motion_est_mmx.c
Go to the documentation of this file.
1 /*
2  * MMX optimized motion estimation
3  * Copyright (c) 2001 Fabrice Bellard
4  * Copyright (c) 2002-2004 Michael Niedermayer
5  *
6  * mostly by Michael Niedermayer <michaelni@gmx.at>
7  *
8  * This file is part of Libav.
9  *
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.
14  *
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.
19  *
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
23  */
24 
25 #include "libavutil/x86_cpu.h"
26 #include "libavcodec/dsputil.h"
27 #include "dsputil_mmx.h"
28 
29 DECLARE_ASM_CONST(8, uint64_t, round_tab)[3]={
30 0x0000000000000000ULL,
31 0x0001000100010001ULL,
32 0x0002000200020002ULL,
33 };
34 
35 DECLARE_ASM_CONST(8, uint64_t, bone)= 0x0101010101010101LL;
36 
37 static inline void sad8_1_mmx(uint8_t *blk1, uint8_t *blk2, int stride, int h)
38 {
39  x86_reg len= -(stride*h);
40  __asm__ volatile(
41  ".p2align 4 \n\t"
42  "1: \n\t"
43  "movq (%1, %%"REG_a"), %%mm0 \n\t"
44  "movq (%2, %%"REG_a"), %%mm2 \n\t"
45  "movq (%2, %%"REG_a"), %%mm4 \n\t"
46  "add %3, %%"REG_a" \n\t"
47  "psubusb %%mm0, %%mm2 \n\t"
48  "psubusb %%mm4, %%mm0 \n\t"
49  "movq (%1, %%"REG_a"), %%mm1 \n\t"
50  "movq (%2, %%"REG_a"), %%mm3 \n\t"
51  "movq (%2, %%"REG_a"), %%mm5 \n\t"
52  "psubusb %%mm1, %%mm3 \n\t"
53  "psubusb %%mm5, %%mm1 \n\t"
54  "por %%mm2, %%mm0 \n\t"
55  "por %%mm1, %%mm3 \n\t"
56  "movq %%mm0, %%mm1 \n\t"
57  "movq %%mm3, %%mm2 \n\t"
58  "punpcklbw %%mm7, %%mm0 \n\t"
59  "punpckhbw %%mm7, %%mm1 \n\t"
60  "punpcklbw %%mm7, %%mm3 \n\t"
61  "punpckhbw %%mm7, %%mm2 \n\t"
62  "paddw %%mm1, %%mm0 \n\t"
63  "paddw %%mm3, %%mm2 \n\t"
64  "paddw %%mm2, %%mm0 \n\t"
65  "paddw %%mm0, %%mm6 \n\t"
66  "add %3, %%"REG_a" \n\t"
67  " js 1b \n\t"
68  : "+a" (len)
69  : "r" (blk1 - len), "r" (blk2 - len), "r" ((x86_reg)stride)
70  );
71 }
72 
73 static inline void sad8_1_mmx2(uint8_t *blk1, uint8_t *blk2, int stride, int h)
74 {
75  __asm__ volatile(
76  ".p2align 4 \n\t"
77  "1: \n\t"
78  "movq (%1), %%mm0 \n\t"
79  "movq (%1, %3), %%mm1 \n\t"
80  "psadbw (%2), %%mm0 \n\t"
81  "psadbw (%2, %3), %%mm1 \n\t"
82  "paddw %%mm0, %%mm6 \n\t"
83  "paddw %%mm1, %%mm6 \n\t"
84  "lea (%1,%3,2), %1 \n\t"
85  "lea (%2,%3,2), %2 \n\t"
86  "sub $2, %0 \n\t"
87  " jg 1b \n\t"
88  : "+r" (h), "+r" (blk1), "+r" (blk2)
89  : "r" ((x86_reg)stride)
90  );
91 }
92 
93 static int sad16_sse2(void *v, uint8_t *blk2, uint8_t *blk1, int stride, int h)
94 {
95  int ret;
96  __asm__ volatile(
97  "pxor %%xmm2, %%xmm2 \n\t"
98  ".p2align 4 \n\t"
99  "1: \n\t"
