motion_est_mmx.c 16 KB
Newer Older
1 2
/*
 * MMX optimized motion estimation
3
 * Copyright (c) 2001 Fabrice Bellard
4
 * Copyright (c) 2002-2004 Michael Niedermayer
5
 *
6 7
 * mostly by Michael Niedermayer <michaelni@gmx.at>
 *
8 9 10
 * This file is part of FFmpeg.
 *
 * FFmpeg is free software; you can redistribute it and/or
11 12
 * modify it under the terms of the GNU Lesser General Public
 * 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
 * FFmpeg is distributed in the hope that it will be useful,
16
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 18
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
 * 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 FFmpeg; if not, write to the Free Software
22
 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
23
 */
24 25 26

#include "libavutil/x86_cpu.h"
#include "libavcodec/dsputil.h"
27
#include "dsputil_mmx.h"
28

29
DECLARE_ASM_CONST(8, uint64_t, round_tab)[3]={
30 31 32
0x0000000000000000ULL,
0x0001000100010001ULL,
0x0002000200020002ULL,
33
};
34

35
DECLARE_ASM_CONST(8, uint64_t, bone)= 0x0101010101010101LL;
Michael Niedermayer's avatar
Michael Niedermayer committed
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 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67
        "1:                             \n\t"
        "movq (%1, %%"REG_a"), %%mm0    \n\t"
        "movq (%2, %%"REG_a"), %%mm2    \n\t"
        "movq (%2, %%"REG_a"), %%mm4    \n\t"
        "add %3, %%"REG_a"              \n\t"
        "psubusb %%mm0, %%mm2           \n\t"
        "psubusb %%mm4, %%mm0           \n\t"
        "movq (%1, %%"REG_a"), %%mm1    \n\t"
        "movq (%2, %%"REG_a"), %%mm3    \n\t"
        "movq (%2, %%"REG_a"), %%mm5    \n\t"
        "psubusb %%mm1, %%mm3           \n\t"
        "psubusb %%mm5, %%mm1           \n\t"
        "por %%mm2, %%mm0               \n\t"
        "por %%mm1, %%mm3               \n\t"
        "movq %%mm0, %%mm1              \n\t"
        "movq %%mm3, %%mm2              \n\t"
        "punpcklbw %%mm7, %%mm0         \n\t"
        "punpckhbw %%mm7, %%mm1         \n\t"
        "punpcklbw %%mm7, %%mm3         \n\t"
        "punpckhbw %%mm7, %%mm2         \n\t"
        "paddw %%mm1, %%mm0             \n\t"
        "paddw %%mm3, %%mm2             \n\t"
        "paddw %%mm2, %%mm0             \n\t"
        "paddw %%mm0, %%mm6             \n\t"
        "add %3, %%"REG_a"              \n\t"
        " 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"
Loren Merritt's avatar
Loren Merritt committed
78 79 80 81
        "movq (%1), %%mm0               \n\t"
        "movq (%1, %3), %%mm1           \n\t"
        "psadbw (%2), %%mm0             \n\t"
        "psadbw (%2, %3), %%mm1         \n\t"
82
        "paddw %%mm0, %%mm6             \n\t"
Loren Merritt's avatar
Loren Merritt committed
83 84 85 86 87 88
        "paddw %%mm1, %%mm6             \n\t"
        "lea (%1,%3,2), %1              \n\t"
        "lea (%2,%3,2), %2              \n\t"
        "sub $2, %0                     \n\t"
        " jg 1b                         \n\t"
        : "+r" (h), "+r" (blk1), "+r" (blk2)
89
        : "r" ((x86_reg)stride)
90
    );
91 92
}

