1 /*
2 * Motion estimation
3 * Copyright (c) 2000,2001 Fabrice Bellard
4 * Copyright (c) 2002-2004 Michael Niedermayer
5 *
6 * new motion estimation (X1/EPZS) by Michael Niedermayer <michaelni@gmx.at>
7 *
8 * This file is part of FFmpeg.
9 *
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.
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 * 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 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 * @file
27 * Motion estimation.
28 */
29
30 #include <stdlib.h>
31 #include <stdio.h>
32 #include <limits.h>
33
34 #include "avcodec.h"
35 #include "mathops.h"
36 #include "motion_est.h"
37 #include "mpegutils.h"
38 #include "mpegvideoenc.h"
39
40 #define P_LEFT P[1]
41 #define P_TOP P[2]
42 #define P_TOPRIGHT P[3]
43 #define P_MEDIAN P[4]
44 #define P_MV1 P[9]
45
46 #define ME_MAP_SHIFT 3
47 #define ME_MAP_MV_BITS 11
48
49 static int sad_hpel_motion_search(MpegEncContext * s,
50 int *mx_ptr, int *my_ptr, int dmin,
51 int src_index, int ref_index,
52 int size, int h);
53
update_map_generation(MotionEstContext * c)54 static inline unsigned update_map_generation(MotionEstContext *c)
55 {
56 c->map_generation+= 1<<(ME_MAP_MV_BITS*2);
57 if(c->map_generation==0){
58 c->map_generation= 1<<(ME_MAP_MV_BITS*2);
59 memset(c->map, 0, sizeof(uint32_t)*ME_MAP_SIZE);
60 }
61 return c->map_generation;
62 }
63
64 /* shape adaptive search stuff */
65 typedef struct Minima{
66 int height;
67 int x, y;
68 int checked;
69 }Minima;
70
minima_cmp(const void * a,const void * b)71 static int minima_cmp(const void *a, const void *b){
72 const Minima *da = (const Minima *) a;
73 const Minima *db = (const Minima *) b;
74
75 return da->height - db->height;
76 }
77
78 #define FLAG_QPEL 1 //must be 1
79 #define FLAG_CHROMA 2
80 #define FLAG_DIRECT 4
81
init_ref(MotionEstContext * c,uint8_t * src[3],uint8_t * ref[3],uint8_t * ref2[3],int x,int y,int ref_index)82 static inline void init_ref(MotionEstContext *c, uint8_t *src[3], uint8_t *ref[3], uint8_t *ref2[3], int x, int y, int ref_index){
83 const int offset[3]= {
84 y*c-> stride + x,
85 ((y*c->uvstride + x)>>1),
86 ((y*c->uvstride + x)>>1),
87 };
88 int i;
89 for(i=0; i<3; i++){
90 c->src[0][i]= src [i] + offset[i];
91 c->ref[0][i]= ref [i] + offset[i];
92 }
93 if(ref_index){
94 for(i=0; i<3; i++){
95 c->ref[ref_index][i]= ref2[i] + offset[i];
96 }
97 }
98 }
99
get_flags(MotionEstContext * c,int direct,int chroma)100 static int get_flags(MotionEstContext *c, int direct, int chroma){
101 return ((c->avctx->flags&AV_CODEC_FLAG_QPEL) ? FLAG_QPEL : 0)
102 + (direct ? FLAG_DIRECT : 0)
103 + (chroma ? FLAG_CHROMA : 0);
104 }
105
cmp_direct_inline(MpegEncContext * s,const int x,const int y,const int subx,const int suby,const int size,const int h,int ref_index,int src_index,me_cmp_func cmp_func,me_cmp_func chroma_cmp_func,int qpel)106 static av_always_inline int cmp_direct_inline(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
107 const int size, const int h, int ref_index, int src_index,
108 me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, int qpel){
109 MotionEstContext * const c= &s->me;
110 const int stride= c->stride;
111 const int hx = subx + x * (1 << (1 + qpel));
112 const int hy = suby + y * (1 << (1 + qpel));
113 uint8_t * const * const ref= c->ref[ref_index];
114 uint8_t * const * const src= c->src[src_index];
115 int d;
116 //FIXME check chroma 4mv, (no crashes ...)
117 av_assert2(x >= c->xmin && hx <= c->xmax<<(qpel+1) && y >= c->ymin && hy <= c->ymax<<(qpel+1));
118 if(x >= c->xmin && hx <= c->xmax<<(qpel+1) && y >= c->ymin && hy <= c->ymax<<(qpel+1)){
119 const int time_pp= s->pp_time;
120 const int time_pb= s->pb_time;
121 const int mask= 2*qpel+1;
122 if(s->mv_type==MV_TYPE_8X8){
123 int i;
124 for(i=0; i<4; i++){
125 int fx = c->direct_basis_mv[i][0] + hx;
126 int fy = c->direct_basis_mv[i][1] + hy;
127 int bx = hx ? fx - c->co_located_mv[i][0] : c->co_located_mv[i][0]*(time_pb - time_pp)/time_pp + ((i &1)<<(qpel+4));
128 int by = hy ? fy - c->co_located_mv[i][1] : c->co_located_mv[i][1]*(time_pb - time_pp)/time_pp + ((i>>1)<<(qpel+4));
129 int fxy= (fx&mask) + ((fy&mask)<<(qpel+1));
130 int bxy= (bx&mask) + ((by&mask)<<(qpel+1));
131
132 uint8_t *dst= c->temp + 8*(i&1) + 8*stride*(i>>1);
133 if(qpel){
134 c->qpel_put[1][fxy](dst, ref[0] + (fx>>2) + (fy>>2)*stride, stride);
135 c->qpel_avg[1][bxy](dst, ref[8] + (bx>>2) + (by>>2)*stride, stride);
136 }else{
137 c->hpel_put[1][fxy](dst, ref[0] + (fx>>1) + (fy>>1)*stride, stride, 8);
138 c->hpel_avg[1][bxy](dst, ref[8] + (bx>>1) + (by>>1)*stride, stride, 8);
139 }
140 }
141 }else{
142 int fx = c->direct_basis_mv[0][0] + hx;
143 int fy = c->direct_basis_mv[0][1] + hy;
144 int bx = hx ? fx - c->co_located_mv[0][0] : (c->co_located_mv[0][0]*(time_pb - time_pp)/time_pp);
145 int by = hy ? fy - c->co_located_mv[0][1] : (c->co_located_mv[0][1]*(time_pb - time_pp)/time_pp);
146 int fxy= (fx&mask) + ((fy&mask)<<(qpel+1));
147 int bxy= (bx&mask) + ((by&mask)<<(qpel+1));
148
149 if(qpel){
150 c->qpel_put[1][fxy](c->temp , ref[0] + (fx>>2) + (fy>>2)*stride , stride);
151 c->qpel_put[1][fxy](c->temp + 8 , ref[0] + (fx>>2) + (fy>>2)*stride + 8 , stride);
152 c->qpel_put[1][fxy](c->temp + 8*stride, ref[0] + (fx>>2) + (fy>>2)*stride + 8*stride, stride);
153 c->qpel_put[1][fxy](c->temp + 8 + 8*stride, ref[0] + (fx>>2) + (fy>>2)*stride + 8 + 8*stride, stride);
154 c->qpel_avg[1][bxy](c->temp , ref[8] + (bx>>2) + (by>>2)*stride , stride);
155 c->qpel_avg[1][bxy](c->temp + 8 , ref[8] + (bx>>2) + (by>>2)*stride + 8 , stride);
156 c->qpel_avg[1][bxy](c->temp + 8*stride, ref[8] + (bx>>2) + (by>>2)*stride + 8*stride, stride);
157 c->qpel_avg[1][bxy](c->temp + 8 + 8*stride, ref[8] + (bx>>2) + (by>>2)*stride + 8 + 8*stride, stride);
158 }else{
159 av_assert2((fx>>1) + 16*s->mb_x >= -16);
160 av_assert2((fy>>1) + 16*s->mb_y >= -16);
161 av_assert2((fx>>1) + 16*s->mb_x <= s->width);
162 av_assert2((fy>>1) + 16*s->mb_y <= s->height);
163 av_assert2((bx>>1) + 16*s->mb_x >= -16);
164 av_assert2((by>>1) + 16*s->mb_y >= -16);
165 av_assert2((bx>>1) + 16*s->mb_x <= s->width);
166 av_assert2((by>>1) + 16*s->mb_y <= s->height);
167
168 c->hpel_put[0][fxy](c->temp, ref[0] + (fx>>1) + (fy>>1)*stride, stride, 16);
169 c->hpel_avg[0][bxy](c->temp, ref[8] + (bx>>1) + (by>>1)*stride, stride, 16);
170 }
171 }
172 d = cmp_func(s, c->temp, src[0], stride, 16);
173 }else
174 d= 256*256*256*32;
175 return d;
176 }
177
cmp_inline(MpegEncContext * s,const int x,const int y,const int subx,const int suby,const int size,const int h,int ref_index,int src_index,me_cmp_func cmp_func,me_cmp_func chroma_cmp_func,int qpel,int chroma)178 static av_always_inline int cmp_inline(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
179 const int size, const int h, int ref_index, int src_index,
180 me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, int qpel, int chroma){
181 MotionEstContext * const c= &s->me;
182 const int stride= c->stride;
183 const int uvstride= c->uvstride;
184 const int dxy= subx + (suby<<(1+qpel)); //FIXME log2_subpel?
185 const int hx= subx + x*(1<<(1+qpel));
186 const int hy= suby + y*(1<<(1+qpel));
187 uint8_t * const * const ref= c->ref[ref_index];
188 uint8_t * const * const src= c->src[src_index];
189 int d;
190 //FIXME check chroma 4mv, (no crashes ...)
