2
* SIMD-optimized motion estimation
3
* Copyright (c) 2000, 2001 Fabrice Bellard
4
* Copyright (c) 2002-2004 Michael Niedermayer <michaelni@gmx.at>
6
* MMX optimization by Nick Kurshev <nickols_k@mail.ru>
8
* This file is part of Libav.
10
* Libav is free software; you can redistribute it and/or
11
* modify it under the terms of the GNU Lesser General Public
12
* License as published by the Free Software Foundation; either
13
* version 2.1 of the License, or (at your option) any later version.
15
* Libav 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 GNU
18
* Lesser General Public License for more details.
20
* You should have received a copy of the GNU Lesser General Public
21
* License along with Libav; if not, write to the Free Software
22
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
25
#include "libavutil/attributes.h"
26
#include "libavutil/cpu.h"
27
#include "libavutil/x86/asm.h"
28
#include "libavutil/x86/cpu.h"
29
#include "libavcodec/me_cmp.h"
30
#include "libavcodec/mpegvideo.h"
34
static int sse8_mmx(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
42
"pxor %%mm0, %%mm0 \n" /* mm0 = 0 */
43
"pxor %%mm7, %%mm7 \n" /* mm7 holds the sum */
45
"movq (%0), %%mm1 \n" /* mm1 = pix1[0][0 - 7] */
46
"movq (%1), %%mm2 \n" /* mm2 = pix2[0][0 - 7] */
47
"movq (%0, %3), %%mm3 \n" /* mm3 = pix1[1][0 - 7] */
48
"movq (%1, %3), %%mm4 \n" /* mm4 = pix2[1][0 - 7] */
50
/* todo: mm1-mm2, mm3-mm4 */
51
/* algo: subtract mm1 from mm2 with saturation and vice versa */
52
/* OR the results to get absolute difference */
53
"movq %%mm1, %%mm5 \n"
54
"movq %%mm3, %%mm6 \n"
55
"psubusb %%mm2, %%mm1 \n"
56
"psubusb %%mm4, %%mm3 \n"
57
"psubusb %%mm5, %%mm2 \n"
58
"psubusb %%mm6, %%mm4 \n"
63
/* now convert to 16-bit vectors so we can square them */
64
"movq %%mm2, %%mm1 \n"
65
"movq %%mm4, %%mm3 \n"
67
"punpckhbw %%mm0, %%mm2 \n"
68
"punpckhbw %%mm0, %%mm4 \n"
69
"punpcklbw %%mm0, %%mm1 \n" /* mm1 now spread over (mm1, mm2) */
70
"punpcklbw %%mm0, %%mm3 \n" /* mm4 now spread over (mm3, mm4) */
72
"pmaddwd %%mm2, %%mm2 \n"
73
"pmaddwd %%mm4, %%mm4 \n"
74
"pmaddwd %%mm1, %%mm1 \n"
75
"pmaddwd %%mm3, %%mm3 \n"
77
"lea (%0, %3, 2), %0 \n" /* pix1 += 2 * line_size */
78
"lea (%1, %3, 2), %1 \n" /* pix2 += 2 * line_size */
80
"paddd %%mm2, %%mm1 \n"
81
"paddd %%mm4, %%mm3 \n"
82
"paddd %%mm1, %%mm7 \n"
83
"paddd %%mm3, %%mm7 \n"
88
"movq %%mm7, %%mm1 \n"
89
"psrlq $32, %%mm7 \n" /* shift hi dword to lo */
90
"paddd %%mm7, %%mm1 \n"
92
: "+r" (pix1), "+r" (pix2), "=r" (tmp)
93
: "r" ((x86_reg) line_size), "m" (h)
99
static int sse16_mmx(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
100
int line_size, int h)
106
"pxor %%mm0, %%mm0\n" /* mm0 = 0 */
107
"pxor %%mm7, %%mm7\n" /* mm7 holds the sum */
109
"movq (%0), %%mm1\n" /* mm1 = pix1[0 - 7] */
110
"movq (%1), %%mm2\n" /* mm2 = pix2[0 - 7] */
111
"movq 8(%0), %%mm3\n" /* mm3 = pix1[8 - 15] */
112
"movq 8(%1), %%mm4\n" /* mm4 = pix2[8 - 15] */
114
/* todo: mm1-mm2, mm3-mm4 */
115
/* algo: subtract mm1 from mm2 with saturation and vice versa */
116
/* OR the results to get absolute difference */
117
"movq %%mm1, %%mm5\n"
118
"movq %%mm3, %%mm6\n"
119
"psubusb %%mm2, %%mm1\n"
120
"psubusb %%mm4, %%mm3\n"
121
"psubusb %%mm5, %%mm2\n"
122
"psubusb %%mm6, %%mm4\n"
127
/* now convert to 16-bit vectors so we can square them */
128
"movq %%mm2, %%mm1\n"
129
"movq %%mm4, %%mm3\n"
131
"punpckhbw %%mm0, %%mm2\n"
132
"punpckhbw %%mm0, %%mm4\n"
133
"punpcklbw %%mm0, %%mm1\n" /* mm1 now spread over (mm1, mm2) */
134
"punpcklbw %%mm0, %%mm3\n" /* mm4 now spread over (mm3, mm4) */
136
"pmaddwd %%mm2, %%mm2\n"
137
"pmaddwd %%mm4, %%mm4\n"
138
"pmaddwd %%mm1, %%mm1\n"
