hvirtual/mpeg2enc/mblock_sad_mmx.s

   1 ;;;
   2 ;;;  mblock_sad_mmxe.s:
   3 ;;;
   4 ;;; Enhanced MMX optimized Sum Absolute Differences routines for macroblocks
   5 ;;; (interpolated, 1-pel, 2*2 sub-sampled pel and 4*4 sub-sampled pel)
   6 ;
   7 ;  dist1_* Original Copyright (C) 2000 Chris Atenasio <chris@crud.net>
   8 ;  Enhancements and rest Copyright (C) 2000 Andrew Stevens <as@comlab.ox.ac.uk>
   9
  10 ;
  11 ;  This program is free software; you can redistribute it and/or
  12 ;  modify it under the terms of the GNU General Public License
  13 ;  as published by the Free Software Foundation; either version 2
  14 ;  of the License, or (at your option) any later version.
  15 ;
  16 ;  This program is distributed in the hope that it will be useful,
  17 ;  but WITHOUT ANY WARRANTY; without even the implied warranty of
  18 ;  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  19 ;  GNU General Public License for more details.
  20 ;
  21 ;  You should have received a copy of the GNU General Public License
  22 ;  along with this program; if not, write to the Free Software
  23 ;  Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA  02111-1307, USA.
  24 ;
  25 ;
  26 ;
  27
  28
  29 global dist1_00_mmx
  30
  31 ; int dist1_mmx(unsigned char *blk1,unsigned char *blk2,int lx,int h, int distlim);
  32 ; N.b. distlim is *ignored* as testing for it is more expensive than the
  33 ; occasional saving by aborting the computionation early...
  34 ; esi = p1 (init:               blk1)
  35 ; edi = p2 (init:               blk2)
  36 ; ebx = distlim
  37 ; ecx = rowsleft (init:  h)
  38 ; edx = lx;
  39
  40 ; mm0 = distance accumulators (4 words)
  41 ; mm1 = temp
  42 ; mm2 = temp
  43 ; mm3 = temp
  44 ; mm4 = temp
  45 ; mm5 = temp
  46 ; mm6 = 0
  47 ; mm7 = temp
  48
  49
  50 align 32
  51 dist1_00_mmx:
  52         push ebp                ; save frame pointer
  53         mov ebp, esp
  54
  55         push ebx                ; Saves registers (called saves convention in
  56         push ecx                ; x86 GCC it seems)
  57         push edx                ;
  58         push esi
  59         push edi
  60
  61         pxor mm0, mm0                           ; zero acculumators
  62         pxor mm6, mm6
  63         mov esi, [ebp+8]                        ; get p1
  64         mov edi, [ebp+12]                       ; get p2
  65         mov edx, [ebp+16]                       ; get lx
  66         mov  ecx, [ebp+20]                      ; get rowsleft
  67         ;mov ebx, [ebp+24]              ; distlim
  68         jmp nextrowmm00
  69 align 32
  70 nextrowmm00:
  71         movq mm4, [esi]         ; load first 8 bytes of p1 row
  72         movq mm5, [edi]         ; load first 8 bytes of p2 row
  73
  74         movq mm7, mm4       ; mm5 = abs(mm4-mm5)
  75         psubusb mm7, mm5
  76         psubusb mm5, mm4
  77         paddb   mm5, mm7
  78
  79         ;;  Add the abs(mm4-mm5) bytes to the accumulators
  80         movq mm2, [esi+8]               ; load second 8 bytes of p1 row (interleaved)
  81         movq  mm7,mm5                           ; mm7 := [i :   B0..3, mm1]W
  82         punpcklbw mm7,mm6
  83         movq mm3, [edi+8]               ; load second 8 bytes of p2 row (interleaved)
  84         paddw mm0, mm7
  85         punpckhbw mm5,mm6
  86         paddw mm0, mm5
  87
  88                 ;; This is logically where the mm2, mm3 loads would go...
  89
  90         movq mm7, mm2       ; mm3 = abs(mm2-mm3)
  91         psubusb mm7, mm3
  92         psubusb mm3, mm2
  93         paddb   mm3, mm7
  94
  95         ;;  Add the abs(mm4-mm5) bytes to the accumulators
  96         movq  mm7,mm3                           ; mm7 := [i :   B0..3, mm1]W
  97         punpcklbw mm7,mm6
  98         punpckhbw mm3,mm6
  99         paddw mm0, mm7
 100
 101         add esi, edx            ; update pointer to next row
 102         add edi, edx            ; ditto
 103
 104         paddw mm0, mm3
 105
 106
 107
 108         sub  ecx,1
 109         jnz near nextrowmm00
 110
 111 returnmm00:
 112
 113                 ;; Sum the Accumulators
 114         movq  mm5, mm0                          ; mm5 := [W0+W2,W1+W3, mm0
 115         psrlq mm5, 32
 116         movq  mm4, mm0
 117         paddw mm4, mm5
 118
 119         movq  mm7, mm4              ; mm6 := [W0+W2+W1+W3, mm0]
 120         psrlq mm7, 16
 121         paddw mm4, mm7
 122         movd eax, mm4           ; store return value
 123         and  eax, 0xffff
 124
 125         pop edi
 126         pop esi
 127         pop edx
 128         pop ecx
 129         pop ebx
 130
 131         pop ebp
 132
 133         emms                    ; clear mmx registers
 134         ret
 135 ;
 136 ;;;  dist1_01_mmx.s:  mmx1 optimised 7bit*8 word absolute difference sum
 137 ;;;     We're reduce to seven bits as otherwise we also have to mess
 138 ;;;     horribly with carries and signed only comparisons make the code
 139 ;;;     simply enormous (and probably barely faster than a simple loop).
