Parent Directory
| 
Revision Log
Revision 1583 - (view) (download)
| 1 : | edgomez | 1382 | ;/**************************************************************************** |
| 2 : | Isibaar | 3 | ; * |
| 3 : | edgomez | 1382 | ; * XVID MPEG-4 VIDEO CODEC |
| 4 : | ; * - 8<->16 bit transfer functions - | ||
| 5 : | Isibaar | 3 | ; * |
| 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 | 3 | ; * |
| 10 : | edgomez | 1382 | ; * This program is free software ; you can redistribute it and/or modify |
| 11 : | ; * it 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 | 3 | ; * |
| 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 | 3 | ; * |
| 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 | 3 | ; * |
| 24 : | syskin | 1583 | ; * $Id: mem_transfer_mmx.asm,v 1.16 2004-12-19 13:16:50 syskin Exp $ |
| 25 : | Isibaar | 3 | ; * |
| 26 : | edgomez | 1382 | ; ***************************************************************************/ |
| 27 : | Isibaar | 3 | |
| 28 : | edgomez | 1382 | BITS 32 |
| 29 : | Isibaar | 3 | |
| 30 : | edgomez | 1382 | %macro cglobal 1 |
| 31 : | Isibaar | 3 | %ifdef PREFIX |
| 32 : | edgomez | 1535 | %ifdef MARK_FUNCS |
| 33 : | edgomez | 1540 | global _%1:function %1.endfunc-%1 |
| 34 : | %define %1 _%1:function %1.endfunc-%1 | ||
| 35 : | edgomez | 1535 | %else |
| 36 : | global _%1 | ||
| 37 : | %define %1 _%1 | ||
| 38 : | %endif | ||
| 39 : | Isibaar | 3 | %else |
| 40 : | edgomez | 1535 | %ifdef MARK_FUNCS |
| 41 : | edgomez | 1540 | global %1:function %1.endfunc-%1 |
| 42 : | edgomez | 1535 | %else |
| 43 : | global %1 | ||
| 44 : | %endif | ||
| 45 : | Isibaar | 3 | %endif |
| 46 : | %endmacro | ||
| 47 : | |||
| 48 : | edgomez | 1382 | ;============================================================================= |
| 49 : | ; Read only data | ||
| 50 : | ;============================================================================= | ||
| 51 : | Isibaar | 3 | |
| 52 : | edgomez | 1382 | %ifdef FORMAT_COFF |
| 53 : | edgomez | 1519 | SECTION .rodata |
| 54 : | edgomez | 1382 | %else |
| 55 : | edgomez | 1519 | SECTION .rodata align=16 |
| 56 : | edgomez | 1382 | %endif |
| 57 : | Isibaar | 3 | |
| 58 : | edgomez | 1382 | ALIGN 16 |
| 59 : | mmx_one: | ||
| 60 : | dw 1, 1, 1, 1 | ||
| 61 : | |||
| 62 : | ;============================================================================= | ||
| 63 : | ; Code | ||
| 64 : | ;============================================================================= | ||
| 65 : | |||
| 66 : | SECTION .text | ||
| 67 : | |||
