Parent Directory
| 
Revision Log
Revision 1877 - (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 : | ; * 2002 Pascal Massimino <skal@planet-d.net> | ||
| 8 : | Isibaar | 262 | ; * |
| 9 : | edgomez | 1382 | ; * This program is free software; you can redistribute it and/or modify it |
| 10 : | ; * under the terms of the GNU General Public License as published by | ||
| 11 : | ; * the Free Software Foundation; either version 2 of the License, or | ||
| 12 : | ; * (at your option) any later version. | ||
| 13 : | Isibaar | 262 | ; * |
| 14 : | edgomez | 1382 | ; * This program is distributed in the hope that it will be useful, |
| 15 : | ; * but WITHOUT ANY WARRANTY; without even the implied warranty of | ||
| 16 : | ; * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | ||
| 17 : | ; * GNU General Public License for more details. | ||
| 18 : | Isibaar | 262 | ; * |
| 19 : | edgomez | 1382 | ; * You should have received a copy of the GNU General Public License |
| 20 : | ; * along with this program; if not, write to the Free Software | ||
| 21 : | ; * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA | ||
| 22 : | Isibaar | 262 | ; * |
| 23 : | Isibaar | 1877 | ; * $Id: sad_mmx.asm,v 1.22 2009-09-16 17:07:58 Isibaar Exp $ |
| 24 : | Isibaar | 262 | ; * |
| 25 : | edgomez | 1382 | ; ***************************************************************************/ |
| 26 : | Isibaar | 262 | |
| 27 : | Isibaar | 1795 | %include "nasm.inc" |
| 28 : | Isibaar | 262 | |
| 29 : | edgomez | 1382 | ;============================================================================= |
| 30 : | ; Read only data | ||
| 31 : | ;============================================================================= | ||
| 32 : | Isibaar | 262 | |
| 33 : | Isibaar | 1795 | DATA |
| 34 : | Isibaar | 262 | |
| 35 : | Isibaar | 1795 | ALIGN SECTION_ALIGN |
| 36 : | edgomez | 1382 | mmx_one: |
| 37 : | times 4 dw 1 | ||
| 38 : | Isibaar | 262 | |
| 39 : | edgomez | 1382 | ;============================================================================= |
| 40 : | ; Helper macros | ||
| 41 : | ;============================================================================= | ||
| 42 : | Isibaar | 262 | |
| 43 : | %macro SAD_16x16_MMX 0 | ||
| 44 : | Isibaar | 1795 | movq mm0, [_EAX] |
| 45 : | movq mm1, [TMP1] | ||
| 46 : | Isibaar | 262 | |
| 47 : | Isibaar | 1795 | movq mm2, [_EAX+8] |
| 48 : | movq mm3, [TMP1+8] | ||
| 49 : | Isibaar | 262 | |
| 50 : | edgomez | 1382 | movq mm4, mm0 |
| 51 : | psubusb mm0, mm1 | ||
| 52 : | Isibaar | 1795 | lea _EAX, [_EAX+TMP0] |
| 53 : | edgomez | 1382 | movq mm5, mm2 |
| 54 : | psubusb mm2, mm3 | ||
| 55 : | Isibaar | 262 | |
| 56 : | edgomez | 1382 | psubusb mm1, mm4 |
| 57 : | psubusb mm3, mm5 | ||
| 58 : | suxen_drol | 1599 | por mm0, mm1 |
| 59 : | edgomez | 1382 | por mm2, mm3 |
| 60 : | Isibaar | 262 | |
| 61 : | edgomez | 1382 | movq mm1, mm0 |
| 62 : | punpcklbw mm0,mm7 | ||
| 63 : | suxen_drol | 1599 | movq mm3, mm2 |
| 64 : | edgomez | 1382 | punpckhbw mm1,mm7 |
| 65 : | Isibaar | 1795 | lea TMP1, [TMP1+TMP0] |
| 66 : | edgomez | 1382 | punpcklbw mm2,mm7 |
| 67 : | paddusw mm0, mm1 | ||
| 68 : | suxen_drol | 1599 | punpckhbw mm3,mm7 |
| 69 : | edgomez | 1382 | paddusw mm6, mm0 |
| 70 : | paddusw mm2, mm3 | ||
| 71 : | paddusw mm6, mm2 | ||
| 72 : | suxen_drol | 1599 | |
| 73 : | Isibaar | 262 | %endmacro |
| 74 : | |||
| 75 : | edgomez | 1382 | %macro SAD_8x8_MMX 0 |
| 76 : | Isibaar | 1795 | movq mm0, [_EAX] |