 140 ;;;     Since signals with a bona-fide 8bit res will be rare we simply
 141 ;;;     take the precision hit...
 142 ;;;     Actually we don't worry about carries from the low-order bits
 143 ;;;     either so 1/4 of the time we'll be 1 too low...
 144 ;;;
 145 ;  Copyright (C) 2000 Andrew Stevens <as@comlab.ox.ac.uk>
 146
 147 ;
 148 ;  This program is free software; you can redistribute it and/or
 149 ;  modify it under the terms of the GNU General Public License
 150 ;  as published by the Free Software Foundation; either version 2
 151 ;  of the License, or (at your option) any later version.
 152 ;
 153 ;  This program is distributed in the hope that it will be useful,
 154 ;  but WITHOUT ANY WARRANTY; without even the implied warranty of
 155 ;  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
 156 ;  GNU General Public License for more details.
 157 ;
 158 ;  You should have received a copy of the GNU General Public License
 159 ;  along with this program; if not, write to the Free Software
 160 ;  Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA  02111-1307, USA.
 161 ;
 162 ;
 163 ;
 164
 165
 166 global dist1_01_mmx
 167
 168 ; int dist1_01_mmx(unsigned char *p1,unsigned char *p2,int lx,int h);
 169
 170 ; esi = p1 (init:               blk1)
 171 ; edi = p2 (init:               blk2)
 172 ; ecx = rowsleft (init:  h)
 173 ; edx = lx;
 174
 175 ; mm0 = distance accumulators (4 words)
 176 ; mm1 = bytes p2
 177 ; mm2 = bytes p1
 178 ; mm3 = bytes p1+1
 179 ; mm4 = temp 4 bytes in words interpolating p1, p1+1
 180 ; mm5 = temp 4 bytes in words from p2
 181 ; mm6 = temp comparison bit mask p1,p2
 182 ; mm7 = temp comparison bit mask p2,p1
 183
 184
 185 align 32
 186 dist1_01_mmx:
 187         push ebp                ; save stack pointer
 188         mov ebp, esp    ; so that we can do this
 189
 190         push ebx                ; Saves registers (called saves convention in
 191         push ecx                ; x86 GCC it seems)
 192         push edx                ;
 193         push esi
 194         push edi
 195
 196         pxor mm0, mm0                           ; zero acculumators
 197
 198         mov esi, [ebp+8]                        ; get p1
 199         mov edi, [ebp+12]                       ; get p2
 200         mov edx, [ebp+16]                       ; get lx
 201         mov ecx, [ebp+20]                       ; rowsleft := h
 202         jmp nextrowmm01                                 ; snap to it
 203 align 32
 204 nextrowmm01:
 205
 206                 ;;
 207                 ;; First 8 bytes of row
 208                 ;;
 209
 210                 ;; First 4 bytes of 8
 211
 212         movq mm4, [esi]             ; mm4 := first 4 bytes p1
 213         pxor mm7, mm7
 214         movq mm2, mm4                           ;  mm2 records all 8 bytes
 215         punpcklbw mm4, mm7            ;  First 4 bytes p1 in Words...
 216
 217         movq mm6, [esi+1]                       ;  mm6 := first 4 bytes p1+1
 218         movq mm3, mm6               ;  mm3 records all 8 bytes
 219         punpcklbw mm6, mm7
 220         paddw mm4, mm6              ;  mm4 := First 4 bytes interpolated in words
 221         psrlw mm4, 1
 222
 223         movq mm5, [edi]                         ; mm5:=first 4 bytes of p2 in words
 224         movq mm1, mm5
 225         punpcklbw mm5, mm7
 226
 227         movq  mm7,mm4
 228         pcmpgtw mm7,mm5         ; mm7 := [i : W0..3,mm4>mm5]
 229
 230         movq  mm6,mm4           ; mm6 := [i : W0..3, (mm4-mm5)*(mm4-mm5 > 0)]
 231         psubw mm6,mm5
 232         pand  mm6, mm7
 233
 234         paddw mm0, mm6                          ; Add to accumulator
 235
 236         movq  mm6,mm5       ; mm6 := [i : W0..3,mm5>mm4]
 237         pcmpgtw mm6,mm4
 238         psubw mm5,mm4           ; mm5 := [i : B0..7, (mm5-mm4)*(mm5-mm4 > 0)]
 239         pand  mm5, mm6
 240
 241         paddw mm0, mm5                          ; Add to accumulator
 242
 243                 ;; Second 4 bytes of 8
 244
 245         movq mm4, mm2               ; mm4 := Second 4 bytes p1 in words
 246         pxor  mm7, mm7
 247         punpckhbw mm4, mm7
 248         movq mm6, mm3                   ; mm6 := Second 4 bytes p1+1 in words
 249         punpckhbw mm6, mm7
 250
 251         paddw mm4, mm6          ;  mm4 := First 4 Interpolated bytes in words
 252         psrlw mm4, 1
 253
 254         movq mm5, mm1                   ; mm5:= second 4 bytes of p2 in words
 255         punpckhbw mm5, mm7
 256
 257         movq  mm7,mm4
 258         pcmpgtw mm7,mm5         ; mm7 := [i : W0..3,mm4>mm5]
 259
 260         movq  mm6,mm4           ; mm6 := [i : W0..3, (mm4-mm5)*(mm4-mm5 > 0)]
 261         psubw mm6,mm5
 262         pand  mm6, mm7
 263
 264         paddw mm0, mm6                          ; Add to accumulator
 265
 266         movq  mm6,mm5       ; mm6 := [i : W0..3,mm5>mm4]
 267         pcmpgtw mm6,mm4
 268         psubw mm5,mm4           ; mm5 := [i : B0..7, (mm5-mm4)*(mm5-mm4 > 0)]
 269         pand  mm5, mm6
 270
 271         paddw mm0, mm5                          ; Add to accumulator
 272
 273
 274                 ;;
 275                 ;; Second 8 bytes of row
 276                 ;;
 277                 ;; First 4 bytes of 8
 278
 279         movq mm4, [esi+8]             ; mm4 := first 4 bytes p1+8
 280         pxor mm7, mm7
 281         movq mm2, mm4                           ;  mm2 records all 8 bytes
 282         punpcklbw mm4, mm7            ;  First 4 bytes p1 in Words...
