Parent Directory
| 
Revision Log
Revision 1.10 - (view) (download)
| 1 : | edgomez | 1.3 | ;/**************************************************************************** |
| 2 : | edgomez | 1.2 | ; * |
| 3 : | edgomez | 1.3 | ; * XVID MPEG-4 VIDEO CODEC |
| 4 : | ; * - 8<->16 bit transfer functions - | ||
| 5 : | edgomez | 1.2 | ; * |
| 6 : | edgomez | 1.3 | ; * Copyright (C) 2002 Jaan Kalda |
| 7 : | edgomez | 1.2 | ; * |
| 8 : | edgomez | 1.3 | ; * This program is free software ; you can redistribute it and/or modify |
| 9 : | ; * it under the terms of the GNU General Public License as published by | ||
| 10 : | ; * the Free Software Foundation ; either version 2 of the License, or | ||
| 11 : | ; * (at your option) any later version. | ||
| 12 : | edgomez | 1.2 | ; * |
| 13 : | edgomez | 1.3 | ; * This program is distributed in the hope that it will be useful, |
| 14 : | ; * but WITHOUT ANY WARRANTY ; without even the implied warranty of | ||
| 15 : | ; * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | ||
| 16 : | ; * GNU General Public License for more details. | ||
| 17 : | edgomez | 1.2 | ; * |
| 18 : | edgomez | 1.3 | ; * You should have received a copy of the GNU General Public License |
| 19 : | ; * along with this program ; if not, write to the Free Software | ||
| 20 : | ; * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA | ||
| 21 : | edgomez | 1.2 | ; * |
| 22 : | Isibaar | 1.10 | ; * $Id: mem_transfer_3dne.asm,v 1.9 2008/08/19 09:06:48 Isibaar Exp $ |
| 23 : | edgomez | 1.3 | ; * |
| 24 : | ; ***************************************************************************/ | ||
| 25 : | |||
| 26 : | ; these 3dne functions are compatible with iSSE, but are optimized specifically | ||
| 27 : | ; for K7 pipelines | ||
| 28 : | |||
| 29 : | BITS 32 | ||
| 30 : | edgomez | 1.2 | |
| 31 : | edgomez | 1.3 | %macro cglobal 1 |
| 32 : | %ifdef PREFIX | ||
| 33 : | edgomez | 1.5 | %ifdef MARK_FUNCS |
| 34 : | edgomez | 1.6 | global _%1:function %1.endfunc-%1 |
| 35 : | %define %1 _%1:function %1.endfunc-%1 | ||
| 36 : | Isibaar | 1.10 | %define ENDFUNC .endfunc |
| 37 : | edgomez | 1.5 | %else |
| 38 : | global _%1 | ||
| 39 : | %define %1 _%1 | ||
| 40 : | Isibaar | 1.10 | %define ENDFUNC |
| 41 : | edgomez | 1.5 | %endif |
| 42 : | edgomez | 1.3 | %else |
| 43 : | edgomez | 1.5 | %ifdef MARK_FUNCS |
| 44 : | edgomez | 1.6 | global %1:function %1.endfunc-%1 |
| 45 : | Isibaar | 1.10 | %define ENDFUNC .endfunc |
| 46 : | edgomez | 1.5 | %else |
| 47 : | global %1 | ||
| 48 : | Isibaar | 1.10 | %define ENDFUNC |
| 49 : | edgomez | 1.5 | %endif |
| 50 : | edgomez | 1.3 | %endif |
| 51 : | %endmacro | ||
| 52 : | edgomez | 1.2 | |
| 53 : | edgomez | 1.3 | ;============================================================================= |