| 68 : | Isibaar | 226 | cglobal transfer_8to16copy_mmx |
| 69 : | cglobal transfer_16to8copy_mmx | ||
| 70 : | cglobal transfer_8to16sub_mmx | ||
| 71 : | edgomez | 851 | cglobal transfer_8to16subro_mmx |
| 72 : | Isibaar | 226 | cglobal transfer_8to16sub2_mmx |
| 73 : | cglobal transfer_8to16sub2_xmm | ||
| 74 : | syskin | 1583 | cglobal transfer_8to16sub2ro_xmm |
| 75 : | Isibaar | 226 | cglobal transfer_16to8add_mmx |
| 76 : | cglobal transfer8x8_copy_mmx | ||
| 77 : | Isibaar | 3 | |
| 78 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
| 79 : | Isibaar | 3 | ; |
| 80 : | ; void transfer_8to16copy_mmx(int16_t * const dst, | ||
| 81 : | ; const uint8_t * const src, | ||
| 82 : | ; uint32_t stride); | ||
| 83 : | ; | ||
| 84 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
| 85 : | Isibaar | 3 | |
| 86 : | Isibaar | 226 | %macro COPY_8_TO_16 1 |
| 87 : | movq mm0, [eax] | ||
| 88 : | movq mm1, [eax+edx] | ||
| 89 : | movq mm2, mm0 | ||
| 90 : | movq mm3, mm1 | ||
| 91 : | punpcklbw mm0, mm7 | ||
| 92 : | movq [ecx+%1*32], mm0 | ||
| 93 : | punpcklbw mm1, mm7 | ||
| 94 : | movq [ecx+%1*32+16], mm1 | ||
| 95 : | punpckhbw mm2, mm7 | ||
| 96 : | punpckhbw mm3, mm7 | ||
| 97 : | edgomez | 1382 | lea eax, [eax+2*edx] |
| 98 : | Isibaar | 226 | movq [ecx+%1*32+8], mm2 |
| 99 : | movq [ecx+%1*32+24], mm3 | ||
| 100 : | %endmacro | ||
| 101 : | |||
| 102 : | edgomez | 1382 | ALIGN 16 |
| 103 : | Isibaar | 226 | transfer_8to16copy_mmx: |
| 104 : | Isibaar | 3 | |
| 105 : | Isibaar | 226 | mov ecx, [esp+ 4] ; Dst |
| 106 : | mov eax, [esp+ 8] ; Src | ||
| 107 : | mov edx, [esp+12] ; Stride | ||
| 108 : | edgomez | 1382 | pxor mm7, mm7 |
| 109 : | Isibaar | 3 | |
| 110 : | Isibaar | 226 | COPY_8_TO_16 0 |
| 111 : | COPY_8_TO_16 1 | ||
| 112 : | COPY_8_TO_16 2 | ||
| 113 : | COPY_8_TO_16 3 | ||
| 114 : | ret | ||
| 115 : | edgomez | 1540 | .endfunc |
| 116 : | Isibaar | 3 | |
| 117 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
| 118 : | Isibaar | 3 | ; |
| 119 : | ; void transfer_16to8copy_mmx(uint8_t * const dst, | ||
| 120 : | ; const int16_t * const src, | ||
| 121 : | ; uint32_t stride); | ||
| 122 : | ; | ||
| 123 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
| 124 : | Isibaar | 3 | |
| 125 : | Isibaar | 226 | %macro COPY_16_TO_8 1 |
| 126 : | movq mm0, [eax+%1*32] | ||
| 127 : | movq mm1, [eax+%1*32+8] | ||
| 128 : | packuswb mm0, mm1 | ||
| 129 : | movq [ecx], mm0 | ||
| 130 : | movq mm2, [eax+%1*32+16] | ||
| 131 : | movq mm3, [eax+%1*32+24] | ||
| 132 : | packuswb mm2, mm3 | ||
