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 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/internal.h"
27 #include "libavutil/mem.h"
28 #include "libavutil/x86/asm.h"
29 #include "libavutil/x86/cpu.h"
30 #include "libavcodec/mpegvideo.h"
31 #include "dsputil_x86.h"
35 DECLARE_ASM_CONST(8, uint64_t, round_tab)[3] = {
36 0x0000000000000000ULL,
37 0x0001000100010001ULL,
38 0x0002000200020002ULL,
41 DECLARE_ASM_CONST(8, uint64_t, bone) = 0x0101010101010101LL;
43 static inline void sad8_1_mmx(uint8_t *blk1, uint8_t *blk2, int stride, int h)
45 x86_reg len = -(stride * h);
49 "movq (%1, %%"REG_a"), %%mm0 \n\t"
50 "movq (%2, %%"REG_a"), %%mm2 \n\t"
51 "movq (%2, %%"REG_a"), %%mm4 \n\t"
52 "add %3, %%"REG_a" \n\t"
53 "psubusb %%mm0, %%mm2 \n\t"
54 "psubusb %%mm4, %%mm0 \n\t"
55 "movq (%1, %%"REG_a"), %%mm1 \n\t"
56 "movq (%2, %%"REG_a"), %%mm3 \n\t"
57 "movq (%2, %%"REG_a"), %%mm5 \n\t"
58 "psubusb %%mm1, %%mm3 \n\t"
59 "psubusb %%mm5, %%mm1 \n\t"
60 "por %%mm2, %%mm0 \n\t"
61 "por %%mm1, %%mm3 \n\t"
62 "movq %%mm0, %%mm1 \n\t"
63 "movq %%mm3, %%mm2 \n\t"
64 "punpcklbw %%mm7, %%mm0 \n\t"
65 "punpckhbw %%mm7, %%mm1 \n\t"
66 "punpcklbw %%mm7, %%mm3 \n\t"
67 "punpckhbw %%mm7, %%mm2 \n\t"
68 "paddw %%mm1, %%mm0 \n\t"
69 "paddw %%mm3, %%mm2 \n\t"
70 "paddw %%mm2, %%mm0 \n\t"
71 "paddw %%mm0, %%mm6 \n\t"
72 "add %3, %%"REG_a" \n\t"
75 : "r" (blk1 - len), "r" (blk2 - len), "r" ((x86_reg) stride));
78 static inline void sad8_1_mmxext(uint8_t *blk1, uint8_t *blk2,
84 "movq (%1), %%mm0 \n\t"
85 "movq (%1, %3), %%mm1 \n\t"
86 "psadbw (%2), %%mm0 \n\t"
87 "psadbw (%2, %3), %%mm1 \n\t"
88 "paddw %%mm0, %%mm6 \n\t"
89 "paddw %%mm1, %%mm6 \n\t"
90 "lea (%1,%3,2), %1 \n\t"
91 "lea (%2,%3,2), %2 \n\t"
94 : "+r" (h), "+r" (blk1), "+r" (blk2)
95 : "r" ((x86_reg) stride));
98 static int sad16_sse2(MpegEncContext *v, uint8_t *blk2, uint8_t *blk1,
103 "pxor %%xmm2, %%xmm2 \n\t"
106 "movdqu (%1), %%xmm0 \n\t"
107 "movdqu (%1, %4), %%xmm1 \n\t"
108 "psadbw (%2), %%xmm0 \n\t"
109 "psadbw (%2, %4), %%xmm1 \n\t"
110 "paddw %%xmm0, %%xmm2 \n\t"
111 "paddw %%xmm1, %%xmm2 \n\t"
112 "lea (%1,%4,2), %1 \n\t"
113 "lea (%2,%4,2), %2 \n\t"
116 "movhlps %%xmm2, %%xmm0 \n\t"
117 "paddw %%xmm0, %%xmm2 \n\t"
118 "movd %%xmm2, %3 \n\t"
119 : "+r" (h), "+r" (blk1), "+r" (blk2), "=r" (ret)
120 : "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));
146 static inline void sad8_y2a_mmxext(uint8_t *blk1, uint8_t *blk2,
150 "movq (%1), %%mm0 \n\t"
154 "movq (%1), %%mm1 \n\t"
155 "movq (%1, %3), %%mm2 \n\t"
156 "pavgb %%mm1, %%mm0 \n\t"
