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
25 #include "libavutil/attributes.h"
26 #include "libavutil/avassert.h"
27 #include "libavutil/mem.h"
28 #include "libavutil/x86/asm.h"
29 #include "dsputil_x86.h"
33 DECLARE_ASM_CONST(8, uint64_t, round_tab)[3]={
34 0x0000000000000000ULL,
35 0x0001000100010001ULL,
36 0x0002000200020002ULL,
39 DECLARE_ASM_CONST(8, uint64_t, bone)= 0x0101010101010101LL;
41 static inline void sad8_1_mmx(uint8_t *blk1, uint8_t *blk2, int stride, int h)
43 x86_reg len= -(x86_reg)stride*h;
47 "movq (%1, %%"REG_a"), %%mm0 \n\t"
48 "movq (%2, %%"REG_a"), %%mm2 \n\t"
49 "movq (%2, %%"REG_a"), %%mm4 \n\t"
50 "add %3, %%"REG_a" \n\t"
51 "psubusb %%mm0, %%mm2 \n\t"
52 "psubusb %%mm4, %%mm0 \n\t"
53 "movq (%1, %%"REG_a"), %%mm1 \n\t"
54 "movq (%2, %%"REG_a"), %%mm3 \n\t"
55 "movq (%2, %%"REG_a"), %%mm5 \n\t"
56 "psubusb %%mm1, %%mm3 \n\t"
57 "psubusb %%mm5, %%mm1 \n\t"
58 "por %%mm2, %%mm0 \n\t"
59 "por %%mm1, %%mm3 \n\t"
60 "movq %%mm0, %%mm1 \n\t"
61 "movq %%mm3, %%mm2 \n\t"
62 "punpcklbw %%mm7, %%mm0 \n\t"
63 "punpckhbw %%mm7, %%mm1 \n\t"
64 "punpcklbw %%mm7, %%mm3 \n\t"
65 "punpckhbw %%mm7, %%mm2 \n\t"
66 "paddw %%mm1, %%mm0 \n\t"
67 "paddw %%mm3, %%mm2 \n\t"
68 "paddw %%mm2, %%mm0 \n\t"
69 "paddw %%mm0, %%mm6 \n\t"
70 "add %3, %%"REG_a" \n\t"
73 : "r" (blk1 - len), "r" (blk2 - len), "r" ((x86_reg)stride)
77 static inline void sad8_1_mmxext(uint8_t *blk1, uint8_t *blk2,
83 "movq (%1), %%mm0 \n\t"
84 "movq (%1, %3), %%mm1 \n\t"
85 "psadbw (%2), %%mm0 \n\t"
86 "psadbw (%2, %3), %%mm1 \n\t"
87 "paddw %%mm0, %%mm6 \n\t"
88 "paddw %%mm1, %%mm6 \n\t"
89 "lea (%1,%3,2), %1 \n\t"
90 "lea (%2,%3,2), %2 \n\t"
93 : "+r" (h), "+r" (blk1), "+r" (blk2)
94 : "r" ((x86_reg)stride)
98 static int sad16_sse2(void *v, uint8_t *blk2, uint8_t *blk1, int stride, int h)
102 "pxor %%xmm2, %%xmm2 \n\t"
105 "movdqu (%1), %%xmm0 \n\t"
106 "movdqu (%1, %4), %%xmm1 \n\t"
107 "psadbw (%2), %%xmm0 \n\t"
108 "psadbw (%2, %4), %%xmm1 \n\t"
109 "paddw %%xmm0, %%xmm2 \n\t"
110 "paddw %%xmm1, %%xmm2 \n\t"
111 "lea (%1,%4,2), %1 \n\t"
112 "lea (%2,%4,2), %2 \n\t"
115 "movhlps %%xmm2, %%xmm0 \n\t"
116 "paddw %%xmm0, %%xmm2 \n\t"
117 "movd %%xmm2, %3 \n\t"
118 : "+r" (h), "+r" (blk1), "+r" (blk2), "=r"(ret)
119 : "r" ((x86_reg)stride)
124 static inline void sad8_x2a_mmxext(uint8_t *blk1, uint8_t *blk2,
130 "movq (%1), %%mm0 \n\t"
131 "movq (%1, %3), %%mm1 \n\t"
132 "pavgb 1(%1), %%mm0 \n\t"
133 "pavgb 1(%1, %3), %%mm1 \n\t"
134 "psadbw (%2), %%mm0 \n\t"
135 "psadbw (%2, %3), %%mm1 \n\t"
136 "paddw %%mm0, %%mm6 \n\t"
137 "paddw %%mm1, %%mm6 \n\t"
138 "lea (%1,%3,2), %1 \n\t"
139 "lea (%2,%3,2), %2 \n\t"
