Parent Directory | Revision Log
Revision 652 -
(view)
(download)
Original Path: trunk/xvidcore/src/motion/x86_asm/sad_mmx.asm
1 : | chl | 430 | ;/***************************************************************************** |
2 : | Isibaar | 262 | ; * |
3 : | chl | 430 | ; * XVID MPEG-4 VIDEO CODEC |
4 : | ; * mmx sum of absolute difference | ||
5 : | Isibaar | 262 | ; * |
6 : | chl | 430 | ; * Copyright(C) 2002 Peter Ross <pross@xvid.org> |
7 : | Isibaar | 262 | ; * |
8 : | edgomez | 652 | ; * This file is part of XviD, a free MPEG-4 video encoder/decoder |
9 : | Isibaar | 262 | ; * |
10 : | edgomez | 652 | ; * XviD is free software; you can redistribute it and/or modify it |
11 : | ; * under the terms of the GNU General Public License as published by | ||
12 : | chl | 430 | ; * the Free Software Foundation; either version 2 of the License, or |
13 : | ; * (at your option) any later version. | ||
14 : | Isibaar | 262 | ; * |
15 : | chl | 430 | ; * 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 | 262 | ; * |
20 : | chl | 430 | ; * 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 | 262 | ; * |
24 : | edgomez | 652 | ; * Under section 8 of the GNU General Public License, the copyright |
25 : | ; * holders of XVID explicitly forbid distribution in the following | ||
26 : | ; * countries: | ||
27 : | ; * | ||
28 : | ; * - Japan | ||
29 : | ; * - United States of America | ||
30 : | ; * | ||
31 : | ; * Linking XviD statically or dynamically with other modules is making a | ||
32 : | ; * combined work based on XviD. Thus, the terms and conditions of the | ||
33 : | ; * GNU General Public License cover the whole combination. | ||
34 : | ; * | ||
35 : | ; * As a special exception, the copyright holders of XviD give you | ||
36 : | ; * permission to link XviD with independent modules that communicate with | ||
37 : | ; * XviD solely through the VFW1.1 and DShow interfaces, regardless of the | ||
38 : | ; * license terms of these independent modules, and to copy and distribute | ||
39 : | ; * the resulting combined work under terms of your choice, provided that | ||
40 : | ; * every copy of the combined work is accompanied by a complete copy of | ||
41 : | ; * the source code of XviD (the version of XviD used to produce the | ||
42 : | ; * combined work), being distributed under the terms of the GNU General | ||
43 : | ; * Public License plus this exception. An independent module is a module | ||
44 : | ; * which is not derived from or based on XviD. | ||
45 : | ; * | ||
46 : | ; * Note that people who make modified versions of XviD are not obligated | ||
47 : | ; * to grant this special exception for their modified versions; it is | ||
