Parent Directory
| 
Revision Log
Revision 1790 - (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 | 1790 | ; * $Id: sad_mmx.asm,v 1.18 2008-08-19 09:06:48 Isibaar Exp $ |
| 24 : | Isibaar | 262 | ; * |
| 25 : | edgomez | 1382 | ; ***************************************************************************/ |
| 26 : | Isibaar | 262 | |
| 27 : | edgomez | 1382 | BITS 32 |
| 28 : | Isibaar | 262 | |
| 29 : | edgomez | 1382 | %macro cglobal 1 |
| 30 : | suxen_drol | 329 | %ifdef PREFIX |
| 31 : | edgomez | 1535 | %ifdef MARK_FUNCS |
| 32 : | edgomez | 1540 | global _%1:function %1.endfunc-%1 |
| 33 : | %define %1 _%1:function %1.endfunc-%1 | ||
| 34 : | edgomez | 1535 | %else |
| 35 : | global _%1 | ||
| 36 : | %define %1 _%1 | ||
| 37 : | %endif | ||
| 38 : | Isibaar | 262 | %else |
| 39 : | edgomez | 1535 | %ifdef MARK_FUNCS |
| 40 : | edgomez | 1540 | global %1:function %1.endfunc-%1 |
| 41 : | edgomez | 1535 | %else |
| 42 : | global %1 | ||
| 43 : | %endif | ||
| 44 : | Isibaar | 262 | %endif |
| 45 : | %endmacro | ||
| 46 : | |||
| 47 : | edgomez | 1382 | ;============================================================================= |
| 48 : | ; Read only data | ||
| 49 : | ;============================================================================= | ||
| 50 : | Isibaar | 262 | |
| 51 : | edgomez | 1382 | %ifdef FORMAT_COFF |
| 52 : | edgomez | 1519 | SECTION .rodata |
| 53 : | edgomez | 1382 | %else |
| 54 : | edgomez | 1519 | SECTION .rodata align=16 |
| 55 : | edgomez | 1382 | %endif |
| 56 : | Isibaar | 262 | |
| 57 : | edgomez | 1382 | ALIGN 16 |
| 58 : | mmx_one: | ||
| 59 : | times 4 dw 1 | ||
| 60 : | Isibaar | 262 | |
| 61 : | edgomez | 1382 | ;============================================================================= |
| 62 : | ; Helper macros | ||
| 63 : | ;============================================================================= | ||
| 64 : | Isibaar | 262 | |
| 65 : | %macro SAD_16x16_MMX 0 | ||
| 66 : | edgomez | 1382 | movq mm0, [eax] |
| 67 : | movq mm1, [edx] | ||
| 68 : | Isibaar | 262 | |
| 69 : | edgomez | 1382 | movq mm2, [eax+8] |
| 70 : | movq mm3, [edx+8] | ||
| 71 : | Isibaar | 262 | |
| 72 : | edgomez | 1382 | movq mm4, mm0 |
| 73 : | psubusb mm0, mm1 | ||
| 74 : | lea eax, [eax+ecx] | ||
| 75 : | movq mm5, mm2 | ||
| 76 : | psubusb mm2, mm3 | ||
| 77 : | Isibaar | 262 | |
| 78 : | edgomez | 1382 | psubusb mm1, mm4 |
| 79 : | psubusb mm3, mm5 | ||
| 80 : | suxen_drol | 1599 | por mm0, mm1 |
| 81 : | edgomez | 1382 | por mm2, mm3 |
| 82 : | Isibaar | 262 | |
| 83 : | edgomez | 1382 | movq mm1, mm0 |
| 84 : | punpcklbw mm0,mm7 | ||
| 85 : | suxen_drol | 1599 | movq mm3, mm2 |
| 86 : | edgomez | 1382 | punpckhbw mm1,mm7 |
| 87 : | suxen_drol | 1599 | lea edx, [edx+ecx] |
| 88 : | edgomez | 1382 | punpcklbw mm2,mm7 |
| 89 : | paddusw mm0, mm1 | ||
| 90 : | suxen_drol | 1599 | punpckhbw mm3,mm7 |
| 91 : | edgomez | 1382 | paddusw mm6, mm0 |
| 92 : | paddusw mm2, mm3 | ||
| 93 : | paddusw mm6, mm2 | ||
| 94 : | suxen_drol | 1599 | |
| 95 : | Isibaar | 262 | %endmacro |
| 96 : | |||
| 97 : | edgomez | 1382 | %macro SAD_8x8_MMX 0 |
| 98 : | movq mm0, [eax] | ||
| 99 : | movq mm1, [edx] | ||
| 100 : | Isibaar | 262 | |
| 101 : | edgomez | 1382 | movq mm2, [eax+ecx] |
