus_one times 8 dw 1 |
66 |
|
|
67 |
|
|
68 |
;=========================================================================== |
;=========================================================================== |
69 |
; |
; |
70 |
; subtract by Q/2 table |
; subtract by Q/2 table |
380 |
|
|
381 |
;=========================================================================== |
;=========================================================================== |
382 |
; |
; |
383 |
|
; void quant_intra_sse2(int16_t * coeff, |
384 |
|
; const int16_t const * data, |
385 |
|
; const uint32_t quant, |
386 |
|
; const uint32_t dcscalar); |
387 |
|
; |
388 |
|
;=========================================================================== |
389 |
|
|
390 |
|
align ALIGN |
391 |
|
cglobal quant_intra_sse2 |
392 |
|
quant_intra_sse2 |
393 |
|
|
394 |
|
push esi |
395 |
|
push edi |
396 |
|
|
397 |
|
mov edi, [esp + 8 + 4] ; coeff |
398 |
|
mov esi, [esp + 8 + 8] ; data |
399 |
|
mov eax, [esp + 8 + 12] ; quant |
400 |
|
|
401 |
|
xor ecx, ecx |
402 |
|
cmp al, 1 |
403 |
|
jz near .qas2_q1loop |
404 |
|
|
405 |
|
.qas2_not1 |
406 |
|
movq mm7, [mmx_div + eax*8 - 8] |
407 |
|
movq2dq xmm7, mm7 |
408 |
|
movlhps xmm7, xmm7 |
409 |
|
|
410 |
|
align 16 |
411 |
|
.qas2_loop |
412 |
|
movdqa xmm0, [esi + ecx*8] ; xmm0 = [1st] |
413 |
|
movdqa xmm3, [esi + ecx*8 + 16] ; xmm3 = [2nd] |
414 |
|
pxor xmm1, xmm1 |
415 |
|
pxor xmm4, xmm4 |
416 |
|
pcmpgtw xmm1, xmm0 |
417 |
|
pcmpgtw xmm4, xmm3 |
418 |
|
pxor xmm0, xmm1 |
419 |
|
pxor xmm3, xmm4 |
420 |
|
psubw xmm0, xmm1 |
421 |
|
psubw xmm3, xmm4 |
422 |
|
pmulhw xmm0, xmm7 |
423 |
|
pmulhw xmm3, xmm7 |
424 |
|
pxor xmm0, xmm1 |
425 |
|
pxor xmm3, xmm4 |
426 |
|
psubw xmm0, xmm1 |
427 |
|
psubw xmm3, xmm4 |
428 |
|
movdqa [edi + ecx*8], xmm0 |
429 |
|
movdqa [edi + ecx*8 + 16], xmm3 |
430 |
|
|
431 |
|
add ecx, 4 |
432 |
|
cmp ecx, 16 |
433 |
|
jnz .qas2_loop |
434 |
|
|
435 |
|
.qas2_done |
436 |
|
mov ecx, [esp + 8 + 16] ; dcscalar |
437 |
|
mov edx, ecx |
438 |
|
movsx eax, word [esi] |
439 |
|
shr edx, 1 |
440 |
|
cmp eax, 0 |
441 |
|
jg .qas2_gtzero |
442 |
|
|
443 |
|
sub eax, edx |
444 |
|
jmp short .qas2_mul |
445 |
|
.qas2_gtzero |
446 |
|
add eax, edx |
447 |
|
.qas2_mul |
448 |
|
cdq |
449 |
|
idiv ecx |
450 |
|
|
451 |
|
mov [edi], ax |
452 |
|
|
453 |
|
pop edi |
454 |
|
pop esi |
455 |
|
|
456 |
|
ret |
457 |
|
|
458 |
|
align 16 |
459 |
|
.qas2_q1loop |
460 |
|
movdqa xmm0, [esi + ecx*8] ; xmm0 = [1st] |
461 |
|
movdqa xmm3, [esi + ecx*8 + 16] ; xmm3 = [2nd] |
462 |
|
pxor xmm1, xmm1 |
463 |
|
pxor xmm4, xmm4 |
464 |
|
pcmpgtw xmm1, xmm0 |
465 |
|
pcmpgtw xmm4, xmm3 |
466 |
|
pxor xmm0, xmm1 |
467 |
|
pxor xmm3, xmm4 |
468 |
|
psubw xmm0, xmm1 |
469 |
|
psubw xmm3, xmm4 |
470 |
|
psrlw xmm0, 1 |