48 : | ; * their choice whether to do so. The GNU General Public License gives | ||
49 : | ; * permission to release a modified version without this exception; this | ||
50 : | ; * exception also makes it possible to release a modified version which | ||
51 : | ; * carries forward this exception. | ||
52 : | ; * | ||
53 : | ; * $Id: sad_mmx.asm,v 1.10 2002-11-17 00:32:06 edgomez Exp $ | ||
54 : | ; * | ||
55 : | chl | 431 | ; ****************************************************************************/ |
56 : | Isibaar | 262 | |
57 : | bits 32 | ||
58 : | |||
59 : | %macro cglobal 1 | ||
60 : | suxen_drol | 329 | %ifdef PREFIX |
61 : | Isibaar | 262 | global _%1 |
62 : | %define %1 _%1 | ||
63 : | %else | ||
64 : | global %1 | ||
65 : | %endif | ||
66 : | %endmacro | ||
67 : | |||
68 : | section .data | ||
69 : | |||
70 : | align 16 | ||
71 : | mmx_one times 4 dw 1 | ||
72 : | |||
73 : | section .text | ||
74 : | |||
75 : | cglobal sad16_mmx | ||
76 : | cglobal sad8_mmx | ||
77 : | cglobal sad16bi_mmx | ||
78 : | cglobal sad8bi_mmx | ||
79 : | cglobal dev16_mmx | ||
80 : | |||
81 : | ;=========================================================================== | ||
82 : | ; | ||
83 : | ; uint32_t sad16_mmx(const uint8_t * const cur, | ||
84 : | ; const uint8_t * const ref, | ||
85 : | ; const uint32_t stride, | ||
86 : | suxen_drol | 329 | ; const uint32_t best_sad); |
87 : | ; | ||
88 : | Isibaar | 262 | ; (early termination ignore; slows this down) |
89 : | ; | ||
90 : | ;=========================================================================== | ||
91 : | |||
92 : | %macro SAD_16x16_MMX 0 | ||
93 : | movq mm0, [eax] | ||
94 : | movq mm1, [edx] | ||
95 : | |||
96 : | movq mm2, [eax+8] | ||
97 : | movq mm3, [edx+8] | ||
98 : | |||
99 : | movq mm4, mm0 | ||
100 : | psubusb mm0, mm1 | ||
101 : | lea eax,[eax+ecx] | ||
102 : | movq mm5, mm2 | ||
103 : | psubusb mm2, mm3 | ||
104 : | lea edx,[edx+ecx] | ||
105 : | |||
106 : | psubusb mm1, mm4 | ||
107 : | por mm0, mm1 | ||
108 : | psubusb mm3, mm5 | ||
109 : | por mm2, mm3 | ||
110 : | |||
111 : | movq mm1,mm0 | ||
112 : | movq mm3,mm2 | ||
113 : | |||
114 : | punpcklbw mm0,mm7 | ||
115 : | punpckhbw mm1,mm7 | ||
116 : | punpcklbw mm2,mm7 | ||
117 : | punpckhbw mm3,mm7 | ||
118 : | |||
119 : | paddusw mm0,mm1 | ||
120 : | paddusw mm6,mm0 | ||
121 : | paddusw mm2,mm3 | ||
122 : | paddusw mm6,mm2 | ||
123 : | %endmacro | ||
124 : | |||
125 : | align 16 | ||
126 : | sad16_mmx: | ||
127 : | |||
128 : | mov eax, [esp+ 4] ; Src1 | ||
129 : | mov edx, [esp+ 8] ; Src2 | ||
130 : | mov ecx, [esp+12] ; Stride | ||
131 : | |||
132 : | pxor mm6, mm6 ; accum | ||
133 : | pxor mm7, mm7 ; zero | ||
134 : | |||
135 : | SAD_16x16_MMX | ||
136 : | SAD_16x16_MMX | ||
137 : | SAD_16x16_MMX | ||
