Parent Directory
| 
Revision Log
Revision 1803 - (view) (download)
| 1 : | edgomez | 1382 | ;/**************************************************************************** |
| 2 : | ; * | ||
| 3 : | ; * XVID MPEG-4 VIDEO CODEC | ||
| 4 : | ; * - MMX and XMM YV12->YV12 conversion - | ||
| 5 : | ; * | ||
| 6 : | Isibaar | 1795 | ; * Copyright(C) 2001-2008 Michael Militzer <michael@xvid.org> |
| 7 : | edgomez | 1382 | ; * |
| 8 : | ; * This program is free software; you can redistribute it and/or modify it | ||
| 9 : | ; * 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 : | ; * | ||
| 13 : | ; * 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 : | ; * | ||
| 18 : | ; * 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 : | ; * | ||
| 22 : | Isibaar | 1803 | ; * $Id: colorspace_yuv_mmx.asm,v 1.10 2008-11-26 23:35:50 Isibaar Exp $ |
| 23 : | edgomez | 1382 | ; * |
| 24 : | ; ***************************************************************************/ | ||
| 25 : | edgomez | 851 | |
| 26 : | Isibaar | 1795 | %include "nasm.inc" |
| 27 : | edgomez | 851 | |
| 28 : | edgomez | 1382 | ;============================================================================= |
| 29 : | ; Helper macros | ||
| 30 : | ;============================================================================= | ||
| 31 : | edgomez | 851 | |
| 32 : | ;------------------------------------------------------------------------------ | ||
| 33 : | Skal | 1742 | ; PLANE_COPY ( DST, DST_STRIDE, SRC, SRC_STRIDE, WIDTH, HEIGHT, OPT ) |
| 34 : | edgomez | 851 | ; DST dst buffer |
| 35 : | Skal | 1742 | ; DST_STRIDE dst stride |
| 36 : | edgomez | 851 | ; SRC src destination buffer |
| 37 : | Skal | 1742 | ; SRC_STRIDE src stride |
| 38 : | edgomez | 851 | ; WIDTH width |
| 39 : | ; HEIGHT height | ||
| 40 : | ; OPT 0=plain mmx, 1=xmm | ||
| 41 : | Isibaar | 1795 | ; |
| 42 : | ; | ||
| 43 : | ; Trashes: DST, SRC, WIDTH, HEIGHT, _EBX, _ECX, _EDX | ||
| 44 : | edgomez | 851 | ;------------------------------------------------------------------------------ |
| 45 : | edgomez | 1382 | |
| 46 : | edgomez | 851 | %macro PLANE_COPY 7 |
| 47 : | Isibaar | 1795 | %define DST %1 |
| 48 : | Isibaar | 1803 | %define DST_STRIDE %2 |
| 49 : | Isibaar | 1795 | %define SRC %3 |
| 50 : | Isibaar | 1803 | %define SRC_STRIDE %4 |
| 51 : | %define WIDTH %5 | ||
| 52 : | %define HEIGHT %6 | ||
| 53 : | Isibaar | 1795 | %define OPT %7 |
| 54 : | edgomez | 851 | |
| 55 : | Isibaar | 1795 | mov _EBX, WIDTH |
| 56 : | shr WIDTH, 6 ; $_EAX$ = width / 64 | ||
| 57 : | and _EBX, 63 ; remainder = width % 64 | ||
| 58 : | mov _EDX, _EBX | ||
| 59 : | shr _EBX, 4 ; $_EBX$ = remainder / 16 | ||
| 60 : | and _EDX, 15 ; $_EDX$ = remainder % 16 | ||
| 61 : | edgomez | 851 | |