139
"pmaddwd %%mm3, %%mm3\n"
144
"paddd %%mm2, %%mm1\n"
145
"paddd %%mm4, %%mm3\n"
146
"paddd %%mm1, %%mm7\n"
147
"paddd %%mm3, %%mm7\n"
152
"movq %%mm7, %%mm1\n"
153
"psrlq $32, %%mm7\n" /* shift hi dword to lo */
154
"paddd %%mm7, %%mm1\n"
156
: "+r" (pix1), "+r" (pix2), "=r" (tmp)
157
: "r" ((x86_reg) line_size), "m" (h)
163
static int hf_noise8_mmx(uint8_t *pix1, int line_size, int h)
169
"pxor %%mm7, %%mm7\n"
170
"pxor %%mm6, %%mm6\n"
173
"movq %%mm0, %%mm1\n"
177
"movq %%mm0, %%mm2\n"
178
"movq %%mm1, %%mm3\n"
179
"punpcklbw %%mm7, %%mm0\n"
180
"punpcklbw %%mm7, %%mm1\n"
181
"punpckhbw %%mm7, %%mm2\n"
182
"punpckhbw %%mm7, %%mm3\n"
183
"psubw %%mm1, %%mm0\n"
184
"psubw %%mm3, %%mm2\n"
189
"movq %%mm4, %%mm1\n"
193
"movq %%mm4, %%mm5\n"
194
"movq %%mm1, %%mm3\n"
195
"punpcklbw %%mm7, %%mm4\n"
196
"punpcklbw %%mm7, %%mm1\n"
197
"punpckhbw %%mm7, %%mm5\n"
198
"punpckhbw %%mm7, %%mm3\n"
199
"psubw %%mm1, %%mm4\n"
200
"psubw %%mm3, %%mm5\n"
201
"psubw %%mm4, %%mm0\n"
202
"psubw %%mm5, %%mm2\n"
203
"pxor %%mm3, %%mm3\n"
204
"pxor %%mm1, %%mm1\n"
205
"pcmpgtw %%mm0, %%mm3\n\t"
206
"pcmpgtw %%mm2, %%mm1\n\t"
207
"pxor %%mm3, %%mm0\n"
208
"pxor %%mm1, %%mm2\n"
209
"psubw %%mm3, %%mm0\n"
210
"psubw %%mm1, %%mm2\n"
211
"paddw %%mm0, %%mm2\n"
212
"paddw %%mm2, %%mm6\n"
218
"movq %%mm0, %%mm1\n"
222
"movq %%mm0, %%mm2\n"
223
"movq %%mm1, %%mm3\n"
224
"punpcklbw %%mm7, %%mm0\n"
225
"punpcklbw %%mm7, %%mm1\n"
226
"punpckhbw %%mm7, %%mm2\n"
227
"punpckhbw %%mm7, %%mm3\n"
228
"psubw %%mm1, %%mm0\n"
229
"psubw %%mm3, %%mm2\n"
230
"psubw %%mm0, %%mm4\n"
231
"psubw %%mm2, %%mm5\n"
232
"pxor %%mm3, %%mm3\n"
233
"pxor %%mm1, %%mm1\n"
234
"pcmpgtw %%mm4, %%mm3\n\t"
235
"pcmpgtw %%mm5, %%mm1\n\t"
236
"pxor %%mm3, %%mm4\n"
237
"pxor %%mm1, %%mm5\n"
238
"psubw %%mm3, %%mm4\n"
239
"psubw %%mm1, %%mm5\n"
240
"paddw %%mm4, %%mm5\n"
241
"paddw %%mm5, %%mm6\n"
246
"movq %%mm4, %%mm1\n"
250
"movq %%mm4, %%mm5\n"
251
"movq %%mm1, %%mm3\n"
252
"punpcklbw %%mm7, %%mm4\n"
253
"punpcklbw %%mm7, %%mm1\n"
254
"punpckhbw %%mm7, %%mm5\n"
255
"punpckhbw %%mm7, %%mm3\n"
256
"psubw %%mm1, %%mm4\n"
257
"psubw %%mm3, %%mm5\n"
258
"psubw %%mm4, %%mm0\n"
259
"psubw %%mm5, %%mm2\n"
260
"pxor %%mm3, %%mm3\n"
261
"pxor %%mm1, %%mm1\n"
262
"pcmpgtw %%mm0, %%mm3\n\t"
263
"pcmpgtw %%mm2, %%mm1\n\t"
264
"pxor %%mm3, %%mm0\n"
265
"pxor %%mm1, %%mm2\n"
266
"psubw %%mm3, %%mm0\n"
267
"psubw %%mm1, %%mm2\n"
268
"paddw %%mm0, %%mm2\n"
269
"paddw %%mm2, %%mm6\n"
275
"movq %%mm6, %%mm0\n"
276
"punpcklwd %%mm7, %%mm0\n"
277
"punpckhwd %%mm7, %%mm6\n"
278
"paddd %%mm0, %%mm6\n"
280
"movq %%mm6, %%mm0\n"
282
"paddd %%mm6, %%mm0\n"
284
: "+r" (pix1), "=r" (tmp)
285
: "r" ((x86_reg) line_size), "g" (h - 2)
291
static int hf_noise16_mmx(uint8_t *pix1, int line_size, int h)
298
"pxor %%mm7, %%mm7\n"
299
"pxor %%mm6, %%mm6\n"
302
"movq 1(%0), %%mm1\n"
303
"movq %%mm0, %%mm2\n"
304
"movq %%mm1, %%mm3\n"
305
"punpcklbw %%mm7, %%mm0\n"
306
"punpcklbw %%mm7, %%mm1\n"
307
"punpckhbw %%mm7, %%mm2\n"
308
"punpckhbw %%mm7, %%mm3\n"
309
"psubw %%mm1, %%mm0\n"
310
"psubw %%mm3, %%mm2\n"
315
"movq 1(%0), %%mm1\n"
316
"movq %%mm4, %%mm5\n"
317
"movq %%mm1, %%mm3\n"
318
"punpcklbw %%mm7, %%mm4\n"
319
"punpcklbw %%mm7, %%mm1\n"
320
"punpckhbw %%mm7, %%mm5\n"
321
"punpckhbw %%mm7, %%mm3\n"
322
"psubw %%mm1, %%mm4\n"
323
"psubw %%mm3, %%mm5\n"
324
"psubw %%mm4, %%mm0\n"
325
"psubw %%mm5, %%mm2\n"
326
"pxor %%mm3, %%mm3\n"
327
"pxor %%mm1, %%mm1\n"
328
"pcmpgtw %%mm0, %%mm3\n\t"
329
"pcmpgtw %%mm2, %%mm1\n\t"
330
"pxor %%mm3, %%mm0\n"
331
"pxor %%mm1, %%mm2\n"
332
"psubw %%mm3, %%mm0\n"
333
"psubw %%mm1, %%mm2\n"
334
"paddw %%mm0, %%mm2\n"
335
"paddw %%mm2, %%mm6\n"
341