| 54 : | ; Read only data | ||
| 55 : | ;============================================================================= | ||
| 56 : | edgomez | 1.2 | |
| 57 : | %ifdef FORMAT_COFF | ||
| 58 : | edgomez | 1.4 | SECTION .rodata |
| 59 : | edgomez | 1.2 | %else |
| 60 : | edgomez | 1.4 | SECTION .rodata align=16 |
| 61 : | edgomez | 1.2 | %endif |
| 62 : | |||
| 63 : | edgomez | 1.3 | ALIGN 8 |
| 64 : | edgomez | 1.2 | mm_zero: |
| 65 : | edgomez | 1.3 | dd 0,0 |
| 66 : | ;============================================================================= | ||
| 67 : | ; Macros | ||
| 68 : | ;============================================================================= | ||
| 69 : | edgomez | 1.2 | |
| 70 : | %macro nop4 0 | ||
| 71 : | edgomez | 1.3 | db 08Dh, 074h, 026h, 0 |
| 72 : | edgomez | 1.2 | %endmacro |
| 73 : | |||
| 74 : | edgomez | 1.3 | ;============================================================================= |
| 75 : | ; Code | ||
| 76 : | ;============================================================================= | ||
| 77 : | |||
| 78 : | SECTION .text | ||
| 79 : | edgomez | 1.2 | |
| 80 : | cglobal transfer_8to16copy_3dne | ||
| 81 : | cglobal transfer_16to8copy_3dne | ||
| 82 : | cglobal transfer_8to16sub_3dne | ||
| 83 : | cglobal transfer_8to16subro_3dne | ||
| 84 : | cglobal transfer_8to16sub2_3dne | ||
| 85 : | cglobal transfer_16to8add_3dne | ||
| 86 : | cglobal transfer8x8_copy_3dne | ||
| 87 : | suxen_drol | 1.7 | cglobal transfer8x4_copy_3dne |
| 88 : | edgomez | 1.2 | |
| 89 : | edgomez | 1.3 | ;----------------------------------------------------------------------------- |
| 90 : | edgomez | 1.2 | ; |
| 91 : | ; void transfer_8to16copy_3dne(int16_t * const dst, | ||
| 92 : | ; const uint8_t * const src, | ||
| 93 : | ; uint32_t stride); | ||
| 94 : | ; | ||
| 95 : | edgomez | 1.3 | ;----------------------------------------------------------------------------- |
| 96 : | edgomez | 1.2 | |
| 97 : | edgomez | 1.3 | ALIGN 16 |
| 98 : | edgomez | 1.2 | transfer_8to16copy_3dne: |
| 99 : | |||
| 100 : | mov eax, [esp+ 8] ; Src | ||
| 101 : | mov edx, [esp+12] ; Stride | ||
| 102 : | mov ecx, [esp+ 4] ; Dst | ||
| 103 : | punpcklbw mm0, [byte eax] | ||
| 104 : | punpcklbw mm1, [eax+4] | ||
| 105 : | edgomez | 1.3 | movq mm2, [eax+edx] |
| 106 : | movq mm3, [eax+edx] | ||
| 107 : | pxor mm7, mm7 | ||
| 108 : | lea eax, [eax+2*edx] | ||
| 109 : | punpcklbw mm2, mm7 | ||
| 110 : | punpckhbw mm3, mm7 | ||
| 111 : | psrlw mm0, 8 | ||
| 112 : | psrlw mm1, 8 | ||
| 113 : | edgomez | 1.2 | punpcklbw mm4, [eax] |
| 114 : | punpcklbw mm5, [eax+edx+4] | ||
| 115 : | movq [byte ecx+0*64], mm0 | ||
| 116 : | movq [ecx+0*64+8], mm1 | ||
| 117 : | punpcklbw mm6, [eax+edx] | ||
| 118 : | punpcklbw mm7, [eax+4] | ||