| 133 : | movq [ecx+edx], mm2 | ||
| 134 : | %endmacro | ||
| 135 : | |||
| 136 : | edgomez | 1382 | ALIGN 16 |
| 137 : | Isibaar | 226 | transfer_16to8copy_mmx: |
| 138 : | Isibaar | 3 | |
| 139 : | Isibaar | 226 | mov ecx, [esp+ 4] ; Dst |
| 140 : | mov eax, [esp+ 8] ; Src | ||
| 141 : | mov edx, [esp+12] ; Stride | ||
| 142 : | Isibaar | 3 | |
| 143 : | Isibaar | 226 | COPY_16_TO_8 0 |
| 144 : | lea ecx,[ecx+2*edx] | ||
| 145 : | COPY_16_TO_8 1 | ||
| 146 : | lea ecx,[ecx+2*edx] | ||
| 147 : | COPY_16_TO_8 2 | ||
| 148 : | lea ecx,[ecx+2*edx] | ||
| 149 : | COPY_16_TO_8 3 | ||
| 150 : | ret | ||
| 151 : | edgomez | 1540 | .endfunc |
| 152 : | Isibaar | 3 | |
| 153 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
| 154 : | Isibaar | 3 | ; |
| 155 : | ; void transfer_8to16sub_mmx(int16_t * const dct, | ||
| 156 : | ; uint8_t * const cur, | ||
| 157 : | ; const uint8_t * const ref, | ||
| 158 : | ; const uint32_t stride); | ||
| 159 : | ; | ||
| 160 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
| 161 : | Isibaar | 3 | |
| 162 : | edgomez | 851 | ; when second argument == 1, reference (ebx) block is to current (eax) |
| 163 : | %macro COPY_8_TO_16_SUB 2 | ||
| 164 : | Isibaar | 226 | movq mm0, [eax] ; cur |
| 165 : | movq mm2, [eax+edx] | ||
| 166 : | edgomez | 1425 | movq mm1, mm0 |
| 167 : | movq mm3, mm2 | ||
| 168 : | Isibaar | 3 | |
| 169 : | edgomez | 1425 | punpcklbw mm0, mm7 |
| 170 : | punpcklbw mm2, mm7 | ||
| 171 : | Isibaar | 226 | movq mm4, [ebx] ; ref |
| 172 : | edgomez | 1425 | punpckhbw mm1, mm7 |
| 173 : | punpckhbw mm3, mm7 | ||
| 174 : | Isibaar | 226 | movq mm5, [ebx+edx] ; ref |
| 175 : | Isibaar | 3 | |
| 176 : | edgomez | 1425 | movq mm6, mm4 |
| 177 : | edgomez | 851 | %if %2 == 1 |
| 178 : | Isibaar | 226 | movq [eax], mm4 |
| 179 : | movq [eax+edx], mm5 | ||
| 180 : | edgomez | 851 | %endif |
| 181 : | edgomez | 1425 | punpcklbw mm4, mm7 |
| 182 : | punpckhbw mm6, mm7 | ||
| 183 : | psubsw mm0, mm4 | ||
| 184 : | psubsw mm1, mm6 | ||
| 185 : | movq mm6, mm5 | ||
| 186 : | punpcklbw mm5, mm7 | ||
| 187 : | punpckhbw mm6, mm7 | ||
| 188 : | psubsw mm2, mm5 | ||
| 189 : | edgomez | 1382 | lea eax, [eax+2*edx] |
| 190 : | edgomez | 1425 | psubsw mm3, mm6 |
| 191 : | Isibaar | 226 | lea ebx,[ebx+2*edx] |
| 192 : | Isibaar | 3 | |
| 193 : | Isibaar | 226 | movq [ecx+%1*32+ 0], mm0 ; dst |
| 194 : | edgomez | 851 | movq [ecx+%1*32+ 8], mm1 |
| 195 : | movq [ecx+%1*32+16], mm2 | ||
| 196 : | movq [ecx+%1*32+24], mm3 | ||
| 197 : | Isibaar | 226 | %endmacro |
| 198 : | Isibaar | 3 | |
