hvirtual/mpeg2enc/bdist2_mmx.s

   1 ;
   2 ;  bdist2_mmx.s:  MMX optimized bidirectional squared 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 ; * squared 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 bdist2_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 ;  int s;
  39
  40 ; Handy macros for readbility
  41
  42 %define pf [ebp+8]
  43 %define pb [ebp+12]
  44 %define p2 [ebp+16]
  45 %define lx [ebp+20]
  46 %define hxf [ebp+24]
  47 %define hyf [ebp+28]
  48 %define hxb [ebp+32]
  49 %define hyb [ebp+36]
  50 %define h   [ebp+40]
  51
  52
  53 %define pfa [esp+4]
  54 %define pfb [esp+8]
  55 %define pfc [esp+12]
  56 %define pba [esp+16]
  57 %define pbb [esp+20]
  58 %define pbc [esp+24]
  59
  60 SECTION .text
  61 global bdist2_mmx
  62
  63 align 32
  64 bdist2_mmx:
  65         push ebp                        ; save frame pointer
  66         mov ebp, esp            ; link
  67         push ebx
  68         push ecx
  69         push edx
  70         push esi
  71         push edi
  72
  73         ;;
  74         ;; Make space for local variables on stack
  75         sub       esp, 32
  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 = s (accumulated sym)
 102         mov       eax, esi
 103
 104         mov       edi, h
 105         test      edi, edi  ; h = 0?
 106         jle       near bdist2exit
 107
 108         pxor      mm7, mm7
 109         pxor      mm6, mm6
 110         pcmpeqw   mm5, mm5
 111         psubw     mm6, mm5
 112         psllw     mm6, 1
 113
 114 bdist2top:
 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
 146         mov       eax, pb
 147         mov       ebx, pba
 148         mov       ecx, pbb
 149         mov       edx, pbc
 150         movq      mm2, [eax]
 151         movq      mm3, mm2
 152         punpcklbw mm2, mm7
 153         punpckhbw mm3, mm7
 154         movq      mm4, [ebx]
 155         movq      mm5, mm4
 156         punpcklbw mm4, mm7
 157         punpckhbw mm5, mm7
 158         paddw     mm2, mm4
 159         paddw     mm3, mm5
 160         movq      mm4, [ecx]
 161         movq      mm5, mm4
 162         punpcklbw mm4, mm7
 163         punpckhbw mm5, mm7
 164         paddw     mm2, mm4
 165         paddw     mm3, mm5
 166         movq      mm4, [edx]
 167         movq      mm5, mm4
 168         punpcklbw mm4, mm7
 169         punpckhbw mm5, mm7
 170         paddw     mm2, mm4
 171         paddw     mm3, mm5
 172
 173         paddw     mm2, mm6
 174         paddw     mm3, mm6
 175         psrlw     mm2, 2
 176         psrlw     mm3, 2
 177
 178         paddw     mm0, mm2
 179         paddw     mm1, mm3
 180         psrlw     mm6, 1
 181         paddw     mm0, mm6
 182         paddw     mm1, mm6
 183         psllw     mm6, 1
 184         psrlw     mm0, 1
 185         psrlw     mm1, 1
 186
 187         mov       eax, p2
 188         movq      mm2, [eax]
 189         movq      mm3, mm2
 190         punpcklbw mm2, mm7
 191         punpckhbw mm3, mm7
 192
 193         psubw     mm0, mm2
 194         psubw     mm1, mm3
 195         pmaddwd   mm0, mm0
 196         pmaddwd   mm1, mm1
 197         paddd     mm0, mm1
 198
 199         movd      eax, mm0
 200         psrlq     mm0, 32
 201         movd      ebx, mm0
 202         add       esi, eax
 203         add       esi, ebx
 204
 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
 236         mov       eax, pb
 237         mov       ebx, pba
 238         mov       ecx, pbb
 239         mov       edx, pbc
 240         movq      mm2, [eax+8]
 241         movq      mm3, mm2
 242         punpcklbw mm2, mm7
 243         punpckhbw mm3, mm7
 244         movq      mm4, [ebx+8]
 245         movq      mm5, mm4
 246         punpcklbw mm4, mm7
 247         punpckhbw mm5, mm7
 248         paddw     mm2, mm4
 249         paddw     mm3, mm5
 250         movq      mm4, [ecx+8]
 251         movq      mm5, mm4
 252         punpcklbw mm4, mm7
 253         punpckhbw mm5, mm7
 254         paddw     mm2, mm4
 255         paddw     mm3, mm5
 256         movq      mm4, [edx+8]
 257         movq      mm5, mm4
 258         punpcklbw mm4, mm7
 259         punpckhbw mm5, mm7
 260         paddw     mm2, mm4
 261         paddw     mm3, mm5
 262         paddw     mm2, mm6
 263         paddw     mm3, mm6
 264         psrlw     mm2, 2
 265         psrlw     mm3, 2
 266
 267         paddw     mm0, mm2
 268         paddw     mm1, mm3
 269         psrlw     mm6, 1
 270         paddW     mm0, mm6
 271         paddw     mm1, mm6
 272         psllw     mm6, 1
 273         psrlw     mm0, 1
 274         psrlw     mm1, 1
 275
 276         mov       eax, p2
 277         movq      mm2, [eax+8]
 278         movq      mm3, mm2
 279         punpcklbw mm2, mm7
 280         punpckhbw mm3, mm7
 281
 282         psubw     mm0, mm2
 283         psubw     mm1, mm3
 284         pmaddwd   mm0, mm0
 285         pmaddwd   mm1, mm1
 286         paddd     mm0, mm1
 287
 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 bdist2top
 307     mov       eax, esi
 308
 309 bdist2exit:
 310
 311         ;;
 312         ;; Get rid of local variables
 313         add esp, 32
 314
 315         ;; Retore (callee saves convention...)
 316         ;;
 317         pop edi
 318         pop esi
 319         pop edx
 320         pop ecx
 321         pop ebx
 322
 323         pop ebp                 ; restore stack pointer
 324
 325         emms                    ; clear mmx registers
 326         ret
 327