Parent Directory
| 
Revision Log
Revision 1793 - (view) (download)
| 1 : | edgomez | 1382 | ;/**************************************************************************** |
| 2 : | edgomez | 851 | ; * |
| 3 : | edgomez | 1382 | ; * XVID MPEG-4 VIDEO CODEC |
| 4 : | ; * - K7 optimized SAD operators - | ||
| 5 : | edgomez | 851 | ; * |
| 6 : | edgomez | 1382 | ; * Copyright(C) 2002 Jaan Kalda |
| 7 : | edgomez | 851 | ; * |
| 8 : | edgomez | 1382 | ; * This program is free software; you can redistribute it and/or modify it |
| 9 : | ; * under the terms of the GNU General Public License as published by | ||
| 10 : | ; * the Free Software Foundation; either version 2 of the License, or | ||
| 11 : | ; * (at your option) any later version. | ||
| 12 : | edgomez | 851 | ; * |
| 13 : | edgomez | 1382 | ; * This program is distributed in the hope that it will be useful, |
| 14 : | ; * but WITHOUT ANY WARRANTY; without even the implied warranty of | ||
| 15 : | ; * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | ||
| 16 : | ; * GNU General Public License for more details. | ||
| 17 : | edgomez | 851 | ; * |
| 18 : | edgomez | 1382 | ; * You should have received a copy of the GNU General Public License |
| 19 : | ; * along with this program; if not, write to the Free Software | ||
| 20 : | ; * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA | ||
| 21 : | edgomez | 851 | ; * |
| 22 : | Isibaar | 1793 | ; * $Id: sad_3dne.asm,v 1.9 2008-11-11 20:46:24 Isibaar Exp $ |
| 23 : | edgomez | 1382 | ; * |
| 24 : | ; ***************************************************************************/ | ||
| 25 : | edgomez | 851 | |
| 26 : | edgomez | 1382 | ; these 3dne functions are compatible with iSSE, but are optimized specifically |
| 27 : | ; for K7 pipelines | ||
| 28 : | edgomez | 851 | |
| 29 : | edgomez | 1382 | BITS 32 |
| 30 : | |||
| 31 : | %macro cglobal 1 | ||
| 32 : | edgomez | 851 | %ifdef PREFIX |
| 33 : | edgomez | 1535 | %ifdef MARK_FUNCS |
| 34 : | edgomez | 1540 | global _%1:function %1.endfunc-%1 |
| 35 : | %define %1 _%1:function %1.endfunc-%1 | ||
| 36 : | Isibaar | 1793 | %define ENDFUNC .endfunc |
| 37 : | edgomez | 1535 | %else |
| 38 : | global _%1 | ||
| 39 : | %define %1 _%1 | ||
| 40 : | Isibaar | 1793 | %define ENDFUNC |
| 41 : | edgomez | 1535 | %endif |
| 42 : | edgomez | 851 | %else |
| 43 : | edgomez | 1535 | %ifdef MARK_FUNCS |
| 44 : | edgomez | 1540 | global %1:function %1.endfunc-%1 |
| 45 : | Isibaar | 1793 | %define ENDFUNC .endfunc |
| 46 : | edgomez | 1535 | %else |
| 47 : | global %1 | ||
| 48 : | Isibaar | 1793 | %define ENDFUNC |
| 49 : | edgomez | 1535 | %endif |