157 "pavgb %%mm2, %%mm1 \n\t"
158 "psadbw (%2), %%mm0 \n\t"
159 "psadbw (%2, %3), %%mm1 \n\t"
160 "paddw %%mm0, %%mm6 \n\t"
161 "paddw %%mm1, %%mm6 \n\t"
162 "movq %%mm2, %%mm0 \n\t"
163 "lea (%1,%3,2), %1 \n\t"
164 "lea (%2,%3,2), %2 \n\t"
167 : "+r" (h), "+r" (blk1), "+r" (blk2)
168 : "r" ((x86_reg) stride));
171 static inline void sad8_4_mmxext(uint8_t *blk1, uint8_t *blk2,
175 "movq "MANGLE(bone)", %%mm5 \n\t"
176 "movq (%1), %%mm0 \n\t"
177 "pavgb 1(%1), %%mm0 \n\t"
181 "movq (%1), %%mm1 \n\t"
182 "movq (%1,%3), %%mm2 \n\t"
183 "pavgb 1(%1), %%mm1 \n\t"
184 "pavgb 1(%1,%3), %%mm2 \n\t"
185 "psubusb %%mm5, %%mm1 \n\t"
186 "pavgb %%mm1, %%mm0 \n\t"
187 "pavgb %%mm2, %%mm1 \n\t"
188 "psadbw (%2), %%mm0 \n\t"
189 "psadbw (%2,%3), %%mm1 \n\t"
190 "paddw %%mm0, %%mm6 \n\t"
191 "paddw %%mm1, %%mm6 \n\t"
192 "movq %%mm2, %%mm0 \n\t"
193 "lea (%1,%3,2), %1 \n\t"
194 "lea (%2,%3,2), %2 \n\t"
197 : "+r" (h), "+r" (blk1), "+r" (blk2)
198 : "r" ((x86_reg) stride));
201 static inline void sad8_2_mmx(uint8_t *blk1a, uint8_t *blk1b, uint8_t *blk2,
204 x86_reg len = -(stride * h);
208 "movq (%1, %%"REG_a"), %%mm0 \n\t"
209 "movq (%2, %%"REG_a"), %%mm1 \n\t"
210 "movq (%1, %%"REG_a"), %%mm2 \n\t"
211 "movq (%2, %%"REG_a"), %%mm3 \n\t"
212 "punpcklbw %%mm7, %%mm0 \n\t"
213 "punpcklbw %%mm7, %%mm1 \n\t"
214 "punpckhbw %%mm7, %%mm2 \n\t"
215 "punpckhbw %%mm7, %%mm3 \n\t"
216 "paddw %%mm0, %%mm1 \n\t"
217 "paddw %%mm2, %%mm3 \n\t"
218 "movq (%3, %%"REG_a"), %%mm4 \n\t"
219 "movq (%3, %%"REG_a"), %%mm2 \n\t"
220 "paddw %%mm5, %%mm1 \n\t"
221 "paddw %%mm5, %%mm3 \n\t"
222 "psrlw $1, %%mm1 \n\t"
223 "psrlw $1, %%mm3 \n\t"
224 "packuswb %%mm3, %%mm1 \n\t"
225 "psubusb %%mm1, %%mm4 \n\t"
226 "psubusb %%mm2, %%mm1 \n\t"
227 "por %%mm4, %%mm1 \n\t"
228 "movq %%mm1, %%mm0 \n\t"
229 "punpcklbw %%mm7, %%mm0 \n\t"
230 "punpckhbw %%mm7, %%mm1 \n\t"
231 "paddw %%mm1, %%mm0 \n\t"
232 "paddw %%mm0, %%mm6 \n\t"
233 "add %4, %%"REG_a" \n\t"
236 : "r" (blk1a - len), "r" (blk1b - len), "r" (blk2 - len),
237 "r" ((x86_reg) stride));
240 static inline void sad8_4_mmx(uint8_t *blk1, uint8_t *blk2, int stride, int h)
242 x86_reg len = -(stride * h);
244 "movq (%1, %%"REG_a"), %%mm0 \n\t"
245 "movq 1(%1, %%"REG_a"), %%mm2 \n\t"
246 "movq %%mm0, %%mm1 \n\t"
247 "movq %%mm2, %%mm3 \n\t"
248 "punpcklbw %%mm7, %%mm0 \n\t"
249 "punpckhbw %%mm7, %%mm1 \n\t"
250 "punpcklbw %%mm7, %%mm2 \n\t"
251 "punpckhbw %%mm7, %%mm3 \n\t"
252 "paddw %%mm2, %%mm0 \n\t"
253 "paddw %%mm3, %%mm1 \n\t"
256 "movq (%2, %%"REG_a"), %%mm2 \n\t"
257 "movq 1(%2, %%"REG_a"), %%mm4 \n\t"