93 94 95
static int sad16_sse2(void *v, uint8_t *blk2, uint8_t *blk1, int stride, int h)
{
    int ret;
96
    __asm__ volatile(
97
        "pxor %%xmm2, %%xmm2            \n\t"
98
        ".p2align 4                     \n\t"
99 100
        "1:                             \n\t"
        "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"
104 105
        "paddw %%xmm0, %%xmm2           \n\t"
        "paddw %%xmm1, %%xmm2           \n\t"
106 107
        "lea (%1,%4,2), %1              \n\t"
        "lea (%2,%4,2), %2              \n\t"
108 109
        "sub $2, %0                     \n\t"
        " jg 1b                         \n\t"
110 111
        "movhlps %%xmm2, %%xmm0         \n\t"
        "paddw   %%xmm0, %%xmm2         \n\t"
112 113 114
        "movd    %%xmm2, %3             \n\t"
        : "+r" (h), "+r" (blk1), "+r" (blk2), "=r"(ret)
        : "r" ((x86_reg)stride)
115 116 117 118
    );
    return ret;
}

Loren Merritt's avatar
Loren Merritt committed
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"
Loren Merritt's avatar
Loren Merritt committed
124 125 126 127 128 129
        "movq (%1), %%mm0               \n\t"
        "movq (%1, %3), %%mm1           \n\t"
        "pavgb 1(%1), %%mm0             \n\t"
        "pavgb 1(%1, %3), %%mm1         \n\t"
        "psadbw (%2), %%mm0             \n\t"
        "psadbw (%2, %3), %%mm1         \n\t"
130
        "paddw %%mm0, %%mm6             \n\t"
Loren Merritt's avatar
Loren Merritt committed
131 132 133 134 135 136
        "paddw %%mm1, %%mm6             \n\t"
        "lea (%1,%3,2), %1              \n\t"
        "lea (%2,%3,2), %2              \n\t"
        "sub $2, %0                     \n\t"
        " jg 1b                         \n\t"
        : "+r" (h), "+r" (blk1), "+r" (blk2)
137
        : "r" ((x86_reg)stride)
Loren Merritt's avatar
Loren Merritt committed
138 139 140 141 142
    );
}

