libmpeg3/video/reconmmx.s

   1 ADD_1:          dd 01010101h, 01010101h
   2 MASK_AND:       dd 7f7f7f7fh, 7f7f7f7fh
   3 PLUS_384:       dd 01800180h, 01800180h
   4 PLUS_128:       dd 00800080h, 00800080h
   5
   6 %assign LocalFrameSize  0
   7 %assign RegisterStorageSize  16
   8
   9 ; Arguments:
  10 %assign source                    LocalFrameSize + RegisterStorageSize +  4
  11 %assign dest                      LocalFrameSize + RegisterStorageSize +  8
  12 %assign lx2                       LocalFrameSize + RegisterStorageSize + 12
  13 %assign h                         LocalFrameSize + RegisterStorageSize + 16
  14
  15 ; Locals (on local stack frame)
  16
  17
  18 ; extern void C rec_mmx (
  19 ;                                     unsigned char *source,
  20 ;                                     unsigned char *dest,
  21 ;                                     int lx2,
  22 ;                                     int h
  23 ;
  24 ;  The local variables are on the stack,
  25 ;
  26
  27 global recva_mmx
  28 global recvac_mmx
  29 global rech_mmx
  30 global rechc_mmx
  31 global add_block_mmx
  32 global set_block_mmx
  33
  34
  35   align 16
  36 rech_mmx:
  37   push       esi
  38   push       edi
  39   push       ecx
  40   push       ebx
  41   mov        esi, [esp+source]
  42   mov        edi, [esp+dest]
  43   mov        ecx, [esp+h]
  44   mov        ebx, [esp+lx2]
  45   movq       mm5, [MASK_AND]
  46   movq       mm6, [ADD_1]
  47 .rech1:
  48   movq       mm0,[esi]
  49   movq       mm1,[esi+1]
  50   movq       mm2,[esi+8]
  51   movq       mm3,[esi+9]
  52   psrlw      mm0,1
  53   psrlw      mm1,1
  54   psrlw      mm2,1
  55   psrlw      mm3,1
  56   pand       mm0,mm5
  57   pand       mm1,mm5
  58   pand       mm2,mm5
  59   pand       mm3,mm5
  60   paddusb    mm0,mm1
  61   paddusb    mm2,mm3
  62   paddusb    mm0,mm6
  63   paddusb    mm2,mm6
  64   movq       [edi],mm0
  65   add        esi,ebx
  66   movq       [edi+8],mm2
  67   add        edi,ebx
  68   dec        ecx
  69   jnz        .rech1
  70   emms
  71   pop        ebx
  72   pop        ecx
  73   pop        edi
  74   pop        esi
  75   ret
  76
  77   align 16
  78 rechc_mmx:
  79   push       esi
  80   push       edi
  81   push       ecx
  82   push       ebx
  83 ;  sub        esp, LocalFrameSize
  84   mov        esi, [esp+source]
  85   mov        edi, [esp+dest]
  86   mov        ecx, [esp+h]
  87   mov        ebx, [esp+lx2]
  88   movq       mm5, [MASK_AND]
  89   movq       mm6, [ADD_1]
  90 .rechc1:
  91   movq       mm0,[esi]
  92   movq       mm1,[esi+1]
  93   psrlw      mm0,1
  94   psrlw      mm1,1
  95   pand       mm0,mm5
  96   pand       mm1,mm5
  97   paddusb    mm0,mm1
  98   paddusb    mm0,mm6
  99   movq       [edi],mm0
 100   add        edi,ebx
 101   add        esi,ebx
 102   dec        ecx
 103   jnz        .rechc1
 104   emms
 105 ;  add        esp, LocalFrameSize
 106   pop        ebx
 107   pop        ecx
 108   pop        edi
 109   pop        esi
 110   ret
 111
 112
 113
 114 %assign RegisterStorageSize  20
 115 %assign source                    LocalFrameSize + RegisterStorageSize +  4
 116 %assign dest                      LocalFrameSize + RegisterStorageSize +  8
 117 %assign lx                        LocalFrameSize + RegisterStorageSize + 12
 118 %assign lx2                       LocalFrameSize + RegisterStorageSize + 16
 119 %assign h                         LocalFrameSize + RegisterStorageSize + 20
 120
 121   align 16
 122 recva_mmx:
 123   push       esi
 124   push       edi
 125   push       ecx
 126   push       ebx
 127   push       edx
 128   mov        esi, [esp+source]
 129   mov        edi, [esp+dest]
 130   mov        ecx, [esp+h]
 131   mov        ebx, [esp+lx2]
 132   mov        edx, [esp+lx]
 133   movq       mm7, [MASK_AND]
 134   movq       mm6, [ADD_1]
 135 .recva1:
 136   movq       mm0,[esi]
 137   movq       mm1,[esi+edx]
 138   movq       mm2,[esi+8]
 139   movq       mm3,[esi+edx+8]
 140   movq       mm4,[edi]
