| 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 1.9 2004-03-22 22:36:23 edgomez Exp $ |
; * $Id: cbp_mmx.asm,v 1.16 2008-11-11 20:46:24 Isibaar Exp $ |
| 25 |
; * |
; * |
| 26 |
; ***************************************************************************/ |
; ***************************************************************************/ |
| 27 |
|
|
| 33 |
|
|
| 34 |
%macro cglobal 1 |
%macro cglobal 1 |
| 35 |
%ifdef PREFIX |
%ifdef PREFIX |
| 36 |
|
%ifdef MARK_FUNCS |
| 37 |
|
global _%1:function %1.endfunc-%1 |
| 38 |
|
%define %1 _%1:function %1.endfunc-%1 |
| 39 |
|
%define ENDFUNC .endfunc |
| 40 |
|
%else |
| 41 |
global _%1 |
global _%1 |
| 42 |
%define %1 _%1 |
%define %1 _%1 |
| 43 |
|
%define ENDFUNC |
| 44 |
|
%endif |
| 45 |
|
%else |
| 46 |
|
%ifdef MARK_FUNCS |
| 47 |
|
global %1:function %1.endfunc-%1 |
| 48 |
|
%define ENDFUNC .endfunc |
| 49 |
%else |
%else |
| 50 |
global %1 |
global %1 |
| 51 |
|
%define ENDFUNC |
| 52 |
|
%endif |
| 53 |
%endif |
%endif |
| 54 |
%endmacro |
%endmacro |
| 55 |
|
|
| 58 |
;============================================================================= |
;============================================================================= |
| 59 |
|
|
| 60 |
%ifdef FORMAT_COFF |
%ifdef FORMAT_COFF |
| 61 |
SECTION .rodata data |
SECTION .rodata |
| 62 |
%else |
%else |
| 63 |
SECTION .rodata data align=16 |
SECTION .rodata align=16 |
| 64 |
%endif |
%endif |
| 65 |
|
|
| 66 |
ALIGN 16 |
ALIGN 16 |
| 67 |
|
|
| 68 |
|
mult_mask: |
| 69 |
|
db 0x10,0x20,0x04,0x08,0x01,0x02,0x00,0x00 |
| 70 |
ignore_dc: |
ignore_dc: |
| 71 |
dw 0, -1, -1, -1, -1, -1, -1, -1 |
dw 0, -1, -1, -1 |
| 72 |
|
|
| 73 |
;============================================================================= |
;============================================================================= |
| 74 |
; Code |
; Code |
| 82 |
; uint32_t calc_cbp_mmx(const int16_t coeff[6][64]); |
; uint32_t calc_cbp_mmx(const int16_t coeff[6][64]); |
| 83 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
| 84 |
|
|
| 85 |
|
%macro MAKE_LOAD 1 |
| 86 |
|
por mm0, [eax-128*1+%1*8] |
| 87 |
|
por mm1, [eax+128*0+%1*8] |
| 88 |
|
por mm2, [eax+128*1+%1*8] |
| 89 |
|
por mm3, [eax+128*2+%1*8] |
| 90 |
|
por mm4, [eax+128*3+%1*8] |
| 91 |
|
por mm5, [eax+128*4+%1*8] |
| 92 |
|
%endmacro |
| 93 |
|
|
| 94 |
ALIGN 16 |
ALIGN 16 |
| 95 |
calc_cbp_mmx: |
calc_cbp_mmx: |
| 96 |
push ebx |
mov eax, [esp + 4] ; coeff |
|
push esi |
|
|
|
|
|
mov esi, [esp + 8 + 4] ; coeff |
|
|
xor eax, eax ; cbp = 0 |
|
|
mov edx, (1 << 5) |
|
| 97 |
|
|
| 98 |
movq mm7, [ignore_dc] |
movq mm7, [ignore_dc] |
| 99 |
|
pxor mm6, mm6 ; used only for comparing |
| 100 |
.loop |
movq mm0, [eax+128*0] |
| 101 |
movq mm0, [esi] |
movq mm1, [eax+128*1] |
| 102 |
movq mm1, [esi+8] |
movq mm2, [eax+128*2] |
| 103 |
|
movq mm3, [eax+128*3] |
| 104 |
|
movq mm4, [eax+128*4] |
| 105 |
|
movq mm5, [eax+128*5] |
| 106 |
|
add eax, 8+128 |
| 107 |
pand mm0, mm7 |
pand mm0, mm7 |
| 108 |
|
pand mm1, mm7 |
| 109 |
|
pand mm2, mm7 |
| 110 |
|
pand mm3, mm7 |
| 111 |
|
pand mm4, mm7 |
| 112 |
|
pand mm5, mm7 |
| 113 |
|
|
| 114 |
|
MAKE_LOAD 0 |
| 115 |
|
MAKE_LOAD 1 |
| 116 |
|
MAKE_LOAD 2 |
| 117 |
|
MAKE_LOAD 3 |
| 118 |
|
MAKE_LOAD 4 |
| 119 |
|
MAKE_LOAD 5 |
| 120 |
|
MAKE_LOAD 6 |
| 121 |
|
MAKE_LOAD 7 |
| 122 |
|
MAKE_LOAD 8 |
| 123 |
|
MAKE_LOAD 9 |
| 124 |
|
MAKE_LOAD 10 |
| 125 |
|
MAKE_LOAD 11 |
| 126 |
|
MAKE_LOAD 12 |
| 127 |
|
MAKE_LOAD 13 |
| 128 |
|
MAKE_LOAD 14 |
| 129 |
|
|
| 130 |
|
movq mm7, [mult_mask] |
| 131 |
|
packssdw mm0, mm1 |
| 132 |
|
packssdw mm2, mm3 |
| 133 |
|
packssdw mm4, mm5 |
| 134 |
|
packssdw mm0, mm2 |
| 135 |
|
packssdw mm4, mm6 |
| 136 |
|
pcmpeqw mm0, mm6 |
| 137 |
|
pcmpeqw mm4, mm6 |
| 138 |
|
pcmpeqw mm0, mm6 |
| 139 |
|
pcmpeqw mm4, mm6 |
| 140 |
|
psrlw mm0, 15 |
| 141 |
|
psrlw mm4, 15 |
| 142 |
|
packuswb mm0, mm4 |
| 143 |
|
pmaddwd mm0, mm7 |
| 144 |
|
|
|
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 |
|
| 145 |
movq mm1, mm0 |
movq mm1, mm0 |
| 146 |
psrlq mm1, 32 |
psrlq mm1, 32 |
| 147 |
lea esi, [esi + 128] |
paddusb mm0, mm1 |
| 148 |
|
|
| 149 |
por mm0, mm1 |
movd eax, mm0 |
| 150 |
movd ebx, mm0 |
shr eax, 8 |
| 151 |
|
and eax, 0x3F |
| 152 |
test ebx, ebx |
ret |
| 153 |
jz .next |
ENDFUNC |
|
or eax, edx ; cbp |= 1 << (5-i) |
|
| 154 |
|
|
|
.next |
|
|
shr edx,1 |
|
|
jnc .loop |
|
| 155 |
|
|
| 156 |
pop esi |
%ifidn __OUTPUT_FORMAT__,elf |
| 157 |
pop ebx |
section ".note.GNU-stack" noalloc noexec nowrite progbits |
| 158 |
|
%endif |
| 159 |
|
|
|
ret |
|
Parent Directory
| 
Revision Log
| 
Patch