 283
 284         movq mm6, [esi+9]                       ;  mm6 := first 4 bytes p1+9
 285         movq mm3, mm6               ;  mm3 records all 8 bytes
 286         punpcklbw mm6, mm7
 287         paddw mm4, mm6              ;  mm4 := First 4 bytes interpolated in words
 288         psrlw mm4, 1
 289
 290         movq mm5, [edi+8]                               ; mm5:=first 4 bytes of p2+8 in words
 291         movq mm1, mm5
 292         punpcklbw mm5, mm7
 293
 294         movq  mm7,mm4
 295         pcmpgtw mm7,mm5         ; mm7 := [i : W0..3,mm4>mm5]
 296
 297         movq  mm6,mm4           ; mm6 := [i : W0..3, (mm4-mm5)*(mm4-mm5 > 0)]
 298         psubw mm6,mm5
 299         pand  mm6, mm7
 300
 301         paddw mm0, mm6                          ; Add to accumulator
 302
 303         movq  mm6,mm5       ; mm6 := [i : W0..3,mm5>mm4]
 304         pcmpgtw mm6,mm4
 305         psubw mm5,mm4           ; mm5 := [i : B0..7, (mm5-mm4)*(mm5-mm4 > 0)]
 306         pand  mm5, mm6
 307
 308         paddw mm0, mm5                          ; Add to accumulator
 309
 310                 ;; Second 4 bytes of 8
 311
 312         movq mm4, mm2               ; mm4 := Second 4 bytes p1 in words
 313         pxor  mm7, mm7
 314         punpckhbw mm4, mm7
 315         movq mm6, mm3                   ; mm6 := Second 4 bytes p1+1 in words
 316         punpckhbw mm6, mm7
 317
 318         paddw mm4, mm6          ;  mm4 := First 4 Interpolated bytes in words
 319         psrlw mm4, 1
 320
 321         movq mm5, mm1                   ; mm5:= second 4 bytes of p2 in words
 322         punpckhbw mm5, mm7
 323
 324         movq  mm7,mm4
 325         pcmpgtw mm7,mm5         ; mm7 := [i : W0..3,mm4>mm5]
 326
 327         movq  mm6,mm4           ; mm6 := [i : W0..3, (mm4-mm5)*(mm4-mm5 > 0)]
 328         psubw mm6,mm5
 329         pand  mm6, mm7
 330
 331         paddw mm0, mm6                          ; Add to accumulator
 332
 333         movq  mm6,mm5       ; mm6 := [i : W0..3,mm5>mm4]
 334         pcmpgtw mm6,mm4
 335         psubw mm5,mm4           ; mm5 := [i : B0..7, (mm5-mm4)*(mm5-mm4 > 0)]
 336         pand  mm5, mm6
 337
 338         paddw mm0, mm5                          ; Add to accumulator
 339
 340
 341                 ;;
 342                 ;;      Loop termination condition... and stepping
 343                 ;;
 344
 345         add esi, edx            ; update pointer to next row
 346         add edi, edx            ; ditto
 347
 348         sub  ecx,1
 349         test ecx, ecx           ; check rowsleft
 350         jnz near nextrowmm01
 351
 352
 353                 ;; Sum the Accumulators
 354         movq  mm4, mm0
 355         psrlq mm4, 32
 356         paddw mm0, mm4
 357         movq  mm6, mm0
 358         psrlq mm6, 16
 359         paddw mm0, mm6
 360         movd eax, mm0           ; store return value
 361         and  eax, 0xffff
 362
 363         pop edi
 364         pop esi
 365         pop edx
 366         pop ecx
 367         pop ebx
 368
 369         pop ebp                 ; restore stack pointer
 370
 371         emms                    ; clear mmx registers
 372         ret                     ; we now return you to your regular programming
 373 ;
 374 ;;;  dist1_01_mmx.s:  mmx1 optimised 7bit*8 word absolute difference sum
 375 ;;;     We're reduce to seven bits as otherwise we also have to mess
 376 ;;;     horribly with carries and signed only comparisons make the code
 377 ;;;     simply enormous (and probably barely faster than a simple loop).
 378 ;;;     Since signals with a bona-fide 8bit res will be rare we simply
 379 ;;;     take the precision hit...
 380 ;;;     Actually we don't worry about carries from the low-order bits
 381 ;;;     either so 1/4 of the time we'll be 1 too low...
 382 ;;;
 383 ;  Copyright (C) 2000 Andrew Stevens <as@comlab.ox.ac.uk>
 384
 385 ;
 386 ;  This program is free software; you can redistribute it and/or
 387 ;  modify it under the terms of the GNU General Public License
 388 ;  as published by the Free Software Foundation; either version 2
 389 ;  of the License, or (at your option) any later version.
 390 ;
 391 ;  This program is distributed in the hope that it will be useful,
 392 ;  but WITHOUT ANY WARRANTY; without even the implied warranty of
 393 ;  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
 394 ;  GNU General Public License for more details.
