Parent Directory
| 
Revision Log
Revision 1519 - (view) (download)
| 1 : | edgomez | 1382 | ;/**************************************************************************** |
| 2 : | Isibaar | 262 | ; * |
| 3 : | edgomez | 1382 | ; * XVID MPEG-4 VIDEO CODEC |
| 4 : | ; * - K7 optimized SAD operators - | ||
| 5 : | Isibaar | 262 | ; * |
| 6 : | edgomez | 1382 | ; * Copyright(C) 2001 Peter Ross <pross@xvid.org> |
| 7 : | ; * 2001 Michael Militzer <isibaar@xvid.org> | ||
| 8 : | ; * 2002 Pascal Massimino <skal@planet-d.net> | ||
| 9 : | Isibaar | 262 | ; * |
| 10 : | edgomez | 1382 | ; * This program is free software; you can redistribute it and/or modify it |
| 11 : | ; * under the terms of the GNU General Public License as published by | ||
| 12 : | ; * the Free Software Foundation; either version 2 of the License, or | ||
| 13 : | ; * (at your option) any later version. | ||
| 14 : | Isibaar | 262 | ; * |
| 15 : | edgomez | 1382 | ; * This program is distributed in the hope that it will be useful, |
| 16 : | ; * but WITHOUT ANY WARRANTY; without even the implied warranty of | ||
| 17 : | ; * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | ||
| 18 : | ; * GNU General Public License for more details. | ||
| 19 : | Isibaar | 262 | ; * |
| 20 : | edgomez | 1382 | ; * You should have received a copy of the GNU General Public License |
| 21 : | ; * along with this program; if not, write to the Free Software | ||
| 22 : | ; * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA | ||
| 23 : | Isibaar | 262 | ; * |
| 24 : | edgomez | 1519 | ; * $Id: sad_xmm.asm,v 1.8 2004-07-24 11:46:08 edgomez Exp $ |
| 25 : | Isibaar | 262 | ; * |
| 26 : | edgomez | 1382 | ; ***************************************************************************/ |
| 27 : | Isibaar | 262 | |
| 28 : | edgomez | 1382 | BITS 32 |
| 29 : | suxen_drol | 329 | |
| 30 : | edgomez | 1382 | %macro cglobal 1 |
| 31 : | Isibaar | 262 | %ifdef PREFIX |
| 32 : | edgomez | 1382 | global _%1 |
| 33 : | Isibaar | 262 | %define %1 _%1 |
| 34 : | %else | ||
| 35 : | global %1 | ||
| 36 : | %endif | ||
| 37 : | %endmacro | ||
| 38 : | |||
| 39 : | edgomez | 1382 | ;============================================================================= |
| 40 : | ; Read only data | ||
| 41 : | ;============================================================================= | ||
| 42 : | Isibaar | 262 | |
| 43 : | edgomez | 1382 | %ifdef FORMAT_COFF |
| 44 : | edgomez | 1519 | SECTION .rodata |
| 45 : | edgomez | 1382 | %else |
| 46 : | edgomez | 1519 | SECTION .rodata align=16 |
| 47 : | edgomez | 1382 | %endif |
| 48 : | Isibaar | 262 | |
| 49 : | edgomez | 1382 | ALIGN 16 |