| 50 : | edgomez | 851 | %endif |
| 51 : | %endmacro | ||
| 52 : | |||
| 53 : | edgomez | 1382 | ;============================================================================= |
| 54 : | ; Read only data | ||
| 55 : | ;============================================================================= | ||
| 56 : | |||
| 57 : | edgomez | 851 | %ifdef FORMAT_COFF |
| 58 : | edgomez | 1519 | SECTION .rodata |
| 59 : | edgomez | 851 | %else |
| 60 : | edgomez | 1519 | SECTION .rodata align=16 |
| 61 : | edgomez | 851 | %endif |
| 62 : | |||
| 63 : | edgomez | 1382 | ALIGN 16 |
| 64 : | mmx_one: | ||
| 65 : | times 4 dw 1 | ||
| 66 : | edgomez | 851 | |
| 67 : | edgomez | 1382 | ;============================================================================= |
| 68 : | ; Helper macros | ||
| 69 : | ;============================================================================= | ||
| 70 : | edgomez | 851 | |
| 71 : | edgomez | 1382 | ;; %1 block number (0..4) |
| 72 : | %macro SAD_16x16_SSE 1 | ||
| 73 : | movq mm7, [eax] | ||
| 74 : | movq mm6, [eax+8] | ||
| 75 : | psadbw mm7, [edx] | ||
| 76 : | psadbw mm6, [edx+8] | ||
| 77 : | %if (%1) | ||
| 78 : | paddd mm1, mm5 | ||
| 79 : | %endif | ||
| 80 : | movq mm5, [eax+ecx] | ||
| 81 : | movq mm4, [eax+ecx+8] | ||
| 82 : | psadbw mm5, [edx+ecx] | ||
| 83 : | psadbw mm4, [edx+ecx+8] | ||
| 84 : | movq mm3, [eax+2*ecx] | ||
| 85 : | movq mm2, [eax+2*ecx+8] | ||
| 86 : | psadbw mm3, [edx+2*ecx] | ||
| 87 : | psadbw mm2, [edx+2*ecx+8] | ||
| 88 : | %if (%1) | ||
| 89 : | movd [esp+4*(%1-1)], mm1 | ||
| 90 : | %else | ||
| 91 : | sub esp, byte 12 | ||
| 92 : | %endif | ||
| 93 : | movq mm1, [eax+ebx] | ||
| 94 : | movq mm0, [eax+ebx+8] | ||
| 95 : | psadbw mm1, [edx+ebx] | ||
| 96 : | psadbw mm0, [edx+ebx+8] | ||
| 97 : | lea eax, [eax+4*ecx] | ||
| 98 : | lea edx, [edx+4*ecx] | ||
| 99 : | paddd mm7, mm6 | ||
| 100 : | paddd mm5, mm4 | ||
| 101 : | paddd mm3, mm2 | ||
| 102 : | paddd mm1, mm0 | ||
| 103 : | paddd mm5, mm7 | ||
| 104 : | paddd mm1, mm3 | ||
| 105 : | %endmacro | ||
| 106 : | edgomez | 851 | |
| 107 : | edgomez | 1382 | %macro SADBI_16x16_SSE0 0 |
| 108 : | movq mm2, [edx] | ||
| 109 : | movq mm3, [edx+8] | ||
| 110 : | |||
| 111 : | movq mm5, [byte eax] | ||
| 112 : | movq mm6, [eax+8] | ||
| 113 : | pavgb mm2, [byte ebx] | ||
| 114 : | pavgb mm3, [ebx+8] | ||
| 115 : | |||
| 116 : | add edx, ecx | ||
| 117 : | psadbw mm5, mm2 | ||
| 118 : | psadbw mm6, mm3 | ||
| 119 : | |||
| 120 : | add eax, ecx | ||
| 121 : | add ebx, ecx | ||
| 122 : | movq mm2, [byte edx] | ||
| 123 : | |||
| 124 : | movq mm3, [edx+8] | ||
| 125 : | movq mm0, [byte eax] | ||
| 126 : | |||