| 62 : | Isibaar | 1795 | %%loop64_start_pc: |
| 63 : | push DST | ||
| 64 : | push SRC | ||
| 65 : | edgomez | 851 | |
| 66 : | Isibaar | 1795 | mov _ECX, WIDTH ; width64 |
| 67 : | test WIDTH, WIDTH | ||
| 68 : | Skal | 1742 | jz %%loop16_start_pc |
| 69 : | |||
| 70 : | %%loop64_pc: | ||
| 71 : | edgomez | 1382 | %if OPT == 1 ; xmm |
| 72 : | Isibaar | 1795 | prefetchnta [SRC + 64] ; non temporal prefetch |
| 73 : | prefetchnta [SRC + 96] | ||
| 74 : | edgomez | 851 | %endif |
| 75 : | Isibaar | 1795 | movq mm1, [SRC ] ; read from src |
| 76 : | movq mm2, [SRC + 8] | ||
| 77 : | movq mm3, [SRC + 16] | ||
| 78 : | movq mm4, [SRC + 24] | ||
| 79 : | movq mm5, [SRC + 32] | ||
| 80 : | movq mm6, [SRC + 40] | ||
| 81 : | movq mm7, [SRC + 48] | ||
| 82 : | movq mm0, [SRC + 56] | ||
| 83 : | edgomez | 851 | |
| 84 : | edgomez | 1382 | %if OPT == 0 ; plain mmx |
| 85 : | Isibaar | 1795 | movq [DST ], mm1 ; write to y_out |
| 86 : | movq [DST + 8], mm2 | ||
| 87 : | movq [DST + 16], mm3 | ||
| 88 : | movq [DST + 24], mm4 | ||
| 89 : | movq [DST + 32], mm5 | ||
| 90 : | movq [DST + 40], mm6 | ||
| 91 : | movq [DST + 48], mm7 | ||
| 92 : | movq [DST + 56], mm0 | ||
| 93 : | edgomez | 851 | %else |
| 94 : | Isibaar | 1795 | movntq [DST ], mm1 ; write to y_out |
| 95 : | movntq [DST + 8], mm2 | ||
| 96 : | movntq [DST + 16], mm3 | ||
| 97 : | movntq [DST + 24], mm4 | ||
| 98 : | movntq [DST + 32], mm5 | ||
| 99 : | movntq [DST + 40], mm6 | ||
| 100 : | movntq [DST + 48], mm7 | ||
| 101 : | movntq [DST + 56], mm0 | ||
| 102 : | edgomez | 851 | %endif |
| 103 : | |||
| 104 : | Isibaar | 1795 | add SRC, 64 |
| 105 : | add DST, 64 | ||
| 106 : | Skal | 1742 | loop %%loop64_pc |
| 107 : | edgomez | 851 | |
| 108 : | Skal | 1742 | %%loop16_start_pc: |
| 109 : | Isibaar | 1795 | mov _ECX, _EBX ; width16 |
| 110 : | test _EBX, _EBX | ||
| 111 : | Skal | 1742 | jz %%loop1_start_pc |
| 112 : | |||
| 113 : | %%loop16_pc: | ||
| 114 : | Isibaar | 1795 | movq mm1, [SRC] |
| 115 : | movq mm2, [SRC + 8] | ||
| 116 : | edgomez | 1382 | %if OPT == 0 ; plain mmx |
| 117 : | Isibaar | 1795 | movq [DST], mm1 |
| 118 : | movq [DST + 8], mm2 | ||
| 119 : | edgomez | 851 | %else |
| 120 : | Isibaar | 1795 | movntq [DST], mm1 |
| 121 : | movntq [DST + 8], mm2 | ||
| 122 : | edgomez | 851 | %endif |
| 123 : | |||
| 124 : | Isibaar | 1795 | add SRC, 16 |
| 125 : | add DST, 16 | ||
| 126 : | Skal | 1742 | loop %%loop16_pc |
| 127 : | edgomez | 851 | |
| 128 : | Skal | 1742 | %%loop1_start_pc: |
| 129 : | Isibaar | 1795 | mov _ECX, _EDX |
| 130 : | edgomez | 1382 | rep movsb |
| 131 : | edgomez | 851 | |
| 132 : | Isibaar | 1795 | pop SRC |
| 133 : | pop DST | ||
| 134 : | |||
| 135 : | %ifdef ARCH_IS_X86_64 | ||
| 136 : | movsx _ECX, SRC_STRIDE | ||
| 137 : | add SRC, _ECX | ||