142 : "+r" (h), "+r" (blk1), "+r" (blk2)
143 : "r" ((x86_reg)stride)
147 static inline void sad8_y2a_mmxext(uint8_t *blk1, uint8_t *blk2,
151 "movq (%1), %%mm0 \n\t"
155 "movq (%1), %%mm1 \n\t"
156 "movq (%1, %3), %%mm2 \n\t"
157 "pavgb %%mm1, %%mm0 \n\t"
158 "pavgb %%mm2, %%mm1 \n\t"
159 "psadbw (%2), %%mm0 \n\t"
160 "psadbw (%2, %3), %%mm1 \n\t"
161 "paddw %%mm0, %%mm6 \n\t"
162 "paddw %%mm1, %%mm6 \n\t"
163 "movq %%mm2, %%mm0 \n\t"
164 "lea (%1,%3,2), %1 \n\t"
165 "lea (%2,%3,2), %2 \n\t"
168 : "+r" (h), "+r" (blk1), "+r" (blk2)
169 : "r" ((x86_reg)stride)
173 static inline void sad8_4_mmxext(uint8_t *blk1, uint8_t *blk2,
177 "movq "MANGLE(bone)", %%mm5 \n\t"
178 "movq (%1), %%mm0 \n\t"
179 "pavgb 1(%1), %%mm0 \n\t"
183 "movq (%1), %%mm1 \n\t"
184 "movq (%1,%3), %%mm2 \n\t"
185 "pavgb 1(%1), %%mm1 \n\t"
186 "pavgb 1(%1,%3), %%mm2 \n\t"
187 "psubusb %%mm5, %%mm1 \n\t"
188 "pavgb %%mm1, %%mm0 \n\t"
189 "pavgb %%mm2, %%mm1 \n\t"
190 "psadbw (%2), %%mm0 \n\t"
191 "psadbw (%2,%3), %%mm1 \n\t"
192 "paddw %%mm0, %%mm6 \n\t"
193 "paddw %%mm1, %%mm6 \n\t"
194 "movq %%mm2, %%mm0 \n\t"
195 "lea (%1,%3,2), %1 \n\t"
196 "lea (%2,%3,2), %2 \n\t"
199 : "+r" (h), "+r" (blk1), "+r" (blk2)
200 : "r" ((x86_reg)stride)
204 static inline void sad8_2_mmx(uint8_t *blk1a, uint8_t *blk1b, uint8_t *blk2, int stride, int h)
206 x86_reg len= -(x86_reg)stride*h;
210 "movq (%1, %%"REG_a"), %%mm0 \n\t"
211 "movq (%2, %%"REG_a"), %%mm1 \n\t"
212 "movq (%1, %%"REG_a"), %%mm2 \n\t"
213 "movq (%2, %%"REG_a"), %%mm3 \n\t"
214 "punpcklbw %%mm7, %%mm0 \n\t"
215 "punpcklbw %%mm7, %%mm1 \n\t"
216 "punpckhbw %%mm7, %%mm2 \n\t"
217 "punpckhbw %%mm7, %%mm3 \n\t"
218 "paddw %%mm0, %%mm1 \n\t"
219 "paddw %%mm2, %%mm3 \n\t"
220 "movq (%3, %%"REG_a"), %%mm4 \n\t"
221 "movq (%3, %%"REG_a"), %%mm2 \n\t"
222 "paddw %%mm5, %%mm1 \n\t"
223 "paddw %%mm5, %%mm3 \n\t"
224 "psrlw $1, %%mm1 \n\t"
225 "psrlw $1, %%mm3 \n\t"
226 "packuswb %%mm3, %%mm1 \n\t"
227 "psubusb %%mm1, %%mm4 \n\t"
228 "psubusb %%mm2, %%mm1 \n\t"
229 "por %%mm4, %%mm1 \n\t"
230 "movq %%mm1, %%mm0 \n\t"
231 "punpcklbw %%mm7, %%mm0 \n\t"
232 "punpckhbw %%mm7, %%mm1 \n\t"
233 "paddw %%mm1, %%mm0 \n\t"
234 "paddw %%mm0, %%mm6 \n\t"
235 "add %4, %%"REG_a" \n\t"
238 : "r" (blk1a - len), "r" (blk1b -len), "r" (blk2 - len), "r" ((x86_reg)stride)
242 static inline void sad8_4_mmx(uint8_t *blk1, uint8_t *blk2, int stride, int h)
244 x86_reg len= -(x86_reg)stride*h;
246 "movq (%1, %%"REG_a"), %%mm0 \n\t"
247 "movq 1(%1, %%"REG_a"), %%mm2 \n\t"
248 "movq %%mm0, %%mm1 \n\t"
249 "movq %%mm2, %%mm3 \n\t"
250 "punpcklbw %%mm7, %%mm0 \n\t"
251 "punpckhbw %%mm7, %%mm1 \n\t"