138 : | SAD_16x16_MMX | ||
139 : | SAD_16x16_MMX | ||
140 : | SAD_16x16_MMX | ||
141 : | SAD_16x16_MMX | ||
142 : | SAD_16x16_MMX | ||
143 : | |||
144 : | SAD_16x16_MMX | ||
145 : | SAD_16x16_MMX | ||
146 : | SAD_16x16_MMX | ||
147 : | SAD_16x16_MMX | ||
148 : | SAD_16x16_MMX | ||
149 : | SAD_16x16_MMX | ||
150 : | SAD_16x16_MMX | ||
151 : | SAD_16x16_MMX | ||
152 : | |||
153 : | pmaddwd mm6, [mmx_one] ; collapse | ||
154 : | movq mm7, mm6 | ||
155 : | psrlq mm7, 32 | ||
156 : | paddd mm6, mm7 | ||
157 : | |||
158 : | movd eax, mm6 | ||
159 : | |||
160 : | ret | ||
161 : | |||
162 : | |||
163 : | ;=========================================================================== | ||
164 : | ; | ||
165 : | ; uint32_t sad8_mmx(const uint8_t * const cur, | ||
166 : | ; const uint8_t * const ref, | ||
167 : | ; const uint32_t stride); | ||
168 : | ; | ||
169 : | ;=========================================================================== | ||
170 : | |||
171 : | %macro SAD_8x8_MMX 0 | ||
172 : | movq mm0, [eax] | ||
173 : | movq mm1, [edx] | ||
174 : | |||
175 : | movq mm2, [eax+ecx] | ||
176 : | movq mm3, [edx+ecx] | ||
177 : | |||
178 : | lea eax,[eax+2*ecx] | ||
179 : | lea edx,[edx+2*ecx] | ||
180 : | |||
181 : | movq mm4, mm0 | ||
182 : | psubusb mm0, mm1 | ||
183 : | movq mm5, mm2 | ||
184 : | psubusb mm2, mm3 | ||
185 : | |||
186 : | psubusb mm1, mm4 | ||
187 : | por mm0, mm1 | ||
188 : | psubusb mm3, mm5 | ||
189 : | por mm2, mm3 | ||
190 : | |||
191 : | movq mm1,mm0 | ||
192 : | movq mm3,mm2 | ||
193 : | |||
194 : | punpcklbw mm0,mm7 | ||
195 : | punpckhbw mm1,mm7 | ||
196 : | punpcklbw mm2,mm7 | ||
197 : | punpckhbw mm3,mm7 | ||
198 : | |||
199 : | paddusw mm0,mm1 | ||
200 : | paddusw mm6,mm0 | ||
201 : | paddusw mm2,mm3 | ||
202 : | paddusw mm6,mm2 | ||
203 : | %endmacro | ||
204 : | |||
205 : | align 16 | ||
206 : | sad8_mmx: | ||
207 : | |||
208 : | mov eax, [esp+ 4] ; Src1 | ||
209 : | mov edx, [esp+ 8] ; Src2 | ||
210 : | mov ecx, [esp+12] ; Stride | ||
211 : | |||
212 : | pxor mm6, mm6 ; accum | ||
213 : | pxor mm7, mm7 ; zero | ||
214 : | |||
215 : | SAD_8x8_MMX | ||
216 : | SAD_8x8_MMX | ||
217 : | SAD_8x8_MMX | ||
218 : | SAD_8x8_MMX | ||
219 : | |||
220 : | pmaddwd mm6, [mmx_one] ; collapse | ||
221 : | movq mm7, mm6 | ||
222 : | psrlq mm7, 32 | ||
223 : | paddd mm6, mm7 | ||
224 : | |||
225 : | movd eax, mm6 | ||
226 : | |||
227 : | ret | ||
228 : | |||
229 : | |||
230 : | |||
231 : | |||
232 : | |||
233 : | ;=========================================================================== | ||
234 : | ; | ||
235 : | ; uint32_t sad16bi_mmx(const uint8_t * const cur, | ||
236 : | ; const uint8_t * const ref1, | ||
237 : | suxen_drol | 329 | ; const uint8_t * const ref2, |