| 102 : | movq mm3, [edx+ecx] | ||
| 103 : | Isibaar | 262 | |
| 104 : | edgomez | 1382 | lea eax,[eax+2*ecx] |
| 105 : | lea edx,[edx+2*ecx] | ||
| 106 : | Isibaar | 262 | |
| 107 : | edgomez | 1382 | movq mm4, mm0 |
| 108 : | psubusb mm0, mm1 | ||
| 109 : | movq mm5, mm2 | ||
| 110 : | psubusb mm2, mm3 | ||
| 111 : | Isibaar | 262 | |
| 112 : | edgomez | 1382 | psubusb mm1, mm4 |
| 113 : | psubusb mm3, mm5 | ||
| 114 : | suxen_drol | 1599 | por mm0, mm1 |
| 115 : | edgomez | 1382 | por mm2, mm3 |
| 116 : | Isibaar | 262 | |
| 117 : | edgomez | 1382 | movq mm1,mm0 |
| 118 : | punpcklbw mm0,mm7 | ||
| 119 : | suxen_drol | 1599 | movq mm3,mm2 |
| 120 : | edgomez | 1382 | punpckhbw mm1,mm7 |
| 121 : | punpcklbw mm2,mm7 | ||
| 122 : | paddusw mm0,mm1 | ||
| 123 : | suxen_drol | 1599 | punpckhbw mm3,mm7 |
| 124 : | edgomez | 1382 | paddusw mm6,mm0 |
| 125 : | paddusw mm2,mm3 | ||
| 126 : | paddusw mm6,mm2 | ||
| 127 : | %endmacro | ||
| 128 : | Isibaar | 262 | |
| 129 : | suxen_drol | 1599 | |
| 130 : | edgomez | 1382 | %macro SADV_16x16_MMX 0 |
| 131 : | movq mm0, [eax] | ||
| 132 : | movq mm1, [edx] | ||
| 133 : | Isibaar | 262 | |
| 134 : | edgomez | 1382 | movq mm2, [eax+8] |
| 135 : | movq mm4, mm0 | ||
| 136 : | suxen_drol | 1599 | movq mm3, [edx+8] |
| 137 : | edgomez | 1382 | psubusb mm0, mm1 |
| 138 : | Isibaar | 262 | |
| 139 : | edgomez | 1382 | psubusb mm1, mm4 |
| 140 : | lea eax,[eax+ecx] | ||
| 141 : | suxen_drol | 1599 | por mm0, mm1 |
| 142 : | Isibaar | 262 | |
| 143 : | edgomez | 1382 | movq mm4, mm2 |
| 144 : | psubusb mm2, mm3 | ||
| 145 : | Isibaar | 262 | |
| 146 : | edgomez | 1382 | psubusb mm3, mm4 |
| 147 : | por mm2, mm3 | ||
| 148 : | suxen_drol | 1599 | |
| 149 : | edgomez | 1382 | movq mm1,mm0 |
| 150 : | punpcklbw mm0,mm7 | ||
| 151 : | suxen_drol | 1599 | movq mm3,mm2 |
| 152 : | edgomez | 1382 | punpckhbw mm1,mm7 |
| 153 : | punpcklbw mm2,mm7 | ||
| 154 : | paddusw mm0,mm1 | ||
| 155 : | suxen_drol | 1599 | punpckhbw mm3,mm7 |
| 156 : | edgomez | 1382 | paddusw mm5, mm0 |
| 157 : | suxen_drol | 1599 | paddusw mm2,mm3 |
| 158 : | lea edx,[edx+ecx] | ||
| 159 : | edgomez | 1382 | paddusw mm6, mm2 |
| 160 : | Isibaar | 262 | %endmacro |
| 161 : | |||
| 162 : | edgomez | 1382 | %macro SADBI_16x16_MMX 2 ; SADBI_16x16_MMX( int_ptr_offset, bool_increment_ptr ); |
| 163 : | Isibaar | 262 | |
| 164 : | edgomez | 1382 | movq mm0, [edx+%1] |
| 165 : | movq mm2, [ebx+%1] | ||
| 166 : | movq mm1, mm0 | ||
| 167 : | movq mm3, mm2 | ||
| 168 : | Isibaar | 262 | |
| 169 : | edgomez | 1382 | %if %2 != 0 |
| 170 : | add edx, ecx | ||
| 171 : | %endif | ||
| 172 : | Isibaar | 262 | |
| 173 : | edgomez | 1382 | punpcklbw mm0, mm7 |
| 174 : | punpckhbw mm1, mm7 | ||
| 175 : | punpcklbw mm2, mm7 | ||
| 176 : | punpckhbw mm3, mm7 | ||
| 177 : | Isibaar | 262 | |
| 178 : | edgomez | 1382 | %if %2 != 0 |
| 179 : | add ebx, ecx | ||
| 180 : | %endif | ||
| 181 : | Isibaar | 262 | |
| 182 : | edgomez | 1382 | paddusw mm0, mm2 ; mm01 = ref1 + ref2 |
| 183 : | paddusw mm1, mm3 | ||
| 184 : | paddusw mm0, [mmx_one] ; mm01 += 1 | ||
| 185 : | paddusw mm1, [mmx_one] | ||
| 186 : | psrlw mm0, 1 ; mm01 >>= 1 | ||
| 187 : | psrlw mm1, 1 | ||
| 188 : | Isibaar | 262 | |
| 189 : | edgomez | 1382 | movq mm2, [eax+%1] |
