Parent Directory
| 
Revision Log
Revision 1.12 - (view) (download)
| 1 : | edgomez | 1.3 | ;/**************************************************************************** |
| 2 : | edgomez | 1.2 | ; * |
| 3 : | edgomez | 1.3 | ; * XVID MPEG-4 VIDEO CODEC |
| 4 : | ; * - K7 optimized SAD operators - | ||
| 5 : | edgomez | 1.2 | ; * |
| 6 : | edgomez | 1.3 | ; * Copyright(C) 2002 Jaan Kalda |
| 7 : | edgomez | 1.2 | ; * |
| 8 : | edgomez | 1.3 | ; * 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 | 1.2 | ; * |
| 13 : | edgomez | 1.3 | ; * 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 | 1.2 | ; * |
| 18 : | edgomez | 1.3 | ; * 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 | 1.2 | ; * |
| 22 : | Isibaar | 1.12 | ; * $Id: sad_3dne.asm,v 1.11 2008/12/04 14:41:50 Isibaar Exp $ |
| 23 : | edgomez | 1.3 | ; * |
| 24 : | ; ***************************************************************************/ | ||
| 25 : | |||
| 26 : | ; these 3dne functions are compatible with iSSE, but are optimized specifically | ||
| 27 : | ; for K7 pipelines | ||
| 28 : | edgomez | 1.2 | |
| 29 : | Isibaar | 1.10 | %include "nasm.inc" |
| 30 : | edgomez | 1.2 | |
| 31 : | edgomez | 1.3 | ;============================================================================= |
| 32 : | ; Read only data | ||
| 33 : | ;============================================================================= | ||
| 34 : | |||
| 35 : | Isibaar | 1.10 | DATA |
| 36 : | edgomez | 1.3 | |
| 37 : | Isibaar | 1.10 | ALIGN SECTION_ALIGN |
| 38 : | edgomez | 1.3 | mmx_one: |
| 39 : | times 4 dw 1 | ||
| 40 : | |||
| 41 : | ;============================================================================= | ||
| 42 : | ; Helper macros | ||
| 43 : | ;============================================================================= | ||
| 44 : | |||
| 45 : | ;; %1 block number (0..4) | ||
| 46 : | %macro SAD_16x16_SSE 1 | ||
| 47 : | Isibaar | 1.10 | movq mm7, [_EAX] |
| 48 : | movq mm6, [_EAX+8] | ||
| 49 : | psadbw mm7, [TMP1] | ||
| 50 : | psadbw mm6, [TMP1+8] | ||
| 51 : | edgomez | 1.3 | %if (%1) |
| 52 : | paddd mm1, mm5 | ||
| 53 : | %endif | ||
| 54 : | Isibaar | 1.10 | movq mm5, [_EAX+TMP0] |
| 55 : | movq mm4, [_EAX+TMP0+8] | ||
| 56 : | psadbw mm5, [TMP1+TMP0] | ||
| 57 : | psadbw mm4, [TMP1+TMP0+8] | ||
| 58 : | movq mm3, [_EAX+2*TMP0] | ||
| 59 : | movq mm2, [_EAX+2*TMP0+8] | ||
| 60 : | psadbw mm3, [TMP1+2*TMP0] | ||
| 61 : | psadbw mm2, [TMP1+2*TMP0+8] | ||
| 62 : | edgomez | 1.3 | %if (%1) |
| 63 : | Isibaar | 1.10 | movd [_ESP+4*(%1-1)], mm1 |
| 64 : | edgomez | 1.2 | %else |