| 50 : | mmx_one: times 4 dw 1 | ||
| 51 : | Isibaar | 262 | |
| 52 : | edgomez | 1382 | ;============================================================================= |
| 53 : | ; Helper macros | ||
| 54 : | ;============================================================================= | ||
| 55 : | Isibaar | 262 | |
| 56 : | edgomez | 1382 | %macro SAD_16x16_SSE 0 |
| 57 : | movq mm0, [eax] | ||
| 58 : | psadbw mm0, [edx] | ||
| 59 : | movq mm1, [eax+8] | ||
| 60 : | add eax, ecx | ||
| 61 : | psadbw mm1, [edx+8] | ||
| 62 : | paddusw mm5, mm0 | ||
| 63 : | add edx, ecx | ||
| 64 : | paddusw mm6, mm1 | ||
| 65 : | %endmacro | ||
| 66 : | |||
| 67 : | %macro SAD_8x8_SSE 0 | ||
| 68 : | movq mm0, [eax] | ||
| 69 : | movq mm1, [eax+ecx] | ||
| 70 : | psadbw mm0, [edx] | ||
| 71 : | psadbw mm1, [edx+ecx] | ||
| 72 : | add eax, ebx | ||
| 73 : | add edx, ebx | ||
| 74 : | paddusw mm5, mm0 | ||
| 75 : | paddusw mm6, mm1 | ||
| 76 : | %endmacro | ||
| 77 : | |||
| 78 : | %macro SADBI_16x16_SSE 0 | ||
| 79 : | movq mm0, [eax] | ||
| 80 : | movq mm1, [eax+8] | ||
| 81 : | movq mm2, [edx] | ||
| 82 : | movq mm3, [edx+8] | ||
| 83 : | pavgb mm2, [ebx] | ||
| 84 : | add edx, ecx | ||
| 85 : | pavgb mm3, [ebx+8] | ||
| 86 : | add ebx, ecx | ||
| 87 : | psadbw mm0, mm2 | ||
| 88 : | add eax, ecx | ||
| 89 : | psadbw mm1, mm3 | ||
| 90 : | paddusw mm5, mm0 | ||
| 91 : | paddusw mm6, mm1 | ||
| 92 : | %endmacro | ||
| 93 : | |||
| 94 : | %macro SADBI_8x8_XMM 0 | ||
| 95 : | movq mm0, [eax] | ||
| 96 : | movq mm1, [eax+ecx] | ||
| 97 : | movq mm2, [edx] | ||
| 98 : | movq mm3, [edx+ecx] | ||
| 99 : | pavgb mm2, [ebx] | ||
| 100 : | lea edx, [edx+2*ecx] | ||
| 101 : | pavgb mm3, [ebx+ecx] | ||
| 102 : | lea ebx, [ebx+2*ecx] | ||
| 103 : | psadbw mm0, mm2 | ||
| 104 : | lea eax, [eax+2*ecx] | ||
| 105 : | psadbw mm1, mm3 | ||
| 106 : | paddusw mm5, mm0 | ||
| 107 : | paddusw mm6, mm1 | ||
| 108 : | %endmacro | ||
| 109 : | |||
| 110 : | %macro MEAN_16x16_SSE 0 | ||
| 111 : | movq mm0, [eax] | ||
| 112 : | movq mm1, [eax+8] | ||
| 113 : | psadbw mm0, mm7 | ||
| 114 : | psadbw mm1, mm7 | ||
| 115 : | add eax, ecx | ||
| 116 : | paddw mm5, mm0 | ||
| 117 : | paddw mm6, mm1 | ||
| 118 : | %endmacro | ||
| 119 : | |||
| 120 : | %macro ABS_16x16_SSE 0 | ||
| 121 : | movq mm0, [eax] | ||
| 122 : | movq mm1, [eax+8] | ||
| 123 : | psadbw mm0, mm4 | ||
| 124 : | psadbw mm1, mm4 | ||
| 125 : | lea eax, [eax+ecx] | ||
| 126 : | paddw mm5, mm0 | ||
| 127 : | paddw mm6, mm1 | ||
| 128 : | %endmacro | ||
| 129 : | |||
| 130 : | ;============================================================================= | ||
| 131 : | ; Code | ||