| 119 : | edgomez | 1.3 | lea eax, [byte eax+2*edx] |
| 120 : | psrlw mm4, 8 | ||
| 121 : | psrlw mm5, 8 | ||
| 122 : | edgomez | 1.2 | punpcklbw mm0, [eax] |
| 123 : | punpcklbw mm1, [eax+edx+4] | ||
| 124 : | movq [ecx+0*64+16], mm2 | ||
| 125 : | movq [ecx+0*64+24], mm3 | ||
| 126 : | edgomez | 1.3 | psrlw mm6, 8 |
| 127 : | psrlw mm7, 8 | ||
| 128 : | edgomez | 1.2 | punpcklbw mm2, [eax+edx] |
| 129 : | punpcklbw mm3, [eax+4] | ||
| 130 : | edgomez | 1.3 | lea eax, [byte eax+2*edx] |
| 131 : | edgomez | 1.2 | movq [byte ecx+0*64+32], mm4 |
| 132 : | movq [ecx+0*64+56], mm5 | ||
| 133 : | edgomez | 1.3 | psrlw mm0, 8 |
| 134 : | psrlw mm1, 8 | ||
| 135 : | edgomez | 1.2 | punpcklbw mm4, [eax] |
| 136 : | punpcklbw mm5, [eax+edx+4] | ||
| 137 : | movq [byte ecx+0*64+48], mm6 | ||
| 138 : | movq [ecx+0*64+40], mm7 | ||
| 139 : | edgomez | 1.3 | psrlw mm2, 8 |
| 140 : | psrlw mm3, 8 | ||
| 141 : | edgomez | 1.2 | punpcklbw mm6, [eax+edx] |
| 142 : | punpcklbw mm7, [eax+4] | ||
| 143 : | movq [byte ecx+1*64], mm0 | ||
| 144 : | movq [ecx+1*64+24], mm1 | ||
| 145 : | edgomez | 1.3 | psrlw mm4, 8 |
| 146 : | psrlw mm5, 8 | ||
| 147 : | edgomez | 1.2 | movq [ecx+1*64+16], mm2 |
| 148 : | movq [ecx+1*64+8], mm3 | ||
| 149 : | edgomez | 1.3 | psrlw mm6, 8 |
| 150 : | psrlw mm7, 8 | ||
| 151 : | edgomez | 1.2 | movq [byte ecx+1*64+32], mm4 |
| 152 : | edgomez | 1.3 | movq [ecx+1*64+56], mm5 |
| 153 : | edgomez | 1.2 | movq [byte ecx+1*64+48], mm6 |
| 154 : | movq [ecx+1*64+40], mm7 | ||
| 155 : | edgomez | 1.3 | ret |
| 156 : | Isibaar | 1.10 | ENDFUNC |
| 157 : | edgomez | 1.2 | |
| 158 : | edgomez | 1.3 | |
| 159 : | ;----------------------------------------------------------------------------- | ||
| 160 : | edgomez | 1.2 | ; |
| 161 : | ; void transfer_16to8copy_3dne(uint8_t * const dst, | ||
| 162 : | ; const int16_t * const src, | ||
| 163 : | ; uint32_t stride); | ||
| 164 : | ; | ||
| 165 : | edgomez | 1.3 | ;----------------------------------------------------------------------------- |
| 166 : | edgomez | 1.2 | |
| 167 : | edgomez | 1.3 | ALIGN 16 |
| 168 : | edgomez | 1.2 | transfer_16to8copy_3dne: |
| 169 : | |||
| 170 : | mov eax, [esp+ 8] ; Src | ||
| 171 : | mov ecx, [esp+ 4] ; Dst | ||
| 172 : | mov edx, [esp+12] ; Stride | ||
| 173 : | |||
| 174 : | movq mm0, [byte eax+0*32] | ||
| 175 : | edgomez | 1.3 | packuswb mm0, [eax+0*32+8] |
| 176 : | edgomez | 1.2 | movq mm1, [eax+0*32+16] |
| 177 : | edgomez | 1.3 | packuswb mm1, [eax+0*32+24] |
| 178 : | edgomez | 1.2 | movq mm5, [eax+2*32+16] |
| 179 : | movq mm2, [eax+1*32] | ||
| 180 : | edgomez | 1.3 | packuswb mm2, [eax+1*32+8] |
| 181 : | edgomez | 1.2 | movq mm3, [eax+1*32+16] |
| 182 : | packuswb mm3, [eax+1*32+24] | ||
| 183 : | edgomez | 1.3 | movq mm6, [eax+3*32] |