| 65 : | Isibaar | 1.10 | sub _ESP, byte 12 |
| 66 : | edgomez | 1.2 | %endif |
| 67 : | Isibaar | 1.10 | movq mm1, [_EAX+_EBX] |
| 68 : | movq mm0, [_EAX+_EBX+8] | ||
| 69 : | psadbw mm1, [TMP1+_EBX] | ||
| 70 : | psadbw mm0, [TMP1+_EBX+8] | ||
| 71 : | lea _EAX, [_EAX+4*TMP0] | ||
| 72 : | lea TMP1, [TMP1+4*TMP0] | ||
| 73 : | edgomez | 1.3 | paddd mm7, mm6 |
| 74 : | paddd mm5, mm4 | ||
| 75 : | paddd mm3, mm2 | ||
| 76 : | paddd mm1, mm0 | ||
| 77 : | paddd mm5, mm7 | ||
| 78 : | paddd mm1, mm3 | ||
| 79 : | %endmacro | ||
| 80 : | |||
| 81 : | %macro SADBI_16x16_SSE0 0 | ||
| 82 : | Isibaar | 1.10 | movq mm2, [TMP1] |
| 83 : | movq mm3, [TMP1+8] | ||
| 84 : | edgomez | 1.3 | |
| 85 : | Isibaar | 1.10 | movq mm5, [byte _EAX] |
| 86 : | movq mm6, [_EAX+8] | ||
| 87 : | pavgb mm2, [byte _EBX] | ||
| 88 : | pavgb mm3, [_EBX+8] | ||
| 89 : | edgomez | 1.3 | |
| 90 : | Isibaar | 1.10 | add TMP1, TMP0 |
| 91 : | edgomez | 1.3 | psadbw mm5, mm2 |
| 92 : | psadbw mm6, mm3 | ||
| 93 : | |||
| 94 : | Isibaar | 1.10 | add _EAX, TMP0 |
| 95 : | add _EBX, TMP0 | ||
| 96 : | movq mm2, [byte TMP1] | ||
| 97 : | edgomez | 1.3 | |
| 98 : | Isibaar | 1.10 | movq mm3, [TMP1+8] |
| 99 : | movq mm0, [byte _EAX] | ||
| 100 : | edgomez | 1.3 | |
| 101 : | Isibaar | 1.10 | movq mm1, [_EAX+8] |
| 102 : | pavgb mm2, [byte _EBX] | ||
| 103 : | edgomez | 1.3 | |
| 104 : | Isibaar | 1.10 | pavgb mm3, [_EBX+8] |
| 105 : | add TMP1, TMP0 | ||
| 106 : | add _EAX, TMP0 | ||
| 107 : | edgomez | 1.3 | |
| 108 : | Isibaar | 1.10 | add _EBX, TMP0 |
| 109 : | edgomez | 1.3 | psadbw mm0, mm2 |
| 110 : | psadbw mm1, mm3 | ||
| 111 : | |||
| 112 : | %endmacro | ||
| 113 : | |||
| 114 : | %macro SADBI_16x16_SSE 0 | ||
| 115 : | Isibaar | 1.10 | movq mm2, [byte TMP1] |
| 116 : | movq mm3, [TMP1+8] | ||
| 117 : | edgomez | 1.3 | paddusw mm5, mm0 |
| 118 : | paddusw mm6, mm1 | ||
| 119 : | Isibaar | 1.10 | movq mm0, [_EAX] |
| 120 : | movq mm1, [_EAX+8] | ||
| 121 : | pavgb mm2, [_EBX] | ||
| 122 : | pavgb mm3, [_EBX+8] | ||
| 123 : | add TMP1, TMP0 | ||
| 124 : | add _EAX, TMP0 | ||
| 125 : | add _EBX, TMP0 | ||
| 126 : | edgomez | 1.3 | psadbw mm0, mm2 |
| 127 : | psadbw mm1, mm3 | ||
| 128 : | %endmacro | ||
| 129 : | edgomez | 1.2 | |
| 130 : | edgomez | 1.3 | %macro SADBI_8x8_3dne 0 |
| 131 : | Isibaar | 1.10 | movq mm2, [TMP1] |
| 132 : | movq mm3, [TMP1+TMP0] | ||
| 133 : | pavgb mm2, [_EAX] | ||
| 134 : | pavgb mm3, [_EAX+TMP0] | ||
| 135 : | lea TMP1, [TMP1+2*TMP0] | ||
| 136 : | lea _EAX, [_EAX+2*TMP0] | ||
| 137 : | edgomez | 1.3 | paddusw mm5, mm0 |
| 138 : | paddusw mm6, mm1 | ||