| 77 : | movq mm1, [TMP1] | ||
| 78 : | Isibaar | 262 | |
| 79 : | Isibaar | 1795 | movq mm2, [_EAX+TMP0] |
| 80 : | movq mm3, [TMP1+TMP0] | ||
| 81 : | Isibaar | 262 | |
| 82 : | Isibaar | 1795 | lea _EAX,[_EAX+2*TMP0] |
| 83 : | lea TMP1,[TMP1+2*TMP0] | ||
| 84 : | Isibaar | 262 | |
| 85 : | edgomez | 1382 | movq mm4, mm0 |
| 86 : | psubusb mm0, mm1 | ||
| 87 : | movq mm5, mm2 | ||
| 88 : | psubusb mm2, mm3 | ||
| 89 : | Isibaar | 262 | |
| 90 : | edgomez | 1382 | psubusb mm1, mm4 |
| 91 : | psubusb mm3, mm5 | ||
| 92 : | suxen_drol | 1599 | por mm0, mm1 |
| 93 : | edgomez | 1382 | por mm2, mm3 |
| 94 : | Isibaar | 262 | |
| 95 : | edgomez | 1382 | movq mm1,mm0 |
| 96 : | punpcklbw mm0,mm7 | ||
| 97 : | suxen_drol | 1599 | movq mm3,mm2 |
| 98 : | edgomez | 1382 | punpckhbw mm1,mm7 |
| 99 : | punpcklbw mm2,mm7 | ||
| 100 : | paddusw mm0,mm1 | ||
| 101 : | suxen_drol | 1599 | punpckhbw mm3,mm7 |
| 102 : | edgomez | 1382 | paddusw mm6,mm0 |
| 103 : | paddusw mm2,mm3 | ||
| 104 : | paddusw mm6,mm2 | ||
| 105 : | %endmacro | ||
| 106 : | Isibaar | 262 | |
| 107 : | suxen_drol | 1599 | |
| 108 : | edgomez | 1382 | %macro SADV_16x16_MMX 0 |
| 109 : | Isibaar | 1795 | movq mm0, [_EAX] |
| 110 : | movq mm1, [TMP1] | ||
| 111 : | Isibaar | 262 | |
| 112 : | Isibaar | 1795 | movq mm2, [_EAX+8] |
| 113 : | edgomez | 1382 | movq mm4, mm0 |
| 114 : | Isibaar | 1795 | movq mm3, [TMP1+8] |
| 115 : | edgomez | 1382 | psubusb mm0, mm1 |
| 116 : | Isibaar | 262 | |
| 117 : | edgomez | 1382 | psubusb mm1, mm4 |
| 118 : | Isibaar | 1795 | lea _EAX,[_EAX+TMP0] |
| 119 : | suxen_drol | 1599 | por mm0, mm1 |
| 120 : | Isibaar | 262 | |
| 121 : | edgomez | 1382 | movq mm4, mm2 |
| 122 : | psubusb mm2, mm3 | ||
| 123 : | Isibaar | 262 | |
| 124 : | edgomez | 1382 | psubusb mm3, mm4 |
| 125 : | por mm2, mm3 | ||
| 126 : | suxen_drol | 1599 | |
| 127 : | edgomez | 1382 | movq mm1,mm0 |
| 128 : | punpcklbw mm0,mm7 | ||
| 129 : | suxen_drol | 1599 | movq mm3,mm2 |
| 130 : | edgomez | 1382 | punpckhbw mm1,mm7 |
| 131 : | punpcklbw mm2,mm7 | ||
| 132 : | paddusw mm0,mm1 | ||
| 133 : | suxen_drol | 1599 | punpckhbw mm3,mm7 |
| 134 : | edgomez | 1382 | paddusw mm5, mm0 |
| 135 : | suxen_drol | 1599 | paddusw mm2,mm3 |
| 136 : | Isibaar | 1795 | lea TMP1,[TMP1+TMP0] |
| 137 : | edgomez | 1382 | paddusw mm6, mm2 |
| 138 : | Isibaar | 262 | %endmacro |
| 139 : | |||
| 140 : | edgomez | 1382 | %macro SADBI_16x16_MMX 2 ; SADBI_16x16_MMX( int_ptr_offset, bool_increment_ptr ); |
| 141 : | Isibaar | 262 | |
| 142 : | Isibaar | 1795 | movq mm0, [TMP1+%1] |
| 143 : | movq mm2, [_EBX+%1] | ||
| 144 : | edgomez | 1382 | movq mm1, mm0 |
| 145 : | movq mm3, mm2 | ||
| 146 : | Isibaar | 262 | |
| 147 : | edgomez | 1382 | %if %2 != 0 |
| 148 : | Isibaar | 1795 | add TMP1, TMP0 |
| 149 : | edgomez | 1382 | %endif |
| 150 : | Isibaar | 262 | |
| 151 : | edgomez | 1382 | punpcklbw mm0, mm7 |
| 152 : | punpckhbw mm1, mm7 | ||
| 153 : | punpcklbw mm2, mm7 | ||
| 154 : | punpckhbw mm3, mm7 | ||
| 155 : | Isibaar | 262 | |
| 156 : | edgomez | 1382 | %if %2 != 0 |
| 157 : | Isibaar | 1795 | add _EBX, TMP0 |
| 158 : | edgomez | 1382 | %endif |
| 159 : | Isibaar | 262 | |
| 160 : | edgomez | 1382 | paddusw mm0, mm2 ; mm01 = ref1 + ref2 |
| 161 : | paddusw mm1, mm3 | ||
| 162 : | paddusw mm0, [mmx_one] ; mm01 += 1 | ||
| 163 : | paddusw mm1, [mmx_one] | ||
| 164 : | psrlw mm0, 1 ; mm01 >>= 1 | ||
| 165 : | psrlw mm1, 1 | ||
| 166 : | Isibaar | 262 | |
