3 |
; * XVID MPEG-4 VIDEO CODEC |
; * XVID MPEG-4 VIDEO CODEC |
4 |
; * - MMX CBP computation - |
; * - MMX CBP computation - |
5 |
; * |
; * |
6 |
; * Copyright (C) 2001-2003 Peter Ross <pross@xvid.org> |
; * Copyright (C) 2005 Carlo Bramini <carlo.bramix@libero.it> |
7 |
|
; * 2001-2003 Peter Ross <pross@xvid.org> |
8 |
; * 2002-2003 Pascal Massimino <skal@planet-d.net> |
; * 2002-2003 Pascal Massimino <skal@planet-d.net> |
9 |
; * |
; * |
10 |
; * This program is free software ; you can redistribute it and/or modify |
; * This program is free software ; you can redistribute it and/or modify |
21 |
; * along with this program ; if not, write to the Free Software |
; * along with this program ; if not, write to the Free Software |
22 |
; * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA |
; * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA |
23 |
; * |
; * |
24 |
; * $Id: cbp_mmx.asm,v 1.10 2004-07-24 11:46:08 edgomez Exp $ |
; * $Id: cbp_mmx.asm,v 1.19 2009-09-16 17:07:58 Isibaar Exp $ |
25 |
; * |
; * |
26 |
; ***************************************************************************/ |
; ***************************************************************************/ |
27 |
|
|
|
BITS 32 |
|
|
|
|
28 |
;============================================================================= |
;============================================================================= |
29 |
; Macros |
; Macros |
30 |
;============================================================================= |
;============================================================================= |
31 |
|
|
32 |
%macro cglobal 1 |
%include "nasm.inc" |
|
%ifdef PREFIX |
|
|
global _%1 |
|
|
%define %1 _%1 |
|
|
%else |
|
|
global %1 |
|
|
%endif |
|
|
%endmacro |
|
33 |
|
|
34 |
;============================================================================= |
;============================================================================= |
35 |
; Local data |
; Local data |
36 |
;============================================================================= |
;============================================================================= |
37 |
|
|
38 |
%ifdef FORMAT_COFF |
DATA |
|
SECTION .rodata |
|
|
%else |
|
|
SECTION .rodata align=16 |
|
|
%endif |
|
39 |
|
|
40 |
ALIGN 16 |
ALIGN SECTION_ALIGN |
41 |
|
|
42 |
|
mult_mask: |
43 |
|
db 0x10,0x20,0x04,0x08,0x01,0x02,0x00,0x00 |
44 |
ignore_dc: |
ignore_dc: |
45 |
dw 0, -1, -1, -1, -1, -1, -1, -1 |
dw 0, -1, -1, -1 |
46 |
|
|
47 |
;============================================================================= |
;============================================================================= |
48 |
; Code |
; Code |
49 |
;============================================================================= |
;============================================================================= |
50 |
|
|
51 |
SECTION .text |
TEXT |
52 |
|
|
53 |
cglobal calc_cbp_mmx |