| 127 : | movq mm1, [eax+8] | ||
| 128 : | pavgb mm2, [byte ebx] | ||
| 129 : | |||
| 130 : | pavgb mm3, [ebx+8] | ||
| 131 : | add edx, ecx | ||
| 132 : | add eax, ecx | ||
| 133 : | |||
| 134 : | add ebx, ecx | ||
| 135 : | psadbw mm0, mm2 | ||
| 136 : | psadbw mm1, mm3 | ||
| 137 : | |||
| 138 : | %endmacro | ||
| 139 : | |||
| 140 : | %macro SADBI_16x16_SSE 0 | ||
| 141 : | movq mm2, [byte edx] | ||
| 142 : | movq mm3, [edx+8] | ||
| 143 : | paddusw mm5, mm0 | ||
| 144 : | paddusw mm6, mm1 | ||
| 145 : | movq mm0, [eax] | ||
| 146 : | movq mm1, [eax+8] | ||
| 147 : | pavgb mm2, [ebx] | ||
| 148 : | pavgb mm3, [ebx+8] | ||
| 149 : | add edx, ecx | ||
| 150 : | add eax, ecx | ||
| 151 : | add ebx, ecx | ||
| 152 : | psadbw mm0, mm2 | ||
| 153 : | psadbw mm1, mm3 | ||
| 154 : | %endmacro | ||
| 155 : | |||
| 156 : | %macro SADBI_8x8_3dne 0 | ||
| 157 : | movq mm2, [edx] | ||
| 158 : | movq mm3, [edx+ecx] | ||
| 159 : | pavgb mm2, [eax] | ||
| 160 : | pavgb mm3, [eax+ecx] | ||
| 161 : | lea edx, [edx+2*ecx] | ||
| 162 : | lea eax, [eax+2*ecx] | ||
| 163 : | paddusw mm5, mm0 | ||
| 164 : | paddusw mm6, mm1 | ||
| 165 : | movq mm0, [ebx] | ||
| 166 : | movq mm1, [ebx+ecx] | ||
| 167 : | lea ebx, [ebx+2*ecx] | ||
| 168 : | psadbw mm0, mm2 | ||
| 169 : | psadbw mm1, mm3 | ||
| 170 : | %endmacro | ||
| 171 : | |||
| 172 : | %macro ABS_16x16_SSE 1 | ||
| 173 : | %if (%1 == 0) | ||
| 174 : | movq mm7, [eax] | ||
| 175 : | psadbw mm7, mm4 | ||
| 176 : | mov esi, esi | ||
| 177 : | movq mm6, [eax+8] | ||
| 178 : | movq mm5, [eax+ecx] | ||
| 179 : | movq mm3, [eax+ecx+8] | ||
| 180 : | psadbw mm6, mm4 | ||
| 181 : | |||
| 182 : | movq mm2, [byte eax+2*ecx] | ||
| 183 : | psadbw mm5, mm4 | ||
| 184 : | movq mm1, [eax+2*ecx+8] | ||
| 185 : | psadbw mm3, mm4 | ||
| 186 : | |||
| 187 : | suxen_drol | 1787 | movq mm0, [eax+edx+0] |
| 188 : | edgomez | 1382 | psadbw mm2, mm4 |
| 189 : | add eax, edx | ||
| 190 : | psadbw mm1, mm4 | ||
| 191 : | %endif | ||
| 192 : | %if (%1 == 1) | ||
| 193 : | psadbw mm0, mm4 | ||
| 194 : | paddd mm7, mm0 | ||
| 195 : | movq mm0, [eax+8] | ||
| 196 : | psadbw mm0, mm4 | ||
| 197 : | paddd mm6, mm0 | ||
| 198 : | |||
| 199 : | movq mm0, [byte eax+ecx] | ||
| 200 : | psadbw mm0, mm4 | ||
| 201 : | |||
| 202 : | paddd mm5, mm0 | ||
| 203 : | movq mm0, [eax+ecx+8] | ||
| 204 : | |||
| 205 : | psadbw mm0, mm4 | ||
| 206 : | paddd mm3, mm0 | ||
| 207 : | movq mm0, [eax+2*ecx] | ||
| 208 : | psadbw mm0, mm4 | ||
| 209 : | paddd mm2, mm0 | ||
| 210 : | |||
| 211 : | movq mm0, [eax+2*ecx+8] | ||
| 212 : | add eax, edx | ||
| 213 : | psadbw mm0, mm4 | ||
| 214 : | paddd mm1, mm0 | ||
| 215 : | movq mm0, [eax] | ||