| 167 : | Isibaar | 1795 | movq mm2, [_EAX+%1] |
| 168 : | edgomez | 1382 | movq mm3, mm2 |
| 169 : | punpcklbw mm2, mm7 ; mm23 = src | ||
| 170 : | punpckhbw mm3, mm7 | ||
| 171 : | Isibaar | 262 | |
| 172 : | edgomez | 1382 | %if %2 != 0 |
| 173 : | Isibaar | 1795 | add _EAX, TMP0 |
| 174 : | edgomez | 1382 | %endif |
| 175 : | Isibaar | 262 | |
| 176 : | edgomez | 1382 | movq mm4, mm0 |
| 177 : | movq mm5, mm1 | ||
| 178 : | psubusw mm0, mm2 | ||
| 179 : | psubusw mm1, mm3 | ||
| 180 : | psubusw mm2, mm4 | ||
| 181 : | psubusw mm3, mm5 | ||
| 182 : | por mm0, mm2 ; mm01 = ABS(mm01 - mm23) | ||
| 183 : | por mm1, mm3 | ||
| 184 : | Isibaar | 262 | |
| 185 : | edgomez | 1382 | paddusw mm6, mm0 ; mm6 += mm01 |
| 186 : | paddusw mm6, mm1 | ||
| 187 : | Isibaar | 262 | |
| 188 : | edgomez | 1382 | %endmacro |
| 189 : | edgomez | 851 | |
| 190 : | edgomez | 1382 | %macro MEAN_16x16_MMX 0 |
| 191 : | Isibaar | 1795 | movq mm0, [_EAX] |
| 192 : | movq mm2, [_EAX+8] | ||
| 193 : | lea _EAX, [_EAX+TMP0] | ||
| 194 : | edgomez | 1382 | movq mm1, mm0 |
| 195 : | punpcklbw mm0, mm7 | ||
| 196 : | suxen_drol | 1599 | movq mm3, mm2 |
| 197 : | edgomez | 1382 | punpckhbw mm1, mm7 |
| 198 : | paddw mm5, mm0 | ||
| 199 : | suxen_drol | 1599 | punpcklbw mm2, mm7 |
| 200 : | edgomez | 1382 | paddw mm6, mm1 |
| 201 : | suxen_drol | 1599 | punpckhbw mm3, mm7 |
| 202 : | edgomez | 1382 | paddw mm5, mm2 |
| 203 : | paddw mm6, mm3 | ||
| 204 : | %endmacro | ||
| 205 : | edgomez | 851 | |
| 206 : | edgomez | 1382 | %macro ABS_16x16_MMX 0 |
| 207 : | Isibaar | 1795 | movq mm0, [_EAX] |
| 208 : | movq mm2, [_EAX+8] | ||
| 209 : | lea _EAX, [_EAX+TMP0] | ||
| 210 : | edgomez | 1382 | movq mm1, mm0 |
| 211 : | movq mm3, mm2 | ||
| 212 : | punpcklbw mm0, mm7 | ||
| 213 : | punpcklbw mm2, mm7 | ||
| 214 : | punpckhbw mm1, mm7 | ||
| 215 : | punpckhbw mm3, mm7 | ||
| 216 : | movq mm4, mm6 | ||
| 217 : | psubusw mm4, mm0 | ||
| 218 : | edgomez | 851 | |
| 219 : | edgomez | 1382 | psubusw mm0, mm6 |
| 220 : | por mm0, mm4 | ||
| 221 : | movq mm4, mm6 | ||
| 222 : | psubusw mm4, mm1 | ||
| 223 : | psubusw mm1, mm6 | ||
| 224 : | por mm1, mm4 | ||
| 225 : | edgomez | 851 | |
| 226 : | edgomez | 1382 | movq mm4, mm6 |
| 227 : | psubusw mm4, mm2 | ||
| 228 : | psubusw mm2, mm6 | ||
| 229 : | por mm2, mm4 | ||
| 230 : | movq mm4, mm6 | ||
| 231 : | psubusw mm4, mm3 | ||
| 232 : | psubusw mm3, mm6 | ||
| 233 : | por mm3, mm4 | ||
| 234 : | edgomez | 851 | |
| 235 : | edgomez | 1382 | paddw mm0, mm1 |
| 236 : | paddw mm2, mm3 | ||
| 237 : | paddw mm5, mm0 | ||
| 238 : | paddw mm5, mm2 | ||
| 239 : | %endmacro | ||
| 240 : | edgomez | 851 | |
| 241 : | edgomez | 1382 | ;============================================================================= |
| 242 : | ; Code | ||
| 243 : | ;============================================================================= | ||
| 244 : | edgomez | 851 | |
| 245 : | Isibaar | 1844 | TEXT |
| 246 : | edgomez | 851 | |
| 247 : | edgomez | 1382 | cglobal sad16_mmx |
| 248 : | cglobal sad16v_mmx | ||
| 249 : | cglobal sad8_mmx | ||
| 250 : | cglobal sad16bi_mmx | ||
| 251 : | cglobal sad8bi_mmx | ||
| 252 : | cglobal dev16_mmx | ||
| 253 : | cglobal sse8_16bit_mmx | ||
| 254 : | edgomez | 1424 | cglobal sse8_8bit_mmx |
| 255 : | edgomez | 1382 | |
| 256 : | ;----------------------------------------------------------------------------- | ||
| 257 : | ; | ||
| 258 : | ; uint32_t sad16_mmx(const uint8_t * const cur, | ||