| 190 : | movq mm3, mm2 | ||
| 191 : | punpcklbw mm2, mm7 ; mm23 = src | ||
| 192 : | punpckhbw mm3, mm7 | ||
| 193 : | Isibaar | 262 | |
| 194 : | edgomez | 1382 | %if %2 != 0 |
| 195 : | add eax, ecx | ||
| 196 : | %endif | ||
| 197 : | Isibaar | 262 | |
| 198 : | edgomez | 1382 | movq mm4, mm0 |
| 199 : | movq mm5, mm1 | ||
| 200 : | psubusw mm0, mm2 | ||
| 201 : | psubusw mm1, mm3 | ||
| 202 : | psubusw mm2, mm4 | ||
| 203 : | psubusw mm3, mm5 | ||
| 204 : | por mm0, mm2 ; mm01 = ABS(mm01 - mm23) | ||
| 205 : | por mm1, mm3 | ||
| 206 : | Isibaar | 262 | |
| 207 : | edgomez | 1382 | paddusw mm6, mm0 ; mm6 += mm01 |
| 208 : | paddusw mm6, mm1 | ||
| 209 : | Isibaar | 262 | |
| 210 : | edgomez | 1382 | %endmacro |
| 211 : | edgomez | 851 | |
| 212 : | edgomez | 1382 | %macro MEAN_16x16_MMX 0 |
| 213 : | movq mm0, [eax] | ||
| 214 : | movq mm2, [eax+8] | ||
| 215 : | lea eax, [eax+ecx] | ||
| 216 : | movq mm1, mm0 | ||
| 217 : | punpcklbw mm0, mm7 | ||
| 218 : | suxen_drol | 1599 | movq mm3, mm2 |
| 219 : | edgomez | 1382 | punpckhbw mm1, mm7 |
| 220 : | paddw mm5, mm0 | ||
| 221 : | suxen_drol | 1599 | punpcklbw mm2, mm7 |
| 222 : | edgomez | 1382 | paddw mm6, mm1 |
| 223 : | suxen_drol | 1599 | punpckhbw mm3, mm7 |
| 224 : | edgomez | 1382 | paddw mm5, mm2 |
| 225 : | paddw mm6, mm3 | ||
| 226 : | %endmacro | ||
| 227 : | edgomez | 851 | |
| 228 : | edgomez | 1382 | %macro ABS_16x16_MMX 0 |
| 229 : | movq mm0, [eax] | ||
| 230 : | movq mm2, [eax+8] | ||
| 231 : | lea eax, [eax+ecx] | ||
| 232 : | movq mm1, mm0 | ||
| 233 : | movq mm3, mm2 | ||
| 234 : | punpcklbw mm0, mm7 | ||
| 235 : | punpcklbw mm2, mm7 | ||
| 236 : | punpckhbw mm1, mm7 | ||
| 237 : | punpckhbw mm3, mm7 | ||
| 238 : | movq mm4, mm6 | ||
| 239 : | psubusw mm4, mm0 | ||
| 240 : | edgomez | 851 | |
| 241 : | edgomez | 1382 | psubusw mm0, mm6 |
| 242 : | por mm0, mm4 | ||
| 243 : | movq mm4, mm6 | ||
| 244 : | psubusw mm4, mm1 | ||
| 245 : | psubusw mm1, mm6 | ||
| 246 : | por mm1, mm4 | ||
| 247 : | edgomez | 851 | |
| 248 : | edgomez | 1382 | movq mm4, mm6 |
| 249 : | psubusw mm4, mm2 | ||
| 250 : | psubusw mm2, mm6 | ||
| 251 : | por mm2, mm4 | ||
| 252 : | movq mm4, mm6 | ||
| 253 : | psubusw mm4, mm3 | ||
| 254 : | psubusw mm3, mm6 | ||
| 255 : | por mm3, mm4 | ||
| 256 : | edgomez | 851 | |
| 257 : | edgomez | 1382 | paddw mm0, mm1 |
| 258 : | paddw mm2, mm3 | ||
| 259 : | paddw mm5, mm0 | ||
| 260 : | paddw mm5, mm2 | ||
| 261 : | %endmacro | ||
| 262 : | edgomez | 851 | |
| 263 : | edgomez | 1382 | ;============================================================================= |
| 264 : | ; Code | ||
| 265 : | ;============================================================================= | ||
| 266 : | edgomez | 851 | |
| 267 : | edgomez | 1382 | SECTION .text |
| 268 : | edgomez | 851 | |
| 269 : | edgomez | 1382 | cglobal sad16_mmx |
| 270 : | cglobal sad16v_mmx | ||
| 271 : | cglobal sad8_mmx | ||
| 272 : | cglobal sad16bi_mmx | ||
| 273 : | cglobal sad8bi_mmx | ||
| 274 : | cglobal dev16_mmx | ||
| 275 : | cglobal sse8_16bit_mmx | ||
| 276 : | edgomez | 1424 | cglobal sse8_8bit_mmx |
| 277 : | edgomez | 1382 | |
| 278 : | ;----------------------------------------------------------------------------- | ||
| 279 : | ; | ||