| 139 : | Isibaar | 1.10 | movq mm0, [_EBX] |
| 140 : | movq mm1, [_EBX+TMP0] | ||
| 141 : | lea _EBX, [_EBX+2*TMP0] | ||
| 142 : | edgomez | 1.3 | psadbw mm0, mm2 |
| 143 : | psadbw mm1, mm3 | ||
| 144 : | %endmacro | ||
| 145 : | edgomez | 1.2 | |
| 146 : | edgomez | 1.3 | %macro ABS_16x16_SSE 1 |
| 147 : | %if (%1 == 0) | ||
| 148 : | Isibaar | 1.10 | movq mm7, [_EAX] |
| 149 : | edgomez | 1.3 | psadbw mm7, mm4 |
| 150 : | mov esi, esi | ||
| 151 : | Isibaar | 1.10 | movq mm6, [_EAX+8] |
| 152 : | movq mm5, [_EAX+TMP0] | ||
| 153 : | movq mm3, [_EAX+TMP0+8] | ||
| 154 : | edgomez | 1.3 | psadbw mm6, mm4 |
| 155 : | |||
| 156 : | Isibaar | 1.10 | movq mm2, [byte _EAX+2*TMP0] |
| 157 : | edgomez | 1.3 | psadbw mm5, mm4 |
| 158 : | Isibaar | 1.10 | movq mm1, [_EAX+2*TMP0+8] |
| 159 : | edgomez | 1.3 | psadbw mm3, mm4 |
| 160 : | |||
| 161 : | Isibaar | 1.10 | movq mm0, [_EAX+TMP1+0] |
| 162 : | edgomez | 1.3 | psadbw mm2, mm4 |
| 163 : | Isibaar | 1.10 | add _EAX, TMP1 |
| 164 : | edgomez | 1.3 | psadbw mm1, mm4 |
| 165 : | %endif | ||
| 166 : | %if (%1 == 1) | ||
| 167 : | psadbw mm0, mm4 | ||
| 168 : | paddd mm7, mm0 | ||
| 169 : | Isibaar | 1.10 | movq mm0, [_EAX+8] |
| 170 : | edgomez | 1.3 | psadbw mm0, mm4 |
| 171 : | paddd mm6, mm0 | ||
| 172 : | |||
| 173 : | Isibaar | 1.10 | movq mm0, [byte _EAX+TMP0] |
| 174 : | edgomez | 1.3 | psadbw mm0, mm4 |
| 175 : | |||
| 176 : | paddd mm5, mm0 | ||
| 177 : | Isibaar | 1.10 | movq mm0, [_EAX+TMP0+8] |
| 178 : | edgomez | 1.3 | |
| 179 : | psadbw mm0, mm4 | ||
| 180 : | paddd mm3, mm0 | ||
| 181 : | Isibaar | 1.10 | movq mm0, [_EAX+2*TMP0] |
| 182 : | edgomez | 1.3 | psadbw mm0, mm4 |
| 183 : | paddd mm2, mm0 | ||
| 184 : | |||
| 185 : | Isibaar | 1.10 | movq mm0, [_EAX+2*TMP0+8] |
| 186 : | add _EAX, TMP1 | ||
| 187 : | edgomez | 1.3 | psadbw mm0, mm4 |
| 188 : | paddd mm1, mm0 | ||
| 189 : | Isibaar | 1.10 | movq mm0, [_EAX] |
| 190 : | edgomez | 1.3 | %endif |
| 191 : | %if (%1 == 2) | ||
| 192 : | psadbw mm0, mm4 | ||
| 193 : | paddd mm7, mm0 | ||
| 194 : | Isibaar | 1.10 | movq mm0, [_EAX+8] |
| 195 : | edgomez | 1.3 | psadbw mm0, mm4 |
| 196 : | paddd mm6, mm0 | ||
| 197 : | %endif | ||
| 198 : | %endmacro | ||
| 199 : | edgomez | 1.2 | |
| 200 : | edgomez | 1.3 | ;============================================================================= |
| 201 : | ; Code | ||
| 202 : | ;============================================================================= | ||
| 203 : | |||
| 204 : | Isibaar | 1.11 | TEXT |
| 205 : | edgomez | 1.3 | |
| 206 : | cglobal sad16_3dne | ||
| 207 : | cglobal sad8_3dne | ||