| 259 : | ; const uint8_t * const ref, | ||
| 260 : | ; const uint32_t stride, | ||
| 261 : | ; const uint32_t best_sad); | ||
| 262 : | ; | ||
| 263 : | ; (early termination ignore; slows this down) | ||
| 264 : | ; | ||
| 265 : | ;----------------------------------------------------------------------------- | ||
| 266 : | edgomez | 851 | |
| 267 : | Isibaar | 1795 | ALIGN SECTION_ALIGN |
| 268 : | edgomez | 1382 | sad16_mmx: |
| 269 : | edgomez | 851 | |
| 270 : | Isibaar | 1795 | mov _EAX, prm1 ; Src1 |
| 271 : | mov TMP1, prm2 ; Src2 | ||
| 272 : | mov TMP0, prm3 ; Stride | ||
| 273 : | edgomez | 851 | |
| 274 : | edgomez | 1382 | pxor mm6, mm6 ; accum |
| 275 : | pxor mm7, mm7 ; zero | ||
| 276 : | edgomez | 851 | |
| 277 : | edgomez | 1382 | SAD_16x16_MMX |
| 278 : | SAD_16x16_MMX | ||
| 279 : | SAD_16x16_MMX | ||
| 280 : | SAD_16x16_MMX | ||
| 281 : | SAD_16x16_MMX | ||
| 282 : | SAD_16x16_MMX | ||
| 283 : | SAD_16x16_MMX | ||
| 284 : | SAD_16x16_MMX | ||
| 285 : | edgomez | 851 | |
| 286 : | edgomez | 1382 | SAD_16x16_MMX |
| 287 : | SAD_16x16_MMX | ||
| 288 : | SAD_16x16_MMX | ||
| 289 : | SAD_16x16_MMX | ||
| 290 : | SAD_16x16_MMX | ||
| 291 : | SAD_16x16_MMX | ||
| 292 : | SAD_16x16_MMX | ||
| 293 : | SAD_16x16_MMX | ||
| 294 : | edgomez | 851 | |
| 295 : | edgomez | 1382 | pmaddwd mm6, [mmx_one] ; collapse |
| 296 : | movq mm7, mm6 | ||
| 297 : | psrlq mm7, 32 | ||
| 298 : | paddd mm6, mm7 | ||
| 299 : | edgomez | 851 | |
| 300 : | edgomez | 1382 | movd eax, mm6 |
| 301 : | edgomez | 851 | |
| 302 : | edgomez | 1382 | ret |
| 303 : | Isibaar | 1793 | ENDFUNC |
| 304 : | edgomez | 851 | |
| 305 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
| 306 : | ; | ||
| 307 : | ; uint32_t sad8_mmx(const uint8_t * const cur, | ||
| 308 : | ; const uint8_t * const ref, | ||
| 309 : | ; const uint32_t stride); | ||
| 310 : | ; | ||
| 311 : | ;----------------------------------------------------------------------------- | ||
| 312 : | edgomez | 851 | |
| 313 : | Isibaar | 1795 | ALIGN SECTION_ALIGN |
| 314 : | edgomez | 1382 | sad8_mmx: |
| 315 : | edgomez | 851 | |
| 316 : | Isibaar | 1795 | mov _EAX, prm1 ; Src1 |
| 317 : | mov TMP1, prm2 ; Src2 | ||
| 318 : | mov TMP0, prm3 ; Stride | ||
| 319 : | edgomez | 851 | |
| 320 : | edgomez | 1382 | pxor mm6, mm6 ; accum |
| 321 : | pxor mm7, mm7 ; zero | ||
| 322 : | edgomez | 851 | |
| 323 : | edgomez | 1382 | SAD_8x8_MMX |
| 324 : | SAD_8x8_MMX | ||
| 325 : | SAD_8x8_MMX | ||
| 326 : | SAD_8x8_MMX | ||
| 327 : | edgomez | 851 | |
| 328 : | edgomez | 1382 | pmaddwd mm6, [mmx_one] ; collapse |
| 329 : | movq mm7, mm6 | ||
| 330 : | psrlq mm7, 32 | ||
| 331 : | paddd mm6, mm7 | ||
| 332 : | edgomez | 851 | |
| 333 : | edgomez | 1382 | movd eax, mm6 |
| 334 : | edgomez | 851 | |
| 335 : | edgomez | 1382 | ret |
| 336 : | Isibaar | 1793 | ENDFUNC |
| 337 : | edgomez | 851 | |
| 338 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
| 339 : | ; | ||
| 340 : | ; uint32_t sad16v_mmx(const uint8_t * const cur, | ||
| 341 : | ; const uint8_t * const ref, | ||
| 342 : | ; const uint32_t stride, | ||
| 343 : | ; int32_t *sad); | ||
| 344 : | ; | ||
| 345 : | ;----------------------------------------------------------------------------- | ||
| 346 : | edgomez | 851 | |
| 347 : | Isibaar | 1795 | ALIGN SECTION_ALIGN |
| 348 : | edgomez | 1382 | sad16v_mmx: |
| 349 : | edgomez | 851 | |
| 350 : | Isibaar | 1795 | mov _EAX, prm1 ; Src1 |
| 351 : | mov TMP1, prm2 ; Src2 | ||