"movq 1(%0), %%mm1\n"
342
"movq %%mm0, %%mm2\n"
343
"movq %%mm1, %%mm3\n"
344
"punpcklbw %%mm7, %%mm0\n"
345
"punpcklbw %%mm7, %%mm1\n"
346
"punpckhbw %%mm7, %%mm2\n"
347
"punpckhbw %%mm7, %%mm3\n"
348
"psubw %%mm1, %%mm0\n"
349
"psubw %%mm3, %%mm2\n"
350
"psubw %%mm0, %%mm4\n"
351
"psubw %%mm2, %%mm5\n"
352
"pxor %%mm3, %%mm3\n"
353
"pxor %%mm1, %%mm1\n"
354
"pcmpgtw %%mm4, %%mm3\n\t"
355
"pcmpgtw %%mm5, %%mm1\n\t"
356
"pxor %%mm3, %%mm4\n"
357
"pxor %%mm1, %%mm5\n"
358
"psubw %%mm3, %%mm4\n"
359
"psubw %%mm1, %%mm5\n"
360
"paddw %%mm4, %%mm5\n"
361
"paddw %%mm5, %%mm6\n"
366
"movq 1(%0), %%mm1\n"
367
"movq %%mm4, %%mm5\n"
368
"movq %%mm1, %%mm3\n"
369
"punpcklbw %%mm7, %%mm4\n"
370
"punpcklbw %%mm7, %%mm1\n"
371
"punpckhbw %%mm7, %%mm5\n"
372
"punpckhbw %%mm7, %%mm3\n"
373
"psubw %%mm1, %%mm4\n"
374
"psubw %%mm3, %%mm5\n"
375
"psubw %%mm4, %%mm0\n"
376
"psubw %%mm5, %%mm2\n"
377
"pxor %%mm3, %%mm3\n"
378
"pxor %%mm1, %%mm1\n"
379
"pcmpgtw %%mm0, %%mm3\n\t"
380
"pcmpgtw %%mm2, %%mm1\n\t"
381
"pxor %%mm3, %%mm0\n"
382
"pxor %%mm1, %%mm2\n"
383
"psubw %%mm3, %%mm0\n"
384
"psubw %%mm1, %%mm2\n"
385
"paddw %%mm0, %%mm2\n"
386
"paddw %%mm2, %%mm6\n"
392
"movq %%mm6, %%mm0\n"
393
"punpcklwd %%mm7, %%mm0\n"
394
"punpckhwd %%mm7, %%mm6\n"
395
"paddd %%mm0, %%mm6\n"
397
"movq %%mm6, %%mm0\n"
399
"paddd %%mm6, %%mm0\n"
401
: "+r" (pix1), "=r" (tmp)
402
: "r" ((x86_reg) line_size), "g" (h - 2)
405
return tmp + hf_noise8_mmx(pix + 8, line_size, h);
408
static int nsse16_mmx(MpegEncContext *c, uint8_t *pix1, uint8_t *pix2,
409
int line_size, int h)
414
score1 = c->mecc.sse[0](c, pix1, pix2, line_size, h);
416
score1 = sse16_mmx(c, pix1, pix2, line_size, h);
417
score2 = hf_noise16_mmx(pix1, line_size, h) -
418
hf_noise16_mmx(pix2, line_size, h);
421
return score1 + FFABS(score2) * c->avctx->nsse_weight;
423
return score1 + FFABS(score2) * 8;
426
static int nsse8_mmx(MpegEncContext *c, uint8_t *pix1, uint8_t *pix2,
427
int line_size, int h)
429
int score1 = sse8_mmx(c, pix1, pix2, line_size, h);
430
int score2 = hf_noise8_mmx(pix1, line_size, h) -
431
hf_noise8_mmx(pix2, line_size, h);
434
return score1 + FFABS(score2) * c->avctx->nsse_weight;
436
return score1 + FFABS(score2) * 8;
439
static int vsad_intra16_mmx(MpegEncContext *v, uint8_t *pix, uint8_t *dummy,
440
int line_size, int h)
444
assert((((int) pix) & 7) == 0);
445
assert((line_size & 7) == 0);
447
#define SUM(in0, in1, out0, out1) \
448
"movq (%0), %%mm2\n" \
449
"movq 8(%0), %%mm3\n" \
451
"movq %%mm2, " #out0 "\n" \
452
"movq %%mm3, " #out1 "\n" \
453
"psubusb " #in0 ", %%mm2\n" \
454
"psubusb " #in1 ", %%mm3\n" \
455
"psubusb " #out0 ", " #in0 "\n" \
456
"psubusb " #out1 ", " #in1 "\n" \
457
"por %%mm2, " #in0 "\n" \
458
"por %%mm3, " #in1 "\n" \
459
"movq " #in0 ", %%mm2\n" \
460
"movq " #in1 ", %%mm3\n" \
461
"punpcklbw %%mm7, " #in0 "\n" \
462
"punpcklbw %%mm7, " #in1 "\n" \
463
"punpckhbw %%mm7, %%mm2\n" \
464
"punpckhbw %%mm7, %%mm3\n" \
465
"paddw " #in1 ", " #in0 "\n" \
466
"paddw %%mm3, %%mm2\n" \
467
"paddw %%mm2, " #in0 "\n" \
468
"paddw " #in0 ", %%mm6\n"
473
"pxor %%mm6, %%mm6\n"
474
"pxor %%mm7, %%mm7\n"
476
"movq 8(%0), %%mm1\n"
481
SUM(%%mm4, %%mm5, %%mm0, %%mm1)
483
SUM(%%mm0, %%mm1, %%mm4, %%mm5)
488
"movq %%mm6, %%mm0\n"
490
"paddw %%mm6, %%mm0\n"
491
"movq %%mm0, %%mm6\n"
493
"paddw %%mm6, %%mm0\n"
495
: "+r" (pix), "=r" (tmp)
496
: "r" ((x86_reg) line_size), "m" (h)
503
static int vsad_intra16_mmxext(MpegEncContext *v, uint8_t *pix, uint8_t *dummy,
504
int line_size, int h)
508
assert((((int) pix) & 7) == 0);
509
assert((line_size & 7) == 0);
511
#define SUM(in0, in1, out0, out1) \
512
"movq (%0), " #out0 "\n" \
513
