2 * MMX optimized motion estimation
3 * Copyright (c) 2001 Fabrice Bellard.
4 * Copyright (c) 2002-2004 Michael Niedermayer
6 * mostly by Michael Niedermayer <michaelni@gmx.at>
8 * This file is part of FFmpeg.
10 * FFmpeg 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 * FFmpeg 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 FFmpeg; if not, write to the Free Software
22 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
27 static const __attribute__ ((aligned(8))) uint64_t round_tab[3]={
28 0x0000000000000000ULL,
29 0x0001000100010001ULL,
30 0x0002000200020002ULL,
33 static attribute_used __attribute__ ((aligned(8))) uint64_t bone= 0x0101010101010101LL;
35 static inline void sad8_1_mmx(uint8_t *blk1, uint8_t *blk2, int stride, int h)
37 long len= -(stride*h);
41 "movq (%1, %%"REG_a"), %%mm0 \n\t"
42 "movq (%2, %%"REG_a"), %%mm2 \n\t"
43 "movq (%2, %%"REG_a"), %%mm4 \n\t"
44 "add %3, %%"REG_a" \n\t"
45 "psubusb %%mm0, %%mm2 \n\t"
46 "psubusb %%mm4, %%mm0 \n\t"
47 "movq (%1, %%"REG_a"), %%mm1 \n\t"
48 "movq (%2, %%"REG_a"), %%mm3 \n\t"
49 "movq (%2, %%"REG_a"), %%mm5 \n\t"
50 "psubusb %%mm1, %%mm3 \n\t"
51 "psubusb %%mm5, %%mm1 \n\t"
52 "por %%mm2, %%mm0 \n\t"
53 "por %%mm1, %%mm3 \n\t"
54 "movq %%mm0, %%mm1 \n\t"
55 "movq %%mm3, %%mm2 \n\t"
56 "punpcklbw %%mm7, %%mm0 \n\t"
57 "punpckhbw %%mm7, %%mm1 \n\t"
58 "punpcklbw %%mm7, %%mm3 \n\t"
59 "punpckhbw %%mm7, %%mm2 \n\t"
60 "paddw %%mm1, %%mm0 \n\t"
61 "paddw %%mm3, %%mm2 \n\t"
62 "paddw %%mm2, %%mm0 \n\t"
63 "paddw %%mm0, %%mm6 \n\t"
64 "add %3, %%"REG_a" \n\t"
67 : "r" (blk1 - len), "r" (blk2 - len), "r" ((long)stride)
71 static inline void sad8_1_mmx2(uint8_t *blk1, uint8_t *blk2, int stride, int h)
76 "movq (%1), %%mm0 \n\t"
77 "movq (%1, %3), %%mm1 \n\t"
78 "psadbw (%2), %%mm0 \n\t"
79 "psadbw (%2, %3), %%mm1 \n\t"
80 "paddw %%mm0, %%mm6 \n\t"
81 "paddw %%mm1, %%mm6 \n\t"
82 "lea (%1,%3,2), %1 \n\t"
83 "lea (%2,%3,2), %2 \n\t"
86 : "+r" (h), "+r" (blk1), "+r" (blk2)
91 static int sad16_sse2(void *v, uint8_t *blk2, uint8_t *blk1, int stride, int h)
95 "pxor %%xmm6, %%xmm6 \n\t"
98 "movdqu (%1), %%xmm0 \n\t"
99 "movdqu (%1, %3), %%xmm1 \n\t"
100 "psadbw (%2), %%xmm0 \n\t"
101 "psadbw (%2, %3), %%xmm1 \n\t"
102 "paddw %%xmm0, %%xmm6 \n\t"
103 "paddw %%xmm1, %%xmm6 \n\t"
104 "lea (%1,%3,2), %1 \n\t"
105 "lea (%2,%3,2), %2 \n\t"
108 : "+r" (h), "+r" (blk1), "+r" (blk2)
112 "movhlps %%xmm6, %%xmm0 \n\t"
113 "paddw %%xmm0, %%xmm6 \n\t"
114 "movd %%xmm6, %0 \n\t"
120 static inline void sad8_x2a_mmx2(uint8_t *blk1, uint8_t *blk2, int stride, int h)
125 "movq (%1), %%mm0 \n\t"
126 "movq (%1, %3), %%mm1 \n\t"