191 int uvdxy; /* no, it might not be used uninitialized */
192 if(dxy){
193 if(qpel){
194 if (h << size == 16) {
195 c->qpel_put[size][dxy](c->temp, ref[0] + x + y*stride, stride); //FIXME prototype (add h)
196 } else if (size == 0 && h == 8) {
197 c->qpel_put[1][dxy](c->temp , ref[0] + x + y*stride , stride);
198 c->qpel_put[1][dxy](c->temp + 8, ref[0] + x + y*stride + 8, stride);
199 } else
200 av_assert2(0);
201 if(chroma){
202 int cx= hx/2;
203 int cy= hy/2;
204 cx= (cx>>1)|(cx&1);
205 cy= (cy>>1)|(cy&1);
206 uvdxy= (cx&1) + 2*(cy&1);
207 // FIXME x/y wrong, but MPEG-4 qpel is sick anyway, we should drop as much of it as possible in favor for H.264
208 }
209 }else{
210 c->hpel_put[size][dxy](c->temp, ref[0] + x + y*stride, stride, h);
211 if(chroma)
212 uvdxy= dxy | (x&1) | (2*(y&1));
213 }
214 d = cmp_func(s, c->temp, src[0], stride, h);
215 }else{
216 d = cmp_func(s, src[0], ref[0] + x + y*stride, stride, h);
217 if(chroma)
218 uvdxy= (x&1) + 2*(y&1);
219 }
220 if(chroma){
221 uint8_t * const uvtemp= c->temp + 16*stride;
222 c->hpel_put[size+1][uvdxy](uvtemp , ref[1] + (x>>1) + (y>>1)*uvstride, uvstride, h>>1);
223 c->hpel_put[size+1][uvdxy](uvtemp+8, ref[2] + (x>>1) + (y>>1)*uvstride, uvstride, h>>1);
224 d += chroma_cmp_func(s, uvtemp , src[1], uvstride, h>>1);
225 d += chroma_cmp_func(s, uvtemp+8, src[2], uvstride, h>>1);
226 }
227 return d;
228 }
229
cmp_simple(MpegEncContext * s,const int x,const int y,int ref_index,int src_index,me_cmp_func cmp_func,me_cmp_func chroma_cmp_func)230 static int cmp_simple(MpegEncContext *s, const int x, const int y,
231 int ref_index, int src_index,
232 me_cmp_func cmp_func, me_cmp_func chroma_cmp_func){
233 return cmp_inline(s,x,y,0,0,0,16,ref_index,src_index, cmp_func, chroma_cmp_func, 0, 0);
234 }
235
cmp_fpel_internal(MpegEncContext * s,const int x,const int y,const int size,const int h,int ref_index,int src_index,me_cmp_func cmp_func,me_cmp_func chroma_cmp_func,const int flags)236 static int cmp_fpel_internal(MpegEncContext *s, const int x, const int y,
237 const int size, const int h, int ref_index, int src_index,
238 me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){
239 if(flags&FLAG_DIRECT){
240 return cmp_direct_inline(s,x,y,0,0,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, flags&FLAG_QPEL);
241 }else{
242 return cmp_inline(s,x,y,0,0,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, 0, flags&FLAG_CHROMA);
243 }
244 }
245
cmp_internal(MpegEncContext * s,const int x,const int y,const int subx,const int suby,const int size,const int h,int ref_index,int src_index,me_cmp_func cmp_func,me_cmp_func chroma_cmp_func,const int flags)246 static int cmp_internal(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
247 const int size, const int h, int ref_index, int src_index,
248 me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){
249 if(flags&FLAG_DIRECT){
250 return cmp_direct_inline(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, flags&FLAG_QPEL);
251 }else{
252 return cmp_inline(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, flags&FLAG_QPEL, flags&FLAG_CHROMA);
253 }
254 }
255
256 /** @brief compares a block (either a full macroblock or a partition thereof)
257 against a proposed motion-compensated prediction of that block
258 */
cmp(MpegEncContext * s,const int x,const int y,const int subx,const int suby,const int size,const int h,int ref_index,int src_index,me_cmp_func cmp_func,me_cmp_func chroma_cmp_func,const int flags)259 static av_always_inline int cmp(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
260 const int size, const int h, int ref_index, int src_index,
261 me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){
262 if(av_builtin_constant_p(flags) && av_builtin_constant_p(h) && av_builtin_constant_p(size)
263 && av_builtin_constant_p(subx) && av_builtin_constant_p(suby)
264 && flags==0 && h==16 && size==0 && subx==0 && suby==0){
265 return cmp_simple(s,x,y,ref_index,src_index, cmp_func, chroma_cmp_func);
266 }else if(av_builtin_constant_p(subx) && av_builtin_constant_p(suby)
267 && subx==0 && suby==0){
268 return cmp_fpel_internal(s,x,y,size,h,ref_index,src_index, cmp_func, chroma_cmp_func,flags);
269 }else{
270 return cmp_internal(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, flags);
271 }
272 }
273
cmp_hpel(MpegEncContext * s,const int x,const int y,const int subx,const int suby,const int size,const int h,int ref_index,int src_index,me_cmp_func cmp_func,me_cmp_func chroma_cmp_func,const int flags)274 static int cmp_hpel(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
275 const int size, const int h, int ref_index, int src_index,
276 me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){
277 if(flags&FLAG_DIRECT){
278 return cmp_direct_inline(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, 0);
279 }else{
280 return cmp_inline(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, 0, flags&FLAG_CHROMA);
281 }
282 }
283
cmp_qpel(MpegEncContext * s,const int x,const int y,const int subx,const int suby,const int size,const int h,int ref_index,int src_index,me_cmp_func cmp_func,me_cmp_func chroma_cmp_func,const int flags)284 static int cmp_qpel(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
285 const int size, const int h, int ref_index, int src_index,
286 me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){
287 if(flags&FLAG_DIRECT){
288 return cmp_direct_inline(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, 1);
289 }else{
290 return cmp_inline(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, 1, flags&FLAG_CHROMA);
291 }
292 }
293
294 #include "motion_est_template.c"
295
zero_cmp(MpegEncContext * s,uint8_t * a,uint8_t * b,ptrdiff_t stride,int h)296 static int zero_cmp(MpegEncContext *s, uint8_t *a, uint8_t *b,
297 ptrdiff_t stride, int h)
298 {
299 return 0;
300 }
301
zero_hpel(uint8_t * a,const uint8_t * b,ptrdiff_t stride,int h)302 static void zero_hpel(uint8_t *a, const uint8_t *b, ptrdiff_t stride, int h){
303 }
304
ff_init_me(MpegEncContext * s)305 int ff_init_me(MpegEncContext *s){
306 MotionEstContext * const c= &s->me;
307 int cache_size= FFMIN(ME_MAP_SIZE>>ME_MAP_SHIFT, 1<<ME_MAP_SHIFT);
308 int dia_size= FFMAX(FFABS(s->avctx->dia_size)&255, FFABS(s->avctx->pre_dia_size)&255);
309 int ret;
310
311 if(FFMIN(s->avctx->dia_size, s->avctx->pre_dia_size) < -FFMIN(ME_MAP_SIZE, MAX_SAB_SIZE)){
312 av_log(s->avctx, AV_LOG_ERROR, "ME_MAP size is too small for SAB diamond\n");
313 return -1;
314 }
315
316 c->avctx= s->avctx;
317
318 if(s->codec_id == AV_CODEC_ID_H261)
319 c->avctx->me_sub_cmp = c->avctx->me_cmp;
320
321 if(cache_size < 2*dia_size && !c->stride){
322 av_log(s->avctx, AV_LOG_INFO, "ME_MAP size may be a little small for the selected diamond size\n");
323 }
324
325 ret = ff_set_cmp(&s->mecc, s->mecc.me_pre_cmp, c->avctx->me_pre_cmp);
326 ret |= ff_set_cmp(&s->mecc, s->mecc.me_cmp, c->avctx->me_cmp);
327 ret |= ff_set_cmp(&s->mecc, s->mecc.me_sub_cmp, c->avctx->me_sub_cmp);
328 ret |= ff_set_cmp(&s->mecc, s->mecc.mb_cmp, c->avctx->mb_cmp);
329 if (ret < 0)
330 return ret;
331
332 c->flags = get_flags(c, 0, c->avctx->me_cmp &FF_CMP_CHROMA);
333 c->sub_flags= get_flags(c, 0, c->avctx->me_sub_cmp&FF_CMP_CHROMA);
334 c->mb_flags = get_flags(c, 0, c->avctx->mb_cmp &FF_CMP_CHROMA);
335
336 /*FIXME s->no_rounding b_type*/
337 if (s->avctx->flags & AV_CODEC_FLAG_QPEL) {
338 c->sub_motion_search= qpel_motion_search;
339 c->qpel_avg = s->qdsp.avg_qpel_pixels_tab;
340 if (s->no_rounding)
341 c->qpel_put = s->qdsp.put_no_rnd_qpel_pixels_tab;
342 else
343 c->qpel_put = s->qdsp.put_qpel_pixels_tab;
344 }else{
345 if(c->avctx->me_sub_cmp&FF_CMP_CHROMA)
346 c->sub_motion_search= hpel_motion_search;
347 else if( c->avctx->me_sub_cmp == FF_CMP_SAD
348 && c->avctx-> me_cmp == FF_CMP_SAD
349 && c->avctx-> mb_cmp == FF_CMP_SAD)
350 c->sub_motion_search= sad_hpel_motion_search; // 2050 vs. 2450 cycles
351 else
352 c->sub_motion_search= hpel_motion_search;
353 }
354 c->hpel_avg = s->hdsp.avg_pixels_tab;
355 if (s->no_rounding)
356 c->hpel_put = s->hdsp.put_no_rnd_pixels_tab;
357 else
358 c->hpel_put = s->hdsp.put_pixels_tab;
359
360 if(s->linesize){
361 c->stride = s->linesize;
362 c->uvstride= s->uvlinesize;
363 }else{
364 c->stride = 16*s->mb_width + 32;
365 c->uvstride= 8*s->mb_width + 16;
366 }
367
368 /* 8x8 fullpel search would need a 4x4 chroma compare, which we do
369 * not have yet, and even if we had, the motion estimation code
370 * does not expect it. */
371 if (s->codec_id != AV_CODEC_ID_SNOW) {
372 if ((c->avctx->me_cmp & FF_CMP_CHROMA) /* && !s->mecc.me_cmp[2] */)
373 s->mecc.me_cmp[2] = zero_cmp;
374 if ((c->avctx->me_sub_cmp & FF_CMP_CHROMA) && !s->mecc.me_sub_cmp[2])
375 s->mecc.me_sub_cmp[2] = zero_cmp;
376 c->hpel_put[2][0]= c->hpel_put[2][1]=
377 c->hpel_put[2][2]= c->hpel_put[2][3]= zero_hpel;
378 }
379
380 if(s->codec_id == AV_CODEC_ID_H261){
381 c->sub_motion_search= no_sub_motion_search;
382 }
383
384 return 0;
385 }
386
387 #define CHECK_SAD_HALF_MV(suffix, x, y) \
388 {\