"movq 8(%0), " #out1 "\n" \
515
"psadbw " #out0 ", " #in0 "\n" \
516
"psadbw " #out1 ", " #in1 "\n" \
517
"paddw " #in1 ", " #in0 "\n" \
518
"paddw " #in0 ", %%mm6\n"
522
"pxor %%mm6, %%mm6\n"
523
"pxor %%mm7, %%mm7\n"
525
"movq 8(%0), %%mm1\n"
530
SUM(%%mm4, %%mm5, %%mm0, %%mm1)
532
SUM(%%mm0, %%mm1, %%mm4, %%mm5)
538
: "+r" (pix), "=r" (tmp)
539
: "r" ((x86_reg) line_size), "m" (h)
546
static int vsad16_mmx(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
547
int line_size, int h)
551
assert((((int) pix1) & 7) == 0);
552
assert((((int) pix2) & 7) == 0);
553
assert((line_size & 7) == 0);
555
#define SUM(in0, in1, out0, out1) \
556
"movq (%0), %%mm2\n" \
557
"movq (%1), " #out0 "\n" \
558
"movq 8(%0), %%mm3\n" \
559
"movq 8(%1), " #out1 "\n" \
562
"psubb " #out0 ", %%mm2\n" \
563
"psubb " #out1 ", %%mm3\n" \
564
"pxor %%mm7, %%mm2\n" \
565
"pxor %%mm7, %%mm3\n" \
566
"movq %%mm2, " #out0 "\n" \
567
"movq %%mm3, " #out1 "\n" \
568
"psubusb " #in0 ", %%mm2\n" \
569
"psubusb " #in1 ", %%mm3\n" \
570
"psubusb " #out0 ", " #in0 "\n" \
571
"psubusb " #out1 ", " #in1 "\n" \
572
"por %%mm2, " #in0 "\n" \
573
"por %%mm3, " #in1 "\n" \
574
"movq " #in0 ", %%mm2\n" \
575
"movq " #in1 ", %%mm3\n" \
576
"punpcklbw %%mm7, " #in0 "\n" \
577
"punpcklbw %%mm7, " #in1 "\n" \
578
"punpckhbw %%mm7, %%mm2\n" \
579
"punpckhbw %%mm7, %%mm3\n" \
580
"paddw " #in1 ", " #in0 "\n" \
581
"paddw %%mm3, %%mm2\n" \
582
"paddw %%mm2, " #in0 "\n" \
583
"paddw " #in0 ", %%mm6\n"
588
"pxor %%mm6, %%mm6\n"
589
"pcmpeqw %%mm7, %%mm7\n"
591
"packsswb %%mm7, %%mm7\n"
594
"movq 8(%0), %%mm1\n"
595
"movq 8(%1), %%mm3\n"
598
"psubb %%mm2, %%mm0\n"
599
"psubb %%mm3, %%mm1\n"
600
"pxor %%mm7, %%mm0\n"
601
"pxor %%mm7, %%mm1\n"
605
SUM(%%mm4, %%mm5, %%mm0, %%mm1)
607
SUM(%%mm0, %%mm1, %%mm4, %%mm5)
612
"movq %%mm6, %%mm0\n"
614
"paddw %%mm6, %%mm0\n"
615
"movq %%mm0, %%mm6\n"
617
"paddw %%mm6, %%mm0\n"
619
: "+r" (pix1), "+r" (pix2), "=r" (tmp)
620
: "r" ((x86_reg) line_size), "m" (h)
627
static int vsad16_mmxext(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
628
int line_size, int h)
632
assert((((int) pix1) & 7) == 0);
633
assert((((int) pix2) & 7) == 0);
634
assert((line_size & 7) == 0);
636
#define SUM(in0, in1, out0, out1) \
637
"movq (%0), " #out0 "\n" \
638
"movq (%1), %%mm2\n" \
639
"movq 8(%0), " #out1 "\n" \
640
"movq 8(%1), %%mm3\n" \
643
"psubb %%mm2, " #out0 "\n" \
644
"psubb %%mm3, " #out1 "\n" \
645
"pxor %%mm7, " #out0 "\n" \
646
"pxor %%mm7, " #out1 "\n" \
647
"psadbw " #out0 ", " #in0 "\n" \
648
"psadbw " #out1 ", " #in1 "\n" \
649
"paddw " #in1 ", " #in0 "\n" \
650
"paddw " #in0 ", %%mm6\n "
654
"pxor %%mm6, %%mm6\n"
655
"pcmpeqw %%mm7, %%mm7\n"
657
"packsswb %%mm7, %%mm7\n"
660
"movq 8(%0), %%mm1\n"
661
"movq 8(%1), %%mm3\n"
664
"psubb %%mm2, %%mm0\n"
665
"psubb %%mm3, %%mm1\n"
666
"pxor %%mm7, %%mm0\n"
667
"pxor %%mm7, %%mm1\n"
671
SUM(%%mm4, %%mm5, %%mm0, %%mm1)
673
SUM(%%mm0, %%mm1, %%mm4, %%mm5)
679
: "+r" (pix1), "+r" (pix2), "=r" (tmp)
680
: "r" ((x86_reg) line_size), "m" (h)
687
#define MMABS_MMX(a,z) \
688
"pxor " #z ", " #z " \n\t" \
689
"pcmpgtw " #a ", " #z " \n\t" \
690
"pxor " #z ", " #a " \n\t" \
691
"psubw " #z ", " #a " \n\t"
693
#define MMABS_MMXEXT(a, z) \
694
"pxor " #z ", " #z " \n\t" \
695
"psubw " #a ", " #z " \n\t" \
696
"pmaxsw " #z ", " #a " \n\t"
698
#define MMABS_SSSE3(a,z) \
699
"pabsw " #a ", " #a " \n\t"
701
#define MMABS_SUM(a,z, sum) \
703
"paddusw " #a ", " #sum " \n\t"
705
/* FIXME: HSUM_* saturates at 64k, while an 8x8 hadamard or dct block can get
706
* up to about 100k on extreme inputs. But that's very unlikely to occur in
707
* natural video, and it's even more unlikely to not have any alternative