 141   movq       mm5,[edi+8]
 142   psrlw      mm0,1
 143   psrlw      mm1,1
 144   psrlw      mm2,1
 145   psrlw      mm3,1
 146   psrlw      mm4,1
 147   psrlw      mm5,1
 148   pand       mm0,mm7
 149   pand       mm1,mm7
 150   pand       mm2,mm7
 151   pand       mm3,mm7
 152   pand       mm4,mm7
 153   pand       mm5,mm7
 154   paddusb    mm0,mm1
 155   paddusb    mm2,mm3
 156   paddusb    mm0,mm6
 157   paddusb    mm2,mm6
 158   psrlw      mm0,1
 159   psrlw      mm2,1
 160   pand       mm0,mm7
 161   pand       mm2,mm7
 162   paddusb    mm4,mm0
 163   paddusb    mm5,mm2
 164   paddusb    mm4,mm6
 165   paddusb    mm5,mm6
 166   movq       [edi],mm4
 167   movq       [edi+8],mm5
 168   add        edi,ebx
 169   add        esi,ebx
 170   dec        ecx
 171   jnz        near .recva1
 172   emms
 173   pop        edx
 174   pop        ebx
 175   pop        ecx
 176   pop        edi
 177   pop        esi
 178   ret
 179
 180   align 16
 181 recvac_mmx:
 182   push       esi
 183   push       edi
 184   push       ecx
 185   push       ebx
 186   push       edx
 187   mov        esi, [esp+source]
 188   mov        edi, [esp+dest]
 189   mov        ecx, [esp+h]
 190   mov        ebx, [esp+lx2]
 191   mov        edx, [esp+lx]
 192   movq       mm5, [MASK_AND]
 193   movq       mm6, [ADD_1]
 194 .recvac1:
 195   movq       mm0,[esi]
 196   movq       mm1,[esi+edx]
 197   movq       mm4,[edi]
 198   psrlw      mm0,1
 199   psrlw      mm1,1
 200   psrlw      mm4,1
 201   pand       mm0,mm5
 202   pand       mm1,mm5
 203   pand       mm4,mm5
 204   paddusb    mm0,mm1
 205   paddusb    mm0,mm6
 206   psrlw      mm0,1
 207   pand       mm0,mm5
 208   paddusb    mm4,mm0
 209   paddusb    mm4,mm6
 210   movq       [edi],mm4
 211   add        edi,ebx
 212   add        esi,ebx
 213   dec        ecx
 214   jnz        .recvac1
 215   emms
 216   pop        edx
 217   pop        ebx
 218   pop        ecx
 219   pop        edi
 220   pop        esi
 221   ret
 222
 223 %assign RegisterStorageSize  20
 224 %assign rfp                       LocalFrameSize + RegisterStorageSize +  4
 225 %assign bp                        LocalFrameSize + RegisterStorageSize +  8
 226 %assign iincr                     LocalFrameSize + RegisterStorageSize + 12
 227
 228 ; FIXME clipping needs to be done
 229
 230   align 16
 231 add_block_mmx:
 232   push       esi
 233   push       edi
 234   push       ecx
 235   push       ebx
 236   push       edx
 237   mov        esi, [esp+bp]
 238   mov        edi, [esp+rfp]
 239   mov        ebx, [esp+iincr]
 240 ;  movq       mm7, [PLUS_384]
 241   mov        ecx,8
 242   pxor        mm2,mm2           ; clear
 243 %rep 8
 244   movq       mm0, [edi]         ; get dest
 245   movq       mm1,mm0
 246   punpcklbw  mm0,mm2
 247   punpckhbw  mm1,mm2
 248   paddsw     mm0, [esi]
 249   paddsw     mm1, [esi+8]
 250 ;  paddsw     mm0, mm7
 251 ;  paddsw     mm1, mm7
 252   packuswb   mm0,mm1
 253   movq       [edi], mm0
 254   add        edi,ebx
 255   add        esi,16
 256 %endrep
 257   emms
 258   pop        edx
 259   pop        ebx
 260   pop        ecx
 261   pop        edi
 262   pop        esi
 263   ret
 264
 265   align 16
 266 set_block_mmx:
 267   push       esi
 268   push       edi
 269   push       ecx
 270   push       ebx
 271   push       edx
 272   mov        esi, [esp+bp]
 273   mov        edi, [esp+rfp]
 274   mov        ebx, [esp+iincr]
 275   movq       mm7, [PLUS_128]
 276 %rep 4
 277   movq       mm0, [esi]
 278   movq       mm1, [esi+8]
 279   paddsw     mm0, mm7
 280   movq       mm2, [esi+16]
 281   paddsw     mm1, mm7
 282   movq       mm3, [esi+24]
 283   paddsw     mm2, mm7
 284   packuswb   mm0, mm1
 285   paddsw     mm3, mm7
 286   movq       [edi], mm0
 287   packuswb   mm2, mm3
 288   add        edi, ebx
 289   add        esi, 32
 290   movq       [edi], mm2
 291   add        edi, ebx
 292 %endrep
 293   emms
 294   pop        edx
 295   pop        ebx
 296   pop        ecx
 297   pop        edi
 298   pop        esi
 299   ret
 300
 301