 395 ;
 396 ;  You should have received a copy of the GNU General Public License
 397 ;  along with this program; if not, write to the Free Software
 398 ;  Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA  02111-1307, USA.
 399 ;
 400 ;
 401 ;
 402
 403
 404 global dist1_10_mmx
 405
 406 ; int dist1_10_mmx(unsigned char *p1,unsigned char *p2,int lx,int h);
 407
 408 ; esi = p1 (init:               blk1)
 409 ; edi = p2 (init:               blk2)
 410 ; ebx = p1+lx
 411 ; ecx = rowsleft (init:  h)
 412 ; edx = lx;
 413
 414 ; mm0 = distance accumulators (4 words)
 415 ; mm1 = bytes p2
 416 ; mm2 = bytes p1
 417 ; mm3 = bytes p1+1
 418 ; mm4 = temp 4 bytes in words interpolating p1, p1+1
 419 ; mm5 = temp 4 bytes in words from p2
 420 ; mm6 = temp comparison bit mask p1,p2
 421 ; mm7 = temp comparison bit mask p2,p1
 422
 423
 424 align 32
 425 dist1_10_mmx:
 426         push ebp                ; save stack pointer
 427         mov ebp, esp    ; so that we can do this
 428
 429         push ebx                ; Saves registers (called saves convention in
 430         push ecx                ; x86 GCC it seems)
 431         push edx                ;
 432         push esi
 433         push edi
 434
 435         pxor mm0, mm0                           ; zero acculumators
 436
 437         mov esi, [ebp+8]                        ; get p1
 438         mov edi, [ebp+12]                       ; get p2
 439         mov edx, [ebp+16]                       ; get lx
 440         mov ecx, [ebp+20]                       ; rowsleft := h
 441         mov ebx, esi
 442     add ebx, edx
 443         jmp nextrowmm10                                 ; snap to it
 444 align 32
 445 nextrowmm10:
 446
 447                 ;;
 448                 ;; First 8 bytes of row
 449                 ;;
 450
 451                 ;; First 4 bytes of 8
 452
 453         movq mm4, [esi]             ; mm4 := first 4 bytes p1
 454         pxor mm7, mm7
 455         movq mm2, mm4                           ;  mm2 records all 8 bytes
 456         punpcklbw mm4, mm7            ;  First 4 bytes p1 in Words...
 457
 458         movq mm6, [ebx]                     ;  mm6 := first 4 bytes p1+lx
 459         movq mm3, mm6               ;  mm3 records all 8 bytes
 460         punpcklbw mm6, mm7
 461         paddw mm4, mm6              ;  mm4 := First 4 bytes interpolated in words
 462         psrlw mm4, 1
 463
 464         movq mm5, [edi]                         ; mm5:=first 4 bytes of p2 in words
 465         movq mm1, mm5
 466         punpcklbw mm5, mm7
 467
 468         movq  mm7,mm4
 469         pcmpgtw mm7,mm5         ; mm7 := [i : W0..3,mm4>mm5]
 470
 471         movq  mm6,mm4           ; mm6 := [i : W0..3, (mm4-mm5)*(mm4-mm5 > 0)]
 472         psubw mm6,mm5
 473         pand  mm6, mm7
 474
 475         paddw mm0, mm6                          ; Add to accumulator
 476
 477         movq  mm6,mm5       ; mm6 := [i : W0..3,mm5>mm4]
 478         pcmpgtw mm6,mm4
 479         psubw mm5,mm4           ; mm5 := [i : B0..7, (mm5-mm4)*(mm5-mm4 > 0)]
 480         pand  mm5, mm6
 481
 482         paddw mm0, mm5                          ; Add to accumulator
 483
 484                 ;; Second 4 bytes of 8
 485
 486         movq mm4, mm2               ; mm4 := Second 4 bytes p1 in words
 487         pxor  mm7, mm7
 488         punpckhbw mm4, mm7
 489         movq mm6, mm3                   ; mm6 := Second 4 bytes p1+1 in words
 490         punpckhbw mm6, mm7
 491
 492         paddw mm4, mm6          ;  mm4 := First 4 Interpolated bytes in words
 493         psrlw mm4, 1
 494
 495         movq mm5, mm1                   ; mm5:= second 4 bytes of p2 in words
 496         punpckhbw mm5, mm7
 497
 498         movq  mm7,mm4
 499         pcmpgtw mm7,mm5         ; mm7 := [i : W0..3,mm4>mm5]
 500
 501         movq  mm6,mm4           ; mm6 := [i : W0..3, (mm4-mm5)*(mm4-mm5 > 0)]
 502         psubw mm6,mm5
 503         pand  mm6, mm7
 504
 505         paddw mm0, mm6                          ; Add to accumulator
 506
 507         movq  mm6,mm5       ; mm6 := [i : W0..3,mm5>mm4]
 508         pcmpgtw mm6,mm4
 509         psubw mm5,mm4           ; mm5 := [i : B0..7, (mm5-mm4)*(mm5-mm4 > 0)]
 510         pand  mm5, mm6
 511
 512         paddw mm0, mm5                          ; Add to accumulator
 513
 514
 515                 ;;
 516                 ;; Second 8 bytes of row
 517                 ;;
 518                 ;; First 4 bytes of 8
 519
 520         movq mm4, [esi+8]             ; mm4 := first 4 bytes p1+8
 521         pxor mm7, mm7
 522         movq mm2, mm4                           ;  mm2 records all 8 bytes
 523         punpcklbw mm4, mm7            ;  First 4 bytes p1 in Words...
