projects / libav.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
Change bidir refine hash code so we only need to perform a single
[libav.git] / libavcodec / motion_est.c
CommitLineData
de6d9b64 1/*
115329f1 2 * Motion estimation
406792e7 3 * Copyright (c) 2000,2001 Fabrice Bellard
8f2ab833 4 * Copyright (c) 2002-2004 Michael Niedermayer
115329f1 5 *
7b94177e 6 * new motion estimation (X1/EPZS) by Michael Niedermayer <michaelni@gmx.at>
de6d9b64 7 *
b78e7197
DB		 8 * This file is part of FFmpeg.
9 *
10 * FFmpeg is free software; you can redistribute it and/or
ff4ec49e
FB		 11 * modify it under the terms of the GNU Lesser General Public
12 * License as published by the Free Software Foundation; either
b78e7197 13 * version 2.1 of the License, or (at your option) any later version.
de6d9b64 14 *
b78e7197 15 * FFmpeg is distributed in the hope that it will be useful,
de6d9b64 16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
ff4ec49e
FB		 17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * Lesser General Public License for more details.
de6d9b64 19 *
ff4ec49e 20 * You should have received a copy of the GNU Lesser General Public
b78e7197 21 * License along with FFmpeg; if not, write to the Free Software
5509bffa 22 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
de6d9b64 23 */
115329f1 24
983e3246 25/**
bad5537e 26 * @file libavcodec/motion_est.c
983e3246
MN		 27 * Motion estimation.
28 */
115329f1 29
de6d9b64
FB		 30#include <stdlib.h>
31#include <stdio.h>
2d5e962b 32#include <limits.h>
94ca624f 33#include "libavutil/intmath.h"
de6d9b64
FB		 34#include "avcodec.h"
35#include "dsputil.h"
199436b9 36#include "mathops.h"
de6d9b64
FB		 37#include "mpegvideo.h"
38
2750b827
MN		 39#undef NDEBUG
40#include <assert.h>
1457ab52 41
0d21a846 42#define SQ(a) ((a)*(a))
b3987771 43
0d21a846
MN		 44#define P_LEFT P[1]
45#define P_TOP P[2]
46#define P_TOPRIGHT P[3]
47#define P_MEDIAN P[4]
0d21a846
MN		 48#define P_MV1 P[9]
49
1457ab52 50static inline int sad_hpel_motion_search(MpegEncContext * s,
bb270c08 51 int *mx_ptr, int *my_ptr, int dmin,
2750b827
MN		 52 int src_index, int ref_index,
53 int size, int h);
1457ab52 54
af4091f1 55static inline int update_map_generation(MotionEstContext *c)
1457ab52 56{
af4091f1
MN		 57 c->map_generation+= 1<<(ME_MAP_MV_BITS*2);
58 if(c->map_generation==0){
59 c->map_generation= 1<<(ME_MAP_MV_BITS*2);
60 memset(c->map, 0, sizeof(uint32_t)*ME_MAP_SIZE);
1457ab52 61 }
af4091f1 62 return c->map_generation;
1457ab52
MN		 63}
64
b07a5980
MN		 65/* shape adaptive search stuff */
66typedef struct Minima{
67 int height;
68 int x, y;
69 int checked;
70}Minima;
1457ab52 71
b07a5980 72static int minima_cmp(const void *a, const void *b){
5c91a675
ZK		 73 const Minima *da = (const Minima *) a;
74 const Minima *db = (const Minima *) b;
115329f1 75
b07a5980
MN		 76 return da->height - db->height;
77}
1457ab52 78
2750b827
MN		 79#define FLAG_QPEL 1 //must be 1
80#define FLAG_CHROMA 2
81#define FLAG_DIRECT 4
1457ab52 82
af4091f1 83static 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){
2750b827
MN		 84 const int offset[3]= {
85 y*c-> stride + x,
86 ((y*c->uvstride + x)>>1),
87 ((y*c->uvstride + x)>>1),
88 };
89 int i;
90 for(i=0; i<3; i++){
91 c->src[0][i]= src [i] + offset[i];
92 c->ref[0][i]= ref [i] + offset[i];
93 }
94 if(ref_index){
95 for(i=0; i<3; i++){
96 c->ref[ref_index][i]= ref2[i] + offset[i];
97 }
98 }
1457ab52
MN		 99}
100
1f202b0d
MN		 101static int get_flags(MotionEstContext *c, int direct, int chroma){
102 return ((c->avctx->flags&CODEC_FLAG_QPEL) ? FLAG_QPEL : 0)
115329f1 103 + (direct ? FLAG_DIRECT : 0)
2750b827 104 + (chroma ? FLAG_CHROMA : 0);
1457ab52
MN		 105}
106
919e7497 107static av_always_inline int cmp_direct_inline(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
2750b827 108 const int size, const int h, int ref_index, int src_index,
919e7497 109 me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, int qpel){
2750b827
MN		 110 MotionEstContext * const c= &s->me;
111 const int stride= c->stride;
2750b827
MN		 112 const int hx= subx + (x<<(1+qpel));
113 const int hy= suby + (y<<(1+qpel));
114 uint8_t * const * const ref= c->ref[ref_index];
115 uint8_t * const * const src= c->src[src_index];
116 int d;
117 //FIXME check chroma 4mv, (no crashes ...)
3b9cf3bd 118 assert(x >= c->xmin && hx <= c->xmax<<(qpel+1) && y >= c->ymin && hy <= c->ymax<<(qpel+1));
2750b827
MN		 119 if(x >= c->xmin && hx <= c->xmax<<(qpel+1) && y >= c->ymin && hy <= c->ymax<<(qpel+1)){
120 const int time_pp= s->pp_time;
121 const int time_pb= s->pb_time;
122 const int mask= 2*qpel+1;
123 if(s->mv_type==MV_TYPE_8X8){
124 int i;
125 for(i=0; i<4; i++){
126 int fx = c->direct_basis_mv[i][0] + hx;
127 int fy = c->direct_basis_mv[i][1] + hy;
128 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));
129 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));
130 int fxy= (fx&mask) + ((fy&mask)<<(qpel+1));
131 int bxy= (bx&mask) + ((by&mask)<<(qpel+1));
115329f1 132
2750b827
MN		 133 uint8_t *dst= c->temp + 8*(i&1) + 8*stride*(i>>1);
134 if(qpel){
135 c->qpel_put[1][fxy](dst, ref[0] + (fx>>2) + (fy>>2)*stride, stride);
136 c->qpel_avg[1][bxy](dst, ref[8] + (bx>>2) + (by>>2)*stride, stride);
137 }else{
138 c->hpel_put[1][fxy](dst, ref[0] + (fx>>1) + (fy>>1)*stride, stride, 8);
139 c->hpel_avg[1][bxy](dst, ref[8] + (bx>>1) + (by>>1)*stride, stride, 8);
140 }
141 }
142 }else{
143 int fx = c->direct_basis_mv[0][0] + hx;
144 int fy = c->direct_basis_mv[0][1] + hy;
145 int bx = hx ? fx - c->co_located_mv[0][0] : (c->co_located_mv[0][0]*(time_pb - time_pp)/time_pp);
146 int by = hy ? fy - c->co_located_mv[0][1] : (c->co_located_mv[0][1]*(time_pb - time_pp)/time_pp);
147 int fxy= (fx&mask) + ((fy&mask)<<(qpel+1));
148 int bxy= (bx&mask) + ((by&mask)<<(qpel+1));
115329f1 149
2750b827
MN		 150 if(qpel){
151 c->qpel_put[1][fxy](c->temp , ref[0] + (fx>>2) + (fy>>2)*stride , stride);
152 c->qpel_put[1][fxy](c->temp + 8 , ref[0] + (fx>>2) + (fy>>2)*stride + 8 , stride);
153 c->qpel_put[1][fxy](c->temp + 8*stride, ref[0] + (fx>>2) + (fy>>2)*stride + 8*stride, stride);
154 c->qpel_put[1][fxy](c->temp + 8 + 8*stride, ref[0] + (fx>>2) + (fy>>2)*stride + 8 + 8*stride, stride);
155 c->qpel_avg[1][bxy](c->temp , ref[8] + (bx>>2) + (by>>2)*stride , stride);
156 c->qpel_avg[1][bxy](c->temp + 8 , ref[8] + (bx>>2) + (by>>2)*stride + 8 , stride);
157 c->qpel_avg[1][bxy](c->temp + 8*stride, ref[8] + (bx>>2) + (by>>2)*stride + 8*stride, stride);
158 c->qpel_avg[1][bxy](c->temp + 8 + 8*stride, ref[8] + (bx>>2) + (by>>2)*stride + 8 + 8*stride, stride);
115329f1 159 }else{
2750b827
MN		 160 assert((fx>>1) + 16*s->mb_x >= -16);
161 assert((fy>>1) + 16*s->mb_y >= -16);
162 assert((fx>>1) + 16*s->mb_x <= s->width);
163 assert((fy>>1) + 16*s->mb_y <= s->height);
164 assert((bx>>1) + 16*s->mb_x >= -16);
165 assert((by>>1) + 16*s->mb_y >= -16);
166 assert((bx>>1) + 16*s->mb_x <= s->width);
167 assert((by>>1) + 16*s->mb_y <= s->height);
168
169 c->hpel_put[0][fxy](c->temp, ref[0] + (fx>>1) + (fy>>1)*stride, stride, 16);
170 c->hpel_avg[0][bxy](c->temp, ref[8] + (bx>>1) + (by>>1)*stride, stride, 16);
171 }
172 }
173 d = cmp_func(s, c->temp, src[0], stride, 16);
174 }else
175 d= 256*256*256*32;
919e7497
MN		 176 return d;
177}
178
179static av_always_inline int cmp_inline(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
180 const int size, const int h, int ref_index, int src_index,
181 me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, int qpel, int chroma){
182 MotionEstContext * const c= &s->me;
183 const int stride= c->stride;
184 const int uvstride= c->uvstride;
185 const int dxy= subx + (suby<<(1+qpel)); //FIXME log2_subpel?
186 const int hx= subx + (x<<(1+qpel));
187 const int hy= suby + (y<<(1+qpel));
188 uint8_t * const * const ref= c->ref[ref_index];
189 uint8_t * const * const src= c->src[src_index];
190 int d;
191 //FIXME check chroma 4mv, (no crashes ...)