| 138 : | mov ecx, DST_STRIDE | ||
| 139 : | add DST, _ECX | ||
| 140 : | %else | ||
| 141 : | add SRC, SRC_STRIDE | ||
| 142 : | add DST, DST_STRIDE | ||
| 143 : | %endif | ||
| 144 : | Isibaar | 1803 | |
| 145 : | Isibaar | 1795 | dec HEIGHT |
| 146 : | Skal | 1742 | jg near %%loop64_start_pc |
| 147 : | Isibaar | 1795 | |
| 148 : | %undef DST | ||
| 149 : | %undef DST_STRIDE | ||
| 150 : | %undef SRC | ||
| 151 : | %undef SRC_STRIDE | ||
| 152 : | %undef WIDTH | ||
| 153 : | %undef HEIGHT | ||
| 154 : | %undef OPT | ||
| 155 : | edgomez | 851 | %endmacro |
| 156 : | |||
| 157 : | ;------------------------------------------------------------------------------ | ||
| 158 : | Skal | 1742 | ; PLANE_FILL ( DST, DST_STRIDE, WIDTH, HEIGHT, OPT ) |
| 159 : | ; DST dst buffer | ||
| 160 : | ; DST_STRIDE dst stride | ||
| 161 : | ; WIDTH width | ||
| 162 : | ; HEIGHT height | ||
| 163 : | ; OPT 0=plain mmx, 1=xmm | ||
| 164 : | Isibaar | 1795 | ; |
| 165 : | ; Trashes: DST, WIDTH, HEIGHT, _EBX, _ECX, _EDX, _EAX | ||
| 166 : | Skal | 1742 | ;------------------------------------------------------------------------------ |
| 167 : | |||
| 168 : | %macro PLANE_FILL 5 | ||
| 169 : | %define DST %1 | ||
| 170 : | %define DST_STRIDE %2 | ||
| 171 : | %define WIDTH %3 | ||
| 172 : | %define HEIGHT %4 | ||
| 173 : | %define OPT %5 | ||
| 174 : | |||
| 175 : | Isibaar | 1795 | mov _EAX, 0x80808080 |
| 176 : | mov _EBX, WIDTH | ||
| 177 : | shr WIDTH, 6 ; $_ESI$ = width / 64 | ||
| 178 : | and _EBX, 63 ; _EBX = remainder = width % 64 | ||
| 179 : | movd mm0, eax | ||
| 180 : | mov _EDX, _EBX | ||
| 181 : | shr _EBX, 4 ; $_EBX$ = remainder / 16 | ||
| 182 : | and _EDX, 15 ; $_EDX$ = remainder % 16 | ||
| 183 : | Skal | 1742 | punpckldq mm0, mm0 |
| 184 : | |||
| 185 : | %%loop64_start_pf: | ||
| 186 : | Isibaar | 1795 | push DST |
| 187 : | mov _ECX, WIDTH ; width64 | ||
| 188 : | test WIDTH, WIDTH | ||
| 189 : | Skal | 1742 | jz %%loop16_start_pf |
| 190 : | |||
| 191 : | %%loop64_pf: | ||
| 192 : | |||
| 193 : | Isibaar | 1795 | %if OPT == 0 ; plain mmx |
| 194 : | movq [DST ], mm0 ; write to y_out | ||
| 195 : | movq [DST + 8], mm0 | ||
| 196 : | movq [DST + 16], mm0 | ||
| 197 : | movq [DST + 24], mm0 | ||
| 198 : | movq [DST + 32], mm0 | ||
| 199 : | movq [DST + 40], mm0 | ||
| 200 : | movq [DST + 48], mm0 | ||
| 201 : | movq [DST + 56], mm0 | ||
| 202 : | Skal | 1742 | %else |
| 203 : | Isibaar | 1795 | movntq [DST ], mm0 ; write to y_out |
| 204 : | movntq [DST + 8], mm0 | ||
| 205 : | movntq [DST + 16], mm0 | ||
| 206 : | movntq [DST + 24], mm0 | ||
| 207 : | movntq [DST + 32], mm0 | ||
| 208 : | movntq [DST + 40], mm0 | ||
| 209 : | movntq [DST + 48], mm0 | ||
| 210 : | movntq [DST + 56], mm0 | ||
| 211 : | Skal | 1742 | %endif |
| 212 : | |||