258 "movq %%mm2, %%mm3 \n\t"
259 "movq %%mm4, %%mm5 \n\t"
260 "punpcklbw %%mm7, %%mm2 \n\t"
261 "punpckhbw %%mm7, %%mm3 \n\t"
262 "punpcklbw %%mm7, %%mm4 \n\t"
263 "punpckhbw %%mm7, %%mm5 \n\t"
264 "paddw %%mm4, %%mm2 \n\t"
265 "paddw %%mm5, %%mm3 \n\t"
266 "movq 16+"MANGLE(round_tab)", %%mm5 \n\t"
267 "paddw %%mm2, %%mm0 \n\t"
268 "paddw %%mm3, %%mm1 \n\t"
269 "paddw %%mm5, %%mm0 \n\t"
270 "paddw %%mm5, %%mm1 \n\t"
271 "movq (%3, %%"REG_a"), %%mm4 \n\t"
272 "movq (%3, %%"REG_a"), %%mm5 \n\t"
273 "psrlw $2, %%mm0 \n\t"
274 "psrlw $2, %%mm1 \n\t"
275 "packuswb %%mm1, %%mm0 \n\t"
276 "psubusb %%mm0, %%mm4 \n\t"
277 "psubusb %%mm5, %%mm0 \n\t"
278 "por %%mm4, %%mm0 \n\t"
279 "movq %%mm0, %%mm4 \n\t"
280 "punpcklbw %%mm7, %%mm0 \n\t"
281 "punpckhbw %%mm7, %%mm4 \n\t"
282 "paddw %%mm0, %%mm6 \n\t"
283 "paddw %%mm4, %%mm6 \n\t"
284 "movq %%mm2, %%mm0 \n\t"
285 "movq %%mm3, %%mm1 \n\t"
286 "add %4, %%"REG_a" \n\t"
289 : "r" (blk1 - len), "r" (blk1 - len + stride), "r" (blk2 - len),
290 "r" ((x86_reg) stride));
293 static inline int sum_mmx(void)
297 "movq %%mm6, %%mm0 \n\t"
298 "psrlq $32, %%mm6 \n\t"
299 "paddw %%mm0, %%mm6 \n\t"
300 "movq %%mm6, %%mm0 \n\t"
301 "psrlq $16, %%mm6 \n\t"
302 "paddw %%mm0, %%mm6 \n\t"
303 "movd %%mm6, %0 \n\t"
308 static inline int sum_mmxext(void)
312 "movd %%mm6, %0 \n\t"
317 static inline void sad8_x2a_mmx(uint8_t *blk1, uint8_t *blk2, int stride, int h)
319 sad8_2_mmx(blk1, blk1 + 1, blk2, stride, h);
322 static inline void sad8_y2a_mmx(uint8_t *blk1, uint8_t *blk2, int stride, int h)
324 sad8_2_mmx(blk1, blk1 + stride, blk2, stride, h);
327 #define PIX_SAD(suf) \
328 static int sad8_ ## suf(MpegEncContext *v, uint8_t *blk2, \
329 uint8_t *blk1, int stride, int h) \
333 "pxor %%mm7, %%mm7 \n\t" \
334 "pxor %%mm6, %%mm6 \n\t" \
337 sad8_1_ ## suf(blk1, blk2, stride, 8); \
339 return sum_ ## suf(); \
342 static int sad8_x2_ ## suf(MpegEncContext *v, uint8_t *blk2, \
343 uint8_t *blk1, int stride, int h) \
347 "pxor %%mm7, %%mm7 \n\t" \
348 "pxor %%mm6, %%mm6 \n\t" \
349 "movq %0, %%mm5 \n\t" \
350 :: "m" (round_tab[1])); \
352 sad8_x2a_ ## suf(blk1, blk2, stride, 8); \
354 return sum_ ## suf(); \
357 static int sad8_y2_ ## suf(MpegEncContext *v, uint8_t *blk2, \
358 uint8_t *blk1, int stride, int h) \
362 "pxor %%mm7, %%mm7 \n\t" \
363 "pxor %%mm6, %%mm6 \n\t" \
364 "movq %0, %%mm5 \n\t" \
365 :: "m" (round_tab[1])); \
367 sad8_y2a_ ## suf(blk1, blk2, stride, 8); \
369 return sum_ ## suf(); \
372 static int sad8_xy2_ ## suf(MpegEncContext *v, uint8_t *blk2, \
373 uint8_t *blk1, int stride, int h) \