bf4e3bd2 192 int uvdxy; /* no, it might not be used uninitialized */
2750b827
MN		 193 if(dxy){
194 if(qpel){
195 c->qpel_put[size][dxy](c->temp, ref[0] + x + y*stride, stride); //FIXME prototype (add h)
196 if(chroma){
197 int cx= hx/2;
198 int cy= hy/2;
199 cx= (cx>>1)|(cx&1);
200 cy= (cy>>1)|(cy&1);
201 uvdxy= (cx&1) + 2*(cy&1);
202 //FIXME x/y wrong, but mpeg4 qpel is sick anyway, we should drop as much of it as possible in favor for h264
203 }
204 }else{
205 c->hpel_put[size][dxy](c->temp, ref[0] + x + y*stride, stride, h);
206 if(chroma)
207 uvdxy= dxy | (x&1) | (2*(y&1));
208 }
115329f1 209 d = cmp_func(s, c->temp, src[0], stride, h);
2750b827 210 }else{
115329f1 211 d = cmp_func(s, src[0], ref[0] + x + y*stride, stride, h);
2750b827
MN		 212 if(chroma)
213 uvdxy= (x&1) + 2*(y&1);
214 }
215 if(chroma){
216 uint8_t * const uvtemp= c->temp + 16*stride;
217 c->hpel_put[size+1][uvdxy](uvtemp , ref[1] + (x>>1) + (y>>1)*uvstride, uvstride, h>>1);
218 c->hpel_put[size+1][uvdxy](uvtemp+8, ref[2] + (x>>1) + (y>>1)*uvstride, uvstride, h>>1);
115329f1
DB		 219 d += chroma_cmp_func(s, uvtemp , src[1], uvstride, h>>1);
220 d += chroma_cmp_func(s, uvtemp+8, src[2], uvstride, h>>1);
2750b827 221 }
919e7497
MN		 222 return d;
223}
224
225static int cmp_simple(MpegEncContext *s, const int x, const int y,
226 int ref_index, int src_index,
227 me_cmp_func cmp_func, me_cmp_func chroma_cmp_func){
228 return cmp_inline(s,x,y,0,0,0,16,ref_index,src_index, cmp_func, chroma_cmp_func, 0, 0);
229}
230
231static int cmp_fpel_internal(MpegEncContext *s, const int x, const int y,
232 const int size, const int h, int ref_index, int src_index,
233 me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){
234 if(flags&FLAG_DIRECT){
235 return cmp_direct_inline(s,x,y,0,0,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, flags&FLAG_QPEL);
236 }else{
237 return cmp_inline(s,x,y,0,0,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, 0, flags&FLAG_CHROMA);
2750b827 238 }
919e7497
MN		 239}
240
241static int cmp_internal(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
242 const int size, const int h, int ref_index, int src_index,
243 me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){
244 if(flags&FLAG_DIRECT){
245 return cmp_direct_inline(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, flags&FLAG_QPEL);
246 }else{
247 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);
2750b827 248 }
919e7497 249}
2750b827 250
e056d2a2
MN		 251/*! \brief compares a block (either a full macroblock or a partition thereof)
252 against a proposed motion-compensated prediction of that block
253 */
919e7497
MN		 254static av_always_inline int cmp(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
255 const int size, const int h, int ref_index, int src_index,
256 me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){
257 if(av_builtin_constant_p(flags) && av_builtin_constant_p(h) && av_builtin_constant_p(size)
258 && av_builtin_constant_p(subx) && av_builtin_constant_p(suby)
259 && flags==0 && h==16 && size==0 && subx==0 && suby==0){
260 return cmp_simple(s,x,y,ref_index,src_index, cmp_func, chroma_cmp_func);
261 }else if(av_builtin_constant_p(subx) && av_builtin_constant_p(suby)
262 && subx==0 && suby==0){
263 return cmp_fpel_internal(s,x,y,size,h,ref_index,src_index, cmp_func, chroma_cmp_func,flags);
264 }else{
265 return cmp_internal(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, flags);
266 }
267}
268
269static int cmp_hpel(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
270 const int size, const int h, int ref_index, int src_index,
271 me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){
272 if(flags&FLAG_DIRECT){
273 return cmp_direct_inline(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, 0);
274 }else{
275 return cmp_inline(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, 0, flags&FLAG_CHROMA);
276 }
277}
278
279static int cmp_qpel(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
280 const int size, const int h, int ref_index, int src_index,
281 me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){
282 if(flags&FLAG_DIRECT){
283 return cmp_direct_inline(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, 1);
284 }else{
285 return cmp_inline(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, 1, flags&FLAG_CHROMA);
286 }
2750b827 287}
1457ab52
MN		 288
289#include "motion_est_template.c"
1457ab52 290
dbc56b39
MN		 291static int zero_cmp(void *s, uint8_t *a, uint8_t *b, int stride, int h){
292 return 0;
293}
294
295static void zero_hpel(uint8_t *a, const uint8_t *b, int stride, int h){
296}
297
c192426f 298int ff_init_me(MpegEncContext *s){
9964df63 299 MotionEstContext * const c= &s->me;
af6a403a
MN		 300 int cache_size= FFMIN(ME_MAP_SIZE>>ME_MAP_SHIFT, 1<<ME_MAP_SHIFT);
301 int dia_size= FFMAX(FFABS(s->avctx->dia_size)&255, FFABS(s->avctx->pre_dia_size)&255);
540c5220
MN		 302
303 if(FFMIN(s->avctx->dia_size, s->avctx->pre_dia_size) < -ME_MAP_SIZE){
304 av_log(s->avctx, AV_LOG_ERROR, "ME_MAP size is too small for SAB diamond\n");
305 return -1;
306 }
6f55b11e
MN		 307 //special case of snow is needed because snow uses its own iterative ME code
308 if(s->me_method!=ME_ZERO && s->me_method!=ME_EPZS && s->me_method!=ME_X1 && s->avctx->codec_id != CODEC_ID_SNOW){
b74ec693
MN		 309 av_log(s->avctx, AV_LOG_ERROR, "me_method is only allowed to be set to zero and epzs; for hex,umh,full and others see dia_size\n");
310 return -1;
311 }
540c5220 312
95cefcb4 313 c->avctx= s->avctx;
af6a403a
MN		 314
315 if(cache_size < 2*dia_size && !c->stride){
316 av_log(s->avctx, AV_LOG_INFO, "ME_MAP size may be a little small for the selected diamond size\n");
317 }
9964df63 318
1f202b0d
MN		 319 ff_set_cmp(&s->dsp, s->dsp.me_pre_cmp, c->avctx->me_pre_cmp);
320 ff_set_cmp(&s->dsp, s->dsp.me_cmp, c->avctx->me_cmp);
321 ff_set_cmp(&s->dsp, s->dsp.me_sub_cmp, c->avctx->me_sub_cmp);
322 ff_set_cmp(&s->dsp, s->dsp.mb_cmp, c->avctx->mb_cmp);
115329f1 323
1f202b0d
MN		 324 c->flags = get_flags(c, 0, c->avctx->me_cmp &FF_CMP_CHROMA);
325 c->sub_flags= get_flags(c, 0, c->avctx->me_sub_cmp&FF_CMP_CHROMA);
326 c->mb_flags = get_flags(c, 0, c->avctx->mb_cmp &FF_CMP_CHROMA);
de6d9b64 327
9964df63 328/*FIXME s->no_rounding b_type*/
1457ab52 329 if(s->flags&CODEC_FLAG_QPEL){
af4091f1 330 c->sub_motion_search= qpel_motion_search;
9964df63
MN		 331 c->qpel_avg= s->dsp.avg_qpel_pixels_tab;
332 if(s->no_rounding) c->qpel_put= s->dsp.put_no_rnd_qpel_pixels_tab;
333 else c->qpel_put= s->dsp.put_qpel_pixels_tab;
1457ab52 334 }else{
1f202b0d 335 if(c->avctx->me_sub_cmp&FF_CMP_CHROMA)
af4091f1 336 c->sub_motion_search= hpel_motion_search;
115329f1
DB		 337 else if( c->avctx->me_sub_cmp == FF_CMP_SAD
338 && c->avctx-> me_cmp == FF_CMP_SAD
1f202b0d 339 && c->avctx-> mb_cmp == FF_CMP_SAD)
af4091f1 340 c->sub_motion_search= sad_hpel_motion_search; // 2050 vs. 2450 cycles
1457ab52 341 else
af4091f1 342 c->sub_motion_search= hpel_motion_search;
826f429a 343 }
dbc56b39
MN		 344 c->hpel_avg= s->dsp.avg_pixels_tab;
345 if(s->no_rounding) c->hpel_put= s->dsp.put_no_rnd_pixels_tab;
346 else c->hpel_put= s->dsp.put_pixels_tab;
347
2750b827 348 if(s->linesize){
115329f1 349 c->stride = s->linesize;
af4091f1 350 c->uvstride= s->uvlinesize;
67725183 351 }else{
af4091f1
MN		 352 c->stride = 16*s->mb_width + 32;
353 c->uvstride= 8*s->mb_width + 16;
67725183 354 }
9964df63 355
755bfeab
DB		 356 /* 8x8 fullpel search would need a 4x4 chroma compare, which we do
357 * not have yet, and even if we had, the motion estimation code
358 * does not expect it. */
155ec6ed 359 if(s->codec_id != CODEC_ID_SNOW){
725de25c 360 if((c->avctx->me_cmp&FF_CMP_CHROMA)/* && !s->dsp.me_cmp[2]*/){
155ec6ed
MN		 361 s->dsp.me_cmp[2]= zero_cmp;
362 }
363 if((c->avctx->me_sub_cmp&FF_CMP_CHROMA) && !s->dsp.me_sub_cmp[2]){
364 s->dsp.me_sub_cmp[2]= zero_cmp;
365 }
366 c->hpel_put[2][0]= c->hpel_put[2][1]=
367 c->hpel_put[2][2]= c->hpel_put[2][3]= zero_hpel;
dbc56b39 368 }
dbc56b39 369
1c3990db
MN		 370 if(s->codec_id == CODEC_ID_H261){
371 c->sub_motion_search= no_sub_motion_search;
372 }
373
c192426f 374 return 0;
1457ab52 375}
115329f1 376
802f454e 377#if 0
0c1a9eda 378static int pix_dev(uint8_t * pix, int line_size, int mean)
b3987771
MN		 379{
380 int s, i, j;
381
382 s = 0;
383 for (i = 0; i < 16; i++) {
bb270c08 384 for (j = 0; j < 16; j += 8) {
c26abfa5
DB		 385 s += FFABS(pix[0]-mean);
386 s += FFABS(pix[1]-mean);
387 s += FFABS(pix[2]-mean);
388 s += FFABS(pix[3]-mean);
389 s += FFABS(pix[4]-mean);
390 s += FFABS(pix[5]-mean);
391 s += FFABS(pix[6]-mean);
392 s += FFABS(pix[7]-mean);
bb270c08
DB		 393 pix += 8;
394 }
395 pix += line_size - 16;
b3987771
MN		 396 }
397 return s;
398}
802f454e 399#endif
b3987771 400
eb4b3dd3 401static inline void no_motion_search(MpegEncContext * s,
bb270c08 402 int *mx_ptr, int *my_ptr)
de6d9b64
FB		 403{
404 *mx_ptr = 16 * s->mb_x;
405 *my_ptr = 16 * s->mb_y;
406}
407
45870f57
MN		 408#define Z_THRESHOLD 256
409
1457ab52 410#define CHECK_SAD_HALF_MV(suffix, x, y) \
0d21a846 411{\
bb198e19 412 d= s->dsp.pix_abs[size][(x?1:0)+(y?2:0)](NULL, pix, ptr+((x)>>1), stride, h);\
1457ab52 413 d += (mv_penalty[pen_x + x] + mv_penalty[pen_y + y])*penalty_factor;\
0d21a846
MN		 414 COPY3_IF_LT(dminh, d, dx, x, dy, y)\
415}
45870f57 416
1457ab52 417static inline int sad_hpel_motion_search(MpegEncContext * s,
bb270c08 418 int *mx_ptr, int *my_ptr, int dmin,
2750b827
MN		 419 int src_index, int ref_index,
420 int size, int h)
de6d9b64 421{
af4091f1
MN		 422 MotionEstContext * const c= &s->me;
423 const int penalty_factor= c->sub_penalty_factor;
bb198e19 424 int mx, my, dminh;
0c1a9eda 425 uint8_t *pix, *ptr;
af4091f1
MN		 426 int stride= c->stride;
427 const int flags= c->sub_flags;
2750b827 428 LOAD_COMMON
115329f1 429
2750b827 430 assert(flags == 0);
de6d9b64 431
af4091f1 432 if(c->skip){
0d21a846
MN		 433// printf("S");
434 *mx_ptr = 0;
435 *my_ptr = 0;
436 return dmin;
437 }
438// printf("N");
115329f1 439
af4091f1 440 pix = c->src[src_index][0];
0d21a846 441
ba6802de
MN		 442 mx = *mx_ptr;
443 my = *my_ptr;
af4091f1 444 ptr = c->ref[ref_index][0] + (my * stride) + mx;
115329f1 445
ba6802de
MN		 446 dminh = dmin;
447
115329f1 448 if (mx > xmin && mx < xmax &&
ba6802de 449 my > ymin && my < ymax) {
0d21a846 450 int dx=0, dy=0;
115329f1 451 int d, pen_x, pen_y;
0d21a846
MN		 452 const int index= (my<<ME_MAP_SHIFT) + mx;
453 const int t= score_map[(index-(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
454 const int l= score_map[(index- 1 )&(ME_MAP_SIZE-1)];
455 const int r= score_map[(index+ 1 )&(ME_MAP_SIZE-1)];
456 const int b= score_map[(index+(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
457 mx<<=1;
458 my<<=1;
ba6802de 459
115329f1 460
ba6802de
MN		 461 pen_x= pred_x + mx;
462 pen_y= pred_y + my;
463
bb198e19 464 ptr-= stride;
0d21a846 465 if(t<=b){
1457ab52 466 CHECK_SAD_HALF_MV(y2 , 0, -1)
0d21a846 467 if(l<=r){
1457ab52 468 CHECK_SAD_HALF_MV(xy2, -1, -1)
0d21a846 469 if(t+r<=b+l){
1457ab52 470 CHECK_SAD_HALF_MV(xy2, +1, -1)
bb198e19 471 ptr+= stride;
0d21a846 472 }else{
bb198e19 473 ptr+= stride;
1457ab52 474 CHECK_SAD_HALF_MV(xy2, -1, +1)
0d21a846 475 }
1457ab52 476 CHECK_SAD_HALF_MV(x2 , -1, 0)
0d21a846 477 }else{
1457ab52 478 CHECK_SAD_HALF_MV(xy2, +1, -1)
0d21a846 479 if(t+l<=b+r){
1457ab52 480 CHECK_SAD_HALF_MV(xy2, -1, -1)
bb198e19 481 ptr+= stride;
0d21a846 482 }else{
bb198e19 483 ptr+= stride;
1457ab52 484 CHECK_SAD_HALF_MV(xy2, +1, +1)
0d21a846 485 }
1457ab52 486 CHECK_SAD_HALF_MV(x2 , +1, 0)
0d21a846
MN		 487 }
488 }else{
489 if(l<=r){
490 if(t+l<=b+r){
1457ab52 491 CHECK_SAD_HALF_MV(xy2, -1, -1)
bb198e19 492 ptr+= stride;
0d21a846 493 }else{
bb198e19 494 ptr+= stride;
1457ab52 495 CHECK_SAD_HALF_MV(xy2, +1, +1)
0d21a846 496 }
1457ab52
MN		 497 CHECK_SAD_HALF_MV(x2 , -1, 0)
498 CHECK_SAD_HALF_MV(xy2, -1, +1)
0d21a846
MN		 499 }else{
500 if(t+r<=b+l){
1457ab52 501 CHECK_SAD_HALF_MV(xy2, +1, -1)
bb198e19 502 ptr+= stride;
0d21a846 503 }else{
bb198e19 504 ptr+= stride;
1457ab52 505 CHECK_SAD_HALF_MV(xy2, -1, +1)
0d21a846 506 }
1457ab52
MN		 507 CHECK_SAD_HALF_MV(x2 , +1, 0)
508 CHECK_SAD_HALF_MV(xy2, +1, +1)
0d21a846 509 }
1457ab52 510 CHECK_SAD_HALF_MV(y2 , 0, +1)
0d21a846
MN		 511 }
512 mx+=dx;
513 my+=dy;
ba6802de
MN		 514
515 }else{
0d21a846
MN		 516 mx<<=1;
517 my<<=1;
ba6802de
MN		 518 }
519
520 *mx_ptr = mx;
521 *my_ptr = my;
0d21a846 522 return dminh;
ba6802de
MN		 523}
524
0d21a846 525static inline void set_p_mv_tables(MpegEncContext * s, int mx, int my, int mv4)
ba6802de 526{
7bc9090a 527 const int xy= s->mb_x + s->mb_y*s->mb_stride;
115329f1 528
9dbcbd92
MN		 529 s->p_mv_table[xy][0] = mx;
530 s->p_mv_table[xy][1] = my;
ba6802de 531
bb628dae 532 /* has already been set to the 4 MV if 4MV is done */
0d21a846 533 if(mv4){
ba6802de
MN		 534 int mot_xy= s->block_index[0];
535
8d7ec294
WH		 536 s->current_picture.motion_val[0][mot_xy ][0]= mx;
537 s->current_picture.motion_val[0][mot_xy ][1]= my;
538 s->current_picture.motion_val[0][mot_xy+1][0]= mx;
539 s->current_picture.motion_val[0][mot_xy+1][1]= my;
ba6802de 540
137c8468 541 mot_xy += s->b8_stride;
8d7ec294
WH		 542 s->current_picture.motion_val[0][mot_xy ][0]= mx;
543 s->current_picture.motion_val[0][mot_xy ][1]= my;
544 s->current_picture.motion_val[0][mot_xy+1][0]= mx;
545 s->current_picture.motion_val[0][mot_xy+1][1]= my;
ba6802de
MN		 546 }
547}
548
ebbcdc9a
MN		 549/**
550 * get fullpel ME search limits.