252 "punpcklbw %%mm7, %%mm2 \n\t"
253 "punpckhbw %%mm7, %%mm3 \n\t"
254 "paddw %%mm2, %%mm0 \n\t"
255 "paddw %%mm3, %%mm1 \n\t"
258 "movq (%2, %%"REG_a"), %%mm2 \n\t"
259 "movq 1(%2, %%"REG_a"), %%mm4 \n\t"
260 "movq %%mm2, %%mm3 \n\t"
261 "movq %%mm4, %%mm5 \n\t"
262 "punpcklbw %%mm7, %%mm2 \n\t"
263 "punpckhbw %%mm7, %%mm3 \n\t"
264 "punpcklbw %%mm7, %%mm4 \n\t"
265 "punpckhbw %%mm7, %%mm5 \n\t"
266 "paddw %%mm4, %%mm2 \n\t"
267 "paddw %%mm5, %%mm3 \n\t"
268 "movq 16+"MANGLE(round_tab)", %%mm5 \n\t"
269 "paddw %%mm2, %%mm0 \n\t"
270 "paddw %%mm3, %%mm1 \n\t"
271 "paddw %%mm5, %%mm0 \n\t"
272 "paddw %%mm5, %%mm1 \n\t"
273 "movq (%3, %%"REG_a"), %%mm4 \n\t"
274 "movq (%3, %%"REG_a"), %%mm5 \n\t"
275 "psrlw $2, %%mm0 \n\t"
276 "psrlw $2, %%mm1 \n\t"
277 "packuswb %%mm1, %%mm0 \n\t"
278 "psubusb %%mm0, %%mm4 \n\t"
279 "psubusb %%mm5, %%mm0 \n\t"
280 "por %%mm4, %%mm0 \n\t"
281 "movq %%mm0, %%mm4 \n\t"
282 "punpcklbw %%mm7, %%mm0 \n\t"
283 "punpckhbw %%mm7, %%mm4 \n\t"
284 "paddw %%mm0, %%mm6 \n\t"
285 "paddw %%mm4, %%mm6 \n\t"
286 "movq %%mm2, %%mm0 \n\t"
287 "movq %%mm3, %%mm1 \n\t"
288 "add %4, %%"REG_a" \n\t"
291 : "r" (blk1 - len), "r" (blk1 -len + stride), "r" (blk2 - len), "r" ((x86_reg)stride)
295 static inline int sum_mmx(void)
299 "movq %%mm6, %%mm0 \n\t"
300 "psrlq $32, %%mm6 \n\t"
301 "paddw %%mm0, %%mm6 \n\t"
302 "movq %%mm6, %%mm0 \n\t"
303 "psrlq $16, %%mm6 \n\t"
304 "paddw %%mm0, %%mm6 \n\t"
305 "movd %%mm6, %0 \n\t"
311 static inline int sum_mmxext(void)
315 "movd %%mm6, %0 \n\t"
321 static inline void sad8_x2a_mmx(uint8_t *blk1, uint8_t *blk2, int stride, int h)
323 sad8_2_mmx(blk1, blk1+1, blk2, stride, h);
325 static inline void sad8_y2a_mmx(uint8_t *blk1, uint8_t *blk2, int stride, int h)
327 sad8_2_mmx(blk1, blk1+stride, blk2, stride, h);
331 #define PIX_SAD(suf)\
332 static int sad8_ ## suf(void *v, uint8_t *blk2, uint8_t *blk1, int stride, int h)\
335 __asm__ volatile("pxor %%mm7, %%mm7 \n\t"\
336 "pxor %%mm6, %%mm6 \n\t":);\
338 sad8_1_ ## suf(blk1, blk2, stride, 8);\
340 return sum_ ## suf();\
342 static int sad8_x2_ ## suf(void *v, uint8_t *blk2, uint8_t *blk1, int stride, int h)\
345 __asm__ volatile("pxor %%mm7, %%mm7 \n\t"\
346 "pxor %%mm6, %%mm6 \n\t"\
347 "movq %0, %%mm5 \n\t"\
348 :: "m"(round_tab[1]) \
351 sad8_x2a_ ## suf(blk1, blk2, stride, 8);\
353 return sum_ ## suf();\
356 static int sad8_y2_ ## suf(void *v, uint8_t *blk2, uint8_t *blk1, int stride, int h)\
359 __asm__ volatile("pxor %%mm7, %%mm7 \n\t"\
360 "pxor %%mm6, %%mm6 \n\t"\
361 "movq %0, %%mm5 \n\t"\
362 :: "m"(round_tab[1]) \
365 sad8_y2a_ ## suf(blk1, blk2, stride, 8);\
367 return sum_ ## suf();\