471 |
|
psrlw xmm3, 1 |
472 |
|
pxor xmm0, xmm1 |
473 |
|
pxor xmm3, xmm4 |
474 |
|
psubw xmm0, xmm1 |
475 |
|
psubw xmm3, xmm4 |
476 |
|
movdqa [edi + ecx*8], xmm0 |
477 |
|
movdqa [edi + ecx*8 + 16], xmm3 |
478 |
|
|
479 |
|
add ecx, 4 |
480 |
|
cmp ecx, 16 |
481 |
|
jnz .qas2_q1loop |
482 |
|
jmp near .qas2_done |
483 |
|
|
484 |
|
|
485 |
|
|
486 |
|
;=========================================================================== |
487 |
|
; |
488 |
; uint32_t quant_inter_mmx(int16_t * coeff, |
; uint32_t quant_inter_mmx(int16_t * coeff, |
489 |
; const int16_t const * data, |
; const int16_t const * data, |
490 |
; const uint32_t quant); |
; const uint32_t quant); |
616 |
movlhps xmm6, xmm6 ; duplicate into high 8 bytes |
movlhps xmm6, xmm6 ; duplicate into high 8 bytes |
617 |
|
|
618 |
cmp al, 1 |
cmp al, 1 |
619 |
jnz .not1 |
jz near .qes2_q1loop |
|
jmp .q1loop |
|
620 |
|
|
621 |
.not1 |
.qes2_not1 |
622 |
movq mm0, [mmx_div + eax*8 - 8] ; divider |
movq mm0, [mmx_div + eax*8 - 8] ; divider |
623 |
movq2dq xmm7, mm0 |
movq2dq xmm7, mm0 |
624 |
movlhps xmm7, xmm7 |
movlhps xmm7, xmm7 |
625 |
|
|
626 |
align 16 |
align 16 |
627 |
.loop |
.qes2_loop |
628 |
movdqa xmm0, [esi + ecx*8] ; xmm0 = [1st] |
movdqa xmm0, [esi + ecx*8] ; xmm0 = [1st] |
629 |
movdqa xmm3, [esi + ecx*8 + 16] ; xmm3 = [2nd] |
movdqa xmm3, [esi + ecx*8 + 16] ; xmm3 = [2nd] |
630 |
pxor xmm1, xmm1 ; xmm1 = 0 |
pxor xmm1, xmm1 |
631 |
pxor xmm4, xmm4 |
pxor xmm4, xmm4 |
632 |
pcmpgtw xmm1, xmm0 ; xmm1 = (0 > xmm0) |
pcmpgtw xmm1, xmm0 |
633 |
pcmpgtw xmm4, xmm3 |
pcmpgtw xmm4, xmm3 |
634 |
pxor xmm0, xmm1 ; xmm0 = |xmm0| |
pxor xmm0, xmm1 |
635 |
pxor xmm3, xmm4 |
pxor xmm3, xmm4 |
636 |
psubw xmm0, xmm1 ; displace |
psubw xmm0, xmm1 |
637 |
psubw xmm3, xmm4 |
psubw xmm3, xmm4 |
638 |
psubusw xmm0, xmm6 ; xmm0 -= sub (unsigned, dont go < 0) |
psubusw xmm0, xmm6 |
639 |
psubusw xmm3, xmm6 |
psubusw xmm3, xmm6 |
640 |
pmulhw xmm0, xmm7 ; xmm0 = (xmm0 / 2Q) >> 16 |
pmulhw xmm0, xmm7 |
641 |
pmulhw xmm3, xmm7 |
pmulhw xmm3, xmm7 |
642 |
paddw xmm5, xmm0 ; sum += xmm0 |
paddw xmm5, xmm0 |
643 |
pxor xmm0, xmm1 ; xmm0 *= sign(xmm0) |
pxor xmm0, xmm1 |
644 |
paddw xmm5, xmm3 |
paddw xmm5, xmm3 |
645 |
pxor xmm3, xmm4 |
pxor xmm3, xmm4 |
646 |
psubw xmm0, xmm1 ; undisplace |
psubw xmm0, xmm1 |
647 |
psubw xmm3, xmm4 |
psubw xmm3, xmm4 |
648 |
movdqa [edi + ecx*8], xmm0 |
movdqa [edi + ecx*8], xmm0 |
649 |
movdqa [edi + ecx*8 + 16], xmm3 |
movdqa [edi + ecx*8 + 16], xmm3 |
650 |
|
|
651 |
add ecx, 4 |
add ecx, 4 |
652 |
cmp ecx, 16 |
cmp ecx, 16 |
653 |
jnz .loop |
jnz .qes2_loop |
654 |
|
|
655 |
.done |
.qes2_done |
656 |
pmaddwd xmm5, [plus_one] |
movdqu xmm6, [plus_one] |
657 |
|
pmaddwd xmm5, xmm6 |
658 |
movhlps xmm6, xmm5 |
movhlps xmm6, xmm5 |