| 199 : | edgomez | 1382 | ALIGN 16 |
| 200 : | Isibaar | 226 | transfer_8to16sub_mmx: |
| 201 : | mov ecx, [esp + 4] ; Dst | ||
| 202 : | mov eax, [esp + 8] ; Cur | ||
| 203 : | push ebx | ||
| 204 : | mov ebx, [esp+4+12] ; Ref | ||
| 205 : | mov edx, [esp+4+16] ; Stride | ||
| 206 : | pxor mm7, mm7 | ||
| 207 : | Isibaar | 3 | |
| 208 : | edgomez | 851 | COPY_8_TO_16_SUB 0, 1 |
| 209 : | COPY_8_TO_16_SUB 1, 1 | ||
| 210 : | COPY_8_TO_16_SUB 2, 1 | ||
| 211 : | COPY_8_TO_16_SUB 3, 1 | ||
| 212 : | Isibaar | 3 | |
| 213 : | Isibaar | 226 | pop ebx |
| 214 : | ret | ||
| 215 : | edgomez | 1540 | .endfunc |
| 216 : | Isibaar | 3 | |
| 217 : | edgomez | 851 | |
| 218 : | edgomez | 1382 | ALIGN 16 |
| 219 : | edgomez | 851 | transfer_8to16subro_mmx: |
| 220 : | mov ecx, [esp + 4] ; Dst | ||
| 221 : | mov eax, [esp + 8] ; Cur | ||
| 222 : | push ebx | ||
| 223 : | mov ebx, [esp+4+12] ; Ref | ||
| 224 : | mov edx, [esp+4+16] ; Stride | ||
| 225 : | pxor mm7, mm7 | ||
| 226 : | |||
| 227 : | COPY_8_TO_16_SUB 0, 0 | ||
| 228 : | COPY_8_TO_16_SUB 1, 0 | ||
| 229 : | COPY_8_TO_16_SUB 2, 0 | ||
| 230 : | COPY_8_TO_16_SUB 3, 0 | ||
| 231 : | |||
| 232 : | pop ebx | ||
| 233 : | ret | ||
| 234 : | edgomez | 1540 | .endfunc |
| 235 : | edgomez | 851 | |
| 236 : | |||
| 237 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
| 238 : | Isibaar | 226 | ; |
| 239 : | ; void transfer_8to16sub2_mmx(int16_t * const dct, | ||
| 240 : | ; uint8_t * const cur, | ||
| 241 : | ; const uint8_t * ref1, | ||
| 242 : | ; const uint8_t * ref2, | ||
| 243 : | ; const uint32_t stride) | ||
| 244 : | ; | ||
| 245 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
| 246 : | Isibaar | 3 | |
| 247 : | Isibaar | 226 | %macro COPY_8_TO_16_SUB2_MMX 1 |
| 248 : | movq mm0, [eax] ; cur | ||
| 249 : | movq mm2, [eax+edx] | ||
| 250 : | edgomez | 1425 | |
| 251 : | edgomez | 1382 | ; mm4 <- (ref1+ref2+1) / 2 |
| 252 : | Isibaar | 226 | movq mm4, [ebx] ; ref1 |
| 253 : | movq mm1, [esi] ; ref2 | ||
| 254 : | edgomez | 1425 | movq mm6, mm4 |
| 255 : | movq mm3, mm1 | ||
| 256 : | punpcklbw mm4, mm7 | ||
| 257 : | punpcklbw mm1, mm7 | ||
| 258 : | punpckhbw mm6, mm7 | ||
| 259 : | punpckhbw mm3, mm7 | ||
| 260 : | paddusw mm4, mm1 | ||
| 261 : | paddusw mm6, mm3 | ||
| 262 : | paddusw mm4, [mmx_one] | ||
| 263 : | paddusw mm6, [mmx_one] | ||
| 264 : | psrlw mm4, 1 | ||
| 265 : | psrlw mm6, 1 | ||
| 266 : | packuswb mm4, mm6 | ||
| 267 : | movq [eax], mm4 | ||
| 268 : | |||
| 269 : | ; mm5 <- (ref1+ref2+1) / 2 | ||
| 270 : | Isibaar | 226 | movq mm5, [ebx+edx] ; ref1 |