ebbcdc9a 551 */
bb198e19 552static inline void get_limits(MpegEncContext *s, int x, int y)
de6d9b64 553{
af4091f1 554 MotionEstContext * const c= &s->me;
a6daaf7c 555 int range= c->avctx->me_range >> (1 + !!(c->flags&FLAG_QPEL));
bb198e19 556/*
1f202b0d 557 if(c->avctx->me_range) c->range= c->avctx->me_range >> 1;
af4091f1 558 else c->range= 16;
bb198e19 559*/
de6d9b64 560 if (s->unrestricted_mv) {
af4091f1
MN		 561 c->xmin = - x - 16;
562 c->ymin = - y - 16;
563 c->xmax = - x + s->mb_width *16;
564 c->ymax = - y + s->mb_height*16;
1c3990db
MN		 565 } else if (s->out_format == FMT_H261){
566 // Search range of H261 is different from other codec standards
567 c->xmin = (x > 15) ? - 15 : 0;
568 c->ymin = (y > 15) ? - 15 : 0;
115329f1 569 c->xmax = (x < s->mb_width * 16 - 16) ? 15 : 0;
1c3990db 570 c->ymax = (y < s->mb_height * 16 - 16) ? 15 : 0;
de6d9b64 571 } else {
af4091f1
MN		 572 c->xmin = - x;
573 c->ymin = - y;
574 c->xmax = - x + s->mb_width *16 - 16;
575 c->ymax = - y + s->mb_height*16 - 16;
de6d9b64 576 }
a6daaf7c
MN		 577 if(range){
578 c->xmin = FFMAX(c->xmin,-range);
579 c->xmax = FFMIN(c->xmax, range);
580 c->ymin = FFMAX(c->ymin,-range);
581 c->ymax = FFMIN(c->ymax, range);
582 }
9dbcbd92
MN		 583}
584
1f202b0d
MN		 585static inline void init_mv4_ref(MotionEstContext *c){
586 const int stride= c->stride;
9964df63
MN		 587
588 c->ref[1][0] = c->ref[0][0] + 8;
589 c->ref[2][0] = c->ref[0][0] + 8*stride;
590 c->ref[3][0] = c->ref[2][0] + 8;
591 c->src[1][0] = c->src[0][0] + 8;
592 c->src[2][0] = c->src[0][0] + 8*stride;
593 c->src[3][0] = c->src[2][0] + 8;
594}
595
bb198e19 596static inline int h263_mv4_search(MpegEncContext *s, int mx, int my, int shift)
0d21a846 597{
2750b827 598 MotionEstContext * const c= &s->me;
bb198e19
MN		 599 const int size= 1;
600 const int h=8;
0d21a846
MN		 601 int block;
602 int P[10][2];
67725183 603 int dmin_sum=0, mx4_sum=0, my4_sum=0;
ca7d05d5 604 int same=1;
1f202b0d 605 const int stride= c->stride;
af4091f1 606 uint8_t *mv_penalty= c->current_mv_penalty;
2750b827 607
1f202b0d 608 init_mv4_ref(c);
115329f1 609
0d21a846
MN		 610 for(block=0; block<4; block++){
611 int mx4, my4;
612 int pred_x4, pred_y4;
613 int dmin4;
614 static const int off[4]= {2, 1, 1, -1};
137c8468 615 const int mot_stride = s->b8_stride;
0d21a846 616 const int mot_xy = s->block_index[block];
bb198e19 617
8d7ec294
WH		 618 P_LEFT[0] = s->current_picture.motion_val[0][mot_xy - 1][0];
619 P_LEFT[1] = s->current_picture.motion_val[0][mot_xy - 1][1];
0d21a846 620
af4091f1 621 if(P_LEFT[0] > (c->xmax<<shift)) P_LEFT[0] = (c->xmax<<shift);
0d21a846
MN		 622
623 /* special case for first line */
9c3d33d6 624 if (s->first_slice_line && block<2) {
af4091f1
MN		 625 c->pred_x= pred_x4= P_LEFT[0];
626 c->pred_y= pred_y4= P_LEFT[1];
0d21a846 627 } else {
8d7ec294
WH		 628 P_TOP[0] = s->current_picture.motion_val[0][mot_xy - mot_stride ][0];
629 P_TOP[1] = s->current_picture.motion_val[0][mot_xy - mot_stride ][1];
630 P_TOPRIGHT[0] = s->current_picture.motion_val[0][mot_xy - mot_stride + off[block]][0];
631 P_TOPRIGHT[1] = s->current_picture.motion_val[0][mot_xy - mot_stride + off[block]][1];
af4091f1
MN		 632 if(P_TOP[1] > (c->ymax<<shift)) P_TOP[1] = (c->ymax<<shift);
633 if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
634 if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift);
635 if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
115329f1 636
0d21a846
MN		 637 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
638 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
639
af4091f1
MN		 640 c->pred_x= pred_x4 = P_MEDIAN[0];
641 c->pred_y= pred_y4 = P_MEDIAN[1];
0d21a846
MN		 642 }
643 P_MV1[0]= mx;
644 P_MV1[1]= my;
645
2750b827 646 dmin4 = epzs_motion_search4(s, &mx4, &my4, P, block, block, s->p_mv_table, (1<<16)>>shift);
0d21a846 647
af4091f1 648 dmin4= c->sub_motion_search(s, &mx4, &my4, dmin4, block, block, size, h);
115329f1 649
2750b827 650 if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
67725183 651 int dxy;
bb198e19 652 const int offset= ((block&1) + (block>>1)*stride)*8;
af4091f1 653 uint8_t *dest_y = c->scratchpad + offset;
67725183 654 if(s->quarter_sample){
2750b827 655 uint8_t *ref= c->ref[block][0] + (mx4>>2) + (my4>>2)*stride;
67725183
MN		 656 dxy = ((my4 & 3) << 2) | (mx4 & 3);
657
658 if(s->no_rounding)
9c3d33d6 659 s->dsp.put_no_rnd_qpel_pixels_tab[1][dxy](dest_y , ref , stride);
67725183 660 else
bb198e19 661 s->dsp.put_qpel_pixels_tab [1][dxy](dest_y , ref , stride);
67725183 662 }else{
2750b827 663 uint8_t *ref= c->ref[block][0] + (mx4>>1) + (my4>>1)*stride;
67725183
MN		 664 dxy = ((my4 & 1) << 1) | (mx4 & 1);
665
666 if(s->no_rounding)
bb198e19 667 s->dsp.put_no_rnd_pixels_tab[1][dxy](dest_y , ref , stride, h);
67725183 668 else
bb198e19 669 s->dsp.put_pixels_tab [1][dxy](dest_y , ref , stride, h);
67725183 670 }
af4091f1 671 dmin_sum+= (mv_penalty[mx4-pred_x4] + mv_penalty[my4-pred_y4])*c->mb_penalty_factor;
67725183
MN		 672 }else
673 dmin_sum+= dmin4;
674
675 if(s->quarter_sample){
676 mx4_sum+= mx4/2;
677 my4_sum+= my4/2;
678 }else{
679 mx4_sum+= mx4;
680 my4_sum+= my4;
681 }
115329f1 682
8d7ec294
WH		 683 s->current_picture.motion_val[0][ s->block_index[block] ][0]= mx4;
684 s->current_picture.motion_val[0][ s->block_index[block] ][1]= my4;
ca7d05d5
MN		 685
686 if(mx4 != mx || my4 != my) same=0;
0d21a846 687 }
115329f1 688
ca7d05d5
MN		 689 if(same)
690 return INT_MAX;
115329f1 691
cb009a8c 692 if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
af4091f1 693 dmin_sum += s->dsp.mb_cmp[0](s, s->new_picture.data[0] + s->mb_x*16 + s->mb_y*16*stride, c->scratchpad, stride, 16);
67725183 694 }
115329f1 695
1f202b0d 696 if(c->avctx->mb_cmp&FF_CMP_CHROMA){
67725183
MN		 697 int dxy;
698 int mx, my;
699 int offset;
700
701 mx= ff_h263_round_chroma(mx4_sum);
702 my= ff_h263_round_chroma(my4_sum);
703 dxy = ((my & 1) << 1) | (mx & 1);
115329f1 704
67725183 705 offset= (s->mb_x*8 + (mx>>1)) + (s->mb_y*8 + (my>>1))*s->uvlinesize;
115329f1 706
67725183 707 if(s->no_rounding){
af4091f1
MN		 708 s->dsp.put_no_rnd_pixels_tab[1][dxy](c->scratchpad , s->last_picture.data[1] + offset, s->uvlinesize, 8);
709 s->dsp.put_no_rnd_pixels_tab[1][dxy](c->scratchpad+8 , s->last_picture.data[2] + offset, s->uvlinesize, 8);
67725183 710 }else{
af4091f1
MN		 711 s->dsp.put_pixels_tab [1][dxy](c->scratchpad , s->last_picture.data[1] + offset, s->uvlinesize, 8);
712 s->dsp.put_pixels_tab [1][dxy](c->scratchpad+8 , s->last_picture.data[2] + offset, s->uvlinesize, 8);
67725183
MN		 713 }
714
af4091f1
MN		 715 dmin_sum += s->dsp.mb_cmp[1](s, s->new_picture.data[1] + s->mb_x*8 + s->mb_y*8*s->uvlinesize, c->scratchpad , s->uvlinesize, 8);
716 dmin_sum += s->dsp.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);
67725183 717 }
115329f1 718
af4091f1
MN		 719 c->pred_x= mx;
720 c->pred_y= my;
67725183 721
1f202b0d 722 switch(c->avctx->mb_cmp&0xFF){
67725183
MN		 723 /*case FF_CMP_SSE:
724 return dmin_sum+ 32*s->qscale*s->qscale;*/
725 case FF_CMP_RD:
726 return dmin_sum;
727 default:
af4091f1 728 return dmin_sum+ 11*c->mb_penalty_factor;
67725183 729 }
0d21a846
MN		 730}
731
9964df63
MN		 732static inline void init_interlaced_ref(MpegEncContext *s, int ref_index){
733 MotionEstContext * const c= &s->me;
734
735 c->ref[1+ref_index][0] = c->ref[0+ref_index][0] + s->linesize;
736 c->src[1][0] = c->src[0][0] + s->linesize;
737 if(c->flags & FLAG_CHROMA){
738 c->ref[1+ref_index][1] = c->ref[0+ref_index][1] + s->uvlinesize;
739 c->ref[1+ref_index][2] = c->ref[0+ref_index][2] + s->uvlinesize;
740 c->src[1][1] = c->src[0][1] + s->uvlinesize;
741 c->src[1][2] = c->src[0][2] + s->uvlinesize;
742 }
743}
744
115329f1 745static int interlaced_search(MpegEncContext *s, int ref_index,
f20f8a8b 746 int16_t (*mv_tables[2][2])[2], uint8_t *field_select_tables[2], int mx, int my, int user_field_select)
bb198e19 747{
2750b827 748 MotionEstContext * const c= &s->me;
bb198e19
MN		 749 const int size=0;
750 const int h=8;
751 int block;
752 int P[10][2];
2750b827 753 uint8_t * const mv_penalty= c->current_mv_penalty;
bb198e19
MN		 754 int same=1;
755 const int stride= 2*s->linesize;
bb198e19
MN		 756 int dmin_sum= 0;
757 const int mot_stride= s->mb_stride;
758 const int xy= s->mb_x + s->mb_y*mot_stride;
115329f1 759
2750b827
MN		 760 c->ymin>>=1;
761 c->ymax>>=1;
762 c->stride<<=1;
763 c->uvstride<<=1;
9964df63 764 init_interlaced_ref(s, ref_index);
115329f1 765
bb198e19
MN		 766 for(block=0; block<2; block++){
767 int field_select;
768 int best_dmin= INT_MAX;
769 int best_field= -1;
770
bb198e19 771 for(field_select=0; field_select<2; field_select++){
2750b827 772 int dmin, mx_i, my_i;
bb198e19 773 int16_t (*mv_table)[2]= mv_tables[block][field_select];
115329f1 774
f20f8a8b 775 if(user_field_select){
8437ff75
MN		 776 assert(field_select==0 || field_select==1);
777 assert(field_select_tables[block][xy]==0 || field_select_tables[block][xy]==1);
f20f8a8b
MN		 778 if(field_select_tables[block][xy] != field_select)
779 continue;
780 }
115329f1 781
bb198e19
MN		 782 P_LEFT[0] = mv_table[xy - 1][0];
783 P_LEFT[1] = mv_table[xy - 1][1];
2750b827 784 if(P_LEFT[0] > (c->xmax<<1)) P_LEFT[0] = (c->xmax<<1);
115329f1 785
af4091f1
