mpeg2enc/bdist1_mmx.s

   1 ;
   2 ;  bdist1_mmx.s:  mmX optimized bidirectional absolute distance sum
   3 ;
   4 ;  Original believed to be Copyright (C) 2000 Brent Byeler
   5 ;
   6 ;  This program is free software; you can reaxstribute it and/or
   7 ;  modify it under the terms of the GNU General Public License
   8 ;  as published by the Free Software Foundation; either version 2
   9 ;  of the License, or (at your option) any later version.
  10 ;
  11 ;  This program is distributed in the hope that it will be useful,
  12 ;  but WITHOUT ANY WARRANTY; without even the implied warranty of
  13 ;  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  14 ;  GNU General Public License for more details.
  15 ;
  16 ;  You should have received a copy of the GNU General Public License
  17 ;  along with this program; if not, write to the Free Software
  18 ;  Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA  02111-1307, USA.
  19 ;
  20
  21 ;/*
  22 ; * absolute difference error between a (16*h) block and a bidirectional
  23 ; * prediction
  24 ; *
  25 ; * p2: address of top left pel of block
  26 ; * pf,hxf,hyf: address and half pel flags of forward ref. block
  27 ; * pb,hxb,hyb: address and half pel flags of backward ref. block
  28 ; * h: height of block
  29 ; * lx: distance (in bytes) of vertically adjacent pels in p2,pf,pb
  30 ; * mmX version
  31 ; */
  32
  33 ;int bdist1_mmx(
  34 ;unsigned char *pf, unsigned char *pb, unsigned char *p2,
  35 ;int lx, int hxf, int hyf, int hxb, int hyb, int h)
  36 ;{
  37 ;  unsigned char *pfa,*pfb,*pfc,*pba,*pbb,*pbc;
  38
  39 ; Handy macros for readbility
  40
  41 %define pf [ebp+8]
  42 %define pb [ebp+12]
  43 %define p2 [ebp+16]
  44 %define lx [ebp+20]
  45 %define hxf [ebp+24]
  46 %define hyf [ebp+28]
  47 %define hxb [ebp+32]
  48 %define hyb [ebp+36]
  49 %define h   [ebp+40]
  50
  51
  52 %define pfa [esp+4]
  53 %define pfb [esp+8]
  54 %define pfc [esp+12]
  55 %define pba [esp+16]
  56 %define pbb [esp+20]
  57 %define pbc [esp+24]
  58
  59 SECTION .text
  60 global bdist1_mmx
  61
  62 align 32
  63 bdist1_mmx:
  64         push ebp                        ; save frame pointer
  65         mov ebp, esp            ; link
  66         push ebx
  67         push ecx
  68         push edx
  69         push esi
  70         push edi
  71
  72         ;;
  73         ;; Make space for local variables on stack
  74         sub       esp, 32
  75
  76
  77         mov       edx, hxb
  78         mov       eax, hxf
  79         mov       esi, lx
  80
  81         mov       ecx, pf
  82         add       ecx, eax
  83         mov       pfa, ecx
  84         mov       ecx, esi
  85         imul      ecx, hyf
  86         mov       ebx, pf
  87         add       ecx, ebx
  88         mov       pfb, ecx
  89         add       eax, ecx
  90         mov       pfc, eax
  91         mov       eax, pb
  92         add       eax, edx
  93         mov       pba, eax
  94         mov       eax, esi
  95         imul      eax, hyb
  96         mov       ecx, pb
  97         add       eax, ecx
  98         mov       pbb, eax
  99         add       edx, eax
 100         mov       pbc, edx
 101         xor       esi, esi         ; esi is "s" the accumulator
 102         mov       eax, esi
 103
 104         mov       edi, h
 105         test      edi, edi  ; h = 0?
 106         jle       near bdist1exit
 107
 108         pxor      mm7, mm7
 109         pxor      mm6, mm6
 110         pcmpeqw   mm5, mm5
 111         psubw     mm6, mm5
 112         psllw     mm6, 1
 113
 114 bdist1top:
 115         mov       eax, pf
 116         mov       ebx, pfa
 117         mov       ecx, pfb
 118         mov       edx, pfc
 119         movq      mm0, [eax]
 120         movq      mm1, mm0
 121         punpcklbw mm0, mm7
 122         punpckhbw mm1, mm7
 123         movq      mm2, [ebx]
 124         movq      mm3, mm2
 125         punpcklbw mm2, mm7
 126         punpckhbw mm3, mm7
 127         paddw     mm0, mm2
 128         paddw     mm1, mm3
 129         movq      mm2, [ecx]
 130         movq      mm3, mm2
 131         punpcklbw mm2, mm7
 132         punpckhbw mm3, mm7
 133         paddw     mm0, mm2
 134         paddw     mm1, mm3
 135         movq      mm2, [edx]
 136         movq      mm3, mm2
 137         punpcklbw mm2, mm7
 138         punpckhbw mm3, mm7
 139         paddw     mm0, mm2
 140         paddw     mm1, mm3
 141         paddw     mm0, mm6
 142         paddw     mm1, mm6
 143         psrlw     mm0, 2
 144         psrlw     mm1, 2
 145         mov       eax, pb
 146         mov       ebx, pba
 147         mov       ecx, pbb
 148         mov       edx, pbc
 149         movq      mm2, [eax]
 150         movq      mm3, mm2
 151         punpcklbw mm2, mm7
 152         punpckhbw mm3, mm7
 153         movq      mm4, [ebx]
 154         movq      mm5, mm4
 155         punpcklbw mm4, mm7
 156         punpckhbw mm5, mm7
 157         paddw     mm2, mm4
 158         paddw     mm3, mm5
 159         movq      mm4, [ecx]