| 280 : | ; uint32_t sad16_mmx(const uint8_t * const cur, | ||
| 281 : | ; const uint8_t * const ref, | ||
| 282 : | ; const uint32_t stride, | ||
| 283 : | ; const uint32_t best_sad); | ||
| 284 : | ; | ||
| 285 : | ; (early termination ignore; slows this down) | ||
| 286 : | ; | ||
| 287 : | ;----------------------------------------------------------------------------- | ||
| 288 : | edgomez | 851 | |
| 289 : | edgomez | 1382 | ALIGN 16 |
| 290 : | sad16_mmx: | ||
| 291 : | edgomez | 851 | |
| 292 : | edgomez | 1382 | mov eax, [esp+ 4] ; Src1 |
| 293 : | mov edx, [esp+ 8] ; Src2 | ||
| 294 : | mov ecx, [esp+12] ; Stride | ||
| 295 : | edgomez | 851 | |
| 296 : | edgomez | 1382 | pxor mm6, mm6 ; accum |
| 297 : | pxor mm7, mm7 ; zero | ||
| 298 : | edgomez | 851 | |
| 299 : | edgomez | 1382 | SAD_16x16_MMX |
| 300 : | SAD_16x16_MMX | ||
| 301 : | SAD_16x16_MMX | ||
| 302 : | SAD_16x16_MMX | ||
| 303 : | SAD_16x16_MMX | ||
| 304 : | SAD_16x16_MMX | ||
| 305 : | SAD_16x16_MMX | ||
| 306 : | SAD_16x16_MMX | ||
| 307 : | edgomez | 851 | |
| 308 : | edgomez | 1382 | SAD_16x16_MMX |
| 309 : | SAD_16x16_MMX | ||
| 310 : | SAD_16x16_MMX | ||
| 311 : | SAD_16x16_MMX | ||
| 312 : | SAD_16x16_MMX | ||
| 313 : | SAD_16x16_MMX | ||
| 314 : | SAD_16x16_MMX | ||
| 315 : | SAD_16x16_MMX | ||
| 316 : | edgomez | 851 | |
| 317 : | edgomez | 1382 | pmaddwd mm6, [mmx_one] ; collapse |
| 318 : | movq mm7, mm6 | ||
| 319 : | psrlq mm7, 32 | ||
| 320 : | paddd mm6, mm7 | ||
| 321 : | edgomez | 851 | |
| 322 : | edgomez | 1382 | movd eax, mm6 |
| 323 : | edgomez | 851 | |
| 324 : | edgomez | 1382 | ret |
| 325 : | edgomez | 1540 | .endfunc |
| 326 : | edgomez | 851 | |
| 327 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
| 328 : | ; | ||
| 329 : | ; uint32_t sad8_mmx(const uint8_t * const cur, | ||
| 330 : | ; const uint8_t * const ref, | ||
| 331 : | ; const uint32_t stride); | ||
| 332 : | ; | ||
| 333 : | ;----------------------------------------------------------------------------- | ||
| 334 : | edgomez | 851 | |
| 335 : | edgomez | 1382 | ALIGN 16 |
| 336 : | sad8_mmx: | ||
| 337 : | edgomez | 851 | |
| 338 : | edgomez | 1382 | mov eax, [esp+ 4] ; Src1 |
| 339 : | mov edx, [esp+ 8] ; Src2 | ||
| 340 : | mov ecx, [esp+12] ; Stride | ||
| 341 : | edgomez | 851 | |
| 342 : | edgomez | 1382 | pxor mm6, mm6 ; accum |
| 343 : | pxor mm7, mm7 ; zero | ||
| 344 : | edgomez | 851 | |
| 345 : | edgomez | 1382 | SAD_8x8_MMX |
| 346 : | SAD_8x8_MMX | ||
| 347 : | SAD_8x8_MMX | ||
| 348 : | SAD_8x8_MMX | ||
| 349 : | edgomez | 851 | |
| 350 : | edgomez | 1382 | pmaddwd mm6, [mmx_one] ; collapse |
| 351 : | movq mm7, mm6 | ||
| 352 : | psrlq mm7, 32 | ||
| 353 : | paddd mm6, mm7 | ||
| 354 : | edgomez | 851 | |
| 355 : | edgomez | 1382 | movd eax, mm6 |
| 356 : | edgomez | 851 | |
| 357 : | edgomez | 1382 | ret |
| 358 : | edgomez | 1540 | .endfunc |
| 359 : | edgomez | 851 | |
| 360 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
| 361 : | ; | ||
| 362 : | ; uint32_t sad16v_mmx(const uint8_t * const cur, | ||
| 363 : | ; const uint8_t * const ref, | ||
| 364 : | ; const uint32_t stride, | ||
| 365 : | ; int32_t *sad); | ||
| 366 : | ; | ||
| 367 : | ;----------------------------------------------------------------------------- | ||