| 132 : | ;============================================================================= | ||
| 133 : | |||
| 134 : | SECTION .text | ||
| 135 : | |||
| 136 : | cglobal sad16_xmm | ||
| 137 : | cglobal sad8_xmm | ||
| 138 : | cglobal sad16bi_xmm | ||
| 139 : | cglobal sad8bi_xmm | ||
| 140 : | cglobal dev16_xmm | ||
| 141 : | cglobal sad16v_xmm | ||
| 142 : | |||
| 143 : | ;----------------------------------------------------------------------------- | ||
| 144 : | Isibaar | 262 | ; |
| 145 : | ; uint32_t sad16_xmm(const uint8_t * const cur, | ||
| 146 : | ; const uint8_t * const ref, | ||
| 147 : | suxen_drol | 329 | ; const uint32_t stride, |
| 148 : | Isibaar | 262 | ; const uint32_t best_sad); |
| 149 : | suxen_drol | 329 | ; |
| 150 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
| 151 : | Isibaar | 262 | |
| 152 : | edgomez | 1382 | ALIGN 16 |
| 153 : | Isibaar | 262 | sad16_xmm: |
| 154 : | |||
| 155 : | edgomez | 1382 | mov eax, [esp+ 4] ; Src1 |
| 156 : | mov edx, [esp+ 8] ; Src2 | ||
| 157 : | mov ecx, [esp+12] ; Stride | ||
| 158 : | Isibaar | 262 | |
| 159 : | edgomez | 1382 | pxor mm5, mm5 ; accum1 |
| 160 : | pxor mm6, mm6 ; accum2 | ||
| 161 : | Isibaar | 262 | |
| 162 : | edgomez | 1382 | SAD_16x16_SSE |
| 163 : | SAD_16x16_SSE | ||
| 164 : | SAD_16x16_SSE | ||
| 165 : | SAD_16x16_SSE | ||
| 166 : | SAD_16x16_SSE | ||
| 167 : | SAD_16x16_SSE | ||
| 168 : | SAD_16x16_SSE | ||
| 169 : | SAD_16x16_SSE | ||
| 170 : | Isibaar | 262 | |
| 171 : | edgomez | 1382 | SAD_16x16_SSE |
| 172 : | SAD_16x16_SSE | ||
| 173 : | SAD_16x16_SSE | ||
| 174 : | SAD_16x16_SSE | ||
| 175 : | SAD_16x16_SSE | ||
| 176 : | SAD_16x16_SSE | ||
| 177 : | SAD_16x16_SSE | ||
| 178 : | SAD_16x16_SSE | ||
| 179 : | Isibaar | 262 | |
| 180 : | edgomez | 1382 | paddusw mm6,mm5 |
| 181 : | movd eax, mm6 | ||
| 182 : | ret | ||
| 183 : | Isibaar | 262 | |
| 184 : | |||
| 185 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
| 186 : | Isibaar | 262 | ; |
| 187 : | ; uint32_t sad8_xmm(const uint8_t * const cur, | ||
| 188 : | ; const uint8_t * const ref, | ||
| 189 : | ; const uint32_t stride); | ||
| 190 : | ; | ||
| 191 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
| 192 : | Isibaar | 262 | |
| 193 : | edgomez | 1382 | ALIGN 16 |
| 194 : | suxen_drol | 329 | sad8_xmm: |
| 195 : | |||
| 196 : | edgomez | 1382 | mov eax, [esp+ 4] ; Src1 |
| 197 : | mov edx, [esp+ 8] ; Src2 | ||
| 198 : | mov ecx, [esp+12] ; Stride | ||
| 199 : | push ebx | ||
| 200 : | lea ebx, [ecx+ecx] | ||
| 201 : | suxen_drol | 329 | |
| 202 : | edgomez | 1382 | pxor mm5, mm5 ; accum1 |
| 203 : | pxor mm6, mm6 ; accum2 | ||
| 204 : | suxen_drol | 329 | |