static inline void sad8_y2a_mmx2(uint8_t *blk1, uint8_t *blk2, int stride, int h)
{
143
    __asm__ volatile(
Loren Merritt's avatar
Loren Merritt committed
144 145
        "movq (%1), %%mm0               \n\t"
        "add %3, %1                     \n\t"
146
        ".p2align 4                     \n\t"
Loren Merritt's avatar
Loren Merritt committed
147 148 149 150 151 152 153 154 155 156 157 158 159 160 161
        "1:                             \n\t"
        "movq (%1), %%mm1               \n\t"
        "movq (%1, %3), %%mm2           \n\t"
        "pavgb %%mm1, %%mm0             \n\t"
        "pavgb %%mm2, %%mm1             \n\t"
        "psadbw (%2), %%mm0             \n\t"
        "psadbw (%2, %3), %%mm1         \n\t"
        "paddw %%mm0, %%mm6             \n\t"
        "paddw %%mm1, %%mm6             \n\t"
        "movq %%mm2, %%mm0              \n\t"
        "lea (%1,%3,2), %1              \n\t"
        "lea (%2,%3,2), %2              \n\t"
        "sub $2, %0                     \n\t"
        " jg 1b                         \n\t"
        : "+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)
Loren Merritt's avatar
Loren Merritt committed
167
{
168
    __asm__ volatile(
169
        "movq "MANGLE(bone)", %%mm5     \n\t"
Loren Merritt's avatar
Loren Merritt committed
170 171 172
        "movq (%1), %%mm0               \n\t"
        "pavgb 1(%1), %%mm0             \n\t"
        "add %3, %1                     \n\t"
173
        ".p2align 4                     \n\t"
174
        "1:                             \n\t"
Loren Merritt's avatar
Loren Merritt committed
175 176 177 178
        "movq (%1), %%mm1               \n\t"
        "movq (%1,%3), %%mm2            \n\t"
        "pavgb 1(%1), %%mm1             \n\t"
        "pavgb 1(%1,%3), %%mm2          \n\t"
Loren Merritt's avatar
Loren Merritt committed
179 180
        "psubusb %%mm5, %%mm1           \n\t"
        "pavgb %%mm1, %%mm0             \n\t"
Loren Merritt's avatar
Loren Merritt committed
181 182 183
        "pavgb %%mm2, %%mm1             \n\t"
        "psadbw (%2), %%mm0             \n\t"
        "psadbw (%2,%3), %%mm1          \n\t"
184
        "paddw %%mm0, %%mm6             \n\t"
Loren Merritt's avatar
Loren Merritt committed
185 186 187 188 189 190 191
        "paddw %%mm1, %%mm6             \n\t"
        "movq %%mm2, %%mm0              \n\t"
        "lea (%1,%3,2), %1              \n\t"
        "lea (%2,%3,2), %2              \n\t"
        "sub $2, %0                     \n\t"
        " jg 1b                         \n\t"
        : "+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"
201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228
        "1:                             \n\t"
        "movq (%1, %%"REG_a"), %%mm0    \n\t"
        "movq (%2, %%"REG_a"), %%mm1    \n\t"
        "movq (%1, %%"REG_a"), %%mm2    \n\t"
        "movq (%2, %%"REG_a"), %%mm3    \n\t"
        "punpcklbw %%mm7, %%mm0         \n\t"
        "punpcklbw %%mm7, %%mm1         \n\t"
        "punpckhbw %%mm7, %%mm2         \n\t"
        "punpckhbw %%mm7, %%mm3         \n\t"
        "paddw %%mm0, %%mm1             \n\t"
        "paddw %%mm2, %%mm3             \n\t"
        "movq (%3, %%"REG_a"), %%mm4    \n\t"
        "movq (%3, %%"REG_a"), %%mm2    \n\t"
        "paddw %%mm5, %%mm1             \n\t"
        "paddw %%mm5, %%mm3             \n\t"
        "psrlw $1, %%mm1                \n\t"
        "psrlw $1, %%mm3                \n\t"
        "packuswb %%mm3, %%mm1          \n\t"
        "psubusb %%mm1, %%mm4           \n\t"