| 352 : | mov TMP0, prm3 ; Stride | ||
| 353 : | edgomez | 851 | |
| 354 : | Isibaar | 1795 | push _EBX |
| 355 : | push _EDI | ||
| 356 : | %ifdef ARCH_IS_X86_64 | ||
| 357 : | mov _EBX, prm4 | ||
| 358 : | %else | ||
| 359 : | mov _EBX, [_ESP + 8 + 16] ; sad ptr | ||
| 360 : | %endif | ||
| 361 : | edgomez | 851 | |
| 362 : | edgomez | 1382 | pxor mm5, mm5 ; accum |
| 363 : | pxor mm6, mm6 ; accum | ||
| 364 : | pxor mm7, mm7 ; zero | ||
| 365 : | suxen_drol | 329 | |
| 366 : | edgomez | 1382 | SADV_16x16_MMX |
| 367 : | SADV_16x16_MMX | ||
| 368 : | SADV_16x16_MMX | ||
| 369 : | SADV_16x16_MMX | ||
| 370 : | SADV_16x16_MMX | ||
| 371 : | SADV_16x16_MMX | ||
| 372 : | SADV_16x16_MMX | ||
| 373 : | SADV_16x16_MMX | ||
| 374 : | suxen_drol | 329 | |
| 375 : | edgomez | 1382 | pmaddwd mm5, [mmx_one] ; collapse |
| 376 : | pmaddwd mm6, [mmx_one] ; collapse | ||
| 377 : | suxen_drol | 329 | |
| 378 : | edgomez | 1382 | movq mm2, mm5 |
| 379 : | movq mm3, mm6 | ||
| 380 : | suxen_drol | 329 | |
| 381 : | edgomez | 1382 | psrlq mm2, 32 |
| 382 : | psrlq mm3, 32 | ||
| 383 : | suxen_drol | 329 | |
| 384 : | edgomez | 1382 | paddd mm5, mm2 |
| 385 : | paddd mm6, mm3 | ||
| 386 : | suxen_drol | 329 | |
| 387 : | Isibaar | 1795 | movd [_EBX], mm5 |
| 388 : | movd [_EBX + 4], mm6 | ||
| 389 : | suxen_drol | 329 | |
| 390 : | edgomez | 1382 | paddd mm5, mm6 |
| 391 : | suxen_drol | 329 | |
| 392 : | edgomez | 1382 | movd edi, mm5 |
| 393 : | suxen_drol | 329 | |
| 394 : | edgomez | 1382 | pxor mm5, mm5 |
| 395 : | pxor mm6, mm6 | ||
| 396 : | suxen_drol | 329 | |
| 397 : | edgomez | 1382 | SADV_16x16_MMX |
| 398 : | SADV_16x16_MMX | ||
| 399 : | SADV_16x16_MMX | ||
| 400 : | SADV_16x16_MMX | ||
| 401 : | SADV_16x16_MMX | ||
| 402 : | SADV_16x16_MMX | ||
| 403 : | SADV_16x16_MMX | ||
| 404 : | SADV_16x16_MMX | ||
| 405 : | suxen_drol | 329 | |
| 406 : | edgomez | 1382 | pmaddwd mm5, [mmx_one] ; collapse |
| 407 : | pmaddwd mm6, [mmx_one] ; collapse | ||
| 408 : | suxen_drol | 329 | |
| 409 : | edgomez | 1382 | movq mm2, mm5 |
| 410 : | movq mm3, mm6 | ||
| 411 : | suxen_drol | 329 | |
| 412 : | edgomez | 1382 | psrlq mm2, 32 |
| 413 : | psrlq mm3, 32 | ||
| 414 : | suxen_drol | 329 | |
| 415 : | edgomez | 1382 | paddd mm5, mm2 |
| 416 : | paddd mm6, mm3 | ||
| 417 : | suxen_drol | 329 | |
| 418 : | Isibaar | 1795 | movd [_EBX + 8], mm5 |
| 419 : | movd [_EBX + 12], mm6 | ||
| 420 : | suxen_drol | 329 | |
| 421 : | edgomez | 1382 | paddd mm5, mm6 |
| 422 : | suxen_drol | 329 | |
| 423 : | edgomez | 1382 | movd eax, mm5 |
| 424 : | suxen_drol | 329 | |
| 425 : | Isibaar | 1795 | add _EAX, _EDI |
| 426 : | suxen_drol | 329 | |
| 427 : | Isibaar | 1795 | pop _EDI |
| 428 : | pop _EBX | ||
| 429 : | suxen_drol | 329 | |
| 430 : | edgomez | 1382 | ret |
| 431 : | Isibaar | 1793 | ENDFUNC |
| 432 : | suxen_drol | 329 | |
| 433 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
| 434 : | ; | ||
| 435 : | ; uint32_t sad16bi_mmx(const uint8_t * const cur, | ||
| 436 : | ; const uint8_t * const ref1, | ||
| 437 : | ; const uint8_t * const ref2, | ||
| 438 : | ; const uint32_t stride); | ||
| 439 : | ; | ||
| 440 : | ;----------------------------------------------------------------------------- | ||
| 441 : | suxen_drol | 329 | |
| 442 : | Isibaar | 1795 | ALIGN SECTION_ALIGN |
| 443 : | edgomez | 1382 | sad16bi_mmx: |
| 444 : | Isibaar | 1795 | mov _EAX, prm1 ; Src |
| 445 : | mov TMP1, prm2 ; Ref1 | ||