389 d = s->mecc.pix_abs[size][(x ? 1 : 0) + (y ? 2 : 0)](NULL, pix, ptr + ((x) >> 1), stride, h); \
390 d += (mv_penalty[pen_x + x] + mv_penalty[pen_y + y])*penalty_factor;\
391 COPY3_IF_LT(dminh, d, dx, x, dy, y)\
392 }
393
sad_hpel_motion_search(MpegEncContext * s,int * mx_ptr,int * my_ptr,int dmin,int src_index,int ref_index,int size,int h)394 static int sad_hpel_motion_search(MpegEncContext * s,
395 int *mx_ptr, int *my_ptr, int dmin,
396 int src_index, int ref_index,
397 int size, int h)
398 {
399 MotionEstContext * const c= &s->me;
400 const int penalty_factor= c->sub_penalty_factor;
401 int mx, my, dminh;
402 uint8_t *pix, *ptr;
403 int stride= c->stride;
404 LOAD_COMMON
405
406 av_assert2(c->sub_flags == 0);
407
408 if(c->skip){
409 *mx_ptr = 0;
410 *my_ptr = 0;
411 return dmin;
412 }
413
414 pix = c->src[src_index][0];
415
416 mx = *mx_ptr;
417 my = *my_ptr;
418 ptr = c->ref[ref_index][0] + (my * stride) + mx;
419
420 dminh = dmin;
421
422 if (mx > xmin && mx < xmax &&
423 my > ymin && my < ymax) {
424 int dx=0, dy=0;
425 int d, pen_x, pen_y;
426 const int index= my*(1<<ME_MAP_SHIFT) + mx;
427 const int t= score_map[(index-(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
428 const int l= score_map[(index- 1 )&(ME_MAP_SIZE-1)];
429 const int r= score_map[(index+ 1 )&(ME_MAP_SIZE-1)];
430 const int b= score_map[(index+(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
431 mx += mx;
432 my += my;
433
434
435 pen_x= pred_x + mx;
436 pen_y= pred_y + my;
437
438 ptr-= stride;
439 if(t<=b){
440 CHECK_SAD_HALF_MV(y2 , 0, -1)
441 if(l<=r){
442 CHECK_SAD_HALF_MV(xy2, -1, -1)
443 if(t+r<=b+l){
444 CHECK_SAD_HALF_MV(xy2, +1, -1)
445 ptr+= stride;
446 }else{
447 ptr+= stride;
448 CHECK_SAD_HALF_MV(xy2, -1, +1)
449 }
450 CHECK_SAD_HALF_MV(x2 , -1, 0)
451 }else{
452 CHECK_SAD_HALF_MV(xy2, +1, -1)
453 if(t+l<=b+r){
454 CHECK_SAD_HALF_MV(xy2, -1, -1)
455 ptr+= stride;
456 }else{
457 ptr+= stride;
458 CHECK_SAD_HALF_MV(xy2, +1, +1)
459 }
460 CHECK_SAD_HALF_MV(x2 , +1, 0)
461 }
462 }else{
463 if(l<=r){
464 if(t+l<=b+r){
465 CHECK_SAD_HALF_MV(xy2, -1, -1)
466 ptr+= stride;
467 }else{
468 ptr+= stride;
469 CHECK_SAD_HALF_MV(xy2, +1, +1)
470 }
471 CHECK_SAD_HALF_MV(x2 , -1, 0)
472 CHECK_SAD_HALF_MV(xy2, -1, +1)
473 }else{
474 if(t+r<=b+l){
475 CHECK_SAD_HALF_MV(xy2, +1, -1)
476 ptr+= stride;
477 }else{
478 ptr+= stride;
479 CHECK_SAD_HALF_MV(xy2, -1, +1)
480 }
481 CHECK_SAD_HALF_MV(x2 , +1, 0)
482 CHECK_SAD_HALF_MV(xy2, +1, +1)
483 }
484 CHECK_SAD_HALF_MV(y2 , 0, +1)
485 }
486 mx+=dx;
487 my+=dy;
488
489 }else{
490 mx += mx;
491 my += my;
492 }
493
494 *mx_ptr = mx;
495 *my_ptr = my;
496 return dminh;
497 }
498
set_p_mv_tables(MpegEncContext * s,int mx,int my,int mv4)499 static inline void set_p_mv_tables(MpegEncContext * s, int mx, int my, int mv4)
500 {
501 const int xy= s->mb_x + s->mb_y*s->mb_stride;
502
503 s->p_mv_table[xy][0] = mx;
504 s->p_mv_table[xy][1] = my;
505
506 /* has already been set to the 4 MV if 4MV is done */
507 if(mv4){
508 int mot_xy= s->block_index[0];
509
510 s->current_picture.motion_val[0][mot_xy ][0] = mx;
511 s->current_picture.motion_val[0][mot_xy ][1] = my;
512 s->current_picture.motion_val[0][mot_xy + 1][0] = mx;
513 s->current_picture.motion_val[0][mot_xy + 1][1] = my;
514
515 mot_xy += s->b8_stride;
516 s->current_picture.motion_val[0][mot_xy ][0] = mx;
517 s->current_picture.motion_val[0][mot_xy ][1] = my;
518 s->current_picture.motion_val[0][mot_xy + 1][0] = mx;
519 s->current_picture.motion_val[0][mot_xy + 1][1] = my;
520 }
521 }
522
523 /**
524 * get fullpel ME search limits.
525 */
get_limits(MpegEncContext * s,int x,int y)526 static inline void get_limits(MpegEncContext *s, int x, int y)
527 {
528 MotionEstContext * const c= &s->me;
529 int range= c->avctx->me_range >> (1 + !!(c->flags&FLAG_QPEL));
530 int max_range = MAX_MV >> (1 + !!(c->flags&FLAG_QPEL));
531 /*
532 if(c->avctx->me_range) c->range= c->avctx->me_range >> 1;
533 else c->range= 16;
534 */
535 if (s->unrestricted_mv) {
536 c->xmin = - x - 16;
537 c->ymin = - y - 16;
538 c->xmax = - x + s->width;
539 c->ymax = - y + s->height;
540 } else if (s->out_format == FMT_H261){
541 // Search range of H.261 is different from other codec standards
542 c->xmin = (x > 15) ? - 15 : 0;
543 c->ymin = (y > 15) ? - 15 : 0;
544 c->xmax = (x < s->mb_width * 16 - 16) ? 15 : 0;
545 c->ymax = (y < s->mb_height * 16 - 16) ? 15 : 0;
546 } else {
547 c->xmin = - x;
548 c->ymin = - y;
549 c->xmax = - x + s->mb_width *16 - 16;
550 c->ymax = - y + s->mb_height*16 - 16;
551 }
552 if(!range || range > max_range)
553 range = max_range;
554 if(range){
555 c->xmin = FFMAX(c->xmin,-range);
556 c->xmax = FFMIN(c->xmax, range);
557 c->ymin = FFMAX(c->ymin,-range);
558 c->ymax = FFMIN(c->ymax, range);
559 }
560 }
561
init_mv4_ref(MotionEstContext * c)562 static inline void init_mv4_ref(MotionEstContext *c){
563 const int stride= c->stride;
564
565 c->ref[1][0] = c->ref[0][0] + 8;
566 c->ref[2][0] = c->ref[0][0] + 8*stride;
567 c->ref[3][0] = c->ref[2][0] + 8;
568 c->src[1][0] = c->src[0][0] + 8;
569 c->src[2][0] = c->src[0][0] + 8*stride;
570 c->src[3][0] = c->src[2][0] + 8;
571 }
572
h263_mv4_search(MpegEncContext * s,int mx,int my,int shift)573 static inline int h263_mv4_search(MpegEncContext *s, int mx, int my, int shift)
574 {
575 MotionEstContext * const c= &s->me;
576 const int size= 1;
577 const int h=8;
578 int block;
579 int P[10][2];
580 int dmin_sum=0, mx4_sum=0, my4_sum=0, i;
581 int same=1;
582 const int stride= c->stride;
583 const uint8_t *mv_penalty = c->current_mv_penalty;
584 int safety_clipping= s->unrestricted_mv && (s->width&15) && (s->height&15);
585
586 init_mv4_ref(c);
587
588 for(block=0; block<4; block++){
589 int mx4, my4;
590 int pred_x4, pred_y4;
591 int dmin4;
592 static const int off[4]= {2, 1, 1, -1};
593 const int mot_stride = s->b8_stride;
594 const int mot_xy = s->block_index[block];
595
596 if(safety_clipping){
597 c->xmax = - 16*s->mb_x + s->width - 8*(block &1);
598 c->ymax = - 16*s->mb_y + s->height - 8*(block>>1);
599 }
600
601 P_LEFT[0] = s->current_picture.motion_val[0][mot_xy - 1][0];
602 P_LEFT[1] = s->current_picture.motion_val[0][mot_xy - 1][1];
603
604 if (P_LEFT[0] > c->xmax * (1 << shift)) P_LEFT[0] = c->xmax * (1 << shift);
605
606 /* special case for first line */
607 if (s->first_slice_line && block<2) {
608 c->pred_x= pred_x4= P_LEFT[0];
609 c->pred_y= pred_y4= P_LEFT[1];
610 } else {
611 P_TOP[0] = s->current_picture.motion_val[0][mot_xy - mot_stride ][0];
612 P_TOP[1] = s->current_picture.motion_val[0][mot_xy - mot_stride ][1];
613 P_TOPRIGHT[0] = s->current_picture.motion_val[0][mot_xy - mot_stride + off[block]][0];
614 P_TOPRIGHT[1] = s->current_picture.motion_val[0][mot_xy - mot_stride + off[block]][1];
615 if (P_TOP[1] > c->ymax * (1 << shift)) P_TOP[1] = c->ymax * (1 << shift);
616 if (P_TOPRIGHT[0] < c->xmin * (1 << shift)) P_TOPRIGHT[0] = c->xmin * (1 << shift);
617 if (P_TOPRIGHT[0] > c->xmax * (1 << shift)) P_TOPRIGHT[0] = c->xmax * (1 << shift);
618 if (P_TOPRIGHT[1] > c->ymax * (1 << shift)) P_TOPRIGHT[1] = c->ymax * (1 << shift);
619
620 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
621 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
622
623 c->pred_x= pred_x4 = P_MEDIAN[0];
624 c->pred_y= pred_y4 = P_MEDIAN[1];
625 }
626 P_MV1[0]= mx;
627 P_MV1[1]= my;
628 if(safety_clipping)
629 for(i=1; i<10; i++){
630 if (s->first_slice_line && block<2 && i>1 && i<9)
631 continue;
632 if (i>4 && i<9)
633 continue;
634 if (P[i][0] > c->xmax * (1 << shift)) P[i][0] = c->xmax * (1 << shift);
635 if (P[i][1] > c->ymax * (1 << shift)) P[i][1] = c->ymax * (1 <<shift );
636 }
637
638 dmin4 = epzs_motion_search2(s, &mx4, &my4, P, block, block, s->p_mv_table, (1<<16)>>shift, 1);
639
640 dmin4= c->sub_motion_search(s, &mx4, &my4, dmin4, block, block, size, h);
641
642 if (s->mecc.me_sub_cmp[0] != s->mecc.mb_cmp[0]) {
643 int dxy;
644 const int offset= ((block&1) + (block>>1)*stride)*8;
645 uint8_t *dest_y = c->scratchpad + offset;
646 if(s->quarter_sample){
647 uint8_t *ref= c->ref[block][0] + (mx4>>2) + (my4>>2)*stride;
648 dxy = ((my4 & 3) << 2) | (mx4 & 3);
649
650 if(s->no_rounding)
651 s->qdsp.put_no_rnd_qpel_pixels_tab[1][dxy](dest_y, ref, stride);
652 else
653 s->qdsp.put_qpel_pixels_tab[1][dxy](dest_y, ref, stride);
654 }else{
655 uint8_t *ref= c->ref[block][0] + (mx4>>1) + (my4>>1)*stride;
656 dxy = ((my4 & 1) << 1) | (mx4 & 1);
657
658 if(s->no_rounding)
659 s->hdsp.put_no_rnd_pixels_tab[1][dxy](dest_y , ref , stride, h);
660 else
661 s->hdsp.put_pixels_tab [1][dxy](dest_y , ref , stride, h);
662 }
663 dmin_sum+= (mv_penalty[mx4-pred_x4] + mv_penalty[my4-pred_y4])*c->mb_penalty_factor;
664 }else
665 dmin_sum+= dmin4;
666
667 if(s->quarter_sample){
668 mx4_sum+= mx4/2;
669 my4_sum+= my4/2;
670 }else{
671 mx4_sum+= mx4;
672 my4_sum+= my4;
673 }
674
675 s->current_picture.motion_val[0][s->block_index[block]][0] = mx4;
676 s->current_picture.motion_val[0][s->block_index[block]][1] = my4;
677
678 if(mx4 != mx || my4 != my) same=0;
679 }
680
681 if(same)