| 184 : | edgomez | 1.2 | movq mm4, [eax+2*32] |
| 185 : | edgomez | 1.3 | packuswb mm4, [eax+2*32+8] |
| 186 : | edgomez | 1.2 | packuswb mm5, [eax+2*32+24] |
| 187 : | movq mm7, [eax+3*32+16] | ||
| 188 : | packuswb mm7, [eax+3*32+24] | ||
| 189 : | edgomez | 1.3 | packuswb mm6, [eax+3*32+8] |
| 190 : | edgomez | 1.2 | movq [ecx], mm0 |
| 191 : | edgomez | 1.3 | lea eax, [3*edx] |
| 192 : | add eax, ecx | ||
| 193 : | edgomez | 1.2 | movq [ecx+edx], mm1 |
| 194 : | movq [ecx+2*edx], mm2 | ||
| 195 : | movq [byte eax], mm3 | ||
| 196 : | movq [ecx+4*edx], mm4 | ||
| 197 : | edgomez | 1.3 | lea ecx, [byte ecx+4*edx] |
| 198 : | edgomez | 1.2 | movq [eax+2*edx], mm5 |
| 199 : | movq [eax+4*edx], mm7 | ||
| 200 : | movq [ecx+2*edx], mm6 | ||
| 201 : | ret | ||
| 202 : | Isibaar | 1.10 | ENDFUNC |
| 203 : | edgomez | 1.2 | |
| 204 : | edgomez | 1.3 | ;----------------------------------------------------------------------------- |
| 205 : | edgomez | 1.2 | ; |
| 206 : | ; void transfer_8to16sub_3dne(int16_t * const dct, | ||
| 207 : | ; uint8_t * const cur, | ||
| 208 : | ; const uint8_t * const ref, | ||
| 209 : | ; const uint32_t stride); | ||
| 210 : | ; | ||
| 211 : | edgomez | 1.3 | ;----------------------------------------------------------------------------- |
| 212 : | edgomez | 1.2 | |
| 213 : | ; when second argument == 1, reference (ebx) block is to current (eax) | ||
| 214 : | %macro COPY_8_TO_16_SUB 2 | ||
| 215 : | movq mm1, [eax] ; cur | ||
| 216 : | movq mm0, mm1 | ||
| 217 : | movq mm4, [ecx] ; ref | ||
| 218 : | movq mm6, mm4 | ||
| 219 : | %if %2 == 1 | ||
| 220 : | movq [eax], mm4 | ||
| 221 : | %endif | ||
| 222 : | punpckhbw mm1, mm7 | ||
| 223 : | punpckhbw mm6, mm7 | ||
| 224 : | punpcklbw mm4, mm7 | ||
| 225 : | edgomez | 1.3 | ALIGN 8 |
| 226 : | edgomez | 1.2 | movq mm2, [byte eax+edx] |
| 227 : | punpcklbw mm0, mm7 | ||
| 228 : | movq mm3, [byte eax+edx] | ||
| 229 : | punpcklbw mm2, mm7 | ||
| 230 : | movq mm5, [byte ecx+edx] ; ref | ||
| 231 : | punpckhbw mm3, mm7 | ||
| 232 : | %if %2 == 1 | ||
| 233 : | movq [byte eax+edx], mm5 | ||
| 234 : | %endif | ||
| 235 : | psubsw mm1, mm6 | ||
| 236 : | |||
| 237 : | movq mm6, mm5 | ||
| 238 : | psubsw mm0, mm4 | ||
| 239 : | edgomez | 1.3 | %if (%1 < 3) |
| 240 : | edgomez | 1.2 | lea eax,[eax+2*edx] |
| 241 : | lea ecx,[ecx+2*edx] | ||
| 242 : | %else | ||
| 243 : | mov ecx,[esp] | ||
| 244 : | add esp,byte 4 | ||
| 245 : | %endif | ||
| 246 : | movq [edi+%1*32+ 8], mm1 | ||
| 247 : | movq [byte edi+%1*32+ 0], mm0 ; dst | ||
| 248 : | punpcklbw mm5, mm7 | ||
| 249 : | punpckhbw mm6, mm7 | ||
| 250 : | psubsw mm2, mm5 | ||
| 251 : | psubsw mm3, mm6 | ||
| 252 : | movq [edi+%1*32+16], mm2 | ||
| 253 : | movq [edi+%1*32+24], mm3 | ||
| 254 : | %endmacro | ||
| 255 : | |||