708
* mvs/modes with lower cost. */
709
#define HSUM_MMX(a, t, dst) \
710
"movq " #a ", " #t " \n\t" \
711
"psrlq $32, " #a " \n\t" \
712
"paddusw " #t ", " #a " \n\t" \
713
"movq " #a ", " #t " \n\t" \
714
"psrlq $16, " #a " \n\t" \
715
"paddusw " #t ", " #a " \n\t" \
716
"movd " #a ", " #dst " \n\t" \
718
#define HSUM_MMXEXT(a, t, dst) \
719
"pshufw $0x0E, " #a ", " #t " \n\t" \
720
"paddusw " #t ", " #a " \n\t" \
721
"pshufw $0x01, " #a ", " #t " \n\t" \
722
"paddusw " #t ", " #a " \n\t" \
723
"movd " #a ", " #dst " \n\t" \
725
#define HSUM_SSE2(a, t, dst) \
726
"movhlps " #a ", " #t " \n\t" \
727
"paddusw " #t ", " #a " \n\t" \
728
"pshuflw $0x0E, " #a ", " #t " \n\t" \
729
"paddusw " #t ", " #a " \n\t" \
730
"pshuflw $0x01, " #a ", " #t " \n\t" \
731
"paddusw " #t ", " #a " \n\t" \
732
"movd " #a ", " #dst " \n\t" \
734
#define DCT_SAD4(m, mm, o) \
735
"mov"#m" "#o" + 0(%1), " #mm "2 \n\t" \
736
"mov"#m" "#o" + 16(%1), " #mm "3 \n\t" \
737
"mov"#m" "#o" + 32(%1), " #mm "4 \n\t" \
738
"mov"#m" "#o" + 48(%1), " #mm "5 \n\t" \
739
MMABS_SUM(mm ## 2, mm ## 6, mm ## 0) \
740
MMABS_SUM(mm ## 3, mm ## 7, mm ## 1) \
741
MMABS_SUM(mm ## 4, mm ## 6, mm ## 0) \
742
MMABS_SUM(mm ## 5, mm ## 7, mm ## 1) \
744
#define DCT_SAD_MMX \
745
"pxor %%mm0, %%mm0 \n\t" \
746
"pxor %%mm1, %%mm1 \n\t" \
747
DCT_SAD4(q, %%mm, 0) \
748
DCT_SAD4(q, %%mm, 8) \
749
DCT_SAD4(q, %%mm, 64) \
750
DCT_SAD4(q, %%mm, 72) \
751
"paddusw %%mm1, %%mm0 \n\t" \
752
HSUM(%%mm0, %%mm1, %0)
754
#define DCT_SAD_SSE2 \
755
"pxor %%xmm0, %%xmm0 \n\t" \
756
"pxor %%xmm1, %%xmm1 \n\t" \
757
DCT_SAD4(dqa, %%xmm, 0) \
758
DCT_SAD4(dqa, %%xmm, 64) \
759
"paddusw %%xmm1, %%xmm0 \n\t" \
760
HSUM(%%xmm0, %%xmm1, %0)
762
#define DCT_SAD_FUNC(cpu) \
763
static int sum_abs_dctelem_ ## cpu(int16_t *block) \
770
return sum & 0xFFFF; \
773
#define DCT_SAD DCT_SAD_MMX
774
#define HSUM(a, t, dst) HSUM_MMX(a, t, dst)
775
#define MMABS(a, z) MMABS_MMX(a, z)
780
#define HSUM(a, t, dst) HSUM_MMXEXT(a, t, dst)
781
#define MMABS(a, z) MMABS_MMXEXT(a, z)
786
#define DCT_SAD DCT_SAD_SSE2
787
#define HSUM(a, t, dst) HSUM_SSE2(a, t, dst)
791
#if HAVE_SSSE3_INLINE
792
#define MMABS(a, z) MMABS_SSSE3(a, z)
800
DECLARE_ASM_CONST(8, uint64_t, round_tab)[3] = {
801
0x0000000000000000ULL,
802
0x0001000100010001ULL,
803
0x0002000200020002ULL,
806
DECLARE_ASM_CONST(8, uint64_t, bone) = 0x0101010101010101LL;
808
static inline void sad8_1_mmx(uint8_t *blk1, uint8_t *blk2, int stride, int h)
810
x86_reg len = -(stride * h);
814
"movq (%1, %%"REG_a"), %%mm0 \n\t"
815
"movq (%2, %%"REG_a"), %%mm2 \n\t"
816
"movq (%2, %%"REG_a"), %%mm4 \n\t"
817
"add %3, %%"REG_a" \n\t"
818
"psubusb %%mm0, %%mm2 \n\t"
819
"psubusb %%mm4, %%mm0 \n\t"
820
"movq (%1, %%"REG_a"), %%mm1 \n\t"
821
"movq (%2, %%"REG_a"), %%mm3 \n\t"
822
"movq (%2, %%"REG_a"), %%mm5 \n\t"
823
"psubusb %%mm1, %%mm3 \n\t"
824
"psubusb %%mm5, %%mm1 \n\t"
825
"por %%mm2, %%mm0 \n\t"
826
"por %%mm1, %%mm3 \n\t"
827
"movq %%mm0, %%mm1 \n\t"
828
"movq %%mm3, %%mm2 \n\t"
829
"punpcklbw %%mm7, %%mm0 \n\t"
830
"punpckhbw %%mm7, %%mm1 \n\t"
831
"punpcklbw %%mm7, %%mm3 \n\t"
832
"punpckhbw %%mm7, %%mm2 \n\t"
833
"paddw %%mm1, %%mm0 \n\t"
834
"paddw %%mm3, %%mm2 \n\t"
835
"paddw %%mm2, %%mm0 \n\t"
836
"paddw %%mm0, %%mm6 \n\t"
837
"add %3, %%"REG_a" \n\t"
840
: "r" (blk1 - len), "r" (blk2 - len), "r" ((x86_reg) stride));
843
static inline void sad8_1_mmxext(uint8_t *blk1, uint8_t *blk2,
849
"movq (%1), %%mm0 \n\t"
850
"movq (%1, %3), %%mm1 \n\t"
851
"psadbw (%2), %%mm0 \n\t"
852
"psadbw (%2, %3), %%mm1 \n\t"
853
"paddw %%mm0, %%mm6 \n\t"
854
"paddw %%mm1, %%mm6 \n\t"
855
"lea (%1,%3,2), %1 \n\t"
856
"lea (%2,%3,2), %2 \n\t"