682 return INT_MAX;
683
684 if (s->mecc.me_sub_cmp[0] != s->mecc.mb_cmp[0]) {
685 dmin_sum += s->mecc.mb_cmp[0](s,
686 s->new_picture->data[0] +
687 s->mb_x * 16 + s->mb_y * 16 * stride,
688 c->scratchpad, stride, 16);
689 }
690
691 if(c->avctx->mb_cmp&FF_CMP_CHROMA){
692 int dxy;
693 int mx, my;
694 int offset;
695
696 mx= ff_h263_round_chroma(mx4_sum);
697 my= ff_h263_round_chroma(my4_sum);
698 dxy = ((my & 1) << 1) | (mx & 1);
699
700 offset= (s->mb_x*8 + (mx>>1)) + (s->mb_y*8 + (my>>1))*s->uvlinesize;
701
702 if(s->no_rounding){
703 s->hdsp.put_no_rnd_pixels_tab[1][dxy](c->scratchpad , s->last_picture.f->data[1] + offset, s->uvlinesize, 8);
704 s->hdsp.put_no_rnd_pixels_tab[1][dxy](c->scratchpad + 8, s->last_picture.f->data[2] + offset, s->uvlinesize, 8);
705 }else{
706 s->hdsp.put_pixels_tab [1][dxy](c->scratchpad , s->last_picture.f->data[1] + offset, s->uvlinesize, 8);
707 s->hdsp.put_pixels_tab [1][dxy](c->scratchpad + 8, s->last_picture.f->data[2] + offset, s->uvlinesize, 8);
708 }
709
710 dmin_sum += s->mecc.mb_cmp[1](s, s->new_picture->data[1] + s->mb_x * 8 + s->mb_y * 8 * s->uvlinesize, c->scratchpad, s->uvlinesize, 8);
711 dmin_sum += s->mecc.mb_cmp[1](s, s->new_picture->data[2] + s->mb_x * 8 + s->mb_y * 8 * s->uvlinesize, c->scratchpad + 8, s->uvlinesize, 8);
712 }
713
714 c->pred_x= mx;
715 c->pred_y= my;
716
717 switch(c->avctx->mb_cmp&0xFF){
718 /*case FF_CMP_SSE:
719 return dmin_sum+ 32*s->qscale*s->qscale;*/
720 case FF_CMP_RD:
721 return dmin_sum;
722 default:
723 return dmin_sum+ 11*c->mb_penalty_factor;
724 }
725 }
726
init_interlaced_ref(MpegEncContext * s,int ref_index)727 static inline void init_interlaced_ref(MpegEncContext *s, int ref_index){
728 MotionEstContext * const c= &s->me;
729
730 c->ref[1+ref_index][0] = c->ref[0+ref_index][0] + s->linesize;
731 c->src[1][0] = c->src[0][0] + s->linesize;
732 if(c->flags & FLAG_CHROMA){
733 c->ref[1+ref_index][1] = c->ref[0+ref_index][1] + s->uvlinesize;
734 c->ref[1+ref_index][2] = c->ref[0+ref_index][2] + s->uvlinesize;
735 c->src[1][1] = c->src[0][1] + s->uvlinesize;
736 c->src[1][2] = c->src[0][2] + s->uvlinesize;
737 }
738 }
739
interlaced_search(MpegEncContext * s,int ref_index,int16_t (* mv_tables[2][2])[2],uint8_t * field_select_tables[2],int mx,int my,int user_field_select)740 static int interlaced_search(MpegEncContext *s, int ref_index,
741 int16_t (*mv_tables[2][2])[2], uint8_t *field_select_tables[2], int mx, int my, int user_field_select)
742 {
743 MotionEstContext * const c= &s->me;
744 const int size=0;
745 const int h=8;
746 int block;
747 int P[10][2];
748 const uint8_t * const mv_penalty = c->current_mv_penalty;
749 int same=1;
750 const int stride= 2*s->linesize;
751 int dmin_sum= 0;
752 const int mot_stride= s->mb_stride;
753 const int xy= s->mb_x + s->mb_y*mot_stride;
754
755 c->ymin>>=1;
756 c->ymax>>=1;
757 c->stride<<=1;
758 c->uvstride<<=1;
759 init_interlaced_ref(s, ref_index);
760
761 for(block=0; block<2; block++){
762 int field_select;
763 int best_dmin= INT_MAX;
764 int best_field= -1;
765
766 for(field_select=0; field_select<2; field_select++){
767 int dmin, mx_i, my_i;
768 int16_t (*mv_table)[2]= mv_tables[block][field_select];
769
770 if(user_field_select){
771 av_assert1(field_select==0 || field_select==1);
772 av_assert1(field_select_tables[block][xy]==0 || field_select_tables[block][xy]==1);
773 if(field_select_tables[block][xy] != field_select)
774 continue;
775 }
776
777 P_LEFT[0] = mv_table[xy - 1][0];
778 P_LEFT[1] = mv_table[xy - 1][1];
779 if(P_LEFT[0] > (c->xmax<<1)) P_LEFT[0] = (c->xmax<<1);
780
781 c->pred_x= P_LEFT[0];
782 c->pred_y= P_LEFT[1];
783
784 if(!s->first_slice_line){
785 P_TOP[0] = mv_table[xy - mot_stride][0];
786 P_TOP[1] = mv_table[xy - mot_stride][1];
787 P_TOPRIGHT[0] = mv_table[xy - mot_stride + 1][0];
788 P_TOPRIGHT[1] = mv_table[xy - mot_stride + 1][1];
789 if(P_TOP[1] > (c->ymax<<1)) P_TOP[1] = (c->ymax<<1);
790 if (P_TOPRIGHT[0] < c->xmin * (1 << 1)) P_TOPRIGHT[0] = c->xmin * (1 << 1);
791 if(P_TOPRIGHT[0] > (c->xmax<<1)) P_TOPRIGHT[0]= (c->xmax<<1);
792 if(P_TOPRIGHT[1] > (c->ymax<<1)) P_TOPRIGHT[1]= (c->ymax<<1);
793
794 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
795 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
796 }
797 P_MV1[0]= mx; //FIXME not correct if block != field_select
798 P_MV1[1]= my / 2;
799
800 dmin = epzs_motion_search2(s, &mx_i, &my_i, P, block, field_select+ref_index, mv_table, (1<<16)>>1, 0);
801
802 dmin= c->sub_motion_search(s, &mx_i, &my_i, dmin, block, field_select+ref_index, size, h);
803
804 mv_table[xy][0]= mx_i;
805 mv_table[xy][1]= my_i;
806
807 if (s->mecc.me_sub_cmp[0] != s->mecc.mb_cmp[0]) {
808 int dxy;
809
810 //FIXME chroma ME
811 uint8_t *ref= c->ref[field_select+ref_index][0] + (mx_i>>1) + (my_i>>1)*stride;
812 dxy = ((my_i & 1) << 1) | (mx_i & 1);
813
814 if(s->no_rounding){
815 s->hdsp.put_no_rnd_pixels_tab[size][dxy](c->scratchpad, ref , stride, h);
816 }else{
817 s->hdsp.put_pixels_tab [size][dxy](c->scratchpad, ref , stride, h);
818 }
819 dmin = s->mecc.mb_cmp[size](s, c->src[block][0], c->scratchpad, stride, h);
820 dmin+= (mv_penalty[mx_i-c->pred_x] + mv_penalty[my_i-c->pred_y] + 1)*c->mb_penalty_factor;
821 }else
822 dmin+= c->mb_penalty_factor; //field_select bits
823
824 dmin += field_select != block; //slightly prefer same field
825
826 if(dmin < best_dmin){
827 best_dmin= dmin;
828 best_field= field_select;
829 }
830 }
831 {
832 int16_t (*mv_table)[2]= mv_tables[block][best_field];
833
834 if(mv_table[xy][0] != mx) same=0; //FIXME check if these checks work and are any good at all
835 if(mv_table[xy][1]&1) same=0;
836 if(mv_table[xy][1]*2 != my) same=0;
837 if(best_field != block) same=0;
838 }
839
840 field_select_tables[block][xy]= best_field;
841 dmin_sum += best_dmin;
842 }
843
844 c->ymin *= 2;
845 c->ymax<<=1;
846 c->stride>>=1;
847 c->uvstride>>=1;
848
849 if(same)
850 return INT_MAX;
851
852 switch(c->avctx->mb_cmp&0xFF){
853 /*case FF_CMP_SSE:
854 return dmin_sum+ 32*s->qscale*s->qscale;*/
855 case FF_CMP_RD:
856 return dmin_sum;
857 default:
858 return dmin_sum+ 11*c->mb_penalty_factor;
859 }
860 }
861
get_penalty_factor(int lambda,int lambda2,int type)862 static inline int get_penalty_factor(int lambda, int lambda2, int type){
863 switch(type&0xFF){
864 default:
865 case FF_CMP_SAD:
866 return lambda>>FF_LAMBDA_SHIFT;
867 case FF_CMP_DCT:
868 return (3*lambda)>>(FF_LAMBDA_SHIFT+1);
869 case FF_CMP_W53:
870 return (4*lambda)>>(FF_LAMBDA_SHIFT);
871 case FF_CMP_W97:
872 return (2*lambda)>>(FF_LAMBDA_SHIFT);
873 case FF_CMP_SATD:
874 case FF_CMP_DCT264:
875 return (2*lambda)>>FF_LAMBDA_SHIFT;
876 case FF_CMP_RD:
877 case FF_CMP_PSNR:
878 case FF_CMP_SSE:
879 case FF_CMP_NSSE:
880 return lambda2>>FF_LAMBDA_SHIFT;
881 case FF_CMP_BIT:
882 case FF_CMP_MEDIAN_SAD:
883 return 1;
884 }
885 }
886
ff_estimate_p_frame_motion(MpegEncContext * s,int mb_x,int mb_y)887 void ff_estimate_p_frame_motion(MpegEncContext * s,
888 int mb_x, int mb_y)
889 {
890 MotionEstContext * const c= &s->me;
891 uint8_t *pix, *ppix;
892 int sum, mx = 0, my = 0, dmin = 0;
893 int varc; ///< the variance of the block (sum of squared (p[y][x]-average))
894 int vard; ///< sum of squared differences with the estimated motion vector
895 int P[10][2];
896 const int shift= 1+s->quarter_sample;
897 int mb_type=0;
898 Picture * const pic= &s->current_picture;
899
900 init_ref(c, s->new_picture->data, s->last_picture.f->data, NULL, 16*mb_x, 16*mb_y, 0);
901
902 av_assert0(s->quarter_sample==0 || s->quarter_sample==1);
903 av_assert0(s->linesize == c->stride);
904 av_assert0(s->uvlinesize == c->uvstride);
905
906 c->penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
907 c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
908 c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
909 c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_DMV;
910
911 get_limits(s, 16*mb_x, 16*mb_y);
912 c->skip=0;
913
914 /* intra / predictive decision */
915 pix = c->src[0][0];
916 sum = s->mpvencdsp.pix_sum(pix, s->linesize);
917 varc = s->mpvencdsp.pix_norm1(pix, s->linesize) -
918 (((unsigned) sum * sum) >> 8) + 500;
919
920 pic->mb_mean[s->mb_stride * mb_y + mb_x] = (sum+128)>>8;
921 pic->mb_var [s->mb_stride * mb_y + mb_x] = (varc+128)>>8;
922 c->mb_var_sum_temp += (varc+128)>>8;
923
924 if (s->motion_est != FF_ME_ZERO) {
925 const int mot_stride = s->b8_stride;
926 const int mot_xy = s->block_index[0];
927
928 P_LEFT[0] = s->current_picture.motion_val[0][mot_xy - 1][0];
929 P_LEFT[1] = s->current_picture.motion_val[0][mot_xy - 1][1];
930
931 if (P_LEFT[0] > (c->xmax << shift))
932 P_LEFT[0] = c->xmax << shift;
933
934 if (!s->first_slice_line) {
935 P_TOP[0] = s->current_picture.motion_val[0][mot_xy - mot_stride ][0];
936 P_TOP[1] = s->current_picture.motion_val[0][mot_xy - mot_stride ][1];
937 P_TOPRIGHT[0] = s->current_picture.motion_val[0][mot_xy - mot_stride + 2][0];
938 P_TOPRIGHT[1] = s->current_picture.motion_val[0][mot_xy - mot_stride + 2][1];
939 if (P_TOP[1] > (c->ymax << shift))