127 "pavgb 1(%1), %%mm0 \n\t"
128 "pavgb 1(%1, %3), %%mm1 \n\t"
129 "psadbw (%2), %%mm0 \n\t"
130 "psadbw (%2, %3), %%mm1 \n\t"
131 "paddw %%mm0, %%mm6 \n\t"
132 "paddw %%mm1, %%mm6 \n\t"
133 "lea (%1,%3,2), %1 \n\t"
134 "lea (%2,%3,2), %2 \n\t"
137 : "+r" (h), "+r" (blk1), "+r" (blk2)
142 static inline void sad8_y2a_mmx2(uint8_t *blk1, uint8_t *blk2, int stride, int h)
145 "movq (%1), %%mm0 \n\t"
149 "movq (%1), %%mm1 \n\t"
150 "movq (%1, %3), %%mm2 \n\t"
151 "pavgb %%mm1, %%mm0 \n\t"
152 "pavgb %%mm2, %%mm1 \n\t"
153 "psadbw (%2), %%mm0 \n\t"
154 "psadbw (%2, %3), %%mm1 \n\t"
155 "paddw %%mm0, %%mm6 \n\t"
156 "paddw %%mm1, %%mm6 \n\t"
157 "movq %%mm2, %%mm0 \n\t"
158 "lea (%1,%3,2), %1 \n\t"
159 "lea (%2,%3,2), %2 \n\t"
162 : "+r" (h), "+r" (blk1), "+r" (blk2)
167 static inline void sad8_4_mmx2(uint8_t *blk1, uint8_t *blk2, int stride, int h)
170 "movq "MANGLE(bone)", %%mm5 \n\t"
171 "movq (%1), %%mm0 \n\t"
172 "pavgb 1(%1), %%mm0 \n\t"
176 "movq (%1), %%mm1 \n\t"
177 "movq (%1,%3), %%mm2 \n\t"
178 "pavgb 1(%1), %%mm1 \n\t"
179 "pavgb 1(%1,%3), %%mm2 \n\t"
180 "psubusb %%mm5, %%mm1 \n\t"
181 "pavgb %%mm1, %%mm0 \n\t"
182 "pavgb %%mm2, %%mm1 \n\t"
183 "psadbw (%2), %%mm0 \n\t"
184 "psadbw (%2,%3), %%mm1 \n\t"
185 "paddw %%mm0, %%mm6 \n\t"
186 "paddw %%mm1, %%mm6 \n\t"
187 "movq %%mm2, %%mm0 \n\t"
188 "lea (%1,%3,2), %1 \n\t"
189 "lea (%2,%3,2), %2 \n\t"
192 : "+r" (h), "+r" (blk1), "+r" (blk2)
197 static inline void sad8_2_mmx(uint8_t *blk1a, uint8_t *blk1b, uint8_t *blk2, int stride, int h)
199 long len= -(stride*h);
203 "movq (%1, %%"REG_a"), %%mm0 \n\t"
204 "movq (%2, %%"REG_a"), %%mm1 \n\t"
205 "movq (%1, %%"REG_a"), %%mm2 \n\t"
206 "movq (%2, %%"REG_a"), %%mm3 \n\t"
207 "punpcklbw %%mm7, %%mm0 \n\t"
208 "punpcklbw %%mm7, %%mm1 \n\t"
209 "punpckhbw %%mm7, %%mm2 \n\t"
210 "punpckhbw %%mm7, %%mm3 \n\t"
211 "paddw %%mm0, %%mm1 \n\t"
212 "paddw %%mm2, %%mm3 \n\t"
213 "movq (%3, %%"REG_a"), %%mm4 \n\t"
214 "movq (%3, %%"REG_a"), %%mm2 \n\t"
215 "paddw %%mm5, %%mm1 \n\t"
216 "paddw %%mm5, %%mm3 \n\t"
217 "psrlw $1, %%mm1 \n\t"
218 "psrlw $1, %%mm3 \n\t"
219 "packuswb %%mm3, %%mm1 \n\t"
220 "psubusb %%mm1, %%mm4 \n\t"
221 "psubusb %%mm2, %%mm1 \n\t"
222 "por %%mm4, %%mm1 \n\t"
223 "movq %%mm1, %%mm0 \n\t"
224 "punpcklbw %%mm7, %%mm0 \n\t"
225 "punpckhbw %%mm7, %%mm1 \n\t"
226 "paddw %%mm1, %%mm0 \n\t"
227 "paddw %%mm0, %%mm6 \n\t"
228 "add %4, %%"REG_a" \n\t"
231 : "r" (blk1a - len), "r" (blk1b -len), "r" (blk2 - len), "r" ((long)stride)
235 static inline void sad8_4_mmx(uint8_t *blk1, uint8_t *blk2, int stride, int h)
237 long len= -(stride*h);
239 "movq (%1, %%"REG_a"), %%mm0 \n\t"