| 208 : | cglobal sad16bi_3dne | ||
| 209 : | cglobal sad8bi_3dne | ||
| 210 : | cglobal dev16_3dne | ||
| 211 : | edgomez | 1.2 | |
| 212 : | edgomez | 1.3 | ;----------------------------------------------------------------------------- |
| 213 : | edgomez | 1.2 | ; |
| 214 : | ; uint32_t sad16_3dne(const uint8_t * const cur, | ||
| 215 : | edgomez | 1.3 | ; const uint8_t * const ref, |
| 216 : | ; const uint32_t stride, | ||
| 217 : | ; const uint32_t best_sad); | ||
| 218 : | edgomez | 1.2 | ; |
| 219 : | edgomez | 1.3 | ;----------------------------------------------------------------------------- |
| 220 : | |||
| 221 : | edgomez | 1.2 | ; optimization: 21% faster |
| 222 : | edgomez | 1.3 | |
| 223 : | Isibaar | 1.10 | ALIGN SECTION_ALIGN |
| 224 : | edgomez | 1.2 | sad16_3dne: |
| 225 : | Isibaar | 1.10 | mov _EAX, prm1 ; Src1 |
| 226 : | mov TMP1, prm2 ; Src2 | ||
| 227 : | mov TMP0, prm3 ; Stride | ||
| 228 : | |||
| 229 : | push _EBX | ||
| 230 : | lea _EBX, [2*TMP0+TMP0] | ||
| 231 : | edgomez | 1.3 | |
| 232 : | SAD_16x16_SSE 0 | ||
| 233 : | SAD_16x16_SSE 1 | ||
| 234 : | SAD_16x16_SSE 2 | ||
| 235 : | SAD_16x16_SSE 3 | ||
| 236 : | |||
| 237 : | paddd mm1, mm5 | ||
| 238 : | movd eax, mm1 | ||
| 239 : | Isibaar | 1.10 | add eax, dword [_ESP] |
| 240 : | add eax, dword [_ESP+4] | ||
| 241 : | add eax, dword [_ESP+8] | ||
| 242 : | mov _EBX, [_ESP+12] | ||
| 243 : | add _ESP, byte PTR_SIZE+12 | ||
| 244 : | edgomez | 1.2 | |
| 245 : | edgomez | 1.3 | ret |
| 246 : | Isibaar | 1.9 | ENDFUNC |
| 247 : | edgomez | 1.2 | |
| 248 : | |||
| 249 : | edgomez | 1.3 | ;----------------------------------------------------------------------------- |
| 250 : | edgomez | 1.2 | ; |
| 251 : | ; uint32_t sad8_3dne(const uint8_t * const cur, | ||
| 252 : | ; const uint8_t * const ref, | ||
| 253 : | ; const uint32_t stride); | ||
| 254 : | ; | ||
| 255 : | edgomez | 1.3 | ;----------------------------------------------------------------------------- |
| 256 : | |||
| 257 : | Isibaar | 1.10 | ALIGN SECTION_ALIGN |
| 258 : | edgomez | 1.2 | sad8_3dne: |
| 259 : | |||
| 260 : | Isibaar | 1.10 | mov _EAX, prm1 ; Src1 |
| 261 : | mov TMP0, prm3 ; Stride | ||
| 262 : | mov TMP1, prm2 ; Src2 | ||
| 263 : | push _EBX | ||
| 264 : | lea _EBX, [TMP0+2*TMP0] | ||
| 265 : | |||
| 266 : | movq mm0, [byte _EAX] ;0 | ||
| 267 : | psadbw mm0, [byte TMP1] | ||
| 268 : | movq mm1, [_EAX+TMP0] ;1 | ||
| 269 : | psadbw mm1, [TMP1+TMP0] | ||
| 270 : | |||
| 271 : | movq mm2, [_EAX+2*TMP0] ;2 | ||
| 272 : | psadbw mm2, [TMP1+2*TMP0] | ||
| 273 : | movq mm3, [_EAX+_EBX] ;3 | ||