940 P_TOP[1] = c->ymax << shift;
941 if (P_TOPRIGHT[0] < (c->xmin * (1 << shift)))
942 P_TOPRIGHT[0] = c->xmin * (1 << shift);
943 if (P_TOPRIGHT[1] > (c->ymax * (1 << shift)))
944 P_TOPRIGHT[1] = c->ymax * (1 << shift);
945
946 P_MEDIAN[0] = mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
947 P_MEDIAN[1] = mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
948
949 if (s->out_format == FMT_H263) {
950 c->pred_x = P_MEDIAN[0];
951 c->pred_y = P_MEDIAN[1];
952 } else { /* MPEG-1 at least */
953 c->pred_x = P_LEFT[0];
954 c->pred_y = P_LEFT[1];
955 }
956 } else {
957 c->pred_x = P_LEFT[0];
958 c->pred_y = P_LEFT[1];
959 }
960 dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift, 0, 16);
961 }
962
963 /* At this point (mx,my) are full-pell and the relative displacement */
964 ppix = c->ref[0][0] + (my * s->linesize) + mx;
965
966 vard = s->mecc.sse[0](NULL, pix, ppix, s->linesize, 16);
967
968 pic->mc_mb_var[s->mb_stride * mb_y + mb_x] = (vard+128)>>8;
969 c->mc_mb_var_sum_temp += (vard+128)>>8;
970
971 if (c->avctx->mb_decision > FF_MB_DECISION_SIMPLE) {
972 int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
973 int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
974 c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
975
976 if (vard*2 + 200*256 > varc && !s->intra_penalty)
977 mb_type|= CANDIDATE_MB_TYPE_INTRA;
978 if (varc*2 + 200*256 > vard || s->qscale > 24){
979 // if (varc*2 + 200*256 + 50*(s->lambda2>>FF_LAMBDA_SHIFT) > vard){
980 mb_type|= CANDIDATE_MB_TYPE_INTER;
981 c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
982 if (s->mpv_flags & FF_MPV_FLAG_MV0)
983 if(mx || my)
984 mb_type |= CANDIDATE_MB_TYPE_SKIPPED; //FIXME check difference
985 }else{
986 mx *= 1 << shift;
987 my *= 1 << shift;
988 }
989 if ((s->avctx->flags & AV_CODEC_FLAG_4MV)
990 && !c->skip && varc>50<<8 && vard>10<<8){
991 if(h263_mv4_search(s, mx, my, shift) < INT_MAX)
992 mb_type|=CANDIDATE_MB_TYPE_INTER4V;
993
994 set_p_mv_tables(s, mx, my, 0);
995 }else
996 set_p_mv_tables(s, mx, my, 1);
997 if ((s->avctx->flags & AV_CODEC_FLAG_INTERLACED_ME)
998 && !c->skip){ //FIXME varc/d checks
999 if(interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 0) < INT_MAX)
1000 mb_type |= CANDIDATE_MB_TYPE_INTER_I;
1001 }
1002 }else{
1003 int intra_score, i;
1004 mb_type= CANDIDATE_MB_TYPE_INTER;
1005
1006 dmin= c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1007 if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
1008 dmin= get_mb_score(s, mx, my, 0, 0, 0, 16, 1);
1009
1010 if ((s->avctx->flags & AV_CODEC_FLAG_4MV)
1011 && !c->skip && varc>50<<8 && vard>10<<8){
1012 int dmin4= h263_mv4_search(s, mx, my, shift);
1013 if(dmin4 < dmin){
1014 mb_type= CANDIDATE_MB_TYPE_INTER4V;
1015 dmin=dmin4;
1016 }
1017 }
1018 if ((s->avctx->flags & AV_CODEC_FLAG_INTERLACED_ME)
1019 && !c->skip){ //FIXME varc/d checks
1020 int dmin_i= interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 0);
1021 if(dmin_i < dmin){
1022 mb_type = CANDIDATE_MB_TYPE_INTER_I;
1023 dmin= dmin_i;
1024 }
1025 }
1026
1027 set_p_mv_tables(s, mx, my, mb_type!=CANDIDATE_MB_TYPE_INTER4V);
1028
1029 /* get intra luma score */
1030 if((c->avctx->mb_cmp&0xFF)==FF_CMP_SSE){
1031 intra_score= varc - 500;
1032 }else{
1033 unsigned mean = (sum+128)>>8;
1034 mean*= 0x01010101;
1035
1036 for(i=0; i<16; i++){
1037 *(uint32_t*)(&c->scratchpad[i*s->linesize+ 0]) = mean;
1038 *(uint32_t*)(&c->scratchpad[i*s->linesize+ 4]) = mean;
1039 *(uint32_t*)(&c->scratchpad[i*s->linesize+ 8]) = mean;
1040 *(uint32_t*)(&c->scratchpad[i*s->linesize+12]) = mean;
1041 }
1042
1043 intra_score= s->mecc.mb_cmp[0](s, c->scratchpad, pix, s->linesize, 16);
1044 }
1045 intra_score += c->mb_penalty_factor*16 + s->intra_penalty;
1046
1047 if(intra_score < dmin){
1048 mb_type= CANDIDATE_MB_TYPE_INTRA;
1049 s->current_picture.mb_type[mb_y*s->mb_stride + mb_x] = CANDIDATE_MB_TYPE_INTRA; //FIXME cleanup
1050 }else
1051 s->current_picture.mb_type[mb_y*s->mb_stride + mb_x] = 0;
1052
1053 {
1054 int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
1055 int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
1056 c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
1057 }
1058 }
1059
1060 s->mb_type[mb_y*s->mb_stride + mb_x]= mb_type;
1061 }
1062
ff_pre_estimate_p_frame_motion(MpegEncContext * s,int mb_x,int mb_y)1063 int ff_pre_estimate_p_frame_motion(MpegEncContext * s,
1064 int mb_x, int mb_y)
1065 {
1066 MotionEstContext * const c= &s->me;
1067 int mx, my, dmin;
1068 int P[10][2];
1069 const int shift= 1+s->quarter_sample;
1070 const int xy= mb_x + mb_y*s->mb_stride;
1071 init_ref(c, s->new_picture->data, s->last_picture.f->data, NULL, 16*mb_x, 16*mb_y, 0);
1072
1073 av_assert0(s->quarter_sample==0 || s->quarter_sample==1);
1074
1075 c->pre_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_pre_cmp);
1076 c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_DMV;
1077
1078 get_limits(s, 16*mb_x, 16*mb_y);
1079 c->skip=0;
1080
1081 P_LEFT[0] = s->p_mv_table[xy + 1][0];
1082 P_LEFT[1] = s->p_mv_table[xy + 1][1];
1083
1084 if(P_LEFT[0] < (c->xmin<<shift)) P_LEFT[0] = (c->xmin<<shift);
1085
1086 /* special case for first line */
1087 if (s->first_slice_line) {
1088 c->pred_x= P_LEFT[0];
1089 c->pred_y= P_LEFT[1];
1090 P_TOP[0]= P_TOPRIGHT[0]= P_MEDIAN[0]=
1091 P_TOP[1]= P_TOPRIGHT[1]= P_MEDIAN[1]= 0; //FIXME
1092 } else {
1093 P_TOP[0] = s->p_mv_table[xy + s->mb_stride ][0];
1094 P_TOP[1] = s->p_mv_table[xy + s->mb_stride ][1];
1095 P_TOPRIGHT[0] = s->p_mv_table[xy + s->mb_stride - 1][0];
1096 P_TOPRIGHT[1] = s->p_mv_table[xy + s->mb_stride - 1][1];
1097 if(P_TOP[1] < (c->ymin<<shift)) P_TOP[1] = (c->ymin<<shift);
1098 if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift);
1099 if(P_TOPRIGHT[1] < (c->ymin<<shift)) P_TOPRIGHT[1]= (c->ymin<<shift);
1100
1101 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1102 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1103
1104 c->pred_x = P_MEDIAN[0];
1105 c->pred_y = P_MEDIAN[1];
1106 }
1107
1108 dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift, 0, 16);
1109
1110 s->p_mv_table[xy][0] = mx<<shift;
1111 s->p_mv_table[xy][1] = my<<shift;
1112
1113 return dmin;
1114 }
1115
estimate_motion_b(MpegEncContext * s,int mb_x,int mb_y,int16_t (* mv_table)[2],int ref_index,int f_code)1116 static int estimate_motion_b(MpegEncContext *s, int mb_x, int mb_y,
1117 int16_t (*mv_table)[2], int ref_index, int f_code)
1118 {
1119 MotionEstContext * const c= &s->me;
1120 int mx = 0, my = 0, dmin = 0;
1121 int P[10][2];
1122 const int shift= 1+s->quarter_sample;
1123 const int mot_stride = s->mb_stride;
1124 const int mot_xy = mb_y*mot_stride + mb_x;
1125 const uint8_t * const mv_penalty = c->mv_penalty[f_code] + MAX_DMV;
1126 int mv_scale;
1127
1128 c->penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
1129 c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
1130 c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
1131 c->current_mv_penalty= mv_penalty;
1132
1133 get_limits(s, 16*mb_x, 16*mb_y);
1134
1135 if (s->motion_est != FF_ME_ZERO) {
1136 P_LEFT[0] = mv_table[mot_xy - 1][0];
1137 P_LEFT[1] = mv_table[mot_xy - 1][1];
1138
1139 if (P_LEFT[0] > (c->xmax << shift)) P_LEFT[0] = (c->xmax << shift);
1140
1141 /* special case for first line */
1142 if (!s->first_slice_line) {
1143 P_TOP[0] = mv_table[mot_xy - mot_stride ][0];
1144 P_TOP[1] = mv_table[mot_xy - mot_stride ][1];
1145 P_TOPRIGHT[0] = mv_table[mot_xy - mot_stride + 1][0];
1146 P_TOPRIGHT[1] = mv_table[mot_xy - mot_stride + 1][1];
1147 if (P_TOP[1] > (c->ymax << shift)) P_TOP[1] = (c->ymax << shift);
1148 if (P_TOPRIGHT[0] < c->xmin * (1 << shift)) P_TOPRIGHT[0] = c->xmin * (1 << shift);
1149 if (P_TOPRIGHT[1] > (c->ymax << shift)) P_TOPRIGHT[1] = (c->ymax << shift);
1150
1151 P_MEDIAN[0] = mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1152 P_MEDIAN[1] = mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1153 }
1154 c->pred_x = P_LEFT[0];
1155 c->pred_y = P_LEFT[1];
1156
1157 if(mv_table == s->b_forw_mv_table){
1158 mv_scale= (s->pb_time<<16) / (s->pp_time<<shift);
1159 }else{
1160 mv_scale = ((s->pb_time - s->pp_time) * (1 << 16)) / (s->pp_time<<shift);
1161 }
1162
1163 dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, ref_index, s->p_mv_table, mv_scale, 0, 16);
1164 }
1165
1166 dmin= c->sub_motion_search(s, &mx, &my, dmin, 0, ref_index, 0, 16);
1167
1168 if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
1169 dmin= get_mb_score(s, mx, my, 0, ref_index, 0, 16, 1);
1170
1171 // s->mb_type[mb_y*s->mb_width + mb_x]= mb_type;
1172 mv_table[mot_xy][0]= mx;
1173 mv_table[mot_xy][1]= my;
1174
1175 return dmin;
1176 }
1177
check_bidir_mv(MpegEncContext * s,int motion_fx,int motion_fy,int motion_bx,int motion_by,int pred_fx,int pred_fy,int pred_bx,int pred_by,int size,int h)1178 static inline int check_bidir_mv(MpegEncContext * s,
1179 int motion_fx, int motion_fy,
1180 int motion_bx, int motion_by,
1181 int pred_fx, int pred_fy,
1182 int pred_bx, int pred_by,
1183 int size, int h)
1184 {
1185 //FIXME optimize?