240 "movq 1(%1, %%"REG_a"), %%mm2 \n\t"
241 "movq %%mm0, %%mm1 \n\t"
242 "movq %%mm2, %%mm3 \n\t"
243 "punpcklbw %%mm7, %%mm0 \n\t"
244 "punpckhbw %%mm7, %%mm1 \n\t"
245 "punpcklbw %%mm7, %%mm2 \n\t"
246 "punpckhbw %%mm7, %%mm3 \n\t"
247 "paddw %%mm2, %%mm0 \n\t"
248 "paddw %%mm3, %%mm1 \n\t"
251 "movq (%2, %%"REG_a"), %%mm2 \n\t"
252 "movq 1(%2, %%"REG_a"), %%mm4 \n\t"
253 "movq %%mm2, %%mm3 \n\t"
254 "movq %%mm4, %%mm5 \n\t"
255 "punpcklbw %%mm7, %%mm2 \n\t"
256 "punpckhbw %%mm7, %%mm3 \n\t"
257 "punpcklbw %%mm7, %%mm4 \n\t"
258 "punpckhbw %%mm7, %%mm5 \n\t"
259 "paddw %%mm4, %%mm2 \n\t"
260 "paddw %%mm5, %%mm3 \n\t"
261 "movq 16+"MANGLE(round_tab)", %%mm5 \n\t"
262 "paddw %%mm2, %%mm0 \n\t"
263 "paddw %%mm3, %%mm1 \n\t"
264 "paddw %%mm5, %%mm0 \n\t"
265 "paddw %%mm5, %%mm1 \n\t"
266 "movq (%3, %%"REG_a"), %%mm4 \n\t"
267 "movq (%3, %%"REG_a"), %%mm5 \n\t"
268 "psrlw $2, %%mm0 \n\t"
269 "psrlw $2, %%mm1 \n\t"
270 "packuswb %%mm1, %%mm0 \n\t"
271 "psubusb %%mm0, %%mm4 \n\t"
272 "psubusb %%mm5, %%mm0 \n\t"
273 "por %%mm4, %%mm0 \n\t"
274 "movq %%mm0, %%mm4 \n\t"
275 "punpcklbw %%mm7, %%mm0 \n\t"
276 "punpckhbw %%mm7, %%mm4 \n\t"
277 "paddw %%mm0, %%mm6 \n\t"
278 "paddw %%mm4, %%mm6 \n\t"
279 "movq %%mm2, %%mm0 \n\t"
280 "movq %%mm3, %%mm1 \n\t"
281 "add %4, %%"REG_a" \n\t"
284 : "r" (blk1 - len), "r" (blk1 -len + stride), "r" (blk2 - len), "r" ((long)stride)
288 static inline int sum_mmx(void)
292 "movq %%mm6, %%mm0 \n\t"
293 "psrlq $32, %%mm6 \n\t"
294 "paddw %%mm0, %%mm6 \n\t"
295 "movq %%mm6, %%mm0 \n\t"
296 "psrlq $16, %%mm6 \n\t"
297 "paddw %%mm0, %%mm6 \n\t"
298 "movd %%mm6, %0 \n\t"
304 static inline int sum_mmx2(void)
308 "movd %%mm6, %0 \n\t"
314 static inline void sad8_x2a_mmx(uint8_t *blk1, uint8_t *blk2, int stride, int h)
316 sad8_2_mmx(blk1, blk1+1, blk2, stride, h);
318 static inline void sad8_y2a_mmx(uint8_t *blk1, uint8_t *blk2, int stride, int h)
320 sad8_2_mmx(blk1, blk1+stride, blk2, stride, h);
324 #define PIX_SAD(suf)\
325 static int sad8_ ## suf(void *v, uint8_t *blk2, uint8_t *blk1, int stride, int h)\
328 asm volatile("pxor %%mm7, %%mm7 \n\t"\
329 "pxor %%mm6, %%mm6 \n\t":);\
331 sad8_1_ ## suf(blk1, blk2, stride, 8);\
333 return sum_ ## suf();\
335 static int sad8_x2_ ## suf(void *v, uint8_t *blk2, uint8_t *blk1, int stride, int h)\
338 asm volatile("pxor %%mm7, %%mm7 \n\t"\
339 "pxor %%mm6, %%mm6 \n\t"\
340 "movq %0, %%mm5 \n\t"\
341 :: "m"(round_tab[1]) \
344 sad8_x2a_ ## suf(blk1, blk2, stride, 8);\
346 return sum_ ## suf();\
349 static int sad8_y2_ ## suf(void *v, uint8_t *blk2, uint8_t *blk1, int stride, int h)\
