42 |
; * * |
; * * |
43 |
; * Revision history: * |
; * Revision history: * |
44 |
; * * |
; * * |
45 |
|
; * 14.06.2002 mmx dequant4_* funcs revamped -Skal- * |
46 |
; * 22.01.2002 initial version * |
; * 22.01.2002 initial version * |
47 |
; * * |
; * * |
48 |
; ******************************************************************************/ |
; ******************************************************************************/ |
54 |
|
|
55 |
bits 32 |
bits 32 |
56 |
|
|
57 |
section .data |
%ifdef FORMAT_COFF |
58 |
|
SECTION .data data |
59 |
|
%else |
60 |
|
SECTION .data data align=8 |
61 |
|
%endif |
62 |
|
|
63 |
|
|
64 |
%macro cglobal 1 |
%macro cglobal 1 |
65 |
%ifdef PREFIX |
%ifdef PREFIX |
243 |
;=========================================================================== |
;=========================================================================== |
244 |
|
|
245 |
align 16 |
align 16 |
246 |
mmx_32768_minus_2048 times 4 dw (32768-2048) |
|
247 |
mmx_32767_minus_2047 times 4 dw (32767-2047) |
mmx_32767_minus_2047 times 4 dw (32767-2047) |
248 |
|
mmx_32768_minus_2048 times 4 dw (32768-2048) |
249 |
|
mmx_2047 times 4 dw 2047 |
250 |
|
mmx_minus_2048 times 4 dw (-2048) |
251 |
|
zero times 4 dw 0 |
252 |
|
|
253 |
section .text |
section .text |
254 |
|
|
677 |
; |
; |
678 |
;=========================================================================== |
;=========================================================================== |
679 |
|
|
680 |
align 16 |
; Note: in order to saturate 'easily', we pre-shift the quantifier |
681 |
cglobal dequant4_intra_mmx |
; by 4. Then, the high-word of (coeff[]*matrix[i]*quant) are used to |
682 |
dequant4_intra_mmx |
; build a saturating mask. It is non-zero only when an overflow occured. |
683 |
|
; We thus avoid packing/unpacking toward double-word. |
684 |
|
; Moreover, we perform the mult (matrix[i]*quant) first, instead of, e.g., |
685 |
|
; (coeff[i]*matrix[i]). This is less prone to overflow if coeff[] are not |
686 |
|
; checked. Input ranges are: coeff in [-127,127], inter_matrix in [1..255],a |
687 |
|
; and quant in [1..31]. |
688 |
|
; |
689 |
|
; The original loop is: |
690 |
|
; |
691 |
|
%if 0 |
692 |
|
movq mm0, [ecx+8*eax + 8*16] ; mm0 = coeff[i] |
693 |
|
pxor mm1, mm1 |
694 |
|
pcmpgtw mm1, mm0 |
695 |
|
pxor mm0, mm1 ; change sign if negative |
696 |
|
psubw mm0, mm1 ; -> mm0 = abs(coeff[i]), mm1 = sign of coeff[i] |
697 |
|
|
698 |
|
movq mm2, mm7 ; mm2 = quant |
699 |
|