1186 //FIXME better f_code prediction (max mv & distance)
1187 //FIXME pointers
1188 MotionEstContext * const c= &s->me;
1189 const uint8_t * const mv_penalty_f = c->mv_penalty[s->f_code] + MAX_DMV; // f_code of the prev frame
1190 const uint8_t * const mv_penalty_b = c->mv_penalty[s->b_code] + MAX_DMV; // f_code of the prev frame
1191 int stride= c->stride;
1192 uint8_t *dest_y = c->scratchpad;
1193 uint8_t *ptr;
1194 int dxy;
1195 int src_x, src_y;
1196 int fbmin;
1197 uint8_t **src_data= c->src[0];
1198 uint8_t **ref_data= c->ref[0];
1199 uint8_t **ref2_data= c->ref[2];
1200
1201 if(s->quarter_sample){
1202 dxy = ((motion_fy & 3) << 2) | (motion_fx & 3);
1203 src_x = motion_fx >> 2;
1204 src_y = motion_fy >> 2;
1205
1206 ptr = ref_data[0] + (src_y * stride) + src_x;
1207 s->qdsp.put_qpel_pixels_tab[0][dxy](dest_y, ptr, stride);
1208
1209 dxy = ((motion_by & 3) << 2) | (motion_bx & 3);
1210 src_x = motion_bx >> 2;
1211 src_y = motion_by >> 2;
1212
1213 ptr = ref2_data[0] + (src_y * stride) + src_x;
1214 s->qdsp.avg_qpel_pixels_tab[size][dxy](dest_y, ptr, stride);
1215 }else{
1216 dxy = ((motion_fy & 1) << 1) | (motion_fx & 1);
1217 src_x = motion_fx >> 1;
1218 src_y = motion_fy >> 1;
1219
1220 ptr = ref_data[0] + (src_y * stride) + src_x;
1221 s->hdsp.put_pixels_tab[size][dxy](dest_y , ptr , stride, h);
1222
1223 dxy = ((motion_by & 1) << 1) | (motion_bx & 1);
1224 src_x = motion_bx >> 1;
1225 src_y = motion_by >> 1;
1226
1227 ptr = ref2_data[0] + (src_y * stride) + src_x;
1228 s->hdsp.avg_pixels_tab[size][dxy](dest_y , ptr , stride, h);
1229 }
1230
1231 fbmin = (mv_penalty_f[motion_fx-pred_fx] + mv_penalty_f[motion_fy-pred_fy])*c->mb_penalty_factor
1232 +(mv_penalty_b[motion_bx-pred_bx] + mv_penalty_b[motion_by-pred_by])*c->mb_penalty_factor
1233 + s->mecc.mb_cmp[size](s, src_data[0], dest_y, stride, h); // FIXME new_pic
1234
1235 if(c->avctx->mb_cmp&FF_CMP_CHROMA){
1236 }
1237 //FIXME CHROMA !!!
1238
1239 return fbmin;
1240 }
1241
1242 /* refine the bidir vectors in hq mode and return the score in both lq & hq mode*/
bidir_refine(MpegEncContext * s,int mb_x,int mb_y)1243 static inline int bidir_refine(MpegEncContext * s, int mb_x, int mb_y)
1244 {
1245 MotionEstContext * const c= &s->me;
1246 const int mot_stride = s->mb_stride;
1247 const int xy = mb_y *mot_stride + mb_x;
1248 int fbmin;
1249 int pred_fx= s->b_bidir_forw_mv_table[xy-1][0];
1250 int pred_fy= s->b_bidir_forw_mv_table[xy-1][1];
1251 int pred_bx= s->b_bidir_back_mv_table[xy-1][0];
1252 int pred_by= s->b_bidir_back_mv_table[xy-1][1];
1253 int motion_fx= s->b_bidir_forw_mv_table[xy][0]= s->b_forw_mv_table[xy][0];
1254 int motion_fy= s->b_bidir_forw_mv_table[xy][1]= s->b_forw_mv_table[xy][1];
1255 int motion_bx= s->b_bidir_back_mv_table[xy][0]= s->b_back_mv_table[xy][0];
1256 int motion_by= s->b_bidir_back_mv_table[xy][1]= s->b_back_mv_table[xy][1];
1257 const int flags= c->sub_flags;
1258 const int qpel= flags&FLAG_QPEL;
1259 const int shift= 1+qpel;
1260 const int xmin= c->xmin * (1 << shift);
1261 const int ymin= c->ymin * (1 << shift);
1262 const int xmax= c->xmax<<shift;
1263 const int ymax= c->ymax<<shift;
1264 #define HASH(fx,fy,bx,by) ((fx)+17*(fy)+63*(bx)+117*(by))
1265 #define HASH8(fx,fy,bx,by) ((uint8_t)HASH(fx,fy,bx,by))
1266 int hashidx= HASH(motion_fx,motion_fy, motion_bx, motion_by);
1267 uint8_t map[256] = { 0 };
1268
1269 map[hashidx&255] = 1;
1270
1271 fbmin= check_bidir_mv(s, motion_fx, motion_fy,
1272 motion_bx, motion_by,
1273 pred_fx, pred_fy,
1274 pred_bx, pred_by,
1275 0, 16);
1276
1277 if(s->avctx->bidir_refine){
1278 int end;
1279 static const uint8_t limittab[5]={0,8,32,64,80};
1280 const int limit= limittab[s->avctx->bidir_refine];
1281 static const int8_t vect[][4]={
1282 { 0, 0, 0, 1}, { 0, 0, 0,-1}, { 0, 0, 1, 0}, { 0, 0,-1, 0}, { 0, 1, 0, 0}, { 0,-1, 0, 0}, { 1, 0, 0, 0}, {-1, 0, 0, 0},
1283
1284 { 0, 0, 1, 1}, { 0, 0,-1,-1}, { 0, 1, 1, 0}, { 0,-1,-1, 0}, { 1, 1, 0, 0}, {-1,-1, 0, 0}, { 1, 0, 0, 1}, {-1, 0, 0,-1},
1285 { 0, 1, 0, 1}, { 0,-1, 0,-1}, { 1, 0, 1, 0}, {-1, 0,-1, 0},
1286 { 0, 0,-1, 1}, { 0, 0, 1,-1}, { 0,-1, 1, 0}, { 0, 1,-1, 0}, {-1, 1, 0, 0}, { 1,-1, 0, 0}, { 1, 0, 0,-1}, {-1, 0, 0, 1},
1287 { 0,-1, 0, 1}, { 0, 1, 0,-1}, {-1, 0, 1, 0}, { 1, 0,-1, 0},
1288
1289 { 0, 1, 1, 1}, { 0,-1,-1,-1}, { 1, 1, 1, 0}, {-1,-1,-1, 0}, { 1, 1, 0, 1}, {-1,-1, 0,-1}, { 1, 0, 1, 1}, {-1, 0,-1,-1},
1290 { 0,-1, 1, 1}, { 0, 1,-1,-1}, {-1, 1, 1, 0}, { 1,-1,-1, 0}, { 1, 1, 0,-1}, {-1,-1, 0, 1}, { 1, 0,-1, 1}, {-1, 0, 1,-1},
1291 { 0, 1,-1, 1}, { 0,-1, 1,-1}, { 1,-1, 1, 0}, {-1, 1,-1, 0}, {-1, 1, 0, 1}, { 1,-1, 0,-1}, { 1, 0, 1,-1}, {-1, 0,-1, 1},
1292 { 0, 1, 1,-1}, { 0,-1,-1, 1}, { 1, 1,-1, 0}, {-1,-1, 1, 0}, { 1,-1, 0, 1}, {-1, 1, 0,-1}, {-1, 0, 1, 1}, { 1, 0,-1,-1},
1293
1294 { 1, 1, 1, 1}, {-1,-1,-1,-1},
1295 { 1, 1, 1,-1}, {-1,-1,-1, 1}, { 1, 1,-1, 1}, {-1,-1, 1,-1}, { 1,-1, 1, 1}, {-1, 1,-1,-1}, {-1, 1, 1, 1}, { 1,-1,-1,-1},
1296 { 1, 1,-1,-1}, {-1,-1, 1, 1}, { 1,-1,-1, 1}, {-1, 1, 1,-1}, { 1,-1, 1,-1}, {-1, 1,-1, 1},
1297 };
1298 static const uint8_t hash[]={
1299 HASH8( 0, 0, 0, 1), HASH8( 0, 0, 0,-1), HASH8( 0, 0, 1, 0), HASH8( 0, 0,-1, 0), HASH8( 0, 1, 0, 0), HASH8( 0,-1, 0, 0), HASH8( 1, 0, 0, 0), HASH8(-1, 0, 0, 0),
1300
1301 HASH8( 0, 0, 1, 1), HASH8( 0, 0,-1,-1), HASH8( 0, 1, 1, 0), HASH8( 0,-1,-1, 0), HASH8( 1, 1, 0, 0), HASH8(-1,-1, 0, 0), HASH8( 1, 0, 0, 1), HASH8(-1, 0, 0,-1),
1302 HASH8( 0, 1, 0, 1), HASH8( 0,-1, 0,-1), HASH8( 1, 0, 1, 0), HASH8(-1, 0,-1, 0),
1303 HASH8( 0, 0,-1, 1), HASH8( 0, 0, 1,-1), HASH8( 0,-1, 1, 0), HASH8( 0, 1,-1, 0), HASH8(-1, 1, 0, 0), HASH8( 1,-1, 0, 0), HASH8( 1, 0, 0,-1), HASH8(-1, 0, 0, 1),
1304 HASH8( 0,-1, 0, 1), HASH8( 0, 1, 0,-1), HASH8(-1, 0, 1, 0), HASH8( 1, 0,-1, 0),
1305
1306 HASH8( 0, 1, 1, 1), HASH8( 0,-1,-1,-1), HASH8( 1, 1, 1, 0), HASH8(-1,-1,-1, 0), HASH8( 1, 1, 0, 1), HASH8(-1,-1, 0,-1), HASH8( 1, 0, 1, 1), HASH8(-1, 0,-1,-1),
1307 HASH8( 0,-1, 1, 1), HASH8( 0, 1,-1,-1), HASH8(-1, 1, 1, 0), HASH8( 1,-1,-1, 0), HASH8( 1, 1, 0,-1), HASH8(-1,-1, 0, 1), HASH8( 1, 0,-1, 1), HASH8(-1, 0, 1,-1),
1308 HASH8( 0, 1,-1, 1), HASH8( 0,-1, 1,-1), HASH8( 1,-1, 1, 0), HASH8(-1, 1,-1, 0), HASH8(-1, 1, 0, 1), HASH8( 1,-1, 0,-1), HASH8( 1, 0, 1,-1), HASH8(-1, 0,-1, 1),
1309 HASH8( 0, 1, 1,-1), HASH8( 0,-1,-1, 1), HASH8( 1, 1,-1, 0), HASH8(-1,-1, 1, 0), HASH8( 1,-1, 0, 1), HASH8(-1, 1, 0,-1), HASH8(-1, 0, 1, 1), HASH8( 1, 0,-1,-1),
1310
1311 HASH8( 1, 1, 1, 1), HASH8(-1,-1,-1,-1),
1312 HASH8( 1, 1, 1,-1), HASH8(-1,-1,-1, 1), HASH8( 1, 1,-1, 1), HASH8(-1,-1, 1,-1), HASH8( 1,-1, 1, 1), HASH8(-1, 1,-1,-1), HASH8(-1, 1, 1, 1), HASH8( 1,-1,-1,-1),
1313 HASH8( 1, 1,-1,-1), HASH8(-1,-1, 1, 1), HASH8( 1,-1,-1, 1), HASH8(-1, 1, 1,-1), HASH8( 1,-1, 1,-1), HASH8(-1, 1,-1, 1),
1314 };
1315
1316 #define CHECK_BIDIR(fx,fy,bx,by)\
1317 if( !map[(hashidx+HASH(fx,fy,bx,by))&255]\
1318 &&(fx<=0 || motion_fx+fx<=xmax) && (fy<=0 || motion_fy+fy<=ymax) && (bx<=0 || motion_bx+bx<=xmax) && (by<=0 || motion_by+by<=ymax)\
1319 &&(fx>=0 || motion_fx+fx>=xmin) && (fy>=0 || motion_fy+fy>=ymin) && (bx>=0 || motion_bx+bx>=xmin) && (by>=0 || motion_by+by>=ymin)){\
1320 int score;\
1321 map[(hashidx+HASH(fx,fy,bx,by))&255] = 1;\