| 274 : | psadbw mm3, [TMP1+_EBX] | ||
| 275 : | edgomez | 1.3 | |
| 276 : | paddd mm0, mm1 | ||
| 277 : | |||
| 278 : | Isibaar | 1.10 | movq mm4, [byte _EAX+4*TMP0];4 |
| 279 : | psadbw mm4, [TMP1+4*TMP0] | ||
| 280 : | movq mm5, [_EAX+2*_EBX] ;6 | ||
| 281 : | psadbw mm5, [TMP1+2*_EBX] | ||
| 282 : | edgomez | 1.3 | |
| 283 : | paddd mm2, mm3 | ||
| 284 : | paddd mm0, mm2 | ||
| 285 : | |||
| 286 : | Isibaar | 1.10 | lea _EBX, [_EBX+4*TMP0] ;3+4=7 |
| 287 : | lea TMP0, [TMP0+4*TMP0] ; 5 | ||
| 288 : | movq mm6, [_EAX+TMP0] ;5 | ||
| 289 : | psadbw mm6, [TMP1+TMP0] | ||
| 290 : | movq mm7, [_EAX+_EBX] ;7 | ||
| 291 : | psadbw mm7, [TMP1+_EBX] | ||
| 292 : | edgomez | 1.3 | paddd mm4, mm5 |
| 293 : | paddd mm6, mm7 | ||
| 294 : | paddd mm0, mm4 | ||
| 295 : | Isibaar | 1.10 | mov _EBX, [_ESP] |
| 296 : | add _ESP, byte PTR_SIZE | ||
| 297 : | edgomez | 1.3 | paddd mm0, mm6 |
| 298 : | movd eax, mm0 | ||
| 299 : | edgomez | 1.2 | |
| 300 : | edgomez | 1.3 | ret |
| 301 : | Isibaar | 1.9 | ENDFUNC |
| 302 : | edgomez | 1.2 | |
| 303 : | |||
| 304 : | edgomez | 1.3 | ;----------------------------------------------------------------------------- |
| 305 : | edgomez | 1.2 | ; |
| 306 : | ; uint32_t sad16bi_3dne(const uint8_t * const cur, | ||
| 307 : | ; const uint8_t * const ref1, | ||
| 308 : | ; const uint8_t * const ref2, | ||
| 309 : | ; const uint32_t stride); | ||
| 310 : | ; | ||
| 311 : | edgomez | 1.3 | ;----------------------------------------------------------------------------- |
| 312 : | edgomez | 1.2 | ;optimization: 14% faster |
| 313 : | |||
| 314 : | Isibaar | 1.10 | ALIGN SECTION_ALIGN |
| 315 : | edgomez | 1.3 | sad16bi_3dne: |
| 316 : | Isibaar | 1.10 | mov _EAX, prm1 ; Src |
| 317 : | mov TMP1, prm2 ; Ref1 | ||
| 318 : | mov TMP0, prm4 ; Stride | ||
| 319 : | |||
| 320 : | push _EBX | ||
| 321 : | %ifdef ARCH_IS_X86_64 | ||
| 322 : | mov _EBX, prm3 | ||
| 323 : | %else | ||
| 324 : | mov _EBX, [_ESP+4+12] ; Ref2 | ||
| 325 : | %endif | ||
| 326 : | edgomez | 1.3 | |
| 327 : | SADBI_16x16_SSE0 | ||
| 328 : | SADBI_16x16_SSE | ||
| 329 : | SADBI_16x16_SSE | ||
| 330 : | SADBI_16x16_SSE | ||
| 331 : | SADBI_16x16_SSE | ||
| 332 : | SADBI_16x16_SSE | ||
| 333 : | SADBI_16x16_SSE | ||
| 334 : | |||
| 335 : | SADBI_16x16_SSE | ||
| 336 : | SADBI_16x16_SSE | ||
| 337 : | SADBI_16x16_SSE | ||
| 338 : | SADBI_16x16_SSE | ||
| 339 : | SADBI_16x16_SSE | ||
| 340 : | SADBI_16x16_SSE | ||
| 341 : | SADBI_16x16_SSE | ||
| 342 : | SADBI_16x16_SSE | ||
| 343 : | paddusw mm5,mm0 | ||
| 344 : | paddusw mm6,mm1 | ||
| 345 : | |||