MN		 786 c->pred_x= P_LEFT[0];
787 c->pred_y= P_LEFT[1];
115329f1 788
9c3d33d6 789 if(!s->first_slice_line){
bb198e19
MN		 790 P_TOP[0] = mv_table[xy - mot_stride][0];
791 P_TOP[1] = mv_table[xy - mot_stride][1];
792 P_TOPRIGHT[0] = mv_table[xy - mot_stride + 1][0];
793 P_TOPRIGHT[1] = mv_table[xy - mot_stride + 1][1];
2750b827
MN		 794 if(P_TOP[1] > (c->ymax<<1)) P_TOP[1] = (c->ymax<<1);
795 if(P_TOPRIGHT[0] < (c->xmin<<1)) P_TOPRIGHT[0]= (c->xmin<<1);
796 if(P_TOPRIGHT[0] > (c->xmax<<1)) P_TOPRIGHT[0]= (c->xmax<<1);
797 if(P_TOPRIGHT[1] > (c->ymax<<1)) P_TOPRIGHT[1]= (c->ymax<<1);
115329f1 798
bb198e19
MN		 799 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
800 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
801 }
802 P_MV1[0]= mx; //FIXME not correct if block != field_select
803 P_MV1[1]= my / 2;
115329f1 804
2750b827 805 dmin = epzs_motion_search2(s, &mx_i, &my_i, P, block, field_select+ref_index, mv_table, (1<<16)>>1);
bb198e19 806
2750b827 807 dmin= c->sub_motion_search(s, &mx_i, &my_i, dmin, block, field_select+ref_index, size, h);
115329f1 808
bb198e19
MN		 809 mv_table[xy][0]= mx_i;
810 mv_table[xy][1]= my_i;
115329f1 811
2750b827 812 if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
bb198e19
MN		 813 int dxy;
814
815 //FIXME chroma ME
2750b827 816 uint8_t *ref= c->ref[field_select+ref_index][0] + (mx_i>>1) + (my_i>>1)*stride;
bb198e19
MN		 817 dxy = ((my_i & 1) << 1) | (mx_i & 1);
818
819 if(s->no_rounding){
2750b827 820 s->dsp.put_no_rnd_pixels_tab[size][dxy](c->scratchpad, ref , stride, h);
bb198e19 821 }else{
2750b827 822 s->dsp.put_pixels_tab [size][dxy](c->scratchpad, ref , stride, h);
bb198e19 823 }
2750b827 824 dmin= s->dsp.mb_cmp[size](s, c->src[block][0], c->scratchpad, stride, h);
af4091f1 825 dmin+= (mv_penalty[mx_i-c->pred_x] + mv_penalty[my_i-c->pred_y] + 1)*c->mb_penalty_factor;
bb198e19 826 }else
2750b827 827 dmin+= c->mb_penalty_factor; //field_select bits
115329f1 828
bb198e19 829 dmin += field_select != block; //slightly prefer same field
115329f1 830
bb198e19
MN		 831 if(dmin < best_dmin){
832 best_dmin= dmin;
833 best_field= field_select;
834 }
835 }
836 {
837 int16_t (*mv_table)[2]= mv_tables[block][best_field];
838
839 if(mv_table[xy][0] != mx) same=0; //FIXME check if these checks work and are any good at all
840 if(mv_table[xy][1]&1) same=0;
115329f1 841 if(mv_table[xy][1]*2 != my) same=0;
bb198e19
MN		 842 if(best_field != block) same=0;
843 }
844
845 field_select_tables[block][xy]= best_field;
846 dmin_sum += best_dmin;
847 }
115329f1 848
2750b827
MN		 849 c->ymin<<=1;
850 c->ymax<<=1;
851 c->stride>>=1;
852 c->uvstride>>=1;
bb198e19
MN		 853
854 if(same)
855 return INT_MAX;
115329f1 856
1f202b0d 857 switch(c->avctx->mb_cmp&0xFF){
bb198e19
MN		 858 /*case FF_CMP_SSE:
859 return dmin_sum+ 32*s->qscale*s->qscale;*/
860 case FF_CMP_RD:
861 return dmin_sum;
862 default:
2750b827 863 return dmin_sum+ 11*c->mb_penalty_factor;
bb198e19
MN		 864 }
865}
866
ae1dbde1
MN		 867static void clip_input_mv(MpegEncContext * s, int16_t *mv, int interlaced){
868 int ymax= s->me.ymax>>interlaced;
869 int ymin= s->me.ymin>>interlaced;
115329f1 870
ae1dbde1
MN		 871 if(mv[0] < s->me.xmin) mv[0] = s->me.xmin;
872 if(mv[0] > s->me.xmax) mv[0] = s->me.xmax;
873 if(mv[1] < ymin) mv[1] = ymin;
874 if(mv[1] > ymax) mv[1] = ymax;
875}
876
f4f3223f
MN		 877static inline int check_input_motion(MpegEncContext * s, int mb_x, int mb_y, int p_type){
878 MotionEstContext * const c= &s->me;
879 Picture *p= s->current_picture_ptr;
880 int mb_xy= mb_x + mb_y*s->mb_stride;
881 int xy= 2*mb_x + 2*mb_y*s->b8_stride;
882 int mb_type= s->current_picture.mb_type[mb_xy];
883 int flags= c->flags;
884 int shift= (flags&FLAG_QPEL) + 1;
885 int mask= (1<<shift)-1;
9964df63 886 int x, y, i;
f4f3223f
MN		 887 int d=0;
888 me_cmp_func cmpf= s->dsp.sse[0];
889 me_cmp_func chroma_cmpf= s->dsp.sse[1];
115329f1 890
ae1dbde1
MN		 891 if(p_type && USES_LIST(mb_type, 1)){
892 av_log(c->avctx, AV_LOG_ERROR, "backward motion vector in P frame\n");
cec1f05f 893 return INT_MAX/2;
ae1dbde1 894 }
f4f3223f 895 assert(IS_INTRA(mb_type) || USES_LIST(mb_type,0) || USES_LIST(mb_type,1));
115329f1 896
ae1dbde1
MN		 897 for(i=0; i<4; i++){
898 int xy= s->block_index[i];
899 clip_input_mv(s, p->motion_val[0][xy], !!IS_INTERLACED(mb_type));
900 clip_input_mv(s, p->motion_val[1][xy], !!IS_INTERLACED(mb_type));
901 }
902
f4f3223f
MN		 903 if(IS_INTERLACED(mb_type)){
904 int xy2= xy + s->b8_stride;
905 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTRA;
906 c->stride<<=1;
907 c->uvstride<<=1;
115329f1 908
1545acea 909 if(!(s->flags & CODEC_FLAG_INTERLACED_ME)){
1f202b0d 910 av_log(c->avctx, AV_LOG_ERROR, "Interlaced macroblock selected but interlaced motion estimation disabled\n");
cec1f05f 911 return INT_MAX/2;
1545acea 912 }
9964df63 913
f4f3223f 914 if(USES_LIST(mb_type, 0)){
fc4a2d1e
MN		 915 int field_select0= p->ref_index[0][4*mb_xy ];
916 int field_select1= p->ref_index[0][4*mb_xy+2];
f4f3223f
MN		 917 assert(field_select0==0 ||field_select0==1);
918 assert(field_select1==0 ||field_select1==1);
f20f8a8b
MN		 919 init_interlaced_ref(s, 0);
920
f4f3223f
MN		 921 if(p_type){
922 s->p_field_select_table[0][mb_xy]= field_select0;
923 s->p_field_select_table[1][mb_xy]= field_select1;
924 *(uint32_t*)s->p_field_mv_table[0][field_select0][mb_xy]= *(uint32_t*)p->motion_val[0][xy ];
925 *(uint32_t*)s->p_field_mv_table[1][field_select1][mb_xy]= *(uint32_t*)p->motion_val[0][xy2];
926 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTER_I;
927 }else{
928 s->b_field_select_table[0][0][mb_xy]= field_select0;
929 s->b_field_select_table[0][1][mb_xy]= field_select1;
930 *(uint32_t*)s->b_field_mv_table[0][0][field_select0][mb_xy]= *(uint32_t*)p->motion_val[0][xy ];
931 *(uint32_t*)s->b_field_mv_table[0][1][field_select1][mb_xy]= *(uint32_t*)p->motion_val[0][xy2];
932 s->mb_type[mb_xy]= CANDIDATE_MB_TYPE_FORWARD_I;
933 }
934
115329f1 935 x= p->motion_val[0][xy ][0];
f4f3223f
MN		 936 y= p->motion_val[0][xy ][1];
937 d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select0, 0, cmpf, chroma_cmpf, flags);
115329f1 938 x= p->motion_val[0][xy2][0];
f4f3223f
MN		 939 y= p->motion_val[0][xy2][1];
940 d+= cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select1, 1, cmpf, chroma_cmpf, flags);
941 }
942 if(USES_LIST(mb_type, 1)){
fc4a2d1e
MN		 943 int field_select0= p->ref_index[1][4*mb_xy ];
944 int field_select1= p->ref_index[1][4*mb_xy+2];
f4f3223f
MN		 945 assert(field_select0==0 ||field_select0==1);
946 assert(field_select1==0 ||field_select1==1);
f20f8a8b
MN		 947 init_interlaced_ref(s, 2);
948
f4f3223f
MN		 949 s->b_field_select_table[1][0][mb_xy]= field_select0;
950 s->b_field_select_table[1][1][mb_xy]= field_select1;
951 *(uint32_t*)s->b_field_mv_table[1][0][field_select0][mb_xy]= *(uint32_t*)p->motion_val[1][xy ];
952 *(uint32_t*)s->b_field_mv_table[1][1][field_select1][mb_xy]= *(uint32_t*)p->motion_val[1][xy2];
953 if(USES_LIST(mb_type, 0)){
954 s->mb_type[mb_xy]= CANDIDATE_MB_TYPE_BIDIR_I;
955 }else{
956 s->mb_type[mb_xy]= CANDIDATE_MB_TYPE_BACKWARD_I;
957 }
958
115329f1 959 x= p->motion_val[1][xy ][0];
f4f3223f
MN		 960 y= p->motion_val[1][xy ][1];
961 d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select0+2, 0, cmpf, chroma_cmpf, flags);
115329f1 962 x= p->motion_val[1][xy2][0];
f4f3223f
MN		 963 y= p->motion_val[1][xy2][1];
964 d+= cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select1+2, 1, cmpf, chroma_cmpf, flags);
965 //FIXME bidir scores
966 }
967 c->stride>>=1;
968 c->uvstride>>=1;
9964df63 969 }else if(IS_8X8(mb_type)){
1545acea 970 if(!(s->flags & CODEC_FLAG_4MV)){
1f202b0d 971 av_log(c->avctx, AV_LOG_ERROR, "4MV macroblock selected but 4MV encoding disabled\n");
cec1f05f 972 return INT_MAX/2;
1545acea 973 }
9964df63
MN		 974 cmpf= s->dsp.sse[1];
975 chroma_cmpf= s->dsp.sse[1];
1f202b0d 976 init_mv4_ref(c);
9964df63
MN		 977 for(i=0; i<4; i++){
978 xy= s->block_index[i];
115329f1 979 x= p->motion_val[0][xy][0];
9964df63
MN		 980 y= p->motion_val[0][xy][1];
981 d+= cmp(s, x>>shift, y>>shift, x&mask, y&mask, 1, 8, i, i, cmpf, chroma_cmpf, flags);
982 }
983 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTER4V;
f4f3223f
MN		 984 }else{
985 if(USES_LIST(mb_type, 0)){
986 if(p_type){
987 *(uint32_t*)s->p_mv_table[mb_xy]= *(uint32_t*)p->motion_val[0][xy];
988 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTER;
989 }else if(USES_LIST(mb_type, 1)){
990 *(uint32_t*)s->b_bidir_forw_mv_table[mb_xy]= *(uint32_t*)p->motion_val[0][xy];
991 *(uint32_t*)s->b_bidir_back_mv_table[mb_xy]= *(uint32_t*)p->motion_val[1][xy];
992 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_BIDIR;
993 }else{
994 *(uint32_t*)s->b_forw_mv_table[mb_xy]= *(uint32_t*)p->motion_val[0][xy];
995 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_FORWARD;
996 }
115329f1 997 x= p->motion_val[0][xy][0];
f4f3223f
MN		 998 y= p->motion_val[0][xy][1];
999 d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 16, 0, 0, cmpf, chroma_cmpf, flags);
1000 }else if(USES_LIST(mb_type, 1)){
1001 *(uint32_t*)s->b_back_mv_table[mb_xy]= *(uint32_t*)p->motion_val[1][xy];
1002 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_BACKWARD;
115329f1
DB		 1003