859
: "+r" (h), "+r" (blk1), "+r" (blk2)
860
: "r" ((x86_reg) stride));
863
static int sad16_sse2(MpegEncContext *v, uint8_t *blk2, uint8_t *blk1,
868
"pxor %%xmm2, %%xmm2 \n\t"
871
"movdqu (%1), %%xmm0 \n\t"
872
"movdqu (%1, %4), %%xmm1 \n\t"
873
"psadbw (%2), %%xmm0 \n\t"
874
"psadbw (%2, %4), %%xmm1 \n\t"
875
"paddw %%xmm0, %%xmm2 \n\t"
876
"paddw %%xmm1, %%xmm2 \n\t"
877
"lea (%1,%4,2), %1 \n\t"
878
"lea (%2,%4,2), %2 \n\t"
881
"movhlps %%xmm2, %%xmm0 \n\t"
882
"paddw %%xmm0, %%xmm2 \n\t"
883
"movd %%xmm2, %3 \n\t"
884
: "+r" (h), "+r" (blk1), "+r" (blk2), "=r" (ret)
885
: "r" ((x86_reg) stride));
889
static inline void sad8_x2a_mmxext(uint8_t *blk1, uint8_t *blk2,
895
"movq (%1), %%mm0 \n\t"
896
"movq (%1, %3), %%mm1 \n\t"
897
"pavgb 1(%1), %%mm0 \n\t"
898
"pavgb 1(%1, %3), %%mm1 \n\t"
899
"psadbw (%2), %%mm0 \n\t"
900
"psadbw (%2, %3), %%mm1 \n\t"
901
"paddw %%mm0, %%mm6 \n\t"
902
"paddw %%mm1, %%mm6 \n\t"
903
"lea (%1,%3,2), %1 \n\t"
904
"lea (%2,%3,2), %2 \n\t"
907
: "+r" (h), "+r" (blk1), "+r" (blk2)
908
: "r" ((x86_reg) stride));
911
static inline void sad8_y2a_mmxext(uint8_t *blk1, uint8_t *blk2,
915
"movq (%1), %%mm0 \n\t"
919
"movq (%1), %%mm1 \n\t"
920
"movq (%1, %3), %%mm2 \n\t"
921
"pavgb %%mm1, %%mm0 \n\t"
922
"pavgb %%mm2, %%mm1 \n\t"
923
"psadbw (%2), %%mm0 \n\t"
924
"psadbw (%2, %3), %%mm1 \n\t"
925
"paddw %%mm0, %%mm6 \n\t"
926
"paddw %%mm1, %%mm6 \n\t"
927
"movq %%mm2, %%mm0 \n\t"
928
"lea (%1,%3,2), %1 \n\t"
929
"lea (%2,%3,2), %2 \n\t"
932
: "+r" (h), "+r" (blk1), "+r" (blk2)
933
: "r" ((x86_reg) stride));
936
static inline void sad8_4_mmxext(uint8_t *blk1, uint8_t *blk2,
940
"movq "MANGLE(bone)", %%mm5 \n\t"
941
"movq (%1), %%mm0 \n\t"
942
"pavgb 1(%1), %%mm0 \n\t"
946
"movq (%1), %%mm1 \n\t"
947
"movq (%1,%3), %%mm2 \n\t"
948
"pavgb 1(%1), %%mm1 \n\t"
949
"pavgb 1(%1,%3), %%mm2 \n\t"
950
"psubusb %%mm5, %%mm1 \n\t"
951
"pavgb %%mm1, %%mm0 \n\t"
952
"pavgb %%mm2, %%mm1 \n\t"
953
"psadbw (%2), %%mm0 \n\t"
954
"psadbw (%2,%3), %%mm1 \n\t"
955
"paddw %%mm0, %%mm6 \n\t"
956
"paddw %%mm1, %%mm6 \n\t"
957
"movq %%mm2, %%mm0 \n\t"
958
"lea (%1,%3,2), %1 \n\t"
959
"lea (%2,%3,2), %2 \n\t"
962
: "+r" (h), "+r" (blk1), "+r" (blk2)
963
: "r" ((x86_reg) stride));
966
static inline void sad8_2_mmx(uint8_t *blk1a, uint8_t *blk1b, uint8_t *blk2,
969
x86_reg len = -(stride * h);
973
"movq (%1, %%"REG_a"), %%mm0 \n\t"
974
"movq (%2, %%"REG_a"), %%mm1 \n\t"
975
"movq (%1, %%"REG_a"), %%mm2 \n\t"
976
"movq (%2, %%"REG_a"), %%mm3 \n\t"
977
"punpcklbw %%mm7, %%mm0 \n\t"
978
"punpcklbw %%mm7, %%mm1 \n\t"
979
"punpckhbw %%mm7, %%mm2 \n\t"
980
"punpckhbw %%mm7, %%mm3 \n\t"
981
"paddw %%mm0, %%mm1 \n\t"
982
"paddw %%mm2, %%mm3 \n\t"
983
"movq (%3, %%"REG_a"), %%mm4 \n\t"
984
"movq (%3, %%"REG_a"), %%mm2 \n\t"
985
"paddw %%mm5, %%mm1 \n\t"
986
"paddw %%mm5, %%mm3 \n\t"
987
"psrlw $1, %%mm1 \n\t"
988
"psrlw $1, %%mm3 \n\t"
989
"packuswb %%mm3, %%mm1 \n\t"
990
"psubusb %%mm1, %%mm4 \n\t"
991
"psubusb %%mm2, %%mm1 \n\t"
992
"por %%mm4, %%mm1 \n\t"
993
"movq %%mm1, %%mm0 \n\t"
994
"punpcklbw %%mm7, %%mm0 \n\t"
995
"punpckhbw %%mm7, %%mm1 \n\t"
996
"paddw %%mm1, %%mm0 \n\t"
997
"paddw %%mm0, %%mm6 \n\t"
998
"add %4, %%"REG_a" \n\t"
1001
: "r" (blk1a - len), "r" (blk1b - len), "r" (blk2 - len),
1002
"r" ((x86_reg) stride));
1005
static inline void sad8_4_mmx(uint8_t *blk1, uint8_t *blk2, int stride, int h)
1007
x86_reg len = -(stride * h);
1009
"movq (%1, %%"REG_a"), %%mm0 \n\t"
1010
"movq 1(%1, %%"REG_a"), %%mm2 \n\t"
1011
"movq %%mm0, %%mm1 \n\t"
1012
"movq %%mm2, %%mm3 \n\t"
1013
"punpcklbw %%mm7, %%mm0 \n\t"
1014
"punpckhbw %%mm7, %%mm1 \n\t"
1015
"punpcklbw %%mm7, %%mm2 \n\t"
1016