352 asm volatile("pxor %%mm7, %%mm7 \n\t"\
353 "pxor %%mm6, %%mm6 \n\t"\
354 "movq %0, %%mm5 \n\t"\
355 :: "m"(round_tab[1]) \
358 sad8_y2a_ ## suf(blk1, blk2, stride, 8);\
360 return sum_ ## suf();\
363 static int sad8_xy2_ ## suf(void *v, uint8_t *blk2, uint8_t *blk1, int stride, int h)\
366 asm volatile("pxor %%mm7, %%mm7 \n\t"\
367 "pxor %%mm6, %%mm6 \n\t"\
370 sad8_4_ ## suf(blk1, blk2, stride, 8);\
372 return sum_ ## suf();\
375 static int sad16_ ## suf(void *v, uint8_t *blk2, uint8_t *blk1, int stride, int h)\
377 asm volatile("pxor %%mm7, %%mm7 \n\t"\
378 "pxor %%mm6, %%mm6 \n\t":);\
380 sad8_1_ ## suf(blk1 , blk2 , stride, h);\
381 sad8_1_ ## suf(blk1+8, blk2+8, stride, h);\
383 return sum_ ## suf();\
385 static int sad16_x2_ ## suf(void *v, uint8_t *blk2, uint8_t *blk1, int stride, int h)\
387 asm volatile("pxor %%mm7, %%mm7 \n\t"\
388 "pxor %%mm6, %%mm6 \n\t"\
389 "movq %0, %%mm5 \n\t"\
390 :: "m"(round_tab[1]) \
393 sad8_x2a_ ## suf(blk1 , blk2 , stride, h);\
394 sad8_x2a_ ## suf(blk1+8, blk2+8, stride, h);\
396 return sum_ ## suf();\
398 static int sad16_y2_ ## suf(void *v, uint8_t *blk2, uint8_t *blk1, int stride, int h)\
400 asm volatile("pxor %%mm7, %%mm7 \n\t"\
401 "pxor %%mm6, %%mm6 \n\t"\
402 "movq %0, %%mm5 \n\t"\
403 :: "m"(round_tab[1]) \
406 sad8_y2a_ ## suf(blk1 , blk2 , stride, h);\
407 sad8_y2a_ ## suf(blk1+8, blk2+8, stride, h);\
409 return sum_ ## suf();\
411 static int sad16_xy2_ ## suf(void *v, uint8_t *blk2, uint8_t *blk1, int stride, int h)\
413 asm volatile("pxor %%mm7, %%mm7 \n\t"\
414 "pxor %%mm6, %%mm6 \n\t"\
417 sad8_4_ ## suf(blk1 , blk2 , stride, h);\
418 sad8_4_ ## suf(blk1+8, blk2+8, stride, h);\
420 return sum_ ## suf();\
426 void dsputil_init_pix_mmx(DSPContext* c, AVCodecContext *avctx)
428 if (mm_flags & MM_MMX) {
429 c->pix_abs[0][0] = sad16_mmx;
430 c->pix_abs[0][1] = sad16_x2_mmx;
431 c->pix_abs[0][2] = sad16_y2_mmx;
432 c->pix_abs[0][3] = sad16_xy2_mmx;
433 c->pix_abs[1][0] = sad8_mmx;
434 c->pix_abs[1][1] = sad8_x2_mmx;
435 c->pix_abs[1][2] = sad8_y2_mmx;
436 c->pix_abs[1][3] = sad8_xy2_mmx;
438 c->sad[0]= sad16_mmx;
441 if (mm_flags & MM_MMXEXT) {
442 c->pix_abs[0][0] = sad16_mmx2;
443 c->pix_abs[1][0] = sad8_mmx2;
445 c->sad[0]= sad16_mmx2;
446 c->sad[1]= sad8_mmx2;
448 if(!(avctx->flags & CODEC_FLAG_BITEXACT)){
449 c->pix_abs[0][1] = sad16_x2_mmx2;
450 c->pix_abs[0][2] = sad16_y2_mmx2;
451 c->pix_abs[0][3] = sad16_xy2_mmx2;
452 c->pix_abs[1][1] = sad8_x2_mmx2;
453 c->pix_abs[1][2] = sad8_y2_mmx2;
454 c->pix_abs[1][3] = sad8_xy2_mmx2;
457 if ((mm_flags & MM_SSE2) && !(mm_flags & MM_3DNOW)) {
458 c->sad[0]= sad16_sse2;