| 346 : | Isibaar | 1.10 | pop _EBX |
| 347 : | edgomez | 1.3 | paddusw mm6,mm5 |
| 348 : | movd eax, mm6 | ||
| 349 : | |||
| 350 : | ret | ||
| 351 : | Isibaar | 1.9 | ENDFUNC |
| 352 : | edgomez | 1.3 | |
| 353 : | ;----------------------------------------------------------------------------- | ||
| 354 : | ; | ||
| 355 : | ; uint32_t sad8bi_3dne(const uint8_t * const cur, | ||
| 356 : | ; const uint8_t * const ref1, | ||
| 357 : | ; const uint8_t * const ref2, | ||
| 358 : | ; const uint32_t stride); | ||
| 359 : | ; | ||
| 360 : | ;----------------------------------------------------------------------------- | ||
| 361 : | edgomez | 1.2 | |
| 362 : | Isibaar | 1.10 | ALIGN SECTION_ALIGN |
| 363 : | edgomez | 1.3 | sad8bi_3dne: |
| 364 : | Isibaar | 1.10 | mov _EAX, prm3 ; Ref2 |
| 365 : | mov TMP1, prm2 ; Ref1 | ||
| 366 : | mov TMP0, prm4 ; Stride | ||
| 367 : | |||
| 368 : | push _EBX | ||
| 369 : | %ifdef ARCH_IS_X86_64 | ||
| 370 : | mov _EBX, prm1 | ||
| 371 : | %else | ||
| 372 : | mov _EBX, [_ESP+4+ 4] ; Src | ||
| 373 : | %endif | ||
| 374 : | |||
| 375 : | movq mm2, [TMP1] | ||
| 376 : | movq mm3, [TMP1+TMP0] | ||
| 377 : | pavgb mm2, [_EAX] | ||
| 378 : | pavgb mm3, [_EAX+TMP0] | ||
| 379 : | lea TMP1, [TMP1+2*TMP0] | ||
| 380 : | lea _EAX, [_EAX+2*TMP0] | ||
| 381 : | movq mm5, [_EBX] | ||
| 382 : | movq mm6, [_EBX+TMP0] | ||
| 383 : | lea _EBX, [_EBX+2*TMP0] | ||
| 384 : | edgomez | 1.3 | psadbw mm5, mm2 |
| 385 : | psadbw mm6, mm3 | ||
| 386 : | |||
| 387 : | Isibaar | 1.10 | movq mm2, [TMP1] |
| 388 : | movq mm3, [TMP1+TMP0] | ||
| 389 : | pavgb mm2, [_EAX] | ||
| 390 : | pavgb mm3, [_EAX+TMP0] | ||
| 391 : | lea TMP1, [TMP1+2*TMP0] | ||
| 392 : | lea _EAX, [_EAX+2*TMP0] | ||
| 393 : | movq mm0, [_EBX] | ||
| 394 : | movq mm1, [_EBX+TMP0] | ||
| 395 : | lea _EBX, [_EBX+2*TMP0] | ||
| 396 : | edgomez | 1.3 | psadbw mm0, mm2 |
| 397 : | psadbw mm1, mm3 | ||
| 398 : | |||
| 399 : | Isibaar | 1.10 | movq mm2, [TMP1] |
| 400 : | movq mm3, [TMP1+TMP0] | ||
| 401 : | pavgb mm2, [_EAX] | ||
| 402 : | pavgb mm3, [_EAX+TMP0] | ||
| 403 : | lea TMP1, [TMP1+2*TMP0] | ||
| 404 : | lea _EAX, [_EAX+2*TMP0] | ||
| 405 : | edgomez | 1.3 | paddusw mm5,mm0 |
| 406 : | paddusw mm6,mm1 | ||
| 407 : | Isibaar | 1.10 | movq mm0, [_EBX] |
| 408 : | movq mm1, [_EBX+TMP0] | ||
| 409 : | lea _EBX, [_EBX+2*TMP0] | ||
| 410 : | edgomez | 1.3 | psadbw mm0, mm2 |
| 411 : | psadbw mm1, mm3 | ||
| 412 : | |||
| 413 : | Isibaar | 1.10 | movq mm2, [TMP1] |
| 414 : | movq mm3, [TMP1+TMP0] | ||
| 415 : | pavgb mm2, [_EAX] | ||
| 416 : | pavgb mm3, [_EAX+TMP0] | ||