 524
 525         movq mm6, [ebx+8]                       ;  mm6 := first 4 bytes p1+lx+8
 526         movq mm3, mm6               ;  mm3 records all 8 bytes
 527         punpcklbw mm6, mm7
 528         paddw mm4, mm6              ;  mm4 := First 4 bytes interpolated in words
 529         psrlw mm4, 1
 530
 531         movq mm5, [edi+8]                               ; mm5:=first 4 bytes of p2+8 in words
 532         movq mm1, mm5
 533         punpcklbw mm5, mm7
 534
 535         movq  mm7,mm4
 536         pcmpgtw mm7,mm5         ; mm7 := [i : W0..3,mm4>mm5]
 537
 538         movq  mm6,mm4           ; mm6 := [i : W0..3, (mm4-mm5)*(mm4-mm5 > 0)]
 539         psubw mm6,mm5
 540         pand  mm6, mm7
 541
 542         paddw mm0, mm6                          ; Add to accumulator
 543
 544         movq  mm6,mm5       ; mm6 := [i : W0..3,mm5>mm4]
 545         pcmpgtw mm6,mm4
 546         psubw mm5,mm4           ; mm5 := [i : B0..7, (mm5-mm4)*(mm5-mm4 > 0)]
 547         pand  mm5, mm6
 548
 549         paddw mm0, mm5                          ; Add to accumulator
 550
 551                 ;; Second 4 bytes of 8
 552
 553         movq mm4, mm2               ; mm4 := Second 4 bytes p1 in words
 554         pxor  mm7, mm7
 555         punpckhbw mm4, mm7
 556         movq mm6, mm3                   ; mm6 := Second 4 bytes p1+1 in words
 557         punpckhbw mm6, mm7
 558
 559         paddw mm4, mm6          ;  mm4 := First 4 Interpolated bytes in words
 560         psrlw mm4, 1
 561
 562         movq mm5, mm1                   ; mm5:= second 4 bytes of p2 in words
 563         punpckhbw mm5, mm7
 564
 565         movq  mm7,mm4
 566         pcmpgtw mm7,mm5         ; mm7 := [i : W0..3,mm4>mm5]
 567
 568         movq  mm6,mm4           ; mm6 := [i : W0..3, (mm4-mm5)*(mm4-mm5 > 0)]
 569         psubw mm6,mm5
 570         pand  mm6, mm7
 571
 572         paddw mm0, mm6                          ; Add to accumulator
 573
 574         movq  mm6,mm5       ; mm6 := [i : W0..3,mm5>mm4]
 575         pcmpgtw mm6,mm4
 576         psubw mm5,mm4           ; mm5 := [i : B0..7, (mm5-mm4)*(mm5-mm4 > 0)]
 577         pand  mm5, mm6
 578
 579         paddw mm0, mm5                          ; Add to accumulator
 580
 581
 582                 ;;
 583                 ;;      Loop termination condition... and stepping
 584                 ;;
 585
 586         add esi, edx            ; update pointer to next row
 587         add edi, edx            ; ditto
 588         add ebx, edx
 589
 590         sub  ecx,1
 591         test ecx, ecx           ; check rowsleft
 592         jnz near nextrowmm10
 593
 594                 ;; Sum the Accumulators
 595         movq  mm4, mm0
 596         psrlq mm4, 32
 597         paddw mm0, mm4
 598         movq  mm6, mm0
 599         psrlq mm6, 16
 600         paddw mm0, mm6
 601         movd eax, mm0           ; store return value
 602         and  eax, 0xffff
 603
 604
 605         pop edi
 606         pop esi
 607         pop edx
 608         pop ecx
 609         pop ebx
 610
 611         pop ebp                 ; restore stack pointer
 612
 613         emms                    ; clear mmx registers
 614         ret                     ; we now return you to your regular programming
 615 ;
 616 ;;;  dist1_01_mmx.s:
 617 ;;;
 618 ;  Copyright (C) 2000 Andrew Stevens <as@comlab.ox.ac.uk>
 619
 620 ;
 621 ;  This program is free software; you can redistribute it and/or
 622 ;  modify it under the terms of the GNU General Public License
 623 ;  as published by the Free Software Foundation; either version 2
 624 ;  of the License, or (at your option) any later version.
 625 ;
 626 ;  This program is distributed in the hope that it will be useful,
 627 ;  but WITHOUT ANY WARRANTY; without even the implied warranty of
 628 ;  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
 629 ;  GNU General Public License for more details.
 630 ;
 631 ;  You should have received a copy of the GNU General Public License
 632 ;  along with this program; if not, write to the Free Software
 633 ;  Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA  02111-1307, USA.
 634 ;
 635 ;
 636 ;
 637
 638
 639 global dist1_11_mmx
 640
 641 ; int dist1_11_mmx(unsigned char *p1,unsigned char *p2,int lx,int h);
 642
 643 ; esi = p1 (init:               blk1)
 644 ; edi = p2 (init:               blk2)
 645 ; ebx = p1+lx
 646 ; ecx = rowsleft (init:  h)
 647 ; edx = lx;
 648
 649 ; mm0 = distance accumulators (4 words)
 650 ; mm1 = bytes p2
 651 ; mm2 = bytes p1
 652 ; mm3 = bytes p1+lx
 653 ; I'd love to find someplace to stash p1+1 and p1+lx+1's bytes
 654 ; but I don't think thats going to happen in iA32-land...