pmullw mm2, [intra_matrix + 8*eax + 8*16 ] ; matrix[i]*quant. |
700 |
|
|
701 |
|
movq mm6, mm2 |
702 |
|
pmulhw mm2, mm0 ; high of coeff*(matrix*quant) (should be 0 if no overflow) |
703 |
|
pmullw mm0, mm6 ; low of coeff*(matrix*quant) |
704 |
|
|
705 |
|
pxor mm5, mm5 |
706 |
|
pcmpgtw mm2, mm5 ; otherflow? |
707 |
|
psrlw mm2, 5 ; =0 if no clamp, 2047 otherwise |
708 |
|
psrlw mm0, 5 |
709 |
|
paddw mm0, mm1 ; start restoring sign |
710 |
|
por mm0, mm2 ; saturate to 2047 if needed |
711 |
|
pxor mm0, mm1 ; finish negating back |
712 |
|
|
713 |
push esi |
movq [edx + 8*eax + 8*16], mm0 ; data[i] |
714 |
push edi |
add eax, 1 |
715 |
|
%endif |
716 |
|
|
717 |
mov edi, [esp + 8 + 4] ; data |
;******************************************************************** |
|
mov esi, [esp + 8 + 8] ; coeff |
|
|
mov eax, [esp + 8 + 12] ; quant |
|
718 |
|
|
719 |
movq mm7, [mmx_mul_quant + eax*8 - 8] |
align 16 |
720 |
|
cglobal dequant4_intra_mmx |
721 |
|
dequant4_intra_mmx: |
722 |
|
|
723 |
xor eax, eax |
mov edx, [esp+4] ; data |
724 |
|
mov ecx, [esp+8] ; coeff |
725 |
|
mov eax, [esp+12] ; quant |
726 |
|
|
727 |
|
movq mm7, [mmx_mul_quant + eax*8 - 8] |
728 |
|
mov eax, -16 ; to keep aligned, we regularly process coeff[0] |
729 |
|
psllw mm7, 2 ; << 2. See comment. |
730 |
|
pxor mm6, mm6 ; this is a NOP |
731 |
|
|
732 |
align 16 |
align 16 |
733 |
.loop |
.loop |
734 |
movq mm0, [esi + 8*eax] ; mm0 = [coeff] |
movq mm0, [ecx+8*eax + 8*16] ; mm0 = c = coeff[i] |
735 |
|
movq mm3, [ecx+8*eax + 8*16 +8]; mm3 = c' = coeff[i+1] |
736 |
|
pxor mm1, mm1 |
737 |
|
pxor mm4, mm4 |
738 |
|
pcmpgtw mm1, mm0 ; mm1 = sgn(c) |
739 |
|
movq mm2, mm7 ; mm2 = quant |
740 |
|
|
741 |
pxor mm1, mm1 ; mm1 = 0 |
pcmpgtw mm4, mm3 ; mm4 = sgn(c') |
742 |
pcmpeqw mm1, mm0 ; mm1 = (0 == mm0) |
pmullw mm2, [intra_matrix + 8*eax + 8*16 ] ; matrix[i]*quant |
743 |
|
|
744 |
pxor mm2, mm2 ; mm2 = 0 |
pxor mm0, mm1 ; negate if negative |
745 |
pcmpgtw mm2, mm0 ; mm2 = (0 > mm0) |
pxor mm3, mm4 ; negate if negative |
|
pxor mm0, mm2 ; mm0 = |mm0| |
|
|
psubw mm0, mm2 ; displace |
|
746 |
|
|
747 |
pmullw mm0, mm7 ; mm0 *= quant |
psubw mm0, mm1 |
748 |
|
psubw mm3, mm4 |
749 |
|
|
750 |
movq mm3, [intra_matrix + 8*eax] |
; we're short on register, here. Poor pairing... |
751 |
|
|
752 |
movq mm4, mm0 ; |
movq mm5, mm2 |
753 |
pmullw mm0, mm3 ; mm0 = low(mm0 * mm3) |
pmullw mm2, mm0 ; low of coeff*(matrix*quant) |
|
pmulhw mm3, mm4 ; mm3 = high(mm0 * mm3) |
|
754 |
|
|
755 |