| 256 : | edgomez | 1.3 | ALIGN 16 |
| 257 : | edgomez | 1.2 | transfer_8to16sub_3dne: |
| 258 : | mov eax, [esp + 8] ; Cur | ||
| 259 : | mov ecx, [esp +12] ; Ref | ||
| 260 : | push edi | ||
| 261 : | suxen_drol | 1.8 | mov edx, [esp+4+16] ; Stride |
| 262 : | edgomez | 1.2 | mov edi, [esp+4+ 4] ; Dst |
| 263 : | pxor mm7, mm7 | ||
| 264 : | nop | ||
| 265 : | edgomez | 1.3 | ALIGN 4 |
| 266 : | edgomez | 1.2 | COPY_8_TO_16_SUB 0, 1 |
| 267 : | COPY_8_TO_16_SUB 1, 1 | ||
| 268 : | COPY_8_TO_16_SUB 2, 1 | ||
| 269 : | COPY_8_TO_16_SUB 3, 1 | ||
| 270 : | edgomez | 1.3 | mov edi, ecx |
| 271 : | edgomez | 1.2 | ret |
| 272 : | Isibaar | 1.10 | ENDFUNC |
| 273 : | edgomez | 1.2 | |
| 274 : | edgomez | 1.3 | ALIGN 16 |
| 275 : | edgomez | 1.2 | transfer_8to16subro_3dne: |
| 276 : | mov eax, [esp + 8] ; Cur | ||
| 277 : | mov ecx, [esp +12] ; Ref | ||
| 278 : | push edi | ||
| 279 : | suxen_drol | 1.8 | mov edx, [esp+4+16] ; Stride |
| 280 : | edgomez | 1.2 | mov edi, [esp+4+ 4] ; Dst |
| 281 : | pxor mm7, mm7 | ||
| 282 : | nop | ||
| 283 : | edgomez | 1.3 | ALIGN 4 |
| 284 : | edgomez | 1.2 | COPY_8_TO_16_SUB 0, 0 |
| 285 : | COPY_8_TO_16_SUB 1, 0 | ||
| 286 : | COPY_8_TO_16_SUB 2, 0 | ||
| 287 : | COPY_8_TO_16_SUB 3, 0 | ||
| 288 : | edgomez | 1.3 | mov edi, ecx |
| 289 : | edgomez | 1.2 | ret |
| 290 : | Isibaar | 1.10 | ENDFUNC |
| 291 : | edgomez | 1.2 | |
| 292 : | |||
| 293 : | edgomez | 1.3 | ;----------------------------------------------------------------------------- |
| 294 : | edgomez | 1.2 | ; |
| 295 : | ; void transfer_8to16sub2_3dne(int16_t * const dct, | ||
| 296 : | ; uint8_t * const cur, | ||
| 297 : | ; const uint8_t * ref1, | ||
| 298 : | ; const uint8_t * ref2, | ||
| 299 : | ; const uint32_t stride) | ||
| 300 : | ; | ||
| 301 : | edgomez | 1.3 | ;----------------------------------------------------------------------------- |
| 302 : | edgomez | 1.2 | |
| 303 : | %macro COPY_8_TO_16_SUB2_SSE 1 | ||
| 304 : | edgomez | 1.3 | db 0Fh, 6Fh, 44h, 20h, 00 ;movq mm0, [byte eax] ; cur |
| 305 : | edgomez | 1.2 | punpcklbw mm0, mm7 |
| 306 : | movq mm2, [byte eax+edx] | ||
| 307 : | punpcklbw mm2, mm7 | ||
| 308 : | db 0Fh, 6Fh, 4ch, 20h, 00 ;movq mm1, [byte eax] | ||
| 309 : | punpckhbw mm1, mm7 | ||
| 310 : | movq mm3, [byte eax+edx] | ||
| 311 : | punpckhbw mm3, mm7 | ||
| 312 : | edgomez | 1.3 | |
| 313 : | edgomez | 1.2 | movq mm4, [byte ebx] ; ref1 |
| 314 : | pavgb mm4, [byte esi] ; ref2 | ||
| 315 : | edgomez | 1.3 | movq [eax], mm4 |
| 316 : | edgomez | 1.2 | movq mm5, [ebx+edx] ; ref |
| 317 : | pavgb mm5, [esi+edx] ; ref2 | ||
| 318 : | edgomez | 1.3 | movq [eax+edx], mm5 |
| 319 : | edgomez | 1.2 | movq mm6, mm4 |
| 320 : | punpcklbw mm4, mm7 | ||
| 321 : | punpckhbw mm6, mm7 | ||
| 322 : | %if (%1 < 3) | ||