| 213 : | Isibaar | 1795 | add DST, 64 |
| 214 : | Skal | 1742 | loop %%loop64_pf |
| 215 : | |||
| 216 : | %%loop16_start_pf: | ||
| 217 : | Isibaar | 1795 | mov _ECX, _EBX ; width16 |
| 218 : | test _EBX, _EBX | ||
| 219 : | Skal | 1742 | jz %%loop1_start_pf |
| 220 : | |||
| 221 : | %%loop16_pf: | ||
| 222 : | Isibaar | 1795 | %if OPT == 0 ; plain mmx |
| 223 : | movq [DST ], mm0 | ||
| 224 : | movq [DST + 8], mm0 | ||
| 225 : | Skal | 1742 | %else |
| 226 : | Isibaar | 1795 | movntq [DST ], mm0 |
| 227 : | movntq [DST + 8], mm0 | ||
| 228 : | Skal | 1742 | %endif |
| 229 : | |||
| 230 : | Isibaar | 1795 | add DST, 16 |
| 231 : | Skal | 1742 | loop %%loop16_pf |
| 232 : | |||
| 233 : | %%loop1_start_pf: | ||
| 234 : | Isibaar | 1795 | mov _ECX, _EDX |
| 235 : | Skal | 1742 | rep stosb |
| 236 : | |||
| 237 : | Isibaar | 1795 | pop DST |
| 238 : | |||
| 239 : | %ifdef ARCH_IS_X86_64 | ||
| 240 : | mov ecx, DST_STRIDE | ||
| 241 : | add DST, _ECX | ||
| 242 : | %else | ||
| 243 : | add DST, DST_STRIDE | ||
| 244 : | %endif | ||
| 245 : | |||
| 246 : | dec HEIGHT | ||
| 247 : | Skal | 1742 | jg near %%loop64_start_pf |
| 248 : | Isibaar | 1795 | |
| 249 : | %undef DST | ||
| 250 : | %undef DST_STRIDE | ||
| 251 : | %undef WIDTH | ||
| 252 : | %undef HEIGHT | ||
| 253 : | %undef OPT | ||
| 254 : | Skal | 1742 | %endmacro |
| 255 : | |||
| 256 : | ;------------------------------------------------------------------------------ | ||
| 257 : | edgomez | 851 | ; MAKE_YV12_TO_YV12( NAME, OPT ) |
| 258 : | ; NAME function name | ||
| 259 : | ; OPT 0=plain mmx, 1=xmm | ||
| 260 : | ; | ||
| 261 : | edgomez | 1382 | ; yv12_to_yv12_mmx(uint8_t * y_dst, uint8_t * u_dst, uint8_t * v_dst, |
| 262 : | edgomez | 851 | ; int y_dst_stride, int uv_dst_stride, |
| 263 : | edgomez | 1382 | ; uint8_t * y_src, uint8_t * u_src, uint8_t * v_src, |
| 264 : | edgomez | 851 | ; int y_src_stride, int uv_src_stride, |
| 265 : | ; int width, int height, int vflip) | ||
| 266 : | ;------------------------------------------------------------------------------ | ||
| 267 : | %macro MAKE_YV12_TO_YV12 2 | ||
| 268 : | %define NAME %1 | ||
| 269 : | Isibaar | 1795 | %define XMM_OPT %2 |
| 270 : | ALIGN SECTION_ALIGN | ||
| 271 : | edgomez | 851 | cglobal NAME |
| 272 : | edgomez | 1382 | NAME: |
| 273 : | edgomez | 851 | |
| 274 : | Isibaar | 1795 | push _EBX ; _ESP + localsize + 3*PTR_SIZE |
| 275 : | edgomez | 851 | |
| 276 : | Isibaar | 1795 | %define localsize 2*4 |
| 277 : | edgomez | 851 | |
| 278 : | Isibaar | 1795 | %ifdef ARCH_IS_X86_64 |
| 279 : | edgomez | 851 | |
| 280 : | Isibaar | 1795 | %ifndef WINDOWS |
| 281 : | %define pushsize 2*PTR_SIZE | ||
| 282 : | %define shadow 0 | ||
| 283 : | %else | ||
| 284 : | %define pushsize 4*PTR_SIZE | ||