100  "movdqu (%1), %%xmm0 \n\t"
101  "movdqu (%1, %4), %%xmm1 \n\t"
102  "psadbw (%2), %%xmm0 \n\t"
103  "psadbw (%2, %4), %%xmm1 \n\t"
104  "paddw %%xmm0, %%xmm2 \n\t"
105  "paddw %%xmm1, %%xmm2 \n\t"
106  "lea (%1,%4,2), %1 \n\t"
107  "lea (%2,%4,2), %2 \n\t"
108  "sub $2, %0 \n\t"
109  " jg 1b \n\t"
110  "movhlps %%xmm2, %%xmm0 \n\t"
111  "paddw %%xmm0, %%xmm2 \n\t"
112  "movd %%xmm2, %3 \n\t"
113  : "+r" (h), "+r" (blk1), "+r" (blk2), "=r"(ret)
114  : "r" ((x86_reg)stride)
115  );
116  return ret;
117 }
118 
119 static inline void sad8_x2a_mmx2(uint8_t *blk1, uint8_t *blk2, int stride, int h)
120 {
121  __asm__ volatile(
122  ".p2align 4 \n\t"
123  "1: \n\t"
124  "movq (%1), %%mm0 \n\t"
125  "movq (%1, %3), %%mm1 \n\t"
126  "pavgb 1(%1), %%mm0 \n\t"
127  "pavgb 1(%1, %3), %%mm1 \n\t"
128  "psadbw (%2), %%mm0 \n\t"
129  "psadbw (%2, %3), %%mm1 \n\t"
130  "paddw %%mm0, %%mm6 \n\t"
131  "paddw %%mm1, %%mm6 \n\t"
132  "lea (%1,%3,2), %1 \n\t"
133  "lea (%2,%3,2), %2 \n\t"
134  "sub $2, %0 \n\t"
135  " jg 1b \n\t"
136  : "+r" (h), "+r" (blk1), "+r" (blk2)
137  : "r" ((x86_reg)stride)
138  );
139 }
140 
141 static inline void sad8_y2a_mmx2(uint8_t *blk1, uint8_t *blk2, int stride, int h)
142 {
143  __asm__ volatile(
144  "movq (%1), %%mm0 \n\t"
145  "add %3, %1 \n\t"
146  ".p2align 4 \n\t"
147  "1: \n\t"
148  "movq (%1), %%mm1 \n\t"
149  "movq (%1, %3), %%mm2 \n\t"
150  "pavgb %%mm1, %%mm0 \n\t"
151  "pavgb %%mm2, %%mm1 \n\t"
152  "psadbw (%2), %%mm0 \n\t"
153  "psadbw (%2, %3), %%mm1 \n\t"
154  "paddw %%mm0, %%mm6 \n\t"
155  "paddw %%mm1, %%mm6 \n\t"
156  "movq %%mm2, %%mm0 \n\t"
157  "lea (%1,%3,2), %1 \n\t"
158  "lea (%2,%3,2), %2 \n\t"
159  "sub $2, %0 \n\t"
160  " jg 1b \n\t"
161  : "+r" (h), "+r" (blk1), "+r" (blk2)
162  : "r" ((x86_reg)stride)
163  );
164 }
165 
166 static inline void sad8_4_mmx2(uint8_t *blk1, uint8_t *blk2, int stride, int h)
167 {
168  __asm__ volatile(
169  "movq "MANGLE(bone)", %%mm5 \n\t"
170  "movq (%1), %%mm0 \n\t"
171  "pavgb 1(%1), %%mm0 \n\t"
172  "add %3, %1 \n\t"
173  ".p2align 4 \n\t"
174  "1: \n\t"
175  "movq (%1), %%mm1 \n\t"
176  "movq (%1,%3), %%mm2 \n\t"
177  "pavgb 1(%1), %%mm1 \n\t"
178  "pavgb 1(%1,%3), %%mm2 \n\t"
179  "psubusb %%mm5, %%mm1 \n\t"
180  "pavgb %%mm1, %%mm0 \n\t"
181  "pavgb %%mm2, %%mm1 \n\t"
182  "psadbw (%2), %%mm0 \n\t"
183  "psadbw (%2,%3), %%mm1 \n\t"
184  "paddw %%mm0, %%mm6 \n\t"
185  "paddw %%mm1, %%mm6 \n\t"
186  "movq %%mm2, %%mm0 \n\t"
187  "lea (%1,%3,2), %1 \n\t"
188  "lea (%2,%3,2), %2 \n\t"
189  "sub $2, %0 \n\t"
190  " jg 1b \n\t"