| 216 : | %endif | ||
| 217 : | %if (%1 == 2) | ||
| 218 : | psadbw mm0, mm4 | ||
| 219 : | paddd mm7, mm0 | ||
| 220 : | movq mm0, [eax+8] | ||
| 221 : | psadbw mm0, mm4 | ||
| 222 : | paddd mm6, mm0 | ||
| 223 : | %endif | ||
| 224 : | %endmacro | ||
| 225 : | |||
| 226 : | ;============================================================================= | ||
| 227 : | ; Code | ||
| 228 : | ;============================================================================= | ||
| 229 : | |||
| 230 : | SECTION .text | ||
| 231 : | |||
| 232 : | cglobal sad16_3dne | ||
| 233 : | cglobal sad8_3dne | ||
| 234 : | cglobal sad16bi_3dne | ||
| 235 : | cglobal sad8bi_3dne | ||
| 236 : | cglobal dev16_3dne | ||
| 237 : | |||
| 238 : | ;----------------------------------------------------------------------------- | ||
| 239 : | edgomez | 851 | ; |
| 240 : | ; uint32_t sad16_3dne(const uint8_t * const cur, | ||
| 241 : | edgomez | 1382 | ; const uint8_t * const ref, |
| 242 : | ; const uint32_t stride, | ||
| 243 : | ; const uint32_t best_sad); | ||
| 244 : | edgomez | 851 | ; |
| 245 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
| 246 : | |||
| 247 : | edgomez | 851 | ; optimization: 21% faster |
| 248 : | edgomez | 1382 | |
| 249 : | ALIGN 16 | ||
| 250 : | edgomez | 851 | sad16_3dne: |
| 251 : | edgomez | 1382 | mov eax, [esp+ 4] ; Src1 |
| 252 : | mov edx, [esp+ 8] ; Src2 | ||
| 253 : | mov ecx, [esp+12] ; Stride | ||
| 254 : | push ebx | ||
| 255 : | lea ebx, [2*ecx+ecx] | ||
| 256 : | edgomez | 851 | |
| 257 : | edgomez | 1382 | SAD_16x16_SSE 0 |
| 258 : | SAD_16x16_SSE 1 | ||
| 259 : | SAD_16x16_SSE 2 | ||
| 260 : | SAD_16x16_SSE 3 | ||
| 261 : | edgomez | 851 | |
| 262 : | edgomez | 1382 | mov ecx, [esp] |
| 263 : | add ecx, [esp+4] | ||
| 264 : | add ecx, [esp+8] | ||
| 265 : | paddd mm1, mm5 | ||
| 266 : | mov ebx, [esp+12] | ||
| 267 : | add esp, byte 4+12 | ||
| 268 : | movd eax, mm1 | ||
| 269 : | add eax, ecx | ||
| 270 : | edgomez | 851 | |
| 271 : | edgomez | 1382 | ret |
| 272 : | Isibaar | 1793 | ENDFUNC |
| 273 : | edgomez | 1382 | |
| 274 : | |||
| 275 : | ;----------------------------------------------------------------------------- | ||
| 276 : | edgomez | 851 | ; |
| 277 : | ; uint32_t sad8_3dne(const uint8_t * const cur, | ||
| 278 : | ; const uint8_t * const ref, | ||
| 279 : | ; const uint32_t stride); | ||
| 280 : | ; | ||
| 281 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
| 282 : | |||
| 283 : | ALIGN 16 | ||
| 284 : | edgomez | 851 | sad8_3dne: |
| 285 : | |||
| 286 : | edgomez | 1382 | mov eax, [esp+ 4] ; Src1 |
| 287 : | mov ecx, [esp+12] ; Stride | ||