| 368 : | edgomez | 851 | |
| 369 : | edgomez | 1382 | ALIGN 16 |
| 370 : | sad16v_mmx: | ||
| 371 : | edgomez | 851 | |
| 372 : | edgomez | 1382 | push ebx |
| 373 : | push edi | ||
| 374 : | edgomez | 851 | |
| 375 : | edgomez | 1382 | mov eax, [esp + 8 + 4] ; Src1 |
| 376 : | mov edx, [esp + 8 + 8] ; Src2 | ||
| 377 : | mov ecx, [esp + 8 + 12] ; Stride | ||
| 378 : | mov ebx, [esp + 8 + 16] ; sad ptr | ||
| 379 : | edgomez | 851 | |
| 380 : | edgomez | 1382 | pxor mm5, mm5 ; accum |
| 381 : | pxor mm6, mm6 ; accum | ||
| 382 : | pxor mm7, mm7 ; zero | ||
| 383 : | suxen_drol | 329 | |
| 384 : | edgomez | 1382 | SADV_16x16_MMX |
| 385 : | SADV_16x16_MMX | ||
| 386 : | SADV_16x16_MMX | ||
| 387 : | SADV_16x16_MMX | ||
| 388 : | SADV_16x16_MMX | ||
| 389 : | SADV_16x16_MMX | ||
| 390 : | SADV_16x16_MMX | ||
| 391 : | SADV_16x16_MMX | ||
| 392 : | suxen_drol | 329 | |
| 393 : | edgomez | 1382 | pmaddwd mm5, [mmx_one] ; collapse |
| 394 : | pmaddwd mm6, [mmx_one] ; collapse | ||
| 395 : | suxen_drol | 329 | |
| 396 : | edgomez | 1382 | movq mm2, mm5 |
| 397 : | movq mm3, mm6 | ||
| 398 : | suxen_drol | 329 | |
| 399 : | edgomez | 1382 | psrlq mm2, 32 |
| 400 : | psrlq mm3, 32 | ||
| 401 : | suxen_drol | 329 | |
| 402 : | edgomez | 1382 | paddd mm5, mm2 |
| 403 : | paddd mm6, mm3 | ||
| 404 : | suxen_drol | 329 | |
| 405 : | edgomez | 1382 | movd [ebx], mm5 |
| 406 : | movd [ebx + 4], mm6 | ||
| 407 : | suxen_drol | 329 | |
| 408 : | edgomez | 1382 | paddd mm5, mm6 |
| 409 : | suxen_drol | 329 | |
| 410 : | edgomez | 1382 | movd edi, mm5 |
| 411 : | suxen_drol | 329 | |
| 412 : | edgomez | 1382 | pxor mm5, mm5 |
| 413 : | pxor mm6, mm6 | ||
| 414 : | suxen_drol | 329 | |
| 415 : | edgomez | 1382 | SADV_16x16_MMX |
| 416 : | SADV_16x16_MMX | ||
| 417 : | SADV_16x16_MMX | ||
| 418 : | SADV_16x16_MMX | ||
| 419 : | SADV_16x16_MMX | ||
| 420 : | SADV_16x16_MMX | ||
| 421 : | SADV_16x16_MMX | ||
| 422 : | SADV_16x16_MMX | ||
| 423 : | suxen_drol | 329 | |
| 424 : | edgomez | 1382 | pmaddwd mm5, [mmx_one] ; collapse |
| 425 : | pmaddwd mm6, [mmx_one] ; collapse | ||
| 426 : | suxen_drol | 329 | |
| 427 : | edgomez | 1382 | movq mm2, mm5 |
| 428 : | movq mm3, mm6 | ||
| 429 : | suxen_drol | 329 | |
| 430 : | edgomez | 1382 | psrlq mm2, 32 |
| 431 : | psrlq mm3, 32 | ||
| 432 : | suxen_drol | 329 | |
| 433 : | edgomez | 1382 | paddd mm5, mm2 |
| 434 : | paddd mm6, mm3 | ||
| 435 : | suxen_drol | 329 | |
| 436 : | edgomez | 1382 | movd [ebx + 8], mm5 |
| 437 : | movd [ebx + 12], mm6 | ||
| 438 : | suxen_drol | 329 | |
| 439 : | edgomez | 1382 | paddd mm5, mm6 |
| 440 : | suxen_drol | 329 | |
| 441 : | edgomez | 1382 | movd eax, mm5 |
| 442 : | suxen_drol | 329 | |
| 443 : | edgomez | 1382 | add eax, edi |
| 444 : | suxen_drol | 329 | |
| 445 : | edgomez | 1382 | pop edi |
| 446 : | pop ebx | ||
| 447 : | suxen_drol | 329 | |
| 448 : | edgomez | 1382 | ret |
| 449 : | edgomez | 1540 | .endfunc |
| 450 : | suxen_drol | 329 | |
| 451 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
| 452 : | ; | ||
| 453 : | ; uint32_t sad16bi_mmx(const uint8_t * const cur, | ||
| 454 : | ; const uint8_t * const ref1, | ||
| 455 : | ; const uint8_t * const ref2, | ||
| 456 : | ; const uint32_t stride); | ||
| 457 : | ; | ||