370 static int sad8_xy2_ ## suf(void *v, uint8_t *blk2, uint8_t *blk1, int stride, int h)\
373 __asm__ volatile("pxor %%mm7, %%mm7 \n\t"\
374 "pxor %%mm6, %%mm6 \n\t"\
377 sad8_4_ ## suf(blk1, blk2, stride, 8);\
379 return sum_ ## suf();\
382 static int sad16_ ## suf(void *v, uint8_t *blk2, uint8_t *blk1, int stride, int h)\
384 __asm__ volatile("pxor %%mm7, %%mm7 \n\t"\
385 "pxor %%mm6, %%mm6 \n\t":);\
387 sad8_1_ ## suf(blk1 , blk2 , stride, h);\
388 sad8_1_ ## suf(blk1+8, blk2+8, stride, h);\
390 return sum_ ## suf();\
392 static int sad16_x2_ ## suf(void *v, uint8_t *blk2, uint8_t *blk1, int stride, int h)\
394 __asm__ volatile("pxor %%mm7, %%mm7 \n\t"\
395 "pxor %%mm6, %%mm6 \n\t"\
396 "movq %0, %%mm5 \n\t"\
397 :: "m"(round_tab[1]) \
400 sad8_x2a_ ## suf(blk1 , blk2 , stride, h);\
401 sad8_x2a_ ## suf(blk1+8, blk2+8, stride, h);\
403 return sum_ ## suf();\
405 static int sad16_y2_ ## suf(void *v, uint8_t *blk2, uint8_t *blk1, int stride, int h)\
407 __asm__ volatile("pxor %%mm7, %%mm7 \n\t"\
408 "pxor %%mm6, %%mm6 \n\t"\
409 "movq %0, %%mm5 \n\t"\
410 :: "m"(round_tab[1]) \
413 sad8_y2a_ ## suf(blk1 , blk2 , stride, h);\
414 sad8_y2a_ ## suf(blk1+8, blk2+8, stride, h);\
416 return sum_ ## suf();\
418 static int sad16_xy2_ ## suf(void *v, uint8_t *blk2, uint8_t *blk1, int stride, int h)\
420 __asm__ volatile("pxor %%mm7, %%mm7 \n\t"\
421 "pxor %%mm6, %%mm6 \n\t"\
424 sad8_4_ ## suf(blk1 , blk2 , stride, h);\
425 sad8_4_ ## suf(blk1+8, blk2+8, stride, h);\
427 return sum_ ## suf();\
433 #endif /* HAVE_INLINE_ASM */
435 av_cold void ff_dsputil_init_pix_mmx(DSPContext *c, AVCodecContext *avctx)
438 int mm_flags = av_get_cpu_flags();
440 if (mm_flags & AV_CPU_FLAG_MMX) {
441 c->pix_abs[0][0] = sad16_mmx;
442 c->pix_abs[0][1] = sad16_x2_mmx;
443 c->pix_abs[0][2] = sad16_y2_mmx;
444 c->pix_abs[0][3] = sad16_xy2_mmx;
445 c->pix_abs[1][0] = sad8_mmx;
446 c->pix_abs[1][1] = sad8_x2_mmx;
447 c->pix_abs[1][2] = sad8_y2_mmx;
448 c->pix_abs[1][3] = sad8_xy2_mmx;
450 c->sad[0]= sad16_mmx;
453 if (mm_flags & AV_CPU_FLAG_MMXEXT) {
454 c->pix_abs[0][0] = sad16_mmxext;
455 c->pix_abs[1][0] = sad8_mmxext;
457 c->sad[0] = sad16_mmxext;
458 c->sad[1] = sad8_mmxext;
460 if(!(avctx->flags & CODEC_FLAG_BITEXACT)){
461 c->pix_abs[0][1] = sad16_x2_mmxext;
462 c->pix_abs[0][2] = sad16_y2_mmxext;
463 c->pix_abs[0][3] = sad16_xy2_mmxext;
464 c->pix_abs[1][1] = sad8_x2_mmxext;
465 c->pix_abs[1][2] = sad8_y2_mmxext;
466 c->pix_abs[1][3] = sad8_xy2_mmxext;
469 if ((mm_flags & AV_CPU_FLAG_SSE2) && !(mm_flags & AV_CPU_FLAG_3DNOW) && avctx->codec_id != AV_CODEC_ID_SNOW) {
470 c->sad[0]= sad16_sse2;
472 #endif /* HAVE_INLINE_ASM */