20 |
; * along with this program; if not, write to the Free Software |
; * along with this program; if not, write to the Free Software |
21 |
; * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA |
; * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA |
22 |
; * |
; * |
23 |
; * $Id: deintl_sse.asm,v 1.3 2008-11-11 20:46:24 Isibaar Exp $ |
; * $Id: deintl_sse.asm,v 1.4 2008-11-26 01:04:34 Isibaar Exp $ |
24 |
; * |
; * |
25 |
; *************************************************************************/ |
; *************************************************************************/ |
26 |
|
|
32 |
; * |
; * |
33 |
; *************************************************************************/ |
; *************************************************************************/ |
34 |
|
|
35 |
bits 32 |
%include "nasm.inc" |
|
|
|
|
%macro cglobal 1 |
|
|
%ifdef PREFIX |
|
|
%ifdef MARK_FUNCS |
|
|
global _%1:function %1.endfunc-%1 |
|
|
%define %1 _%1:function %1.endfunc-%1 |
|
|
%define ENDFUNC .endfunc |
|
|
%else |
|
|
global _%1 |
|
|
%define %1 _%1 |
|
|
%define ENDFUNC |
|
|
%endif |
|
|
%else |
|
|
%ifdef MARK_FUNCS |
|
|
global %1:function %1.endfunc-%1 |
|
|
%define ENDFUNC .endfunc |
|
|
%else |
|
|
global %1 |
|
|
%define ENDFUNC |
|
|
%endif |
|
|
%endif |
|
|
%endmacro |
|
36 |
|
|
37 |
;////////////////////////////////////////////////////////////////////// |
;////////////////////////////////////////////////////////////////////// |
38 |
|
|
40 |
|
|
41 |
;////////////////////////////////////////////////////////////////////// |
;////////////////////////////////////////////////////////////////////// |
42 |
|
|
43 |
%ifdef FORMAT_COFF |
DATA |
|
SECTION .rodata |
|
|
%else |
|
|
SECTION .rodata align=16 |
|
|
%endif |
|
44 |
|
|
45 |
align 16 |
align SECTION_ALIGN |
46 |
Mask_6b times 16 db 0x3f |
Mask_6b times 16 db 0x3f |
47 |
Rnd_3b: times 16 db 3 |
Rnd_3b: times 16 db 3 |
48 |
|
|
49 |
SECTION .text |
SECTION .rotext align=SECTION_ALIGN |
50 |
|
|
51 |
;////////////////////////////////////////////////////////////////////// |
;////////////////////////////////////////////////////////////////////// |
52 |
;// sse version |
;// sse version |
53 |
|
|
54 |
align 16 |
align SECTION_ALIGN |
55 |
xvid_deinterlace_sse: |
xvid_deinterlace_sse: |
56 |
|
|
57 |
mov eax, [esp+ 4] ; Pix |
mov _EAX, prm1 ; Pix |
58 |
mov ecx, [esp+12] ; Height |
mov TMP0, prm3 ; Height |
59 |
mov edx, [esp+16] ; BpS |
mov TMP1, prm4 ; BpS |
60 |
|
|
61 |
push ebx |
push _EBX |
62 |
mov ebx, [esp+4+ 8] ; Width |
%ifdef ARCH_IS_X86_64 |
63 |
|
mov _EBX, prm2 ; Width |
64 |
add ebx, 7 |
%else |
65 |
shr ecx, 1 |
mov _EBX, [esp+4+ 8] ; Width |
66 |
shr ebx, 3 ; Width /= 8 |
%endif |
67 |
dec ecx |
|
68 |
|
add _EBX, 7 |
69 |
|
shr TMP0, 1 |
70 |
|
shr _EBX, 3 ; Width /= 8 |
71 |
|
dec TMP0 |
72 |
|
|
73 |
movq mm6, [Mask_6b] |
movq mm6, [Mask_6b] |
74 |
|
|
75 |
.Loop_x: |
.Loop_x: |
76 |
push eax |
push _EAX |
77 |
movq mm1, [eax ] |
movq mm1, [_EAX ] |
78 |
movq mm2, [eax+ edx] |
movq mm2, [_EAX+ TMP1] |
79 |
lea eax, [eax+ edx] |
lea _EAX, [_EAX+ TMP1] |
80 |
movq mm0, mm2 |
movq mm0, mm2 |
81 |
|
|
82 |
push ecx |
push TMP0 |
83 |
|
|
84 |
.Loop: |
.Loop: |
85 |
movq mm3, [eax+ edx] |
movq mm3, [_EAX+ TMP1] |
86 |
movq mm4, [eax+2*edx] |
movq mm4, [_EAX+2*TMP1] |
87 |
movq mm5, mm2 |
movq mm5, mm2 |
88 |
pavgb mm0, mm4 |
pavgb mm0, mm4 |
89 |
pavgb mm1, mm3 |
pavgb mm1, mm3 |
97 |
pand mm0, mm6 |
pand mm0, mm6 |
98 |
paddusb mm1, mm2 |
paddusb mm1, mm2 |
99 |
psubusb mm1, mm0 |
psubusb mm1, mm0 |
100 |
movq [eax], mm1 |
movq [_EAX], mm1 |
101 |
lea eax, [eax+2*edx] |
lea _EAX, [_EAX+2*TMP1] |
102 |
movq mm0, mm5 |
movq mm0, mm5 |
103 |
movq mm1, mm3 |
movq mm1, mm3 |
104 |
movq mm2, mm4 |
movq mm2, mm4 |
105 |
dec ecx |
dec TMP0 |
106 |
jg .Loop |
jg .Loop |
107 |
|
|
108 |
pavgb mm0, mm2 ; p0 += p2 |
pavgb mm0, mm2 ; p0 += p2 |
117 |
pand mm0, mm6 |
pand mm0, mm6 |
118 |
paddusb mm1, mm2 |
paddusb mm1, mm2 |
119 |
psubusb mm1, mm0 |
psubusb mm1, mm0 |
120 |
movq [eax], mm1 |
movq [_EAX], mm1 |
121 |
|
|
122 |
pop ecx |
pop TMP0 |
123 |
pop eax |
pop _EAX |
124 |
add eax, 8 |
add _EAX, 8 |
125 |
|
|
126 |
dec ebx |
dec _EBX |
127 |
jg .Loop_x |
jg .Loop_x |
128 |
|
|
129 |
pop ebx |
pop _EBX |
130 |
ret |
ret |
131 |
ENDFUNC |
ENDFUNC |
132 |
|
|