movq mm4, mm0 ; mm0,mm4 = unpack(mm3, mm0) |
pmulhw mm0, mm5 ; high of coeff*(matrix*quant) |
756 |
punpcklwd mm0, mm3 ; |
movq mm5, mm7 ; mm2 = quant |
|
punpckhwd mm4, mm3 ; |
|
|
psrld mm0, 3 ; mm0,mm4 /= 8 |
|
|
psrld mm4, 3 ; |
|
|
packssdw mm0, mm4 ; mm0 = pack(mm4, mm0) |
|
757 |
|
|
758 |
pxor mm0, mm2 ; mm0 *= sign(mm0) |
pmullw mm5, [intra_matrix + 8*eax + 8*16 +8] ; matrix[i+1]*quant |
|
psubw mm0, mm2 ; undisplace |
|
|
pandn mm1, mm0 ; mm1 = ~(iszero) & mm0 |
|
759 |
|
|
760 |
%ifdef SATURATE |
movq mm6, mm5 |
761 |
movq mm2, [mmx_32767_minus_2047] |
add eax,2 ; z-flag will be tested later |
|
movq mm6, [mmx_32768_minus_2048] |
|
|
paddsw mm1, mm2 |
|
|
psubsw mm1, mm2 |
|
|
psubsw mm1, mm6 |
|
|
paddsw mm1, mm6 |
|
|
%endif |
|
762 |
|
|
763 |
movq [edi + 8*eax], mm1 ; [data] = mm0 |
pmullw mm6, mm3 ; low of coeff*(matrix*quant) |
764 |
|
pmulhw mm3, mm5 ; high of coeff*(matrix*quant) |
765 |
|
|
766 |
add eax, 1 |
pcmpgtw mm0, [zero] |
767 |
cmp eax, 16 |
paddusw mm2, mm0 |
768 |
jnz near .loop |
psrlw mm2, 5 |
769 |
|
|
770 |
mov ax, [esi] ; ax = data[0] |
pcmpgtw mm3, [zero] |
771 |
imul ax, [esp + 8 + 16] ; eax = data[0] * dcscalar |
paddusw mm6, mm3 |
772 |
mov [edi], ax ; data[0] = ax |
psrlw mm6, 5 |
|
|
|
|
%ifdef SATURATE |
|
|
cmp ax, -2048 |
|
|
jl .set_n2048 |
|
|
cmp ax, 2047 |
|
|
jg .set_2047 |
|
|
%endif |
|
773 |
|
|
774 |
pop edi |
pxor mm2, mm1 ; start negating back |
775 |
pop esi |
pxor mm6, mm4 ; start negating back |
|
ret |
|
776 |
|
|
777 |
%ifdef SATURATE |
psubusw mm1, mm0 |
778 |
.set_n2048 |
psubusw mm4, mm3 |
|
mov word [edi], -2048 |
|
|
pop edi |
|
|
pop esi |
|
|
ret |
|
779 |
|
|
780 |
.set_2047 |
psubw mm2, mm1 ; finish negating back |
781 |
mov word [edi], 2047 |
psubw mm6, mm4 ; finish negating back |
|
pop edi |
|
|
pop esi |
|
782 |
|
|
783 |
ret |
movq [edx + 8*eax + 8*16 -2*8 ], mm2 ; data[i] |
784 |
%endif |
movq [edx + 8*eax + 8*16 -2*8 +8], mm6 ; data[i+1] |
785 |
|
|
786 |
|
jnz near .loop |
787 |
|
|
788 |
|
; deal with DC |
789 |
|
|
790 |
|
movd mm0, [ecx] |
791 |
|
pmullw mm0, [esp+16] ; dcscalar |
792 |
|
movq mm2, [mmx_32767_minus_2047] |
793 |
|
paddsw mm0, mm2 |
794 |
|
psubsw mm0, mm2 |
795 |
|
movq mm2, [mmx_32768_minus_2048] |
796 |
|
psubsw mm0, mm2 |
797 |
|
paddsw mm0, mm2 |
798 |
|
movd eax, mm0 |
799 |
|
mov [edx], ax |
800 |
|
|
801 |
|
ret |
802 |
|
|
803 |
;=========================================================================== |
;=========================================================================== |