        "psubusb %%mm2, %%mm1           \n\t"
        "por %%mm4, %%mm1               \n\t"
        "movq %%mm1, %%mm0              \n\t"
        "punpcklbw %%mm7, %%mm0         \n\t"
        "punpckhbw %%mm7, %%mm1         \n\t"
        "paddw %%mm1, %%mm0             \n\t"
        "paddw %%mm0, %%mm6             \n\t"
        "add %4, %%"REG_a"              \n\t"
        " 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(
238 239
        "movq (%1, %%"REG_a"), %%mm0    \n\t"
        "movq 1(%1, %%"REG_a"), %%mm2   \n\t"
Loren Merritt's avatar
Loren Merritt committed
240 241 242
        "movq %%mm0, %%mm1              \n\t"
        "movq %%mm2, %%mm3              \n\t"
        "punpcklbw %%mm7, %%mm0         \n\t"
243
        "punpckhbw %%mm7, %%mm1         \n\t"
Loren Merritt's avatar
Loren Merritt committed
244 245 246 247
        "punpcklbw %%mm7, %%mm2         \n\t"
        "punpckhbw %%mm7, %%mm3         \n\t"
        "paddw %%mm2, %%mm0             \n\t"
        "paddw %%mm3, %%mm1             \n\t"
248
        ".p2align 4                     \n\t"
Loren Merritt's avatar
Loren Merritt committed
249 250 251 252 253 254 255 256 257 258 259 260 261 262 263
        "1:                             \n\t"
        "movq (%2, %%"REG_a"), %%mm2    \n\t"
        "movq 1(%2, %%"REG_a"), %%mm4   \n\t"
        "movq %%mm2, %%mm3              \n\t"
        "movq %%mm4, %%mm5              \n\t"
        "punpcklbw %%mm7, %%mm2         \n\t"
        "punpckhbw %%mm7, %%mm3         \n\t"
        "punpcklbw %%mm7, %%mm4         \n\t"
        "punpckhbw %%mm7, %%mm5         \n\t"
        "paddw %%mm4, %%mm2             \n\t"
        "paddw %%mm5, %%mm3             \n\t"
        "movq 16+"MANGLE(round_tab)", %%mm5 \n\t"
        "paddw %%mm2, %%mm0             \n\t"
        "paddw %%mm3, %%mm1             \n\t"
        "paddw %%mm5, %%mm0             \n\t"
264
        "paddw %%mm5, %%mm1             \n\t"
Loren Merritt's avatar
Loren Merritt committed
265 266 267
        "movq (%3, %%"REG_a"), %%mm4    \n\t"
        "movq (%3, %%"REG_a"), %%mm5    \n\t"
        "psrlw $2, %%mm0                \n\t"
268
        "psrlw $2, %%mm1                \n\t"
Loren Merritt's avatar
Loren Merritt committed
269 270 271 272 273
        "packuswb %%mm1, %%mm0          \n\t"
        "psubusb %%mm0, %%mm4           \n\t"
        "psubusb %%mm5, %%mm0           \n\t"
        "por %%mm4, %%mm0               \n\t"
        "movq %%mm0, %%mm4              \n\t"
274
        "punpcklbw %%mm7, %%mm0         \n\t"
Loren Merritt's avatar
Loren Merritt committed
275
        "punpckhbw %%mm7, %%mm4         \n\t"
276
        "paddw %%mm0, %%mm6             \n\t"
Loren Merritt's avatar
Loren Merritt committed
277 278 279
        "paddw %%mm4, %%mm6             \n\t"
        "movq  %%mm2, %%mm0             \n\t"
        "movq  %%mm3, %%mm1             \n\t"
280 281
        "add %4, %%"REG_a"              \n\t"
        " 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(
291 292 293 294 295 296 297
        "movq %%mm6, %%mm0              \n\t"
        "psrlq $32, %%mm6               \n\t"
        "paddw %%mm0, %%mm6             \n\t"
        "movq %%mm6, %%mm0              \n\t"
        "psrlq $16, %%mm6               \n\t"
        "paddw %%mm0, %%mm6             \n\t"
        "movd %%mm6, %0                 \n\t"
298 299 300
        : "=r" (ret)
    );
    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"
308 309 310
        : "=r" (ret)
    );
    return ret;
311 312
}