| 288 : | mov edx, [esp+ 8] ; Src2 | ||
| 289 : | push ebx | ||
| 290 : | lea ebx, [ecx+2*ecx] | ||
| 291 : | edgomez | 851 | |
| 292 : | edgomez | 1382 | movq mm0, [byte eax] ;0 |
| 293 : | psadbw mm0, [byte edx] | ||
| 294 : | movq mm1, [eax+ecx] ;1 | ||
| 295 : | psadbw mm1, [edx+ecx] | ||
| 296 : | edgomez | 851 | |
| 297 : | edgomez | 1382 | movq mm2, [eax+2*ecx] ;2 |
| 298 : | psadbw mm2, [edx+2*ecx] | ||
| 299 : | movq mm3, [eax+ebx] ;3 | ||
| 300 : | psadbw mm3, [edx+ebx] | ||
| 301 : | edgomez | 851 | |
| 302 : | edgomez | 1382 | paddd mm0, mm1 |
| 303 : | edgomez | 851 | |
| 304 : | edgomez | 1382 | movq mm4, [byte eax+4*ecx];4 |
| 305 : | psadbw mm4, [edx+4*ecx] | ||
| 306 : | movq mm5, [eax+2*ebx] ;6 | ||
| 307 : | psadbw mm5, [edx+2*ebx] | ||
| 308 : | edgomez | 851 | |
| 309 : | edgomez | 1382 | paddd mm2, mm3 |
| 310 : | paddd mm0, mm2 | ||
| 311 : | edgomez | 851 | |
| 312 : | edgomez | 1382 | lea ebx, [ebx+4*ecx] ;3+4=7 |
| 313 : | lea ecx, [ecx+4*ecx] ; 5 | ||
| 314 : | movq mm6, [eax+ecx] ;5 | ||
| 315 : | psadbw mm6, [edx+ecx] | ||
| 316 : | movq mm7, [eax+ebx] ;7 | ||
| 317 : | psadbw mm7, [edx+ebx] | ||
| 318 : | paddd mm4, mm5 | ||
| 319 : | paddd mm6, mm7 | ||
| 320 : | paddd mm0, mm4 | ||
| 321 : | mov ebx, [esp] | ||
| 322 : | add esp, byte 4 | ||
| 323 : | paddd mm0, mm6 | ||
| 324 : | movd eax, mm0 | ||
| 325 : | edgomez | 851 | |
| 326 : | edgomez | 1382 | ret |
| 327 : | Isibaar | 1793 | ENDFUNC |
| 328 : | edgomez | 851 | |
| 329 : | edgomez | 1382 | |
| 330 : | ;----------------------------------------------------------------------------- | ||
| 331 : | edgomez | 851 | ; |
| 332 : | ; uint32_t sad16bi_3dne(const uint8_t * const cur, | ||
| 333 : | ; const uint8_t * const ref1, | ||
| 334 : | ; const uint8_t * const ref2, | ||
| 335 : | ; const uint32_t stride); | ||
| 336 : | ; | ||
| 337 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
| 338 : | edgomez | 851 | ;optimization: 14% faster |
| 339 : | |||
| 340 : | edgomez | 1382 | ALIGN 16 |
| 341 : | sad16bi_3dne: | ||
| 342 : | mov eax, [esp+ 4] ; Src | ||
| 343 : | mov edx, [esp+ 8] ; Ref1 | ||
| 344 : | mov ecx, [esp+16] ; Stride | ||
| 345 : | push ebx | ||
| 346 : | mov ebx, [esp+4+12] ; Ref2 | ||
| 347 : | edgomez | 851 | |
| 348 : | edgomez | 1382 | SADBI_16x16_SSE0 |
| 349 : | SADBI_16x16_SSE | ||
| 350 : | SADBI_16x16_SSE | ||
| 351 : | SADBI_16x16_SSE | ||
| 352 : | SADBI_16x16_SSE | ||
| 353 : | SADBI_16x16_SSE | ||
| 354 : | SADBI_16x16_SSE | ||
| 355 : | edgomez | 851 | |
| 356 : | edgomez | 1382 | SADBI_16x16_SSE |
| 357 : | SADBI_16x16_SSE | ||
| 358 : | SADBI_16x16_SSE | ||
| 359 : | SADBI_16x16_SSE | ||