1004 x= p->motion_val[1][xy][0];
f4f3223f
MN		 1005 y= p->motion_val[1][xy][1];
1006 d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 16, 2, 0, cmpf, chroma_cmpf, flags);
1007 }else
1008 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTRA;
1009 }
1010 return d;
1011}
1012
9dbcbd92
MN		 1013void ff_estimate_p_frame_motion(MpegEncContext * s,
1014 int mb_x, int mb_y)
1015{
2750b827 1016 MotionEstContext * const c= &s->me;
0c1a9eda 1017 uint8_t *pix, *ppix;
e828d257
GP		 1018 int sum, mx, my, dmin;
1019 int varc; ///< the variance of the block (sum of squared (p[y][x]-average))
1020 int vard; ///< sum of squared differences with the estimated motion vector
0d21a846 1021 int P[10][2];
9dbcbd92
MN		 1022 const int shift= 1+s->quarter_sample;
1023 int mb_type=0;
1e491e29 1024 Picture * const pic= &s->current_picture;
115329f1 1025
af4091f1 1026 init_ref(c, s->new_picture.data, s->last_picture.data, NULL, 16*mb_x, 16*mb_y, 0);
bb198e19 1027
1457ab52 1028 assert(s->quarter_sample==0 || s->quarter_sample==1);
af4091f1
MN		 1029 assert(s->linesize == c->stride);
1030 assert(s->uvlinesize == c->uvstride);
1457ab52 1031
26efc54e
MN		 1032 c->penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
1033 c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
1034 c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
af4091f1 1035 c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
9dbcbd92 1036
bb198e19 1037 get_limits(s, 16*mb_x, 16*mb_y);
af4091f1 1038 c->skip=0;
9dbcbd92 1039
f20f8a8b
MN		 1040 /* intra / predictive decision */
1041 pix = c->src[0][0];
1042 sum = s->dsp.pix_sum(pix, s->linesize);
aaaa6f15 1043 varc = s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)(sum*sum))>>8) + 500;
f20f8a8b
MN		 1044
1045 pic->mb_mean[s->mb_stride * mb_y + mb_x] = (sum+128)>>8;
aaaa6f15
MN		 1046 pic->mb_var [s->mb_stride * mb_y + mb_x] = (varc+128)>>8;
1047 c->mb_var_sum_temp += (varc+128)>>8;
f20f8a8b 1048
1f202b0d 1049 if(c->avctx->me_threshold){
aaaa6f15 1050 vard= check_input_motion(s, mb_x, mb_y, 1);
115329f1 1051
aaaa6f15 1052 if((vard+128)>>8 < c->avctx->me_threshold){
dd7e46e7
MN		 1053 int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
1054 int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
aaaa6f15
MN		 1055 pic->mc_mb_var[s->mb_stride * mb_y + mb_x] = (vard+128)>>8;
1056 c->mc_mb_var_sum_temp += (vard+128)>>8;
dd7e46e7 1057 c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
f4f3223f
MN		 1058 return;
1059 }
aaaa6f15 1060 if((vard+128)>>8 < c->avctx->mb_threshold)
f20f8a8b 1061 mb_type= s->mb_type[mb_x + mb_y*s->mb_stride];
f4f3223f
MN		 1062 }
1063
e4986da9 1064 switch(s->me_method) {
de6d9b64
FB		 1065 case ME_ZERO:
1066 default:
bb270c08 1067 no_motion_search(s, &mx, &my);
0d21a846
MN		 1068 mx-= mb_x*16;
1069 my-= mb_y*16;
de6d9b64
FB		 1070 dmin = 0;
1071 break;
174489bd 1072 case ME_X1:
45870f57 1073 case ME_EPZS:
174489bd 1074 {
137c8468 1075 const int mot_stride = s->b8_stride;
ba6802de 1076 const int mot_xy = s->block_index[0];
174489bd 1077
8d7ec294
WH		 1078 P_LEFT[0] = s->current_picture.motion_val[0][mot_xy - 1][0];
1079 P_LEFT[1] = s->current_picture.motion_val[0][mot_xy - 1][1];
174489bd 1080
af4091f1 1081 if(P_LEFT[0] > (c->xmax<<shift)) P_LEFT[0] = (c->xmax<<shift);
c5b1c10a 1082
9c3d33d6 1083 if(!s->first_slice_line) {
8d7ec294
WH		 1084 P_TOP[0] = s->current_picture.motion_val[0][mot_xy - mot_stride ][0];
1085 P_TOP[1] = s->current_picture.motion_val[0][mot_xy - mot_stride ][1];
1086 P_TOPRIGHT[0] = s->current_picture.motion_val[0][mot_xy - mot_stride + 2][0];
1087 P_TOPRIGHT[1] = s->current_picture.motion_val[0][mot_xy - mot_stride + 2][1];
af4091f1
MN		 1088 if(P_TOP[1] > (c->ymax<<shift)) P_TOP[1] = (c->ymax<<shift);
1089 if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
1090 if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
115329f1 1091
0d21a846
MN		 1092 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1093 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1094
1095 if(s->out_format == FMT_H263){
2750b827
MN		 1096 c->pred_x = P_MEDIAN[0];
1097 c->pred_y = P_MEDIAN[1];
0d21a846 1098 }else { /* mpeg1 at least */
2750b827
MN		 1099 c->pred_x= P_LEFT[0];
1100 c->pred_y= P_LEFT[1];
0d21a846 1101 }
f931ff7b 1102 }else{
2750b827
MN		 1103 c->pred_x= P_LEFT[0];
1104 c->pred_y= P_LEFT[1];
174489bd 1105 }
f931ff7b 1106
c5b1c10a 1107 }
115329f1 1108 dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift, 0, 16);
2750b827 1109
45870f57 1110 break;
de6d9b64 1111 }
de6d9b64 1112
0d21a846 1113 /* At this point (mx,my) are full-pell and the relative displacement */
2750b827 1114 ppix = c->ref[0][0] + (my * s->linesize) + mx;
115329f1 1115
aaaa6f15 1116 vard = s->dsp.sse[0](NULL, pix, ppix, s->linesize, 16);
1fb4890b 1117
aaaa6f15 1118 pic->mc_mb_var[s->mb_stride * mb_y + mb_x] = (vard+128)>>8;
115329f1 1119// pic->mb_cmp_score[s->mb_stride * mb_y + mb_x] = dmin;
aaaa6f15 1120 c->mc_mb_var_sum_temp += (vard+128)>>8;
115329f1 1121
de6d9b64 1122#if 0
e03c341e 1123 printf("varc=%4d avg_var=%4d (sum=%4d) vard=%4d mx=%2d my=%2d\n",
bb270c08 1124 varc, s->avg_mb_var, sum, vard, mx - xx, my - yy);
de6d9b64 1125#endif
f20f8a8b 1126 if(mb_type){
dd7e46e7
MN		 1127 int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
1128 int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
1129 c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
f20f8a8b
MN		 1130
1131 if(mb_type == CANDIDATE_MB_TYPE_INTER){
af4091f1 1132 c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
f20f8a8b
MN		 1133 set_p_mv_tables(s, mx, my, 1);
1134 }else{
1135 mx <<=shift;
1136 my <<=shift;
1137 }
1138 if(mb_type == CANDIDATE_MB_TYPE_INTER4V){
1139 h263_mv4_search(s, mx, my, shift);
1140
1141 set_p_mv_tables(s, mx, my, 0);
1142 }
1143 if(mb_type == CANDIDATE_MB_TYPE_INTER_I){
1144 interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 1);
1145 }
1f202b0d 1146 }else if(c->avctx->mb_decision > FF_MB_DECISION_SIMPLE){
dd7e46e7
MN		 1147 int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
1148 int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
1149 c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
1fb4890b 1150
aaaa6f15 1151 if (vard*2 + 200*256 > varc)
bb198e19 1152 mb_type|= CANDIDATE_MB_TYPE_INTRA;
95da34cd
MN		 1153 if (varc*2 + 200*256 > vard || s->qscale > 24){
1154// if (varc*2 + 200*256 + 50*(s->lambda2>>FF_LAMBDA_SHIFT) > vard){
bb198e19 1155 mb_type|= CANDIDATE_MB_TYPE_INTER;
af4091f1 1156 c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
ca7d05d5
MN		 1157 if(s->flags&CODEC_FLAG_MV0)
1158 if(mx || my)
160d679c 1159 mb_type |= CANDIDATE_MB_TYPE_SKIPPED; //FIXME check difference
c60cf138 1160 }else{
1457ab52
MN		 1161 mx <<=shift;
1162 my <<=shift;
de6d9b64 1163 }
0d21a846 1164 if((s->flags&CODEC_FLAG_4MV)
aaaa6f15 1165 && !c->skip && varc>50<<8 && vard>10<<8){
bb198e19
MN		 1166 if(h263_mv4_search(s, mx, my, shift) < INT_MAX)
1167 mb_type|=CANDIDATE_MB_TYPE_INTER4V;
0d21a846
MN		 1168
1169 set_p_mv_tables(s, mx, my, 0);
1170 }else
1171 set_p_mv_tables(s, mx, my, 1);
bb198e19 1172 if((s->flags&CODEC_FLAG_INTERLACED_ME)
af4091f1 1173 && !c->skip){ //FIXME varc/d checks
f20f8a8b 1174 if(interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 0) < INT_MAX)
bb198e19
MN		 1175 mb_type |= CANDIDATE_MB_TYPE_INTER_I;
1176 }
ba6802de 1177 }else{
f7190f73 1178 int intra_score, i;
bb198e19 1179 mb_type= CANDIDATE_MB_TYPE_INTER;
0d21a846 1180
af4091f1 1181 dmin= c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1f202b0d 1182 if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
155ec6ed 1183 dmin= ff_get_mb_score(s, mx, my, 0, 0, 0, 16, 1);
67725183
MN		 1184
1185 if((s->flags&CODEC_FLAG_4MV)
aaaa6f15 1186 && !c->skip && varc>50<<8 && vard>10<<8){
bb198e19 1187 int dmin4= h263_mv4_search(s, mx, my, shift);
67725183 1188 if(dmin4 < dmin){
bb198e19 1189 mb_type= CANDIDATE_MB_TYPE_INTER4V;
67725183 1190 dmin=dmin4;
e4986da9 1191 }
67725183 1192 }
bb198e19 1193 if((s->flags&CODEC_FLAG_INTERLACED_ME)
af4091f1 1194 && !c->skip){ //FIXME varc/d checks
f20f8a8b 1195 int dmin_i= interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 0);
bb198e19
MN		 1196 if(dmin_i < dmin){
1197 mb_type = CANDIDATE_MB_TYPE_INTER_I;
1198 dmin= dmin_i;
1199 }
1200 }
115329f1
DB		 1201
1202// pic->mb_cmp_score[s->mb_stride * mb_y + mb_x] = dmin;
bb198e19 1203 set_p_mv_tables(s, mx, my, mb_type!=CANDIDATE_MB_TYPE_INTER4V);
f7190f73
MN		 1204
1205 /* get intra luma score */
1f202b0d 1206 if((c->avctx->mb_cmp&0xFF)==FF_CMP_SSE){
aaaa6f15 1207 intra_score= varc - 500;
f7190f73
MN		 1208 }else{
1209 int mean= (sum+128)>>8;
1210 mean*= 0x01010101;
115329f1 1211
f7190f73 1212 for(i=0; i<16; i++){
af4091f1
MN		 1213 *(uint32_t*)(&c->scratchpad[i*s->linesize+ 0]) = mean;
1214 *(uint32_t*)(&c->scratchpad[i*s->linesize+ 4]) = mean;
1215 *(uint32_t*)(&c->scratchpad[i*s->linesize+ 8]) = mean;
1216 *(uint32_t*)(&c->scratchpad[i*s->linesize+12]) = mean;
f7190f73
MN		 1217 }
1218
af4091f1 1219 intra_score= s->dsp.mb_cmp[0](s, c->scratchpad, pix, s->linesize, 16);
f7190f73
MN		 1220 }
1221#if 0 //FIXME
1222 /* get chroma score */
1f202b0d 1223 if(c->avctx->mb_cmp&FF_CMP_CHROMA){
f7190f73
MN		 1224 for(i=1; i<3; i++){
1225 uint8_t *dest_c;