| 417 : | edgomez | 1.3 | paddusw mm5,mm0 |
| 418 : | paddusw mm6,mm1 | ||
| 419 : | Isibaar | 1.10 | movq mm0, [_EBX] |
| 420 : | movq mm1, [_EBX+TMP0] | ||
| 421 : | edgomez | 1.3 | psadbw mm0, mm2 |
| 422 : | psadbw mm1, mm3 | ||
| 423 : | paddusw mm5,mm0 | ||
| 424 : | paddusw mm6,mm1 | ||
| 425 : | |||
| 426 : | paddusw mm6,mm5 | ||
| 427 : | Isibaar | 1.10 | mov _EBX,[_ESP] |
| 428 : | add _ESP,byte PTR_SIZE | ||
| 429 : | edgomez | 1.3 | movd eax, mm6 |
| 430 : | edgomez | 1.2 | |
| 431 : | edgomez | 1.3 | ret |
| 432 : | Isibaar | 1.9 | ENDFUNC |
| 433 : | edgomez | 1.2 | |
| 434 : | |||
| 435 : | ;=========================================================================== | ||
| 436 : | ; | ||
| 437 : | ; uint32_t dev16_3dne(const uint8_t * const cur, | ||
| 438 : | ; const uint32_t stride); | ||
| 439 : | ; | ||
| 440 : | ;=========================================================================== | ||
| 441 : | ; optimization: 25 % faster | ||
| 442 : | |||
| 443 : | Isibaar | 1.10 | ALIGN SECTION_ALIGN |
| 444 : | edgomez | 1.2 | dev16_3dne: |
| 445 : | |||
| 446 : | Isibaar | 1.10 | mov _EAX, prm1 ; Src |
| 447 : | mov TMP0, prm2 ; Stride | ||
| 448 : | lea TMP1, [TMP0+2*TMP0] | ||
| 449 : | edgomez | 1.3 | |
| 450 : | pxor mm4, mm4 | ||
| 451 : | |||
| 452 : | Isibaar | 1.10 | ALIGN SECTION_ALIGN |
| 453 : | edgomez | 1.3 | ABS_16x16_SSE 0 |
| 454 : | ABS_16x16_SSE 1 | ||
| 455 : | ABS_16x16_SSE 1 | ||
| 456 : | ABS_16x16_SSE 1 | ||
| 457 : | ABS_16x16_SSE 1 | ||
| 458 : | |||
| 459 : | paddd mm1, mm2 | ||
| 460 : | paddd mm3, mm5 | ||
| 461 : | |||
| 462 : | ABS_16x16_SSE 2 | ||
| 463 : | |||
| 464 : | paddd mm7, mm6 | ||
| 465 : | paddd mm1, mm3 | ||
| 466 : | Isibaar | 1.10 | mov _EAX, prm1 ; Src |
| 467 : | edgomez | 1.3 | paddd mm7, mm1 |
| 468 : | punpcklbw mm7, mm7 ;xxyyaazz | ||
| 469 : | pshufw mm4, mm7, 055h ; mm4 contains the mean | ||
| 470 : | |||
| 471 : | |||
| 472 : | pxor mm1, mm1 | ||
| 473 : | |||
| 474 : | ABS_16x16_SSE 0 | ||
| 475 : | ABS_16x16_SSE 1 | ||
| 476 : | ABS_16x16_SSE 1 | ||
| 477 : | ABS_16x16_SSE 1 | ||
| 478 : | ABS_16x16_SSE 1 | ||
| 479 : | |||
| 480 : | paddd mm1, mm2 | ||
| 481 : | paddd mm3, mm5 | ||
| 482 : | |||
| 483 : | ABS_16x16_SSE 2 | ||
| 484 : | |||
| 485 : | paddd mm7, mm6 | ||
| 486 : | paddd mm1, mm3 | ||
| 487 : | paddd mm7, mm1 | ||
| 488 : | movd eax, mm7 | ||
| 489 : | |||
| 490 : | ret | ||
| 491 : | Isibaar | 1.9 | ENDFUNC |
| 492 : | edgomez | 1.6 | |
| 493 : | Isibaar | 1.12 | NON_EXEC_STACK |
| No admin address has been configured | ViewVC Help |
| Powered by ViewVC 1.0.4 |