| 446 : | mov TMP0, prm4 ; Stride | ||
| 447 : | suxen_drol | 329 | |
| 448 : | Isibaar | 1795 | push _EBX |
| 449 : | %ifdef ARCH_IS_X86_64 | ||
| 450 : | mov _EBX, prm3 ; Ref2 | ||
| 451 : | %else | ||
| 452 : | mov _EBX, [_ESP+4+12] ; Ref2 | ||
| 453 : | %endif | ||
| 454 : | |||
| 455 : | edgomez | 1382 | pxor mm6, mm6 ; accum2 |
| 456 : | pxor mm7, mm7 | ||
| 457 : | Isibaar | 1793 | .Loop: |
| 458 : | edgomez | 1382 | SADBI_16x16_MMX 0, 0 |
| 459 : | SADBI_16x16_MMX 8, 1 | ||
| 460 : | SADBI_16x16_MMX 0, 0 | ||
| 461 : | SADBI_16x16_MMX 8, 1 | ||
| 462 : | SADBI_16x16_MMX 0, 0 | ||
| 463 : | SADBI_16x16_MMX 8, 1 | ||
| 464 : | SADBI_16x16_MMX 0, 0 | ||
| 465 : | SADBI_16x16_MMX 8, 1 | ||
| 466 : | SADBI_16x16_MMX 0, 0 | ||
| 467 : | SADBI_16x16_MMX 8, 1 | ||
| 468 : | SADBI_16x16_MMX 0, 0 | ||
| 469 : | SADBI_16x16_MMX 8, 1 | ||
| 470 : | SADBI_16x16_MMX 0, 0 | ||
| 471 : | SADBI_16x16_MMX 8, 1 | ||
| 472 : | SADBI_16x16_MMX 0, 0 | ||
| 473 : | SADBI_16x16_MMX 8, 1 | ||
| 474 : | |||
| 475 : | SADBI_16x16_MMX 0, 0 | ||
| 476 : | SADBI_16x16_MMX 8, 1 | ||
| 477 : | SADBI_16x16_MMX 0, 0 | ||
| 478 : | SADBI_16x16_MMX 8, 1 | ||
| 479 : | SADBI_16x16_MMX 0, 0 | ||
| 480 : | SADBI_16x16_MMX 8, 1 | ||
| 481 : | SADBI_16x16_MMX 0, 0 | ||
| 482 : | SADBI_16x16_MMX 8, 1 | ||
| 483 : | SADBI_16x16_MMX 0, 0 | ||
| 484 : | SADBI_16x16_MMX 8, 1 | ||
| 485 : | SADBI_16x16_MMX 0, 0 | ||
| 486 : | SADBI_16x16_MMX 8, 1 | ||
| 487 : | SADBI_16x16_MMX 0, 0 | ||
| 488 : | SADBI_16x16_MMX 8, 1 | ||
| 489 : | SADBI_16x16_MMX 0, 0 | ||
| 490 : | SADBI_16x16_MMX 8, 1 | ||
| 491 : | |||
| 492 : | pmaddwd mm6, [mmx_one] ; collapse | ||
| 493 : | movq mm7, mm6 | ||
| 494 : | psrlq mm7, 32 | ||
| 495 : | paddd mm6, mm7 | ||
| 496 : | |||
| 497 : | movd eax, mm6 | ||
| 498 : | Isibaar | 1795 | pop _EBX |
| 499 : | edgomez | 1382 | |
| 500 : | ret | ||
| 501 : | Isibaar | 1793 | ENDFUNC |
| 502 : | edgomez | 1382 | |
| 503 : | ;----------------------------------------------------------------------------- | ||
| 504 : | suxen_drol | 329 | ; |
| 505 : | edgomez | 1382 | ; uint32_t sad8bi_mmx(const uint8_t * const cur, |
| 506 : | ; const uint8_t * const ref1, | ||
| 507 : | ; const uint8_t * const ref2, | ||
| 508 : | ; const uint32_t stride); | ||
| 509 : | suxen_drol | 329 | ; |
| 510 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
| 511 : | suxen_drol | 329 | |
| 512 : | Isibaar | 1795 | ALIGN SECTION_ALIGN |
| 513 : | edgomez | 1382 | sad8bi_mmx: |
| 514 : | Isibaar | 1795 | mov _EAX, prm1 ; Src |
| 515 : | mov TMP1, prm2 ; Ref1 | ||
| 516 : | mov TMP0, prm4 ; Stride | ||
| 517 : | Isibaar | 262 | |
| 518 : | Isibaar | 1795 | push _EBX |
| 519 : | %ifdef ARCH_IS_X86_64 | ||
| 520 : | mov _EBX, prm3 | ||
| 521 : | %else | ||
| 522 : | mov _EBX, [_ESP+4+12] ; Ref2 | ||
| 523 : | %endif | ||
| 524 : | |||
| 525 : | edgomez | 1382 | pxor mm6, mm6 ; accum2 |
| 526 : | pxor mm7, mm7 | ||
| 527 : | Isibaar | 1793 | .Loop: |
| 528 : | edgomez | 1382 | SADBI_16x16_MMX 0, 1 |
| 529 : | SADBI_16x16_MMX 0, 1 | ||
| 530 : | SADBI_16x16_MMX 0, 1 | ||
| 531 : | SADBI_16x16_MMX 0, 1 | ||
| 532 : | SADBI_16x16_MMX 0, 1 | ||
| 533 : | SADBI_16x16_MMX 0, 1 | ||
| 534 : | SADBI_16x16_MMX 0, 1 | ||
| 535 : | SADBI_16x16_MMX 0, 1 | ||
| 536 : | Isibaar | 262 | |
| 537 : | edgomez | 1382 | pmaddwd mm6, [mmx_one] ; collapse |
| 538 : | movq mm7, mm6 | ||
| 539 : | psrlq mm7, 32 | ||
| 540 : | paddd mm6, mm7 | ||
| 541 : | Isibaar | 262 | |
| 542 : | edgomez | 1382 | movd eax, mm6 |