| 458 : | ;----------------------------------------------------------------------------- | ||
| 459 : | suxen_drol | 329 | |
| 460 : | edgomez | 1382 | ALIGN 16 |
| 461 : | sad16bi_mmx: | ||
| 462 : | push ebx | ||
| 463 : | mov eax, [esp+4+ 4] ; Src | ||
| 464 : | mov edx, [esp+4+ 8] ; Ref1 | ||
| 465 : | mov ebx, [esp+4+12] ; Ref2 | ||
| 466 : | mov ecx, [esp+4+16] ; Stride | ||
| 467 : | suxen_drol | 329 | |
| 468 : | edgomez | 1382 | pxor mm6, mm6 ; accum2 |
| 469 : | pxor mm7, mm7 | ||
| 470 : | .Loop | ||
| 471 : | SADBI_16x16_MMX 0, 0 | ||
| 472 : | SADBI_16x16_MMX 8, 1 | ||
| 473 : | SADBI_16x16_MMX 0, 0 | ||
| 474 : | SADBI_16x16_MMX 8, 1 | ||
| 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 : | |||
| 488 : | SADBI_16x16_MMX 0, 0 | ||
| 489 : | SADBI_16x16_MMX 8, 1 | ||
| 490 : | SADBI_16x16_MMX 0, 0 | ||
| 491 : | SADBI_16x16_MMX 8, 1 | ||
| 492 : | SADBI_16x16_MMX 0, 0 | ||
| 493 : | SADBI_16x16_MMX 8, 1 | ||
| 494 : | SADBI_16x16_MMX 0, 0 | ||
| 495 : | SADBI_16x16_MMX 8, 1 | ||
| 496 : | SADBI_16x16_MMX 0, 0 | ||
| 497 : | SADBI_16x16_MMX 8, 1 | ||
| 498 : | SADBI_16x16_MMX 0, 0 | ||
| 499 : | SADBI_16x16_MMX 8, 1 | ||
| 500 : | SADBI_16x16_MMX 0, 0 | ||
| 501 : | SADBI_16x16_MMX 8, 1 | ||
| 502 : | SADBI_16x16_MMX 0, 0 | ||
| 503 : | SADBI_16x16_MMX 8, 1 | ||
| 504 : | |||
| 505 : | pmaddwd mm6, [mmx_one] ; collapse | ||
| 506 : | movq mm7, mm6 | ||
| 507 : | psrlq mm7, 32 | ||
| 508 : | paddd mm6, mm7 | ||
| 509 : | |||
| 510 : | movd eax, mm6 | ||
| 511 : | pop ebx | ||
| 512 : | |||
| 513 : | ret | ||
| 514 : | edgomez | 1540 | .endfunc |
| 515 : | edgomez | 1382 | |
| 516 : | ;----------------------------------------------------------------------------- | ||
| 517 : | suxen_drol | 329 | ; |
| 518 : | edgomez | 1382 | ; uint32_t sad8bi_mmx(const uint8_t * const cur, |
| 519 : | ; const uint8_t * const ref1, | ||
| 520 : | ; const uint8_t * const ref2, | ||
| 521 : | ; const uint32_t stride); | ||
| 522 : | suxen_drol | 329 | ; |
| 523 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
| 524 : | suxen_drol | 329 | |
| 525 : | edgomez | 1382 | ALIGN 16 |
| 526 : | sad8bi_mmx: | ||
| 527 : | push ebx | ||
| 528 : | mov eax, [esp+4+ 4] ; Src | ||
| 529 : | mov edx, [esp+4+ 8] ; Ref1 | ||
| 530 : | mov ebx, [esp+4+12] ; Ref2 | ||
| 531 : | mov ecx, [esp+4+16] ; Stride | ||
| 532 : | Isibaar | 262 | |
| 533 : | edgomez | 1382 | pxor mm6, mm6 ; accum2 |
| 534 : | pxor mm7, mm7 | ||
| 535 : | .Loop | ||
| 536 : | SADBI_16x16_MMX 0, 1 | ||
| 537 : | SADBI_16x16_MMX 0, 1 | ||
| 538 : | SADBI_16x16_MMX 0, 1 | ||
| 539 : | SADBI_16x16_MMX 0, 1 | ||
| 540 : | SADBI_16x16_MMX 0, 1 | ||
| 541 : | SADBI_16x16_MMX 0, 1 | ||
| 542 : | SADBI_16x16_MMX 0, 1 | ||
| 543 : | SADBI_16x16_MMX 0, 1 | ||
| 544 : | Isibaar | 262 | |
| 545 : | edgomez | 1382 | pmaddwd mm6, [mmx_one] ; collapse |
| 546 : | movq mm7, mm6 | ||
| 547 : | psrlq mm7, 32 | ||
| 548 : | paddd mm6, mm7 | ||
| 549 : | Isibaar | 262 | |
| 550 : | edgomez | 1382 | movd eax, mm6 |
| 551 : | pop ebx | ||
| 552 : | ret | ||
| 553 : | edgomez | 1540 | .endfunc |
| 554 : | Isibaar | 262 | |
| 555 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
| 556 : | ; | ||
| 557 : | ; uint32_t dev16_mmx(const uint8_t * const cur, | ||
| 558 : | ; const uint32_t stride); | ||