| 360 : | SADBI_16x16_SSE | ||
| 361 : | SADBI_16x16_SSE | ||
| 362 : | SADBI_16x16_SSE | ||
| 363 : | SADBI_16x16_SSE | ||
| 364 : | paddusw mm5,mm0 | ||
| 365 : | paddusw mm6,mm1 | ||
| 366 : | edgomez | 851 | |
| 367 : | edgomez | 1382 | pop ebx |
| 368 : | paddusw mm6,mm5 | ||
| 369 : | movd eax, mm6 | ||
| 370 : | edgomez | 851 | |
| 371 : | edgomez | 1382 | ret |
| 372 : | Isibaar | 1793 | ENDFUNC |
| 373 : | edgomez | 851 | |
| 374 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
| 375 : | ; | ||
| 376 : | ; uint32_t sad8bi_3dne(const uint8_t * const cur, | ||
| 377 : | ; const uint8_t * const ref1, | ||
| 378 : | ; const uint8_t * const ref2, | ||
| 379 : | ; const uint32_t stride); | ||
| 380 : | ; | ||
| 381 : | ;----------------------------------------------------------------------------- | ||
| 382 : | edgomez | 851 | |
| 383 : | edgomez | 1382 | ALIGN 16 |
| 384 : | sad8bi_3dne: | ||
| 385 : | mov eax, [esp+12] ; Ref2 | ||
| 386 : | mov edx, [esp+ 8] ; Ref1 | ||
| 387 : | mov ecx, [esp+16] ; Stride | ||
| 388 : | push ebx | ||
| 389 : | mov ebx, [esp+4+ 4] ; Src | ||
| 390 : | edgomez | 851 | |
| 391 : | edgomez | 1382 | movq mm2, [edx] |
| 392 : | movq mm3, [edx+ecx] | ||
| 393 : | pavgb mm2, [eax] | ||
| 394 : | pavgb mm3, [eax+ecx] | ||
| 395 : | lea edx, [edx+2*ecx] | ||
| 396 : | lea eax, [eax+2*ecx] | ||
| 397 : | movq mm5, [ebx] | ||
| 398 : | movq mm6, [ebx+ecx] | ||
| 399 : | lea ebx, [ebx+2*ecx] | ||
| 400 : | psadbw mm5, mm2 | ||
| 401 : | psadbw mm6, mm3 | ||
| 402 : | edgomez | 851 | |
| 403 : | edgomez | 1382 | movq mm2, [edx] |
| 404 : | movq mm3, [edx+ecx] | ||
| 405 : | pavgb mm2, [eax] | ||
| 406 : | pavgb mm3, [eax+ecx] | ||
| 407 : | lea edx, [edx+2*ecx] | ||
| 408 : | lea eax, [eax+2*ecx] | ||
| 409 : | movq mm0, [ebx] | ||
| 410 : | movq mm1, [ebx+ecx] | ||
| 411 : | lea ebx, [ebx+2*ecx] | ||
| 412 : | psadbw mm0, mm2 | ||
| 413 : | psadbw mm1, mm3 | ||
| 414 : | edgomez | 851 | |
| 415 : | edgomez | 1382 | movq mm2, [edx] |
| 416 : | movq mm3, [edx+ecx] | ||
| 417 : | pavgb mm2, [eax] | ||
| 418 : | pavgb mm3, [eax+ecx] | ||
| 419 : | lea edx, [edx+2*ecx] | ||
| 420 : | lea eax, [eax+2*ecx] | ||
| 421 : | paddusw mm5,mm0 | ||
| 422 : | paddusw mm6,mm1 | ||
| 423 : | movq mm0, [ebx] | ||
| 424 : | movq mm1, [ebx+ecx] | ||
| 425 : | lea ebx, [ebx+2*ecx] | ||
| 426 : | psadbw mm0, mm2 | ||
| 427 : | psadbw mm1, mm3 | ||
| 428 : | edgomez | 851 | |
| 429 : | edgomez | 1382 | movq mm2, [edx] |
| 430 : | movq mm3, [edx+ecx] | ||
| 431 : | pavgb mm2, [eax] | ||
| 432 : | pavgb mm3, [eax+ecx] | ||
| 433 : | paddusw mm5,mm0 | ||
| 434 : | paddusw mm6,mm1 | ||
| 435 : | movq mm0, [ebx] | ||