 655 ; mm4 = temp 4 bytes in words interpolating p1, p1+1
 656 ; mm5 = temp 4 bytes in words from p2
 657 ; mm6 = temp comparison bit mask p1,p2
 658 ; mm7 = temp comparison bit mask p2,p1
 659
 660
 661 align 32
 662 dist1_11_mmx:
 663         push ebp                ; save stack pointer
 664         mov ebp, esp    ; so that we can do this
 665
 666         push ebx                ; Saves registers (called saves convention in
 667         push ecx                ; x86 GCC it seems)
 668         push edx                ;
 669         push esi
 670         push edi
 671
 672         pxor mm0, mm0                           ; zero acculumators
 673
 674         mov esi, [ebp+8]                        ; get p1
 675         mov edi, [ebp+12]                       ; get p2
 676         mov edx, [ebp+16]                       ; get lx
 677         mov ecx, [ebp+20]                       ; rowsleft := h
 678         mov ebx, esi
 679     add ebx, edx
 680         jmp nextrowmm11                                 ; snap to it
 681 align 32
 682 nextrowmm11:
 683
 684                 ;;
 685                 ;; First 8 bytes of row
 686                 ;;
 687
 688                 ;; First 4 bytes of 8
 689
 690         movq mm4, [esi]             ; mm4 := first 4 bytes p1
 691         pxor mm7, mm7
 692         movq mm2, mm4                           ;  mm2 records all 8 bytes
 693         punpcklbw mm4, mm7            ;  First 4 bytes p1 in Words...
 694
 695         movq mm6, [ebx]                     ;  mm6 := first 4 bytes p1+lx
 696         movq mm3, mm6               ;  mm3 records all 8 bytes
 697         punpcklbw mm6, mm7
 698         paddw mm4, mm6
 699
 700
 701         movq mm5, [esi+1]                       ; mm5 := first 4 bytes p1+1
 702         punpcklbw mm5, mm7            ;  First 4 bytes p1 in Words...
 703         paddw mm4, mm5
 704         movq mm6, [ebx+1]           ;  mm6 := first 4 bytes p1+lx+1
 705         punpcklbw mm6, mm7
 706         paddw mm4, mm6
 707
 708         psrlw mm4, 2                ; mm4 := First 4 bytes interpolated in words
 709
 710         movq mm5, [edi]                         ; mm5:=first 4 bytes of p2 in words
 711         movq mm1, mm5
 712         punpcklbw mm5, mm7
 713
 714         movq  mm7,mm4
 715         pcmpgtw mm7,mm5         ; mm7 := [i : W0..3,mm4>mm5]
 716
 717         movq  mm6,mm4           ; mm6 := [i : W0..3, (mm4-mm5)*(mm4-mm5 > 0)]
 718         psubw mm6,mm5
 719         pand  mm6, mm7
 720
 721         paddw mm0, mm6                          ; Add to accumulator
 722
 723         movq  mm6,mm5       ; mm6 := [i : W0..3,mm5>mm4]
 724         pcmpgtw mm6,mm4
 725         psubw mm5,mm4           ; mm5 := [i : B0..7, (mm5-mm4)*(mm5-mm4 > 0)]
 726         pand  mm5, mm6
 727
 728         paddw mm0, mm5                          ; Add to accumulator
 729
 730                 ;; Second 4 bytes of 8
 731
 732         movq mm4, mm2               ; mm4 := Second 4 bytes p1 in words
 733         pxor  mm7, mm7
 734         punpckhbw mm4, mm7
 735         movq mm6, mm3                   ; mm6 := Second 4 bytes p1+1 in words
 736         punpckhbw mm6, mm7
 737         paddw mm4, mm6
 738
 739         movq mm5, [esi+1]                       ; mm5 := first 4 bytes p1+1
 740         punpckhbw mm5, mm7          ;  First 4 bytes p1 in Words...
 741         paddw mm4, mm5
 742         movq mm6, [ebx+1]           ;  mm6 := first 4 bytes p1+lx+1
 743         punpckhbw mm6, mm7
 744         paddw mm4, mm6
 745
 746         psrlw mm4, 2                ; mm4 := First 4 bytes interpolated in words
 747
 748         movq mm5, mm1                   ; mm5:= second 4 bytes of p2 in words
 749         punpckhbw mm5, mm7
 750
 751         movq  mm7,mm4
 752         pcmpgtw mm7,mm5         ; mm7 := [i : W0..3,mm4>mm5]
 753
 754         movq  mm6,mm4           ; mm6 := [i : W0..3, (mm4-mm5)*(mm4-mm5 > 0)]
 755         psubw mm6,mm5
 756         pand  mm6, mm7
 757
 758         paddw mm0, mm6                          ; Add to accumulator
 759
 760         movq  mm6,mm5       ; mm6 := [i : W0..3,mm5>mm4]
 761         pcmpgtw mm6,mm4
 762         psubw mm5,mm4           ; mm5 := [i : B0..7, (mm5-mm4)*(mm5-mm4 > 0)]
 763         pand  mm5, mm6
 764
 765         paddw mm0, mm5                          ; Add to accumulator
 766
 767
 768                 ;;
 769                 ;; Second 8 bytes of row
 770                 ;;
 771                 ;; First 4 bytes of 8
 772
 773         movq mm4, [esi+8]             ; mm4 := first 4 bytes p1+8
 774         pxor mm7, mm7
 775         movq mm2, mm4                           ;  mm2 records all 8 bytes
 776         punpcklbw mm4, mm7            ;  First 4 bytes p1 in Words...
 777
 778         movq mm6, [ebx+8]                           ;  mm6 := first 4 bytes p1+lx+8
 779         movq mm3, mm6               ;  mm3 records all 8 bytes
 780         punpcklbw mm6, mm7
 781         paddw mm4, mm6
 782
 783
 784         movq mm5, [esi+9]                       ; mm5 := first 4 bytes p1+9
 785         punpcklbw mm5, mm7            ;  First 4 bytes p1 in Words...