 160         movq      mm5, mm4
 161         punpcklbw mm4, mm7
 162         punpckhbw mm5, mm7
 163         paddw     mm2, mm4
 164         paddw     mm3, mm5
 165         movq      mm4, [edx]
 166         movq      mm5, mm4
 167         punpcklbw mm4, mm7
 168         punpckhbw mm5, mm7
 169         paddw     mm2, mm4
 170         paddw     mm3, mm5
 171         paddw     mm2, mm6
 172         paddw     mm3, mm6
 173         psrlw     mm2, 2
 174         psrlw     mm3, 2
 175         paddw     mm0, mm2
 176         paddw     mm1, mm3
 177         psrlw     mm6, 1
 178         paddw     mm0, mm6
 179         paddw     mm1, mm6
 180         psllw     mm6, 1
 181         psrlw     mm0, 1
 182         psrlw     mm1, 1
 183         packuswb  mm0, mm1
 184
 185         mov       eax, p2
 186         movq      mm1, [eax]
 187         movq      mm2, mm0
 188         psubusb   mm0, mm1
 189         psubusb   mm1, mm2
 190         por       mm0, mm1
 191         movq      mm1, mm0
 192         punpcklbw mm0, mm7
 193         punpckhbw mm1, mm7
 194         paddw     mm0, mm1
 195         movq      mm1, mm0
 196         punpcklwd mm0, mm7
 197         punpckhwd mm1, mm7
 198
 199         paddd     mm0, mm1
 200         movd      eax, mm0
 201         psrlq     mm0, 32
 202         movd      ebx, mm0
 203         add       esi, eax
 204         add       esi, ebx
 205         mov       eax, pf
 206         mov       ebx, pfa
 207         mov       ecx, pfb
 208         mov       edx, pfc
 209         movq      mm0, [eax+8]
 210         movq      mm1, mm0
 211         punpcklbw mm0, mm7
 212         punpckhbw mm1, mm7
 213         movq      mm2, [ebx+8]
 214         movq      mm3, mm2
 215         punpcklbw mm2, mm7
 216         punpckhbw mm3, mm7
 217         paddw     mm0, mm2
 218         paddw     mm1, mm3
 219         movq      mm2, [ecx+8]
 220         movq      mm3, mm2
 221         punpcklbw mm2, mm7
 222         punpckhbw mm3, mm7
 223         paddw     mm0, mm2
 224         paddw     mm1, mm3
 225         movq      mm2, [edx+8]
 226         movq      mm3, mm2
 227         punpcklbw mm2, mm7
 228         punpckhbw mm3, mm7
 229         paddw     mm0, mm2
 230         paddw     mm1, mm3
 231         paddw     mm0, mm6
 232         paddw     mm1, mm6
 233         psrlw     mm0, 2
 234         psrlw     mm1, 2
 235         mov       eax, pb
 236         mov       ebx, pba
 237         mov       ecx, pbb
 238         mov       edx, pbc
 239         movq      mm2, [eax+8]
 240         movq      mm3, mm2
 241         punpcklbw mm2, mm7
 242         punpckhbw mm3, mm7
 243         movq      mm4, [ebx+8]
 244         movq      mm5, mm4
 245         punpcklbw mm4, mm7
 246         punpckhbw mm5, mm7
 247         paddw     mm2, mm4
 248         paddw     mm3, mm5
 249         movq      mm4, [ecx+8]
 250         movq      mm5, mm4
 251         punpcklbw mm4, mm7
 252         punpckhbw mm5, mm7
 253         paddw     mm2, mm4
 254         paddw     mm3, mm5
 255         movq      mm4, [edx+8]
 256         movq      mm5, mm4
 257         punpcklbw mm4, mm7
 258         punpckhbw mm5, mm7
 259         paddw     mm2, mm4
 260         paddw     mm3, mm5
 261         paddw     mm2, mm6
 262         paddw     mm3, mm6
 263         psrlw     mm2, 2
 264         psrlw     mm3, 2
 265         paddw     mm0, mm2
 266         paddw     mm1, mm3
 267         psrlw     mm6, 1
 268         paddW     mm0, mm6
 269         paddw     mm1, mm6
 270         psllw     mm6, 1
 271         psrlw     mm0, 1
 272         psrlw     mm1, 1
 273         packuswb  mm0, mm1
 274         mov       eax, p2
 275         movq      mm1, [eax+8]
 276         movq      mm2, mm0
 277         psubusb   mm0, mm1
 278         psubusb   mm1, mm2
 279         por       mm0, mm1
 280         movq      mm1, mm0
 281         punpcklbw mm0, mm7
 282         punpckhbw mm1, mm7
 283         paddw     mm0, mm1
 284         movq      mm1, mm0
 285         punpcklwd mm0, mm7
 286         punpckhwd mm1, mm7
 287         paddd     mm0, mm1
 288         movd      eax, mm0
 289         psrlq     mm0, 32
 290         movd      ebx, mm0
 291         add       esi, eax
 292         add       esi, ebx
 293
 294     mov       eax, lx
 295         add       p2, eax
 296         add       pf, eax
 297         add       pfa, eax
 298         add       pfb, eax
 299         add       pfc, eax
 300         add       pb, eax
 301         add       pba, eax
 302         add       pbb, eax
 303         add       pbc, eax
 304
 305         dec       edi
 306         jg        near bdist1top
 307     mov       eax, esi
 308
 309 bdist1exit:
 310
 311
 312         ;;
 313         ;; Get rid of local variables
 314         add esp, 32
 315
 316         ;; Retore (callee saves convention...)
 317         ;;
 318         pop edi
 319         pop esi
 320         pop edx
 321         pop ecx
 322         pop ebx
 323
 324         pop ebp                 ; restore stack pointer
 325
 326         emms                    ; clear mmx registers
 327         ret
 328
 329