| 271 : | movq mm1, [esi+edx] ; ref2 | ||
| 272 : | edgomez | 1425 | movq mm6, mm5 |
| 273 : | movq mm3, mm1 | ||
| 274 : | punpcklbw mm5, mm7 | ||
| 275 : | punpcklbw mm1, mm7 | ||
| 276 : | punpckhbw mm6, mm7 | ||
| 277 : | punpckhbw mm3, mm7 | ||
| 278 : | paddusw mm5, mm1 | ||
| 279 : | paddusw mm6, mm3 | ||
| 280 : | paddusw mm5, [mmx_one] | ||
| 281 : | paddusw mm6, [mmx_one] | ||
| 282 : | edgomez | 1382 | lea esi, [esi+2*edx] |
| 283 : | edgomez | 1425 | psrlw mm5, 1 |
| 284 : | psrlw mm6, 1 | ||
| 285 : | packuswb mm5, mm6 | ||
| 286 : | movq [eax+edx], mm5 | ||
| 287 : | |||
| 288 : | movq mm1, mm0 | ||
| 289 : | movq mm3, mm2 | ||
| 290 : | punpcklbw mm0, mm7 | ||
| 291 : | punpcklbw mm2, mm7 | ||
| 292 : | punpckhbw mm1, mm7 | ||
| 293 : | punpckhbw mm3, mm7 | ||
| 294 : | |||
| 295 : | movq mm6, mm4 | ||
| 296 : | punpcklbw mm4, mm7 | ||
| 297 : | punpckhbw mm6, mm7 | ||
| 298 : | psubsw mm0, mm4 | ||
| 299 : | psubsw mm1, mm6 | ||
| 300 : | movq mm6, mm5 | ||
| 301 : | punpcklbw mm5, mm7 | ||
| 302 : | punpckhbw mm6, mm7 | ||
| 303 : | psubsw mm2, mm5 | ||
| 304 : | edgomez | 1382 | lea eax, [eax+2*edx] |
| 305 : | edgomez | 1425 | psubsw mm3, mm6 |
| 306 : | edgomez | 1382 | lea ebx, [ebx+2*edx] |
| 307 : | edgomez | 1425 | |
| 308 : | Isibaar | 226 | movq [ecx+%1*32+ 0], mm0 ; dst |
| 309 : | edgomez | 1382 | movq [ecx+%1*32+ 8], mm1 |
| 310 : | movq [ecx+%1*32+16], mm2 | ||
| 311 : | movq [ecx+%1*32+24], mm3 | ||
| 312 : | Isibaar | 226 | %endmacro |
| 313 : | Isibaar | 3 | |
| 314 : | edgomez | 1382 | ALIGN 16 |
| 315 : | Isibaar | 226 | transfer_8to16sub2_mmx: |
| 316 : | mov ecx, [esp + 4] ; Dst | ||
| 317 : | mov eax, [esp + 8] ; Cur | ||
| 318 : | push ebx | ||
| 319 : | mov ebx, [esp+4+12] ; Ref1 | ||
| 320 : | push esi | ||
| 321 : | mov esi, [esp+8+16] ; Ref2 | ||
| 322 : | mov edx, [esp+8+20] ; Stride | ||
| 323 : | pxor mm7, mm7 | ||
| 324 : | Isibaar | 3 | |
| 325 : | Isibaar | 226 | COPY_8_TO_16_SUB2_MMX 0 |
| 326 : | COPY_8_TO_16_SUB2_MMX 1 | ||
| 327 : | COPY_8_TO_16_SUB2_MMX 2 | ||
| 328 : | COPY_8_TO_16_SUB2_MMX 3 | ||
| 329 : | Isibaar | 3 | |
| 330 : | Isibaar | 226 | pop esi |
| 331 : | pop ebx | ||
| 332 : | ret | ||
| 333 : | edgomez | 1540 | .endfunc |
| 334 : | Isibaar | 3 | |
| 335 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
| 336 : | edgomez | 215 | ; |
| 337 : | ; void transfer_8to16sub2_xmm(int16_t * const dct, | ||
| 338 : | Isibaar | 226 | ; uint8_t * const cur, |
| 339 : | ; const uint8_t * ref1, | ||
| 340 : | ; const uint8_t * ref2, | ||