377 "pxor %%mm7, %%mm7 \n\t" \
378 "pxor %%mm6, %%mm6 \n\t" \
381 sad8_4_ ## suf(blk1, blk2, stride, 8); \
383 return sum_ ## suf(); \
386 static int sad16_ ## suf(MpegEncContext *v, uint8_t *blk2, \
387 uint8_t *blk1, int stride, int h) \
390 "pxor %%mm7, %%mm7 \n\t" \
391 "pxor %%mm6, %%mm6 \n\t" \
394 sad8_1_ ## suf(blk1, blk2, stride, h); \
395 sad8_1_ ## suf(blk1 + 8, blk2 + 8, stride, h); \
397 return sum_ ## suf(); \
400 static int sad16_x2_ ## suf(MpegEncContext *v, uint8_t *blk2, \
401 uint8_t *blk1, int stride, int h) \
404 "pxor %%mm7, %%mm7 \n\t" \
405 "pxor %%mm6, %%mm6 \n\t" \
406 "movq %0, %%mm5 \n\t" \
407 :: "m" (round_tab[1])); \
409 sad8_x2a_ ## suf(blk1, blk2, stride, h); \
410 sad8_x2a_ ## suf(blk1 + 8, blk2 + 8, stride, h); \
412 return sum_ ## suf(); \
415 static int sad16_y2_ ## suf(MpegEncContext *v, uint8_t *blk2, \
416 uint8_t *blk1, int stride, int h) \
419 "pxor %%mm7, %%mm7 \n\t" \
420 "pxor %%mm6, %%mm6 \n\t" \
421 "movq %0, %%mm5 \n\t" \
422 :: "m" (round_tab[1])); \
424 sad8_y2a_ ## suf(blk1, blk2, stride, h); \
425 sad8_y2a_ ## suf(blk1 + 8, blk2 + 8, stride, h); \
427 return sum_ ## suf(); \
430 static int sad16_xy2_ ## suf(MpegEncContext *v, uint8_t *blk2, \
431 uint8_t *blk1, int stride, int h) \
434 "pxor %%mm7, %%mm7 \n\t" \
435 "pxor %%mm6, %%mm6 \n\t" \
438 sad8_4_ ## suf(blk1, blk2, stride, h); \
439 sad8_4_ ## suf(blk1 + 8, blk2 + 8, stride, h); \
441 return sum_ ## suf(); \
447 #endif /* HAVE_INLINE_ASM */
449 av_cold void ff_dsputil_init_pix_mmx(DSPContext *c, AVCodecContext *avctx)
452 int cpu_flags = av_get_cpu_flags();
454 if (INLINE_MMX(cpu_flags)) {
455 c->pix_abs[0][0] = sad16_mmx;
456 c->pix_abs[0][1] = sad16_x2_mmx;
457 c->pix_abs[0][2] = sad16_y2_mmx;
458 c->pix_abs[0][3] = sad16_xy2_mmx;
459 c->pix_abs[1][0] = sad8_mmx;
460 c->pix_abs[1][1] = sad8_x2_mmx;
461 c->pix_abs[1][2] = sad8_y2_mmx;
462 c->pix_abs[1][3] = sad8_xy2_mmx;
464 c->sad[0] = sad16_mmx;
465 c->sad[1] = sad8_mmx;
467 if (INLINE_MMXEXT(cpu_flags)) {
468 c->pix_abs[0][0] = sad16_mmxext;
469 c->pix_abs[1][0] = sad8_mmxext;
471 c->sad[0] = sad16_mmxext;
472 c->sad[1] = sad8_mmxext;
474 if (!(avctx->flags & CODEC_FLAG_BITEXACT)) {
475 c->pix_abs[0][1] = sad16_x2_mmxext;
476 c->pix_abs[0][2] = sad16_y2_mmxext;
477 c->pix_abs[0][3] = sad16_xy2_mmxext;
478 c->pix_abs[1][1] = sad8_x2_mmxext;
479 c->pix_abs[1][2] = sad8_y2_mmxext;
480 c->pix_abs[1][3] = sad8_xy2_mmxext;
483 if (INLINE_SSE2(cpu_flags) && !(cpu_flags & AV_CPU_FLAG_3DNOW)) {
484 c->sad[0] = sad16_sse2;
486 #endif /* HAVE_INLINE_ASM */