| 543 : | Isibaar | 1795 | pop _EBX |
| 544 : | edgomez | 1382 | ret |
| 545 : | Isibaar | 1793 | ENDFUNC |
| 546 : | Isibaar | 262 | |
| 547 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
| 548 : | ; | ||
| 549 : | ; uint32_t dev16_mmx(const uint8_t * const cur, | ||
| 550 : | ; const uint32_t stride); | ||
| 551 : | ; | ||
| 552 : | ;----------------------------------------------------------------------------- | ||
| 553 : | |||
| 554 : | Isibaar | 1795 | ALIGN SECTION_ALIGN |
| 555 : | Isibaar | 262 | dev16_mmx: |
| 556 : | Isibaar | 1795 | mov _EAX, prm1 ; Src |
| 557 : | mov TMP0, prm2 ; Stride | ||
| 558 : | Isibaar | 262 | |
| 559 : | edgomez | 1382 | pxor mm7, mm7 ; zero |
| 560 : | pxor mm5, mm5 ; accum1 | ||
| 561 : | pxor mm6, mm6 ; accum2 | ||
| 562 : | Isibaar | 262 | |
| 563 : | edgomez | 1382 | MEAN_16x16_MMX |
| 564 : | MEAN_16x16_MMX | ||
| 565 : | MEAN_16x16_MMX | ||
| 566 : | MEAN_16x16_MMX | ||
| 567 : | MEAN_16x16_MMX | ||
| 568 : | MEAN_16x16_MMX | ||
| 569 : | MEAN_16x16_MMX | ||
| 570 : | MEAN_16x16_MMX | ||
| 571 : | Isibaar | 262 | |
| 572 : | edgomez | 1382 | MEAN_16x16_MMX |
| 573 : | MEAN_16x16_MMX | ||
| 574 : | MEAN_16x16_MMX | ||
| 575 : | MEAN_16x16_MMX | ||
| 576 : | MEAN_16x16_MMX | ||
| 577 : | MEAN_16x16_MMX | ||
| 578 : | MEAN_16x16_MMX | ||
| 579 : | MEAN_16x16_MMX | ||
| 580 : | Isibaar | 262 | |
| 581 : | edgomez | 1382 | paddusw mm6, mm5 |
| 582 : | pmaddwd mm6, [mmx_one] ; collapse | ||
| 583 : | movq mm5, mm6 | ||
| 584 : | psrlq mm5, 32 | ||
| 585 : | paddd mm6, mm5 | ||
| 586 : | Isibaar | 262 | |
| 587 : | edgomez | 1382 | psllq mm6, 32 ; blank upper dword |
| 588 : | psrlq mm6, 32 + 8 ; /= (16*16) | ||
| 589 : | Isibaar | 262 | |
| 590 : | edgomez | 1382 | punpckldq mm6, mm6 |
| 591 : | packssdw mm6, mm6 | ||
| 592 : | |||
| 593 : | Isibaar | 262 | ; mm6 contains the mean |
| 594 : | ; mm5 is the new accum | ||
| 595 : | |||
| 596 : | edgomez | 1382 | pxor mm5, mm5 |
| 597 : | Isibaar | 1795 | mov _EAX, prm1 ; Src |
| 598 : | Isibaar | 262 | |
| 599 : | edgomez | 1382 | ABS_16x16_MMX |
| 600 : | ABS_16x16_MMX | ||
| 601 : | ABS_16x16_MMX | ||
| 602 : | ABS_16x16_MMX | ||
| 603 : | ABS_16x16_MMX | ||
| 604 : | ABS_16x16_MMX | ||
| 605 : | ABS_16x16_MMX | ||
| 606 : | ABS_16x16_MMX | ||
| 607 : | Isibaar | 262 | |
| 608 : | edgomez | 1382 | ABS_16x16_MMX |
| 609 : | ABS_16x16_MMX | ||
| 610 : | ABS_16x16_MMX | ||
| 611 : | ABS_16x16_MMX | ||
| 612 : | ABS_16x16_MMX | ||
| 613 : | ABS_16x16_MMX | ||
| 614 : | ABS_16x16_MMX | ||
| 615 : | ABS_16x16_MMX | ||
| 616 : | Isibaar | 262 | |
| 617 : | edgomez | 1382 | pmaddwd mm5, [mmx_one] ; collapse |
| 618 : | movq mm6, mm5 | ||
| 619 : | psrlq mm6, 32 | ||
| 620 : | paddd mm6, mm5 | ||
| 621 : | Isibaar | 262 | |
| 622 : | edgomez | 1382 | movd eax, mm6 |
| 623 : | |||
| 624 : | ret | ||
| 625 : | Isibaar | 1793 | ENDFUNC |
| 626 : | edgomez | 1382 | |
| 627 : | ;----------------------------------------------------------------------------- | ||
| 628 : | ; | ||
| 629 : | ; uint32_t sse8_16bit_mmx(const int16_t *b1, | ||
| 630 : | ; const int16_t *b2, | ||
| 631 : | ; const uint32_t stride); | ||
| 632 : | ; | ||
| 633 : | ;----------------------------------------------------------------------------- | ||
| 634 : | |||
| 635 : | edgomez | 1424 | %macro ROW_SSE_16bit_MMX 2 |
| 636 : | edgomez | 1382 | movq mm0, [%1] |
| 637 : | movq mm1, [%1+8] | ||
| 638 : | psubw mm0, [%2] | ||