659 |
paddd xmm5, xmm6 |
paddd xmm5, xmm6 |
660 |
movdq2q mm0, xmm5 |
movdq2q mm0, xmm5 |
670 |
ret |
ret |
671 |
|
|
672 |
align 16 |
align 16 |
673 |
.q1loop |
.qes2_q1loop |
674 |
movq mm0, [esi + 8*ecx] ; mm0 = [1st] |
movdqa xmm0, [esi + ecx*8] ; xmm0 = [1st] |
675 |
movq mm3, [esi + 8*ecx+ 8] ; |
movdqa xmm3, [esi + ecx*8 + 16] ; xmm3 = [2nd] |
676 |
pxor mm1, mm1 ; mm1 = 0 |
pxor xmm1, xmm1 |
677 |
pxor mm4, mm4 ; |
pxor xmm4, xmm4 |
678 |
pcmpgtw mm1, mm0 ; mm1 = (0 > mm0) |
pcmpgtw xmm1, xmm0 |
679 |
pcmpgtw mm4, mm3 ; |
pcmpgtw xmm4, xmm3 |
680 |
pxor mm0, mm1 ; mm0 = |mm0| |
pxor xmm0, xmm1 |
681 |
pxor mm3, mm4 ; |
pxor xmm3, xmm4 |
682 |
psubw mm0, mm1 ; displace |
psubw xmm0, xmm1 |
683 |
psubw mm3, mm4 ; |
psubw xmm3, xmm4 |
684 |
psubusw mm0, mm6 ; mm0 -= sub (unsigned, dont go < 0) |
psubusw xmm0, xmm6 |
685 |
psubusw mm3, mm6 ; |
psubusw xmm3, xmm6 |
686 |
psrlw mm0, 1 ; mm0 >>= 1 (/2) |
psrlw xmm0, 1 |
687 |
psrlw mm3, 1 ; |
psrlw xmm3, 1 |
688 |
paddw mm5, mm0 ; sum += mm0 |
paddw xmm5, xmm0 |
689 |
pxor mm0, mm1 ; mm0 *= sign(mm0) |
pxor xmm0, xmm1 |
690 |
paddw mm5, mm3 ; |
paddw xmm5, xmm3 |
691 |
pxor mm3, mm4 ; |
pxor xmm3, xmm4 |
692 |
psubw mm0, mm1 ; undisplace |
psubw xmm0, xmm1 |
693 |
psubw mm3, mm4 |
psubw xmm3, xmm4 |
694 |
movq [edi + 8*ecx], mm0 |
movdqa [edi + ecx*8], xmm0 |
695 |
movq [edi + 8*ecx + 8], mm3 |
movdqa [edi + ecx*8 + 16], xmm3 |
696 |
|
|
697 |
add ecx,2 |
add ecx,4 |
698 |
cmp ecx,16 |
cmp ecx,16 |
699 |
jnz .q1loop |
jnz .qes2_q1loop |
700 |
|
jmp .qes2_done |
|
jmp .done |
|
701 |
|
|
702 |
|
|
703 |
|
|
803 |
ret |
ret |
804 |
%endif |
%endif |
805 |
|
|
806 |
|
|
807 |
|
|
808 |
|
;=========================================================================== |
809 |
|
; |
810 |
|
; void dequant_intra_sse2(int16_t *data, |
811 |
|
; const int16_t const *coeff, |
812 |
|
; const uint32_t quant, |
813 |
|
; const uint32_t dcscalar); |
814 |
|
; |
815 |
|
;=========================================================================== |
816 |
|
|
817 |
|
align 16 |
818 |
|
cglobal dequant_intra_sse2 |
819 |
|
dequant_intra_sse2 |
820 |
|
|
821 |
|
push esi |
822 |
|
push edi |
823 |
|
|
824 |
|
mov edi, [esp + 8 + 4] ; data |
825 |
|
mov esi, [esp + 8 + 8] ; coeff |
826 |
|
mov eax, [esp + 8 + 12] ; quant |
827 |
|
|
828 |
|
movq mm6, [mmx_add + eax*8 - 8] |
829 |
|
movq mm7, [mmx_mul + eax*8 - 8] |
830 |
|
movq2dq xmm6, mm6 |
831 |
|
movq2dq xmm7, mm7 |
832 |
|
movlhps xmm6, xmm6 |
833 |
|
movlhps xmm7, xmm7 |
834 |
|
|
835 |
|
xor eax, eax |
836 |
|
|
837 |
|
align 16 |
838 |
|
.das2_loop |
839 |
|
movdqa xmm0, [esi + eax*8] |
840 |
|
movdqa xmm3, [esi + eax*8 + 16] |
841 |
|
pxor xmm1, xmm1 |
842 |
|
pxor xmm4, xmm4 |
843 |
|
pcmpgtw xmm1, xmm0 |
844 |
|
pcmpgtw xmm4, xmm3 |