191  : "+r" (h), "+r" (blk1), "+r" (blk2)
192  : "r" ((x86_reg)stride)
193  );
194 }
195 
196 static inline void sad8_2_mmx(uint8_t *blk1a, uint8_t *blk1b, uint8_t *blk2, int stride, int h)
197 {
198  x86_reg len= -(stride*h);
199  __asm__ volatile(
200  ".p2align 4 \n\t"
201  "1: \n\t"
202  "movq (%1, %%"REG_a"), %%mm0 \n\t"
203  "movq (%2, %%"REG_a"), %%mm1 \n\t"
204  "movq (%1, %%"REG_a"), %%mm2 \n\t"
205  "movq (%2, %%"REG_a"), %%mm3 \n\t"
206  "punpcklbw %%mm7, %%mm0 \n\t"
207  "punpcklbw %%mm7, %%mm1 \n\t"
208  "punpckhbw %%mm7, %%mm2 \n\t"
209  "punpckhbw %%mm7, %%mm3 \n\t"
210  "paddw %%mm0, %%mm1 \n\t"
211  "paddw %%mm2, %%mm3 \n\t"
212  "movq (%3, %%"REG_a"), %%mm4 \n\t"
213  "movq (%3, %%"REG_a"), %%mm2 \n\t"
214  "paddw %%mm5, %%mm1 \n\t"
215  "paddw %%mm5, %%mm3 \n\t"
216  "psrlw $1, %%mm1 \n\t"
217  "psrlw $1, %%mm3 \n\t"
218  "packuswb %%mm3, %%mm1 \n\t"
219  "psubusb %%mm1, %%mm4 \n\t"
220  "psubusb %%mm2, %%mm1 \n\t"
221  "por %%mm4, %%mm1 \n\t"
222  "movq %%mm1, %%mm0 \n\t"
223  "punpcklbw %%mm7, %%mm0 \n\t"
224  "punpckhbw %%mm7, %%mm1 \n\t"
225  "paddw %%mm1, %%mm0 \n\t"
226  "paddw %%mm0, %%mm6 \n\t"
227  "add %4, %%"REG_a" \n\t"
228  " js 1b \n\t"
229  : "+a" (len)
230  : "r" (blk1a - len), "r" (blk1b -len), "r" (blk2 - len), "r" ((x86_reg)stride)
231  );
232 }
233 
234 static inline void sad8_4_mmx(uint8_t *blk1, uint8_t *blk2, int stride, int h)
235 {
236  x86_reg len= -(stride*h);
237  __asm__ volatile(
238  "movq (%1, %%"REG_a"), %%mm0 \n\t"
239  "movq 1(%1, %%"REG_a"), %%mm2 \n\t"
240  "movq %%mm0, %%mm1 \n\t"
241  "movq %%mm2, %%mm3 \n\t"
242  "punpcklbw %%mm7, %%mm0 \n\t"
243  "punpckhbw %%mm7, %%mm1 \n\t"
244  "punpcklbw %%mm7, %%mm2 \n\t"
245  "punpckhbw %%mm7, %%mm3 \n\t"
246  "paddw %%mm2, %%mm0 \n\t"
247  "paddw %%mm3, %%mm1 \n\t"
248  ".p2align 4 \n\t"
249  "1: \n\t"
250  "movq (%2, %%"REG_a"), %%mm2 \n\t"
251  "movq 1(%2, %%"REG_a"), %%mm4 \n\t"
252  "movq %%mm2, %%mm3 \n\t"
253  "movq %%mm4, %%mm5 \n\t"
254  "punpcklbw %%mm7, %%mm2 \n\t"
255  "punpckhbw %%mm7, %%mm3 \n\t"
256  "punpcklbw %%mm7, %%mm4 \n\t"
257  "punpckhbw %%mm7, %%mm5 \n\t"
258  "paddw %%mm4, %%mm2 \n\t"
259  "paddw %%mm5, %%mm3 \n\t"
260  "movq 16+"MANGLE(round_tab)", %%mm5 \n\t"
261  "paddw %%mm2, %%mm0 \n\t"
262  "paddw %%mm3, %%mm1 \n\t"
263  "paddw %%mm5, %%mm0 \n\t"
264  "paddw %%mm5, %%mm1 \n\t"
265  "movq (%3, %%"REG_a"), %%mm4 \n\t"
266  "movq (%3, %%"REG_a"), %%mm5 \n\t"
267  "psrlw $2, %%mm0 \n\t"
268  "psrlw $2, %%mm1 \n\t"
269  "packuswb %%mm1, %%mm0 \n\t"
270  "psubusb %%mm0, %%mm4 \n\t"
271  "psubusb %%mm5, %%mm0 \n\t"