804 |
; |
; |
808 |
; |
; |
809 |
;=========================================================================== |
;=========================================================================== |
810 |
|
|
811 |
|
; Note: We use (2*c + sgn(c) - sgn(-c)) as multiplier |
812 |
|
; so we handle the 3 cases: c<0, c==0, and c>0 in one shot. |
813 |
|
; sgn(x) is the result of 'pcmpgtw 0,x': 0 if x>=0, -1 if x<0. |
814 |
|
; It's mixed with the extraction of the absolute value. |
815 |
|
|
816 |
align 16 |
align 16 |
817 |
cglobal dequant4_inter_mmx |
cglobal dequant4_inter_mmx |
818 |
dequant4_inter_mmx |
dequant4_inter_mmx: |
819 |
|
|
820 |
push esi |
mov edx, [esp+ 4] ; data |
821 |
push edi |
mov ecx, [esp+ 8] ; coeff |
822 |
|
mov eax, [esp+12] ; quant |
|
mov edi, [esp + 8 + 4] ; data |
|
|
mov esi, [esp + 8 + 8] ; coeff |
|
|
mov eax, [esp + 8 + 12] ; quant |
|
823 |
movq mm7, [mmx_mul_quant + eax*8 - 8] |
movq mm7, [mmx_mul_quant + eax*8 - 8] |
824 |
movq mm6, [mmx_one] |
mov eax, -16 |
825 |
xor eax, eax |
paddw mm7, mm7 ; << 1 |
826 |
pxor mm5, mm5 ; mismatch sum |
pxor mm6, mm6 ; mismatch sum |
|
|
|
827 |
|
|
828 |
align 16 |
align 16 |
829 |
.loop |
.loop |
830 |
movq mm0, [esi + 8*eax] ; mm0 = [coeff] |
movq mm0, [ecx+8*eax + 8*16 ] ; mm0 = coeff[i] |
831 |
|
movq mm2, [ecx+8*eax + 8*16 +8] ; mm2 = coeff[i+1] |
832 |
pxor mm1, mm1 ; mm1 = 0 |
add eax,2 |
833 |
pcmpeqw mm1, mm0 ; mm1 = (0 == mm0) |
|
834 |
|
pxor mm1, mm1 |
835 |
pxor mm2, mm2 ; mm2 = 0 |
pxor mm3, mm3 |
836 |
pcmpgtw mm2, mm0 ; mm2 = (0 > mm0) |
pcmpgtw mm1, mm0 ; mm1 = sgn(c) (preserved) |
837 |
pxor mm0, mm2 ; mm0 = |mm0| |
pcmpgtw mm3, mm2 ; mm3 = sgn(c') (preserved) |
838 |
psubw mm0, mm2 ; displace |
paddsw mm0, mm1 ; c += sgn(c) |
839 |
|
paddsw mm2, mm3 ; c += sgn(c') |
840 |
psllw mm0, 1 ; |
paddw mm0, mm0 ; c *= 2 |
841 |
paddsw mm0, mm6 ; mm0 = 2*mm0 + 1 |
paddw mm2, mm2 ; c'*= 2 |
|
pmullw mm0, mm7 ; mm0 *= quant |
|
|
|
|
|
movq mm3, [inter_matrix + 8*eax] |
|
|
|
|
|
movq mm4, mm0 |
|
|
pmullw mm0, mm3 ; mm0 = low(mm0 * mm3) |
|
|
pmulhw mm3, mm4 ; mm3 = high(mm0 * mm3) |
|
842 |
|
|
843 |
movq mm4, mm0 ; mm0,mm4 = unpack(mm3, mm0) |
pxor mm4, mm4 |
844 |
punpcklwd mm0, mm3 ; |
pxor mm5, mm5 |
845 |
punpckhwd mm4, mm3 ; |
psubw mm4, mm0 ; -c |
846 |
|
psubw mm5, mm2 ; -c' |
847 |
psrad mm0, 4 ; mm0,mm4 /= 16 |
psraw mm4, 16 ; mm4 = sgn(-c) |
848 |
psrad mm4, 4 ; |
psraw mm5, 16 ; mm5 = sgn(-c') |
849 |
packssdw mm0, mm4 ; mm0 = pack(mm4, mm0) |