1226 int mean;
115329f1 1227
f7190f73 1228 if(s->out_format == FMT_H263){
137c8468 1229 mean= (s->dc_val[i][mb_x + mb_y*s->b8_stride] + 4)>>3; //FIXME not exact but simple ;)
f7190f73
MN		 1230 }else{
1231 mean= (s->last_dc[i] + 4)>>3;
1232 }
1233 dest_c = s->new_picture.data[i] + (mb_y * 8 * (s->uvlinesize)) + mb_x * 8;
115329f1 1234
f7190f73
MN		 1235 mean*= 0x01010101;
1236 for(i=0; i<8; i++){
af4091f1
MN		 1237 *(uint32_t*)(&c->scratchpad[i*s->uvlinesize+ 0]) = mean;
1238 *(uint32_t*)(&c->scratchpad[i*s->uvlinesize+ 4]) = mean;
f7190f73 1239 }
115329f1 1240
af4091f1 1241 intra_score+= s->dsp.mb_cmp[1](s, c->scratchpad, dest_c, s->uvlinesize);
115329f1 1242 }
f7190f73
MN		 1243 }
1244#endif
af4091f1 1245 intra_score += c->mb_penalty_factor*16;
115329f1 1246
f7190f73 1247 if(intra_score < dmin){
bb198e19
MN		 1248 mb_type= CANDIDATE_MB_TYPE_INTRA;
1249 s->current_picture.mb_type[mb_y*s->mb_stride + mb_x]= CANDIDATE_MB_TYPE_INTRA; //FIXME cleanup
f7190f73
MN		 1250 }else
1251 s->current_picture.mb_type[mb_y*s->mb_stride + mb_x]= 0;
115329f1 1252
dd7e46e7
MN		 1253 {
1254 int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
1255 int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
1256 c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
ba6802de 1257 }
de6d9b64 1258 }
ba6802de 1259
7bc9090a 1260 s->mb_type[mb_y*s->mb_stride + mb_x]= mb_type;
9dbcbd92
MN		 1261}
1262
f5fb6b34
MN		 1263int ff_pre_estimate_p_frame_motion(MpegEncContext * s,
1264 int mb_x, int mb_y)
1265{
2750b827 1266 MotionEstContext * const c= &s->me;
bb198e19 1267 int mx, my, dmin;
f5fb6b34
MN		 1268 int P[10][2];
1269 const int shift= 1+s->quarter_sample;
7bc9090a 1270 const int xy= mb_x + mb_y*s->mb_stride;
af4091f1 1271 init_ref(c, s->new_picture.data, s->last_picture.data, NULL, 16*mb_x, 16*mb_y, 0);
115329f1 1272
f5fb6b34
MN		 1273 assert(s->quarter_sample==0 || s->quarter_sample==1);
1274
26efc54e 1275 c->pre_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_pre_cmp);
af4091f1 1276 c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
f5fb6b34 1277
bb198e19 1278 get_limits(s, 16*mb_x, 16*mb_y);
af4091f1 1279 c->skip=0;
f5fb6b34
MN		 1280
1281 P_LEFT[0] = s->p_mv_table[xy + 1][0];
1282 P_LEFT[1] = s->p_mv_table[xy + 1][1];
1283
af4091f1 1284 if(P_LEFT[0] < (c->xmin<<shift)) P_LEFT[0] = (c->xmin<<shift);
f5fb6b34
MN		 1285
1286 /* special case for first line */
9c3d33d6 1287 if (s->first_slice_line) {
2750b827
MN		 1288 c->pred_x= P_LEFT[0];
1289 c->pred_y= P_LEFT[1];
f931ff7b 1290 P_TOP[0]= P_TOPRIGHT[0]= P_MEDIAN[0]=
115329f1 1291 P_TOP[1]= P_TOPRIGHT[1]= P_MEDIAN[1]= 0; //FIXME
f5fb6b34 1292 } else {
7bc9090a
MN		 1293 P_TOP[0] = s->p_mv_table[xy + s->mb_stride ][0];
1294 P_TOP[1] = s->p_mv_table[xy + s->mb_stride ][1];
1295 P_TOPRIGHT[0] = s->p_mv_table[xy + s->mb_stride - 1][0];
1296 P_TOPRIGHT[1] = s->p_mv_table[xy + s->mb_stride - 1][1];
af4091f1
MN		 1297 if(P_TOP[1] < (c->ymin<<shift)) P_TOP[1] = (c->ymin<<shift);
1298 if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift);
1299 if(P_TOPRIGHT[1] < (c->ymin<<shift)) P_TOPRIGHT[1]= (c->ymin<<shift);
115329f1 1300
f5fb6b34
MN		 1301 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1302 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1303
2750b827
MN		 1304 c->pred_x = P_MEDIAN[0];
1305 c->pred_y = P_MEDIAN[1];
f5fb6b34 1306 }
2750b827 1307
115329f1 1308 dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift, 0, 16);
f931ff7b 1309
f5fb6b34
MN		 1310 s->p_mv_table[xy][0] = mx<<shift;
1311 s->p_mv_table[xy][1] = my<<shift;
115329f1 1312
f5fb6b34
MN		 1313 return dmin;
1314}
1315
5c91a675 1316static int ff_estimate_motion_b(MpegEncContext * s,
2750b827 1317 int mb_x, int mb_y, int16_t (*mv_table)[2], int ref_index, int f_code)
9dbcbd92 1318{
af4091f1 1319 MotionEstContext * const c= &s->me;
bb198e19 1320 int mx, my, dmin;
0d21a846 1321 int P[10][2];
9dbcbd92 1322 const int shift= 1+s->quarter_sample;
7bc9090a
MN		 1323 const int mot_stride = s->mb_stride;
1324 const int mot_xy = mb_y*mot_stride + mb_x;
af4091f1 1325 uint8_t * const mv_penalty= c->mv_penalty[f_code] + MAX_MV;
b07a5980 1326 int mv_scale;
115329f1 1327
26efc54e
MN		 1328 c->penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
1329 c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
1330 c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
af4091f1 1331 c->current_mv_penalty= mv_penalty;
1457ab52 1332
bb198e19 1333 get_limits(s, 16*mb_x, 16*mb_y);
9dbcbd92
MN		 1334
1335 switch(s->me_method) {
1336 case ME_ZERO:
1337 default:
bb270c08 1338 no_motion_search(s, &mx, &my);
9dbcbd92 1339 dmin = 0;
0d21a846
MN		 1340 mx-= mb_x*16;
1341 my-= mb_y*16;
9dbcbd92 1342 break;
9dbcbd92
MN		 1343 case ME_X1:
1344 case ME_EPZS:
1345 {
0d21a846
MN		 1346 P_LEFT[0] = mv_table[mot_xy - 1][0];
1347 P_LEFT[1] = mv_table[mot_xy - 1][1];
9dbcbd92 1348
af4091f1 1349 if(P_LEFT[0] > (c->xmax<<shift)) P_LEFT[0] = (c->xmax<<shift);
9dbcbd92
MN		 1350
1351 /* special case for first line */
9c3d33d6 1352 if (!s->first_slice_line) {
0d21a846
MN		 1353 P_TOP[0] = mv_table[mot_xy - mot_stride ][0];
1354 P_TOP[1] = mv_table[mot_xy - mot_stride ][1];
1355 P_TOPRIGHT[0] = mv_table[mot_xy - mot_stride + 1 ][0];
1356 P_TOPRIGHT[1] = mv_table[mot_xy - mot_stride + 1 ][1];
af4091f1
MN		 1357 if(P_TOP[1] > (c->ymax<<shift)) P_TOP[1]= (c->ymax<<shift);
1358 if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
1359 if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
115329f1 1360
0d21a846
MN		 1361 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1362 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
9dbcbd92 1363 }
af4091f1
MN		 1364 c->pred_x= P_LEFT[0];
1365 c->pred_y= P_LEFT[1];
9dbcbd92 1366 }
115329f1 1367
b07a5980
MN		 1368 if(mv_table == s->b_forw_mv_table){
1369 mv_scale= (s->pb_time<<16) / (s->pp_time<<shift);
1370 }else{
1371 mv_scale= ((s->pb_time - s->pp_time)<<16) / (s->pp_time<<shift);
1372 }
115329f1 1373
26efc54e 1374 dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, ref_index, s->p_mv_table, mv_scale, 0, 16);
115329f1 1375
9dbcbd92
MN		 1376 break;
1377 }
115329f1 1378
af4091f1 1379 dmin= c->sub_motion_search(s, &mx, &my, dmin, 0, ref_index, 0, 16);
115329f1 1380
1f202b0d 1381 if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
155ec6ed 1382 dmin= ff_get_mb_score(s, mx, my, 0, ref_index, 0, 16, 1);
67725183 1383
0d21a846 1384//printf("%d %d %d %d//", s->mb_x, s->mb_y, mx, my);
9dbcbd92
MN		 1385// s->mb_type[mb_y*s->mb_width + mb_x]= mb_type;
1386 mv_table[mot_xy][0]= mx;
1387 mv_table[mot_xy][1]= my;
1457ab52 1388
91029be7 1389 return dmin;
de6d9b64
FB		 1390}
1391
2750b827 1392static inline int check_bidir_mv(MpegEncContext * s,
91029be7
MN		 1393 int motion_fx, int motion_fy,
1394 int motion_bx, int motion_by,
1395 int pred_fx, int pred_fy,
bb198e19
MN		 1396 int pred_bx, int pred_by,
1397 int size, int h)
de6d9b64 1398{
91029be7 1399 //FIXME optimize?
1457ab52 1400 //FIXME better f_code prediction (max mv & distance)
bb198e19 1401 //FIXME pointers
2750b827 1402 MotionEstContext * const c= &s->me;
377798d6
MN		 1403 uint8_t * const mv_penalty_f= c->mv_penalty[s->f_code] + MAX_MV; // f_code of the prev frame
1404 uint8_t * const mv_penalty_b= c->mv_penalty[s->b_code] + MAX_MV; // f_code of the prev frame
af4091f1 1405 int stride= c->stride;
af4091f1 1406 uint8_t *dest_y = c->scratchpad;
91029be7
MN		 1407 uint8_t *ptr;
1408 int dxy;
1409 int src_x, src_y;
1410 int fbmin;
2750b827
MN		 1411 uint8_t **src_data= c->src[0];
1412 uint8_t **ref_data= c->ref[0];
1413 uint8_t **ref2_data= c->ref[2];
91029be7 1414
1457ab52
MN		 1415 if(s->quarter_sample){
1416 dxy = ((motion_fy & 3) << 2) | (motion_fx & 3);
bb198e19
MN		 1417 src_x = motion_fx >> 2;
1418 src_y = motion_fy >> 2;
1457ab52 1419
bb198e19
MN		 1420 ptr = ref_data[0] + (src_y * stride) + src_x;
1421 s->dsp.put_qpel_pixels_tab[0][dxy](dest_y , ptr , stride);
1457ab52
MN		 1422
1423 dxy = ((motion_by & 3) << 2) | (motion_bx & 3);
bb198e19
MN		 1424 src_x = motion_bx >> 2;
1425 src_y = motion_by >> 2;
115329f1 1426
2750b827 1427 ptr = ref2_data[0] + (src_y * stride) + src_x;
bb198e19 1428 s->dsp.avg_qpel_pixels_tab[size][dxy](dest_y , ptr , stride);
1457ab52
MN		 1429 }else{
1430 dxy = ((motion_fy & 1) << 1) | (motion_fx & 1);
bb198e19
MN		 1431 src_x = motion_fx >> 1;
1432 src_y = motion_fy >> 1;
1457ab52 1433
bb198e19
MN		 1434 ptr = ref_data[0] + (src_y * stride) + src_x;
1435 s->dsp.put_pixels_tab[size][dxy](dest_y , ptr , stride, h);
1457ab52
MN		 1436
1437 dxy = ((motion_by & 1) << 1) | (motion_bx & 1);
bb198e19
MN		 1438 src_x = motion_bx >> 1;
1439 src_y = motion_by >> 1;
115329f1 1440
2750b827 1441 ptr = ref2_data[0] + (src_y * stride) + src_x;
bb198e19 1442 s->dsp.avg_pixels_tab[size][dxy](dest_y , ptr , stride, h);
1457ab52
MN		 1443 }
1444
377798d6
MN		 1445 fbmin = (mv_penalty_f[motion_fx-pred_fx] + mv_penalty_f[motion_fy-pred_fy])*c->mb_penalty_factor
1446 +(mv_penalty_b[motion_bx-pred_bx] + mv_penalty_b[motion_by-pred_by])*c->mb_penalty_factor
bb198e19 1447 + s->dsp.mb_cmp[size](s, src_data[0], dest_y, stride, h); //FIXME new_pic
115329f1 1448
1f202b0d 1449 if(c->avctx->mb_cmp&FF_CMP_CHROMA){
67725183
MN		 1450 }
1451 //FIXME CHROMA !!!