"punpckhbw %%mm7, %%mm3 \n\t"
1017
"paddw %%mm2, %%mm0 \n\t"
1018
"paddw %%mm3, %%mm1 \n\t"
1021
"movq (%2, %%"REG_a"), %%mm2 \n\t"
1022
"movq 1(%2, %%"REG_a"), %%mm4 \n\t"
1023
"movq %%mm2, %%mm3 \n\t"
1024
"movq %%mm4, %%mm5 \n\t"
1025
"punpcklbw %%mm7, %%mm2 \n\t"
1026
"punpckhbw %%mm7, %%mm3 \n\t"
1027
"punpcklbw %%mm7, %%mm4 \n\t"
1028
"punpckhbw %%mm7, %%mm5 \n\t"
1029
"paddw %%mm4, %%mm2 \n\t"
1030
"paddw %%mm5, %%mm3 \n\t"
1031
"movq 16+"MANGLE(round_tab)", %%mm5 \n\t"
1032
"paddw %%mm2, %%mm0 \n\t"
1033
"paddw %%mm3, %%mm1 \n\t"
1034
"paddw %%mm5, %%mm0 \n\t"
1035
"paddw %%mm5, %%mm1 \n\t"
1036
"movq (%3, %%"REG_a"), %%mm4 \n\t"
1037
"movq (%3, %%"REG_a"), %%mm5 \n\t"
1038
"psrlw $2, %%mm0 \n\t"
1039
"psrlw $2, %%mm1 \n\t"
1040
"packuswb %%mm1, %%mm0 \n\t"
1041
"psubusb %%mm0, %%mm4 \n\t"
1042
"psubusb %%mm5, %%mm0 \n\t"
1043
"por %%mm4, %%mm0 \n\t"
1044
"movq %%mm0, %%mm4 \n\t"
1045
"punpcklbw %%mm7, %%mm0 \n\t"
1046
"punpckhbw %%mm7, %%mm4 \n\t"
1047
"paddw %%mm0, %%mm6 \n\t"
1048
"paddw %%mm4, %%mm6 \n\t"
1049
"movq %%mm2, %%mm0 \n\t"
1050
"movq %%mm3, %%mm1 \n\t"
1051
"add %4, %%"REG_a" \n\t"
1054
: "r" (blk1 - len), "r" (blk1 - len + stride), "r" (blk2 - len),
1055
"r" ((x86_reg) stride));
1058
static inline int sum_mmx(void)
1062
"movq %%mm6, %%mm0 \n\t"
1063
"psrlq $32, %%mm6 \n\t"
1064
"paddw %%mm0, %%mm6 \n\t"
1065
"movq %%mm6, %%mm0 \n\t"
1066
"psrlq $16, %%mm6 \n\t"
1067
"paddw %%mm0, %%mm6 \n\t"
1068
"movd %%mm6, %0 \n\t"
1070
return ret & 0xFFFF;
1073
static inline int sum_mmxext(void)
1077
"movd %%mm6, %0 \n\t"
1082
static inline void sad8_x2a_mmx(uint8_t *blk1, uint8_t *blk2, int stride, int h)
1084
sad8_2_mmx(blk1, blk1 + 1, blk2, stride, h);
1087
static inline void sad8_y2a_mmx(uint8_t *blk1, uint8_t *blk2, int stride, int h)
1089
sad8_2_mmx(blk1, blk1 + stride, blk2, stride, h);
1092
#define PIX_SAD(suf) \
1093
static int sad8_ ## suf(MpegEncContext *v, uint8_t *blk2, \
1094
uint8_t *blk1, int stride, int h) \
1097
__asm__ volatile ( \
1098
"pxor %%mm7, %%mm7 \n\t" \
1099
"pxor %%mm6, %%mm6 \n\t" \
1102
sad8_1_ ## suf(blk1, blk2, stride, 8); \
1104
return sum_ ## suf(); \
1107
static int sad8_x2_ ## suf(MpegEncContext *v, uint8_t *blk2, \
1108
uint8_t *blk1, int stride, int h) \
1111
__asm__ volatile ( \
1112
"pxor %%mm7, %%mm7 \n\t" \
1113
"pxor %%mm6, %%mm6 \n\t" \
1114
"movq %0, %%mm5 \n\t" \
1115
:: "m" (round_tab[1])); \
1117
sad8_x2a_ ## suf(blk1, blk2, stride, 8); \
1119
return sum_ ## suf(); \
1122
static int sad8_y2_ ## suf(MpegEncContext *v, uint8_t *blk2, \
1123
uint8_t *blk1, int stride, int h) \
1126
__asm__ volatile ( \
1127
"pxor %%mm7, %%mm7 \n\t" \
1128
"pxor %%mm6, %%mm6 \n\t" \
1129
"movq %0, %%mm5 \n\t" \
1130
:: "m" (round_tab[1])); \
1132
sad8_y2a_ ## suf(blk1, blk2, stride, 8); \
1134
return sum_ ## suf(); \
1137
static int sad8_xy2_ ## suf(MpegEncContext *v, uint8_t *blk2, \
1138
uint8_t *blk1, int stride, int h) \
1141
__asm__ volatile ( \
1142
"pxor %%mm7, %%mm7 \n\t" \
1143
"pxor %%mm6, %%mm6 \n\t" \
1146
sad8_4_ ## suf(blk1, blk2, stride, 8); \
1148
return sum_ ## suf(); \
1151
static int sad16_ ## suf(MpegEncContext *v, uint8_t *blk2, \
1152
uint8_t *blk1, int stride, int h) \
1154
__asm__ volatile ( \
1155
"pxor %%mm7, %%mm7 \n\t" \
1156
"pxor %%mm6, %%mm6 \n\t" \
1159
sad8_1_ ## suf(blk1, blk2, stride, h); \
1160
sad8_1_ ## suf(blk1 + 8, blk2 + 8, stride, h); \
1162
return sum_ ## suf(); \
1165
static int sad16_x2_ ## suf(MpegEncContext *v, uint8_t *blk2, \
1166
uint8_t *blk1, int stride, int h) \
1168
__asm__ volatile ( \
1169
"pxor %%mm7, %%mm7 \n\t" \
1170
"pxor %%mm6, %%mm6 \n\t" \
1171
"movq %0, %%mm5 \n\t" \
1172
:: "m" (round_tab[1])); \