| 559 : | ; | ||
| 560 : | ;----------------------------------------------------------------------------- | ||
| 561 : | |||
| 562 : | ALIGN 16 | ||
| 563 : | Isibaar | 262 | dev16_mmx: |
| 564 : | edgomez | 1382 | mov eax, [esp+ 4] ; Src |
| 565 : | mov ecx, [esp+ 8] ; Stride | ||
| 566 : | Isibaar | 262 | |
| 567 : | edgomez | 1382 | pxor mm7, mm7 ; zero |
| 568 : | pxor mm5, mm5 ; accum1 | ||
| 569 : | pxor mm6, mm6 ; accum2 | ||
| 570 : | Isibaar | 262 | |
| 571 : | edgomez | 1382 | MEAN_16x16_MMX |
| 572 : | 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 : | Isibaar | 262 | |
| 580 : | edgomez | 1382 | MEAN_16x16_MMX |
| 581 : | MEAN_16x16_MMX | ||
| 582 : | MEAN_16x16_MMX | ||
| 583 : | MEAN_16x16_MMX | ||
| 584 : | MEAN_16x16_MMX | ||
| 585 : | MEAN_16x16_MMX | ||
| 586 : | MEAN_16x16_MMX | ||
| 587 : | MEAN_16x16_MMX | ||
| 588 : | Isibaar | 262 | |
| 589 : | edgomez | 1382 | paddusw mm6, mm5 |
| 590 : | pmaddwd mm6, [mmx_one] ; collapse | ||
| 591 : | movq mm5, mm6 | ||
| 592 : | psrlq mm5, 32 | ||
| 593 : | paddd mm6, mm5 | ||
| 594 : | Isibaar | 262 | |
| 595 : | edgomez | 1382 | psllq mm6, 32 ; blank upper dword |
| 596 : | psrlq mm6, 32 + 8 ; /= (16*16) | ||
| 597 : | Isibaar | 262 | |
| 598 : | edgomez | 1382 | punpckldq mm6, mm6 |
| 599 : | packssdw mm6, mm6 | ||
| 600 : | |||
| 601 : | Isibaar | 262 | ; mm6 contains the mean |
| 602 : | ; mm5 is the new accum | ||
| 603 : | |||
| 604 : | edgomez | 1382 | pxor mm5, mm5 |
| 605 : | mov eax, [esp+ 4] ; Src | ||
| 606 : | Isibaar | 262 | |
| 607 : | edgomez | 1382 | ABS_16x16_MMX |
| 608 : | 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 : | Isibaar | 262 | |
| 616 : | edgomez | 1382 | ABS_16x16_MMX |
| 617 : | ABS_16x16_MMX | ||
| 618 : | ABS_16x16_MMX | ||
| 619 : | ABS_16x16_MMX | ||
| 620 : | ABS_16x16_MMX | ||
| 621 : | ABS_16x16_MMX | ||
| 622 : | ABS_16x16_MMX | ||
| 623 : | ABS_16x16_MMX | ||
| 624 : | Isibaar | 262 | |
| 625 : | edgomez | 1382 | pmaddwd mm5, [mmx_one] ; collapse |
| 626 : | movq mm6, mm5 | ||
| 627 : | psrlq mm6, 32 | ||
| 628 : | paddd mm6, mm5 | ||
| 629 : | Isibaar | 262 | |
| 630 : | edgomez | 1382 | movd eax, mm6 |
| 631 : | |||
| 632 : | ret | ||
| 633 : | edgomez | 1540 | .endfunc |
| 634 : | edgomez | 1382 | |
| 635 : | ;----------------------------------------------------------------------------- | ||
| 636 : | ; | ||
| 637 : | ; uint32_t sse8_16bit_mmx(const int16_t *b1, | ||
| 638 : | ; const int16_t *b2, | ||
| 639 : | ; const uint32_t stride); | ||
| 640 : | ; | ||
| 641 : | ;----------------------------------------------------------------------------- | ||
| 642 : | |||
| 643 : | edgomez | 1424 | %macro ROW_SSE_16bit_MMX 2 |
| 644 : | edgomez | 1382 | movq mm0, [%1] |
| 645 : | movq mm1, [%1+8] | ||
| 646 : | psubw mm0, [%2] | ||
| 647 : | psubw mm1, [%2+8] | ||
| 648 : | pmaddwd mm0, mm0 | ||
| 649 : | pmaddwd mm1, mm1 | ||
| 650 : | paddd mm2, mm0 | ||
| 651 : | paddd mm2, mm1 | ||
| 652 : | edgomez | 1424 | %endmacro |
| 653 : | edgomez | 1382 | |
| 654 : | sse8_16bit_mmx: | ||
| 655 : | push esi | ||
| 656 : | push edi | ||
| 657 : | |||
| 658 : | ;; Load the function params | ||
| 659 : | mov esi, [esp+8+4] | ||
| 660 : | mov edi, [esp+8+8] | ||
| 661 : | mov edx, [esp+8+12] | ||