cglobal calc_cbp_mmx |
54 |
|
|
56 |
; uint32_t calc_cbp_mmx(const int16_t coeff[6][64]); |
; uint32_t calc_cbp_mmx(const int16_t coeff[6][64]); |
57 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
58 |
|
|
59 |
ALIGN 16 |
%macro MAKE_LOAD 2 |
60 |
calc_cbp_mmx: |
por mm0, [%2-128*1+%1*8] |
61 |
push ebx |
por mm1, [%2+128*0+%1*8] |
62 |
push esi |
por mm2, [%2+128*1+%1*8] |
63 |
|
por mm3, [%2+128*2+%1*8] |
64 |
|
por mm4, [%2+128*3+%1*8] |
65 |
|
por mm5, [%2+128*4+%1*8] |
66 |
|
%endmacro |
67 |
|
|
68 |
mov esi, [esp + 8 + 4] ; coeff |
ALIGN SECTION_ALIGN |
69 |
xor eax, eax ; cbp = 0 |
calc_cbp_mmx: |
70 |
mov edx, (1 << 5) |
mov _EAX, prm1 ; coeff |
71 |
|
|
72 |
movq mm7, [ignore_dc] |
movq mm7, [ignore_dc] |
73 |
|
pxor mm6, mm6 ; used only for comparing |
74 |
.loop |
movq mm0, [_EAX+128*0] |
75 |
movq mm0, [esi] |
movq mm1, [_EAX+128*1] |
76 |
movq mm1, [esi+8] |
movq mm2, [_EAX+128*2] |
77 |
|
movq mm3, [_EAX+128*3] |
78 |
|
movq mm4, [_EAX+128*4] |
79 |
|
movq mm5, [_EAX+128*5] |
80 |
|
add _EAX, 8+128 |
81 |
pand mm0, mm7 |
pand mm0, mm7 |
82 |
|
pand mm1, mm7 |
83 |
|
pand mm2, mm7 |
84 |
|
pand mm3, mm7 |
85 |
|
pand mm4, mm7 |
86 |
|
pand mm5, mm7 |
87 |
|
|
88 |
|
MAKE_LOAD 0, _EAX |
89 |
|
MAKE_LOAD 1, _EAX |
90 |
|
MAKE_LOAD 2, _EAX |
91 |
|
MAKE_LOAD 3, _EAX |
92 |
|
MAKE_LOAD 4, _EAX |
93 |
|
MAKE_LOAD 5, _EAX |
94 |
|
MAKE_LOAD 6, _EAX |
95 |
|
MAKE_LOAD 7, _EAX |
96 |
|
MAKE_LOAD 8, _EAX |
97 |
|
MAKE_LOAD 9, _EAX |
98 |
|
MAKE_LOAD 10, _EAX |
99 |
|
MAKE_LOAD 11, _EAX |
100 |
|
MAKE_LOAD 12, _EAX |
101 |
|
MAKE_LOAD 13, _EAX |
102 |
|
MAKE_LOAD 14, _EAX |
103 |
|
|
104 |
|
movq mm7, [mult_mask] |
105 |
|
packssdw mm0, mm1 |
106 |
|
packssdw mm2, mm3 |
107 |
|
packssdw mm4, mm5 |
108 |
|
packssdw mm0, mm2 |
109 |
|
packssdw mm4, mm6 |
110 |
|
pcmpeqw mm0, mm6 |
111 |
|
pcmpeqw mm4, mm6 |
112 |
|
pcmpeqw mm0, mm6 |
113 |
|
pcmpeqw mm4, mm6 |
114 |
|
psrlw mm0, 15 |
115 |
|
psrlw mm4, 15 |
116 |
|
packuswb mm0, mm4 |
117 |
|
pmaddwd mm0, mm7 |
118 |
|
|
|
por mm0, [esi+16] |
|
|
por mm1, [esi+24] |
|
|
|
|
|
por mm0, [esi+32] |
|
|
por mm1, [esi+40] |
|
|
|
|
|
por mm0, [esi+48] |
|
|
por mm1, [esi+56] |
|
|
|
|
|
por mm0, [esi+64] |
|
|
por mm1, [esi+72] |
|
|
|
|
|
por mm0, [esi+80] |
|
|
por mm1, [esi+88] |
|
|
|
|
|
por mm0, [esi+96] |
|
|
por mm1, [esi+104] |
|
|
|
|
|
por mm0, [esi+112] |
|
|
por mm1, [esi+120] |
|
|
|
|
|
por mm0, mm1 |
|
119 |
movq mm1, mm0 |
movq mm1, mm0 |
120 |
psrlq mm1, 32 |
psrlq mm1, 32 |
121 |
lea esi, [esi + 128] |
paddusb mm0, mm1 |
|
|
|
|
por mm0, mm1 |
|
|
movd ebx, mm0 |
|
|
|
|
|
test ebx, ebx |
|
|
jz .next |
|
|
or eax, edx ; cbp |= 1 << (5-i) |
|
|
|
|
|
.next |
|
|
shr edx,1 |
|
|
jnc .loop |
|
|
|
|
|
pop esi |
|
|
pop ebx |
|
122 |
|
|
123 |
|
movd eax, mm0 |
124 |
|
shr _EAX, 8 |
125 |
|
and _EAX, 0x3F |
126 |
ret |
ret |
127 |
|
ENDFUNC |
128 |
|
|
129 |
|
NON_EXEC_STACK |