| 341 : | ; const uint32_t stride) | ||
| 342 : | edgomez | 215 | ; |
| 343 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
| 344 : | Isibaar | 3 | |
| 345 : | Isibaar | 226 | %macro COPY_8_TO_16_SUB2_SSE 1 |
| 346 : | movq mm0, [eax] ; cur | ||
| 347 : | movq mm2, [eax+edx] | ||
| 348 : | edgomez | 1425 | movq mm1, mm0 |
| 349 : | movq mm3, mm2 | ||
| 350 : | edgomez | 215 | |
| 351 : | edgomez | 1425 | punpcklbw mm0, mm7 |
| 352 : | punpcklbw mm2, mm7 | ||
| 353 : | edgomez | 1382 | movq mm4, [ebx] ; ref1 |
| 354 : | Isibaar | 226 | pavgb mm4, [esi] ; ref2 |
| 355 : | edgomez | 1425 | movq [eax], mm4 |
| 356 : | punpckhbw mm1, mm7 | ||
| 357 : | punpckhbw mm3, mm7 | ||
| 358 : | edgomez | 1382 | movq mm5, [ebx+edx] ; ref |
| 359 : | Isibaar | 226 | pavgb mm5, [esi+edx] ; ref2 |
| 360 : | edgomez | 1425 | movq [eax+edx], mm5 |
| 361 : | edgomez | 215 | |
| 362 : | edgomez | 1425 | movq mm6, mm4 |
| 363 : | punpcklbw mm4, mm7 | ||
| 364 : | punpckhbw mm6, mm7 | ||
| 365 : | psubsw mm0, mm4 | ||
| 366 : | psubsw mm1, mm6 | ||
| 367 : | edgomez | 1382 | lea esi, [esi+2*edx] |
| 368 : | edgomez | 1425 | movq mm6, mm5 |
| 369 : | punpcklbw mm5, mm7 | ||
| 370 : | punpckhbw mm6, mm7 | ||
| 371 : | psubsw mm2, mm5 | ||
| 372 : | edgomez | 1382 | lea eax, [eax+2*edx] |
| 373 : | edgomez | 1425 | psubsw mm3, mm6 |
| 374 : | edgomez | 1382 | lea ebx, [ebx+2*edx] |
| 375 : | edgomez | 1425 | |
| 376 : | Isibaar | 226 | movq [ecx+%1*32+ 0], mm0 ; dst |
| 377 : | edgomez | 1382 | movq [ecx+%1*32+ 8], mm1 |
| 378 : | movq [ecx+%1*32+16], mm2 | ||
| 379 : | movq [ecx+%1*32+24], mm3 | ||
| 380 : | Isibaar | 226 | %endmacro |
| 381 : | edgomez | 215 | |
| 382 : | edgomez | 1382 | ALIGN 16 |
| 383 : | Isibaar | 226 | transfer_8to16sub2_xmm: |
| 384 : | mov ecx, [esp + 4] ; Dst | ||
| 385 : | mov eax, [esp + 8] ; Cur | ||
| 386 : | push ebx | ||
| 387 : | mov ebx, [esp+4+12] ; Ref1 | ||
| 388 : | push esi | ||
| 389 : | mov esi, [esp+8+16] ; Ref2 | ||
| 390 : | mov edx, [esp+8+20] ; Stride | ||
| 391 : | pxor mm7, mm7 | ||
| 392 : | edgomez | 215 | |
| 393 : | Isibaar | 226 | COPY_8_TO_16_SUB2_SSE 0 |
| 394 : | COPY_8_TO_16_SUB2_SSE 1 | ||
| 395 : | COPY_8_TO_16_SUB2_SSE 2 | ||
| 396 : | COPY_8_TO_16_SUB2_SSE 3 | ||
| 397 : | edgomez | 215 | |
| 398 : | Isibaar | 226 | pop esi |
| 399 : | pop ebx | ||
| 400 : | ret | ||
| 401 : | edgomez | 1540 | .endfunc |
| 402 : | edgomez | 215 | |
| 403 : | syskin | 1583 | |
| 404 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
| 405 : | Isibaar | 3 | ; |