 786         paddw mm4, mm5
 787         movq mm6, [ebx+9]           ;  mm6 := first 4 bytes p1+lx+9
 788         punpcklbw mm6, mm7
 789         paddw mm4, mm6
 790
 791         psrlw mm4, 2                ; mm4 := First 4 bytes interpolated in words
 792
 793         movq mm5, [edi+8]                               ; mm5:=first 4 bytes of p2+8 in words
 794         movq mm1, mm5
 795         punpcklbw mm5, mm7
 796
 797         movq  mm7,mm4
 798         pcmpgtw mm7,mm5         ; mm7 := [i : W0..3,mm4>mm5]
 799
 800         movq  mm6,mm4           ; mm6 := [i : W0..3, (mm4-mm5)*(mm4-mm5 > 0)]
 801         psubw mm6,mm5
 802         pand  mm6, mm7
 803
 804         paddw mm0, mm6                          ; Add to accumulator
 805
 806         movq  mm6,mm5       ; mm6 := [i : W0..3,mm5>mm4]
 807         pcmpgtw mm6,mm4
 808         psubw mm5,mm4           ; mm5 := [i : B0..7, (mm5-mm4)*(mm5-mm4 > 0)]
 809         pand  mm5, mm6
 810
 811         paddw mm0, mm5                          ; Add to accumulator
 812
 813                 ;; Second 4 bytes of 8
 814
 815         movq mm4, mm2               ; mm4 := Second 4 bytes p1 in words
 816         pxor  mm7, mm7
 817         punpckhbw mm4, mm7
 818         movq mm6, mm3                   ; mm6 := Second 4 bytes p1+1 in words
 819         punpckhbw mm6, mm7
 820         paddw mm4, mm6
 821
 822         movq mm5, [esi+9]                       ; mm5 := first 4 bytes p1+1
 823         punpckhbw mm5, mm7          ;  First 4 bytes p1 in Words...
 824         paddw mm4, mm5
 825         movq mm6, [ebx+9]           ;  mm6 := first 4 bytes p1+lx+1
 826         punpckhbw mm6, mm7
 827         paddw mm4, mm6
 828
 829         psrlw mm4, 2                ; mm4 := First 4 bytes interpolated in words
 830
 831         movq mm5, mm1                   ; mm5:= second 4 bytes of p2 in words
 832         punpckhbw mm5, mm7
 833
 834         movq  mm7,mm4
 835         pcmpgtw mm7,mm5         ; mm7 := [i : W0..3,mm4>mm5]
 836
 837         movq  mm6,mm4           ; mm6 := [i : W0..3, (mm4-mm5)*(mm4-mm5 > 0)]
 838         psubw mm6,mm5
 839         pand  mm6, mm7
 840
 841         paddw mm0, mm6                          ; Add to accumulator
 842
 843         movq  mm6,mm5       ; mm6 := [i : W0..3,mm5>mm4]
 844         pcmpgtw mm6,mm4
 845         psubw mm5,mm4           ; mm5 := [i : B0..7, (mm5-mm4)*(mm5-mm4 > 0)]
 846         pand  mm5, mm6
 847
 848         paddw mm0, mm5                          ; Add to accumulator
 849
 850
 851                 ;;
 852                 ;;      Loop termination condition... and stepping
 853                 ;;
 854
 855         add esi, edx            ; update pointer to next row
 856         add edi, edx            ; ditto
 857         add ebx, edx
 858
 859         sub  ecx,1
 860         test ecx, ecx           ; check rowsleft
 861         jnz near nextrowmm11
 862
 863                 ;; Sum the Accumulators
 864         movq  mm4, mm0
 865         psrlq mm4, 32
 866         paddw mm0, mm4
 867         movq  mm6, mm0
 868         psrlq mm6, 16
 869         paddw mm0, mm6
 870         movd eax, mm0           ; store return value
 871         and  eax, 0xffff
 872
 873         pop edi
 874         pop esi
 875         pop edx
 876         pop ecx
 877         pop ebx
 878
 879         pop ebp                 ; restore stack pointer
 880
 881         emms                    ; clear mmx registers
 882         ret                     ; we now return you to your regular programming
 883
 884
 885 global dist22_mmx
 886
 887 ; int dist22_mmx(unsigned char *blk1,unsigned char *blk2,int lx,int h);
 888
 889 ; eax = p1 (init:               blk1)
 890 ; ebx = p2 (init:                blk2)
 891 ; ecx = rowsleft (init:  h)
 892 ; edx = lx;
 893
 894 ; mm0 = distance accumulators (4 words)
 895 ; mm1 = temp
 896 ; mm2 = temp
 897 ; mm3 = temp
 898 ; mm4 = temp
 899 ; mm5 = temp
 900 ; mm6 = 0
 901 ; mm7 = temp
 902
 903
 904 align 32
 905 dist22_mmx:
 906         push ebp                ; save stack pointer
 907         mov ebp, esp    ; so that we can do this
 908
 909         push ebx                ; Saves registers (called saves convention in
 910         push ecx                ; x86 GCC it seems)
 911         push edx                ;
 912
 913         pxor mm0, mm0                           ; zero acculumators
 914         pxor mm6, mm6
 915         mov eax, [ebp+8]                        ; get p1
 916         mov ebx, [ebp+12]                       ; get p2
 917         mov edx, [ebp+16]                       ; get lx
 918
 919         mov  ecx, [ebp+20]                      ; get rowsleft
 920
 921         jmp nextrow                                     ; snap to it
 922 align 32
 923 nextrow:
 924         movq mm4, [eax]         ; load 8 bytes of p1
 925         movq mm5, [ebx]         ; load 8 bytes of p2
 926
 927         movq mm7, mm4       ; mm5 = abs(*p1-*p2)
 928         psubusb mm7, mm5
 929         psubusb mm5, mm4