238 : | ; const uint32_t stride); | ||
239 : | ; | ||
240 : | ;=========================================================================== | ||
241 : | %macro SADBI_16x16_MMX 2 ; SADBI_16x16_MMX( int_ptr_offset, bool_increment_ptr ); | ||
242 : | |||
243 : | movq mm0, [edx+%1] | ||
244 : | movq mm2, [ebx+%1] | ||
245 : | movq mm1, mm0 | ||
246 : | movq mm3, mm2 | ||
247 : | |||
248 : | %if %2 != 0 | ||
249 : | add edx, ecx | ||
250 : | %endif | ||
251 : | |||
252 : | punpcklbw mm0, mm7 | ||
253 : | punpckhbw mm1, mm7 | ||
254 : | punpcklbw mm2, mm7 | ||
255 : | punpckhbw mm3, mm7 | ||
256 : | |||
257 : | %if %2 != 0 | ||
258 : | add ebx, ecx | ||
259 : | %endif | ||
260 : | |||
261 : | paddusw mm0, mm2 ; mm01 = ref1 + ref2 | ||
262 : | paddusw mm1, mm3 | ||
263 : | paddusw mm0, [mmx_one] ; mm01 += 1 | ||
264 : | paddusw mm1, [mmx_one] | ||
265 : | psrlw mm0, 1 ; mm01 >>= 1 | ||
266 : | psrlw mm1, 1 | ||
267 : | |||
268 : | movq mm2, [eax+%1] | ||
269 : | movq mm3, mm2 | ||
270 : | punpcklbw mm2, mm7 ; mm23 = src | ||
271 : | punpckhbw mm3, mm7 | ||
272 : | |||
273 : | %if %2 != 0 | ||
274 : | add eax, ecx | ||
275 : | %endif | ||
276 : | |||
277 : | movq mm4, mm0 | ||
278 : | movq mm5, mm1 | ||
279 : | psubusw mm0, mm2 | ||
280 : | psubusw mm1, mm3 | ||
281 : | psubusw mm2, mm4 | ||
282 : | psubusw mm3, mm5 | ||
283 : | por mm0, mm2 ; mm01 = ABS(mm01 - mm23) | ||
284 : | por mm1, mm3 | ||
285 : | |||
286 : | paddusw mm6,mm0 ; mm6 += mm01 | ||
287 : | paddusw mm6,mm1 | ||
288 : | |||
289 : | %endmacro | ||
290 : | |||
291 : | align 16 | ||
292 : | sad16bi_mmx: | ||
293 : | push ebx | ||
294 : | mov eax, [esp+4+ 4] ; Src | ||
295 : | mov edx, [esp+4+ 8] ; Ref1 | ||
296 : | mov ebx, [esp+4+12] ; Ref2 | ||
297 : | mov ecx, [esp+4+16] ; Stride | ||
298 : | |||
299 : | pxor mm6, mm6 ; accum2 | ||
300 : | pxor mm7, mm7 | ||
301 : | .Loop | ||
302 : | SADBI_16x16_MMX 0, 0 | ||
303 : | SADBI_16x16_MMX 8, 1 | ||
304 : | SADBI_16x16_MMX 0, 0 | ||
305 : | SADBI_16x16_MMX 8, 1 | ||
306 : | SADBI_16x16_MMX 0, 0 | ||
307 : | SADBI_16x16_MMX 8, 1 | ||
308 : | SADBI_16x16_MMX 0, 0 | ||
309 : | SADBI_16x16_MMX 8, 1 | ||
310 : | SADBI_16x16_MMX 0, 0 | ||
311 : | SADBI_16x16_MMX 8, 1 | ||
312 : | SADBI_16x16_MMX 0, 0 | ||
313 : | SADBI_16x16_MMX 8, 1 | ||
314 : | SADBI_16x16_MMX 0, 0 | ||
315 : | SADBI_16x16_MMX 8, 1 | ||
316 : | SADBI_16x16_MMX 0, 0 | ||
317 : | SADBI_16x16_MMX 8, 1 | ||
318 : | |||
319 : | SADBI_16x16_MMX 0, 0 | ||
320 : | SADBI_16x16_MMX 8, 1 | ||
321 : | SADBI_16x16_MMX 0, 0 | ||
322 : | SADBI_16x16_MMX 8, 1 | ||
323 : | SADBI_16x16_MMX 0, 0 | ||
324 : | SADBI_16x16_MMX 8, 1 | ||
325 : | SADBI_16x16_MMX 0, 0 | ||
326 : | SADBI_16x16_MMX 8, 1 | ||