Loren Merritt's avatar
Loren Merritt committed
313 314 315 316 317 318 319 320 321
static inline void sad8_x2a_mmx(uint8_t *blk1, uint8_t *blk2, int stride, int h)
{
    sad8_2_mmx(blk1, blk1+1, blk2, stride, h);
}
static inline void sad8_y2a_mmx(uint8_t *blk1, uint8_t *blk2, int stride, int h)
{
    sad8_2_mmx(blk1, blk1+stride, blk2, stride, h);
}

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);\
331 332 333
\
    return sum_ ## suf();\
}\
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"\
340 341 342
                 :: "m"(round_tab[1]) \
                 );\
\
Loren Merritt's avatar
Loren Merritt committed
343
    sad8_x2a_ ## suf(blk1, blk2, stride, 8);\
344 345 346 347
\
    return sum_ ## suf();\
}\
\
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"\
352 353
                 "pxor %%mm6, %%mm6     \n\t"\
                 "movq %0, %%mm5        \n\t"\
354 355 356
                 :: "m"(round_tab[1]) \
                 );\
\
Loren Merritt's avatar
Loren Merritt committed
357
    sad8_y2a_ ## suf(blk1, blk2, stride, 8);\
358 359 360 361
\
    return sum_ ## suf();\
}\
\
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"\
Loren Merritt's avatar
Loren Merritt committed
367
                 ::);\
368
\
369
    sad8_4_ ## suf(blk1, blk2, stride, 8);\
370 371 372 373
\
    return sum_ ## suf();\
}\
\
374
static int sad16_ ## suf(void *v, uint8_t *blk2, uint8_t *blk1, int stride, int h)\
Michael Niedermayer's avatar
Michael Niedermayer committed
375
{\
376
    __asm__ volatile("pxor %%mm7, %%mm7     \n\t"\
377
                 "pxor %%mm6, %%mm6     \n\t":);\
Michael Niedermayer's avatar
Michael Niedermayer committed
378
\
379 380
    sad8_1_ ## suf(blk1  , blk2  , stride, h);\
    sad8_1_ ## suf(blk1+8, blk2+8, stride, h);\
Michael Niedermayer's avatar
Michael Niedermayer committed
381 382 383
\
    return sum_ ## suf();\
}\
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"\
387 388
                 "pxor %%mm6, %%mm6     \n\t"\
                 "movq %0, %%mm5        \n\t"\
389 390 391
                 :: "m"(round_tab[1]) \
                 );\
\
Loren Merritt's avatar
Loren Merritt committed
392 393
    sad8_x2a_ ## suf(blk1  , blk2  , stride, h);\
    sad8_x2a_ ## suf(blk1+8, blk2+8, stride, h);\
394 395 396
\
    return sum_ ## suf();\
}\
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"\
400 401
                 "pxor %%mm6, %%mm6     \n\t"\
                 "movq %0, %%mm5        \n\t"\
402 403 404
                 :: "m"(round_tab[1]) \
                 );\
\
Loren Merritt's avatar
Loren Merritt committed
405 406
    sad8_y2a_ ## suf(blk1  , blk2  , stride, h);\
    sad8_y2a_ ## suf(blk1+8, blk2+8, stride, h);\
407 408 409
\
    return sum_ ## suf();\
}\
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"\
Loren Merritt's avatar
Loren Merritt committed
414
                 ::);\
415
\
416 417
    sad8_4_ ## suf(blk1  , blk2  , stride, h);\
    sad8_4_ ## suf(blk1+8, blk2+8, stride, h);\
418 419 420
\
    return sum_ ## suf();\
}\
421

422 423
PIX_SAD(mmx)
PIX_SAD(mmx2)
424

425
void dsputil_init_pix_mmx(DSPContext* c, AVCodecContext *avctx)
426
{
427
    int mm_flags = av_get_cpu_flags();
428

429
    if (mm_flags & AV_CPU_FLAG_MMX) {
430 431 432 433 434 435 436 437
        c->pix_abs[0][0] = sad16_mmx;
        c->pix_abs[0][1] = sad16_x2_mmx;
        c->pix_abs[0][2] = sad16_y2_mmx;
        c->pix_abs[0][3] = sad16_xy2_mmx;
        c->pix_abs[1][0] = sad8_mmx;
        c->pix_abs[1][1] = sad8_x2_mmx;
        c->pix_abs[1][2] = sad8_y2_mmx;
        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) {
443 444
        c->pix_abs[0][0] = sad16_mmx2;
        c->pix_abs[1][0] = sad8_mmx2;
445

446 447
        c->sad[0]= sad16_mmx2;
        c->sad[1]= sad8_mmx2;
448

449
        if(!(avctx->flags & CODEC_FLAG_BITEXACT)){
450 451 452 453 454 455
            c->pix_abs[0][1] = sad16_x2_mmx2;
            c->pix_abs[0][2] = sad16_y2_mmx2;
            c->pix_abs[0][3] = sad16_xy2_mmx2;
            c->pix_abs[1][1] = sad8_x2_mmx2;
            c->pix_abs[1][2] = sad8_y2_mmx2;
            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) {
459 460
        c->sad[0]= sad16_sse2;
    }
461
}