| 406 : | syskin | 1583 | ; void transfer_8to16sub2ro_xmm(int16_t * const dct, |
| 407 : | ; const uint8_t * const cur, | ||
| 408 : | ; const uint8_t * ref1, | ||
| 409 : | ; const uint8_t * ref2, | ||
| 410 : | ; const uint32_t stride) | ||
| 411 : | ; | ||
| 412 : | ;----------------------------------------------------------------------------- | ||
| 413 : | |||
| 414 : | %macro COPY_8_TO_16_SUB2RO_SSE 1 | ||
| 415 : | movq mm0, [eax] ; cur | ||
| 416 : | movq mm2, [eax+edx] | ||
| 417 : | movq mm1, mm0 | ||
| 418 : | movq mm3, mm2 | ||
| 419 : | |||
| 420 : | punpcklbw mm0, mm7 | ||
| 421 : | punpcklbw mm2, mm7 | ||
| 422 : | movq mm4, [ebx] ; ref1 | ||
| 423 : | pavgb mm4, [esi] ; ref2 | ||
| 424 : | punpckhbw mm1, mm7 | ||
| 425 : | punpckhbw mm3, mm7 | ||
| 426 : | movq mm5, [ebx+edx] ; ref | ||
| 427 : | pavgb mm5, [esi+edx] ; ref2 | ||
| 428 : | |||
| 429 : | movq mm6, mm4 | ||
| 430 : | punpcklbw mm4, mm7 | ||
| 431 : | punpckhbw mm6, mm7 | ||
| 432 : | psubsw mm0, mm4 | ||
| 433 : | psubsw mm1, mm6 | ||
| 434 : | lea esi, [esi+2*edx] | ||
| 435 : | movq mm6, mm5 | ||
| 436 : | punpcklbw mm5, mm7 | ||
| 437 : | punpckhbw mm6, mm7 | ||
| 438 : | psubsw mm2, mm5 | ||
| 439 : | lea eax, [eax+2*edx] | ||
| 440 : | psubsw mm3, mm6 | ||
| 441 : | lea ebx, [ebx+2*edx] | ||
| 442 : | |||
| 443 : | movq [ecx+%1*32+ 0], mm0 ; dst | ||
| 444 : | movq [ecx+%1*32+ 8], mm1 | ||
| 445 : | movq [ecx+%1*32+16], mm2 | ||
| 446 : | movq [ecx+%1*32+24], mm3 | ||
| 447 : | %endmacro | ||
| 448 : | |||
| 449 : | ALIGN 16 | ||
| 450 : | transfer_8to16sub2ro_xmm: | ||
| 451 : | pxor mm7, mm7 | ||
| 452 : | mov ecx, [esp + 4] ; Dst | ||
| 453 : | mov eax, [esp + 8] ; Cur | ||
| 454 : | push ebx | ||
| 455 : | mov ebx, [esp+4+12] ; Ref1 | ||
| 456 : | push esi | ||
| 457 : | mov esi, [esp+8+16] ; Ref2 | ||
| 458 : | mov edx, [esp+8+20] ; Stride | ||
| 459 : | |||
| 460 : | COPY_8_TO_16_SUB2RO_SSE 0 | ||
| 461 : | COPY_8_TO_16_SUB2RO_SSE 1 | ||
| 462 : | COPY_8_TO_16_SUB2RO_SSE 2 | ||
| 463 : | COPY_8_TO_16_SUB2RO_SSE 3 | ||
| 464 : | |||
| 465 : | pop esi | ||
| 466 : | pop ebx | ||
| 467 : | ret | ||
| 468 : | .endfunc | ||
| 469 : | |||
| 470 : | |||
| 471 : | ;----------------------------------------------------------------------------- | ||
| 472 : | ; | ||
| 473 : | Isibaar | 3 | ; void transfer_16to8add_mmx(uint8_t * const dst, |
| 474 : | ; const int16_t * const src, | ||
| 475 : | ; uint32_t stride); | ||
| 476 : | ; | ||
| 477 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
| 478 : | Isibaar | 3 | |