1322 score= check_bidir_mv(s, motion_fx+fx, motion_fy+fy, motion_bx+bx, motion_by+by, pred_fx, pred_fy, pred_bx, pred_by, 0, 16);\
1323 if(score < fbmin){\
1324 hashidx += HASH(fx,fy,bx,by);\
1325 fbmin= score;\
1326 motion_fx+=fx;\
1327 motion_fy+=fy;\
1328 motion_bx+=bx;\
1329 motion_by+=by;\
1330 end=0;\
1331 }\
1332 }
1333 #define CHECK_BIDIR2(a,b,c,d)\
1334 CHECK_BIDIR(a,b,c,d)\
1335 CHECK_BIDIR(-(a),-(b),-(c),-(d))
1336
1337 do{
1338 int i;
1339 int borderdist=0;
1340 end=1;
1341
1342 CHECK_BIDIR2(0,0,0,1)
1343 CHECK_BIDIR2(0,0,1,0)
1344 CHECK_BIDIR2(0,1,0,0)
1345 CHECK_BIDIR2(1,0,0,0)
1346
1347 for(i=8; i<limit; i++){
1348 int fx= motion_fx+vect[i][0];
1349 int fy= motion_fy+vect[i][1];
1350 int bx= motion_bx+vect[i][2];
1351 int by= motion_by+vect[i][3];
1352 if(borderdist<=0){
1353 int a= (xmax - FFMAX(fx,bx))|(FFMIN(fx,bx) - xmin);
1354 int b= (ymax - FFMAX(fy,by))|(FFMIN(fy,by) - ymin);
1355 if((a|b) < 0)
1356 map[(hashidx+hash[i])&255] = 1;
1357 }
1358 if(!map[(hashidx+hash[i])&255]){
1359 int score;
1360 map[(hashidx+hash[i])&255] = 1;
1361 score= check_bidir_mv(s, fx, fy, bx, by, pred_fx, pred_fy, pred_bx, pred_by, 0, 16);
1362 if(score < fbmin){
1363 hashidx += hash[i];
1364 fbmin= score;
1365 motion_fx=fx;
1366 motion_fy=fy;
1367 motion_bx=bx;
1368 motion_by=by;
1369 end=0;
1370 borderdist--;
1371 if(borderdist<=0){
1372 int a= FFMIN(xmax - FFMAX(fx,bx), FFMIN(fx,bx) - xmin);
1373 int b= FFMIN(ymax - FFMAX(fy,by), FFMIN(fy,by) - ymin);
1374 borderdist= FFMIN(a,b);
1375 }
1376 }
1377 }
1378 }
1379 }while(!end);
1380 }
1381
1382 s->b_bidir_forw_mv_table[xy][0]= motion_fx;
1383 s->b_bidir_forw_mv_table[xy][1]= motion_fy;
1384 s->b_bidir_back_mv_table[xy][0]= motion_bx;
1385 s->b_bidir_back_mv_table[xy][1]= motion_by;
1386
1387 return fbmin;
1388 }
1389
direct_search(MpegEncContext * s,int mb_x,int mb_y)1390 static inline int direct_search(MpegEncContext * s, int mb_x, int mb_y)
1391 {
1392 MotionEstContext * const c= &s->me;
1393 int P[10][2];
1394 const int mot_stride = s->mb_stride;
1395 const int mot_xy = mb_y*mot_stride + mb_x;
1396 const int shift= 1+s->quarter_sample;
1397 int dmin, i;
1398 const int time_pp= s->pp_time;
1399 const int time_pb= s->pb_time;
1400 int mx, my, xmin, xmax, ymin, ymax;
1401 int16_t (*mv_table)[2]= s->b_direct_mv_table;
1402
1403 c->current_mv_penalty= c->mv_penalty[1] + MAX_DMV;
1404 ymin= xmin=(-32)>>shift;
1405 ymax= xmax= 31>>shift;
1406
1407 if (IS_8X8(s->next_picture.mb_type[mot_xy])) {
1408 s->mv_type= MV_TYPE_8X8;
1409 }else{
1410 s->mv_type= MV_TYPE_16X16;
1411 }
1412
1413 for(i=0; i<4; i++){
1414 int index= s->block_index[i];
1415 int min, max;
1416
1417 c->co_located_mv[i][0] = s->next_picture.motion_val[0][index][0];
1418 c->co_located_mv[i][1] = s->next_picture.motion_val[0][index][1];
1419 c->direct_basis_mv[i][0]= c->co_located_mv[i][0]*time_pb/time_pp + ((i& 1)<<(shift+3));
1420 c->direct_basis_mv[i][1]= c->co_located_mv[i][1]*time_pb/time_pp + ((i>>1)<<(shift+3));
1421 // c->direct_basis_mv[1][i][0]= c->co_located_mv[i][0]*(time_pb - time_pp)/time_pp + ((i &1)<<(shift+3);
1422 // c->direct_basis_mv[1][i][1]= c->co_located_mv[i][1]*(time_pb - time_pp)/time_pp + ((i>>1)<<(shift+3);
1423
1424 max= FFMAX(c->direct_basis_mv[i][0], c->direct_basis_mv[i][0] - c->co_located_mv[i][0])>>shift;
1425 min= FFMIN(c->direct_basis_mv[i][0], c->direct_basis_mv[i][0] - c->co_located_mv[i][0])>>shift;
1426 max+= 16*mb_x + 1; // +-1 is for the simpler rounding
1427 min+= 16*mb_x - 1;
1428 xmax= FFMIN(xmax, s->width - max);
1429 xmin= FFMAX(xmin, - 16 - min);
1430
1431 max= FFMAX(c->direct_basis_mv[i][1], c->direct_basis_mv[i][1] - c->co_located_mv[i][1])>>shift;
1432 min= FFMIN(c->direct_basis_mv[i][1], c->direct_basis_mv[i][1] - c->co_located_mv[i][1])>>shift;
1433 max+= 16*mb_y + 1; // +-1 is for the simpler rounding
1434 min+= 16*mb_y - 1;
1435 ymax= FFMIN(ymax, s->height - max);
1436 ymin= FFMAX(ymin, - 16 - min);
1437
1438 if(s->mv_type == MV_TYPE_16X16) break;
1439 }
1440
1441 av_assert2(xmax <= 15 && ymax <= 15 && xmin >= -16 && ymin >= -16);
1442
1443 if(xmax < 0 || xmin >0 || ymax < 0 || ymin > 0){
1444 s->b_direct_mv_table[mot_xy][0]= 0;
1445 s->b_direct_mv_table[mot_xy][1]= 0;
1446
1447 return 256*256*256*64;
1448 }
1449
1450 c->xmin= xmin;
1451 c->ymin= ymin;
1452 c->xmax= xmax;
1453 c->ymax= ymax;
1454 c->flags |= FLAG_DIRECT;
1455 c->sub_flags |= FLAG_DIRECT;
1456 c->pred_x=0;
1457 c->pred_y=0;
1458
1459 P_LEFT[0] = av_clip(mv_table[mot_xy - 1][0], xmin * (1 << shift), xmax << shift);
1460 P_LEFT[1] = av_clip(mv_table[mot_xy - 1][1], ymin * (1 << shift), ymax << shift);
1461
1462 /* special case for first line */
1463 if (!s->first_slice_line) { //FIXME maybe allow this over thread boundary as it is clipped
1464 P_TOP[0] = av_clip(mv_table[mot_xy - mot_stride ][0], xmin * (1 << shift), xmax << shift);
1465 P_TOP[1] = av_clip(mv_table[mot_xy - mot_stride ][1], ymin * (1 << shift), ymax << shift);
1466 P_TOPRIGHT[0] = av_clip(mv_table[mot_xy - mot_stride + 1][0], xmin * (1 << shift), xmax << shift);
1467 P_TOPRIGHT[1] = av_clip(mv_table[mot_xy - mot_stride + 1][1], ymin * (1 << shift), ymax << shift);
1468
1469 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1470 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1471 }
1472
1473 dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, mv_table, 1<<(16-shift), 0, 16);
1474 if(c->sub_flags&FLAG_QPEL)
1475 dmin = qpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1476 else
1477 dmin = hpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1478
1479 if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
1480 dmin= get_mb_score(s, mx, my, 0, 0, 0, 16, 1);
1481
1482 get_limits(s, 16*mb_x, 16*mb_y); //restore c->?min/max, maybe not needed
1483
1484 mv_table[mot_xy][0]= mx;
1485 mv_table[mot_xy][1]= my;
1486 c->flags &= ~FLAG_DIRECT;
1487 c->sub_flags &= ~FLAG_DIRECT;
1488
1489 return dmin;
1490 }
1491
ff_estimate_b_frame_motion(MpegEncContext * s,int mb_x,int mb_y)1492 void ff_estimate_b_frame_motion(MpegEncContext * s,
1493 int mb_x, int mb_y)
1494 {
1495 MotionEstContext * const c= &s->me;
1496 const int penalty_factor= c->mb_penalty_factor;
1497 int fmin, bmin, dmin, fbmin, bimin, fimin;
1498 int type=0;
1499 const int xy = mb_y*s->mb_stride + mb_x;
1500 init_ref(c, s->new_picture->data, s->last_picture.f->data,
1501 s->next_picture.f->data, 16 * mb_x, 16 * mb_y, 2);
1502
1503 get_limits(s, 16*mb_x, 16*mb_y);
1504
1505 c->skip=0;
1506
1507 if (s->codec_id == AV_CODEC_ID_MPEG4 && s->next_picture.mbskip_table[xy]) {
1508 int score= direct_search(s, mb_x, mb_y); //FIXME just check 0,0
1509
1510 score= ((unsigned)(score*score + 128*256))>>16;
1511 c->mc_mb_var_sum_temp += score;
1512 s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE
1513 s->mb_type[mb_y*s->mb_stride + mb_x]= CANDIDATE_MB_TYPE_DIRECT0;
1514
1515 return;
1516 }
1517
1518 if (s->codec_id == AV_CODEC_ID_MPEG4)
1519 dmin= direct_search(s, mb_x, mb_y);
1520 else
1521 dmin= INT_MAX;
1522 // FIXME penalty stuff for non-MPEG-4
1523 c->skip=0;
1524 fmin = estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, 0, s->f_code) +