272  "por %%mm4, %%mm0 \n\t"
273  "movq %%mm0, %%mm4 \n\t"
274  "punpcklbw %%mm7, %%mm0 \n\t"
275  "punpckhbw %%mm7, %%mm4 \n\t"
276  "paddw %%mm0, %%mm6 \n\t"
277  "paddw %%mm4, %%mm6 \n\t"
278  "movq %%mm2, %%mm0 \n\t"
279  "movq %%mm3, %%mm1 \n\t"
280  "add %4, %%"REG_a" \n\t"
281  " js 1b \n\t"
282  : "+a" (len)
283  : "r" (blk1 - len), "r" (blk1 -len + stride), "r" (blk2 - len), "r" ((x86_reg)stride)
284  );
285 }
286 
287 static inline int sum_mmx(void)
288 {
289  int ret;
290  __asm__ volatile(
291  "movq %%mm6, %%mm0 \n\t"
292  "psrlq $32, %%mm6 \n\t"
293  "paddw %%mm0, %%mm6 \n\t"
294  "movq %%mm6, %%mm0 \n\t"
295  "psrlq $16, %%mm6 \n\t"
296  "paddw %%mm0, %%mm6 \n\t"
297  "movd %%mm6, %0 \n\t"
298  : "=r" (ret)
299  );
300  return ret&0xFFFF;
301 }
302 
303 static inline int sum_mmx2(void)
304 {
305  int ret;
306  __asm__ volatile(
307  "movd %%mm6, %0 \n\t"
308  : "=r" (ret)
309  );
310  return ret;
311 }
312 
313 static inline void sad8_x2a_mmx(uint8_t *blk1, uint8_t *blk2, int stride, int h)
314 {
315  sad8_2_mmx(blk1, blk1+1, blk2, stride, h);
316 }
317 static inline void sad8_y2a_mmx(uint8_t *blk1, uint8_t *blk2, int stride, int h)
318 {
319  sad8_2_mmx(blk1, blk1+stride, blk2, stride, h);
320 }
321 
322 
323 #define PIX_SAD(suf)\
324 static int sad8_ ## suf(void *v, uint8_t *blk2, uint8_t *blk1, int stride, int h)\
325 {\
326  assert(h==8);\
327  __asm__ volatile("pxor %%mm7, %%mm7 \n\t"\
328  "pxor %%mm6, %%mm6 \n\t":);\
329 \
330  sad8_1_ ## suf(blk1, blk2, stride, 8);\
331 \
332  return sum_ ## suf();\
333 }\
334 static int sad8_x2_ ## suf(void *v, uint8_t *blk2, uint8_t *blk1, int stride, int h)\
335 {\
336  assert(h==8);\
337  __asm__ volatile("pxor %%mm7, %%mm7 \n\t"\
338  "pxor %%mm6, %%mm6 \n\t"\
339  "movq %0, %%mm5 \n\t"\
340  :: "m"(round_tab[1]) \
341  );\
342 \
343  sad8_x2a_ ## suf(blk1, blk2, stride, 8);\
344 \
345  return sum_ ## suf();\
346 }\
347 \
348 static int sad8_y2_ ## suf(void *v, uint8_t *blk2, uint8_t *blk1, int stride, int h)\
349 {\
350  assert(h==8);\
351  __asm__ volatile("pxor %%mm7, %%mm7 \n\t"\
352  "pxor %%mm6, %%mm6 \n\t"\
353  "movq %0, %%mm5 \n\t"\
354  :: "m"(round_tab[1]) \
355  );\
356 \
357  sad8_y2a_ ## suf(blk1, blk2, stride, 8);\
358 \
359  return sum_ ## suf();\
360 }\
361 \
362 static int sad8_xy2_ ## suf(void *v, uint8_t *blk2, uint8_t *blk1, int stride, int h)\
363 {\
364  assert(h==8);\
365  __asm__ volatile("pxor %%mm7, %%mm7 \n\t"\
366  "pxor %%mm6, %%mm6 \n\t"\
367  ::);\
368 \
369  sad8_4_ ## suf(blk1, blk2, stride, 8);\
370 \
371  return sum_ ## suf();\