| 662 : | |||
| 663 : | ;; Reset the sse accumulator | ||
| 664 : | pxor mm2, mm2 | ||
| 665 : | |||
| 666 : | ;; Let's go | ||
| 667 : | edgomez | 1424 | %rep 8 |
| 668 : | ROW_SSE_16bit_MMX esi, edi | ||
| 669 : | edgomez | 1382 | lea esi, [esi+edx] |
| 670 : | lea edi, [edi+edx] | ||
| 671 : | edgomez | 1424 | %endrep |
| 672 : | edgomez | 1382 | |
| 673 : | ;; Finish adding each dword of the accumulator | ||
| 674 : | movq mm3, mm2 | ||
| 675 : | psrlq mm2, 32 | ||
| 676 : | paddd mm2, mm3 | ||
| 677 : | movd eax, mm2 | ||
| 678 : | |||
| 679 : | ;; All done | ||
| 680 : | pop edi | ||
| 681 : | pop esi | ||
| 682 : | ret | ||
| 683 : | edgomez | 1540 | .endfunc |
| 684 : | edgomez | 1424 | |
| 685 : | ;----------------------------------------------------------------------------- | ||
| 686 : | ; | ||
| 687 : | ; uint32_t sse8_8bit_mmx(const int8_t *b1, | ||
| 688 : | ; const int8_t *b2, | ||
| 689 : | ; const uint32_t stride); | ||
| 690 : | ; | ||
| 691 : | ;----------------------------------------------------------------------------- | ||
| 692 : | |||
| 693 : | %macro ROW_SSE_8bit_MMX 2 | ||
| 694 : | movq mm0, [%1] ; load a row | ||
| 695 : | movq mm2, [%2] ; load a row | ||
| 696 : | |||
| 697 : | movq mm1, mm0 ; copy row | ||
| 698 : | movq mm3, mm2 ; copy row | ||
| 699 : | |||
| 700 : | punpcklbw mm0, mm7 ; turn the 4low elements into 16bit | ||
| 701 : | punpckhbw mm1, mm7 ; turn the 4high elements into 16bit | ||
| 702 : | |||
| 703 : | punpcklbw mm2, mm7 ; turn the 4low elements into 16bit | ||
| 704 : | punpckhbw mm3, mm7 ; turn the 4high elements into 16bit | ||
| 705 : | |||
| 706 : | psubw mm0, mm2 ; low part of src-dst | ||
| 707 : | psubw mm1, mm3 ; high part of src-dst | ||
| 708 : | |||
| 709 : | pmaddwd mm0, mm0 ; compute the square sum | ||
| 710 : | pmaddwd mm1, mm1 ; compute the square sum | ||
| 711 : | |||
| 712 : | paddd mm6, mm0 ; add to the accumulator | ||
| 713 : | paddd mm6, mm1 ; add to the accumulator | ||
| 714 : | %endmacro | ||
| 715 : | |||
| 716 : | sse8_8bit_mmx: | ||
| 717 : | push esi | ||
| 718 : | push edi | ||
| 719 : | |||
| 720 : | ;; Load the function params | ||
| 721 : | mov esi, [esp+8+4] | ||
| 722 : | mov edi, [esp+8+8] | ||
| 723 : | mov edx, [esp+8+12] | ||
| 724 : | |||
| 725 : | ;; Reset the sse accumulator | ||
| 726 : | pxor mm6, mm6 | ||
| 727 : | |||
| 728 : | ;; Used to interleave 8bit data with 0x00 values | ||
| 729 : | pxor mm7, mm7 | ||
| 730 : | |||
| 731 : | ;; Let's go | ||
| 732 : | %rep 8 | ||
| 733 : | ROW_SSE_8bit_MMX esi, edi | ||
| 734 : | lea esi, [esi+edx] | ||
| 735 : | lea edi, [edi+edx] | ||
| 736 : | %endrep | ||
| 737 : | |||
| 738 : | ;; Finish adding each dword of the accumulator | ||
| 739 : | movq mm7, mm6 | ||
| 740 : | psrlq mm6, 32 | ||
| 741 : | paddd mm6, mm7 | ||
| 742 : | movd eax, mm6 | ||
| 743 : | |||
| 744 : | ;; All done | ||
| 745 : | pop edi | ||
| 746 : | pop esi | ||
| 747 : | ret | ||
| 748 : | edgomez | 1540 | .endfunc |
| 749 : | |||
| 750 : | Isibaar | 1790 | |
| 751 : | %ifidn __OUTPUT_FORMAT__,elf | ||
| 752 : | section ".note.GNU-stack" noalloc noexec nowrite progbits | ||
| 753 : | %endif | ||
| 754 : |
| No admin address has been configured | ViewVC Help |
| Powered by ViewVC 1.0.4 |