845 |
|
pxor xmm2, xmm2 |
846 |
|
pxor xmm5, xmm5 |
847 |
|
pcmpeqw xmm2, xmm0 |
848 |
|
pcmpeqw xmm5, xmm3 |
849 |
|
pandn xmm2, xmm6 |
850 |
|
pandn xmm5, xmm6 |
851 |
|
pxor xmm0, xmm1 |
852 |
|
pxor xmm3, xmm4 |
853 |
|
psubw xmm0, xmm1 |
854 |
|
psubw xmm3, xmm4 |
855 |
|
pmullw xmm0, xmm7 |
856 |
|
pmullw xmm3, xmm7 |
857 |
|
paddw xmm0, xmm2 |
858 |
|
paddw xmm3, xmm5 |
859 |
|
pxor xmm0, xmm1 |
860 |
|
pxor xmm3, xmm4 |
861 |
|
psubw xmm0, xmm1 |
862 |
|
psubw xmm3, xmm4 |
863 |
|
|
864 |
|
%ifdef SATURATE |
865 |
|
movdqu xmm2, [sse2_pos_2047] |
866 |
|
movdqu xmm4, [sse2_neg_2048] |
867 |
|
pminsw xmm0, xmm2 |
868 |
|
pminsw xmm3, xmm2 |
869 |
|
pmaxsw xmm0, xmm4 |
870 |
|
pmaxsw xmm3, xmm4 |
871 |
|
%endif |
872 |
|
|
873 |
|
movdqa [edi + eax*8], xmm0 |
874 |
|
movdqa [edi + eax*8 + 16], xmm3 |
875 |
|
|
876 |
|
add eax, 4 |
877 |
|
cmp eax, 16 |
878 |
|
jnz near .das2_loop |
879 |
|
|
880 |
|
mov ax, [esi] ; ax = data[0] |
881 |
|
imul ax, [esp + 8 + 16] ; eax = data[0] * dcscalar |
882 |
|
|
883 |
|
%ifdef SATURATE |
884 |
|
cmp ax, -2048 |
885 |
|
jl .das2_set_n2048 |
886 |
|
cmp ax, 2047 |
887 |
|
jg .das2_set_2047 |
888 |
|
%endif |
889 |
|
mov [edi], ax |
890 |
|
|
891 |
|
pop edi |
892 |
|
pop esi |
893 |
|
ret |
894 |
|
|
895 |
|
%ifdef SATURATE |
896 |
|
align 16 |
897 |
|
.das2_set_n2048 |
898 |
|
mov word [edi], -2048 |
899 |
|
pop edi |
900 |
|
pop esi |
901 |
|
ret |
902 |
|
|
903 |
|
align 16 |
904 |
|
.das2_set_2047 |
905 |
|
mov word [edi], 2047 |
906 |
|
pop edi |
907 |
|
pop esi |
908 |
|
ret |
909 |
|
%endif |
910 |
|
|
911 |
|
|
912 |
|
|
913 |
;=========================================================================== |
;=========================================================================== |
914 |
; |
; |
915 |
; void dequant_inter_mmx(int16_t * data, |
; void dequant_inter_mmx(int16_t * data, |
1015 |
xor eax, eax |
xor eax, eax |
1016 |
|
|
1017 |
align 16 |
align 16 |
1018 |
.loop |
.des2_loop |
1019 |
movdqa xmm0, [esi + eax*8] ; xmm0 = [coeff] |
movdqa xmm0, [esi + eax*8] ; xmm0 = [coeff] |
1020 |
movdqa xmm3, [esi + eax*8 + 16] |
movdqa xmm3, [esi + eax*8 + 16] |
1021 |
pxor xmm1, xmm1 |
pxor xmm1, xmm1 |
1042 |
psubw xmm3, xmm4 |
psubw xmm3, xmm4 |
1043 |
|
|
1044 |
%ifdef SATURATE |
%ifdef SATURATE |
1045 |
movdqa xmm2, [sse2_pos_2047] |
movdqu xmm2, [sse2_pos_2047] |
1046 |
movdqa xmm4, [sse2_neg_2048] |
movdqu xmm4, [sse2_neg_2048] |
1047 |
pminsw xmm0, xmm2 |
pminsw xmm0, xmm2 |
1048 |
pminsw xmm3, xmm2 |
pminsw xmm3, xmm2 |
1049 |
pmaxsw xmm0, xmm4 |
pmaxsw xmm0, xmm4 |
1055 |
|
|
1056 |
add eax, 4 |
add eax, 4 |
1057 |
cmp eax, 16 |
cmp eax, 16 |
1058 |
jnz near .loop |
jnz near .des2_loop |
1059 |
|
|
1060 |
pop edi |
pop edi |
1061 |
pop esi |
pop esi |