372 }\
373 \
374 static int sad16_ ## suf(void *v, uint8_t *blk2, uint8_t *blk1, int stride, int h)\
375 {\
376  __asm__ volatile("pxor %%mm7, %%mm7 \n\t"\
377  "pxor %%mm6, %%mm6 \n\t":);\
378 \
379  sad8_1_ ## suf(blk1 , blk2 , stride, h);\
380  sad8_1_ ## suf(blk1+8, blk2+8, stride, h);\
381 \
382  return sum_ ## suf();\
383 }\
384 static int sad16_x2_ ## suf(void *v, uint8_t *blk2, uint8_t *blk1, int stride, int h)\
385 {\
386  __asm__ volatile("pxor %%mm7, %%mm7 \n\t"\
387  "pxor %%mm6, %%mm6 \n\t"\
388  "movq %0, %%mm5 \n\t"\
389  :: "m"(round_tab[1]) \
390  );\
391 \
392  sad8_x2a_ ## suf(blk1 , blk2 , stride, h);\
393  sad8_x2a_ ## suf(blk1+8, blk2+8, stride, h);\
394 \
395  return sum_ ## suf();\
396 }\
397 static int sad16_y2_ ## suf(void *v, uint8_t *blk2, uint8_t *blk1, int stride, int h)\
398 {\
399  __asm__ volatile("pxor %%mm7, %%mm7 \n\t"\
400  "pxor %%mm6, %%mm6 \n\t"\
401  "movq %0, %%mm5 \n\t"\
402  :: "m"(round_tab[1]) \
403  );\
404 \
405  sad8_y2a_ ## suf(blk1 , blk2 , stride, h);\
406  sad8_y2a_ ## suf(blk1+8, blk2+8, stride, h);\
407 \
408  return sum_ ## suf();\
409 }\
410 static int sad16_xy2_ ## suf(void *v, uint8_t *blk2, uint8_t *blk1, int stride, int h)\
411 {\
412  __asm__ volatile("pxor %%mm7, %%mm7 \n\t"\
413  "pxor %%mm6, %%mm6 \n\t"\
414  ::);\
415 \
416  sad8_4_ ## suf(blk1 , blk2 , stride, h);\
417  sad8_4_ ## suf(blk1+8, blk2+8, stride, h);\
418 \
419  return sum_ ## suf();\
420 }\
421 
422 PIX_SAD(mmx)
423 PIX_SAD(mmx2)
424 
426 {
427  int mm_flags = av_get_cpu_flags();
428 
429  if (mm_flags & AV_CPU_FLAG_MMX) {
430  c->pix_abs[0][0] = sad16_mmx;
431  c->pix_abs[0][1] = sad16_x2_mmx;
432  c->pix_abs[0][2] = sad16_y2_mmx;
433  c->pix_abs[0][3] = sad16_xy2_mmx;
434  c->pix_abs[1][0] = sad8_mmx;
435  c->pix_abs[1][1] = sad8_x2_mmx;
436  c->pix_abs[1][2] = sad8_y2_mmx;
437  c->pix_abs[1][3] = sad8_xy2_mmx;
438 
439  c->sad[0]= sad16_mmx;
440  c->sad[1]= sad8_mmx;
441  }
442  if (mm_flags & AV_CPU_FLAG_MMX2) {
443  c->pix_abs[0][0] = sad16_mmx2;
444  c->pix_abs[1][0] = sad8_mmx2;
445 
446  c->sad[0]= sad16_mmx2;
447  c->sad[1]= sad8_mmx2;
448 
449  if(!(avctx->flags & CODEC_FLAG_BITEXACT)){
450  c->pix_abs[0][1] = sad16_x2_mmx2;
451  c->pix_abs[0][2] = sad16_y2_mmx2;
452  c->pix_abs[0][3] = sad16_xy2_mmx2;
453  c->pix_abs[1][1] = sad8_x2_mmx2;
454  c->pix_abs[1][2] = sad8_y2_mmx2;
455  c->pix_abs[1][3] = sad8_xy2_mmx2;
456  }
457  }
458  if ((mm_flags & AV_CPU_FLAG_SSE2) && !(mm_flags & AV_CPU_FLAG_3DNOW) && avctx->codec_id != CODEC_ID_SNOW) {
459  c->sad[0]= sad16_sse2;
460  }
461 }