| 323 : | lea esi,[esi+2*edx] | ||
| 324 : | lea ebx,[byte ebx+2*edx] | ||
| 325 : | lea eax,[eax+2*edx] | ||
| 326 : | %else | ||
| 327 : | mov esi,[esp] | ||
| 328 : | mov ebx,[esp+4] | ||
| 329 : | add esp,byte 8 | ||
| 330 : | edgomez | 1.3 | %endif |
| 331 : | edgomez | 1.2 | psubsw mm0, mm4 |
| 332 : | psubsw mm1, mm6 | ||
| 333 : | movq mm6, mm5 | ||
| 334 : | punpcklbw mm5, mm7 | ||
| 335 : | punpckhbw mm6, mm7 | ||
| 336 : | psubsw mm2, mm5 | ||
| 337 : | psubsw mm3, mm6 | ||
| 338 : | movq [byte ecx+%1*32+ 0], mm0 ; dst | ||
| 339 : | movq [ecx+%1*32+ 8], mm1 | ||
| 340 : | movq [ecx+%1*32+16], mm2 | ||
| 341 : | movq [ecx+%1*32+24], mm3 | ||
| 342 : | %endmacro | ||
| 343 : | |||
| 344 : | edgomez | 1.3 | ALIGN 16 |
| 345 : | edgomez | 1.2 | transfer_8to16sub2_3dne: |
| 346 : | mov edx, [esp +20] ; Stride | ||
| 347 : | mov ecx, [esp + 4] ; Dst | ||
| 348 : | mov eax, [esp + 8] ; Cur | ||
| 349 : | push ebx | ||
| 350 : | lea ebp,[byte ebp] | ||
| 351 : | mov ebx, [esp+4+12] ; Ref1 | ||
| 352 : | push esi | ||
| 353 : | pxor mm7, mm7 | ||
| 354 : | mov esi, [esp+8+16] ; Ref2 | ||
| 355 : | edgomez | 1.3 | nop4 |
| 356 : | edgomez | 1.2 | COPY_8_TO_16_SUB2_SSE 0 |
| 357 : | COPY_8_TO_16_SUB2_SSE 1 | ||
| 358 : | COPY_8_TO_16_SUB2_SSE 2 | ||
| 359 : | COPY_8_TO_16_SUB2_SSE 3 | ||
| 360 : | |||
| 361 : | ret | ||
| 362 : | Isibaar | 1.10 | ENDFUNC |
| 363 : | edgomez | 1.2 | |
| 364 : | |||
| 365 : | edgomez | 1.3 | ;----------------------------------------------------------------------------- |
| 366 : | edgomez | 1.2 | ; |
| 367 : | ; void transfer_16to8add_3dne(uint8_t * const dst, | ||
| 368 : | ; const int16_t * const src, | ||
| 369 : | ; uint32_t stride); | ||
| 370 : | ; | ||
| 371 : | edgomez | 1.3 | ;----------------------------------------------------------------------------- |
| 372 : | edgomez | 1.2 | |
| 373 : | %macro COPY_16_TO_8_ADD 1 | ||
| 374 : | db 0Fh, 6Fh, 44h, 21h, 00 ;movq mm0, [byte ecx] | ||
| 375 : | punpcklbw mm0, mm7 | ||
| 376 : | movq mm2, [byte ecx+edx] | ||
| 377 : | punpcklbw mm2, mm7 | ||
| 378 : | db 0Fh, 6Fh, 4ch, 21h, 00 ;movq mm1, [byte ecx] | ||
| 379 : | punpckhbw mm1, mm7 | ||
| 380 : | movq mm3, [byte ecx+edx] | ||
| 381 : | punpckhbw mm3, mm7 | ||
| 382 : | paddsw mm0, [byte eax+%1*32+ 0] | ||
| 383 : | paddsw mm1, [eax+%1*32+ 8] | ||
| 384 : | paddsw mm2, [eax+%1*32+16] | ||
| 385 : | paddsw mm3, [eax+%1*32+24] | ||
| 386 : | packuswb mm0, mm1 | ||
| 387 : | packuswb mm2, mm3 | ||
| 388 : | edgomez | 1.3 | mov esp,esp |
| 389 : | edgomez | 1.2 | movq [byte ecx], mm0 |
| 390 : | movq [ecx+edx], mm2 | ||
| 391 : | %endmacro | ||
| 392 : | |||
| 393 : | |||
| 394 : | edgomez | 1.3 | ALIGN 16 |
| 395 : | edgomez | 1.2 | transfer_16to8add_3dne: |
| 396 : | mov ecx, [esp+ 4] ; Dst | ||