| 639 : | psubw mm1, [%2+8] | ||
| 640 : | pmaddwd mm0, mm0 | ||
| 641 : | pmaddwd mm1, mm1 | ||
| 642 : | paddd mm2, mm0 | ||
| 643 : | paddd mm2, mm1 | ||
| 644 : | edgomez | 1424 | %endmacro |
| 645 : | edgomez | 1382 | |
| 646 : | sse8_16bit_mmx: | ||
| 647 : | |||
| 648 : | ;; Load the function params | ||
| 649 : | Isibaar | 1795 | mov _EAX, prm1 |
| 650 : | mov TMP0, prm2 | ||
| 651 : | mov TMP1, prm3 | ||
| 652 : | edgomez | 1382 | |
| 653 : | ;; Reset the sse accumulator | ||
| 654 : | pxor mm2, mm2 | ||
| 655 : | |||
| 656 : | ;; Let's go | ||
| 657 : | edgomez | 1424 | %rep 8 |
| 658 : | Isibaar | 1795 | ROW_SSE_16bit_MMX _EAX, TMP0 |
| 659 : | lea _EAX, [_EAX+TMP1] | ||
| 660 : | lea TMP0, [TMP0+TMP1] | ||
| 661 : | edgomez | 1424 | %endrep |
| 662 : | edgomez | 1382 | |
| 663 : | ;; Finish adding each dword of the accumulator | ||
| 664 : | movq mm3, mm2 | ||
| 665 : | psrlq mm2, 32 | ||
| 666 : | paddd mm2, mm3 | ||
| 667 : | movd eax, mm2 | ||
| 668 : | |||
| 669 : | ;; All done | ||
| 670 : | ret | ||
| 671 : | Isibaar | 1793 | ENDFUNC |
| 672 : | edgomez | 1424 | |
| 673 : | ;----------------------------------------------------------------------------- | ||
| 674 : | ; | ||
| 675 : | ; uint32_t sse8_8bit_mmx(const int8_t *b1, | ||
| 676 : | ; const int8_t *b2, | ||
| 677 : | ; const uint32_t stride); | ||
| 678 : | ; | ||
| 679 : | ;----------------------------------------------------------------------------- | ||
| 680 : | |||
| 681 : | %macro ROW_SSE_8bit_MMX 2 | ||
| 682 : | movq mm0, [%1] ; load a row | ||
| 683 : | movq mm2, [%2] ; load a row | ||
| 684 : | |||
| 685 : | movq mm1, mm0 ; copy row | ||
| 686 : | movq mm3, mm2 ; copy row | ||
| 687 : | |||
| 688 : | punpcklbw mm0, mm7 ; turn the 4low elements into 16bit | ||
| 689 : | punpckhbw mm1, mm7 ; turn the 4high elements into 16bit | ||
| 690 : | |||
| 691 : | punpcklbw mm2, mm7 ; turn the 4low elements into 16bit | ||
| 692 : | punpckhbw mm3, mm7 ; turn the 4high elements into 16bit | ||
| 693 : | |||
| 694 : | psubw mm0, mm2 ; low part of src-dst | ||
| 695 : | psubw mm1, mm3 ; high part of src-dst | ||
| 696 : | |||
| 697 : | pmaddwd mm0, mm0 ; compute the square sum | ||
| 698 : | pmaddwd mm1, mm1 ; compute the square sum | ||
| 699 : | |||
| 700 : | paddd mm6, mm0 ; add to the accumulator | ||
| 701 : | paddd mm6, mm1 ; add to the accumulator | ||
| 702 : | %endmacro | ||
| 703 : | |||
| 704 : | sse8_8bit_mmx: | ||
| 705 : | |||
| 706 : | ;; Load the function params | ||
| 707 : | Isibaar | 1795 | mov _EAX, prm1 |
| 708 : | mov TMP0, prm2 | ||
| 709 : | mov TMP1, prm3 | ||
| 710 : | edgomez | 1424 | |
| 711 : | ;; Reset the sse accumulator | ||
| 712 : | pxor mm6, mm6 | ||
| 713 : | |||
| 714 : | ;; Used to interleave 8bit data with 0x00 values | ||
| 715 : | pxor mm7, mm7 | ||
| 716 : | |||
| 717 : | ;; Let's go | ||
| 718 : | %rep 8 | ||
| 719 : | Isibaar | 1795 | ROW_SSE_8bit_MMX _EAX, TMP0 |
| 720 : | lea _EAX, [_EAX+TMP1] | ||
| 721 : | lea TMP0, [TMP0+TMP1] | ||
| 722 : | edgomez | 1424 | %endrep |
| 723 : | |||
| 724 : | ;; Finish adding each dword of the accumulator | ||
| 725 : | movq mm7, mm6 | ||
| 726 : | psrlq mm6, 32 | ||
| 727 : | paddd mm6, mm7 | ||
| 728 : | movd eax, mm6 | ||
| 729 : | |||
| 730 : | ;; All done | ||
| 731 : | ret | ||
| 732 : | Isibaar | 1793 | ENDFUNC |
| 733 : | edgomez | 1540 | |
| 734 : | Isibaar | 1877 | NON_EXEC_STACK |
| No admin address has been configured | ViewVC Help |
| Powered by ViewVC 1.0.4 |