327 : | SADBI_16x16_MMX 0, 0 | ||
328 : | SADBI_16x16_MMX 8, 1 | ||
329 : | SADBI_16x16_MMX 0, 0 | ||
330 : | SADBI_16x16_MMX 8, 1 | ||
331 : | SADBI_16x16_MMX 0, 0 | ||
332 : | SADBI_16x16_MMX 8, 1 | ||
333 : | SADBI_16x16_MMX 0, 0 | ||
334 : | SADBI_16x16_MMX 8, 1 | ||
335 : | |||
336 : | pmaddwd mm6, [mmx_one] ; collapse | ||
337 : | movq mm7, mm6 | ||
338 : | psrlq mm7, 32 | ||
339 : | paddd mm6, mm7 | ||
340 : | |||
341 : | movd eax, mm6 | ||
342 : | pop ebx | ||
343 : | ret | ||
344 : | |||
345 : | ;=========================================================================== | ||
346 : | ; | ||
347 : | ; uint32_t sad8bi_mmx(const uint8_t * const cur, | ||
348 : | ; const uint8_t * const ref1, | ||
349 : | ; const uint8_t * const ref2, | ||
350 : | ; const uint32_t stride); | ||
351 : | ; | ||
352 : | ;=========================================================================== | ||
353 : | align 16 | ||
354 : | sad8bi_mmx: | ||
355 : | push ebx | ||
356 : | mov eax, [esp+4+ 4] ; Src | ||
357 : | mov edx, [esp+4+ 8] ; Ref1 | ||
358 : | mov ebx, [esp+4+12] ; Ref2 | ||
359 : | mov ecx, [esp+4+16] ; Stride | ||
360 : | |||
361 : | pxor mm6, mm6 ; accum2 | ||
362 : | pxor mm7, mm7 | ||
363 : | .Loop | ||
364 : | SADBI_16x16_MMX 0, 1 | ||
365 : | SADBI_16x16_MMX 0, 1 | ||
366 : | SADBI_16x16_MMX 0, 1 | ||
367 : | SADBI_16x16_MMX 0, 1 | ||
368 : | SADBI_16x16_MMX 0, 1 | ||
369 : | SADBI_16x16_MMX 0, 1 | ||
370 : | SADBI_16x16_MMX 0, 1 | ||
371 : | SADBI_16x16_MMX 0, 1 | ||
372 : | |||
373 : | pmaddwd mm6, [mmx_one] ; collapse | ||
374 : | movq mm7, mm6 | ||
375 : | psrlq mm7, 32 | ||
376 : | paddd mm6, mm7 | ||
377 : | |||
378 : | movd eax, mm6 | ||
379 : | pop ebx | ||
380 : | ret | ||
381 : | |||
382 : | |||
383 : | |||
384 : | |||
385 : | ;=========================================================================== | ||
386 : | ; | ||
387 : | ; uint32_t dev16_mmx(const uint8_t * const cur, | ||
388 : | ; const uint32_t stride); | ||
389 : | ; | ||
390 : | ;=========================================================================== | ||
391 : | |||
392 : | %macro MEAN_16x16_MMX 0 | ||
393 : | Isibaar | 262 | movq mm0, [eax] |
394 : | movq mm2, [eax+8] | ||
395 : | lea eax,[eax+ecx] | ||
396 : | movq mm1, mm0 | ||
397 : | movq mm3, mm2 | ||
398 : | punpcklbw mm0,mm7 | ||
399 : | punpcklbw mm2,mm7 | ||
400 : | punpckhbw mm1,mm7 | ||
401 : | punpckhbw mm3,mm7 | ||
402 : | paddw mm5, mm0 | ||
403 : | paddw mm6, mm1 | ||
404 : | paddw mm5, mm2 | ||
405 : | paddw mm6, mm3 | ||
406 : | %endmacro | ||
407 : | |||
408 : | %macro ABS_16x16_MMX 0 | ||
409 : | movq mm0, [eax] | ||
410 : | movq mm2, [eax+8] | ||
411 : | lea eax,[eax+ecx] | ||
412 : | movq mm1, mm0 | ||
413 : | movq mm3, mm2 | ||