| 479 : | Isibaar | 226 | %macro COPY_16_TO_8_ADD 1 |
| 480 : | edgomez | 1425 | movq mm0, [ecx] |
| 481 : | movq mm2, [ecx+edx] | ||
| 482 : | movq mm1, mm0 | ||
| 483 : | movq mm3, mm2 | ||
| 484 : | punpcklbw mm0, mm7 | ||
| 485 : | punpcklbw mm2, mm7 | ||
| 486 : | punpckhbw mm1, mm7 | ||
| 487 : | punpckhbw mm3, mm7 | ||
| 488 : | paddsw mm0, [eax+%1*32+ 0] | ||
| 489 : | paddsw mm1, [eax+%1*32+ 8] | ||
| 490 : | paddsw mm2, [eax+%1*32+16] | ||
| 491 : | paddsw mm3, [eax+%1*32+24] | ||
| 492 : | packuswb mm0, mm1 | ||
| 493 : | movq [ecx], mm0 | ||
| 494 : | packuswb mm2, mm3 | ||
| 495 : | movq [ecx+edx], mm2 | ||
| 496 : | Isibaar | 226 | %endmacro |
| 497 : | Isibaar | 3 | |
| 498 : | |||
| 499 : | edgomez | 1382 | ALIGN 16 |
| 500 : | Isibaar | 226 | transfer_16to8add_mmx: |
| 501 : | mov ecx, [esp+ 4] ; Dst | ||
| 502 : | mov eax, [esp+ 8] ; Src | ||
| 503 : | mov edx, [esp+12] ; Stride | ||
| 504 : | edgomez | 1425 | pxor mm7, mm7 |
| 505 : | Isibaar | 3 | |
| 506 : | Isibaar | 226 | COPY_16_TO_8_ADD 0 |
| 507 : | lea ecx,[ecx+2*edx] | ||
| 508 : | COPY_16_TO_8_ADD 1 | ||
| 509 : | lea ecx,[ecx+2*edx] | ||
| 510 : | COPY_16_TO_8_ADD 2 | ||
| 511 : | lea ecx,[ecx+2*edx] | ||
| 512 : | COPY_16_TO_8_ADD 3 | ||
| 513 : | ret | ||
| 514 : | edgomez | 1540 | .endfunc |
| 515 : | Isibaar | 3 | |
| 516 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
| 517 : | Isibaar | 3 | ; |
| 518 : | ; void transfer8x8_copy_mmx(uint8_t * const dst, | ||
| 519 : | ; const uint8_t * const src, | ||
| 520 : | ; const uint32_t stride); | ||
| 521 : | ; | ||
| 522 : | ; | ||
| 523 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
| 524 : | Isibaar | 3 | |
| 525 : | edgomez | 1425 | %macro COPY_8_TO_8 0 |
| 526 : | movq mm0, [eax] | ||
| 527 : | movq mm1, [eax+edx] | ||
| 528 : | movq [ecx], mm0 | ||
| 529 : | lea eax, [eax+2*edx] | ||
| 530 : | movq [ecx+edx], mm1 | ||
| 531 : | %endmacro | ||
| 532 : | |||
| 533 : | edgomez | 1382 | ALIGN 16 |
| 534 : | Isibaar | 226 | transfer8x8_copy_mmx: |
| 535 : | edgomez | 1425 | mov ecx, [esp+ 4] ; Dst |
| 536 : | Isibaar | 226 | mov eax, [esp+ 8] ; Src |
| 537 : | mov edx, [esp+12] ; Stride | ||
| 538 : | Isibaar | 3 | |
| 539 : | edgomez | 1425 | COPY_8_TO_8 |
| 540 : | lea ecx,[ecx+2*edx] | ||
| 541 : | COPY_8_TO_8 | ||
| 542 : | lea ecx,[ecx+2*edx] | ||
| 543 : | COPY_8_TO_8 | ||
| 544 : | lea ecx,[ecx+2*edx] | ||
| 545 : | COPY_8_TO_8 | ||
| 546 : | ret | ||
| 547 : | edgomez | 1540 | .endfunc |
| 548 : |
| No admin address has been configured | ViewVC Help |
| Powered by ViewVC 1.0.4 |