| 205 : | edgomez | 1382 | SAD_8x8_SSE |
| 206 : | SAD_8x8_SSE | ||
| 207 : | SAD_8x8_SSE | ||
| 208 : | suxen_drol | 329 | |
| 209 : | edgomez | 1382 | movq mm0, [eax] |
| 210 : | movq mm1, [eax+ecx] | ||
| 211 : | psadbw mm0, [edx] | ||
| 212 : | psadbw mm1, [edx+ecx] | ||
| 213 : | suxen_drol | 329 | |
| 214 : | edgomez | 1382 | pop ebx |
| 215 : | suxen_drol | 329 | |
| 216 : | edgomez | 1382 | paddusw mm5,mm0 |
| 217 : | paddusw mm6,mm1 | ||
| 218 : | suxen_drol | 329 | |
| 219 : | edgomez | 1382 | paddusw mm6,mm5 |
| 220 : | movd eax, mm6 | ||
| 221 : | suxen_drol | 329 | |
| 222 : | edgomez | 1382 | ret |
| 223 : | suxen_drol | 329 | |
| 224 : | edgomez | 1382 | |
| 225 : | ;----------------------------------------------------------------------------- | ||
| 226 : | suxen_drol | 329 | ; |
| 227 : | ; uint32_t sad16bi_xmm(const uint8_t * const cur, | ||
| 228 : | ; const uint8_t * const ref1, | ||
| 229 : | ; const uint8_t * const ref2, | ||
| 230 : | ; const uint32_t stride); | ||
| 231 : | ; | ||
| 232 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
| 233 : | suxen_drol | 329 | |
| 234 : | edgomez | 1382 | ALIGN 16 |
| 235 : | Isibaar | 262 | sad16bi_xmm: |
| 236 : | edgomez | 1382 | push ebx |
| 237 : | mov eax, [esp+4+ 4] ; Src | ||
| 238 : | mov edx, [esp+4+ 8] ; Ref1 | ||
| 239 : | mov ebx, [esp+4+12] ; Ref2 | ||
| 240 : | mov ecx, [esp+4+16] ; Stride | ||
| 241 : | Isibaar | 262 | |
| 242 : | edgomez | 1382 | pxor mm5, mm5 ; accum1 |
| 243 : | pxor mm6, mm6 ; accum2 | ||
| 244 : | Isibaar | 262 | |
| 245 : | edgomez | 1382 | SADBI_16x16_SSE |
| 246 : | SADBI_16x16_SSE | ||
| 247 : | SADBI_16x16_SSE | ||
| 248 : | SADBI_16x16_SSE | ||
| 249 : | SADBI_16x16_SSE | ||
| 250 : | SADBI_16x16_SSE | ||
| 251 : | SADBI_16x16_SSE | ||
| 252 : | SADBI_16x16_SSE | ||
| 253 : | Isibaar | 262 | |
| 254 : | edgomez | 1382 | SADBI_16x16_SSE |
| 255 : | SADBI_16x16_SSE | ||
| 256 : | SADBI_16x16_SSE | ||
| 257 : | SADBI_16x16_SSE | ||
| 258 : | SADBI_16x16_SSE | ||
| 259 : | SADBI_16x16_SSE | ||
| 260 : | SADBI_16x16_SSE | ||
| 261 : | SADBI_16x16_SSE | ||
| 262 : | Isibaar | 262 | |
| 263 : | edgomez | 1382 | paddusw mm6,mm5 |
| 264 : | movd eax, mm6 | ||
| 265 : | pop ebx | ||
| 266 : | ret | ||
| 267 : | Isibaar | 262 | |
| 268 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
| 269 : | ; | ||
| 270 : | ; uint32_t sad8bi_xmm(const uint8_t * const cur, | ||
| 271 : | ; const uint8_t * const ref1, | ||
| 272 : | ; const uint8_t * const ref2, | ||
| 273 : | ; const uint32_t stride); | ||
| 274 : | ; | ||
| 275 : | ;----------------------------------------------------------------------------- | ||