414 : | punpcklbw mm0, mm7 | ||
415 : | punpcklbw mm2, mm7 | ||
416 : | punpckhbw mm1, mm7 | ||
417 : | punpckhbw mm3, mm7 | ||
418 : | movq mm4, mm6 | ||
419 : | psubusw mm4, mm0 | ||
420 : | |||
421 : | psubusw mm0, mm6 | ||
422 : | por mm0, mm4 | ||
423 : | movq mm4, mm6 | ||
424 : | psubusw mm4, mm1 | ||
425 : | psubusw mm1, mm6 | ||
426 : | por mm1, mm4 | ||
427 : | |||
428 : | movq mm4, mm6 | ||
429 : | psubusw mm4, mm2 | ||
430 : | psubusw mm2, mm6 | ||
431 : | por mm2, mm4 | ||
432 : | movq mm4, mm6 | ||
433 : | psubusw mm4, mm3 | ||
434 : | psubusw mm3, mm6 | ||
435 : | por mm3, mm4 | ||
436 : | |||
437 : | paddw mm0, mm1 | ||
438 : | paddw mm2, mm3 | ||
439 : | paddw mm5, mm0 | ||
440 : | paddw mm5, mm2 | ||
441 : | %endmacro | ||
442 : | |||
443 : | align 16 | ||
444 : | dev16_mmx: | ||
445 : | mov eax, [esp+ 4] ; Src | ||
446 : | mov ecx, [esp+ 8] ; Stride | ||
447 : | |||
448 : | pxor mm7, mm7 ; zero | ||
449 : | pxor mm5, mm5 ; accum1 | ||
450 : | pxor mm6, mm6 ; accum2 | ||
451 : | |||
452 : | MEAN_16x16_MMX | ||
453 : | MEAN_16x16_MMX | ||
454 : | MEAN_16x16_MMX | ||
455 : | MEAN_16x16_MMX | ||
456 : | MEAN_16x16_MMX | ||
457 : | MEAN_16x16_MMX | ||
458 : | MEAN_16x16_MMX | ||
459 : | MEAN_16x16_MMX | ||
460 : | |||
461 : | MEAN_16x16_MMX | ||
462 : | MEAN_16x16_MMX | ||
463 : | MEAN_16x16_MMX | ||
464 : | MEAN_16x16_MMX | ||
465 : | MEAN_16x16_MMX | ||
466 : | MEAN_16x16_MMX | ||
467 : | MEAN_16x16_MMX | ||
468 : | MEAN_16x16_MMX | ||
469 : | |||
470 : | paddusw mm6, mm5 | ||
471 : | pmaddwd mm6, [mmx_one] ; collapse | ||
472 : | movq mm5, mm6 | ||
473 : | psrlq mm5, 32 | ||
474 : | paddd mm6, mm5 | ||
475 : | |||
476 : | psllq mm6, 32 ; blank upper dword | ||
477 : | psrlq mm6, 32 + 8 ; /= (16*16) | ||
478 : | |||
479 : | punpckldq mm6, mm6 | ||
480 : | packssdw mm6, mm6 | ||
481 : | |||
482 : | ; mm6 contains the mean | ||
483 : | ; mm5 is the new accum | ||
484 : | |||
485 : | pxor mm5, mm5 | ||
486 : | mov eax, [esp+ 4] ; Src | ||
487 : | |||
488 : | ABS_16x16_MMX | ||
489 : | ABS_16x16_MMX | ||
490 : | ABS_16x16_MMX | ||
491 : | ABS_16x16_MMX | ||
492 : | ABS_16x16_MMX | ||
493 : | ABS_16x16_MMX | ||
494 : | ABS_16x16_MMX | ||
495 : | ABS_16x16_MMX | ||
496 : | |||
497 : | ABS_16x16_MMX | ||
498 : | ABS_16x16_MMX | ||
499 : | ABS_16x16_MMX | ||
500 : | ABS_16x16_MMX | ||
501 : | ABS_16x16_MMX | ||
502 : | ABS_16x16_MMX | ||
503 : | ABS_16x16_MMX | ||
504 : | ABS_16x16_MMX | ||
505 : | |||
506 : | pmaddwd mm5, [mmx_one] ; collapse | ||
507 : | movq mm6, mm5 | ||
508 : | psrlq mm6, 32 | ||
509 : | paddd mm6, mm5 | ||
510 : | |||
511 : | movd eax, mm6 | ||
512 : | chl | 430 | ret |
513 : |
No admin address has been configured | ViewVC Help |
Powered by ViewVC 1.0.4 |