| 397 : | mov edx, [esp+12] ; Stride | ||
| 398 : | mov eax, [esp+ 8] ; Src | ||
| 399 : | pxor mm7, mm7 | ||
| 400 : | nop | ||
| 401 : | |||
| 402 : | COPY_16_TO_8_ADD 0 | ||
| 403 : | lea ecx,[byte ecx+2*edx] | ||
| 404 : | COPY_16_TO_8_ADD 1 | ||
| 405 : | lea ecx,[byte ecx+2*edx] | ||
| 406 : | COPY_16_TO_8_ADD 2 | ||
| 407 : | lea ecx,[byte ecx+2*edx] | ||
| 408 : | COPY_16_TO_8_ADD 3 | ||
| 409 : | ret | ||
| 410 : | Isibaar | 1.10 | ENDFUNC |
| 411 : | edgomez | 1.2 | |
| 412 : | edgomez | 1.3 | ;----------------------------------------------------------------------------- |
| 413 : | edgomez | 1.2 | ; |
| 414 : | ; void transfer8x8_copy_3dne(uint8_t * const dst, | ||
| 415 : | ; const uint8_t * const src, | ||
| 416 : | ; const uint32_t stride); | ||
| 417 : | ; | ||
| 418 : | ; | ||
| 419 : | edgomez | 1.3 | ;----------------------------------------------------------------------------- |
| 420 : | edgomez | 1.2 | |
| 421 : | %macro COPY_8_TO_8 0 | ||
| 422 : | movq mm0, [byte eax] | ||
| 423 : | movq mm1, [eax+edx] | ||
| 424 : | movq [byte ecx], mm0 | ||
| 425 : | lea eax,[byte eax+2*edx] | ||
| 426 : | movq [ecx+edx], mm1 | ||
| 427 : | %endmacro | ||
| 428 : | |||
| 429 : | edgomez | 1.3 | ALIGN 16 |
| 430 : | edgomez | 1.2 | transfer8x8_copy_3dne: |
| 431 : | mov eax, [esp+ 8] ; Src | ||
| 432 : | mov edx, [esp+12] ; Stride | ||
| 433 : | mov ecx, [esp+ 4] ; Dst | ||
| 434 : | |||
| 435 : | COPY_8_TO_8 | ||
| 436 : | lea ecx,[byte ecx+2*edx] | ||
| 437 : | COPY_8_TO_8 | ||
| 438 : | lea ecx,[byte ecx+2*edx] | ||
| 439 : | COPY_8_TO_8 | ||
| 440 : | lea ecx,[byte ecx+2*edx] | ||
| 441 : | COPY_8_TO_8 | ||
| 442 : | ret | ||
| 443 : | Isibaar | 1.10 | ENDFUNC |
| 444 : | edgomez | 1.6 | |
| 445 : | suxen_drol | 1.7 | ;----------------------------------------------------------------------------- |
| 446 : | ; | ||
| 447 : | ; void transfer8x4_copy_3dne(uint8_t * const dst, | ||
| 448 : | ; const uint8_t * const src, | ||
| 449 : | ; const uint32_t stride); | ||
| 450 : | ; | ||
| 451 : | ; | ||
| 452 : | ;----------------------------------------------------------------------------- | ||
| 453 : | |||
| 454 : | ALIGN 16 | ||
| 455 : | transfer8x4_copy_3dne: | ||
| 456 : | mov eax, [esp+ 8] ; Src | ||
| 457 : | mov edx, [esp+12] ; Stride | ||
| 458 : | mov ecx, [esp+ 4] ; Dst | ||
| 459 : | |||
| 460 : | COPY_8_TO_8 | ||
| 461 : | lea ecx,[byte ecx+2*edx] | ||
| 462 : | COPY_8_TO_8 | ||
| 463 : | ret | ||
| 464 : | Isibaar | 1.10 | ENDFUNC |
| 465 : | suxen_drol | 1.7 | |
| 466 : | Isibaar | 1.9 | |
| 467 : | %ifidn __OUTPUT_FORMAT__,elf | ||
| 468 : | section ".note.GNU-stack" noalloc noexec nowrite progbits | ||
| 469 : | %endif | ||
| 470 : |
| No admin address has been configured | ViewVC Help |
| Powered by ViewVC 1.0.4 |