| 285 : | Isibaar | 1803 | %define shadow 32 + 2*PTR_SIZE |
| 286 : | Isibaar | 1795 | %endif |
| 287 : | edgomez | 851 | |
| 288 : | Isibaar | 1795 | %define prm_vflip dword [_ESP + localsize + pushsize + shadow + 7*PTR_SIZE] |
| 289 : | %define prm_height dword [_ESP + localsize + pushsize + shadow + 6*PTR_SIZE] | ||
| 290 : | %define prm_width dword [_ESP + localsize + pushsize + shadow + 5*PTR_SIZE] | ||
| 291 : | %define prm_uv_src_stride dword [_ESP + localsize + pushsize + shadow + 4*PTR_SIZE] | ||
| 292 : | %define prm_y_src_stride dword [_ESP + localsize + pushsize + shadow + 3*PTR_SIZE] | ||
| 293 : | %define prm_v_src [_ESP + localsize + pushsize + shadow + 2*PTR_SIZE] | ||
| 294 : | %define prm_u_src [_ESP + localsize + pushsize + shadow + 1*PTR_SIZE] | ||
| 295 : | |||
| 296 : | %ifdef WINDOWS | ||
| 297 : | push _ESI ; _ESP + localsize + 2*PTR_SIZE | ||
| 298 : | push _EDI ; _ESP + localsize + 1*PTR_SIZE | ||
| 299 : | push _EBP ; _ESP + localsize + 0*PTR_SIZE | ||
| 300 : | |||
| 301 : | sub _ESP, localsize | ||
| 302 : | |||
| 303 : | %define prm_y_src _ESI | ||
| 304 : | %define prm_uv_dst_stride TMP0d | ||
| 305 : | %define prm_y_dst_stride prm4d | ||
| 306 : | %define prm_v_dst prm3 | ||
| 307 : | %define prm_u_dst TMP1 | ||
| 308 : | %define prm_y_dst _EDI | ||
| 309 : | |||
| 310 : | mov _EDI, prm1 | ||
| 311 : | mov TMP1, prm2 | ||
| 312 : | |||
| 313 : | Isibaar | 1803 | mov _ESI, [_ESP + localsize + pushsize + shadow + 0*PTR_SIZE] |
| 314 : | mov TMP0d, dword [_ESP + localsize + pushsize + shadow - 1*PTR_SIZE] | ||
| 315 : | Isibaar | 1795 | |
| 316 : | %else | ||
| 317 : | push _EBP ; _ESP + localsize + 0*PTR_SIZE | ||
| 318 : | |||
| 319 : | sub _ESP, localsize | ||
| 320 : | |||
| 321 : | %define prm_y_src _ESI | ||
| 322 : | %define prm_uv_dst_stride prm5d | ||
| 323 : | %define prm_y_dst_stride TMP1d | ||
| 324 : | %define prm_v_dst prm6 | ||
| 325 : | %define prm_u_dst TMP0 | ||
| 326 : | %define prm_y_dst _EDI | ||
| 327 : | |||
| 328 : | mov TMP0, prm2 | ||
| 329 : | mov _ESI, prm6 | ||
| 330 : | |||
| 331 : | mov prm6, prm3 | ||
| 332 : | mov TMP1d, prm4d | ||
| 333 : | %endif | ||
| 334 : | |||
| 335 : | %define _ip _ESP + localsize + pushsize + 0 | ||
| 336 : | |||
| 337 : | %else | ||
| 338 : | |||
| 339 : | %define pushsize 4*PTR_SIZE | ||
| 340 : | |||
| 341 : | %define prm_vflip [_ESP + localsize + pushsize + 13*PTR_SIZE] | ||
| 342 : | %define prm_height [_ESP + localsize + pushsize + 12*PTR_SIZE] | ||
| 343 : | %define prm_width [_ESP + localsize + pushsize + 11*PTR_SIZE] | ||
| 344 : | %define prm_uv_src_stride [_ESP + localsize + pushsize + 10*PTR_SIZE] | ||
| 345 : | %define prm_y_src_stride [_ESP + localsize + pushsize + 9*PTR_SIZE] | ||
| 346 : | %define prm_v_src [_ESP + localsize + pushsize + 8*PTR_SIZE] | ||