| 436 : | movq mm1, [ebx+ecx] | ||
| 437 : | psadbw mm0, mm2 | ||
| 438 : | psadbw mm1, mm3 | ||
| 439 : | paddusw mm5,mm0 | ||
| 440 : | paddusw mm6,mm1 | ||
| 441 : | edgomez | 851 | |
| 442 : | edgomez | 1382 | paddusw mm6,mm5 |
| 443 : | mov ebx,[esp] | ||
| 444 : | add esp,byte 4 | ||
| 445 : | movd eax, mm6 | ||
| 446 : | edgomez | 851 | |
| 447 : | edgomez | 1382 | ret |
| 448 : | Isibaar | 1793 | ENDFUNC |
| 449 : | edgomez | 851 | |
| 450 : | |||
| 451 : | ;=========================================================================== | ||
| 452 : | ; | ||
| 453 : | ; uint32_t dev16_3dne(const uint8_t * const cur, | ||
| 454 : | ; const uint32_t stride); | ||
| 455 : | ; | ||
| 456 : | ;=========================================================================== | ||
| 457 : | ; optimization: 25 % faster | ||
| 458 : | |||
| 459 : | edgomez | 1382 | ALIGN 16 |
| 460 : | dev16_3dne: | ||
| 461 : | edgomez | 851 | |
| 462 : | edgomez | 1382 | mov eax, [esp+ 4] ; Src |
| 463 : | mov ecx, [esp+ 8] ; Stride | ||
| 464 : | lea edx, [ecx+2*ecx] | ||
| 465 : | edgomez | 851 | |
| 466 : | edgomez | 1382 | pxor mm4, mm4 |
| 467 : | edgomez | 851 | |
| 468 : | edgomez | 1382 | ALIGN 8 |
| 469 : | ABS_16x16_SSE 0 | ||
| 470 : | ABS_16x16_SSE 1 | ||
| 471 : | ABS_16x16_SSE 1 | ||
| 472 : | ABS_16x16_SSE 1 | ||
| 473 : | ABS_16x16_SSE 1 | ||
| 474 : | edgomez | 851 | |
| 475 : | edgomez | 1382 | paddd mm1, mm2 |
| 476 : | paddd mm3, mm5 | ||
| 477 : | edgomez | 851 | |
| 478 : | edgomez | 1382 | ABS_16x16_SSE 2 |
| 479 : | edgomez | 851 | |
| 480 : | edgomez | 1382 | paddd mm7, mm6 |
| 481 : | paddd mm1, mm3 | ||
| 482 : | mov eax, [esp+ 4] ; Src | ||
| 483 : | paddd mm7, mm1 | ||
| 484 : | punpcklbw mm7, mm7 ;xxyyaazz | ||
| 485 : | pshufw mm4, mm7, 055h ; mm4 contains the mean | ||
| 486 : | |||
| 487 : | |||
| 488 : | pxor mm1, mm1 | ||
| 489 : | |||
| 490 : | ABS_16x16_SSE 0 | ||
| 491 : | ABS_16x16_SSE 1 | ||
| 492 : | ABS_16x16_SSE 1 | ||
| 493 : | ABS_16x16_SSE 1 | ||
| 494 : | ABS_16x16_SSE 1 | ||
| 495 : | |||
| 496 : | paddd mm1, mm2 | ||
| 497 : | paddd mm3, mm5 | ||
| 498 : | |||
| 499 : | ABS_16x16_SSE 2 | ||
| 500 : | |||
| 501 : | paddd mm7, mm6 | ||
| 502 : | paddd mm1, mm3 | ||
| 503 : | paddd mm7, mm1 | ||
| 504 : | movd eax, mm7 | ||
| 505 : | |||
| 506 : | ret | ||
| 507 : | Isibaar | 1793 | ENDFUNC |
| 508 : | edgomez | 1540 | |
| 509 : | Isibaar | 1790 | |
| 510 : | %ifidn __OUTPUT_FORMAT__,elf | ||
| 511 : | section ".note.GNU-stack" noalloc noexec nowrite progbits | ||
| 512 : | %endif | ||
| 513 : |
| No admin address has been configured | ViewVC Help |
| Powered by ViewVC 1.0.4 |