1525 3 * penalty_factor;
1526
1527 c->skip=0;
1528 bmin = estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, 2, s->b_code) +
1529 2 * penalty_factor;
1530 ff_dlog(s, " %d %d ", s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1]);
1531
1532 c->skip=0;
1533 fbmin= bidir_refine(s, mb_x, mb_y) + penalty_factor;
1534 ff_dlog(s, "%d %d %d %d\n", dmin, fmin, bmin, fbmin);
1535
1536 if (s->avctx->flags & AV_CODEC_FLAG_INTERLACED_ME) {
1537 //FIXME mb type penalty
1538 c->skip=0;
1539 c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_DMV;
1540 fimin= interlaced_search(s, 0,
1541 s->b_field_mv_table[0], s->b_field_select_table[0],
1542 s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1], 0);
1543 c->current_mv_penalty= c->mv_penalty[s->b_code] + MAX_DMV;
1544 bimin= interlaced_search(s, 2,
1545 s->b_field_mv_table[1], s->b_field_select_table[1],
1546 s->b_back_mv_table[xy][0], s->b_back_mv_table[xy][1], 0);
1547 }else
1548 fimin= bimin= INT_MAX;
1549
1550 {
1551 int score= fmin;
1552 type = CANDIDATE_MB_TYPE_FORWARD;
1553
1554 if (dmin <= score){
1555 score = dmin;
1556 type = CANDIDATE_MB_TYPE_DIRECT;
1557 }
1558 if(bmin<score){
1559 score=bmin;
1560 type= CANDIDATE_MB_TYPE_BACKWARD;
1561 }
1562 if(fbmin<score){
1563 score=fbmin;
1564 type= CANDIDATE_MB_TYPE_BIDIR;
1565 }
1566 if(fimin<score){
1567 score=fimin;
1568 type= CANDIDATE_MB_TYPE_FORWARD_I;
1569 }
1570 if(bimin<score){
1571 score=bimin;
1572 type= CANDIDATE_MB_TYPE_BACKWARD_I;
1573 }
1574
1575 score= ((unsigned)(score*score + 128*256))>>16;
1576 c->mc_mb_var_sum_temp += score;
1577 s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE
1578 }
1579
1580 if(c->avctx->mb_decision > FF_MB_DECISION_SIMPLE){
1581 type= CANDIDATE_MB_TYPE_FORWARD | CANDIDATE_MB_TYPE_BACKWARD | CANDIDATE_MB_TYPE_BIDIR | CANDIDATE_MB_TYPE_DIRECT;
1582 if(fimin < INT_MAX)
1583 type |= CANDIDATE_MB_TYPE_FORWARD_I;
1584 if(bimin < INT_MAX)
1585 type |= CANDIDATE_MB_TYPE_BACKWARD_I;
1586 if(fimin < INT_MAX && bimin < INT_MAX){
1587 type |= CANDIDATE_MB_TYPE_BIDIR_I;
1588 }
1589 //FIXME something smarter
1590 if(dmin>256*256*16) type&= ~CANDIDATE_MB_TYPE_DIRECT; //do not try direct mode if it is invalid for this MB
1591 if (s->codec_id == AV_CODEC_ID_MPEG4 && type&CANDIDATE_MB_TYPE_DIRECT &&
1592 s->mpv_flags & FF_MPV_FLAG_MV0 && *(uint32_t*)s->b_direct_mv_table[xy])
1593 type |= CANDIDATE_MB_TYPE_DIRECT0;
1594 }
1595
1596 s->mb_type[mb_y*s->mb_stride + mb_x]= type;
1597 }
1598
1599 /* find best f_code for ME which do unlimited searches */
ff_get_best_fcode(MpegEncContext * s,const int16_t (* mv_table)[2],int type)1600 int ff_get_best_fcode(MpegEncContext * s, const int16_t (*mv_table)[2], int type)
1601 {
1602 if (s->motion_est != FF_ME_ZERO) {
1603 int score[8];
1604 int i, y, range= s->avctx->me_range ? s->avctx->me_range : (INT_MAX/2);
1605 const uint8_t * fcode_tab = s->fcode_tab;
1606 int best_fcode=-1;
1607 int best_score=-10000000;
1608
1609 if(s->msmpeg4_version)
1610 range= FFMIN(range, 16);
1611 else if(s->codec_id == AV_CODEC_ID_MPEG2VIDEO && s->avctx->strict_std_compliance >= FF_COMPLIANCE_NORMAL)
1612 range= FFMIN(range, 256);
1613
1614 for(i=0; i<8; i++) score[i]= s->mb_num*(8-i);
1615
1616 for(y=0; y<s->mb_height; y++){
1617 int x;
1618 int xy= y*s->mb_stride;
1619 for(x=0; x<s->mb_width; x++, xy++){
1620 if(s->mb_type[xy] & type){
1621 int mx= mv_table[xy][0];
1622 int my= mv_table[xy][1];
1623 int fcode= FFMAX(fcode_tab[mx + MAX_MV],
1624 fcode_tab[my + MAX_MV]);
1625 int j;
1626
1627 if (mx >= range || mx < -range ||
1628 my >= range || my < -range)
1629 continue;
1630
1631 for(j=0; j<fcode && j<8; j++){
1632 if(s->pict_type==AV_PICTURE_TYPE_B || s->current_picture.mc_mb_var[xy] < s->current_picture.mb_var[xy])
1633 score[j]-= 170;
1634 }
1635 }
1636 }
1637 }
1638
1639 for(i=1; i<8; i++){
1640 if(score[i] > best_score){
1641 best_score= score[i];
1642 best_fcode= i;
1643 }
1644 }
1645
1646 return best_fcode;
1647 }else{
1648 return 1;
1649 }
1650 }
1651
ff_fix_long_p_mvs(MpegEncContext * s,int type)1652 void ff_fix_long_p_mvs(MpegEncContext * s, int type)
1653 {
1654 MotionEstContext * const c= &s->me;
1655 const int f_code= s->f_code;
1656 int y, range;
1657 av_assert0(s->pict_type==AV_PICTURE_TYPE_P);
1658
1659 range = (((s->out_format == FMT_MPEG1 || s->msmpeg4_version) ? 8 : 16) << f_code);
1660
1661 av_assert0(range <= 16 || !s->msmpeg4_version);
1662 av_assert0(range <=256 || !(s->codec_id == AV_CODEC_ID_MPEG2VIDEO && s->avctx->strict_std_compliance >= FF_COMPLIANCE_NORMAL));
1663
1664 if(c->avctx->me_range && range > c->avctx->me_range) range= c->avctx->me_range;
1665
1666 if (s->avctx->flags & AV_CODEC_FLAG_4MV) {
1667 const int wrap= s->b8_stride;
1668
1669 /* clip / convert to intra 8x8 type MVs */
1670 for(y=0; y<s->mb_height; y++){
1671 int xy= y*2*wrap;
1672 int i= y*s->mb_stride;
1673 int x;
1674
1675 for(x=0; x<s->mb_width; x++){
1676 if(s->mb_type[i]&CANDIDATE_MB_TYPE_INTER4V){
1677 int block;
1678 for(block=0; block<4; block++){
1679 int off= (block& 1) + (block>>1)*wrap;
1680 int mx = s->current_picture.motion_val[0][ xy + off ][0];
1681 int my = s->current_picture.motion_val[0][ xy + off ][1];
1682
1683 if( mx >=range || mx <-range
1684 || my >=range || my <-range){
1685 s->mb_type[i] &= ~CANDIDATE_MB_TYPE_INTER4V;
1686 s->mb_type[i] |= type;
1687 s->current_picture.mb_type[i] = type;
1688 }
1689 }
1690 }
1691 xy+=2;
1692 i++;
1693 }
1694 }
1695 }
1696 }
1697
1698 /**
1699 * @param truncate 1 for truncation, 0 for using intra
1700 */
ff_fix_long_mvs(MpegEncContext * s,uint8_t * field_select_table,int field_select,int16_t (* mv_table)[2],int f_code,int type,int truncate)1701 void ff_fix_long_mvs(MpegEncContext * s, uint8_t *field_select_table, int field_select,
1702 int16_t (*mv_table)[2], int f_code, int type, int truncate)
1703 {
1704 MotionEstContext * const c= &s->me;
1705 int y, h_range, v_range;
1706
1707 // RAL: 8 in MPEG-1, 16 in MPEG-4
1708 int range = (((s->out_format == FMT_MPEG1 || s->msmpeg4_version) ? 8 : 16) << f_code);
1709
1710 if(c->avctx->me_range && range > c->avctx->me_range) range= c->avctx->me_range;
1711
1712 h_range= range;
1713 v_range= field_select_table ? range>>1 : range;
1714
1715 /* clip / convert to intra 16x16 type MVs */
1716 for(y=0; y<s->mb_height; y++){
1717 int x;
1718 int xy= y*s->mb_stride;
1719 for(x=0; x<s->mb_width; x++){
1720 if (s->mb_type[xy] & type){ // RAL: "type" test added...
1721 if (!field_select_table || field_select_table[xy] == field_select) {
1722 if( mv_table[xy][0] >=h_range || mv_table[xy][0] <-h_range
1723 || mv_table[xy][1] >=v_range || mv_table[xy][1] <-v_range){
1724
1725 if(truncate){
1726 if (mv_table[xy][0] > h_range-1) mv_table[xy][0]= h_range-1;
1727 else if(mv_table[xy][0] < -h_range ) mv_table[xy][0]= -h_range;
1728 if (mv_table[xy][1] > v_range-1) mv_table[xy][1]= v_range-1;
1729 else if(mv_table[xy][1] < -v_range ) mv_table[xy][1]= -v_range;
1730 }else{
1731 s->mb_type[xy] &= ~type;
1732 s->mb_type[xy] |= CANDIDATE_MB_TYPE_INTRA;
1733 mv_table[xy][0]=
1734 mv_table[xy][1]= 0;
1735 }
1736 }
1737 }
1738 }
1739 xy++;
1740 }
1741 }
1742 }
1743