| 276 : | Isibaar | 262 | |
| 277 : | edgomez | 1382 | ALIGN 16 |
| 278 : | sad8bi_xmm: | ||
| 279 : | push ebx | ||
| 280 : | mov eax, [esp+4+ 4] ; Src | ||
| 281 : | mov edx, [esp+4+ 8] ; Ref1 | ||
| 282 : | mov ebx, [esp+4+12] ; Ref2 | ||
| 283 : | mov ecx, [esp+4+16] ; Stride | ||
| 284 : | Isibaar | 262 | |
| 285 : | edgomez | 1382 | pxor mm5, mm5 ; accum1 |
| 286 : | pxor mm6, mm6 ; accum2 | ||
| 287 : | .Loop | ||
| 288 : | SADBI_8x8_XMM | ||
| 289 : | SADBI_8x8_XMM | ||
| 290 : | SADBI_8x8_XMM | ||
| 291 : | SADBI_8x8_XMM | ||
| 292 : | Isibaar | 262 | |
| 293 : | edgomez | 1382 | paddusw mm6,mm5 |
| 294 : | movd eax, mm6 | ||
| 295 : | pop ebx | ||
| 296 : | ret | ||
| 297 : | Isibaar | 262 | |
| 298 : | |||
| 299 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
| 300 : | suxen_drol | 329 | ; |
| 301 : | ; uint32_t dev16_xmm(const uint8_t * const cur, | ||
| 302 : | ; const uint32_t stride); | ||
| 303 : | ; | ||
| 304 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
| 305 : | suxen_drol | 329 | |
| 306 : | edgomez | 1382 | ALIGN 16 |
| 307 : | Isibaar | 262 | dev16_xmm: |
| 308 : | |||
| 309 : | edgomez | 1382 | mov eax, [esp+ 4] ; Src |
| 310 : | mov ecx, [esp+ 8] ; Stride | ||
| 311 : | Isibaar | 262 | |
| 312 : | edgomez | 1382 | pxor mm7, mm7 ; zero |
| 313 : | pxor mm5, mm5 ; mean accums | ||
| 314 : | pxor mm6, mm6 | ||
| 315 : | Isibaar | 262 | |
| 316 : | edgomez | 1382 | MEAN_16x16_SSE |
| 317 : | MEAN_16x16_SSE | ||
| 318 : | MEAN_16x16_SSE | ||
| 319 : | MEAN_16x16_SSE | ||
| 320 : | MEAN_16x16_SSE | ||
| 321 : | MEAN_16x16_SSE | ||
| 322 : | MEAN_16x16_SSE | ||
| 323 : | MEAN_16x16_SSE | ||
| 324 : | Isibaar | 262 | |
| 325 : | edgomez | 1382 | MEAN_16x16_SSE |
| 326 : | MEAN_16x16_SSE | ||
| 327 : | MEAN_16x16_SSE | ||
| 328 : | MEAN_16x16_SSE | ||
| 329 : | MEAN_16x16_SSE | ||
| 330 : | MEAN_16x16_SSE | ||
| 331 : | MEAN_16x16_SSE | ||
| 332 : | MEAN_16x16_SSE | ||
| 333 : | Isibaar | 262 | |
| 334 : | edgomez | 1382 | paddusw mm6, mm5 |
| 335 : | Isibaar | 262 | |
| 336 : | edgomez | 1382 | movq mm4, mm6 |
| 337 : | psllq mm4, 32 | ||
| 338 : | paddd mm4, mm6 | ||
| 339 : | psrld mm4, 8 ; /= (16*16) | ||
| 340 : | Isibaar | 262 | |
| 341 : | edgomez | 1382 | packssdw mm4, mm4 |
| 342 : | packuswb mm4, mm4 | ||
| 343 : | Isibaar | 262 | |
| 344 : | edgomez | 1382 | ; mm4 contains the mean |
| 345 : | Isibaar | 262 | |
| 346 : | edgomez | 1382 | mov eax, [esp+ 4] ; Src |
| 347 : | edgomez | 851 | |
| 348 : | Isibaar | 262 | |
| 349 : | edgomez | 1382 | pxor mm5, mm5 ; sums |
| 350 : | pxor mm6, mm6 | ||
| 351 : | Isibaar | 262 | |
| 352 : | edgomez | 1382 | ABS_16x16_SSE |
