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
29DECLARE_ASM_CONST(8, uint64_t, round_tab)[3]={
300x0000000000000000ULL,
310x0001000100010001ULL,
320x0002000200020002ULL,
33};
34
35DECLARE_ASM_CONST(8, uint64_t, bone)= 0x0101010101010101LL;
36
37static 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
73static 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
93static 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
119static 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
141static 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
166static 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
196static 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
234static 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
287static 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
303static 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
313static 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}
317static 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)\
324static 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}\
334static 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\
348static 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\
362static 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\
374static 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}\
384static 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}\
397static 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}\
410static 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
422PIX_SAD(mmx)
423PIX_SAD(mmx2)
424
425void 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        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}
462