115329f1 1452
91029be7
MN		 1453 return fbmin;
1454}
9dbcbd92 1455
91029be7 1456/* refine the bidir vectors in hq mode and return the score in both lq & hq mode*/
2750b827 1457static inline int bidir_refine(MpegEncContext * s, int mb_x, int mb_y)
91029be7 1458{
316a2ec8 1459 MotionEstContext * const c= &s->me;
7bc9090a
MN		 1460 const int mot_stride = s->mb_stride;
1461 const int xy = mb_y *mot_stride + mb_x;
91029be7
MN		 1462 int fbmin;
1463 int pred_fx= s->b_bidir_forw_mv_table[xy-1][0];
1464 int pred_fy= s->b_bidir_forw_mv_table[xy-1][1];
1465 int pred_bx= s->b_bidir_back_mv_table[xy-1][0];
1466 int pred_by= s->b_bidir_back_mv_table[xy-1][1];
1467 int motion_fx= s->b_bidir_forw_mv_table[xy][0]= s->b_forw_mv_table[xy][0];
1468 int motion_fy= s->b_bidir_forw_mv_table[xy][1]= s->b_forw_mv_table[xy][1];
1469 int motion_bx= s->b_bidir_back_mv_table[xy][0]= s->b_back_mv_table[xy][0];
1470 int motion_by= s->b_bidir_back_mv_table[xy][1]= s->b_back_mv_table[xy][1];
316a2ec8
MN		 1471 const int flags= c->sub_flags;
1472 const int qpel= flags&FLAG_QPEL;
1473 const int shift= 1+qpel;
1474 const int xmin= c->xmin<<shift;
1475 const int ymin= c->ymin<<shift;
1476 const int xmax= c->xmax<<shift;
1477 const int ymax= c->ymax<<shift;
a69220cc
MN		 1478#define HASH(fx,fy,bx,by) ((fx)+17*(fy)+63*(bx)+117*(by))
1479 int hashidx= HASH(motion_fx,motion_fy, motion_bx, motion_by);
2d603902 1480 uint8_t map[256];
8bf755d4
LM		 1481
1482 memset(map,0,sizeof(map));
a69220cc 1483 map[hashidx&255] = 1;
115329f1 1484
2750b827 1485 fbmin= check_bidir_mv(s, motion_fx, motion_fy,
91029be7
MN		 1486 motion_bx, motion_by,
1487 pred_fx, pred_fy,
bb198e19
MN		 1488 pred_bx, pred_by,
1489 0, 16);
91029be7 1490
316a2ec8
MN		 1491 if(s->avctx->bidir_refine){
1492 int score, end;
1493#define CHECK_BIDIR(fx,fy,bx,by)\
a69220cc 1494 if( !map[(hashidx+HASH(fx,fy,bx,by))&255]\
8bf755d4 1495 &&(fx<=0 || motion_fx+fx<=xmax) && (fy<=0 || motion_fy+fy<=ymax) && (bx<=0 || motion_bx+bx<=xmax) && (by<=0 || motion_by+by<=ymax)\
d4f0c2fc 1496 &&(fx>=0 || motion_fx+fx>=xmin) && (fy>=0 || motion_fy+fy>=ymin) && (bx>=0 || motion_bx+bx>=xmin) && (by>=0 || motion_by+by>=ymin)){\
a69220cc 1497 map[(hashidx+HASH(fx,fy,bx,by))&255] = 1;\
d4f0c2fc
MN		 1498 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);\
1499 if(score < fbmin){\
a69220cc 1500 hashidx += HASH(fx,fy,bx,by);\
d4f0c2fc
MN		 1501 fbmin= score;\
1502 motion_fx+=fx;\
1503 motion_fy+=fy;\
1504 motion_bx+=bx;\
1505 motion_by+=by;\
1506 end=0;\
1507 }\
316a2ec8
MN		 1508 }
1509#define CHECK_BIDIR2(a,b,c,d)\
1510CHECK_BIDIR(a,b,c,d)\
8226ecaa 1511CHECK_BIDIR(-(a),-(b),-(c),-(d))
316a2ec8
MN		 1512
1513#define CHECK_BIDIRR(a,b,c,d)\
1514CHECK_BIDIR2(a,b,c,d)\
1515CHECK_BIDIR2(b,c,d,a)\
1516CHECK_BIDIR2(c,d,a,b)\
1517CHECK_BIDIR2(d,a,b,c)
1518
1519 do{
1520 end=1;
1521
316a2ec8
MN		 1522 CHECK_BIDIRR( 0, 0, 0, 1)
1523 if(s->avctx->bidir_refine > 1){
1524 CHECK_BIDIRR( 0, 0, 1, 1)
1525 CHECK_BIDIR2( 0, 1, 0, 1)
1526 CHECK_BIDIR2( 1, 0, 1, 0)
1527 CHECK_BIDIRR( 0, 0,-1, 1)
1528 CHECK_BIDIR2( 0,-1, 0, 1)
1529 CHECK_BIDIR2(-1, 0, 1, 0)
1530 if(s->avctx->bidir_refine > 2){
1531 CHECK_BIDIRR( 0, 1, 1, 1)
1532 CHECK_BIDIRR( 0,-1, 1, 1)
1533 CHECK_BIDIRR( 0, 1,-1, 1)
1534 CHECK_BIDIRR( 0, 1, 1,-1)
1535 if(s->avctx->bidir_refine > 3){
1536 CHECK_BIDIR2( 1, 1, 1, 1)
1537 CHECK_BIDIRR( 1, 1, 1,-1)
1538 CHECK_BIDIR2( 1, 1,-1,-1)
1539 CHECK_BIDIR2( 1,-1,-1, 1)
1540 CHECK_BIDIR2( 1,-1, 1,-1)
1541 }
1542 }
1543 }
1544 }while(!end);
1545 }
1546
8bf755d4
LM		 1547 s->b_bidir_forw_mv_table[xy][0]= motion_fx;
1548 s->b_bidir_forw_mv_table[xy][1]= motion_fy;
1549 s->b_bidir_back_mv_table[xy][0]= motion_bx;
1550 s->b_bidir_back_mv_table[xy][1]= motion_by;
1551
316a2ec8 1552 return fbmin;
91029be7
MN		 1553}
1554
2750b827 1555static inline int direct_search(MpegEncContext * s, int mb_x, int mb_y)
91029be7 1556{
af4091f1 1557 MotionEstContext * const c= &s->me;
0d21a846 1558 int P[10][2];
7bc9090a
MN		 1559 const int mot_stride = s->mb_stride;
1560 const int mot_xy = mb_y*mot_stride + mb_x;
1457ab52
MN		 1561 const int shift= 1+s->quarter_sample;
1562 int dmin, i;
91029be7 1563 const int time_pp= s->pp_time;
bea669e5 1564 const int time_pb= s->pb_time;
1457ab52 1565 int mx, my, xmin, xmax, ymin, ymax;
91029be7 1566 int16_t (*mv_table)[2]= s->b_direct_mv_table;
115329f1 1567
af4091f1 1568 c->current_mv_penalty= c->mv_penalty[1] + MAX_MV;
1457ab52
MN		 1569 ymin= xmin=(-32)>>shift;
1570 ymax= xmax= 31>>shift;
1571
7bc9090a 1572 if(IS_8X8(s->next_picture.mb_type[mot_xy])){
1457ab52
MN		 1573 s->mv_type= MV_TYPE_8X8;
1574 }else{
1575 s->mv_type= MV_TYPE_16X16;
91029be7 1576 }
1457ab52
MN		 1577
1578 for(i=0; i<4; i++){
1579 int index= s->block_index[i];
1580 int min, max;
115329f1 1581
af4091f1
MN		 1582 c->co_located_mv[i][0]= s->next_picture.motion_val[0][index][0];
1583 c->co_located_mv[i][1]= s->next_picture.motion_val[0][index][1];
1584 c->direct_basis_mv[i][0]= c->co_located_mv[i][0]*time_pb/time_pp + ((i& 1)<<(shift+3));
1585 c->direct_basis_mv[i][1]= c->co_located_mv[i][1]*time_pb/time_pp + ((i>>1)<<(shift+3));
1586// c->direct_basis_mv[1][i][0]= c->co_located_mv[i][0]*(time_pb - time_pp)/time_pp + ((i &1)<<(shift+3);
1587// c->direct_basis_mv[1][i][1]= c->co_located_mv[i][1]*(time_pb - time_pp)/time_pp + ((i>>1)<<(shift+3);
1588
1589 max= FFMAX(c->direct_basis_mv[i][0], c->direct_basis_mv[i][0] - c->co_located_mv[i][0])>>shift;
1590 min= FFMIN(c->direct_basis_mv[i][0], c->direct_basis_mv[i][0] - c->co_located_mv[i][0])>>shift;
5a603607
MN		 1591 max+= 16*mb_x + 1; // +-1 is for the simpler rounding
1592 min+= 16*mb_x - 1;
4a711c33
MN		 1593 xmax= FFMIN(xmax, s->width - max);
1594 xmin= FFMAX(xmin, - 16 - min);
1457ab52 1595
af4091f1
MN		 1596 max= FFMAX(c->direct_basis_mv[i][1], c->direct_basis_mv[i][1] - c->co_located_mv[i][1])>>shift;
1597 min= FFMIN(c->direct_basis_mv[i][1], c->direct_basis_mv[i][1] - c->co_located_mv[i][1])>>shift;
5a603607
MN		 1598 max+= 16*mb_y + 1; // +-1 is for the simpler rounding
1599 min+= 16*mb_y - 1;
4a711c33
MN		 1600 ymax= FFMIN(ymax, s->height - max);
1601 ymin= FFMAX(ymin, - 16 - min);
115329f1 1602
1457ab52 1603 if(s->mv_type == MV_TYPE_16X16) break;
91029be7 1604 }
115329f1 1605
1457ab52 1606 assert(xmax <= 15 && ymax <= 15 && xmin >= -16 && ymin >= -16);
115329f1 1607
1457ab52
MN		 1608 if(xmax < 0 || xmin >0 || ymax < 0 || ymin > 0){
1609 s->b_direct_mv_table[mot_xy][0]= 0;
1610 s->b_direct_mv_table[mot_xy][1]= 0;
1611
1612 return 256*256*256*64;
1613 }
115329f1 1614
af4091f1
MN		 1615 c->xmin= xmin;
1616 c->ymin= ymin;
1617 c->xmax= xmax;
1618 c->ymax= ymax;
1619 c->flags |= FLAG_DIRECT;
1620 c->sub_flags |= FLAG_DIRECT;
1621 c->pred_x=0;
1622 c->pred_y=0;
1457ab52 1623
f66e4f5f
RD		 1624 P_LEFT[0] = av_clip(mv_table[mot_xy - 1][0], xmin<<shift, xmax<<shift);
1625 P_LEFT[1] = av_clip(mv_table[mot_xy - 1][1], ymin<<shift, ymax<<shift);
0a13093d
MN		 1626
1627 /* special case for first line */
90b5b51e 1628 if (!s->first_slice_line) { //FIXME maybe allow this over thread boundary as it is clipped
f66e4f5f
RD		 1629 P_TOP[0] = av_clip(mv_table[mot_xy - mot_stride ][0], xmin<<shift, xmax<<shift);
1630 P_TOP[1] = av_clip(mv_table[mot_xy - mot_stride ][1], ymin<<shift, ymax<<shift);
1631 P_TOPRIGHT[0] = av_clip(mv_table[mot_xy - mot_stride + 1 ][0], xmin<<shift, xmax<<shift);
1632 P_TOPRIGHT[1] = av_clip(mv_table[mot_xy - mot_stride + 1 ][1], ymin<<shift, ymax<<shift);
115329f1 1633
0a13093d
MN		 1634 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1635 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1636 }
115329f1 1637
26efc54e 1638 dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, mv_table, 1<<(16-shift), 0, 16);
115329f1 1639 if(c->sub_flags&FLAG_QPEL)
2750b827
MN		 1640 dmin = qpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1641 else
1642 dmin = hpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
115329f1 1643
1f202b0d 1644 if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
155ec6ed 1645 dmin= ff_get_mb_score(s, mx, my, 0, 0, 0, 16, 1);
115329f1 1646
af4091f1 1647 get_limits(s, 16*mb_x, 16*mb_y); //restore c->?min/max, maybe not needed
91029be7 1648
2175b80b
MN		 1649 mv_table[mot_xy][0]= mx;
1650 mv_table[mot_xy][1]= my;
af4091f1
MN		 1651 c->flags &= ~FLAG_DIRECT;
1652 c->sub_flags &= ~FLAG_DIRECT;
2750b827 1653
91029be7 1654 return dmin;
9dbcbd92
MN		 1655}
1656
1657void ff_estimate_b_frame_motion(MpegEncContext * s,
1658 int mb_x, int mb_y)
1659{
af4091f1
MN		 1660 MotionEstContext * const c= &s->me;
1661 const int penalty_factor= c->mb_penalty_factor;
bb198e19 1662 int fmin, bmin, dmin, fbmin, bimin, fimin;
91029be7 1663 int type=0;
f20f8a8b 1664 const int xy = mb_y*s->mb_stride + mb_x;
af4091f1 1665 init_ref(c, s->new_picture.data, s->last_picture.data, s->next_picture.data, 16*mb_x, 16*mb_y, 2);
f20f8a8b 1666
ae1dbde1 1667 get_limits(s, 16*mb_x, 16*mb_y);
115329f1 1668
af4091f1 1669 c->skip=0;
2f16af06
MN		 1670
1671 if(s->codec_id == CODEC_ID_MPEG4 && s->next_picture.mbskip_table[xy]){
1672 int score= direct_search(s, mb_x, mb_y); //FIXME just check 0,0
1673
1674 score= ((unsigned)(score*score + 128*256))>>16;
1675 c->mc_mb_var_sum_temp += score;
1676 s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE
1677 s->mb_type[mb_y*s->mb_stride + mb_x]= CANDIDATE_MB_TYPE_DIRECT0;
1678
1679 return;
1680 }
1681
1f202b0d 1682 if(c->avctx->me_threshold){
aaaa6f15 1683 int vard= check_input_motion(s, mb_x, mb_y, 0);
115329f1 1684
aaaa6f15 1685 if((vard+128)>>8 < c->avctx->me_threshold){
f4f3223f
MN		 1686// pix = c->src[0][0];
1687// sum = s->dsp.pix_sum(pix, s->linesize);
aaaa6f15 1688// varc = s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)(sum*sum))>>8) + 500;
115329f1 1689
aaaa6f15
MN		 1690// pic->mb_var [s->mb_stride * mb_y + mb_x] = (varc+128)>>8;
1691 s->current_picture.mc_mb_var[s->mb_stride * mb_y + mb_x] = (vard+128)>>8;
f4f3223f 1692/* pic->mb_mean [s->mb_stride * mb_y + mb_x] = (sum+128)>>8;
aaaa6f15
MN		 1693 c->mb_var_sum_temp += (varc+128)>>8;*/
1694 c->mc_mb_var_sum_temp += (vard+128)>>8;
1695/* if (vard <= 64<<8 || vard < varc) {
1696 c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
f4f3223f 1697 }else{