| 353 : | ABS_16x16_SSE | ||
| 354 : | ABS_16x16_SSE | ||
| 355 : | ABS_16x16_SSE | ||
| 356 : | ABS_16x16_SSE | ||
| 357 : | ABS_16x16_SSE | ||
| 358 : | ABS_16x16_SSE | ||
| 359 : | ABS_16x16_SSE | ||
| 360 : | Isibaar | 262 | |
| 361 : | edgomez | 1382 | ABS_16x16_SSE |
| 362 : | ABS_16x16_SSE | ||
| 363 : | ABS_16x16_SSE | ||
| 364 : | ABS_16x16_SSE | ||
| 365 : | ABS_16x16_SSE | ||
| 366 : | ABS_16x16_SSE | ||
| 367 : | ABS_16x16_SSE | ||
| 368 : | ABS_16x16_SSE | ||
| 369 : | Isibaar | 262 | |
| 370 : | edgomez | 1382 | paddusw mm6, mm5 |
| 371 : | movq mm7, mm6 | ||
| 372 : | psllq mm7, 32 | ||
| 373 : | paddd mm6, mm7 | ||
| 374 : | edgomez | 851 | |
| 375 : | edgomez | 1382 | movd eax, mm6 |
| 376 : | ret | ||
| 377 : | edgomez | 851 | |
| 378 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
| 379 : | edgomez | 851 | ;int sad16v_xmm(const uint8_t * const cur, |
| 380 : | ; const uint8_t * const ref, | ||
| 381 : | ; const uint32_t stride, | ||
| 382 : | ; int* sad8); | ||
| 383 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
| 384 : | |||
| 385 : | ALIGN 16 | ||
| 386 : | edgomez | 851 | sad16v_xmm: |
| 387 : | edgomez | 1382 | push ebx |
| 388 : | mov eax, [esp+4+ 4] ; Src1 | ||
| 389 : | mov edx, [esp+4+ 8] ; Src2 | ||
| 390 : | mov ecx, [esp+4+12] ; Stride | ||
| 391 : | mov ebx, [esp+4+16] ; sad ptr | ||
| 392 : | edgomez | 851 | |
| 393 : | edgomez | 1382 | pxor mm5, mm5 ; accum1 |
| 394 : | pxor mm6, mm6 ; accum2 | ||
| 395 : | pxor mm7, mm7 ; total | ||
| 396 : | edgomez | 851 | |
| 397 : | edgomez | 1382 | SAD_16x16_SSE |
| 398 : | SAD_16x16_SSE | ||
| 399 : | SAD_16x16_SSE | ||
| 400 : | SAD_16x16_SSE | ||
| 401 : | SAD_16x16_SSE | ||
| 402 : | SAD_16x16_SSE | ||
| 403 : | SAD_16x16_SSE | ||
| 404 : | SAD_16x16_SSE | ||
| 405 : | edgomez | 851 | |
| 406 : | edgomez | 1382 | paddusw mm7, mm5 |
| 407 : | paddusw mm7, mm6 | ||
| 408 : | movd [ebx], mm5 | ||
| 409 : | movd [ebx+4], mm6 | ||
| 410 : | edgomez | 851 | |
| 411 : | edgomez | 1382 | pxor mm5, mm5 ; accum1 |
| 412 : | pxor mm6, mm6 ; accum2 | ||
| 413 : | edgomez | 851 | |
| 414 : | edgomez | 1382 | SAD_16x16_SSE |
| 415 : | SAD_16x16_SSE | ||
| 416 : | SAD_16x16_SSE | ||
| 417 : | SAD_16x16_SSE | ||
| 418 : | SAD_16x16_SSE | ||
| 419 : | SAD_16x16_SSE | ||
| 420 : | SAD_16x16_SSE | ||
| 421 : | SAD_16x16_SSE | ||
| 422 : | |||
| 423 : | paddusw mm7, mm5 | ||
| 424 : | paddusw mm7, mm6 | ||
| 425 : | movd [ebx+8], mm5 | ||
| 426 : | movd [ebx+12], mm6 | ||
| 427 : | |||
| 428 : | movd eax, mm7 | ||
| 429 : | pop ebx | ||
| 430 : | ret |
| No admin address has been configured | ViewVC Help |
| Powered by ViewVC 1.0.4 |