psubsw mm0, mm4 ; c -= sgn(-c) |
850 |
|
psubsw mm2, mm5 ; c' -= sgn(-c') |
851 |
pxor mm0, mm2 ; mm0 *= sign(mm0) |
pxor mm0, mm1 ; finish changing sign if needed |
852 |
psubw mm0, mm2 ; undisplace |
pxor mm2, mm3 ; finish changing sign if needed |
853 |
pandn mm1, mm0 ; mm1 = ~(iszero) & mm0 |
|
854 |
|
; we're short on register, here. Poor pairing... |
855 |
|
|
856 |
;%ifdef SATURATE |
movq mm4, mm7 ; (matrix*quant) |
857 |
movq mm2, [mmx_32767_minus_2047] |
pmullw mm4, [inter_matrix + 8*eax + 8*16 -2*8] |
858 |
movq mm4, [mmx_32768_minus_2048] |
movq mm5, mm4 |
859 |
paddsw mm1, mm2 |
pmulhw mm5, mm0 ; high of c*(matrix*quant) |
860 |
psubsw mm1, mm2 |
pmullw mm0, mm4 ; low of c*(matrix*quant) |
861 |
psubsw mm1, mm4 |
|
862 |
paddsw mm1, mm4 |
movq mm4, mm7 ; (matrix*quant) |
863 |
;%endif |
pmullw mm4, [inter_matrix + 8*eax + 8*16 -2*8 + 8] |
864 |
|
|
865 |
pxor mm5, mm1 ; mismatch |
pcmpgtw mm5, [zero] |
866 |
|
paddusw mm0, mm5 |
867 |
movq [edi + 8*eax], mm1 ; [data] = mm0 |
psrlw mm0, 5 |
868 |
|
pxor mm0, mm1 ; start restoring sign |
869 |
|
psubusw mm1, mm5 |
870 |
|
|
871 |
|
movq mm5, mm4 |
872 |
|
pmulhw mm5, mm2 ; high of c*(matrix*quant) |
873 |
|
pmullw mm2, mm4 ; low of c*(matrix*quant) |
874 |
|
psubw mm0, mm1 ; finish restoring sign |
875 |
|
|
876 |
|
pcmpgtw mm5, [zero] |
877 |
|
paddusw mm2, mm5 |
878 |
|
psrlw mm2, 5 |
879 |
|
pxor mm2, mm3 ; start restoring sign |
880 |
|
psubusw mm3, mm5 |
881 |
|
psubw mm2, mm3 ; finish restoring sign |
882 |
|
|
883 |
|
pxor mm6, mm0 ; mismatch control |
884 |
|
movq [edx + 8*eax + 8*16 -2*8 ], mm0 ; data[i] |
885 |
|
pxor mm6, mm2 ; mismatch control |
886 |
|
movq [edx + 8*eax + 8*16 -2*8 +8], mm2 ; data[i+1] |
887 |
|
|
|
add eax, 1 |
|
|
cmp eax, 16 |
|
888 |
jnz near .loop |
jnz near .loop |
889 |
|
|
890 |
; mismatch control |
; mismatch control |
891 |
|
|
892 |
movq mm0, mm5 |
movq mm0, mm6 |
|
movq mm1, mm5 |
|
|
movq mm2, mm5 |
|
893 |
psrlq mm0, 48 |
psrlq mm0, 48 |
894 |
|
movq mm1, mm6 |
895 |
|
movq mm2, mm6 |
896 |
psrlq mm1, 32 |
psrlq mm1, 32 |
897 |
|
pxor mm6, mm0 |
898 |
psrlq mm2, 16 |
psrlq mm2, 16 |
899 |
pxor mm5, mm0 |
pxor mm6, mm1 |
900 |
pxor mm5, mm1 |
pxor mm6, mm2 |
901 |
pxor mm5, mm2 |
movd eax, mm6 |
902 |
|
and eax, 1 |
903 |
movd eax, mm5 |
xor eax, 1 |
904 |
test eax, 0x1 |
xor word [edx + 2*63], ax |
|
jnz .done |
|
|
|
|
|
xor word [edi + 2*63], 1 |
|
|
|
|
|
.done |
|
|
pop edi |
|
|
pop esi |
|
905 |
|
|
906 |
ret |
ret |
907 |
|
|