| 347 : | %define prm_u_src [_ESP + localsize + pushsize + 7*PTR_SIZE] | ||
| 348 : | |||
| 349 : | %define prm_y_src _ESI | ||
| 350 : | %define prm_uv_dst_stride [_ESP + localsize + pushsize + 5*PTR_SIZE] | ||
| 351 : | %define prm_y_dst_stride [_ESP + localsize + pushsize + 4*PTR_SIZE] | ||
| 352 : | %define prm_v_dst [_ESP + localsize + pushsize + 3*PTR_SIZE] | ||
| 353 : | %define prm_u_dst [_ESP + localsize + pushsize + 2*PTR_SIZE] | ||
| 354 : | %define prm_y_dst _EDI | ||
| 355 : | |||
| 356 : | %define _ip _ESP + localsize + pushsize + 0 | ||
| 357 : | |||
| 358 : | push _ESI ; _ESP + localsize + 2*PTR_SIZE | ||
| 359 : | push _EDI ; _ESP + localsize + 1*PTR_SIZE | ||
| 360 : | push _EBP ; _ESP + localsize + 0*PTR_SIZE | ||
| 361 : | |||
| 362 : | sub _ESP, localsize | ||
| 363 : | |||
| 364 : | mov _ESI, [_ESP + localsize + pushsize + 6*PTR_SIZE] | ||
| 365 : | mov _EDI, [_ESP + localsize + pushsize + 1*PTR_SIZE] | ||
| 366 : | |||
| 367 : | %endif | ||
| 368 : | |||
| 369 : | %define width2 dword [_ESP + localsize - 1*4] | ||
| 370 : | %define height2 dword [_ESP + localsize - 2*4] | ||
| 371 : | |||
| 372 : | mov eax, prm_width | ||
| 373 : | mov ebx, prm_height | ||
| 374 : | edgomez | 1382 | shr eax, 1 ; calculate widht/2, heigh/2 |
| 375 : | shr ebx, 1 | ||
| 376 : | Isibaar | 1795 | mov width2, eax |
| 377 : | mov height2, ebx | ||
| 378 : | edgomez | 851 | |
| 379 : | Isibaar | 1795 | mov eax, prm_vflip |
| 380 : | test eax, eax | ||
| 381 : | Skal | 1742 | jz near .go |
| 382 : | edgomez | 851 | |
| 383 : | Skal | 1742 | ; flipping support |
| 384 : | Isibaar | 1795 | mov eax, prm_height |
| 385 : | mov ecx, prm_y_src_stride | ||
| 386 : | Skal | 1742 | sub eax, 1 |
| 387 : | Isibaar | 1795 | imul eax, ecx |
| 388 : | Isibaar | 1803 | add _ESI, _EAX ; y_src += (height-1) * y_src_stride |
| 389 : | Isibaar | 1795 | neg ecx |
| 390 : | mov prm_y_src_stride, ecx ; y_src_stride = -y_src_stride | ||
| 391 : | edgomez | 851 | |
| 392 : | Isibaar | 1795 | mov eax, height2 |
| 393 : | mov _EDX, prm_u_src | ||
| 394 : | mov _EBP, prm_v_src | ||
| 395 : | mov ecx, prm_uv_src_stride | ||
| 396 : | test _EDX, _EDX | ||
| 397 : | Skal | 1742 | jz .go |
| 398 : | Isibaar | 1795 | test _EBP, _EBP |
| 399 : | Skal | 1742 | jz .go |
| 400 : | Isibaar | 1795 | sub eax, 1 ; _EAX = height2 - 1 |
| 401 : | imul eax, ecx | ||
| 402 : | Isibaar | 1803 | add _EDX, _EAX ; u_src += (height2-1) * uv_src_stride |
| 403 : | add _EBP, _EAX ; v_src += (height2-1) * uv_src_stride | ||
| 404 : | Isibaar | 1795 | neg ecx |
| 405 : | mov prm_u_src, _EDX | ||
| 406 : | mov prm_v_src, _EBP | ||
| 407 : | mov prm_uv_src_stride, ecx ; uv_src_stride = -uv_src_stride | ||
| 408 : | edgomez | 851 | |
| 409 : | Skal | 1742 | .go: |