 930         add eax, edx            ; update pointer to next row
 931         paddb   mm5,mm7
 932
 933                 ;;  Add the mm5 bytes to the accumulatores
 934         movq  mm7,mm5
 935         punpcklbw mm7,mm6
 936         paddw mm0, mm7
 937         punpckhbw mm5,mm6
 938         add ebx, edx            ; update pointer to next row
 939         paddw mm0, mm5
 940
 941         movq mm4, [eax]         ; load 8 bytes of p1 (next row)
 942         movq mm5, [ebx]         ; load 8 bytes of p2 (next row)
 943
 944         movq mm7, mm4       ; mm5 = abs(*p1-*p2)
 945         psubusb mm7, mm5
 946         psubusb mm5, mm4
 947         add eax, edx            ; update pointer to next row
 948         paddb   mm5,mm7
 949
 950                 ;;  Add the mm5 bytes to the accumulatores
 951         movq  mm7,mm5
 952         punpcklbw mm7,mm6
 953         add ebx, edx            ; update pointer to next row
 954         paddw mm0, mm7
 955         punpckhbw mm5,mm6
 956         sub  ecx,2
 957         paddw mm0, mm5
 958
 959
 960         jnz nextrow
 961
 962                 ;; Sum the Accumulators
 963         movq  mm1, mm0
 964         psrlq mm1, 16
 965         movq  mm2, mm0
 966         psrlq mm2, 32
 967         movq  mm3, mm0
 968         psrlq mm3, 48
 969         paddw mm0, mm1
 970         paddw mm2, mm3
 971         paddw mm0, mm2
 972
 973         movd eax, mm0           ; store return value
 974         and  eax, 0xffff
 975
 976         pop edx                 ; pop pop
 977         pop ecx                 ; fizz fizz
 978         pop ebx                 ; ia86 needs a fizz instruction
 979
 980         pop ebp                 ; restore stack pointer
 981
 982         emms                    ; clear mmx registers
 983         ret                     ; we now return you to your regular programming
 984
 985
 986
 987
 988 global dist44_mmx
 989
 990 ; int dist44_mmx(unsigned char *blk1,unsigned char *blk2,int qlx,int qh);
 991
 992 ; eax = p1
 993 ; ebx = p2
 994 ; ecx = temp
 995 ; edx = qlx;
 996 ; esi = rowsleft
 997
 998 ; mm0 = distance accumulator left block p1
 999 ; mm1 = distance accumulator right block p1
1000 ; mm2 = 0
1001 ; mm3 = right block of p1
1002 ; mm4 = left block of p1
1003 ; mm5 = p2
1004 ; mm6 = temp
1005 ; mm7 = temp
1006
1007 align 32
1008 dist44_mmx:
1009         push ebp                ; save stack pointer
1010         mov ebp, esp            ; so that we can do this
1011
1012         push ebx
1013         push ecx
1014         push edx
1015         push esi
1016
1017         pxor mm0, mm0           ; zero acculumator
1018         pxor mm1, mm1
1019         pxor mm2, mm2
1020         mov eax, [ebp+8]        ; get p1
1021         mov ebx, [ebp+12]       ; get p2
1022         mov edx, [ebp+16]       ; get qlx
1023         mov esi, [ebp+20]       ; get rowsleft
1024         jmp nextrowqd           ; snap to it
1025 align 32
1026 nextrowqd:
1027
1028                 ;;
1029                 ;; Beware loop obfuscated by interleaving to try to
1030                 ;; hide latencies...
1031                 ;;
1032         movq mm4, [eax]                         ; mm4 =  first 4 bytes of p1 in words
1033         movq mm5, [ebx]             ; mm5 = 4 bytes of p2 in words
1034         movq mm3, mm4
1035         punpcklbw mm4, mm2
1036         punpcklbw mm5, mm2
1037
1038         movq mm7, mm4
1039         movq mm6, mm5
1040         psubusw mm7, mm5
1041         psubusw mm6, mm4
1042
1043         add eax, edx                    ; update a pointer to next row
1044 ;       punpckhbw mm3, mm2                      ; mm3 = 2nd 4 bytes of p1 in words
1045
1046         paddw   mm7, mm6
1047         paddw mm0, mm7                          ; Add absolute differences to left block accumulators
1048
1049 ;       movq mm7,mm3
1050 ;       psubusw mm7, mm5
1051 ;       psubusw mm5, mm3
1052
1053         add ebx, edx            ; update a pointer to next row
1054         sub   esi, 1
1055
1056 ;       paddw   mm7, mm5
1057 ;       paddw mm1, mm7                          ; Add absolute differences to right block accumulators
1058
1059
1060
1061         jnz nextrowqd
1062
1063                 ;;              Sum the accumulators
1064
1065         movq  mm4, mm0
1066         psrlq mm4, 32
1067         paddw mm0, mm4
1068         movq  mm6, mm0
1069         psrlq mm6, 16
1070         paddw mm0, mm6
1071         movd eax, mm0           ; store return value
1072
1073 ;       movq  mm4, mm1
1074 ;       psrlq mm4, 32
1075 ;       paddw mm1, mm4
1076 ;       movq  mm6, mm1
1077 ;       psrlq mm6, 16
1078 ;       paddw mm1, mm6
1079 ;       movd ebx, mm1
1080
1081         and  eax, 0xffff
1082 ;       sal  ebx, 16
1083 ;       or   eax, ebx
1084
1085         pop esi
1086         pop edx
1087         pop ecx
1088         pop ebx
1089
1090         pop ebp                 ; restore stack pointer
1091
1092         emms                    ; clear mmx registers
1093         ret                     ; we now return you to your regular programming