1174
sad8_x2a_ ## suf(blk1, blk2, stride, h); \
1175
sad8_x2a_ ## suf(blk1 + 8, blk2 + 8, stride, h); \
1177
return sum_ ## suf(); \
1180
static int sad16_y2_ ## suf(MpegEncContext *v, uint8_t *blk2, \
1181
uint8_t *blk1, int stride, int h) \
1183
__asm__ volatile ( \
1184
"pxor %%mm7, %%mm7 \n\t" \
1185
"pxor %%mm6, %%mm6 \n\t" \
1186
"movq %0, %%mm5 \n\t" \
1187
:: "m" (round_tab[1])); \
1189
sad8_y2a_ ## suf(blk1, blk2, stride, h); \
1190
sad8_y2a_ ## suf(blk1 + 8, blk2 + 8, stride, h); \
1192
return sum_ ## suf(); \
1195
static int sad16_xy2_ ## suf(MpegEncContext *v, uint8_t *blk2, \
1196
uint8_t *blk1, int stride, int h) \
1198
__asm__ volatile ( \
1199
"pxor %%mm7, %%mm7 \n\t" \
1200
"pxor %%mm6, %%mm6 \n\t" \
1203
sad8_4_ ## suf(blk1, blk2, stride, h); \
1204
sad8_4_ ## suf(blk1 + 8, blk2 + 8, stride, h); \
1206
return sum_ ## suf(); \
1212
#endif /* HAVE_INLINE_ASM */
1214
int ff_sse16_sse2(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
1215
int line_size, int h);
1217
#define hadamard_func(cpu) \
1218
int ff_hadamard8_diff_ ## cpu(MpegEncContext *s, uint8_t *src1, \
1219
uint8_t *src2, int stride, int h); \
1220
int ff_hadamard8_diff16_ ## cpu(MpegEncContext *s, uint8_t *src1, \
1221
uint8_t *src2, int stride, int h);
1224
hadamard_func(mmxext)
1226
hadamard_func(ssse3)
1228
av_cold void ff_me_cmp_init_x86(MECmpContext *c, AVCodecContext *avctx)
1230
int cpu_flags = av_get_cpu_flags();
1233
if (INLINE_MMX(cpu_flags)) {
1234
c->sum_abs_dctelem = sum_abs_dctelem_mmx;
1236
c->pix_abs[0][0] = sad16_mmx;
1237
c->pix_abs[0][1] = sad16_x2_mmx;
1238
c->pix_abs[0][2] = sad16_y2_mmx;
1239
c->pix_abs[0][3] = sad16_xy2_mmx;
1240
c->pix_abs[1][0] = sad8_mmx;
1241
c->pix_abs[1][1] = sad8_x2_mmx;
1242
c->pix_abs[1][2] = sad8_y2_mmx;
1243
c->pix_abs[1][3] = sad8_xy2_mmx;
1245
c->sad[0] = sad16_mmx;
1246
c->sad[1] = sad8_mmx;
1248
c->sse[0] = sse16_mmx;
1249
c->sse[1] = sse8_mmx;
1250
c->vsad[4] = vsad_intra16_mmx;
1252
c->nsse[0] = nsse16_mmx;
1253
c->nsse[1] = nsse8_mmx;
1255
if (!(avctx->flags & CODEC_FLAG_BITEXACT)) {
1256
c->vsad[0] = vsad16_mmx;
1260
if (INLINE_MMXEXT(cpu_flags)) {
1261
c->sum_abs_dctelem = sum_abs_dctelem_mmxext;
1263
c->vsad[4] = vsad_intra16_mmxext;
1265
c->pix_abs[0][0] = sad16_mmxext;
1266
c->pix_abs[1][0] = sad8_mmxext;
1268
c->sad[0] = sad16_mmxext;
1269
c->sad[1] = sad8_mmxext;
1271
if (!(avctx->flags & CODEC_FLAG_BITEXACT)) {
1272
c->pix_abs[0][1] = sad16_x2_mmxext;
1273
c->pix_abs[0][2] = sad16_y2_mmxext;
1274
c->pix_abs[0][3] = sad16_xy2_mmxext;
1275
c->pix_abs[1][1] = sad8_x2_mmxext;
1276
c->pix_abs[1][2] = sad8_y2_mmxext;
1277
c->pix_abs[1][3] = sad8_xy2_mmxext;
1279
c->vsad[0] = vsad16_mmxext;
1283
if (INLINE_SSE2(cpu_flags)) {
1284
c->sum_abs_dctelem = sum_abs_dctelem_sse2;
1287
if (INLINE_SSE2(cpu_flags) && !(cpu_flags & AV_CPU_FLAG_3DNOW)) {
1288
c->sad[0] = sad16_sse2;
1291
#if HAVE_SSSE3_INLINE
1292
if (INLINE_SSSE3(cpu_flags)) {
1293
c->sum_abs_dctelem = sum_abs_dctelem_ssse3;
1296
#endif /* HAVE_INLINE_ASM */
1298
if (EXTERNAL_MMX(cpu_flags)) {
1299
c->hadamard8_diff[0] = ff_hadamard8_diff16_mmx;
1300
c->hadamard8_diff[1] = ff_hadamard8_diff_mmx;
1303
if (EXTERNAL_MMXEXT(cpu_flags)) {
1304
c->hadamard8_diff[0] = ff_hadamard8_diff16_mmxext;
1305
c->hadamard8_diff[1] = ff_hadamard8_diff_mmxext;
1308
if (EXTERNAL_SSE2(cpu_flags)) {
1309
c->sse[0] = ff_sse16_sse2;
1311
#if HAVE_ALIGNED_STACK
1312
c->hadamard8_diff[0] = ff_hadamard8_diff16_sse2;
1313
c->hadamard8_diff[1] = ff_hadamard8_diff_sse2;
1317
if (EXTERNAL_SSSE3(cpu_flags) && HAVE_ALIGNED_STACK) {
1318
c->hadamard8_diff[0] = ff_hadamard8_diff16_ssse3;
1319
c->hadamard8_diff[1] = ff_hadamard8_diff_ssse3;