39a64e30 1698 c->scene_change_score+= s->qscale * s->avctx->scenechange_factor;
f4f3223f
MN		 1699 }*/
1700 return;
1701 }
aaaa6f15 1702 if((vard+128)>>8 < c->avctx->mb_threshold){
f20f8a8b
MN		 1703 type= s->mb_type[mb_y*s->mb_stride + mb_x];
1704 if(type == CANDIDATE_MB_TYPE_DIRECT){
1705 direct_search(s, mb_x, mb_y);
1706 }
1707 if(type == CANDIDATE_MB_TYPE_FORWARD || type == CANDIDATE_MB_TYPE_BIDIR){
af4091f1 1708 c->skip=0;
f20f8a8b
MN		 1709 ff_estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, 0, s->f_code);
1710 }
1711 if(type == CANDIDATE_MB_TYPE_BACKWARD || type == CANDIDATE_MB_TYPE_BIDIR){
af4091f1 1712 c->skip=0;
f20f8a8b
MN		 1713 ff_estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, 2, s->b_code);
1714 }
1715 if(type == CANDIDATE_MB_TYPE_FORWARD_I || type == CANDIDATE_MB_TYPE_BIDIR_I){
af4091f1
MN		 1716 c->skip=0;
1717 c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
f20f8a8b
MN		 1718 interlaced_search(s, 0,
1719 s->b_field_mv_table[0], s->b_field_select_table[0],
1720 s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1], 1);
1721 }
1722 if(type == CANDIDATE_MB_TYPE_BACKWARD_I || type == CANDIDATE_MB_TYPE_BIDIR_I){
af4091f1
MN		 1723 c->skip=0;
1724 c->current_mv_penalty= c->mv_penalty[s->b_code] + MAX_MV;
f20f8a8b
MN		 1725 interlaced_search(s, 2,
1726 s->b_field_mv_table[1], s->b_field_select_table[1],
1727 s->b_back_mv_table[xy][0], s->b_back_mv_table[xy][1], 1);
1728 }
1729 return;
1730 }
f4f3223f
MN		 1731 }
1732
2d5e962b 1733 if (s->codec_id == CODEC_ID_MPEG4)
2750b827 1734 dmin= direct_search(s, mb_x, mb_y);
2d5e962b
MN		 1735 else
1736 dmin= INT_MAX;
bb198e19 1737//FIXME penalty stuff for non mpeg4
af4091f1 1738 c->skip=0;
2750b827 1739 fmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, 0, s->f_code) + 3*penalty_factor;
115329f1 1740
af4091f1 1741 c->skip=0;
2750b827 1742 bmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, 2, s->b_code) + 2*penalty_factor;
9dbcbd92 1743//printf(" %d %d ", s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1]);
91029be7 1744
af4091f1 1745 c->skip=0;
2750b827 1746 fbmin= bidir_refine(s, mb_x, mb_y) + penalty_factor;
67725183 1747//printf("%d %d %d %d\n", dmin, fmin, bmin, fbmin);
115329f1 1748
bb198e19 1749 if(s->flags & CODEC_FLAG_INTERLACED_ME){
bb198e19 1750//FIXME mb type penalty
af4091f1
MN		 1751 c->skip=0;
1752 c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
2750b827
MN		 1753 fimin= interlaced_search(s, 0,
1754 s->b_field_mv_table[0], s->b_field_select_table[0],
f20f8a8b 1755 s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1], 0);
af4091f1 1756 c->current_mv_penalty= c->mv_penalty[s->b_code] + MAX_MV;
2750b827
MN		 1757 bimin= interlaced_search(s, 2,
1758 s->b_field_mv_table[1], s->b_field_select_table[1],
f20f8a8b 1759 s->b_back_mv_table[xy][0], s->b_back_mv_table[xy][1], 0);
bb198e19
MN		 1760 }else
1761 fimin= bimin= INT_MAX;
1762
3aa102be 1763 {
63b15e55 1764 int score= fmin;
bb198e19 1765 type = CANDIDATE_MB_TYPE_FORWARD;
115329f1 1766
2d5e962b 1767 if (dmin <= score){
63b15e55 1768 score = dmin;
bb198e19 1769 type = CANDIDATE_MB_TYPE_DIRECT;
91029be7
MN		 1770 }
1771 if(bmin<score){
1772 score=bmin;
115329f1 1773 type= CANDIDATE_MB_TYPE_BACKWARD;
91029be7
MN		 1774 }
1775 if(fbmin<score){
1776 score=fbmin;
bb198e19
MN		 1777 type= CANDIDATE_MB_TYPE_BIDIR;
1778 }
1779 if(fimin<score){
1780 score=fimin;
1781 type= CANDIDATE_MB_TYPE_FORWARD_I;
1782 }
1783 if(bimin<score){
1784 score=bimin;
1785 type= CANDIDATE_MB_TYPE_BACKWARD_I;
91029be7 1786 }
115329f1 1787
0bfacb95 1788 score= ((unsigned)(score*score + 128*256))>>16;
1f202b0d 1789 c->mc_mb_var_sum_temp += score;
7bc9090a 1790 s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE
91029be7 1791 }
3aa102be 1792
1f202b0d 1793 if(c->avctx->mb_decision > FF_MB_DECISION_SIMPLE){
bb198e19
MN		 1794 type= CANDIDATE_MB_TYPE_FORWARD | CANDIDATE_MB_TYPE_BACKWARD | CANDIDATE_MB_TYPE_BIDIR | CANDIDATE_MB_TYPE_DIRECT;
1795 if(fimin < INT_MAX)
1796 type |= CANDIDATE_MB_TYPE_FORWARD_I;
1797 if(bimin < INT_MAX)
1798 type |= CANDIDATE_MB_TYPE_BACKWARD_I;
1799 if(fimin < INT_MAX && bimin < INT_MAX){
1800 type |= CANDIDATE_MB_TYPE_BIDIR_I;
1801 }
1802 //FIXME something smarter
755bfeab 1803 if(dmin>256*256*16) type&= ~CANDIDATE_MB_TYPE_DIRECT; //do not try direct mode if it is invalid for this MB
d951bb9a
MN		 1804 if(s->codec_id == CODEC_ID_MPEG4 && type&CANDIDATE_MB_TYPE_DIRECT && s->flags&CODEC_FLAG_MV0 && *(uint32_t*)s->b_direct_mv_table[xy])
1805 type |= CANDIDATE_MB_TYPE_DIRECT0;
115329f1 1806#if 0
622348f9
MN		 1807 if(s->out_format == FMT_MPEG1)
1808 type |= CANDIDATE_MB_TYPE_INTRA;
1809#endif
3aa102be
MN		 1810 }
1811
7bc9090a 1812 s->mb_type[mb_y*s->mb_stride + mb_x]= type;
9dbcbd92
MN		 1813}
1814
1815/* find best f_code for ME which do unlimited searches */
1816int ff_get_best_fcode(MpegEncContext * s, int16_t (*mv_table)[2], int type)
1817{
9dbcbd92 1818 if(s->me_method>=ME_EPZS){
0d21a846 1819 int score[8];
0cc64d3d 1820 int i, y, range= s->avctx->me_range ? s->avctx->me_range : (INT_MAX/2);
0c1a9eda 1821 uint8_t * fcode_tab= s->fcode_tab;
0d21a846
MN		 1822 int best_fcode=-1;
1823 int best_score=-10000000;
9dbcbd92 1824
115329f1 1825 if(s->msmpeg4_version)
2f300f89
MN		 1826 range= FFMIN(range, 16);
1827 else if(s->codec_id == CODEC_ID_MPEG2VIDEO && s->avctx->strict_std_compliance >= FF_COMPLIANCE_NORMAL)
1828 range= FFMIN(range, 256);
1829
1457ab52 1830 for(i=0; i<8; i++) score[i]= s->mb_num*(8-i);
9dbcbd92
MN		 1831
1832 for(y=0; y<s->mb_height; y++){
1833 int x;
7bc9090a 1834 int xy= y*s->mb_stride;
9dbcbd92 1835 for(x=0; x<s->mb_width; x++){
d3985de7 1836 if(s->mb_type[xy] & type){
7494cac0
MN		 1837 int mx= mv_table[xy][0];
1838 int my= mv_table[xy][1];
1839 int fcode= FFMAX(fcode_tab[mx + MAX_MV],
1840 fcode_tab[my + MAX_MV]);
0d21a846 1841 int j;
115329f1
DB		 1842
1843 if(mx >= range || mx < -range ||
7494cac0
MN		 1844 my >= range || my < -range)
1845 continue;
115329f1 1846
0d21a846 1847 for(j=0; j<fcode && j<8; j++){
9701840b 1848 if(s->pict_type==FF_B_TYPE || s->current_picture.mc_mb_var[xy] < s->current_picture.mb_var[xy])
0d21a846
MN		 1849 score[j]-= 170;
1850 }
9dbcbd92 1851 }
9dbcbd92
MN		 1852 xy++;
1853 }
1854 }
115329f1 1855
0d21a846
MN		 1856 for(i=1; i<8; i++){
1857 if(score[i] > best_score){
1858 best_score= score[i];
1859 best_fcode= i;
1860 }
1861// printf("%d %d\n", i, score[i]);
9dbcbd92 1862 }
0d21a846 1863
9dbcbd92 1864// printf("fcode: %d type: %d\n", i, s->pict_type);
0d21a846 1865 return best_fcode;
9dbcbd92
MN		 1866/* for(i=0; i<=MAX_FCODE; i++){
1867 printf("%d ", mv_num[i]);
1868 }
1869 printf("\n");*/
1870 }else{
1871 return 1;
de6d9b64
FB		 1872 }
1873}
1874
9dbcbd92
MN		 1875void ff_fix_long_p_mvs(MpegEncContext * s)
1876{
1f202b0d 1877 MotionEstContext * const c= &s->me;
9dbcbd92 1878 const int f_code= s->f_code;
ebbcdc9a 1879 int y, range;
9701840b 1880 assert(s->pict_type==FF_P_TYPE);
ebbcdc9a 1881
2f300f89
MN		 1882 range = (((s->out_format == FMT_MPEG1 || s->msmpeg4_version) ? 8 : 16) << f_code);
1883
1884 assert(range <= 16 || !s->msmpeg4_version);
1885 assert(range <=256 || !(s->codec_id == CODEC_ID_MPEG2VIDEO && s->avctx->strict_std_compliance >= FF_COMPLIANCE_NORMAL));
115329f1 1886
1f202b0d 1887 if(c->avctx->me_range && range > c->avctx->me_range) range= c->avctx->me_range;
115329f1 1888
0d21a846 1889//printf("%d no:%d %d//\n", clip, noclip, f_code);
9dbcbd92 1890 if(s->flags&CODEC_FLAG_4MV){
137c8468 1891 const int wrap= s->b8_stride;
9dbcbd92
MN		 1892
1893 /* clip / convert to intra 8x8 type MVs */
1894 for(y=0; y<s->mb_height; y++){
137c8468 1895 int xy= y*2*wrap;
7bc9090a 1896 int i= y*s->mb_stride;
9dbcbd92
MN		 1897 int x;
1898
1899 for(x=0; x<s->mb_width; x++){
bb198e19 1900 if(s->mb_type[i]&CANDIDATE_MB_TYPE_INTER4V){
9dbcbd92
MN		 1901 int block;
1902 for(block=0; block<4; block++){
1903 int off= (block& 1) + (block>>1)*wrap;
8d7ec294
WH		 1904 int mx= s->current_picture.motion_val[0][ xy + off ][0];
1905 int my= s->current_picture.motion_val[0][ xy + off ][1];
9dbcbd92 1906
ebbcdc9a
MN		 1907 if( mx >=range || mx <-range
1908 || my >=range || my <-range){
bb198e19
MN		 1909 s->mb_type[i] &= ~CANDIDATE_MB_TYPE_INTER4V;
1910 s->mb_type[i] |= CANDIDATE_MB_TYPE_INTRA;
1911 s->current_picture.mb_type[i]= CANDIDATE_MB_TYPE_INTRA;
9dbcbd92
MN		 1912 }
1913 }
9dbcbd92 1914 }
f10bd870
MN		 1915 xy+=2;
1916 i++;
9dbcbd92
MN		 1917 }
1918 }
1919 }
1920}
1921
bb198e19
MN		 1922/**
1923 *
1924 * @param truncate 1 for truncation, 0 for using intra
1925 */
115329f1 1926void ff_fix_long_mvs(MpegEncContext * s, uint8_t *field_select_table, int field_select,
bb198e19 1927 int16_t (*mv_table)[2], int f_code, int type, int truncate)
9dbcbd92 1928{
1f202b0d 1929 MotionEstContext * const c= &s->me;
bb198e19 1930 int y, h_range, v_range;
9dbcbd92 1931
63b15e55 1932 // RAL: 8 in MPEG-1, 16 in MPEG-4
2f300f89 1933 int range = (((s->out_format == FMT_MPEG1 || s->msmpeg4_version) ? 8 : 16) << f_code);
bb198e19 1934
1f202b0d 1935 if(c->avctx->me_range && range > c->avctx->me_range) range= c->avctx->me_range;
63b15e55 1936
bb198e19
MN		 1937 h_range= range;
1938 v_range= field_select_table ? range>>1 : range;
1939
9dbcbd92
MN		 1940 /* clip / convert to intra 16x16 type MVs */
1941 for(y=0; y<s->mb_height; y++){
1942 int x;
7bc9090a 1943 int xy= y*s->mb_stride;
ebbcdc9a 1944 for(x=0; x<s->mb_width; x++){
7bc9090a 1945 if (s->mb_type[xy] & type){ // RAL: "type" test added...
bb198e19
MN		 1946 if(field_select_table==NULL || field_select_table[xy] == field_select){
1947 if( mv_table[xy][0] >=h_range || mv_table[xy][0] <-h_range
1948 || mv_table[xy][1] >=v_range || mv_table[xy][1] <-v_range){
1949
1950 if(truncate){
1951 if (mv_table[xy][0] > h_range-1) mv_table[xy][0]= h_range-1;
1952 else if(mv_table[xy][0] < -h_range ) mv_table[xy][0]= -h_range;
1953 if (mv_table[xy][1] > v_range-1) mv_table[xy][1]= v_range-1;
1954 else if(mv_table[xy][1] < -v_range ) mv_table[xy][1]= -v_range;
1955 }else{
1956 s->mb_type[xy] &= ~type;
1957 s->mb_type[xy] |= CANDIDATE_MB_TYPE_INTRA;
1958 mv_table[xy][0]=
1959 mv_table[xy][1]= 0;
1960 }
63b15e55 1961 }
ebbcdc9a 1962 }
9dbcbd92
MN		 1963 }
1964 xy++;
9dbcbd92
MN		 1965 }
1966 }
1967}
Libav master