| 410 : | Isibaar | 1795 | mov eax, prm_width |
| 411 : | mov ebp, prm_height | ||
| 412 : | PLANE_COPY _EDI, prm_y_dst_stride, _ESI, prm_y_src_stride, _EAX, _EBP, XMM_OPT | ||
| 413 : | edgomez | 851 | |
| 414 : | Isibaar | 1795 | mov _EAX, prm_u_src |
| 415 : | or _EAX, prm_v_src | ||
| 416 : | Skal | 1742 | jz near .UVFill_0x80 |
| 417 : | edgomez | 851 | |
| 418 : | Isibaar | 1795 | mov eax, width2 |
| 419 : | mov ebp, height2 | ||
| 420 : | mov _ESI, prm_u_src | ||
| 421 : | mov _EDI, prm_u_dst | ||
| 422 : | PLANE_COPY _EDI, prm_uv_dst_stride, _ESI, prm_uv_src_stride, _EAX, _EBP, XMM_OPT | ||
| 423 : | |||
| 424 : | mov eax, width2 | ||
| 425 : | mov ebp, height2 | ||
| 426 : | mov _ESI, prm_v_src | ||
| 427 : | mov _EDI, prm_v_dst | ||
| 428 : | PLANE_COPY _EDI, prm_uv_dst_stride, _ESI, prm_uv_src_stride, _EAX, _EBP, XMM_OPT | ||
| 429 : | |||
| 430 : | Skal | 1742 | .Done_UVPlane: |
| 431 : | Isibaar | 1795 | add _ESP, localsize |
| 432 : | |||
| 433 : | pop _EBP | ||
| 434 : | %ifndef ARCH_IS_X86_64 | ||
| 435 : | pop _EDI | ||
| 436 : | pop _ESI | ||
| 437 : | %else | ||
| 438 : | %ifdef WINDOWS | ||
| 439 : | pop _EDI | ||
| 440 : | pop _ESI | ||
| 441 : | %endif | ||
| 442 : | %endif | ||
| 443 : | pop _EBX | ||
| 444 : | |||
| 445 : | Skal | 1742 | ret |
| 446 : | edgomez | 851 | |
| 447 : | Skal | 1742 | .UVFill_0x80: |
| 448 : | Isibaar | 1795 | |
| 449 : | mov esi, width2 | ||
| 450 : | mov ebp, height2 | ||
| 451 : | mov _EDI, prm_u_dst | ||
| 452 : | PLANE_FILL _EDI, prm_uv_dst_stride, _ESI, _EBP, XMM_OPT | ||
| 453 : | |||
| 454 : | mov esi, width2 | ||
| 455 : | mov ebp, height2 | ||
| 456 : | mov _EDI, prm_v_dst | ||
| 457 : | PLANE_FILL _EDI, prm_uv_dst_stride, _ESI, _EBP, XMM_OPT | ||
| 458 : | |||
| 459 : | Skal | 1742 | jmp near .Done_UVPlane |
| 460 : | Isibaar | 1795 | |
| 461 : | Isibaar | 1793 | ENDFUNC |
| 462 : | Isibaar | 1795 | |
| 463 : | %undef NAME | ||
| 464 : | %undef XMM_OPT | ||
| 465 : | edgomez | 851 | %endmacro |
| 466 : | |||
| 467 : | edgomez | 1382 | ;============================================================================= |
| 468 : | ; Code | ||
| 469 : | ;============================================================================= | ||
| 470 : | edgomez | 851 | |
| 471 : | Isibaar | 1795 | SECTION .rotext align=SECTION_ALIGN |
| 472 : | edgomez | 1382 | |
| 473 : | edgomez | 851 | MAKE_YV12_TO_YV12 yv12_to_yv12_mmx, 0 |
| 474 : | edgomez | 1382 | |
| 475 : | edgomez | 851 | MAKE_YV12_TO_YV12 yv12_to_yv12_xmm, 1 |
| 476 : | Isibaar | 1790 | |
| 477 : | %ifidn __OUTPUT_FORMAT__,elf | ||
| 478 : | section ".note.GNU-stack" noalloc noexec nowrite progbits | ||
| 479 : | %endif | ||
| 480 : |
| No admin address has been configured | ViewVC Help |
| Powered by ViewVC 1.0.4 |